~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/arch/ia64/kernel/perfmon.c

Version: ~ [ linux-4.15-rc3 ] ~ [ linux-4.14.5 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.68 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.105 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.47 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.87 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.51 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.96 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.27.62 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * This file implements the perfmon-2 subsystem which is used
  3  * to program the IA-64 Performance Monitoring Unit (PMU).
  4  *
  5  * The initial version of perfmon.c was written by
  6  * Ganesh Venkitachalam, IBM Corp.
  7  *
  8  * Then it was modified for perfmon-1.x by Stephane Eranian and
  9  * David Mosberger, Hewlett Packard Co.
 10  *
 11  * Version Perfmon-2.x is a rewrite of perfmon-1.x
 12  * by Stephane Eranian, Hewlett Packard Co.
 13  *
 14  * Copyright (C) 1999-2005  Hewlett Packard Co
 15  *               Stephane Eranian <eranian@hpl.hp.com>
 16  *               David Mosberger-Tang <davidm@hpl.hp.com>
 17  *
 18  * More information about perfmon available at:
 19  *      http://www.hpl.hp.com/research/linux/perfmon
 20  */
 21 
 22 #include <linux/module.h>
 23 #include <linux/kernel.h>
 24 #include <linux/sched.h>
 25 #include <linux/sched/task.h>
 26 #include <linux/sched/task_stack.h>
 27 #include <linux/interrupt.h>
 28 #include <linux/proc_fs.h>
 29 #include <linux/seq_file.h>
 30 #include <linux/init.h>
 31 #include <linux/vmalloc.h>
 32 #include <linux/mm.h>
 33 #include <linux/sysctl.h>
 34 #include <linux/list.h>
 35 #include <linux/file.h>
 36 #include <linux/poll.h>
 37 #include <linux/vfs.h>
 38 #include <linux/smp.h>
 39 #include <linux/pagemap.h>
 40 #include <linux/mount.h>
 41 #include <linux/bitops.h>
 42 #include <linux/capability.h>
 43 #include <linux/rcupdate.h>
 44 #include <linux/completion.h>
 45 #include <linux/tracehook.h>
 46 #include <linux/slab.h>
 47 #include <linux/cpu.h>
 48 
 49 #include <asm/errno.h>
 50 #include <asm/intrinsics.h>
 51 #include <asm/page.h>
 52 #include <asm/perfmon.h>
 53 #include <asm/processor.h>
 54 #include <asm/signal.h>
 55 #include <linux/uaccess.h>
 56 #include <asm/delay.h>
 57 
 58 #ifdef CONFIG_PERFMON
 59 /*
 60  * perfmon context state
 61  */
 62 #define PFM_CTX_UNLOADED        1       /* context is not loaded onto any task */
 63 #define PFM_CTX_LOADED          2       /* context is loaded onto a task */
 64 #define PFM_CTX_MASKED          3       /* context is loaded but monitoring is masked due to overflow */
 65 #define PFM_CTX_ZOMBIE          4       /* owner of the context is closing it */
 66 
 67 #define PFM_INVALID_ACTIVATION  (~0UL)
 68 
 69 #define PFM_NUM_PMC_REGS        64      /* PMC save area for ctxsw */
 70 #define PFM_NUM_PMD_REGS        64      /* PMD save area for ctxsw */
 71 
 72 /*
 73  * depth of message queue
 74  */
 75 #define PFM_MAX_MSGS            32
 76 #define PFM_CTXQ_EMPTY(g)       ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
 77 
 78 /*
 79  * type of a PMU register (bitmask).
 80  * bitmask structure:
 81  *      bit0   : register implemented
 82  *      bit1   : end marker
 83  *      bit2-3 : reserved
 84  *      bit4   : pmc has pmc.pm
 85  *      bit5   : pmc controls a counter (has pmc.oi), pmd is used as counter
 86  *      bit6-7 : register type
 87  *      bit8-31: reserved
 88  */
 89 #define PFM_REG_NOTIMPL         0x0 /* not implemented at all */
 90 #define PFM_REG_IMPL            0x1 /* register implemented */
 91 #define PFM_REG_END             0x2 /* end marker */
 92 #define PFM_REG_MONITOR         (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
 93 #define PFM_REG_COUNTING        (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
 94 #define PFM_REG_CONTROL         (0x4<<4|PFM_REG_IMPL) /* PMU control register */
 95 #define PFM_REG_CONFIG          (0x8<<4|PFM_REG_IMPL) /* configuration register */
 96 #define PFM_REG_BUFFER          (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
 97 
 98 #define PMC_IS_LAST(i)  (pmu_conf->pmc_desc[i].type & PFM_REG_END)
 99 #define PMD_IS_LAST(i)  (pmu_conf->pmd_desc[i].type & PFM_REG_END)
100 
101 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags &  PFM_REGFL_OVFL_NOTIFY)
102 
103 /* i assumed unsigned */
104 #define PMC_IS_IMPL(i)    (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
105 #define PMD_IS_IMPL(i)    (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
106 
107 /* XXX: these assume that register i is implemented */
108 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
109 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
110 #define PMC_IS_MONITOR(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR)  == PFM_REG_MONITOR)
111 #define PMC_IS_CONTROL(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL)  == PFM_REG_CONTROL)
112 
113 #define PMC_DFL_VAL(i)     pmu_conf->pmc_desc[i].default_value
114 #define PMC_RSVD_MASK(i)   pmu_conf->pmc_desc[i].reserved_mask
115 #define PMD_PMD_DEP(i)     pmu_conf->pmd_desc[i].dep_pmd[0]
116 #define PMC_PMD_DEP(i)     pmu_conf->pmc_desc[i].dep_pmd[0]
117 
118 #define PFM_NUM_IBRS      IA64_NUM_DBG_REGS
119 #define PFM_NUM_DBRS      IA64_NUM_DBG_REGS
120 
121 #define CTX_OVFL_NOBLOCK(c)     ((c)->ctx_fl_block == 0)
122 #define CTX_HAS_SMPL(c)         ((c)->ctx_fl_is_sampling)
123 #define PFM_CTX_TASK(h)         (h)->ctx_task
124 
125 #define PMU_PMC_OI              5 /* position of pmc.oi bit */
126 
127 /* XXX: does not support more than 64 PMDs */
128 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
129 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
130 
131 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
132 
133 #define CTX_USED_IBR(ctx,n)     (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
134 #define CTX_USED_DBR(ctx,n)     (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
135 #define CTX_USES_DBREGS(ctx)    (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
136 #define PFM_CODE_RR     0       /* requesting code range restriction */
137 #define PFM_DATA_RR     1       /* requestion data range restriction */
138 
139 #define PFM_CPUINFO_CLEAR(v)    pfm_get_cpu_var(pfm_syst_info) &= ~(v)
140 #define PFM_CPUINFO_SET(v)      pfm_get_cpu_var(pfm_syst_info) |= (v)
141 #define PFM_CPUINFO_GET()       pfm_get_cpu_var(pfm_syst_info)
142 
143 #define RDEP(x) (1UL<<(x))
144 
145 /*
146  * context protection macros
147  * in SMP:
148  *      - we need to protect against CPU concurrency (spin_lock)
149  *      - we need to protect against PMU overflow interrupts (local_irq_disable)
150  * in UP:
151  *      - we need to protect against PMU overflow interrupts (local_irq_disable)
152  *
153  * spin_lock_irqsave()/spin_unlock_irqrestore():
154  *      in SMP: local_irq_disable + spin_lock
155  *      in UP : local_irq_disable
156  *
157  * spin_lock()/spin_lock():
158  *      in UP : removed automatically
159  *      in SMP: protect against context accesses from other CPU. interrupts
160  *              are not masked. This is useful for the PMU interrupt handler
161  *              because we know we will not get PMU concurrency in that code.
162  */
163 #define PROTECT_CTX(c, f) \
164         do {  \
165                 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
166                 spin_lock_irqsave(&(c)->ctx_lock, f); \
167                 DPRINT(("spinlocked ctx %p  by [%d]\n", c, task_pid_nr(current))); \
168         } while(0)
169 
170 #define UNPROTECT_CTX(c, f) \
171         do { \
172                 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
173                 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
174         } while(0)
175 
176 #define PROTECT_CTX_NOPRINT(c, f) \
177         do {  \
178                 spin_lock_irqsave(&(c)->ctx_lock, f); \
179         } while(0)
180 
181 
182 #define UNPROTECT_CTX_NOPRINT(c, f) \
183         do { \
184                 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
185         } while(0)
186 
187 
188 #define PROTECT_CTX_NOIRQ(c) \
189         do {  \
190                 spin_lock(&(c)->ctx_lock); \
191         } while(0)
192 
193 #define UNPROTECT_CTX_NOIRQ(c) \
194         do { \
195                 spin_unlock(&(c)->ctx_lock); \
196         } while(0)
197 
198 
199 #ifdef CONFIG_SMP
200 
201 #define GET_ACTIVATION()        pfm_get_cpu_var(pmu_activation_number)
202 #define INC_ACTIVATION()        pfm_get_cpu_var(pmu_activation_number)++
203 #define SET_ACTIVATION(c)       (c)->ctx_last_activation = GET_ACTIVATION()
204 
205 #else /* !CONFIG_SMP */
206 #define SET_ACTIVATION(t)       do {} while(0)
207 #define GET_ACTIVATION(t)       do {} while(0)
208 #define INC_ACTIVATION(t)       do {} while(0)
209 #endif /* CONFIG_SMP */
210 
211 #define SET_PMU_OWNER(t, c)     do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
212 #define GET_PMU_OWNER()         pfm_get_cpu_var(pmu_owner)
213 #define GET_PMU_CTX()           pfm_get_cpu_var(pmu_ctx)
214 
215 #define LOCK_PFS(g)             spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
216 #define UNLOCK_PFS(g)           spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
217 
218 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
219 
220 /*
221  * cmp0 must be the value of pmc0
222  */
223 #define PMC0_HAS_OVFL(cmp0)  (cmp0 & ~0x1UL)
224 
225 #define PFMFS_MAGIC 0xa0b4d889
226 
227 /*
228  * debugging
229  */
230 #define PFM_DEBUGGING 1
231 #ifdef PFM_DEBUGGING
232 #define DPRINT(a) \
233         do { \
234                 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
235         } while (0)
236 
237 #define DPRINT_ovfl(a) \
238         do { \
239                 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
240         } while (0)
241 #endif
242 
243 /*
244  * 64-bit software counter structure
245  *
246  * the next_reset_type is applied to the next call to pfm_reset_regs()
247  */
248 typedef struct {
249         unsigned long   val;            /* virtual 64bit counter value */
250         unsigned long   lval;           /* last reset value */
251         unsigned long   long_reset;     /* reset value on sampling overflow */
252         unsigned long   short_reset;    /* reset value on overflow */
253         unsigned long   reset_pmds[4];  /* which other pmds to reset when this counter overflows */
254         unsigned long   smpl_pmds[4];   /* which pmds are accessed when counter overflow */
255         unsigned long   seed;           /* seed for random-number generator */
256         unsigned long   mask;           /* mask for random-number generator */
257         unsigned int    flags;          /* notify/do not notify */
258         unsigned long   eventid;        /* overflow event identifier */
259 } pfm_counter_t;
260 
261 /*
262  * context flags
263  */
264 typedef struct {
265         unsigned int block:1;           /* when 1, task will blocked on user notifications */
266         unsigned int system:1;          /* do system wide monitoring */
267         unsigned int using_dbreg:1;     /* using range restrictions (debug registers) */
268         unsigned int is_sampling:1;     /* true if using a custom format */
269         unsigned int excl_idle:1;       /* exclude idle task in system wide session */
270         unsigned int going_zombie:1;    /* context is zombie (MASKED+blocking) */
271         unsigned int trap_reason:2;     /* reason for going into pfm_handle_work() */
272         unsigned int no_msg:1;          /* no message sent on overflow */
273         unsigned int can_restart:1;     /* allowed to issue a PFM_RESTART */
274         unsigned int reserved:22;
275 } pfm_context_flags_t;
276 
277 #define PFM_TRAP_REASON_NONE            0x0     /* default value */
278 #define PFM_TRAP_REASON_BLOCK           0x1     /* we need to block on overflow */
279 #define PFM_TRAP_REASON_RESET           0x2     /* we need to reset PMDs */
280 
281 
282 /*
283  * perfmon context: encapsulates all the state of a monitoring session
284  */
285 
286 typedef struct pfm_context {
287         spinlock_t              ctx_lock;               /* context protection */
288 
289         pfm_context_flags_t     ctx_flags;              /* bitmask of flags  (block reason incl.) */
290         unsigned int            ctx_state;              /* state: active/inactive (no bitfield) */
291 
292         struct task_struct      *ctx_task;              /* task to which context is attached */
293 
294         unsigned long           ctx_ovfl_regs[4];       /* which registers overflowed (notification) */
295 
296         struct completion       ctx_restart_done;       /* use for blocking notification mode */
297 
298         unsigned long           ctx_used_pmds[4];       /* bitmask of PMD used            */
299         unsigned long           ctx_all_pmds[4];        /* bitmask of all accessible PMDs */
300         unsigned long           ctx_reload_pmds[4];     /* bitmask of force reload PMD on ctxsw in */
301 
302         unsigned long           ctx_all_pmcs[4];        /* bitmask of all accessible PMCs */
303         unsigned long           ctx_reload_pmcs[4];     /* bitmask of force reload PMC on ctxsw in */
304         unsigned long           ctx_used_monitors[4];   /* bitmask of monitor PMC being used */
305 
306         unsigned long           ctx_pmcs[PFM_NUM_PMC_REGS];     /*  saved copies of PMC values */
307 
308         unsigned int            ctx_used_ibrs[1];               /* bitmask of used IBR (speedup ctxsw in) */
309         unsigned int            ctx_used_dbrs[1];               /* bitmask of used DBR (speedup ctxsw in) */
310         unsigned long           ctx_dbrs[IA64_NUM_DBG_REGS];    /* DBR values (cache) when not loaded */
311         unsigned long           ctx_ibrs[IA64_NUM_DBG_REGS];    /* IBR values (cache) when not loaded */
312 
313         pfm_counter_t           ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
314 
315         unsigned long           th_pmcs[PFM_NUM_PMC_REGS];      /* PMC thread save state */
316         unsigned long           th_pmds[PFM_NUM_PMD_REGS];      /* PMD thread save state */
317 
318         unsigned long           ctx_saved_psr_up;       /* only contains psr.up value */
319 
320         unsigned long           ctx_last_activation;    /* context last activation number for last_cpu */
321         unsigned int            ctx_last_cpu;           /* CPU id of current or last CPU used (SMP only) */
322         unsigned int            ctx_cpu;                /* cpu to which perfmon is applied (system wide) */
323 
324         int                     ctx_fd;                 /* file descriptor used my this context */
325         pfm_ovfl_arg_t          ctx_ovfl_arg;           /* argument to custom buffer format handler */
326 
327         pfm_buffer_fmt_t        *ctx_buf_fmt;           /* buffer format callbacks */
328         void                    *ctx_smpl_hdr;          /* points to sampling buffer header kernel vaddr */
329         unsigned long           ctx_smpl_size;          /* size of sampling buffer */
330         void                    *ctx_smpl_vaddr;        /* user level virtual address of smpl buffer */
331 
332         wait_queue_head_t       ctx_msgq_wait;
333         pfm_msg_t               ctx_msgq[PFM_MAX_MSGS];
334         int                     ctx_msgq_head;
335         int                     ctx_msgq_tail;
336         struct fasync_struct    *ctx_async_queue;
337 
338         wait_queue_head_t       ctx_zombieq;            /* termination cleanup wait queue */
339 } pfm_context_t;
340 
341 /*
342  * magic number used to verify that structure is really
343  * a perfmon context
344  */
345 #define PFM_IS_FILE(f)          ((f)->f_op == &pfm_file_ops)
346 
347 #define PFM_GET_CTX(t)          ((pfm_context_t *)(t)->thread.pfm_context)
348 
349 #ifdef CONFIG_SMP
350 #define SET_LAST_CPU(ctx, v)    (ctx)->ctx_last_cpu = (v)
351 #define GET_LAST_CPU(ctx)       (ctx)->ctx_last_cpu
352 #else
353 #define SET_LAST_CPU(ctx, v)    do {} while(0)
354 #define GET_LAST_CPU(ctx)       do {} while(0)
355 #endif
356 
357 
358 #define ctx_fl_block            ctx_flags.block
359 #define ctx_fl_system           ctx_flags.system
360 #define ctx_fl_using_dbreg      ctx_flags.using_dbreg
361 #define ctx_fl_is_sampling      ctx_flags.is_sampling
362 #define ctx_fl_excl_idle        ctx_flags.excl_idle
363 #define ctx_fl_going_zombie     ctx_flags.going_zombie
364 #define ctx_fl_trap_reason      ctx_flags.trap_reason
365 #define ctx_fl_no_msg           ctx_flags.no_msg
366 #define ctx_fl_can_restart      ctx_flags.can_restart
367 
368 #define PFM_SET_WORK_PENDING(t, v)      do { (t)->thread.pfm_needs_checking = v; } while(0);
369 #define PFM_GET_WORK_PENDING(t)         (t)->thread.pfm_needs_checking
370 
371 /*
372  * global information about all sessions
373  * mostly used to synchronize between system wide and per-process
374  */
375 typedef struct {
376         spinlock_t              pfs_lock;                  /* lock the structure */
377 
378         unsigned int            pfs_task_sessions;         /* number of per task sessions */
379         unsigned int            pfs_sys_sessions;          /* number of per system wide sessions */
380         unsigned int            pfs_sys_use_dbregs;        /* incremented when a system wide session uses debug regs */
381         unsigned int            pfs_ptrace_use_dbregs;     /* incremented when a process uses debug regs */
382         struct task_struct      *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
383 } pfm_session_t;
384 
385 /*
386  * information about a PMC or PMD.
387  * dep_pmd[]: a bitmask of dependent PMD registers
388  * dep_pmc[]: a bitmask of dependent PMC registers
389  */
390 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
391 typedef struct {
392         unsigned int            type;
393         int                     pm_pos;
394         unsigned long           default_value;  /* power-on default value */
395         unsigned long           reserved_mask;  /* bitmask of reserved bits */
396         pfm_reg_check_t         read_check;
397         pfm_reg_check_t         write_check;
398         unsigned long           dep_pmd[4];
399         unsigned long           dep_pmc[4];
400 } pfm_reg_desc_t;
401 
402 /* assume cnum is a valid monitor */
403 #define PMC_PM(cnum, val)       (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
404 
405 /*
406  * This structure is initialized at boot time and contains
407  * a description of the PMU main characteristics.
408  *
409  * If the probe function is defined, detection is based
410  * on its return value: 
411  *      - 0 means recognized PMU
412  *      - anything else means not supported
413  * When the probe function is not defined, then the pmu_family field
414  * is used and it must match the host CPU family such that:
415  *      - cpu->family & config->pmu_family != 0
416  */
417 typedef struct {
418         unsigned long  ovfl_val;        /* overflow value for counters */
419 
420         pfm_reg_desc_t *pmc_desc;       /* detailed PMC register dependencies descriptions */
421         pfm_reg_desc_t *pmd_desc;       /* detailed PMD register dependencies descriptions */
422 
423         unsigned int   num_pmcs;        /* number of PMCS: computed at init time */
424         unsigned int   num_pmds;        /* number of PMDS: computed at init time */
425         unsigned long  impl_pmcs[4];    /* bitmask of implemented PMCS */
426         unsigned long  impl_pmds[4];    /* bitmask of implemented PMDS */
427 
428         char          *pmu_name;        /* PMU family name */
429         unsigned int  pmu_family;       /* cpuid family pattern used to identify pmu */
430         unsigned int  flags;            /* pmu specific flags */
431         unsigned int  num_ibrs;         /* number of IBRS: computed at init time */
432         unsigned int  num_dbrs;         /* number of DBRS: computed at init time */
433         unsigned int  num_counters;     /* PMC/PMD counting pairs : computed at init time */
434         int           (*probe)(void);   /* customized probe routine */
435         unsigned int  use_rr_dbregs:1;  /* set if debug registers used for range restriction */
436 } pmu_config_t;
437 /*
438  * PMU specific flags
439  */
440 #define PFM_PMU_IRQ_RESEND      1       /* PMU needs explicit IRQ resend */
441 
442 /*
443  * debug register related type definitions
444  */
445 typedef struct {
446         unsigned long ibr_mask:56;
447         unsigned long ibr_plm:4;
448         unsigned long ibr_ig:3;
449         unsigned long ibr_x:1;
450 } ibr_mask_reg_t;
451 
452 typedef struct {
453         unsigned long dbr_mask:56;
454         unsigned long dbr_plm:4;
455         unsigned long dbr_ig:2;
456         unsigned long dbr_w:1;
457         unsigned long dbr_r:1;
458 } dbr_mask_reg_t;
459 
460 typedef union {
461         unsigned long  val;
462         ibr_mask_reg_t ibr;
463         dbr_mask_reg_t dbr;
464 } dbreg_t;
465 
466 
467 /*
468  * perfmon command descriptions
469  */
470 typedef struct {
471         int             (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
472         char            *cmd_name;
473         int             cmd_flags;
474         unsigned int    cmd_narg;
475         size_t          cmd_argsize;
476         int             (*cmd_getsize)(void *arg, size_t *sz);
477 } pfm_cmd_desc_t;
478 
479 #define PFM_CMD_FD              0x01    /* command requires a file descriptor */
480 #define PFM_CMD_ARG_READ        0x02    /* command must read argument(s) */
481 #define PFM_CMD_ARG_RW          0x04    /* command must read/write argument(s) */
482 #define PFM_CMD_STOP            0x08    /* command does not work on zombie context */
483 
484 
485 #define PFM_CMD_NAME(cmd)       pfm_cmd_tab[(cmd)].cmd_name
486 #define PFM_CMD_READ_ARG(cmd)   (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
487 #define PFM_CMD_RW_ARG(cmd)     (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
488 #define PFM_CMD_USE_FD(cmd)     (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
489 #define PFM_CMD_STOPPED(cmd)    (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
490 
491 #define PFM_CMD_ARG_MANY        -1 /* cannot be zero */
492 
493 typedef struct {
494         unsigned long pfm_spurious_ovfl_intr_count;     /* keep track of spurious ovfl interrupts */
495         unsigned long pfm_replay_ovfl_intr_count;       /* keep track of replayed ovfl interrupts */
496         unsigned long pfm_ovfl_intr_count;              /* keep track of ovfl interrupts */
497         unsigned long pfm_ovfl_intr_cycles;             /* cycles spent processing ovfl interrupts */
498         unsigned long pfm_ovfl_intr_cycles_min;         /* min cycles spent processing ovfl interrupts */
499         unsigned long pfm_ovfl_intr_cycles_max;         /* max cycles spent processing ovfl interrupts */
500         unsigned long pfm_smpl_handler_calls;
501         unsigned long pfm_smpl_handler_cycles;
502         char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
503 } pfm_stats_t;
504 
505 /*
506  * perfmon internal variables
507  */
508 static pfm_stats_t              pfm_stats[NR_CPUS];
509 static pfm_session_t            pfm_sessions;   /* global sessions information */
510 
511 static DEFINE_SPINLOCK(pfm_alt_install_check);
512 static pfm_intr_handler_desc_t  *pfm_alt_intr_handler;
513 
514 static struct proc_dir_entry    *perfmon_dir;
515 static pfm_uuid_t               pfm_null_uuid = {0,};
516 
517 static spinlock_t               pfm_buffer_fmt_lock;
518 static LIST_HEAD(pfm_buffer_fmt_list);
519 
520 static pmu_config_t             *pmu_conf;
521 
522 /* sysctl() controls */
523 pfm_sysctl_t pfm_sysctl;
524 EXPORT_SYMBOL(pfm_sysctl);
525 
526 static struct ctl_table pfm_ctl_table[] = {
527         {
528                 .procname       = "debug",
529                 .data           = &pfm_sysctl.debug,
530                 .maxlen         = sizeof(int),
531                 .mode           = 0666,
532                 .proc_handler   = proc_dointvec,
533         },
534         {
535                 .procname       = "debug_ovfl",
536                 .data           = &pfm_sysctl.debug_ovfl,
537                 .maxlen         = sizeof(int),
538                 .mode           = 0666,
539                 .proc_handler   = proc_dointvec,
540         },
541         {
542                 .procname       = "fastctxsw",
543                 .data           = &pfm_sysctl.fastctxsw,
544                 .maxlen         = sizeof(int),
545                 .mode           = 0600,
546                 .proc_handler   = proc_dointvec,
547         },
548         {
549                 .procname       = "expert_mode",
550                 .data           = &pfm_sysctl.expert_mode,
551                 .maxlen         = sizeof(int),
552                 .mode           = 0600,
553                 .proc_handler   = proc_dointvec,
554         },
555         {}
556 };
557 static struct ctl_table pfm_sysctl_dir[] = {
558         {
559                 .procname       = "perfmon",
560                 .mode           = 0555,
561                 .child          = pfm_ctl_table,
562         },
563         {}
564 };
565 static struct ctl_table pfm_sysctl_root[] = {
566         {
567                 .procname       = "kernel",
568                 .mode           = 0555,
569                 .child          = pfm_sysctl_dir,
570         },
571         {}
572 };
573 static struct ctl_table_header *pfm_sysctl_header;
574 
575 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
576 
577 #define pfm_get_cpu_var(v)              __ia64_per_cpu_var(v)
578 #define pfm_get_cpu_data(a,b)           per_cpu(a, b)
579 
580 static inline void
581 pfm_put_task(struct task_struct *task)
582 {
583         if (task != current) put_task_struct(task);
584 }
585 
586 static inline void
587 pfm_reserve_page(unsigned long a)
588 {
589         SetPageReserved(vmalloc_to_page((void *)a));
590 }
591 static inline void
592 pfm_unreserve_page(unsigned long a)
593 {
594         ClearPageReserved(vmalloc_to_page((void*)a));
595 }
596 
597 static inline unsigned long
598 pfm_protect_ctx_ctxsw(pfm_context_t *x)
599 {
600         spin_lock(&(x)->ctx_lock);
601         return 0UL;
602 }
603 
604 static inline void
605 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
606 {
607         spin_unlock(&(x)->ctx_lock);
608 }
609 
610 /* forward declaration */
611 static const struct dentry_operations pfmfs_dentry_operations;
612 
613 static struct dentry *
614 pfmfs_mount(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
615 {
616         return mount_pseudo(fs_type, "pfm:", NULL, &pfmfs_dentry_operations,
617                         PFMFS_MAGIC);
618 }
619 
620 static struct file_system_type pfm_fs_type = {
621         .name     = "pfmfs",
622         .mount    = pfmfs_mount,
623         .kill_sb  = kill_anon_super,
624 };
625 MODULE_ALIAS_FS("pfmfs");
626 
627 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
628 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
629 DEFINE_PER_CPU(pfm_context_t  *, pmu_ctx);
630 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
631 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
632 
633 
634 /* forward declaration */
635 static const struct file_operations pfm_file_ops;
636 
637 /*
638  * forward declarations
639  */
640 #ifndef CONFIG_SMP
641 static void pfm_lazy_save_regs (struct task_struct *ta);
642 #endif
643 
644 void dump_pmu_state(const char *);
645 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
646 
647 #include "perfmon_itanium.h"
648 #include "perfmon_mckinley.h"
649 #include "perfmon_montecito.h"
650 #include "perfmon_generic.h"
651 
652 static pmu_config_t *pmu_confs[]={
653         &pmu_conf_mont,
654         &pmu_conf_mck,
655         &pmu_conf_ita,
656         &pmu_conf_gen, /* must be last */
657         NULL
658 };
659 
660 
661 static int pfm_end_notify_user(pfm_context_t *ctx);
662 
663 static inline void
664 pfm_clear_psr_pp(void)
665 {
666         ia64_rsm(IA64_PSR_PP);
667         ia64_srlz_i();
668 }
669 
670 static inline void
671 pfm_set_psr_pp(void)
672 {
673         ia64_ssm(IA64_PSR_PP);
674         ia64_srlz_i();
675 }
676 
677 static inline void
678 pfm_clear_psr_up(void)
679 {
680         ia64_rsm(IA64_PSR_UP);
681         ia64_srlz_i();
682 }
683 
684 static inline void
685 pfm_set_psr_up(void)
686 {
687         ia64_ssm(IA64_PSR_UP);
688         ia64_srlz_i();
689 }
690 
691 static inline unsigned long
692 pfm_get_psr(void)
693 {
694         unsigned long tmp;
695         tmp = ia64_getreg(_IA64_REG_PSR);
696         ia64_srlz_i();
697         return tmp;
698 }
699 
700 static inline void
701 pfm_set_psr_l(unsigned long val)
702 {
703         ia64_setreg(_IA64_REG_PSR_L, val);
704         ia64_srlz_i();
705 }
706 
707 static inline void
708 pfm_freeze_pmu(void)
709 {
710         ia64_set_pmc(0,1UL);
711         ia64_srlz_d();
712 }
713 
714 static inline void
715 pfm_unfreeze_pmu(void)
716 {
717         ia64_set_pmc(0,0UL);
718         ia64_srlz_d();
719 }
720 
721 static inline void
722 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
723 {
724         int i;
725 
726         for (i=0; i < nibrs; i++) {
727                 ia64_set_ibr(i, ibrs[i]);
728                 ia64_dv_serialize_instruction();
729         }
730         ia64_srlz_i();
731 }
732 
733 static inline void
734 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
735 {
736         int i;
737 
738         for (i=0; i < ndbrs; i++) {
739                 ia64_set_dbr(i, dbrs[i]);
740                 ia64_dv_serialize_data();
741         }
742         ia64_srlz_d();
743 }
744 
745 /*
746  * PMD[i] must be a counter. no check is made
747  */
748 static inline unsigned long
749 pfm_read_soft_counter(pfm_context_t *ctx, int i)
750 {
751         return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
752 }
753 
754 /*
755  * PMD[i] must be a counter. no check is made
756  */
757 static inline void
758 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
759 {
760         unsigned long ovfl_val = pmu_conf->ovfl_val;
761 
762         ctx->ctx_pmds[i].val = val  & ~ovfl_val;
763         /*
764          * writing to unimplemented part is ignore, so we do not need to
765          * mask off top part
766          */
767         ia64_set_pmd(i, val & ovfl_val);
768 }
769 
770 static pfm_msg_t *
771 pfm_get_new_msg(pfm_context_t *ctx)
772 {
773         int idx, next;
774 
775         next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
776 
777         DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
778         if (next == ctx->ctx_msgq_head) return NULL;
779 
780         idx =   ctx->ctx_msgq_tail;
781         ctx->ctx_msgq_tail = next;
782 
783         DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
784 
785         return ctx->ctx_msgq+idx;
786 }
787 
788 static pfm_msg_t *
789 pfm_get_next_msg(pfm_context_t *ctx)
790 {
791         pfm_msg_t *msg;
792 
793         DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
794 
795         if (PFM_CTXQ_EMPTY(ctx)) return NULL;
796 
797         /*
798          * get oldest message
799          */
800         msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
801 
802         /*
803          * and move forward
804          */
805         ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
806 
807         DPRINT(("ctx=%p head=%d tail=%d type=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, msg->pfm_gen_msg.msg_type));
808 
809         return msg;
810 }
811 
812 static void
813 pfm_reset_msgq(pfm_context_t *ctx)
814 {
815         ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
816         DPRINT(("ctx=%p msgq reset\n", ctx));
817 }
818 
819 static void *
820 pfm_rvmalloc(unsigned long size)
821 {
822         void *mem;
823         unsigned long addr;
824 
825         size = PAGE_ALIGN(size);
826         mem  = vzalloc(size);
827         if (mem) {
828                 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
829                 addr = (unsigned long)mem;
830                 while (size > 0) {
831                         pfm_reserve_page(addr);
832                         addr+=PAGE_SIZE;
833                         size-=PAGE_SIZE;
834                 }
835         }
836         return mem;
837 }
838 
839 static void
840 pfm_rvfree(void *mem, unsigned long size)
841 {
842         unsigned long addr;
843 
844         if (mem) {
845                 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
846                 addr = (unsigned long) mem;
847                 while ((long) size > 0) {
848                         pfm_unreserve_page(addr);
849                         addr+=PAGE_SIZE;
850                         size-=PAGE_SIZE;
851                 }
852                 vfree(mem);
853         }
854         return;
855 }
856 
857 static pfm_context_t *
858 pfm_context_alloc(int ctx_flags)
859 {
860         pfm_context_t *ctx;
861 
862         /* 
863          * allocate context descriptor 
864          * must be able to free with interrupts disabled
865          */
866         ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
867         if (ctx) {
868                 DPRINT(("alloc ctx @%p\n", ctx));
869 
870                 /*
871                  * init context protection lock
872                  */
873                 spin_lock_init(&ctx->ctx_lock);
874 
875                 /*
876                  * context is unloaded
877                  */
878                 ctx->ctx_state = PFM_CTX_UNLOADED;
879 
880                 /*
881                  * initialization of context's flags
882                  */
883                 ctx->ctx_fl_block       = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
884                 ctx->ctx_fl_system      = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
885                 ctx->ctx_fl_no_msg      = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
886                 /*
887                  * will move to set properties
888                  * ctx->ctx_fl_excl_idle   = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
889                  */
890 
891                 /*
892                  * init restart semaphore to locked
893                  */
894                 init_completion(&ctx->ctx_restart_done);
895 
896                 /*
897                  * activation is used in SMP only
898                  */
899                 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
900                 SET_LAST_CPU(ctx, -1);
901 
902                 /*
903                  * initialize notification message queue
904                  */
905                 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
906                 init_waitqueue_head(&ctx->ctx_msgq_wait);
907                 init_waitqueue_head(&ctx->ctx_zombieq);
908 
909         }
910         return ctx;
911 }
912 
913 static void
914 pfm_context_free(pfm_context_t *ctx)
915 {
916         if (ctx) {
917                 DPRINT(("free ctx @%p\n", ctx));
918                 kfree(ctx);
919         }
920 }
921 
922 static void
923 pfm_mask_monitoring(struct task_struct *task)
924 {
925         pfm_context_t *ctx = PFM_GET_CTX(task);
926         unsigned long mask, val, ovfl_mask;
927         int i;
928 
929         DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
930 
931         ovfl_mask = pmu_conf->ovfl_val;
932         /*
933          * monitoring can only be masked as a result of a valid
934          * counter overflow. In UP, it means that the PMU still
935          * has an owner. Note that the owner can be different
936          * from the current task. However the PMU state belongs
937          * to the owner.
938          * In SMP, a valid overflow only happens when task is
939          * current. Therefore if we come here, we know that
940          * the PMU state belongs to the current task, therefore
941          * we can access the live registers.
942          *
943          * So in both cases, the live register contains the owner's
944          * state. We can ONLY touch the PMU registers and NOT the PSR.
945          *
946          * As a consequence to this call, the ctx->th_pmds[] array
947          * contains stale information which must be ignored
948          * when context is reloaded AND monitoring is active (see
949          * pfm_restart).
950          */
951         mask = ctx->ctx_used_pmds[0];
952         for (i = 0; mask; i++, mask>>=1) {
953                 /* skip non used pmds */
954                 if ((mask & 0x1) == 0) continue;
955                 val = ia64_get_pmd(i);
956 
957                 if (PMD_IS_COUNTING(i)) {
958                         /*
959                          * we rebuild the full 64 bit value of the counter
960                          */
961                         ctx->ctx_pmds[i].val += (val & ovfl_mask);
962                 } else {
963                         ctx->ctx_pmds[i].val = val;
964                 }
965                 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
966                         i,
967                         ctx->ctx_pmds[i].val,
968                         val & ovfl_mask));
969         }
970         /*
971          * mask monitoring by setting the privilege level to 0
972          * we cannot use psr.pp/psr.up for this, it is controlled by
973          * the user
974          *
975          * if task is current, modify actual registers, otherwise modify
976          * thread save state, i.e., what will be restored in pfm_load_regs()
977          */
978         mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
979         for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
980                 if ((mask & 0x1) == 0UL) continue;
981                 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
982                 ctx->th_pmcs[i] &= ~0xfUL;
983                 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
984         }
985         /*
986          * make all of this visible
987          */
988         ia64_srlz_d();
989 }
990 
991 /*
992  * must always be done with task == current
993  *
994  * context must be in MASKED state when calling
995  */
996 static void
997 pfm_restore_monitoring(struct task_struct *task)
998 {
999         pfm_context_t *ctx = PFM_GET_CTX(task);
1000         unsigned long mask, ovfl_mask;
1001         unsigned long psr, val;
1002         int i, is_system;
1003 
1004         is_system = ctx->ctx_fl_system;
1005         ovfl_mask = pmu_conf->ovfl_val;
1006 
1007         if (task != current) {
1008                 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1009                 return;
1010         }
1011         if (ctx->ctx_state != PFM_CTX_MASKED) {
1012                 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1013                         task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1014                 return;
1015         }
1016         psr = pfm_get_psr();
1017         /*
1018          * monitoring is masked via the PMC.
1019          * As we restore their value, we do not want each counter to
1020          * restart right away. We stop monitoring using the PSR,
1021          * restore the PMC (and PMD) and then re-establish the psr
1022          * as it was. Note that there can be no pending overflow at
1023          * this point, because monitoring was MASKED.
1024          *
1025          * system-wide session are pinned and self-monitoring
1026          */
1027         if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1028                 /* disable dcr pp */
1029                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1030                 pfm_clear_psr_pp();
1031         } else {
1032                 pfm_clear_psr_up();
1033         }
1034         /*
1035          * first, we restore the PMD
1036          */
1037         mask = ctx->ctx_used_pmds[0];
1038         for (i = 0; mask; i++, mask>>=1) {
1039                 /* skip non used pmds */
1040                 if ((mask & 0x1) == 0) continue;
1041 
1042                 if (PMD_IS_COUNTING(i)) {
1043                         /*
1044                          * we split the 64bit value according to
1045                          * counter width
1046                          */
1047                         val = ctx->ctx_pmds[i].val & ovfl_mask;
1048                         ctx->ctx_pmds[i].val &= ~ovfl_mask;
1049                 } else {
1050                         val = ctx->ctx_pmds[i].val;
1051                 }
1052                 ia64_set_pmd(i, val);
1053 
1054                 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1055                         i,
1056                         ctx->ctx_pmds[i].val,
1057                         val));
1058         }
1059         /*
1060          * restore the PMCs
1061          */
1062         mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1063         for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1064                 if ((mask & 0x1) == 0UL) continue;
1065                 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1066                 ia64_set_pmc(i, ctx->th_pmcs[i]);
1067                 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1068                                         task_pid_nr(task), i, ctx->th_pmcs[i]));
1069         }
1070         ia64_srlz_d();
1071 
1072         /*
1073          * must restore DBR/IBR because could be modified while masked
1074          * XXX: need to optimize 
1075          */
1076         if (ctx->ctx_fl_using_dbreg) {
1077                 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1078                 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1079         }
1080 
1081         /*
1082          * now restore PSR
1083          */
1084         if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1085                 /* enable dcr pp */
1086                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1087                 ia64_srlz_i();
1088         }
1089         pfm_set_psr_l(psr);
1090 }
1091 
1092 static inline void
1093 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1094 {
1095         int i;
1096 
1097         ia64_srlz_d();
1098 
1099         for (i=0; mask; i++, mask>>=1) {
1100                 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1101         }
1102 }
1103 
1104 /*
1105  * reload from thread state (used for ctxw only)
1106  */
1107 static inline void
1108 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1109 {
1110         int i;
1111         unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1112 
1113         for (i=0; mask; i++, mask>>=1) {
1114                 if ((mask & 0x1) == 0) continue;
1115                 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1116                 ia64_set_pmd(i, val);
1117         }
1118         ia64_srlz_d();
1119 }
1120 
1121 /*
1122  * propagate PMD from context to thread-state
1123  */
1124 static inline void
1125 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1126 {
1127         unsigned long ovfl_val = pmu_conf->ovfl_val;
1128         unsigned long mask = ctx->ctx_all_pmds[0];
1129         unsigned long val;
1130         int i;
1131 
1132         DPRINT(("mask=0x%lx\n", mask));
1133 
1134         for (i=0; mask; i++, mask>>=1) {
1135 
1136                 val = ctx->ctx_pmds[i].val;
1137 
1138                 /*
1139                  * We break up the 64 bit value into 2 pieces
1140                  * the lower bits go to the machine state in the
1141                  * thread (will be reloaded on ctxsw in).
1142                  * The upper part stays in the soft-counter.
1143                  */
1144                 if (PMD_IS_COUNTING(i)) {
1145                         ctx->ctx_pmds[i].val = val & ~ovfl_val;
1146                          val &= ovfl_val;
1147                 }
1148                 ctx->th_pmds[i] = val;
1149 
1150                 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1151                         i,
1152                         ctx->th_pmds[i],
1153                         ctx->ctx_pmds[i].val));
1154         }
1155 }
1156 
1157 /*
1158  * propagate PMC from context to thread-state
1159  */
1160 static inline void
1161 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1162 {
1163         unsigned long mask = ctx->ctx_all_pmcs[0];
1164         int i;
1165 
1166         DPRINT(("mask=0x%lx\n", mask));
1167 
1168         for (i=0; mask; i++, mask>>=1) {
1169                 /* masking 0 with ovfl_val yields 0 */
1170                 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1171                 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1172         }
1173 }
1174 
1175 
1176 
1177 static inline void
1178 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1179 {
1180         int i;
1181 
1182         for (i=0; mask; i++, mask>>=1) {
1183                 if ((mask & 0x1) == 0) continue;
1184                 ia64_set_pmc(i, pmcs[i]);
1185         }
1186         ia64_srlz_d();
1187 }
1188 
1189 static inline int
1190 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1191 {
1192         return memcmp(a, b, sizeof(pfm_uuid_t));
1193 }
1194 
1195 static inline int
1196 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1197 {
1198         int ret = 0;
1199         if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1200         return ret;
1201 }
1202 
1203 static inline int
1204 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1205 {
1206         int ret = 0;
1207         if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1208         return ret;
1209 }
1210 
1211 
1212 static inline int
1213 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1214                      int cpu, void *arg)
1215 {
1216         int ret = 0;
1217         if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1218         return ret;
1219 }
1220 
1221 static inline int
1222 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1223                      int cpu, void *arg)
1224 {
1225         int ret = 0;
1226         if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1227         return ret;
1228 }
1229 
1230 static inline int
1231 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1232 {
1233         int ret = 0;
1234         if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1235         return ret;
1236 }
1237 
1238 static inline int
1239 pfm_buf_fmt_restart_active(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1240 {
1241         int ret = 0;
1242         if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1243         return ret;
1244 }
1245 
1246 static pfm_buffer_fmt_t *
1247 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1248 {
1249         struct list_head * pos;
1250         pfm_buffer_fmt_t * entry;
1251 
1252         list_for_each(pos, &pfm_buffer_fmt_list) {
1253                 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1254                 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1255                         return entry;
1256         }
1257         return NULL;
1258 }
1259  
1260 /*
1261  * find a buffer format based on its uuid
1262  */
1263 static pfm_buffer_fmt_t *
1264 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1265 {
1266         pfm_buffer_fmt_t * fmt;
1267         spin_lock(&pfm_buffer_fmt_lock);
1268         fmt = __pfm_find_buffer_fmt(uuid);
1269         spin_unlock(&pfm_buffer_fmt_lock);
1270         return fmt;
1271 }
1272  
1273 int
1274 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1275 {
1276         int ret = 0;
1277 
1278         /* some sanity checks */
1279         if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1280 
1281         /* we need at least a handler */
1282         if (fmt->fmt_handler == NULL) return -EINVAL;
1283 
1284         /*
1285          * XXX: need check validity of fmt_arg_size
1286          */
1287 
1288         spin_lock(&pfm_buffer_fmt_lock);
1289 
1290         if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1291                 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1292                 ret = -EBUSY;
1293                 goto out;
1294         } 
1295         list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1296         printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1297 
1298 out:
1299         spin_unlock(&pfm_buffer_fmt_lock);
1300         return ret;
1301 }
1302 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1303 
1304 int
1305 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1306 {
1307         pfm_buffer_fmt_t *fmt;
1308         int ret = 0;
1309 
1310         spin_lock(&pfm_buffer_fmt_lock);
1311 
1312         fmt = __pfm_find_buffer_fmt(uuid);
1313         if (!fmt) {
1314                 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1315                 ret = -EINVAL;
1316                 goto out;
1317         }
1318         list_del_init(&fmt->fmt_list);
1319         printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1320 
1321 out:
1322         spin_unlock(&pfm_buffer_fmt_lock);
1323         return ret;
1324 
1325 }
1326 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1327 
1328 static int
1329 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1330 {
1331         unsigned long flags;
1332         /*
1333          * validity checks on cpu_mask have been done upstream
1334          */
1335         LOCK_PFS(flags);
1336 
1337         DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1338                 pfm_sessions.pfs_sys_sessions,
1339                 pfm_sessions.pfs_task_sessions,
1340                 pfm_sessions.pfs_sys_use_dbregs,
1341                 is_syswide,
1342                 cpu));
1343 
1344         if (is_syswide) {
1345                 /*
1346                  * cannot mix system wide and per-task sessions
1347                  */
1348                 if (pfm_sessions.pfs_task_sessions > 0UL) {
1349                         DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1350                                 pfm_sessions.pfs_task_sessions));
1351                         goto abort;
1352                 }
1353 
1354                 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1355 
1356                 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1357 
1358                 pfm_sessions.pfs_sys_session[cpu] = task;
1359 
1360                 pfm_sessions.pfs_sys_sessions++ ;
1361 
1362         } else {
1363                 if (pfm_sessions.pfs_sys_sessions) goto abort;
1364                 pfm_sessions.pfs_task_sessions++;
1365         }
1366 
1367         DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1368                 pfm_sessions.pfs_sys_sessions,
1369                 pfm_sessions.pfs_task_sessions,
1370                 pfm_sessions.pfs_sys_use_dbregs,
1371                 is_syswide,
1372                 cpu));
1373 
1374         /*
1375          * Force idle() into poll mode
1376          */
1377         cpu_idle_poll_ctrl(true);
1378 
1379         UNLOCK_PFS(flags);
1380 
1381         return 0;
1382 
1383 error_conflict:
1384         DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1385                 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1386                 cpu));
1387 abort:
1388         UNLOCK_PFS(flags);
1389 
1390         return -EBUSY;
1391 
1392 }
1393 
1394 static int
1395 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1396 {
1397         unsigned long flags;
1398         /*
1399          * validity checks on cpu_mask have been done upstream
1400          */
1401         LOCK_PFS(flags);
1402 
1403         DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1404                 pfm_sessions.pfs_sys_sessions,
1405                 pfm_sessions.pfs_task_sessions,
1406                 pfm_sessions.pfs_sys_use_dbregs,
1407                 is_syswide,
1408                 cpu));
1409 
1410 
1411         if (is_syswide) {
1412                 pfm_sessions.pfs_sys_session[cpu] = NULL;
1413                 /*
1414                  * would not work with perfmon+more than one bit in cpu_mask
1415                  */
1416                 if (ctx && ctx->ctx_fl_using_dbreg) {
1417                         if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1418                                 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1419                         } else {
1420                                 pfm_sessions.pfs_sys_use_dbregs--;
1421                         }
1422                 }
1423                 pfm_sessions.pfs_sys_sessions--;
1424         } else {
1425                 pfm_sessions.pfs_task_sessions--;
1426         }
1427         DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1428                 pfm_sessions.pfs_sys_sessions,
1429                 pfm_sessions.pfs_task_sessions,
1430                 pfm_sessions.pfs_sys_use_dbregs,
1431                 is_syswide,
1432                 cpu));
1433 
1434         /* Undo forced polling. Last session reenables pal_halt */
1435         cpu_idle_poll_ctrl(false);
1436 
1437         UNLOCK_PFS(flags);
1438 
1439         return 0;
1440 }
1441 
1442 /*
1443  * removes virtual mapping of the sampling buffer.
1444  * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1445  * a PROTECT_CTX() section.
1446  */
1447 static int
1448 pfm_remove_smpl_mapping(void *vaddr, unsigned long size)
1449 {
1450         struct task_struct *task = current;
1451         int r;
1452 
1453         /* sanity checks */
1454         if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1455                 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1456                 return -EINVAL;
1457         }
1458 
1459         DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1460 
1461         /*
1462          * does the actual unmapping
1463          */
1464         r = vm_munmap((unsigned long)vaddr, size);
1465 
1466         if (r !=0) {
1467                 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1468         }
1469 
1470         DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1471 
1472         return 0;
1473 }
1474 
1475 /*
1476  * free actual physical storage used by sampling buffer
1477  */
1478 #if 0
1479 static int
1480 pfm_free_smpl_buffer(pfm_context_t *ctx)
1481 {
1482         pfm_buffer_fmt_t *fmt;
1483 
1484         if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1485 
1486         /*
1487          * we won't use the buffer format anymore
1488          */
1489         fmt = ctx->ctx_buf_fmt;
1490 
1491         DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1492                 ctx->ctx_smpl_hdr,
1493                 ctx->ctx_smpl_size,
1494                 ctx->ctx_smpl_vaddr));
1495 
1496         pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1497 
1498         /*
1499          * free the buffer
1500          */
1501         pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1502 
1503         ctx->ctx_smpl_hdr  = NULL;
1504         ctx->ctx_smpl_size = 0UL;
1505 
1506         return 0;
1507 
1508 invalid_free:
1509         printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1510         return -EINVAL;
1511 }
1512 #endif
1513 
1514 static inline void
1515 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1516 {
1517         if (fmt == NULL) return;
1518 
1519         pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1520 
1521 }
1522 
1523 /*
1524  * pfmfs should _never_ be mounted by userland - too much of security hassle,
1525  * no real gain from having the whole whorehouse mounted. So we don't need
1526  * any operations on the root directory. However, we need a non-trivial
1527  * d_name - pfm: will go nicely and kill the special-casing in procfs.
1528  */
1529 static struct vfsmount *pfmfs_mnt __read_mostly;
1530 
1531 static int __init
1532 init_pfm_fs(void)
1533 {
1534         int err = register_filesystem(&pfm_fs_type);
1535         if (!err) {
1536                 pfmfs_mnt = kern_mount(&pfm_fs_type);
1537                 err = PTR_ERR(pfmfs_mnt);
1538                 if (IS_ERR(pfmfs_mnt))
1539                         unregister_filesystem(&pfm_fs_type);
1540                 else
1541                         err = 0;
1542         }
1543         return err;
1544 }
1545 
1546 static ssize_t
1547 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1548 {
1549         pfm_context_t *ctx;
1550         pfm_msg_t *msg;
1551         ssize_t ret;
1552         unsigned long flags;
1553         DECLARE_WAITQUEUE(wait, current);
1554         if (PFM_IS_FILE(filp) == 0) {
1555                 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1556                 return -EINVAL;
1557         }
1558 
1559         ctx = filp->private_data;
1560         if (ctx == NULL) {
1561                 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1562                 return -EINVAL;
1563         }
1564 
1565         /*
1566          * check even when there is no message
1567          */
1568         if (size < sizeof(pfm_msg_t)) {
1569                 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1570                 return -EINVAL;
1571         }
1572 
1573         PROTECT_CTX(ctx, flags);
1574 
1575         /*
1576          * put ourselves on the wait queue
1577          */
1578         add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1579 
1580 
1581         for(;;) {
1582                 /*
1583                  * check wait queue
1584                  */
1585 
1586                 set_current_state(TASK_INTERRUPTIBLE);
1587 
1588                 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1589 
1590                 ret = 0;
1591                 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1592 
1593                 UNPROTECT_CTX(ctx, flags);
1594 
1595                 /*
1596                  * check non-blocking read
1597                  */
1598                 ret = -EAGAIN;
1599                 if(filp->f_flags & O_NONBLOCK) break;
1600 
1601                 /*
1602                  * check pending signals
1603                  */
1604                 if(signal_pending(current)) {
1605                         ret = -EINTR;
1606                         break;
1607                 }
1608                 /*
1609                  * no message, so wait
1610                  */
1611                 schedule();
1612 
1613                 PROTECT_CTX(ctx, flags);
1614         }
1615         DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1616         set_current_state(TASK_RUNNING);
1617         remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1618 
1619         if (ret < 0) goto abort;
1620 
1621         ret = -EINVAL;
1622         msg = pfm_get_next_msg(ctx);
1623         if (msg == NULL) {
1624                 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1625                 goto abort_locked;
1626         }
1627 
1628         DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1629 
1630         ret = -EFAULT;
1631         if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1632 
1633 abort_locked:
1634         UNPROTECT_CTX(ctx, flags);
1635 abort:
1636         return ret;
1637 }
1638 
1639 static ssize_t
1640 pfm_write(struct file *file, const char __user *ubuf,
1641                           size_t size, loff_t *ppos)
1642 {
1643         DPRINT(("pfm_write called\n"));
1644         return -EINVAL;
1645 }
1646 
1647 static unsigned int
1648 pfm_poll(struct file *filp, poll_table * wait)
1649 {
1650         pfm_context_t *ctx;
1651         unsigned long flags;
1652         unsigned int mask = 0;
1653 
1654         if (PFM_IS_FILE(filp) == 0) {
1655                 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1656                 return 0;
1657         }
1658 
1659         ctx = filp->private_data;
1660         if (ctx == NULL) {
1661                 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1662                 return 0;
1663         }
1664 
1665 
1666         DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1667 
1668         poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1669 
1670         PROTECT_CTX(ctx, flags);
1671 
1672         if (PFM_CTXQ_EMPTY(ctx) == 0)
1673                 mask =  POLLIN | POLLRDNORM;
1674 
1675         UNPROTECT_CTX(ctx, flags);
1676 
1677         DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1678 
1679         return mask;
1680 }
1681 
1682 static long
1683 pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1684 {
1685         DPRINT(("pfm_ioctl called\n"));
1686         return -EINVAL;
1687 }
1688 
1689 /*
1690  * interrupt cannot be masked when coming here
1691  */
1692 static inline int
1693 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1694 {
1695         int ret;
1696 
1697         ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1698 
1699         DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1700                 task_pid_nr(current),
1701                 fd,
1702                 on,
1703                 ctx->ctx_async_queue, ret));
1704 
1705         return ret;
1706 }
1707 
1708 static int
1709 pfm_fasync(int fd, struct file *filp, int on)
1710 {
1711         pfm_context_t *ctx;
1712         int ret;
1713 
1714         if (PFM_IS_FILE(filp) == 0) {
1715                 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1716                 return -EBADF;
1717         }
1718 
1719         ctx = filp->private_data;
1720         if (ctx == NULL) {
1721                 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1722                 return -EBADF;
1723         }
1724         /*
1725          * we cannot mask interrupts during this call because this may
1726          * may go to sleep if memory is not readily avalaible.
1727          *
1728          * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1729          * done in caller. Serialization of this function is ensured by caller.
1730          */
1731         ret = pfm_do_fasync(fd, filp, ctx, on);
1732 
1733 
1734         DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1735                 fd,
1736                 on,
1737                 ctx->ctx_async_queue, ret));
1738 
1739         return ret;
1740 }
1741 
1742 #ifdef CONFIG_SMP
1743 /*
1744  * this function is exclusively called from pfm_close().
1745  * The context is not protected at that time, nor are interrupts
1746  * on the remote CPU. That's necessary to avoid deadlocks.
1747  */
1748 static void
1749 pfm_syswide_force_stop(void *info)
1750 {
1751         pfm_context_t   *ctx = (pfm_context_t *)info;
1752         struct pt_regs *regs = task_pt_regs(current);
1753         struct task_struct *owner;
1754         unsigned long flags;
1755         int ret;
1756 
1757         if (ctx->ctx_cpu != smp_processor_id()) {
1758                 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d  but on CPU%d\n",
1759                         ctx->ctx_cpu,
1760                         smp_processor_id());
1761                 return;
1762         }
1763         owner = GET_PMU_OWNER();
1764         if (owner != ctx->ctx_task) {
1765                 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1766                         smp_processor_id(),
1767                         task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1768                 return;
1769         }
1770         if (GET_PMU_CTX() != ctx) {
1771                 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1772                         smp_processor_id(),
1773                         GET_PMU_CTX(), ctx);
1774                 return;
1775         }
1776 
1777         DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1778         /*
1779          * the context is already protected in pfm_close(), we simply
1780          * need to mask interrupts to avoid a PMU interrupt race on
1781          * this CPU
1782          */
1783         local_irq_save(flags);
1784 
1785         ret = pfm_context_unload(ctx, NULL, 0, regs);
1786         if (ret) {
1787                 DPRINT(("context_unload returned %d\n", ret));
1788         }
1789 
1790         /*
1791          * unmask interrupts, PMU interrupts are now spurious here
1792          */
1793         local_irq_restore(flags);
1794 }
1795 
1796 static void
1797 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1798 {
1799         int ret;
1800 
1801         DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1802         ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1803         DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1804 }
1805 #endif /* CONFIG_SMP */
1806 
1807 /*
1808  * called for each close(). Partially free resources.
1809  * When caller is self-monitoring, the context is unloaded.
1810  */
1811 static int
1812 pfm_flush(struct file *filp, fl_owner_t id)
1813 {
1814         pfm_context_t *ctx;
1815         struct task_struct *task;
1816         struct pt_regs *regs;
1817         unsigned long flags;
1818         unsigned long smpl_buf_size = 0UL;
1819         void *smpl_buf_vaddr = NULL;
1820         int state, is_system;
1821 
1822         if (PFM_IS_FILE(filp) == 0) {
1823                 DPRINT(("bad magic for\n"));
1824                 return -EBADF;
1825         }
1826 
1827         ctx = filp->private_data;
1828         if (ctx == NULL) {
1829                 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1830                 return -EBADF;
1831         }
1832 
1833         /*
1834          * remove our file from the async queue, if we use this mode.
1835          * This can be done without the context being protected. We come
1836          * here when the context has become unreachable by other tasks.
1837          *
1838          * We may still have active monitoring at this point and we may
1839          * end up in pfm_overflow_handler(). However, fasync_helper()
1840          * operates with interrupts disabled and it cleans up the
1841          * queue. If the PMU handler is called prior to entering
1842          * fasync_helper() then it will send a signal. If it is
1843          * invoked after, it will find an empty queue and no
1844          * signal will be sent. In both case, we are safe
1845          */
1846         PROTECT_CTX(ctx, flags);
1847 
1848         state     = ctx->ctx_state;
1849         is_system = ctx->ctx_fl_system;
1850 
1851         task = PFM_CTX_TASK(ctx);
1852         regs = task_pt_regs(task);
1853 
1854         DPRINT(("ctx_state=%d is_current=%d\n",
1855                 state,
1856                 task == current ? 1 : 0));
1857 
1858         /*
1859          * if state == UNLOADED, then task is NULL
1860          */
1861 
1862         /*
1863          * we must stop and unload because we are losing access to the context.
1864          */
1865         if (task == current) {
1866 #ifdef CONFIG_SMP
1867                 /*
1868                  * the task IS the owner but it migrated to another CPU: that's bad
1869                  * but we must handle this cleanly. Unfortunately, the kernel does
1870                  * not provide a mechanism to block migration (while the context is loaded).
1871                  *
1872                  * We need to release the resource on the ORIGINAL cpu.
1873                  */
1874                 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1875 
1876                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1877                         /*
1878                          * keep context protected but unmask interrupt for IPI
1879                          */
1880                         local_irq_restore(flags);
1881 
1882                         pfm_syswide_cleanup_other_cpu(ctx);
1883 
1884                         /*
1885                          * restore interrupt masking
1886                          */
1887                         local_irq_save(flags);
1888 
1889                         /*
1890                          * context is unloaded at this point
1891                          */
1892                 } else
1893 #endif /* CONFIG_SMP */
1894                 {
1895 
1896                         DPRINT(("forcing unload\n"));
1897                         /*
1898                         * stop and unload, returning with state UNLOADED
1899                         * and session unreserved.
1900                         */
1901                         pfm_context_unload(ctx, NULL, 0, regs);
1902 
1903                         DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1904                 }
1905         }
1906 
1907         /*
1908          * remove virtual mapping, if any, for the calling task.
1909          * cannot reset ctx field until last user is calling close().
1910          *
1911          * ctx_smpl_vaddr must never be cleared because it is needed
1912          * by every task with access to the context
1913          *
1914          * When called from do_exit(), the mm context is gone already, therefore
1915          * mm is NULL, i.e., the VMA is already gone  and we do not have to
1916          * do anything here
1917          */
1918         if (ctx->ctx_smpl_vaddr && current->mm) {
1919                 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1920                 smpl_buf_size  = ctx->ctx_smpl_size;
1921         }
1922 
1923         UNPROTECT_CTX(ctx, flags);
1924 
1925         /*
1926          * if there was a mapping, then we systematically remove it
1927          * at this point. Cannot be done inside critical section
1928          * because some VM function reenables interrupts.
1929          *
1930          */
1931         if (smpl_buf_vaddr) pfm_remove_smpl_mapping(smpl_buf_vaddr, smpl_buf_size);
1932 
1933         return 0;
1934 }
1935 /*
1936  * called either on explicit close() or from exit_files(). 
1937  * Only the LAST user of the file gets to this point, i.e., it is
1938  * called only ONCE.
1939  *
1940  * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero 
1941  * (fput()),i.e, last task to access the file. Nobody else can access the 
1942  * file at this point.
1943  *
1944  * When called from exit_files(), the VMA has been freed because exit_mm()
1945  * is executed before exit_files().
1946  *
1947  * When called from exit_files(), the current task is not yet ZOMBIE but we
1948  * flush the PMU state to the context. 
1949  */
1950 static int
1951 pfm_close(struct inode *inode, struct file *filp)
1952 {
1953         pfm_context_t *ctx;
1954         struct task_struct *task;
1955         struct pt_regs *regs;
1956         DECLARE_WAITQUEUE(wait, current);
1957         unsigned long flags;
1958         unsigned long smpl_buf_size = 0UL;
1959         void *smpl_buf_addr = NULL;
1960         int free_possible = 1;
1961         int state, is_system;
1962 
1963         DPRINT(("pfm_close called private=%p\n", filp->private_data));
1964 
1965         if (PFM_IS_FILE(filp) == 0) {
1966                 DPRINT(("bad magic\n"));
1967                 return -EBADF;
1968         }
1969         
1970         ctx = filp->private_data;
1971         if (ctx == NULL) {
1972                 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1973                 return -EBADF;
1974         }
1975 
1976         PROTECT_CTX(ctx, flags);
1977 
1978         state     = ctx->ctx_state;
1979         is_system = ctx->ctx_fl_system;
1980 
1981         task = PFM_CTX_TASK(ctx);
1982         regs = task_pt_regs(task);
1983 
1984         DPRINT(("ctx_state=%d is_current=%d\n", 
1985                 state,
1986                 task == current ? 1 : 0));
1987 
1988         /*
1989          * if task == current, then pfm_flush() unloaded the context
1990          */
1991         if (state == PFM_CTX_UNLOADED) goto doit;
1992 
1993         /*
1994          * context is loaded/masked and task != current, we need to
1995          * either force an unload or go zombie
1996          */
1997 
1998         /*
1999          * The task is currently blocked or will block after an overflow.
2000          * we must force it to wakeup to get out of the
2001          * MASKED state and transition to the unloaded state by itself.
2002          *
2003          * This situation is only possible for per-task mode
2004          */
2005         if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2006 
2007                 /*
2008                  * set a "partial" zombie state to be checked
2009                  * upon return from down() in pfm_handle_work().
2010                  *
2011                  * We cannot use the ZOMBIE state, because it is checked
2012                  * by pfm_load_regs() which is called upon wakeup from down().
2013                  * In such case, it would free the context and then we would
2014                  * return to pfm_handle_work() which would access the
2015                  * stale context. Instead, we set a flag invisible to pfm_load_regs()
2016                  * but visible to pfm_handle_work().
2017                  *
2018                  * For some window of time, we have a zombie context with
2019                  * ctx_state = MASKED  and not ZOMBIE
2020                  */
2021                 ctx->ctx_fl_going_zombie = 1;
2022 
2023                 /*
2024                  * force task to wake up from MASKED state
2025                  */
2026                 complete(&ctx->ctx_restart_done);
2027 
2028                 DPRINT(("waking up ctx_state=%d\n", state));
2029 
2030                 /*
2031                  * put ourself to sleep waiting for the other
2032                  * task to report completion
2033                  *
2034                  * the context is protected by mutex, therefore there
2035                  * is no risk of being notified of completion before
2036                  * begin actually on the waitq.
2037                  */
2038                 set_current_state(TASK_INTERRUPTIBLE);
2039                 add_wait_queue(&ctx->ctx_zombieq, &wait);
2040 
2041                 UNPROTECT_CTX(ctx, flags);
2042 
2043                 /*
2044                  * XXX: check for signals :
2045                  *      - ok for explicit close
2046                  *      - not ok when coming from exit_files()
2047                  */
2048                 schedule();
2049 
2050 
2051                 PROTECT_CTX(ctx, flags);
2052 
2053 
2054                 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2055                 set_current_state(TASK_RUNNING);
2056 
2057                 /*
2058                  * context is unloaded at this point
2059                  */
2060                 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2061         }
2062         else if (task != current) {
2063 #ifdef CONFIG_SMP
2064                 /*
2065                  * switch context to zombie state
2066                  */
2067                 ctx->ctx_state = PFM_CTX_ZOMBIE;
2068 
2069                 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2070                 /*
2071                  * cannot free the context on the spot. deferred until
2072                  * the task notices the ZOMBIE state
2073                  */
2074                 free_possible = 0;
2075 #else
2076                 pfm_context_unload(ctx, NULL, 0, regs);
2077 #endif
2078         }
2079 
2080 doit:
2081         /* reload state, may have changed during  opening of critical section */
2082         state = ctx->ctx_state;
2083 
2084         /*
2085          * the context is still attached to a task (possibly current)
2086          * we cannot destroy it right now
2087          */
2088 
2089         /*
2090          * we must free the sampling buffer right here because
2091          * we cannot rely on it being cleaned up later by the
2092          * monitored task. It is not possible to free vmalloc'ed
2093          * memory in pfm_load_regs(). Instead, we remove the buffer
2094          * now. should there be subsequent PMU overflow originally
2095          * meant for sampling, the will be converted to spurious
2096          * and that's fine because the monitoring tools is gone anyway.
2097          */
2098         if (ctx->ctx_smpl_hdr) {
2099                 smpl_buf_addr = ctx->ctx_smpl_hdr;
2100                 smpl_buf_size = ctx->ctx_smpl_size;
2101                 /* no more sampling */
2102                 ctx->ctx_smpl_hdr = NULL;
2103                 ctx->ctx_fl_is_sampling = 0;
2104         }
2105 
2106         DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2107                 state,
2108                 free_possible,
2109                 smpl_buf_addr,
2110                 smpl_buf_size));
2111 
2112         if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2113 
2114         /*
2115          * UNLOADED that the session has already been unreserved.
2116          */
2117         if (state == PFM_CTX_ZOMBIE) {
2118                 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2119         }
2120 
2121         /*
2122          * disconnect file descriptor from context must be done
2123          * before we unlock.
2124          */
2125         filp->private_data = NULL;
2126 
2127         /*
2128          * if we free on the spot, the context is now completely unreachable
2129          * from the callers side. The monitored task side is also cut, so we
2130          * can freely cut.
2131          *
2132          * If we have a deferred free, only the caller side is disconnected.
2133          */
2134         UNPROTECT_CTX(ctx, flags);
2135 
2136         /*
2137          * All memory free operations (especially for vmalloc'ed memory)
2138          * MUST be done with interrupts ENABLED.
2139          */
2140         if (smpl_buf_addr)  pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2141 
2142         /*
2143          * return the memory used by the context
2144          */
2145         if (free_possible) pfm_context_free(ctx);
2146 
2147         return 0;
2148 }
2149 
2150 static const struct file_operations pfm_file_ops = {
2151         .llseek         = no_llseek,
2152         .read           = pfm_read,
2153         .write          = pfm_write,
2154         .poll           = pfm_poll,
2155         .unlocked_ioctl = pfm_ioctl,
2156         .fasync         = pfm_fasync,
2157         .release        = pfm_close,
2158         .flush          = pfm_flush
2159 };
2160 
2161 static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2162 {
2163         return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2164                              d_inode(dentry)->i_ino);
2165 }
2166 
2167 static const struct dentry_operations pfmfs_dentry_operations = {
2168         .d_delete = always_delete_dentry,
2169         .d_dname = pfmfs_dname,
2170 };
2171 
2172 
2173 static struct file *
2174 pfm_alloc_file(pfm_context_t *ctx)
2175 {
2176         struct file *file;
2177         struct inode *inode;
2178         struct path path;
2179         struct qstr this = { .name = "" };
2180 
2181         /*
2182          * allocate a new inode
2183          */
2184         inode = new_inode(pfmfs_mnt->mnt_sb);
2185         if (!inode)
2186                 return ERR_PTR(-ENOMEM);
2187 
2188         DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2189 
2190         inode->i_mode = S_IFCHR|S_IRUGO;
2191         inode->i_uid  = current_fsuid();
2192         inode->i_gid  = current_fsgid();
2193 
2194         /*
2195          * allocate a new dcache entry
2196          */
2197         path.dentry = d_alloc(pfmfs_mnt->mnt_root, &this);
2198         if (!path.dentry) {
2199                 iput(inode);
2200                 return ERR_PTR(-ENOMEM);
2201         }
2202         path.mnt = mntget(pfmfs_mnt);
2203 
2204         d_add(path.dentry, inode);
2205 
2206         file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2207         if (IS_ERR(file)) {
2208                 path_put(&path);
2209                 return file;
2210         }
2211 
2212         file->f_flags = O_RDONLY;
2213         file->private_data = ctx;
2214 
2215         return file;
2216 }
2217 
2218 static int
2219 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2220 {
2221         DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2222 
2223         while (size > 0) {
2224                 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2225 
2226 
2227                 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2228                         return -ENOMEM;
2229 
2230                 addr  += PAGE_SIZE;
2231                 buf   += PAGE_SIZE;
2232                 size  -= PAGE_SIZE;
2233         }
2234         return 0;
2235 }
2236 
2237 /*
2238  * allocate a sampling buffer and remaps it into the user address space of the task
2239  */
2240 static int
2241 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2242 {
2243         struct mm_struct *mm = task->mm;
2244         struct vm_area_struct *vma = NULL;
2245         unsigned long size;
2246         void *smpl_buf;
2247 
2248 
2249         /*
2250          * the fixed header + requested size and align to page boundary
2251          */
2252         size = PAGE_ALIGN(rsize);
2253 
2254         DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2255 
2256         /*
2257          * check requested size to avoid Denial-of-service attacks
2258          * XXX: may have to refine this test
2259          * Check against address space limit.
2260          *
2261          * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2262          *      return -ENOMEM;
2263          */
2264         if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2265                 return -ENOMEM;
2266 
2267         /*
2268          * We do the easy to undo allocations first.
2269          *
2270          * pfm_rvmalloc(), clears the buffer, so there is no leak
2271          */
2272         smpl_buf = pfm_rvmalloc(size);
2273         if (smpl_buf == NULL) {
2274                 DPRINT(("Can't allocate sampling buffer\n"));
2275                 return -ENOMEM;
2276         }
2277 
2278         DPRINT(("smpl_buf @%p\n", smpl_buf));
2279 
2280         /* allocate vma */
2281         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2282         if (!vma) {
2283                 DPRINT(("Cannot allocate vma\n"));
2284                 goto error_kmem;
2285         }
2286         INIT_LIST_HEAD(&vma->anon_vma_chain);
2287 
2288         /*
2289          * partially initialize the vma for the sampling buffer
2290          */
2291         vma->vm_mm           = mm;
2292         vma->vm_file         = get_file(filp);
2293         vma->vm_flags        = VM_READ|VM_MAYREAD|VM_DONTEXPAND|VM_DONTDUMP;
2294         vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */
2295 
2296         /*
2297          * Now we have everything we need and we can initialize
2298          * and connect all the data structures
2299          */
2300 
2301         ctx->ctx_smpl_hdr   = smpl_buf;
2302         ctx->ctx_smpl_size  = size; /* aligned size */
2303 
2304         /*
2305          * Let's do the difficult operations next.
2306          *
2307          * now we atomically find some area in the address space and
2308          * remap the buffer in it.
2309          */
2310         down_write(&task->mm->mmap_sem);
2311 
2312         /* find some free area in address space, must have mmap sem held */
2313         vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2314         if (IS_ERR_VALUE(vma->vm_start)) {
2315                 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2316                 up_write(&task->mm->mmap_sem);
2317                 goto error;
2318         }
2319         vma->vm_end = vma->vm_start + size;
2320         vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2321 
2322         DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2323 
2324         /* can only be applied to current task, need to have the mm semaphore held when called */
2325         if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2326                 DPRINT(("Can't remap buffer\n"));
2327                 up_write(&task->mm->mmap_sem);
2328                 goto error;
2329         }
2330 
2331         /*
2332          * now insert the vma in the vm list for the process, must be
2333          * done with mmap lock held
2334          */
2335         insert_vm_struct(mm, vma);
2336 
2337         vm_stat_account(vma->vm_mm, vma->vm_flags, vma_pages(vma));
2338         up_write(&task->mm->mmap_sem);
2339 
2340         /*
2341          * keep track of user level virtual address
2342          */
2343         ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2344         *(unsigned long *)user_vaddr = vma->vm_start;
2345 
2346         return 0;
2347 
2348 error:
2349         kmem_cache_free(vm_area_cachep, vma);
2350 error_kmem:
2351         pfm_rvfree(smpl_buf, size);
2352 
2353         return -ENOMEM;
2354 }
2355 
2356 /*
2357  * XXX: do something better here
2358  */
2359 static int
2360 pfm_bad_permissions(struct task_struct *task)
2361 {
2362         const struct cred *tcred;
2363         kuid_t uid = current_uid();
2364         kgid_t gid = current_gid();
2365         int ret;
2366 
2367         rcu_read_lock();
2368         tcred = __task_cred(task);
2369 
2370         /* inspired by ptrace_attach() */
2371         DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2372                 from_kuid(&init_user_ns, uid),
2373                 from_kgid(&init_user_ns, gid),
2374                 from_kuid(&init_user_ns, tcred->euid),
2375                 from_kuid(&init_user_ns, tcred->suid),
2376                 from_kuid(&init_user_ns, tcred->uid),
2377                 from_kgid(&init_user_ns, tcred->egid),
2378                 from_kgid(&init_user_ns, tcred->sgid)));
2379 
2380         ret = ((!uid_eq(uid, tcred->euid))
2381                || (!uid_eq(uid, tcred->suid))
2382                || (!uid_eq(uid, tcred->uid))
2383                || (!gid_eq(gid, tcred->egid))
2384                || (!gid_eq(gid, tcred->sgid))
2385                || (!gid_eq(gid, tcred->gid))) && !capable(CAP_SYS_PTRACE);
2386 
2387         rcu_read_unlock();
2388         return ret;
2389 }
2390 
2391 static int
2392 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2393 {
2394         int ctx_flags;
2395 
2396         /* valid signal */
2397 
2398         ctx_flags = pfx->ctx_flags;
2399 
2400         if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2401 
2402                 /*
2403                  * cannot block in this mode
2404                  */
2405                 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2406                         DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2407                         return -EINVAL;
2408                 }
2409         } else {
2410         }
2411         /* probably more to add here */
2412 
2413         return 0;
2414 }
2415 
2416 static int
2417 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2418                      unsigned int cpu, pfarg_context_t *arg)
2419 {
2420         pfm_buffer_fmt_t *fmt = NULL;
2421         unsigned long size = 0UL;
2422         void *uaddr = NULL;
2423         void *fmt_arg = NULL;
2424         int ret = 0;
2425 #define PFM_CTXARG_BUF_ARG(a)   (pfm_buffer_fmt_t *)(a+1)
2426 
2427         /* invoke and lock buffer format, if found */
2428         fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2429         if (fmt == NULL) {
2430                 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2431                 return -EINVAL;
2432         }
2433 
2434         /*
2435          * buffer argument MUST be contiguous to pfarg_context_t
2436          */
2437         if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2438 
2439         ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2440 
2441         DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2442 
2443         if (ret) goto error;
2444 
2445         /* link buffer format and context */
2446         ctx->ctx_buf_fmt = fmt;
2447         ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2448 
2449         /*
2450          * check if buffer format wants to use perfmon buffer allocation/mapping service
2451          */
2452         ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2453         if (ret) goto error;
2454 
2455         if (size) {
2456                 /*
2457                  * buffer is always remapped into the caller's address space
2458                  */
2459                 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2460                 if (ret) goto error;
2461 
2462                 /* keep track of user address of buffer */
2463                 arg->ctx_smpl_vaddr = uaddr;
2464         }
2465         ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2466 
2467 error:
2468         return ret;
2469 }
2470 
2471 static void
2472 pfm_reset_pmu_state(pfm_context_t *ctx)
2473 {
2474         int i;
2475 
2476         /*
2477          * install reset values for PMC.
2478          */
2479         for (i=1; PMC_IS_LAST(i) == 0; i++) {
2480                 if (PMC_IS_IMPL(i) == 0) continue;
2481                 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2482                 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2483         }
2484         /*
2485          * PMD registers are set to 0UL when the context in memset()
2486          */
2487 
2488         /*
2489          * On context switched restore, we must restore ALL pmc and ALL pmd even
2490          * when they are not actively used by the task. In UP, the incoming process
2491          * may otherwise pick up left over PMC, PMD state from the previous process.
2492          * As opposed to PMD, stale PMC can cause harm to the incoming
2493          * process because they may change what is being measured.
2494          * Therefore, we must systematically reinstall the entire
2495          * PMC state. In SMP, the same thing is possible on the
2496          * same CPU but also on between 2 CPUs.
2497          *
2498          * The problem with PMD is information leaking especially
2499          * to user level when psr.sp=0
2500          *
2501          * There is unfortunately no easy way to avoid this problem
2502          * on either UP or SMP. This definitively slows down the
2503          * pfm_load_regs() function.
2504          */
2505 
2506          /*
2507           * bitmask of all PMCs accessible to this context
2508           *
2509           * PMC0 is treated differently.
2510           */
2511         ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2512 
2513         /*
2514          * bitmask of all PMDs that are accessible to this context
2515          */
2516         ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2517 
2518         DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2519 
2520         /*
2521          * useful in case of re-enable after disable
2522          */
2523         ctx->ctx_used_ibrs[0] = 0UL;
2524         ctx->ctx_used_dbrs[0] = 0UL;
2525 }
2526 
2527 static int
2528 pfm_ctx_getsize(void *arg, size_t *sz)
2529 {
2530         pfarg_context_t *req = (pfarg_context_t *)arg;
2531         pfm_buffer_fmt_t *fmt;
2532 
2533         *sz = 0;
2534 
2535         if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2536 
2537         fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2538         if (fmt == NULL) {
2539                 DPRINT(("cannot find buffer format\n"));
2540                 return -EINVAL;
2541         }
2542         /* get just enough to copy in user parameters */
2543         *sz = fmt->fmt_arg_size;
2544         DPRINT(("arg_size=%lu\n", *sz));
2545 
2546         return 0;
2547 }
2548 
2549 
2550 
2551 /*
2552  * cannot attach if :
2553  *      - kernel task
2554  *      - task not owned by caller
2555  *      - task incompatible with context mode
2556  */
2557 static int
2558 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2559 {
2560         /*
2561          * no kernel task or task not owner by caller
2562          */
2563         if (task->mm == NULL) {
2564                 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2565                 return -EPERM;
2566         }
2567         if (pfm_bad_permissions(task)) {
2568                 DPRINT(("no permission to attach to  [%d]\n", task_pid_nr(task)));
2569                 return -EPERM;
2570         }
2571         /*
2572          * cannot block in self-monitoring mode
2573          */
2574         if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2575                 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2576                 return -EINVAL;
2577         }
2578 
2579         if (task->exit_state == EXIT_ZOMBIE) {
2580                 DPRINT(("cannot attach to  zombie task [%d]\n", task_pid_nr(task)));
2581                 return -EBUSY;
2582         }
2583 
2584         /*
2585          * always ok for self
2586          */
2587         if (task == current) return 0;
2588 
2589         if (!task_is_stopped_or_traced(task)) {
2590                 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2591                 return -EBUSY;
2592         }
2593         /*
2594          * make sure the task is off any CPU
2595          */
2596         wait_task_inactive(task, 0);
2597 
2598         /* more to come... */
2599 
2600         return 0;
2601 }
2602 
2603 static int
2604 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2605 {
2606         struct task_struct *p = current;
2607         int ret;
2608 
2609         /* XXX: need to add more checks here */
2610         if (pid < 2) return -EPERM;
2611 
2612         if (pid != task_pid_vnr(current)) {
2613 
2614                 read_lock(&tasklist_lock);
2615 
2616                 p = find_task_by_vpid(pid);
2617 
2618                 /* make sure task cannot go away while we operate on it */
2619                 if (p) get_task_struct(p);
2620 
2621                 read_unlock(&tasklist_lock);
2622 
2623                 if (p == NULL) return -ESRCH;
2624         }
2625 
2626         ret = pfm_task_incompatible(ctx, p);
2627         if (ret == 0) {
2628                 *task = p;
2629         } else if (p != current) {
2630                 pfm_put_task(p);
2631         }
2632         return ret;
2633 }
2634 
2635 
2636 
2637 static int
2638 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2639 {
2640         pfarg_context_t *req = (pfarg_context_t *)arg;
2641         struct file *filp;
2642         struct path path;
2643         int ctx_flags;
2644         int fd;
2645         int ret;
2646 
2647         /* let's check the arguments first */
2648         ret = pfarg_is_sane(current, req);
2649         if (ret < 0)
2650                 return ret;
2651 
2652         ctx_flags = req->ctx_flags;
2653 
2654         ret = -ENOMEM;
2655 
2656         fd = get_unused_fd_flags(0);
2657         if (fd < 0)
2658                 return fd;
2659 
2660         ctx = pfm_context_alloc(ctx_flags);
2661         if (!ctx)
2662                 goto error;
2663 
2664         filp = pfm_alloc_file(ctx);
2665         if (IS_ERR(filp)) {
2666                 ret = PTR_ERR(filp);
2667                 goto error_file;
2668         }
2669 
2670         req->ctx_fd = ctx->ctx_fd = fd;
2671 
2672         /*
2673          * does the user want to sample?
2674          */
2675         if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2676                 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2677                 if (ret)
2678                         goto buffer_error;
2679         }
2680 
2681         DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2682                 ctx,
2683                 ctx_flags,
2684                 ctx->ctx_fl_system,
2685                 ctx->ctx_fl_block,
2686                 ctx->ctx_fl_excl_idle,
2687                 ctx->ctx_fl_no_msg,
2688                 ctx->ctx_fd));
2689 
2690         /*
2691          * initialize soft PMU state
2692          */
2693         pfm_reset_pmu_state(ctx);
2694 
2695         fd_install(fd, filp);
2696 
2697         return 0;
2698 
2699 buffer_error:
2700         path = filp->f_path;
2701         put_filp(filp);
2702         path_put(&path);
2703 
2704         if (ctx->ctx_buf_fmt) {
2705                 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2706         }
2707 error_file:
2708         pfm_context_free(ctx);
2709 
2710 error:
2711         put_unused_fd(fd);
2712         return ret;
2713 }
2714 
2715 static inline unsigned long
2716 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2717 {
2718         unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2719         unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2720         extern unsigned long carta_random32 (unsigned long seed);
2721 
2722         if (reg->flags & PFM_REGFL_RANDOM) {
2723                 new_seed = carta_random32(old_seed);
2724                 val -= (old_seed & mask);       /* counter values are negative numbers! */
2725                 if ((mask >> 32) != 0)
2726                         /* construct a full 64-bit random value: */
2727                         new_seed |= carta_random32(old_seed >> 32) << 32;
2728                 reg->seed = new_seed;
2729         }
2730         reg->lval = val;
2731         return val;
2732 }
2733 
2734 static void
2735 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2736 {
2737         unsigned long mask = ovfl_regs[0];
2738         unsigned long reset_others = 0UL;
2739         unsigned long val;
2740         int i;
2741 
2742         /*
2743          * now restore reset value on sampling overflowed counters
2744          */
2745         mask >>= PMU_FIRST_COUNTER;
2746         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2747 
2748                 if ((mask & 0x1UL) == 0UL) continue;
2749 
2750                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2751                 reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2752 
2753                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2754         }
2755 
2756         /*
2757          * Now take care of resetting the other registers
2758          */
2759         for(i = 0; reset_others; i++, reset_others >>= 1) {
2760 
2761                 if ((reset_others & 0x1) == 0) continue;
2762 
2763                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2764 
2765                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2766                           is_long_reset ? "long" : "short", i, val));
2767         }
2768 }
2769 
2770 static void
2771 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2772 {
2773         unsigned long mask = ovfl_regs[0];
2774         unsigned long reset_others = 0UL;
2775         unsigned long val;
2776         int i;
2777 
2778         DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2779 
2780         if (ctx->ctx_state == PFM_CTX_MASKED) {
2781                 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2782                 return;
2783         }
2784 
2785         /*
2786          * now restore reset value on sampling overflowed counters
2787          */
2788         mask >>= PMU_FIRST_COUNTER;
2789         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2790 
2791                 if ((mask & 0x1UL) == 0UL) continue;
2792 
2793                 val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2794                 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2795 
2796                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2797 
2798                 pfm_write_soft_counter(ctx, i, val);
2799         }
2800 
2801         /*
2802          * Now take care of resetting the other registers
2803          */
2804         for(i = 0; reset_others; i++, reset_others >>= 1) {
2805 
2806                 if ((reset_others & 0x1) == 0) continue;
2807 
2808                 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2809 
2810                 if (PMD_IS_COUNTING(i)) {
2811                         pfm_write_soft_counter(ctx, i, val);
2812                 } else {
2813                         ia64_set_pmd(i, val);
2814                 }
2815                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2816                           is_long_reset ? "long" : "short", i, val));
2817         }
2818         ia64_srlz_d();
2819 }
2820 
2821 static int
2822 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2823 {
2824         struct task_struct *task;
2825         pfarg_reg_t *req = (pfarg_reg_t *)arg;
2826         unsigned long value, pmc_pm;
2827         unsigned long smpl_pmds, reset_pmds, impl_pmds;
2828         unsigned int cnum, reg_flags, flags, pmc_type;
2829         int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2830         int is_monitor, is_counting, state;
2831         int ret = -EINVAL;
2832         pfm_reg_check_t wr_func;
2833 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2834 
2835         state     = ctx->ctx_state;
2836         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2837         is_system = ctx->ctx_fl_system;
2838         task      = ctx->ctx_task;
2839         impl_pmds = pmu_conf->impl_pmds[0];
2840 
2841         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2842 
2843         if (is_loaded) {
2844                 /*
2845                  * In system wide and when the context is loaded, access can only happen
2846                  * when the caller is running on the CPU being monitored by the session.
2847                  * It does not have to be the owner (ctx_task) of the context per se.
2848                  */
2849                 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2850                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2851                         return -EBUSY;
2852                 }
2853                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2854         }
2855         expert_mode = pfm_sysctl.expert_mode; 
2856 
2857         for (i = 0; i < count; i++, req++) {
2858 
2859                 cnum       = req->reg_num;
2860                 reg_flags  = req->reg_flags;
2861                 value      = req->reg_value;
2862                 smpl_pmds  = req->reg_smpl_pmds[0];
2863                 reset_pmds = req->reg_reset_pmds[0];
2864                 flags      = 0;
2865 
2866 
2867                 if (cnum >= PMU_MAX_PMCS) {
2868                         DPRINT(("pmc%u is invalid\n", cnum));
2869                         goto error;
2870                 }
2871 
2872                 pmc_type   = pmu_conf->pmc_desc[cnum].type;
2873                 pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2874                 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2875                 is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2876 
2877                 /*
2878                  * we reject all non implemented PMC as well
2879                  * as attempts to modify PMC[0-3] which are used
2880                  * as status registers by the PMU
2881                  */
2882                 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2883                         DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2884                         goto error;
2885                 }
2886                 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2887                 /*
2888                  * If the PMC is a monitor, then if the value is not the default:
2889                  *      - system-wide session: PMCx.pm=1 (privileged monitor)
2890                  *      - per-task           : PMCx.pm=0 (user monitor)
2891                  */
2892                 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2893                         DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2894                                 cnum,
2895                                 pmc_pm,
2896                                 is_system));
2897                         goto error;
2898                 }
2899 
2900                 if (is_counting) {
2901                         /*
2902                          * enforce generation of overflow interrupt. Necessary on all
2903                          * CPUs.
2904                          */
2905                         value |= 1 << PMU_PMC_OI;
2906 
2907                         if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2908                                 flags |= PFM_REGFL_OVFL_NOTIFY;
2909                         }
2910 
2911                         if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2912 
2913                         /* verify validity of smpl_pmds */
2914                         if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2915                                 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2916                                 goto error;
2917                         }
2918 
2919                         /* verify validity of reset_pmds */
2920                         if ((reset_pmds & impl_pmds) != reset_pmds) {
2921                                 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2922                                 goto error;
2923                         }
2924                 } else {
2925                         if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2926                                 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2927                                 goto error;
2928                         }
2929                         /* eventid on non-counting monitors are ignored */
2930                 }
2931 
2932                 /*
2933                  * execute write checker, if any
2934                  */
2935                 if (likely(expert_mode == 0 && wr_func)) {
2936                         ret = (*wr_func)(task, ctx, cnum, &value, regs);
2937                         if (ret) goto error;
2938                         ret = -EINVAL;
2939                 }
2940 
2941                 /*
2942                  * no error on this register
2943                  */
2944                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2945 
2946                 /*
2947                  * Now we commit the changes to the software state
2948                  */
2949 
2950                 /*
2951                  * update overflow information
2952                  */
2953                 if (is_counting) {
2954                         /*
2955                          * full flag update each time a register is programmed
2956                          */
2957                         ctx->ctx_pmds[cnum].flags = flags;
2958 
2959                         ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2960                         ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2961                         ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
2962 
2963                         /*
2964                          * Mark all PMDS to be accessed as used.
2965                          *
2966                          * We do not keep track of PMC because we have to
2967                          * systematically restore ALL of them.
2968                          *
2969                          * We do not update the used_monitors mask, because
2970                          * if we have not programmed them, then will be in
2971                          * a quiescent state, therefore we will not need to
2972                          * mask/restore then when context is MASKED.
2973                          */
2974                         CTX_USED_PMD(ctx, reset_pmds);
2975                         CTX_USED_PMD(ctx, smpl_pmds);
2976                         /*
2977                          * make sure we do not try to reset on
2978                          * restart because we have established new values
2979                          */
2980                         if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
2981                 }
2982                 /*
2983                  * Needed in case the user does not initialize the equivalent
2984                  * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2985                  * possible leak here.
2986                  */
2987                 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
2988 
2989                 /*
2990                  * keep track of the monitor PMC that we are using.
2991                  * we save the value of the pmc in ctx_pmcs[] and if
2992                  * the monitoring is not stopped for the context we also
2993                  * place it in the saved state area so that it will be
2994                  * picked up later by the context switch code.
2995                  *
2996                  * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
2997                  *
2998                  * The value in th_pmcs[] may be modified on overflow, i.e.,  when
2999                  * monitoring needs to be stopped.
3000                  */
3001                 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3002 
3003                 /*
3004                  * update context state
3005                  */
3006                 ctx->ctx_pmcs[cnum] = value;
3007 
3008                 if (is_loaded) {
3009                         /*
3010                          * write thread state
3011                          */
3012                         if (is_system == 0) ctx->th_pmcs[cnum] = value;
3013 
3014                         /*
3015                          * write hardware register if we can
3016                          */
3017                         if (can_access_pmu) {
3018                                 ia64_set_pmc(cnum, value);
3019                         }
3020 #ifdef CONFIG_SMP
3021                         else {
3022                                 /*
3023                                  * per-task SMP only here
3024                                  *
3025                                  * we are guaranteed that the task is not running on the other CPU,
3026                                  * we indicate that this PMD will need to be reloaded if the task
3027                                  * is rescheduled on the CPU it ran last on.
3028                                  */
3029                                 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3030                         }
3031 #endif
3032                 }
3033 
3034                 DPRINT(("pmc[%u]=0x%lx ld=%d apmu=%d flags=0x%x all_pmcs=0x%lx used_pmds=0x%lx eventid=%ld smpl_pmds=0x%lx reset_pmds=0x%lx reloads_pmcs=0x%lx used_monitors=0x%lx ovfl_regs=0x%lx\n",
3035                           cnum,
3036                           value,
3037                           is_loaded,
3038                           can_access_pmu,
3039                           flags,
3040                           ctx->ctx_all_pmcs[0],
3041                           ctx->ctx_used_pmds[0],
3042                           ctx->ctx_pmds[cnum].eventid,
3043                           smpl_pmds,
3044                           reset_pmds,
3045                           ctx->ctx_reload_pmcs[0],
3046                           ctx->ctx_used_monitors[0],
3047                           ctx->ctx_ovfl_regs[0]));
3048         }
3049 
3050         /*
3051          * make sure the changes are visible
3052          */
3053         if (can_access_pmu) ia64_srlz_d();
3054 
3055         return 0;
3056 error:
3057         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3058         return ret;
3059 }
3060 
3061 static int
3062 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3063 {
3064         struct task_struct *task;
3065         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3066         unsigned long value, hw_value, ovfl_mask;
3067         unsigned int cnum;
3068         int i, can_access_pmu = 0, state;
3069         int is_counting, is_loaded, is_system, expert_mode;
3070         int ret = -EINVAL;
3071         pfm_reg_check_t wr_func;
3072 
3073 
3074         state     = ctx->ctx_state;
3075         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3076         is_system = ctx->ctx_fl_system;
3077         ovfl_mask = pmu_conf->ovfl_val;
3078         task      = ctx->ctx_task;
3079 
3080         if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3081 
3082         /*
3083          * on both UP and SMP, we can only write to the PMC when the task is
3084          * the owner of the local PMU.
3085          */
3086         if (likely(is_loaded)) {
3087                 /*
3088                  * In system wide and when the context is loaded, access can only happen
3089                  * when the caller is running on the CPU being monitored by the session.
3090                  * It does not have to be the owner (ctx_task) of the context per se.
3091                  */
3092                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3093                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3094                         return -EBUSY;
3095                 }
3096                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3097         }
3098         expert_mode = pfm_sysctl.expert_mode; 
3099 
3100         for (i = 0; i < count; i++, req++) {
3101 
3102                 cnum  = req->reg_num;
3103                 value = req->reg_value;
3104 
3105                 if (!PMD_IS_IMPL(cnum)) {
3106                         DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3107                         goto abort_mission;
3108                 }
3109                 is_counting = PMD_IS_COUNTING(cnum);
3110                 wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3111 
3112                 /*
3113                  * execute write checker, if any
3114                  */
3115                 if (unlikely(expert_mode == 0 && wr_func)) {
3116                         unsigned long v = value;
3117 
3118                         ret = (*wr_func)(task, ctx, cnum, &v, regs);
3119                         if (ret) goto abort_mission;
3120 
3121                         value = v;
3122                         ret   = -EINVAL;
3123                 }
3124 
3125                 /*
3126                  * no error on this register
3127                  */
3128                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3129 
3130                 /*
3131                  * now commit changes to software state
3132                  */
3133                 hw_value = value;
3134 
3135                 /*
3136                  * update virtualized (64bits) counter
3137                  */
3138                 if (is_counting) {
3139                         /*
3140                          * write context state
3141                          */
3142                         ctx->ctx_pmds[cnum].lval = value;
3143 
3144                         /*
3145                          * when context is load we use the split value
3146                          */
3147                         if (is_loaded) {
3148                                 hw_value = value &  ovfl_mask;
3149                                 value    = value & ~ovfl_mask;
3150                         }
3151                 }
3152                 /*
3153                  * update reset values (not just for counters)
3154                  */
3155                 ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3156                 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3157 
3158                 /*
3159                  * update randomization parameters (not just for counters)
3160                  */
3161                 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3162                 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3163 
3164                 /*
3165                  * update context value
3166                  */
3167                 ctx->ctx_pmds[cnum].val  = value;
3168 
3169                 /*
3170                  * Keep track of what we use
3171                  *
3172                  * We do not keep track of PMC because we have to
3173                  * systematically restore ALL of them.
3174                  */
3175                 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3176 
3177                 /*
3178                  * mark this PMD register used as well
3179                  */
3180                 CTX_USED_PMD(ctx, RDEP(cnum));
3181 
3182                 /*
3183                  * make sure we do not try to reset on
3184                  * restart because we have established new values
3185                  */
3186                 if (is_counting && state == PFM_CTX_MASKED) {
3187                         ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3188                 }
3189 
3190                 if (is_loaded) {
3191                         /*
3192                          * write thread state
3193                          */
3194                         if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3195 
3196                         /*
3197                          * write hardware register if we can
3198                          */
3199                         if (can_access_pmu) {
3200                                 ia64_set_pmd(cnum, hw_value);
3201                         } else {
3202 #ifdef CONFIG_SMP
3203                                 /*
3204                                  * we are guaranteed that the task is not running on the other CPU,
3205                                  * we indicate that this PMD will need to be reloaded if the task
3206                                  * is rescheduled on the CPU it ran last on.
3207                                  */
3208                                 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3209 #endif
3210                         }
3211                 }
3212 
3213                 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3214                           "long_reset=0x%lx notify=%c seed=0x%lx mask=0x%lx used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3215                         cnum,
3216                         value,
3217                         is_loaded,
3218                         can_access_pmu,
3219                         hw_value,
3220                         ctx->ctx_pmds[cnum].val,
3221                         ctx->ctx_pmds[cnum].short_reset,
3222                         ctx->ctx_pmds[cnum].long_reset,
3223                         PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3224                         ctx->ctx_pmds[cnum].seed,
3225                         ctx->ctx_pmds[cnum].mask,
3226                         ctx->ctx_used_pmds[0],
3227                         ctx->ctx_pmds[cnum].reset_pmds[0],
3228                         ctx->ctx_reload_pmds[0],
3229                         ctx->ctx_all_pmds[0],
3230                         ctx->ctx_ovfl_regs[0]));
3231         }
3232 
3233         /*
3234          * make changes visible
3235          */
3236         if (can_access_pmu) ia64_srlz_d();
3237 
3238         return 0;
3239 
3240 abort_mission:
3241         /*
3242          * for now, we have only one possibility for error
3243          */
3244         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3245         return ret;
3246 }
3247 
3248 /*
3249  * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3250  * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3251  * interrupt is delivered during the call, it will be kept pending until we leave, making
3252  * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3253  * guaranteed to return consistent data to the user, it may simply be old. It is not
3254  * trivial to treat the overflow while inside the call because you may end up in
3255  * some module sampling buffer code causing deadlocks.
3256  */
3257 static int
3258 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3259 {
3260         struct task_struct *task;
3261         unsigned long val = 0UL, lval, ovfl_mask, sval;
3262         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3263         unsigned int cnum, reg_flags = 0;
3264         int i, can_access_pmu = 0, state;
3265         int is_loaded, is_system, is_counting, expert_mode;
3266         int ret = -EINVAL;
3267         pfm_reg_check_t rd_func;
3268 
3269         /*
3270          * access is possible when loaded only for
3271          * self-monitoring tasks or in UP mode
3272          */
3273 
3274         state     = ctx->ctx_state;
3275         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3276         is_system = ctx->ctx_fl_system;
3277         ovfl_mask = pmu_conf->ovfl_val;
3278         task      = ctx->ctx_task;
3279 
3280         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3281 
3282         if (likely(is_loaded)) {
3283                 /*
3284                  * In system wide and when the context is loaded, access can only happen
3285                  * when the caller is running on the CPU being monitored by the session.
3286                  * It does not have to be the owner (ctx_task) of the context per se.
3287                  */
3288                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3289                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3290                         return -EBUSY;
3291                 }
3292                 /*
3293                  * this can be true when not self-monitoring only in UP
3294                  */
3295                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3296 
3297                 if (can_access_pmu) ia64_srlz_d();
3298         }
3299         expert_mode = pfm_sysctl.expert_mode; 
3300 
3301         DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3302                 is_loaded,
3303                 can_access_pmu,
3304                 state));
3305 
3306         /*
3307          * on both UP and SMP, we can only read the PMD from the hardware register when
3308          * the task is the owner of the local PMU.
3309          */
3310 
3311         for (i = 0; i < count; i++, req++) {
3312 
3313                 cnum        = req->reg_num;
3314                 reg_flags   = req->reg_flags;
3315 
3316                 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3317                 /*
3318                  * we can only read the register that we use. That includes
3319                  * the one we explicitly initialize AND the one we want included
3320                  * in the sampling buffer (smpl_regs).
3321                  *
3322                  * Having this restriction allows optimization in the ctxsw routine
3323                  * without compromising security (leaks)
3324                  */
3325                 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3326 
3327                 sval        = ctx->ctx_pmds[cnum].val;
3328                 lval        = ctx->ctx_pmds[cnum].lval;
3329                 is_counting = PMD_IS_COUNTING(cnum);
3330 
3331                 /*
3332                  * If the task is not the current one, then we check if the
3333                  * PMU state is still in the local live register due to lazy ctxsw.
3334                  * If true, then we read directly from the registers.
3335                  */
3336                 if (can_access_pmu){
3337                         val = ia64_get_pmd(cnum);
3338                 } else {
3339                         /*
3340                          * context has been saved
3341                          * if context is zombie, then task does not exist anymore.
3342                          * In this case, we use the full value saved in the context (pfm_flush_regs()).
3343                          */
3344                         val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3345                 }
3346                 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3347 
3348                 if (is_counting) {
3349                         /*
3350                          * XXX: need to check for overflow when loaded
3351                          */
3352                         val &= ovfl_mask;
3353                         val += sval;
3354                 }
3355 
3356                 /*
3357                  * execute read checker, if any
3358                  */
3359                 if (unlikely(expert_mode == 0 && rd_func)) {
3360                         unsigned long v = val;
3361                         ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3362                         if (ret) goto error;
3363                         val = v;
3364                         ret = -EINVAL;
3365                 }
3366 
3367                 PFM_REG_RETFLAG_SET(reg_flags, 0);
3368 
3369                 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3370 
3371                 /*
3372                  * update register return value, abort all if problem during copy.
3373                  * we only modify the reg_flags field. no check mode is fine because
3374                  * access has been verified upfront in sys_perfmonctl().
3375                  */
3376                 req->reg_value            = val;
3377                 req->reg_flags            = reg_flags;
3378                 req->reg_last_reset_val   = lval;
3379         }
3380 
3381         return 0;
3382 
3383 error:
3384         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3385         return ret;
3386 }
3387 
3388 int
3389 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3390 {
3391         pfm_context_t *ctx;
3392 
3393         if (req == NULL) return -EINVAL;
3394 
3395         ctx = GET_PMU_CTX();
3396 
3397         if (ctx == NULL) return -EINVAL;
3398 
3399         /*
3400          * for now limit to current task, which is enough when calling
3401          * from overflow handler
3402          */
3403         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3404 
3405         return pfm_write_pmcs(ctx, req, nreq, regs);
3406 }
3407 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3408 
3409 int
3410 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3411 {
3412         pfm_context_t *ctx;
3413 
3414         if (req == NULL) return -EINVAL;
3415 
3416         ctx = GET_PMU_CTX();
3417 
3418         if (ctx == NULL) return -EINVAL;
3419 
3420         /*
3421          * for now limit to current task, which is enough when calling
3422          * from overflow handler
3423          */
3424         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3425 
3426         return pfm_read_pmds(ctx, req, nreq, regs);
3427 }
3428 EXPORT_SYMBOL(pfm_mod_read_pmds);
3429 
3430 /*
3431  * Only call this function when a process it trying to
3432  * write the debug registers (reading is always allowed)
3433  */
3434 int
3435 pfm_use_debug_registers(struct task_struct *task)
3436 {
3437         pfm_context_t *ctx = task->thread.pfm_context;
3438         unsigned long flags;
3439         int ret = 0;
3440 
3441         if (pmu_conf->use_rr_dbregs == 0) return 0;
3442 
3443         DPRINT(("called for [%d]\n", task_pid_nr(task)));
3444 
3445         /*
3446          * do it only once
3447          */
3448         if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3449 
3450         /*
3451          * Even on SMP, we do not need to use an atomic here because
3452          * the only way in is via ptrace() and this is possible only when the
3453          * process is stopped. Even in the case where the ctxsw out is not totally
3454          * completed by the time we come here, there is no way the 'stopped' process
3455          * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3456          * So this is always safe.
3457          */
3458         if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3459 
3460         LOCK_PFS(flags);
3461 
3462         /*
3463          * We cannot allow setting breakpoints when system wide monitoring
3464          * sessions are using the debug registers.
3465          */
3466         if (pfm_sessions.pfs_sys_use_dbregs> 0)
3467                 ret = -1;
3468         else
3469                 pfm_sessions.pfs_ptrace_use_dbregs++;
3470 
3471         DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3472                   pfm_sessions.pfs_ptrace_use_dbregs,
3473                   pfm_sessions.pfs_sys_use_dbregs,
3474                   task_pid_nr(task), ret));
3475 
3476         UNLOCK_PFS(flags);
3477 
3478         return ret;
3479 }
3480 
3481 /*
3482  * This function is called for every task that exits with the
3483  * IA64_THREAD_DBG_VALID set. This indicates a task which was
3484  * able to use the debug registers for debugging purposes via
3485  * ptrace(). Therefore we know it was not using them for
3486  * performance monitoring, so we only decrement the number
3487  * of "ptraced" debug register users to keep the count up to date
3488  */
3489 int
3490 pfm_release_debug_registers(struct task_struct *task)
3491 {
3492         unsigned long flags;
3493         int ret;
3494 
3495         if (pmu_conf->use_rr_dbregs == 0) return 0;
3496 
3497         LOCK_PFS(flags);
3498         if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3499                 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3500                 ret = -1;
3501         }  else {
3502                 pfm_sessions.pfs_ptrace_use_dbregs--;
3503                 ret = 0;
3504         }
3505         UNLOCK_PFS(flags);
3506 
3507         return ret;
3508 }
3509 
3510 static int
3511 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3512 {
3513         struct task_struct *task;
3514         pfm_buffer_fmt_t *fmt;
3515         pfm_ovfl_ctrl_t rst_ctrl;
3516         int state, is_system;
3517         int ret = 0;
3518 
3519         state     = ctx->ctx_state;
3520         fmt       = ctx->ctx_buf_fmt;
3521         is_system = ctx->ctx_fl_system;
3522         task      = PFM_CTX_TASK(ctx);
3523 
3524         switch(state) {
3525                 case PFM_CTX_MASKED:
3526                         break;
3527                 case PFM_CTX_LOADED: 
3528                         if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3529                         /* fall through */
3530                 case PFM_CTX_UNLOADED:
3531                 case PFM_CTX_ZOMBIE:
3532                         DPRINT(("invalid state=%d\n", state));
3533                         return -EBUSY;
3534                 default:
3535                         DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3536                         return -EINVAL;
3537         }
3538 
3539         /*
3540          * In system wide and when the context is loaded, access can only happen
3541          * when the caller is running on the CPU being monitored by the session.
3542          * It does not have to be the owner (ctx_task) of the context per se.
3543          */
3544         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3545                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3546                 return -EBUSY;
3547         }
3548 
3549         /* sanity check */
3550         if (unlikely(task == NULL)) {
3551                 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3552                 return -EINVAL;
3553         }
3554 
3555         if (task == current || is_system) {
3556 
3557                 fmt = ctx->ctx_buf_fmt;
3558 
3559                 DPRINT(("restarting self %d ovfl=0x%lx\n",
3560                         task_pid_nr(task),
3561                         ctx->ctx_ovfl_regs[0]));
3562 
3563                 if (CTX_HAS_SMPL(ctx)) {
3564 
3565                         prefetch(ctx->ctx_smpl_hdr);
3566 
3567                         rst_ctrl.bits.mask_monitoring = 0;
3568                         rst_ctrl.bits.reset_ovfl_pmds = 0;
3569 
3570                         if (state == PFM_CTX_LOADED)
3571                                 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3572                         else
3573                                 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3574                 } else {
3575                         rst_ctrl.bits.mask_monitoring = 0;
3576                         rst_ctrl.bits.reset_ovfl_pmds = 1;
3577                 }
3578 
3579                 if (ret == 0) {
3580                         if (rst_ctrl.bits.reset_ovfl_pmds)
3581                                 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3582 
3583                         if (rst_ctrl.bits.mask_monitoring == 0) {
3584                                 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3585 
3586                                 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3587                         } else {
3588                                 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3589 
3590                                 // cannot use pfm_stop_monitoring(task, regs);
3591                         }
3592                 }
3593                 /*
3594                  * clear overflowed PMD mask to remove any stale information
3595                  */
3596                 ctx->ctx_ovfl_regs[0] = 0UL;
3597 
3598                 /*
3599                  * back to LOADED state
3600                  */
3601                 ctx->ctx_state = PFM_CTX_LOADED;
3602 
3603                 /*
3604                  * XXX: not really useful for self monitoring
3605                  */
3606                 ctx->ctx_fl_can_restart = 0;
3607 
3608                 return 0;
3609         }
3610 
3611         /* 
3612          * restart another task
3613          */
3614 
3615         /*
3616          * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3617          * one is seen by the task.
3618          */
3619         if (state == PFM_CTX_MASKED) {
3620                 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3621                 /*
3622                  * will prevent subsequent restart before this one is
3623                  * seen by other task
3624                  */
3625                 ctx->ctx_fl_can_restart = 0;
3626         }
3627 
3628         /*
3629          * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3630          * the task is blocked or on its way to block. That's the normal
3631          * restart path. If the monitoring is not masked, then the task
3632          * can be actively monitoring and we cannot directly intervene.
3633          * Therefore we use the trap mechanism to catch the task and
3634          * force it to reset the buffer/reset PMDs.
3635          *
3636          * if non-blocking, then we ensure that the task will go into
3637          * pfm_handle_work() before returning to user mode.
3638          *
3639          * We cannot explicitly reset another task, it MUST always
3640          * be done by the task itself. This works for system wide because
3641          * the tool that is controlling the session is logically doing 
3642          * "self-monitoring".
3643          */
3644         if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3645                 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3646                 complete(&ctx->ctx_restart_done);
3647         } else {
3648                 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3649 
3650                 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3651 
3652                 PFM_SET_WORK_PENDING(task, 1);
3653 
3654                 set_notify_resume(task);
3655 
3656                 /*
3657                  * XXX: send reschedule if task runs on another CPU
3658                  */
3659         }
3660         return 0;
3661 }
3662 
3663 static int
3664 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3665 {
3666         unsigned int m = *(unsigned int *)arg;
3667 
3668         pfm_sysctl.debug = m == 0 ? 0 : 1;
3669 
3670         printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3671 
3672         if (m == 0) {
3673                 memset(pfm_stats, 0, sizeof(pfm_stats));
3674                 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3675         }
3676         return 0;
3677 }
3678 
3679 /*
3680  * arg can be NULL and count can be zero for this function
3681  */
3682 static int
3683 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3684 {
3685         struct thread_struct *thread = NULL;
3686         struct task_struct *task;
3687         pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3688         unsigned long flags;
3689         dbreg_t dbreg;
3690         unsigned int rnum;
3691         int first_time;
3692         int ret = 0, state;
3693         int i, can_access_pmu = 0;
3694         int is_system, is_loaded;
3695 
3696         if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3697 
3698         state     = ctx->ctx_state;
3699         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3700         is_system = ctx->ctx_fl_system;
3701         task      = ctx->ctx_task;
3702 
3703         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3704 
3705         /*
3706          * on both UP and SMP, we can only write to the PMC when the task is
3707          * the owner of the local PMU.
3708          */
3709         if (is_loaded) {
3710                 thread = &task->thread;
3711                 /*
3712                  * In system wide and when the context is loaded, access can only happen
3713                  * when the caller is running on the CPU being monitored by the session.
3714                  * It does not have to be the owner (ctx_task) of the context per se.
3715                  */
3716                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3717                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3718                         return -EBUSY;
3719                 }
3720                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3721         }
3722 
3723         /*
3724          * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3725          * ensuring that no real breakpoint can be installed via this call.
3726          *
3727          * IMPORTANT: regs can be NULL in this function
3728          */
3729 
3730         first_time = ctx->ctx_fl_using_dbreg == 0;
3731 
3732         /*
3733          * don't bother if we are loaded and task is being debugged
3734          */
3735         if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3736                 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3737                 return -EBUSY;
3738         }
3739 
3740         /*
3741          * check for debug registers in system wide mode
3742          *
3743          * If though a check is done in pfm_context_load(),
3744          * we must repeat it here, in case the registers are
3745          * written after the context is loaded
3746          */
3747         if (is_loaded) {
3748                 LOCK_PFS(flags);
3749 
3750                 if (first_time && is_system) {
3751                         if (pfm_sessions.pfs_ptrace_use_dbregs)
3752                                 ret = -EBUSY;
3753                         else
3754                                 pfm_sessions.pfs_sys_use_dbregs++;
3755                 }
3756                 UNLOCK_PFS(flags);
3757         }
3758 
3759         if (ret != 0) return ret;
3760 
3761         /*
3762          * mark ourself as user of the debug registers for
3763          * perfmon purposes.
3764          */
3765         ctx->ctx_fl_using_dbreg = 1;
3766 
3767         /*
3768          * clear hardware registers to make sure we don't
3769          * pick up stale state.
3770          *
3771          * for a system wide session, we do not use
3772          * thread.dbr, thread.ibr because this process
3773          * never leaves the current CPU and the state
3774          * is shared by all processes running on it
3775          */
3776         if (first_time && can_access_pmu) {
3777                 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3778                 for (i=0; i < pmu_conf->num_ibrs; i++) {
3779                         ia64_set_ibr(i, 0UL);
3780                         ia64_dv_serialize_instruction();
3781                 }
3782                 ia64_srlz_i();
3783                 for (i=0; i < pmu_conf->num_dbrs; i++) {
3784                         ia64_set_dbr(i, 0UL);
3785                         ia64_dv_serialize_data();
3786                 }
3787                 ia64_srlz_d();
3788         }
3789 
3790         /*
3791          * Now install the values into the registers
3792          */
3793         for (i = 0; i < count; i++, req++) {
3794 
3795                 rnum      = req->dbreg_num;
3796                 dbreg.val = req->dbreg_value;
3797 
3798                 ret = -EINVAL;
3799 
3800                 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3801                         DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3802                                   rnum, dbreg.val, mode, i, count));
3803 
3804                         goto abort_mission;
3805                 }
3806 
3807                 /*
3808                  * make sure we do not install enabled breakpoint
3809                  */
3810                 if (rnum & 0x1) {
3811                         if (mode == PFM_CODE_RR)
3812                                 dbreg.ibr.ibr_x = 0;
3813                         else
3814                                 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3815                 }
3816 
3817                 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3818 
3819                 /*
3820                  * Debug registers, just like PMC, can only be modified
3821                  * by a kernel call. Moreover, perfmon() access to those
3822                  * registers are centralized in this routine. The hardware
3823                  * does not modify the value of these registers, therefore,
3824                  * if we save them as they are written, we can avoid having
3825                  * to save them on context switch out. This is made possible
3826                  * by the fact that when perfmon uses debug registers, ptrace()
3827                  * won't be able to modify them concurrently.
3828                  */
3829                 if (mode == PFM_CODE_RR) {
3830                         CTX_USED_IBR(ctx, rnum);
3831 
3832                         if (can_access_pmu) {
3833                                 ia64_set_ibr(rnum, dbreg.val);
3834                                 ia64_dv_serialize_instruction();
3835                         }
3836 
3837                         ctx->ctx_ibrs[rnum] = dbreg.val;
3838 
3839                         DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3840                                 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3841                 } else {
3842                         CTX_USED_DBR(ctx, rnum);
3843 
3844                         if (can_access_pmu) {
3845                                 ia64_set_dbr(rnum, dbreg.val);
3846                                 ia64_dv_serialize_data();
3847                         }
3848                         ctx->ctx_dbrs[rnum] = dbreg.val;
3849 
3850                         DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3851                                 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3852                 }
3853         }
3854 
3855         return 0;
3856 
3857 abort_mission:
3858         /*
3859          * in case it was our first attempt, we undo the global modifications
3860          */
3861         if (first_time) {
3862                 LOCK_PFS(flags);
3863                 if (ctx->ctx_fl_system) {
3864                         pfm_sessions.pfs_sys_use_dbregs--;
3865                 }
3866                 UNLOCK_PFS(flags);
3867                 ctx->ctx_fl_using_dbreg = 0;
3868         }
3869         /*
3870          * install error return flag
3871          */
3872         PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3873 
3874         return ret;
3875 }
3876 
3877 static int
3878 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3879 {
3880         return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3881 }
3882 
3883 static int
3884 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3885 {
3886         return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3887 }
3888 
3889 int
3890 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3891 {
3892         pfm_context_t *ctx;
3893 
3894         if (req == NULL) return -EINVAL;
3895 
3896         ctx = GET_PMU_CTX();
3897 
3898         if (ctx == NULL) return -EINVAL;
3899 
3900         /*
3901          * for now limit to current task, which is enough when calling
3902          * from overflow handler
3903          */
3904         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3905 
3906         return pfm_write_ibrs(ctx, req, nreq, regs);
3907 }
3908 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3909 
3910 int
3911 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3912 {
3913         pfm_context_t *ctx;
3914 
3915         if (req == NULL) return -EINVAL;
3916 
3917         ctx = GET_PMU_CTX();
3918 
3919         if (ctx == NULL) return -EINVAL;
3920 
3921         /*
3922          * for now limit to current task, which is enough when calling
3923          * from overflow handler
3924          */
3925         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3926 
3927         return pfm_write_dbrs(ctx, req, nreq, regs);
3928 }
3929 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3930 
3931 
3932 static int
3933 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3934 {
3935         pfarg_features_t *req = (pfarg_features_t *)arg;
3936 
3937         req->ft_version = PFM_VERSION;
3938         return 0;
3939 }
3940 
3941 static int
3942 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3943 {
3944         struct pt_regs *tregs;
3945         struct task_struct *task = PFM_CTX_TASK(ctx);
3946         int state, is_system;
3947 
3948         state     = ctx->ctx_state;
3949         is_system = ctx->ctx_fl_system;
3950 
3951         /*
3952          * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3953          */
3954         if (state == PFM_CTX_UNLOADED) return -EINVAL;
3955 
3956         /*
3957          * In system wide and when the context is loaded, access can only happen
3958          * when the caller is running on the CPU being monitored by the session.
3959          * It does not have to be the owner (ctx_task) of the context per se.
3960          */
3961         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3962                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3963                 return -EBUSY;
3964         }
3965         DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3966                 task_pid_nr(PFM_CTX_TASK(ctx)),
3967                 state,
3968                 is_system));
3969         /*
3970          * in system mode, we need to update the PMU directly
3971          * and the user level state of the caller, which may not
3972          * necessarily be the creator of the context.
3973          */
3974         if (is_system) {
3975                 /*
3976                  * Update local PMU first
3977                  *
3978                  * disable dcr pp
3979                  */
3980                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
3981                 ia64_srlz_i();
3982 
3983                 /*
3984                  * update local cpuinfo
3985                  */
3986                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
3987 
3988                 /*
3989                  * stop monitoring, does srlz.i
3990                  */
3991                 pfm_clear_psr_pp();
3992 
3993                 /*
3994                  * stop monitoring in the caller
3995                  */
3996                 ia64_psr(regs)->pp = 0;
3997 
3998                 return 0;
3999         }
4000         /*
4001          * per-task mode
4002          */
4003 
4004         if (task == current) {
4005                 /* stop monitoring  at kernel level */
4006                 pfm_clear_psr_up();
4007 
4008                 /*
4009                  * stop monitoring at the user level
4010                  */
4011                 ia64_psr(regs)->up = 0;
4012         } else {
4013                 tregs = task_pt_regs(task);
4014 
4015                 /*
4016                  * stop monitoring at the user level
4017                  */
4018                 ia64_psr(tregs)->up = 0;
4019 
4020                 /*
4021                  * monitoring disabled in kernel at next reschedule
4022                  */
4023                 ctx->ctx_saved_psr_up = 0;
4024                 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4025         }
4026         return 0;
4027 }
4028 
4029 
4030 static int
4031 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4032 {
4033         struct pt_regs *tregs;
4034         int state, is_system;
4035 
4036         state     = ctx->ctx_state;
4037         is_system = ctx->ctx_fl_system;
4038 
4039         if (state != PFM_CTX_LOADED) return -EINVAL;
4040 
4041         /*
4042          * In system wide and when the context is loaded, access can only happen
4043          * when the caller is running on the CPU being monitored by the session.
4044          * It does not have to be the owner (ctx_task) of the context per se.
4045          */
4046         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4047                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4048                 return -EBUSY;
4049         }
4050 
4051         /*
4052          * in system mode, we need to update the PMU directly
4053          * and the user level state of the caller, which may not
4054          * necessarily be the creator of the context.
4055          */
4056         if (is_system) {
4057 
4058                 /*
4059                  * set user level psr.pp for the caller
4060                  */
4061                 ia64_psr(regs)->pp = 1;
4062 
4063                 /*
4064                  * now update the local PMU and cpuinfo
4065                  */
4066                 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4067 
4068                 /*
4069                  * start monitoring at kernel level
4070                  */
4071                 pfm_set_psr_pp();
4072 
4073                 /* enable dcr pp */
4074                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4075                 ia64_srlz_i();
4076 
4077                 return 0;
4078         }
4079 
4080         /*
4081          * per-process mode
4082          */
4083 
4084         if (ctx->ctx_task == current) {
4085 
4086                 /* start monitoring at kernel level */
4087                 pfm_set_psr_up();
4088 
4089                 /*
4090                  * activate monitoring at user level
4091                  */
4092                 ia64_psr(regs)->up = 1;
4093 
4094         } else {
4095                 tregs = task_pt_regs(ctx->ctx_task);
4096 
4097                 /*
4098                  * start monitoring at the kernel level the next
4099                  * time the task is scheduled
4100                  */
4101                 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4102 
4103                 /*
4104                  * activate monitoring at user level
4105                  */
4106                 ia64_psr(tregs)->up = 1;
4107         }
4108         return 0;
4109 }
4110 
4111 static int
4112 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4113 {
4114         pfarg_reg_t *req = (pfarg_reg_t *)arg;
4115         unsigned int cnum;
4116         int i;
4117         int ret = -EINVAL;
4118 
4119         for (i = 0; i < count; i++, req++) {
4120 
4121                 cnum = req->reg_num;
4122 
4123                 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4124 
4125                 req->reg_value = PMC_DFL_VAL(cnum);
4126 
4127                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4128 
4129                 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4130         }
4131         return 0;
4132 
4133 abort_mission:
4134         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4135         return ret;
4136 }
4137 
4138 static int
4139 pfm_check_task_exist(pfm_context_t *ctx)
4140 {
4141         struct task_struct *g, *t;
4142         int ret = -ESRCH;
4143 
4144         read_lock(&tasklist_lock);
4145 
4146         do_each_thread (g, t) {
4147                 if (t->thread.pfm_context == ctx) {
4148                         ret = 0;
4149                         goto out;
4150                 }
4151         } while_each_thread (g, t);
4152 out:
4153         read_unlock(&tasklist_lock);
4154 
4155         DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4156 
4157         return ret;
4158 }
4159 
4160 static int
4161 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4162 {
4163         struct task_struct *task;
4164         struct thread_struct *thread;
4165         struct pfm_context_t *old;
4166         unsigned long flags;
4167 #ifndef CONFIG_SMP
4168         struct task_struct *owner_task = NULL;
4169 #endif
4170         pfarg_load_t *req = (pfarg_load_t *)arg;
4171         unsigned long *pmcs_source, *pmds_source;
4172         int the_cpu;
4173         int ret = 0;
4174         int state, is_system, set_dbregs = 0;
4175 
4176         state     = ctx->ctx_state;
4177         is_system = ctx->ctx_fl_system;
4178         /*
4179          * can only load from unloaded or terminated state
4180          */
4181         if (state != PFM_CTX_UNLOADED) {
4182                 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4183                         req->load_pid,
4184                         ctx->ctx_state));
4185                 return -EBUSY;
4186         }
4187 
4188         DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4189 
4190         if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4191                 DPRINT(("cannot use blocking mode on self\n"));
4192                 return -EINVAL;
4193         }
4194 
4195         ret = pfm_get_task(ctx, req->load_pid, &task);
4196         if (ret) {
4197                 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4198                 return ret;
4199         }
4200 
4201         ret = -EINVAL;
4202 
4203         /*
4204          * system wide is self monitoring only
4205          */
4206         if (is_system && task != current) {
4207                 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4208                         req->load_pid));
4209                 goto error;
4210         }
4211 
4212         thread = &task->thread;
4213 
4214         ret = 0;
4215         /*
4216          * cannot load a context which is using range restrictions,
4217          * into a task that is being debugged.
4218          */
4219         if (ctx->ctx_fl_using_dbreg) {
4220                 if (thread->flags & IA64_THREAD_DBG_VALID) {
4221                         ret = -EBUSY;
4222                         DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4223                         goto error;
4224                 }
4225                 LOCK_PFS(flags);
4226 
4227                 if (is_system) {
4228                         if (pfm_sessions.pfs_ptrace_use_dbregs) {
4229                                 DPRINT(("cannot load [%d] dbregs in use\n",
4230                                                         task_pid_nr(task)));
4231                                 ret = -EBUSY;
4232                         } else {
4233                                 pfm_sessions.pfs_sys_use_dbregs++;
4234                                 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4235                                 set_dbregs = 1;
4236                         }
4237                 }
4238 
4239                 UNLOCK_PFS(flags);
4240 
4241                 if (ret) goto error;
4242         }
4243 
4244         /*
4245          * SMP system-wide monitoring implies self-monitoring.
4246          *
4247          * The programming model expects the task to
4248          * be pinned on a CPU throughout the session.
4249          * Here we take note of the current CPU at the
4250          * time the context is loaded. No call from
4251          * another CPU will be allowed.
4252          *
4253          * The pinning via shed_setaffinity()
4254          * must be done by the calling task prior
4255          * to this call.
4256          *
4257          * systemwide: keep track of CPU this session is supposed to run on
4258          */
4259         the_cpu = ctx->ctx_cpu = smp_processor_id();
4260 
4261         ret = -EBUSY;
4262         /*
4263          * now reserve the session
4264          */
4265         ret = pfm_reserve_session(current, is_system, the_cpu);
4266         if (ret) goto error;
4267 
4268         /*
4269          * task is necessarily stopped at this point.
4270          *
4271          * If the previous context was zombie, then it got removed in
4272          * pfm_save_regs(). Therefore we should not see it here.
4273          * If we see a context, then this is an active context
4274          *
4275          * XXX: needs to be atomic
4276          */
4277         DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4278                 thread->pfm_context, ctx));
4279 
4280         ret = -EBUSY;
4281         old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4282         if (old != NULL) {
4283                 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4284                 goto error_unres;
4285         }
4286 
4287         pfm_reset_msgq(ctx);
4288 
4289         ctx->ctx_state = PFM_CTX_LOADED;
4290 
4291         /*
4292          * link context to task
4293          */
4294         ctx->ctx_task = task;
4295 
4296         if (is_system) {
4297                 /*
4298                  * we load as stopped
4299                  */
4300                 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4301                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4302 
4303                 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4304         } else {
4305                 thread->flags |= IA64_THREAD_PM_VALID;
4306         }
4307 
4308         /*
4309          * propagate into thread-state
4310          */
4311         pfm_copy_pmds(task, ctx);
4312         pfm_copy_pmcs(task, ctx);
4313 
4314         pmcs_source = ctx->th_pmcs;
4315         pmds_source = ctx->th_pmds;
4316 
4317         /*
4318          * always the case for system-wide
4319          */
4320         if (task == current) {
4321 
4322                 if (is_system == 0) {
4323 
4324                         /* allow user level control */
4325                         ia64_psr(regs)->sp = 0;
4326                         DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4327 
4328                         SET_LAST_CPU(ctx, smp_processor_id());
4329                         INC_ACTIVATION();
4330                         SET_ACTIVATION(ctx);
4331 #ifndef CONFIG_SMP
4332                         /*
4333                          * push the other task out, if any
4334                          */
4335                         owner_task = GET_PMU_OWNER();
4336                         if (owner_task) pfm_lazy_save_regs(owner_task);
4337 #endif
4338                 }
4339                 /*
4340                  * load all PMD from ctx to PMU (as opposed to thread state)
4341                  * restore all PMC from ctx to PMU
4342                  */
4343                 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4344                 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4345 
4346                 ctx->ctx_reload_pmcs[0] = 0UL;
4347                 ctx->ctx_reload_pmds[0] = 0UL;
4348 
4349                 /*
4350                  * guaranteed safe by earlier check against DBG_VALID
4351                  */
4352                 if (ctx->ctx_fl_using_dbreg) {
4353                         pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4354                         pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4355                 }
4356                 /*
4357                  * set new ownership
4358                  */
4359                 SET_PMU_OWNER(task, ctx);
4360 
4361                 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4362         } else {
4363                 /*
4364                  * when not current, task MUST be stopped, so this is safe
4365                  */
4366                 regs = task_pt_regs(task);
4367 
4368                 /* force a full reload */
4369                 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4370                 SET_LAST_CPU(ctx, -1);
4371 
4372                 /* initial saved psr (stopped) */
4373                 ctx->ctx_saved_psr_up = 0UL;
4374                 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4375         }
4376 
4377         ret = 0;
4378 
4379 error_unres:
4380         if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4381 error:
4382         /*
4383          * we must undo the dbregs setting (for system-wide)
4384          */
4385         if (ret && set_dbregs) {
4386                 LOCK_PFS(flags);
4387                 pfm_sessions.pfs_sys_use_dbregs--;
4388                 UNLOCK_PFS(flags);
4389         }
4390         /*
4391          * release task, there is now a link with the context
4392          */
4393         if (is_system == 0 && task != current) {
4394                 pfm_put_task(task);
4395 
4396                 if (ret == 0) {
4397                         ret = pfm_check_task_exist(ctx);
4398                         if (ret) {
4399                                 ctx->ctx_state = PFM_CTX_UNLOADED;
4400                                 ctx->ctx_task  = NULL;
4401                         }
4402                 }
4403         }
4404         return ret;
4405 }
4406 
4407 /*
4408  * in this function, we do not need to increase the use count
4409  * for the task via get_task_struct(), because we hold the
4410  * context lock. If the task were to disappear while having
4411  * a context attached, it would go through pfm_exit_thread()
4412  * which also grabs the context lock  and would therefore be blocked
4413  * until we are here.
4414  */
4415 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4416 
4417 static int
4418 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4419 {
4420         struct task_struct *task = PFM_CTX_TASK(ctx);
4421         struct pt_regs *tregs;
4422         int prev_state, is_system;
4423         int ret;
4424 
4425         DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4426 
4427         prev_state = ctx->ctx_state;
4428         is_system  = ctx->ctx_fl_system;
4429 
4430         /*
4431          * unload only when necessary
4432          */
4433         if (prev_state == PFM_CTX_UNLOADED) {
4434                 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4435                 return 0;
4436         }
4437 
4438         /*
4439          * clear psr and dcr bits
4440          */
4441         ret = pfm_stop(ctx, NULL, 0, regs);
4442         if (ret) return ret;
4443 
4444         ctx->ctx_state = PFM_CTX_UNLOADED;
4445 
4446         /*
4447          * in system mode, we need to update the PMU directly
4448          * and the user level state of the caller, which may not
4449          * necessarily be the creator of the context.
4450          */
4451         if (is_system) {
4452 
4453                 /*
4454                  * Update cpuinfo
4455                  *
4456                  * local PMU is taken care of in pfm_stop()
4457                  */
4458                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4459                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4460 
4461                 /*
4462                  * save PMDs in context
4463                  * release ownership
4464                  */
4465                 pfm_flush_pmds(current, ctx);
4466 
4467                 /*
4468                  * at this point we are done with the PMU
4469                  * so we can unreserve the resource.
4470                  */
4471                 if (prev_state != PFM_CTX_ZOMBIE) 
4472                         pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4473 
4474                 /*
4475                  * disconnect context from task
4476                  */
4477                 task->thread.pfm_context = NULL;
4478                 /*
4479                  * disconnect task from context
4480                  */
4481                 ctx->ctx_task = NULL;
4482 
4483                 /*
4484                  * There is nothing more to cleanup here.
4485                  */
4486                 return 0;
4487         }
4488 
4489         /*
4490          * per-task mode
4491          */
4492         tregs = task == current ? regs : task_pt_regs(task);
4493 
4494         if (task == current) {
4495                 /*
4496                  * cancel user level control
4497                  */
4498                 ia64_psr(regs)->sp = 1;
4499 
4500                 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4501         }
4502         /*
4503          * save PMDs to context
4504          * release ownership
4505          */
4506         pfm_flush_pmds(task, ctx);
4507 
4508         /*
4509          * at this point we are done with the PMU
4510          * so we can unreserve the resource.
4511          *
4512          * when state was ZOMBIE, we have already unreserved.
4513          */
4514         if (prev_state != PFM_CTX_ZOMBIE) 
4515                 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4516 
4517         /*
4518          * reset activation counter and psr
4519          */
4520         ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4521         SET_LAST_CPU(ctx, -1);
4522 
4523         /*
4524          * PMU state will not be restored
4525          */
4526         task->thread.flags &= ~IA64_THREAD_PM_VALID;
4527 
4528         /*
4529          * break links between context and task
4530          */
4531         task->thread.pfm_context  = NULL;
4532         ctx->ctx_task             = NULL;
4533 
4534         PFM_SET_WORK_PENDING(task, 0);
4535 
4536         ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4537         ctx->ctx_fl_can_restart  = 0;
4538         ctx->ctx_fl_going_zombie = 0;
4539 
4540         DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4541 
4542         return 0;
4543 }
4544 
4545 
4546 /*
4547  * called only from exit_thread()
4548  * we come here only if the task has a context attached (loaded or masked)
4549  */
4550 void
4551 pfm_exit_thread(struct task_struct *task)
4552 {
4553         pfm_context_t *ctx;
4554         unsigned long flags;
4555         struct pt_regs *regs = task_pt_regs(task);
4556         int ret, state;
4557         int free_ok = 0;
4558 
4559         ctx = PFM_GET_CTX(task);
4560 
4561         PROTECT_CTX(ctx, flags);
4562 
4563         DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4564 
4565         state = ctx->ctx_state;
4566         switch(state) {
4567                 case PFM_CTX_UNLOADED:
4568                         /*
4569                          * only comes to this function if pfm_context is not NULL, i.e., cannot
4570                          * be in unloaded state
4571                          */
4572                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4573                         break;
4574                 case PFM_CTX_LOADED:
4575                 case PFM_CTX_MASKED:
4576                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4577                         if (ret) {
4578                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4579                         }
4580                         DPRINT(("ctx unloaded for current state was %d\n", state));
4581 
4582                         pfm_end_notify_user(ctx);
4583                         break;
4584                 case PFM_CTX_ZOMBIE:
4585                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4586                         if (ret) {
4587                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4588                         }
4589                         free_ok = 1;
4590                         break;
4591                 default:
4592                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4593                         break;
4594         }
4595         UNPROTECT_CTX(ctx, flags);
4596 
4597         { u64 psr = pfm_get_psr();
4598           BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4599           BUG_ON(GET_PMU_OWNER());
4600           BUG_ON(ia64_psr(regs)->up);
4601           BUG_ON(ia64_psr(regs)->pp);
4602         }
4603 
4604         /*
4605          * All memory free operations (especially for vmalloc'ed memory)
4606          * MUST be done with interrupts ENABLED.
4607          */
4608         if (free_ok) pfm_context_free(ctx);
4609 }
4610 
4611 /*
4612  * functions MUST be listed in the increasing order of their index (see permfon.h)
4613  */
4614 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4615 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4616 #define PFM_CMD_PCLRWS  (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4617 #define PFM_CMD_PCLRW   (PFM_CMD_FD|PFM_CMD_ARG_RW)
4618 #define PFM_CMD_NONE    { NULL, "no-cmd", 0, 0, 0, NULL}
4619 
4620 static pfm_cmd_desc_t pfm_cmd_tab[]={
4621 /* 0  */PFM_CMD_NONE,
4622 /* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4623 /* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4624 /* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4625 /* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4626 /* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4627 /* 6  */PFM_CMD_NONE,
4628 /* 7  */PFM_CMD_NONE,
4629 /* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4630 /* 9  */PFM_CMD_NONE,
4631 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4632 /* 11 */PFM_CMD_NONE,
4633 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4634 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4635 /* 14 */PFM_CMD_NONE,
4636 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4637 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4638 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4639 /* 18 */PFM_CMD_NONE,
4640 /* 19 */PFM_CMD_NONE,
4641 /* 20 */PFM_CMD_NONE,
4642 /* 21 */PFM_CMD_NONE,
4643 /* 22 */PFM_CMD_NONE,
4644 /* 23 */PFM_CMD_NONE,
4645 /* 24 */PFM_CMD_NONE,
4646 /* 25 */PFM_CMD_NONE,
4647 /* 26 */PFM_CMD_NONE,
4648 /* 27 */PFM_CMD_NONE,
4649 /* 28 */PFM_CMD_NONE,
4650 /* 29 */PFM_CMD_NONE,
4651 /* 30 */PFM_CMD_NONE,
4652 /* 31 */PFM_CMD_NONE,
4653 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4654 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4655 };
4656 #define PFM_CMD_COUNT   (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4657 
4658 static int
4659 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4660 {
4661         struct task_struct *task;
4662         int state, old_state;
4663 
4664 recheck:
4665         state = ctx->ctx_state;
4666         task  = ctx->ctx_task;
4667 
4668         if (task == NULL) {
4669                 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4670                 return 0;
4671         }
4672 
4673         DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4674                 ctx->ctx_fd,
4675                 state,
4676                 task_pid_nr(task),
4677                 task->state, PFM_CMD_STOPPED(cmd)));
4678 
4679         /*
4680          * self-monitoring always ok.
4681          *
4682          * for system-wide the caller can either be the creator of the
4683          * context (to one to which the context is attached to) OR
4684          * a task running on the same CPU as the session.
4685          */
4686         if (task == current || ctx->ctx_fl_system) return 0;
4687 
4688         /*
4689          * we are monitoring another thread
4690          */
4691         switch(state) {
4692                 case PFM_CTX_UNLOADED:
4693                         /*
4694                          * if context is UNLOADED we are safe to go
4695                          */
4696                         return 0;
4697                 case PFM_CTX_ZOMBIE:
4698                         /*
4699                          * no command can operate on a zombie context
4700                          */
4701                         DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4702                         return -EINVAL;
4703                 case PFM_CTX_MASKED:
4704                         /*
4705                          * PMU state has been saved to software even though
4706                          * the thread may still be running.
4707                          */
4708                         if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4709         }
4710 
4711         /*
4712          * context is LOADED or MASKED. Some commands may need to have 
4713          * the task stopped.
4714          *
4715          * We could lift this restriction for UP but it would mean that
4716          * the user has no guarantee the task would not run between
4717          * two successive calls to perfmonctl(). That's probably OK.
4718          * If this user wants to ensure the task does not run, then
4719          * the task must be stopped.
4720          */
4721         if (PFM_CMD_STOPPED(cmd)) {
4722                 if (!task_is_stopped_or_traced(task)) {
4723                         DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4724                         return -EBUSY;
4725                 }
4726                 /*
4727                  * task is now stopped, wait for ctxsw out
4728                  *
4729                  * This is an interesting point in the code.
4730                  * We need to unprotect the context because
4731                  * the pfm_save_regs() routines needs to grab
4732                  * the same lock. There are danger in doing
4733                  * this because it leaves a window open for
4734                  * another task to get access to the context
4735                  * and possibly change its state. The one thing
4736                  * that is not possible is for the context to disappear
4737                  * because we are protected by the VFS layer, i.e.,
4738                  * get_fd()/put_fd().
4739                  */
4740                 old_state = state;
4741 
4742                 UNPROTECT_CTX(ctx, flags);
4743 
4744                 wait_task_inactive(task, 0);
4745 
4746                 PROTECT_CTX(ctx, flags);
4747 
4748                 /*
4749                  * we must recheck to verify if state has changed
4750                  */
4751                 if (ctx->ctx_state != old_state) {
4752                         DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4753                         goto recheck;
4754                 }
4755         }
4756         return 0;
4757 }
4758 
4759 /*
4760  * system-call entry point (must return long)
4761  */
4762 asmlinkage long
4763 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4764 {
4765         struct fd f = {NULL, 0};
4766         pfm_context_t *ctx = NULL;
4767         unsigned long flags = 0UL;
4768         void *args_k = NULL;
4769         long ret; /* will expand int return types */
4770         size_t base_sz, sz, xtra_sz = 0;
4771         int narg, completed_args = 0, call_made = 0, cmd_flags;
4772         int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4773         int (*getsize)(void *arg, size_t *sz);
4774 #define PFM_MAX_ARGSIZE 4096
4775 
4776         /*
4777          * reject any call if perfmon was disabled at initialization
4778          */
4779         if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4780 
4781         if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4782                 DPRINT(("invalid cmd=%d\n", cmd));
4783                 return -EINVAL;
4784         }
4785 
4786         func      = pfm_cmd_tab[cmd].cmd_func;
4787         narg      = pfm_cmd_tab[cmd].cmd_narg;
4788         base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4789         getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4790         cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4791 
4792         if (unlikely(func == NULL)) {
4793                 DPRINT(("invalid cmd=%d\n", cmd));
4794                 return -EINVAL;
4795         }
4796 
4797         DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4798                 PFM_CMD_NAME(cmd),
4799                 cmd,
4800                 narg,
4801                 base_sz,
4802                 count));
4803 
4804         /*
4805          * check if number of arguments matches what the command expects
4806          */
4807         if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4808                 return -EINVAL;
4809 
4810 restart_args:
4811         sz = xtra_sz + base_sz*count;
4812         /*
4813          * limit abuse to min page size
4814          */
4815         if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4816                 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4817                 return -E2BIG;
4818         }
4819 
4820         /*
4821          * allocate default-sized argument buffer
4822          */
4823         if (likely(count && args_k == NULL)) {
4824                 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4825                 if (args_k == NULL) return -ENOMEM;
4826         }
4827 
4828         ret = -EFAULT;
4829 
4830         /*
4831          * copy arguments
4832          *
4833          * assume sz = 0 for command without parameters
4834          */
4835         if (sz && copy_from_user(args_k, arg, sz)) {
4836                 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4837                 goto error_args;
4838         }
4839 
4840         /*
4841          * check if command supports extra parameters
4842          */
4843         if (completed_args == 0 && getsize) {
4844                 /*
4845                  * get extra parameters size (based on main argument)
4846                  */
4847                 ret = (*getsize)(args_k, &xtra_sz);
4848                 if (ret) goto error_args;
4849 
4850                 completed_args = 1;
4851 
4852                 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4853 
4854                 /* retry if necessary */
4855                 if (likely(xtra_sz)) goto restart_args;
4856         }
4857 
4858         if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4859 
4860         ret = -EBADF;
4861 
4862         f = fdget(fd);
4863         if (unlikely(f.file == NULL)) {
4864                 DPRINT(("invalid fd %d\n", fd));
4865                 goto error_args;
4866         }
4867         if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4868                 DPRINT(("fd %d not related to perfmon\n", fd));
4869                 goto error_args;
4870         }
4871 
4872         ctx = f.file->private_data;
4873         if (unlikely(ctx == NULL)) {
4874                 DPRINT(("no context for fd %d\n", fd));
4875                 goto error_args;
4876         }
4877         prefetch(&ctx->ctx_state);
4878 
4879         PROTECT_CTX(ctx, flags);
4880 
4881         /*
4882          * check task is stopped
4883          */
4884         ret = pfm_check_task_state(ctx, cmd, flags);
4885         if (unlikely(ret)) goto abort_locked;
4886 
4887 skip_fd:
4888         ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4889 
4890         call_made = 1;
4891 
4892 abort_locked:
4893         if (likely(ctx)) {
4894                 DPRINT(("context unlocked\n"));
4895                 UNPROTECT_CTX(ctx, flags);
4896         }
4897 
4898         /* copy argument back to user, if needed */
4899         if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4900 
4901 error_args:
4902         if (f.file)
4903                 fdput(f);
4904 
4905         kfree(args_k);
4906 
4907         DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4908 
4909         return ret;
4910 }
4911 
4912 static void
4913 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4914 {
4915         pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4916         pfm_ovfl_ctrl_t rst_ctrl;
4917         int state;
4918         int ret = 0;
4919 
4920         state = ctx->ctx_state;
4921         /*
4922          * Unlock sampling buffer and reset index atomically
4923          * XXX: not really needed when blocking
4924          */
4925         if (CTX_HAS_SMPL(ctx)) {
4926 
4927                 rst_ctrl.bits.mask_monitoring = 0;
4928                 rst_ctrl.bits.reset_ovfl_pmds = 0;
4929 
4930                 if (state == PFM_CTX_LOADED)
4931                         ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4932                 else
4933                         ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4934         } else {
4935                 rst_ctrl.bits.mask_monitoring = 0;
4936                 rst_ctrl.bits.reset_ovfl_pmds = 1;
4937         }
4938 
4939         if (ret == 0) {
4940                 if (rst_ctrl.bits.reset_ovfl_pmds) {
4941                         pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4942                 }
4943                 if (rst_ctrl.bits.mask_monitoring == 0) {
4944                         DPRINT(("resuming monitoring\n"));
4945                         if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4946                 } else {
4947                         DPRINT(("stopping monitoring\n"));
4948                         //pfm_stop_monitoring(current, regs);
4949                 }
4950                 ctx->ctx_state = PFM_CTX_LOADED;
4951         }
4952 }
4953 
4954 /*
4955  * context MUST BE LOCKED when calling
4956  * can only be called for current
4957  */
4958 static void
4959 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4960 {
4961         int ret;
4962 
4963         DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4964 
4965         ret = pfm_context_unload(ctx, NULL, 0, regs);
4966         if (ret) {
4967                 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4968         }
4969 
4970         /*
4971          * and wakeup controlling task, indicating we are now disconnected
4972          */
4973         wake_up_interruptible(&ctx->ctx_zombieq);
4974 
4975         /*
4976          * given that context is still locked, the controlling
4977          * task will only get access when we return from
4978          * pfm_handle_work().
4979          */
4980 }
4981 
4982 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4983 
4984  /*
4985   * pfm_handle_work() can be called with interrupts enabled
4986   * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4987   * call may sleep, therefore we must re-enable interrupts
4988   * to avoid deadlocks. It is safe to do so because this function
4989   * is called ONLY when returning to user level (pUStk=1), in which case
4990   * there is no risk of kernel stack overflow due to deep
4991   * interrupt nesting.
4992   */
4993 void
4994 pfm_handle_work(void)
4995 {
4996         pfm_context_t *ctx;
4997         struct pt_regs *regs;
4998         unsigned long flags, dummy_flags;
4999         unsigned long ovfl_regs;
5000         unsigned int reason;
5001         int ret;
5002 
5003         ctx = PFM_GET_CTX(current);
5004         if (ctx == NULL) {
5005                 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5006                         task_pid_nr(current));
5007                 return;
5008         }
5009 
5010         PROTECT_CTX(ctx, flags);
5011 
5012         PFM_SET_WORK_PENDING(current, 0);
5013 
5014         regs = task_pt_regs(current);
5015 
5016         /*
5017          * extract reason for being here and clear
5018          */
5019         reason = ctx->ctx_fl_trap_reason;
5020         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5021         ovfl_regs = ctx->ctx_ovfl_regs[0];
5022 
5023         DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5024 
5025         /*
5026          * must be done before we check for simple-reset mode
5027          */
5028         if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5029                 goto do_zombie;
5030 
5031         //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5032         if (reason == PFM_TRAP_REASON_RESET)
5033                 goto skip_blocking;
5034 
5035         /*
5036          * restore interrupt mask to what it was on entry.
5037          * Could be enabled/diasbled.
5038          */
5039         UNPROTECT_CTX(ctx, flags);
5040 
5041         /*
5042          * force interrupt enable because of down_interruptible()
5043          */
5044         local_irq_enable();
5045 
5046         DPRINT(("before block sleeping\n"));
5047 
5048         /*
5049          * may go through without blocking on SMP systems
5050          * if restart has been received already by the time we call down()
5051          */
5052         ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5053 
5054         DPRINT(("after block sleeping ret=%d\n", ret));
5055 
5056         /*
5057          * lock context and mask interrupts again
5058          * We save flags into a dummy because we may have
5059          * altered interrupts mask compared to entry in this
5060          * function.
5061          */
5062         PROTECT_CTX(ctx, dummy_flags);
5063 
5064         /*
5065          * we need to read the ovfl_regs only after wake-up
5066          * because we may have had pfm_write_pmds() in between
5067          * and that can changed PMD values and therefore 
5068          * ovfl_regs is reset for these new PMD values.
5069          */
5070         ovfl_regs = ctx->ctx_ovfl_regs[0];
5071 
5072         if (ctx->ctx_fl_going_zombie) {
5073 do_zombie:
5074                 DPRINT(("context is zombie, bailing out\n"));
5075                 pfm_context_force_terminate(ctx, regs);
5076                 goto nothing_to_do;
5077         }
5078         /*
5079          * in case of interruption of down() we don't restart anything
5080          */
5081         if (ret < 0)
5082                 goto nothing_to_do;
5083 
5084 skip_blocking:
5085         pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5086         ctx->ctx_ovfl_regs[0] = 0UL;
5087 
5088 nothing_to_do:
5089         /*
5090          * restore flags as they were upon entry
5091          */
5092         UNPROTECT_CTX(ctx, flags);
5093 }
5094 
5095 static int
5096 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5097 {
5098         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5099                 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5100                 return 0;
5101         }
5102 
5103         DPRINT(("waking up somebody\n"));
5104 
5105         if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5106 
5107         /*
5108          * safe, we are not in intr handler, nor in ctxsw when
5109          * we come here
5110          */
5111         kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5112 
5113         return 0;
5114 }
5115 
5116 static int
5117 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5118 {
5119         pfm_msg_t *msg = NULL;
5120 
5121         if (ctx->ctx_fl_no_msg == 0) {
5122                 msg = pfm_get_new_msg(ctx);
5123                 if (msg == NULL) {
5124                         printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5125                         return -1;
5126                 }
5127 
5128                 msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5129                 msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5130                 msg->pfm_ovfl_msg.msg_active_set   = 0;
5131                 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5132                 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5133                 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5134                 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5135                 msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5136         }
5137 
5138         DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5139                 msg,
5140                 ctx->ctx_fl_no_msg,
5141                 ctx->ctx_fd,
5142                 ovfl_pmds));
5143 
5144         return pfm_notify_user(ctx, msg);
5145 }
5146 
5147 static int
5148 pfm_end_notify_user(pfm_context_t *ctx)
5149 {
5150         pfm_msg_t *msg;
5151 
5152         msg = pfm_get_new_msg(ctx);
5153         if (msg == NULL) {
5154                 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5155                 return -1;
5156         }
5157         /* no leak */
5158         memset(msg, 0, sizeof(*msg));
5159 
5160         msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5161         msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5162         msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5163 
5164         DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5165                 msg,
5166                 ctx->ctx_fl_no_msg,
5167                 ctx->ctx_fd));
5168 
5169         return pfm_notify_user(ctx, msg);
5170 }
5171 
5172 /*
5173  * main overflow processing routine.
5174  * it can be called from the interrupt path or explicitly during the context switch code
5175  */
5176 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5177                                 unsigned long pmc0, struct pt_regs *regs)
5178 {
5179         pfm_ovfl_arg_t *ovfl_arg;
5180         unsigned long mask;
5181         unsigned long old_val, ovfl_val, new_val;
5182         unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5183         unsigned long tstamp;
5184         pfm_ovfl_ctrl_t ovfl_ctrl;
5185         unsigned int i, has_smpl;
5186         int must_notify = 0;
5187 
5188         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5189 
5190         /*
5191          * sanity test. Should never happen
5192          */
5193         if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5194 
5195         tstamp   = ia64_get_itc();
5196         mask     = pmc0 >> PMU_FIRST_COUNTER;
5197         ovfl_val = pmu_conf->ovfl_val;
5198         has_smpl = CTX_HAS_SMPL(ctx);
5199 
5200         DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5201                      "used_pmds=0x%lx\n",
5202                         pmc0,
5203                         task ? task_pid_nr(task): -1,
5204                         (regs ? regs->cr_iip : 0),
5205                         CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5206                         ctx->ctx_used_pmds[0]));
5207 
5208 
5209         /*
5210          * first we update the virtual counters
5211          * assume there was a prior ia64_srlz_d() issued
5212          */
5213         for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5214 
5215                 /* skip pmd which did not overflow */
5216                 if ((mask & 0x1) == 0) continue;
5217 
5218                 /*
5219                  * Note that the pmd is not necessarily 0 at this point as qualified events
5220                  * may have happened before the PMU was frozen. The residual count is not
5221                  * taken into consideration here but will be with any read of the pmd via
5222                  * pfm_read_pmds().
5223                  */
5224                 old_val              = new_val = ctx->ctx_pmds[i].val;
5225                 new_val             += 1 + ovfl_val;
5226                 ctx->ctx_pmds[i].val = new_val;
5227 
5228                 /*
5229                  * check for overflow condition
5230                  */
5231                 if (likely(old_val > new_val)) {
5232                         ovfl_pmds |= 1UL << i;
5233                         if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5234                 }
5235 
5236                 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5237                         i,
5238                         new_val,
5239                         old_val,
5240                         ia64_get_pmd(i) & ovfl_val,
5241                         ovfl_pmds,
5242                         ovfl_notify));
5243         }
5244 
5245         /*
5246          * there was no 64-bit overflow, nothing else to do
5247          */
5248         if (ovfl_pmds == 0UL) return;
5249 
5250         /* 
5251          * reset all control bits
5252          */
5253         ovfl_ctrl.val = 0;
5254         reset_pmds    = 0UL;
5255 
5256         /*
5257          * if a sampling format module exists, then we "cache" the overflow by 
5258          * calling the module's handler() routine.
5259          */
5260         if (has_smpl) {
5261                 unsigned long start_cycles, end_cycles;
5262                 unsigned long pmd_mask;
5263                 int j, k, ret = 0;
5264                 int this_cpu = smp_processor_id();
5265 
5266                 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5267                 ovfl_arg = &ctx->ctx_ovfl_arg;
5268 
5269                 prefetch(ctx->ctx_smpl_hdr);
5270 
5271                 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5272 
5273                         mask = 1UL << i;
5274 
5275                         if ((pmd_mask & 0x1) == 0) continue;
5276 
5277                         ovfl_arg->ovfl_pmd      = (unsigned char )i;
5278                         ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5279                         ovfl_arg->active_set    = 0;
5280                         ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5281                         ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5282 
5283                         ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5284                         ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5285                         ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5286 
5287                         /*
5288                          * copy values of pmds of interest. Sampling format may copy them
5289                          * into sampling buffer.
5290                          */
5291                         if (smpl_pmds) {
5292                                 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5293                                         if ((smpl_pmds & 0x1) == 0) continue;
5294                                         ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5295                                         DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5296                                 }
5297                         }
5298 
5299                         pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5300 
5301                         start_cycles = ia64_get_itc();
5302 
5303                         /*
5304                          * call custom buffer format record (handler) routine
5305                          */
5306                         ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5307 
5308                         end_cycles = ia64_get_itc();
5309 
5310                         /*
5311                          * For those controls, we take the union because they have
5312                          * an all or nothing behavior.
5313                          */
5314                         ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5315                         ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5316                         ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5317                         /*
5318                          * build the bitmask of pmds to reset now
5319                          */
5320                         if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5321 
5322                         pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5323                 }
5324                 /*
5325                  * when the module cannot handle the rest of the overflows, we abort right here
5326                  */
5327                 if (ret && pmd_mask) {
5328                         DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5329                                 pmd_mask<<PMU_FIRST_COUNTER));
5330                 }
5331                 /*
5332                  * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5333                  */
5334                 ovfl_pmds &= ~reset_pmds;
5335         } else {
5336                 /*
5337                  * when no sampling module is used, then the default
5338                  * is to notify on overflow if requested by user
5339                  */
5340                 ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5341                 ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5342                 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5343                 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5344                 /*
5345                  * if needed, we reset all overflowed pmds
5346                  */
5347                 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5348         }
5349 
5350         DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5351 
5352         /*
5353          * reset the requested PMD registers using the short reset values
5354          */
5355         if (reset_pmds) {
5356                 unsigned long bm = reset_pmds;
5357                 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5358         }
5359 
5360         if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5361                 /*
5362                  * keep track of what to reset when unblocking
5363                  */
5364                 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5365 
5366                 /*
5367                  * check for blocking context 
5368                  */
5369                 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5370 
5371                         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5372 
5373                         /*
5374                          * set the perfmon specific checking pending work for the task
5375                          */
5376                         PFM_SET_WORK_PENDING(task, 1);
5377 
5378                         /*
5379                          * when coming from ctxsw, current still points to the
5380                          * previous task, therefore we must work with task and not current.
5381                          */
5382                         set_notify_resume(task);
5383                 }
5384                 /*
5385                  * defer until state is changed (shorten spin window). the context is locked
5386                  * anyway, so the signal receiver would come spin for nothing.
5387                  */
5388                 must_notify = 1;
5389         }
5390 
5391         DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5392                         GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5393                         PFM_GET_WORK_PENDING(task),
5394                         ctx->ctx_fl_trap_reason,
5395                         ovfl_pmds,
5396                         ovfl_notify,
5397                         ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5398         /*
5399          * in case monitoring must be stopped, we toggle the psr bits
5400          */
5401         if (ovfl_ctrl.bits.mask_monitoring) {
5402                 pfm_mask_monitoring(task);
5403                 ctx->ctx_state = PFM_CTX_MASKED;
5404                 ctx->ctx_fl_can_restart = 1;
5405         }
5406 
5407         /*
5408          * send notification now
5409          */
5410         if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5411 
5412         return;
5413 
5414 sanity_check:
5415         printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5416                         smp_processor_id(),
5417                         task ? task_pid_nr(task) : -1,
5418                         pmc0);
5419         return;
5420 
5421 stop_monitoring:
5422         /*
5423          * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5424          * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5425          * come here as zombie only if the task is the current task. In which case, we
5426          * can access the PMU  hardware directly.
5427          *
5428          * Note that zombies do have PM_VALID set. So here we do the minimal.
5429          *
5430          * In case the context was zombified it could not be reclaimed at the time
5431          * the monitoring program exited. At this point, the PMU reservation has been
5432          * returned, the sampiing buffer has been freed. We must convert this call
5433          * into a spurious interrupt. However, we must also avoid infinite overflows
5434          * by stopping monitoring for this task. We can only come here for a per-task
5435          * context. All we need to do is to stop monitoring using the psr bits which
5436          * are always task private. By re-enabling secure montioring, we ensure that
5437          * the monitored task will not be able to re-activate monitoring.
5438          * The task will eventually be context switched out, at which point the context
5439          * will be reclaimed (that includes releasing ownership of the PMU).
5440          *
5441          * So there might be a window of time where the number of per-task session is zero
5442          * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5443          * context. This is safe because if a per-task session comes in, it will push this one
5444          * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5445          * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5446          * also push our zombie context out.
5447          *
5448          * Overall pretty hairy stuff....
5449          */
5450         DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5451         pfm_clear_psr_up();
5452         ia64_psr(regs)->up = 0;
5453         ia64_psr(regs)->sp = 1;
5454         return;
5455 }
5456 
5457 static int
5458 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5459 {
5460         struct task_struct *task;
5461         pfm_context_t *ctx;
5462         unsigned long flags;
5463         u64 pmc0;
5464         int this_cpu = smp_processor_id();
5465         int retval = 0;
5466 
5467         pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5468 
5469         /*
5470          * srlz.d done before arriving here
5471          */
5472         pmc0 = ia64_get_pmc(0);
5473 
5474         task = GET_PMU_OWNER();
5475         ctx  = GET_PMU_CTX();
5476 
5477         /*
5478          * if we have some pending bits set
5479          * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5480          */
5481         if (PMC0_HAS_OVFL(pmc0) && task) {
5482                 /*
5483                  * we assume that pmc0.fr is always set here
5484                  */
5485 
5486                 /* sanity check */
5487                 if (!ctx) goto report_spurious1;
5488 
5489                 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5490                         goto report_spurious2;
5491 
5492                 PROTECT_CTX_NOPRINT(ctx, flags);
5493 
5494                 pfm_overflow_handler(task, ctx, pmc0, regs);
5495 
5496                 UNPROTECT_CTX_NOPRINT(ctx, flags);
5497 
5498         } else {
5499                 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5500                 retval = -1;
5501         }
5502         /*
5503          * keep it unfrozen at all times
5504          */
5505         pfm_unfreeze_pmu();
5506 
5507         return retval;
5508 
5509 report_spurious1:
5510         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5511                 this_cpu, task_pid_nr(task));
5512         pfm_unfreeze_pmu();
5513         return -1;
5514 report_spurious2:
5515         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5516                 this_cpu, 
5517                 task_pid_nr(task));
5518         pfm_unfreeze_pmu();
5519         return -1;
5520 }
5521 
5522 static irqreturn_t
5523 pfm_interrupt_handler(int irq, void *arg)
5524 {
5525         unsigned long start_cycles, total_cycles;
5526         unsigned long min, max;
5527         int this_cpu;
5528         int ret;
5529         struct pt_regs *regs = get_irq_regs();
5530 
5531         this_cpu = get_cpu();
5532         if (likely(!pfm_alt_intr_handler)) {
5533                 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5534                 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5535 
5536                 start_cycles = ia64_get_itc();
5537 
5538                 ret = pfm_do_interrupt_handler(arg, regs);
5539 
5540                 total_cycles = ia64_get_itc();
5541 
5542                 /*
5543                  * don't measure spurious interrupts
5544                  */
5545                 if (likely(ret == 0)) {
5546                         total_cycles -= start_cycles;
5547 
5548                         if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5549                         if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5550 
5551                         pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5552                 }
5553         }
5554         else {
5555                 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5556         }
5557 
5558         put_cpu();
5559         return IRQ_HANDLED;
5560 }
5561 
5562 /*
5563  * /proc/perfmon interface, for debug only
5564  */
5565 
5566 #define PFM_PROC_SHOW_HEADER    ((void *)(long)nr_cpu_ids+1)
5567 
5568 static void *
5569 pfm_proc_start(struct seq_file *m, loff_t *pos)
5570 {
5571         if (*pos == 0) {
5572                 return PFM_PROC_SHOW_HEADER;
5573         }
5574 
5575         while (*pos <= nr_cpu_ids) {
5576                 if (cpu_online(*pos - 1)) {
5577                         return (void *)*pos;
5578                 }
5579                 ++*pos;
5580         }
5581         return NULL;
5582 }
5583 
5584 static void *
5585 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5586 {
5587         ++*pos;
5588         return pfm_proc_start(m, pos);
5589 }
5590 
5591 static void
5592 pfm_proc_stop(struct seq_file *m, void *v)
5593 {
5594 }
5595 
5596 static void
5597 pfm_proc_show_header(struct seq_file *m)
5598 {
5599         struct list_head * pos;
5600         pfm_buffer_fmt_t * entry;
5601         unsigned long flags;
5602 
5603         seq_printf(m,
5604                 "perfmon version           : %u.%u\n"
5605                 "model                     : %s\n"
5606                 "fastctxsw                 : %s\n"
5607                 "expert mode               : %s\n"
5608                 "ovfl_mask                 : 0x%lx\n"
5609                 "PMU flags                 : 0x%x\n",
5610                 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5611                 pmu_conf->pmu_name,
5612                 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5613                 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5614                 pmu_conf->ovfl_val,
5615                 pmu_conf->flags);
5616 
5617         LOCK_PFS(flags);
5618 
5619         seq_printf(m,
5620                 "proc_sessions             : %u\n"
5621                 "sys_sessions              : %u\n"
5622                 "sys_use_dbregs            : %u\n"
5623                 "ptrace_use_dbregs         : %u\n",
5624                 pfm_sessions.pfs_task_sessions,
5625                 pfm_sessions.pfs_sys_sessions,
5626                 pfm_sessions.pfs_sys_use_dbregs,
5627                 pfm_sessions.pfs_ptrace_use_dbregs);
5628 
5629         UNLOCK_PFS(flags);
5630 
5631         spin_lock(&pfm_buffer_fmt_lock);
5632 
5633         list_for_each(pos, &pfm_buffer_fmt_list) {
5634                 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5635                 seq_printf(m, "format                    : %16phD %s\n",
5636                            entry->fmt_uuid, entry->fmt_name);
5637         }
5638         spin_unlock(&pfm_buffer_fmt_lock);
5639 
5640 }
5641 
5642 static int
5643 pfm_proc_show(struct seq_file *m, void *v)
5644 {
5645         unsigned long psr;
5646         unsigned int i;
5647         int cpu;
5648 
5649         if (v == PFM_PROC_SHOW_HEADER) {
5650                 pfm_proc_show_header(m);
5651                 return 0;
5652         }
5653 
5654         /* show info for CPU (v - 1) */
5655 
5656         cpu = (long)v - 1;
5657         seq_printf(m,
5658                 "CPU%-2d overflow intrs      : %lu\n"
5659                 "CPU%-2d overflow cycles     : %lu\n"
5660                 "CPU%-2d overflow min        : %lu\n"
5661                 "CPU%-2d overflow max        : %lu\n"
5662                 "CPU%-2d smpl handler calls  : %lu\n"
5663                 "CPU%-2d smpl handler cycles : %lu\n"
5664                 "CPU%-2d spurious intrs      : %lu\n"
5665                 "CPU%-2d replay   intrs      : %lu\n"
5666                 "CPU%-2d syst_wide           : %d\n"
5667                 "CPU%-2d dcr_pp              : %d\n"
5668                 "CPU%-2d exclude idle        : %d\n"
5669                 "CPU%-2d owner               : %d\n"
5670                 "CPU%-2d context             : %p\n"
5671                 "CPU%-2d activations         : %lu\n",
5672                 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5673                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5674                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5675                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5676                 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5677                 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5678                 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5679                 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5680                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5681                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5682                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5683                 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5684                 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5685                 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5686 
5687         if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5688 
5689                 psr = pfm_get_psr();
5690 
5691                 ia64_srlz_d();
5692 
5693                 seq_printf(m, 
5694                         "CPU%-2d psr                 : 0x%lx\n"
5695                         "CPU%-2d pmc0                : 0x%lx\n", 
5696                         cpu, psr,
5697                         cpu, ia64_get_pmc(0));
5698 
5699                 for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5700                         if (PMC_IS_COUNTING(i) == 0) continue;
5701                         seq_printf(m, 
5702                                 "CPU%-2d pmc%u                : 0x%lx\n"
5703                                 "CPU%-2d pmd%u                : 0x%lx\n", 
5704                                 cpu, i, ia64_get_pmc(i),
5705                                 cpu, i, ia64_get_pmd(i));
5706                 }
5707         }
5708         return 0;
5709 }
5710 
5711 const struct seq_operations pfm_seq_ops = {
5712         .start =        pfm_proc_start,
5713         .next =         pfm_proc_next,
5714         .stop =         pfm_proc_stop,
5715         .show =         pfm_proc_show
5716 };
5717 
5718 static int
5719 pfm_proc_open(struct inode *inode, struct file *file)
5720 {
5721         return seq_open(file, &pfm_seq_ops);
5722 }
5723 
5724 
5725 /*
5726  * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5727  * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5728  * is active or inactive based on mode. We must rely on the value in
5729  * local_cpu_data->pfm_syst_info
5730  */
5731 void
5732 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5733 {
5734         struct pt_regs *regs;
5735         unsigned long dcr;
5736         unsigned long dcr_pp;
5737 
5738         dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5739 
5740         /*
5741          * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5742          * on every CPU, so we can rely on the pid to identify the idle task.
5743          */
5744         if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5745                 regs = task_pt_regs(task);
5746                 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5747                 return;
5748         }
5749         /*
5750          * if monitoring has started
5751          */
5752         if (dcr_pp) {
5753                 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5754                 /*
5755                  * context switching in?
5756                  */
5757                 if (is_ctxswin) {
5758                         /* mask monitoring for the idle task */
5759                         ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5760                         pfm_clear_psr_pp();
5761                         ia64_srlz_i();
5762                         return;
5763                 }
5764                 /*
5765                  * context switching out
5766                  * restore monitoring for next task
5767                  *
5768                  * Due to inlining this odd if-then-else construction generates
5769                  * better code.
5770                  */
5771                 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5772                 pfm_set_psr_pp();
5773                 ia64_srlz_i();
5774         }
5775 }
5776 
5777 #ifdef CONFIG_SMP
5778 
5779 static void
5780 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5781 {
5782         struct task_struct *task = ctx->ctx_task;
5783 
5784         ia64_psr(regs)->up = 0;
5785         ia64_psr(regs)->sp = 1;
5786 
5787         if (GET_PMU_OWNER() == task) {
5788                 DPRINT(("cleared ownership for [%d]\n",
5789                                         task_pid_nr(ctx->ctx_task)));
5790                 SET_PMU_OWNER(NULL, NULL);
5791         }
5792 
5793         /*
5794          * disconnect the task from the context and vice-versa
5795          */
5796         PFM_SET_WORK_PENDING(task, 0);
5797 
5798         task->thread.pfm_context  = NULL;
5799         task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5800 
5801         DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5802 }
5803 
5804 
5805 /*
5806  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5807  */
5808 void
5809 pfm_save_regs(struct task_struct *task)
5810 {
5811         pfm_context_t *ctx;
5812         unsigned long flags;
5813         u64 psr;
5814 
5815 
5816         ctx = PFM_GET_CTX(task);
5817         if (ctx == NULL) return;
5818 
5819         /*
5820          * we always come here with interrupts ALREADY disabled by
5821          * the scheduler. So we simply need to protect against concurrent
5822          * access, not CPU concurrency.
5823          */
5824         flags = pfm_protect_ctx_ctxsw(ctx);
5825 
5826         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5827                 struct pt_regs *regs = task_pt_regs(task);
5828 
5829                 pfm_clear_psr_up();
5830 
5831                 pfm_force_cleanup(ctx, regs);
5832 
5833                 BUG_ON(ctx->ctx_smpl_hdr);
5834 
5835                 pfm_unprotect_ctx_ctxsw(ctx, flags);
5836 
5837                 pfm_context_free(ctx);
5838                 return;
5839         }
5840 
5841         /*
5842          * save current PSR: needed because we modify it
5843          */
5844         ia64_srlz_d();
5845         psr = pfm_get_psr();
5846 
5847         BUG_ON(psr & (IA64_PSR_I));
5848 
5849         /*
5850          * stop monitoring:
5851          * This is the last instruction which may generate an overflow
5852          *
5853          * We do not need to set psr.sp because, it is irrelevant in kernel.
5854          * It will be restored from ipsr when going back to user level
5855          */
5856         pfm_clear_psr_up();
5857 
5858         /*
5859          * keep a copy of psr.up (for reload)
5860          */
5861         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5862 
5863         /*
5864          * release ownership of this PMU.
5865          * PM interrupts are masked, so nothing
5866          * can happen.
5867          */
5868         SET_PMU_OWNER(NULL, NULL);
5869 
5870         /*
5871          * we systematically save the PMD as we have no
5872          * guarantee we will be schedule at that same
5873          * CPU again.
5874          */
5875         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5876 
5877         /*
5878          * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5879          * we will need it on the restore path to check
5880          * for pending overflow.
5881          */
5882         ctx->th_pmcs[0] = ia64_get_pmc(0);
5883 
5884         /*
5885          * unfreeze PMU if had pending overflows
5886          */
5887         if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5888 
5889         /*
5890          * finally, allow context access.
5891          * interrupts will still be masked after this call.
5892          */
5893         pfm_unprotect_ctx_ctxsw(ctx, flags);
5894 }
5895 
5896 #else /* !CONFIG_SMP */
5897 void
5898 pfm_save_regs(struct task_struct *task)
5899 {
5900         pfm_context_t *ctx;
5901         u64 psr;
5902 
5903         ctx = PFM_GET_CTX(task);
5904         if (ctx == NULL) return;
5905 
5906         /*
5907          * save current PSR: needed because we modify it
5908          */
5909         psr = pfm_get_psr();
5910 
5911         BUG_ON(psr & (IA64_PSR_I));
5912 
5913         /*
5914          * stop monitoring:
5915          * This is the last instruction which may generate an overflow
5916          *
5917          * We do not need to set psr.sp because, it is irrelevant in kernel.
5918          * It will be restored from ipsr when going back to user level
5919          */
5920         pfm_clear_psr_up();
5921 
5922         /*
5923          * keep a copy of psr.up (for reload)
5924          */
5925         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5926 }
5927 
5928 static void
5929 pfm_lazy_save_regs (struct task_struct *task)
5930 {
5931         pfm_context_t *ctx;
5932         unsigned long flags;
5933 
5934         { u64 psr  = pfm_get_psr();
5935           BUG_ON(psr & IA64_PSR_UP);
5936         }
5937 
5938         ctx = PFM_GET_CTX(task);
5939 
5940         /*
5941          * we need to mask PMU overflow here to
5942          * make sure that we maintain pmc0 until
5943          * we save it. overflow interrupts are
5944          * treated as spurious if there is no
5945          * owner.
5946          *
5947          * XXX: I don't think this is necessary
5948          */
5949         PROTECT_CTX(ctx,flags);
5950 
5951         /*
5952          * release ownership of this PMU.
5953          * must be done before we save the registers.
5954          *
5955          * after this call any PMU interrupt is treated
5956          * as spurious.
5957          */
5958         SET_PMU_OWNER(NULL, NULL);
5959 
5960         /*
5961          * save all the pmds we use
5962          */
5963         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5964 
5965         /*
5966          * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5967          * it is needed to check for pended overflow
5968          * on the restore path
5969          */
5970         ctx->th_pmcs[0] = ia64_get_pmc(0);
5971 
5972         /*
5973          * unfreeze PMU if had pending overflows
5974          */
5975         if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5976 
5977         /*
5978          * now get can unmask PMU interrupts, they will
5979          * be treated as purely spurious and we will not
5980          * lose any information
5981          */
5982         UNPROTECT_CTX(ctx,flags);
5983 }
5984 #endif /* CONFIG_SMP */
5985 
5986 #ifdef CONFIG_SMP
5987 /*
5988  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5989  */
5990 void
5991 pfm_load_regs (struct task_struct *task)
5992 {
5993         pfm_context_t *ctx;
5994         unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
5995         unsigned long flags;
5996         u64 psr, psr_up;
5997         int need_irq_resend;
5998 
5999         ctx = PFM_GET_CTX(task);
6000         if (unlikely(ctx == NULL)) return;
6001 
6002         BUG_ON(GET_PMU_OWNER());
6003 
6004         /*
6005          * possible on unload
6006          */
6007         if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6008 
6009         /*
6010          * we always come here with interrupts ALREADY disabled by
6011          * the scheduler. So we simply need to protect against concurrent
6012          * access, not CPU concurrency.
6013          */
6014         flags = pfm_protect_ctx_ctxsw(ctx);
6015         psr   = pfm_get_psr();
6016 
6017         need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6018 
6019         BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6020         BUG_ON(psr & IA64_PSR_I);
6021 
6022         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6023                 struct pt_regs *regs = task_pt_regs(task);
6024 
6025                 BUG_ON(ctx->ctx_smpl_hdr);
6026 
6027                 pfm_force_cleanup(ctx, regs);
6028 
6029                 pfm_unprotect_ctx_ctxsw(ctx, flags);
6030 
6031                 /*
6032                  * this one (kmalloc'ed) is fine with interrupts disabled
6033                  */
6034                 pfm_context_free(ctx);
6035 
6036                 return;
6037         }
6038 
6039         /*
6040          * we restore ALL the debug registers to avoid picking up
6041          * stale state.
6042          */
6043         if (ctx->ctx_fl_using_dbreg) {
6044                 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6045                 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6046         }
6047         /*
6048          * retrieve saved psr.up
6049          */
6050         psr_up = ctx->ctx_saved_psr_up;
6051 
6052         /*
6053          * if we were the last user of the PMU on that CPU,
6054          * then nothing to do except restore psr
6055          */
6056         if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6057 
6058                 /*
6059                  * retrieve partial reload masks (due to user modifications)
6060                  */
6061                 pmc_mask = ctx->ctx_reload_pmcs[0];
6062                 pmd_mask = ctx->ctx_reload_pmds[0];
6063 
6064         } else {
6065                 /*
6066                  * To avoid leaking information to the user level when psr.sp=0,
6067                  * we must reload ALL implemented pmds (even the ones we don't use).
6068                  * In the kernel we only allow PFM_READ_PMDS on registers which
6069                  * we initialized or requested (sampling) so there is no risk there.
6070                  */
6071                 pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6072 
6073                 /*
6074                  * ALL accessible PMCs are systematically reloaded, unused registers
6075                  * get their default (from pfm_reset_pmu_state()) values to avoid picking
6076                  * up stale configuration.
6077                  *
6078                  * PMC0 is never in the mask. It is always restored separately.
6079                  */
6080                 pmc_mask = ctx->ctx_all_pmcs[0];
6081         }
6082         /*
6083          * when context is MASKED, we will restore PMC with plm=0
6084          * and PMD with stale information, but that's ok, nothing
6085          * will be captured.
6086          *
6087          * XXX: optimize here
6088          */
6089         if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6090         if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6091 
6092         /*
6093          * check for pending overflow at the time the state
6094          * was saved.
6095          */
6096         if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6097                 /*
6098                  * reload pmc0 with the overflow information
6099                  * On McKinley PMU, this will trigger a PMU interrupt
6100                  */
6101                 ia64_set_pmc(0, ctx->th_pmcs[0]);
6102                 ia64_srlz_d();
6103                 ctx->th_pmcs[0] = 0UL;
6104 
6105                 /*
6106                  * will replay the PMU interrupt
6107                  */
6108                 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6109 
6110                 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6111         }
6112 
6113         /*
6114          * we just did a reload, so we reset the partial reload fields
6115          */
6116         ctx->ctx_reload_pmcs[0] = 0UL;
6117         ctx->ctx_reload_pmds[0] = 0UL;
6118 
6119         SET_LAST_CPU(ctx, smp_processor_id());
6120 
6121         /*
6122          * dump activation value for this PMU
6123          */
6124         INC_ACTIVATION();
6125         /*
6126          * record current activation for this context
6127          */
6128         SET_ACTIVATION(ctx);
6129 
6130         /*
6131          * establish new ownership. 
6132          */
6133         SET_PMU_OWNER(task, ctx);
6134 
6135         /*
6136          * restore the psr.up bit. measurement
6137          * is active again.
6138          * no PMU interrupt can happen at this point
6139          * because we still have interrupts disabled.
6140          */
6141         if (likely(psr_up)) pfm_set_psr_up();
6142 
6143         /*
6144          * allow concurrent access to context
6145          */
6146         pfm_unprotect_ctx_ctxsw(ctx, flags);
6147 }
6148 #else /*  !CONFIG_SMP */
6149 /*
6150  * reload PMU state for UP kernels
6151  * in 2.5 we come here with interrupts disabled
6152  */
6153 void
6154 pfm_load_regs (struct task_struct *task)
6155 {
6156         pfm_context_t *ctx;
6157         struct task_struct *owner;
6158         unsigned long pmd_mask, pmc_mask;
6159         u64 psr, psr_up;
6160         int need_irq_resend;
6161 
6162         owner = GET_PMU_OWNER();
6163         ctx   = PFM_GET_CTX(task);
6164         psr   = pfm_get_psr();
6165 
6166         BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6167         BUG_ON(psr & IA64_PSR_I);
6168 
6169         /*
6170          * we restore ALL the debug registers to avoid picking up
6171          * stale state.
6172          *
6173          * This must be done even when the task is still the owner
6174          * as the registers may have been modified via ptrace()
6175          * (not perfmon) by the previous task.
6176          */
6177         if (ctx->ctx_fl_using_dbreg) {
6178                 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6179                 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6180         }
6181 
6182         /*
6183          * retrieved saved psr.up
6184          */
6185         psr_up = ctx->ctx_saved_psr_up;
6186         need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6187 
6188         /*
6189          * short path, our state is still there, just
6190          * need to restore psr and we go
6191          *
6192          * we do not touch either PMC nor PMD. the psr is not touched
6193          * by the overflow_handler. So we are safe w.r.t. to interrupt
6194          * concurrency even without interrupt masking.
6195          */
6196         if (likely(owner == task)) {
6197                 if (likely(psr_up)) pfm_set_psr_up();
6198                 return;
6199         }
6200 
6201         /*
6202          * someone else is still using the PMU, first push it out and
6203          * then we'll be able to install our stuff !
6204          *
6205          * Upon return, there will be no owner for the current PMU
6206          */
6207         if (owner) pfm_lazy_save_regs(owner);
6208 
6209         /*
6210          * To avoid leaking information to the user level when psr.sp=0,
6211          * we must reload ALL implemented pmds (even the ones we don't use).
6212          * In the kernel we only allow PFM_READ_PMDS on registers which
6213          * we initialized or requested (sampling) so there is no risk there.
6214          */
6215         pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6216 
6217         /*
6218          * ALL accessible PMCs are systematically reloaded, unused registers
6219          * get their default (from pfm_reset_pmu_state()) values to avoid picking
6220          * up stale configuration.
6221          *
6222          * PMC0 is never in the mask. It is always restored separately
6223          */
6224         pmc_mask = ctx->ctx_all_pmcs[0];
6225 
6226         pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6227         pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6228 
6229         /*
6230          * check for pending overflow at the time the state
6231          * was saved.
6232          */
6233         if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6234                 /*
6235                  * reload pmc0 with the overflow information
6236                  * On McKinley PMU, this will trigger a PMU interrupt
6237                  */
6238                 ia64_set_pmc(0, ctx->th_pmcs[0]);
6239                 ia64_srlz_d();
6240 
6241                 ctx->th_pmcs[0] = 0UL;
6242 
6243                 /*
6244                  * will replay the PMU interrupt
6245                  */
6246                 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6247 
6248                 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6249         }
6250 
6251         /*
6252          * establish new ownership. 
6253          */
6254         SET_PMU_OWNER(task, ctx);
6255 
6256         /*
6257          * restore the psr.up bit. measurement
6258          * is active again.
6259          * no PMU interrupt can happen at this point
6260          * because we still have interrupts disabled.
6261          */
6262         if (likely(psr_up)) pfm_set_psr_up();
6263 }
6264 #endif /* CONFIG_SMP */
6265 
6266 /*
6267  * this function assumes monitoring is stopped
6268  */
6269 static void
6270 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6271 {
6272         u64 pmc0;
6273         unsigned long mask2, val, pmd_val, ovfl_val;
6274         int i, can_access_pmu = 0;
6275         int is_self;
6276 
6277         /*
6278          * is the caller the task being monitored (or which initiated the
6279          * session for system wide measurements)
6280          */
6281         is_self = ctx->ctx_task == task ? 1 : 0;
6282 
6283         /*
6284          * can access PMU is task is the owner of the PMU state on the current CPU
6285          * or if we are running on the CPU bound to the context in system-wide mode
6286          * (that is not necessarily the task the context is attached to in this mode).
6287          * In system-wide we always have can_access_pmu true because a task running on an
6288          * invalid processor is flagged earlier in the call stack (see pfm_stop).
6289          */
6290         can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6291         if (can_access_pmu) {
6292                 /*
6293                  * Mark the PMU as not owned
6294                  * This will cause the interrupt handler to do nothing in case an overflow
6295                  * interrupt was in-flight
6296                  * This also guarantees that pmc0 will contain the final state
6297                  * It virtually gives us full control on overflow processing from that point
6298                  * on.
6299                  */
6300                 SET_PMU_OWNER(NULL, NULL);
6301                 DPRINT(("releasing ownership\n"));
6302 
6303                 /*
6304                  * read current overflow status:
6305                  *
6306                  * we are guaranteed to read the final stable state
6307                  */
6308                 ia64_srlz_d();
6309                 pmc0 = ia64_get_pmc(0); /* slow */
6310 
6311                 /*
6312                  * reset freeze bit, overflow status information destroyed
6313                  */
6314                 pfm_unfreeze_pmu();
6315         } else {
6316                 pmc0 = ctx->th_pmcs[0];
6317                 /*
6318                  * clear whatever overflow status bits there were
6319                  */
6320                 ctx->th_pmcs[0] = 0;
6321         }
6322         ovfl_val = pmu_conf->ovfl_val;
6323         /*
6324          * we save all the used pmds
6325          * we take care of overflows for counting PMDs
6326          *
6327          * XXX: sampling situation is not taken into account here
6328          */
6329         mask2 = ctx->ctx_used_pmds[0];
6330 
6331         DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6332 
6333         for (i = 0; mask2; i++, mask2>>=1) {
6334 
6335                 /* skip non used pmds */
6336                 if ((mask2 & 0x1) == 0) continue;
6337 
6338                 /*
6339                  * can access PMU always true in system wide mode
6340                  */
6341                 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6342 
6343                 if (PMD_IS_COUNTING(i)) {
6344                         DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6345                                 task_pid_nr(task),
6346                                 i,
6347                                 ctx->ctx_pmds[i].val,
6348                                 val & ovfl_val));
6349 
6350                         /*
6351                          * we rebuild the full 64 bit value of the counter
6352                          */
6353                         val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6354 
6355                         /*
6356                          * now everything is in ctx_pmds[] and we need
6357                          * to clear the saved context from save_regs() such that
6358                          * pfm_read_pmds() gets the correct value
6359                          */
6360                         pmd_val = 0UL;
6361 
6362                         /*
6363                          * take care of overflow inline
6364                          */
6365                         if (pmc0 & (1UL << i)) {
6366                                 val += 1 + ovfl_val;
6367                                 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6368                         }
6369                 }
6370 
6371                 DPRINT(("[%d] ctx_pmd[%d]=0x%lx  pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6372 
6373                 if (is_self) ctx->th_pmds[i] = pmd_val;
6374 
6375                 ctx->ctx_pmds[i].val = val;
6376         }
6377 }
6378 
6379 static struct irqaction perfmon_irqaction = {
6380         .handler = pfm_interrupt_handler,
6381         .name    = "perfmon"
6382 };
6383 
6384 static void
6385 pfm_alt_save_pmu_state(void *data)
6386 {
6387         struct pt_regs *regs;
6388 
6389         regs = task_pt_regs(current);
6390 
6391         DPRINT(("called\n"));
6392 
6393         /*
6394          * should not be necessary but
6395          * let's take not risk
6396          */
6397         pfm_clear_psr_up();
6398         pfm_clear_psr_pp();
6399         ia64_psr(regs)->pp = 0;
6400 
6401         /*
6402          * This call is required
6403          * May cause a spurious interrupt on some processors
6404          */
6405         pfm_freeze_pmu();
6406 
6407         ia64_srlz_d();
6408 }
6409 
6410 void
6411 pfm_alt_restore_pmu_state(void *data)
6412 {
6413         struct pt_regs *regs;
6414 
6415         regs = task_pt_regs(current);
6416 
6417         DPRINT(("called\n"));
6418 
6419         /*
6420          * put PMU back in state expected
6421          * by perfmon
6422          */
6423         pfm_clear_psr_up();
6424         pfm_clear_psr_pp();
6425         ia64_psr(regs)->pp = 0;
6426 
6427         /*
6428          * perfmon runs with PMU unfrozen at all times
6429          */
6430         pfm_unfreeze_pmu();
6431 
6432         ia64_srlz_d();
6433 }
6434 
6435 int
6436 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6437 {
6438         int ret, i;
6439         int reserve_cpu;
6440 
6441         /* some sanity checks */
6442         if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6443 
6444         /* do the easy test first */
6445         if (pfm_alt_intr_handler) return -EBUSY;
6446 
6447         /* one at a time in the install or remove, just fail the others */
6448         if (!spin_trylock(&pfm_alt_install_check)) {
6449                 return -EBUSY;
6450         }
6451 
6452         /* reserve our session */
6453         for_each_online_cpu(reserve_cpu) {
6454                 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6455                 if (ret) goto cleanup_reserve;
6456         }
6457 
6458         /* save the current system wide pmu states */
6459         ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6460         if (ret) {
6461                 DPRINT(("on_each_cpu() failed: %d\n", ret));
6462                 goto cleanup_reserve;
6463         }
6464 
6465         /* officially change to the alternate interrupt handler */
6466         pfm_alt_intr_handler = hdl;
6467 
6468         spin_unlock(&pfm_alt_install_check);
6469 
6470         return 0;
6471 
6472 cleanup_reserve:
6473         for_each_online_cpu(i) {
6474                 /* don't unreserve more than we reserved */
6475                 if (i >= reserve_cpu) break;
6476 
6477                 pfm_unreserve_session(NULL, 1, i);
6478         }
6479 
6480         spin_unlock(&pfm_alt_install_check);
6481 
6482         return ret;
6483 }
6484 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6485 
6486 int
6487 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6488 {
6489         int i;
6490         int ret;
6491 
6492         if (hdl == NULL) return -EINVAL;
6493 
6494         /* cannot remove someone else's handler! */
6495         if (pfm_alt_intr_handler != hdl) return -EINVAL;
6496 
6497         /* one at a time in the install or remove, just fail the others */
6498         if (!spin_trylock(&pfm_alt_install_check)) {
6499                 return -EBUSY;
6500         }
6501 
6502         pfm_alt_intr_handler = NULL;
6503 
6504         ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6505         if (ret) {
6506                 DPRINT(("on_each_cpu() failed: %d\n", ret));
6507         }
6508 
6509         for_each_online_cpu(i) {
6510                 pfm_unreserve_session(NULL, 1, i);
6511         }
6512 
6513         spin_unlock(&pfm_alt_install_check);
6514 
6515         return 0;
6516 }
6517 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6518 
6519 /*
6520  * perfmon initialization routine, called from the initcall() table
6521  */
6522 static int init_pfm_fs(void);
6523 
6524 static int __init
6525 pfm_probe_pmu(void)
6526 {
6527         pmu_config_t **p;
6528         int family;
6529 
6530         family = local_cpu_data->family;
6531         p      = pmu_confs;
6532 
6533         while(*p) {
6534                 if ((*p)->probe) {
6535                         if ((*p)->probe() == 0) goto found;
6536                 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6537                         goto found;
6538                 }
6539                 p++;
6540         }
6541         return -1;
6542 found:
6543         pmu_conf = *p;
6544         return 0;
6545 }
6546 
6547 static const struct file_operations pfm_proc_fops = {
6548         .open           = pfm_proc_open,
6549         .read           = seq_read,
6550         .llseek         = seq_lseek,
6551         .release        = seq_release,
6552 };
6553 
6554 int __init
6555 pfm_init(void)
6556 {
6557         unsigned int n, n_counters, i;
6558 
6559         printk("perfmon: version %u.%u IRQ %u\n",
6560                 PFM_VERSION_MAJ,
6561                 PFM_VERSION_MIN,
6562                 IA64_PERFMON_VECTOR);
6563 
6564         if (pfm_probe_pmu()) {
6565                 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n", 
6566                                 local_cpu_data->family);
6567                 return -ENODEV;
6568         }
6569 
6570         /*
6571          * compute the number of implemented PMD/PMC from the
6572          * description tables
6573          */
6574         n = 0;
6575         for (i=0; PMC_IS_LAST(i) == 0;  i++) {
6576                 if (PMC_IS_IMPL(i) == 0) continue;
6577                 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6578                 n++;
6579         }
6580         pmu_conf->num_pmcs = n;
6581 
6582         n = 0; n_counters = 0;
6583         for (i=0; PMD_IS_LAST(i) == 0;  i++) {
6584                 if (PMD_IS_IMPL(i) == 0) continue;
6585                 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6586                 n++;
6587                 if (PMD_IS_COUNTING(i)) n_counters++;
6588         }
6589         pmu_conf->num_pmds      = n;
6590         pmu_conf->num_counters  = n_counters;
6591 
6592         /*
6593          * sanity checks on the number of debug registers
6594          */
6595         if (pmu_conf->use_rr_dbregs) {
6596                 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6597                         printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6598                         pmu_conf = NULL;
6599                         return -1;
6600                 }
6601                 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6602                         printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6603                         pmu_conf = NULL;
6604                         return -1;
6605                 }
6606         }
6607 
6608         printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6609                pmu_conf->pmu_name,
6610                pmu_conf->num_pmcs,
6611                pmu_conf->num_pmds,
6612                pmu_conf->num_counters,
6613                ffz(pmu_conf->ovfl_val));
6614 
6615         /* sanity check */
6616         if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6617                 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6618                 pmu_conf = NULL;
6619                 return -1;
6620         }
6621 
6622         /*
6623          * create /proc/perfmon (mostly for debugging purposes)
6624          */
6625         perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6626         if (perfmon_dir == NULL) {
6627                 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6628                 pmu_conf = NULL;
6629                 return -1;
6630         }
6631 
6632         /*
6633          * create /proc/sys/kernel/perfmon (for debugging purposes)
6634          */
6635         pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6636 
6637         /*
6638          * initialize all our spinlocks
6639          */
6640         spin_lock_init(&pfm_sessions.pfs_lock);
6641         spin_lock_init(&pfm_buffer_fmt_lock);
6642 
6643         init_pfm_fs();
6644 
6645         for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6646 
6647         return 0;
6648 }
6649 
6650 __initcall(pfm_init);
6651 
6652 /*
6653  * this function is called before pfm_init()
6654  */
6655 void
6656 pfm_init_percpu (void)
6657 {
6658         static int first_time=1;
6659         /*
6660          * make sure no measurement is active
6661          * (may inherit programmed PMCs from EFI).
6662          */
6663         pfm_clear_psr_pp();
6664         pfm_clear_psr_up();
6665 
6666         /*
6667          * we run with the PMU not frozen at all times
6668          */
6669         pfm_unfreeze_pmu();
6670 
6671         if (first_time) {
6672                 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6673                 first_time=0;
6674         }
6675 
6676         ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6677         ia64_srlz_d();
6678 }
6679 
6680 /*
6681  * used for debug purposes only
6682  */
6683 void
6684 dump_pmu_state(const char *from)
6685 {
6686         struct task_struct *task;
6687         struct pt_regs *regs;
6688         pfm_context_t *ctx;
6689         unsigned long psr, dcr, info, flags;
6690         int i, this_cpu;
6691 
6692         local_irq_save(flags);
6693 
6694         this_cpu = smp_processor_id();
6695         regs     = task_pt_regs(current);
6696         info     = PFM_CPUINFO_GET();
6697         dcr      = ia64_getreg(_IA64_REG_CR_DCR);
6698 
6699         if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6700                 local_irq_restore(flags);
6701                 return;
6702         }
6703 
6704         printk("CPU%d from %s() current [%d] iip=0x%lx %s\n", 
6705                 this_cpu, 
6706                 from, 
6707                 task_pid_nr(current),
6708                 regs->cr_iip,
6709                 current->comm);
6710 
6711         task = GET_PMU_OWNER();
6712         ctx  = GET_PMU_CTX();
6713 
6714         printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6715 
6716         psr = pfm_get_psr();
6717 
6718         printk("->CPU%d pmc0=0x%lx psr.pp=%d psr.up=%d dcr.pp=%d syst_info=0x%lx user_psr.up=%d user_psr.pp=%d\n", 
6719                 this_cpu,
6720                 ia64_get_pmc(0),
6721                 psr & IA64_PSR_PP ? 1 : 0,
6722                 psr & IA64_PSR_UP ? 1 : 0,
6723                 dcr & IA64_DCR_PP ? 1 : 0,
6724                 info,
6725                 ia64_psr(regs)->up,
6726                 ia64_psr(regs)->pp);
6727 
6728         ia64_psr(regs)->up = 0;
6729         ia64_psr(regs)->pp = 0;
6730 
6731         for (i=1; PMC_IS_LAST(i) == 0; i++) {
6732                 if (PMC_IS_IMPL(i) == 0) continue;
6733                 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, ctx->th_pmcs[i]);
6734         }
6735 
6736         for (i=1; PMD_IS_LAST(i) == 0; i++) {
6737                 if (PMD_IS_IMPL(i) == 0) continue;
6738                 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, ctx->th_pmds[i]);
6739         }
6740 
6741         if (ctx) {
6742                 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6743                                 this_cpu,
6744                                 ctx->ctx_state,
6745                                 ctx->ctx_smpl_vaddr,
6746                                 ctx->ctx_smpl_hdr,
6747                                 ctx->ctx_msgq_head,
6748                                 ctx->ctx_msgq_tail,
6749                                 ctx->ctx_saved_psr_up);
6750         }
6751         local_irq_restore(flags);
6752 }
6753 
6754 /*
6755  * called from process.c:copy_thread(). task is new child.
6756  */
6757 void
6758 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6759 {
6760         struct thread_struct *thread;
6761 
6762         DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6763 
6764         thread = &task->thread;
6765 
6766         /*
6767          * cut links inherited from parent (current)
6768          */
6769         thread->pfm_context = NULL;
6770 
6771         PFM_SET_WORK_PENDING(task, 0);
6772 
6773         /*
6774          * the psr bits are already set properly in copy_threads()
6775          */
6776 }
6777 #else  /* !CONFIG_PERFMON */
6778 asmlinkage long
6779 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6780 {
6781         return -ENOSYS;
6782 }
6783 #endif /* CONFIG_PERFMON */
6784 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp