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

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

Version: ~ [ linux-4.18-rc5 ] ~ [ linux-4.17.6 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.55 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.112 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.140 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.115 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.57 ] ~ [ 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-3.2.102 ] ~ [ 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 __poll_t
1648 pfm_poll(struct file *filp, poll_table * wait)
1649 {
1650         pfm_context_t *ctx;
1651         unsigned long flags;
1652         __poll_t 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 =  EPOLLIN | EPOLLRDNORM;
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                 /* make sure task cannot go away while we operate on it */
2614                 p = find_get_task_by_vpid(pid);
2615                 if (!p)
2616                         return -ESRCH;
2617         }
2618 
2619         ret = pfm_task_incompatible(ctx, p);
2620         if (ret == 0) {
2621                 *task = p;
2622         } else if (p != current) {
2623                 pfm_put_task(p);
2624         }
2625         return ret;
2626 }
2627 
2628 
2629 
2630 static int
2631 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2632 {
2633         pfarg_context_t *req = (pfarg_context_t *)arg;
2634         struct file *filp;
2635         struct path path;
2636         int ctx_flags;
2637         int fd;
2638         int ret;
2639 
2640         /* let's check the arguments first */
2641         ret = pfarg_is_sane(current, req);
2642         if (ret < 0)
2643                 return ret;
2644 
2645         ctx_flags = req->ctx_flags;
2646 
2647         ret = -ENOMEM;
2648 
2649         fd = get_unused_fd_flags(0);
2650         if (fd < 0)
2651                 return fd;
2652 
2653         ctx = pfm_context_alloc(ctx_flags);
2654         if (!ctx)
2655                 goto error;
2656 
2657         filp = pfm_alloc_file(ctx);
2658         if (IS_ERR(filp)) {
2659                 ret = PTR_ERR(filp);
2660                 goto error_file;
2661         }
2662 
2663         req->ctx_fd = ctx->ctx_fd = fd;
2664 
2665         /*
2666          * does the user want to sample?
2667          */
2668         if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2669                 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2670                 if (ret)
2671                         goto buffer_error;
2672         }
2673 
2674         DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2675                 ctx,
2676                 ctx_flags,
2677                 ctx->ctx_fl_system,
2678                 ctx->ctx_fl_block,
2679                 ctx->ctx_fl_excl_idle,
2680                 ctx->ctx_fl_no_msg,
2681                 ctx->ctx_fd));
2682 
2683         /*
2684          * initialize soft PMU state
2685          */
2686         pfm_reset_pmu_state(ctx);
2687 
2688         fd_install(fd, filp);
2689 
2690         return 0;
2691 
2692 buffer_error:
2693         path = filp->f_path;
2694         put_filp(filp);
2695         path_put(&path);
2696 
2697         if (ctx->ctx_buf_fmt) {
2698                 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2699         }
2700 error_file:
2701         pfm_context_free(ctx);
2702 
2703 error:
2704         put_unused_fd(fd);
2705         return ret;
2706 }
2707 
2708 static inline unsigned long
2709 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2710 {
2711         unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2712         unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2713         extern unsigned long carta_random32 (unsigned long seed);
2714 
2715         if (reg->flags & PFM_REGFL_RANDOM) {
2716                 new_seed = carta_random32(old_seed);
2717                 val -= (old_seed & mask);       /* counter values are negative numbers! */
2718                 if ((mask >> 32) != 0)
2719                         /* construct a full 64-bit random value: */
2720                         new_seed |= carta_random32(old_seed >> 32) << 32;
2721                 reg->seed = new_seed;
2722         }
2723         reg->lval = val;
2724         return val;
2725 }
2726 
2727 static void
2728 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2729 {
2730         unsigned long mask = ovfl_regs[0];
2731         unsigned long reset_others = 0UL;
2732         unsigned long val;
2733         int i;
2734 
2735         /*
2736          * now restore reset value on sampling overflowed counters
2737          */
2738         mask >>= PMU_FIRST_COUNTER;
2739         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2740 
2741                 if ((mask & 0x1UL) == 0UL) continue;
2742 
2743                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2744                 reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2745 
2746                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2747         }
2748 
2749         /*
2750          * Now take care of resetting the other registers
2751          */
2752         for(i = 0; reset_others; i++, reset_others >>= 1) {
2753 
2754                 if ((reset_others & 0x1) == 0) continue;
2755 
2756                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2757 
2758                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2759                           is_long_reset ? "long" : "short", i, val));
2760         }
2761 }
2762 
2763 static void
2764 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2765 {
2766         unsigned long mask = ovfl_regs[0];
2767         unsigned long reset_others = 0UL;
2768         unsigned long val;
2769         int i;
2770 
2771         DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2772 
2773         if (ctx->ctx_state == PFM_CTX_MASKED) {
2774                 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2775                 return;
2776         }
2777 
2778         /*
2779          * now restore reset value on sampling overflowed counters
2780          */
2781         mask >>= PMU_FIRST_COUNTER;
2782         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2783 
2784                 if ((mask & 0x1UL) == 0UL) continue;
2785 
2786                 val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2787                 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2788 
2789                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2790 
2791                 pfm_write_soft_counter(ctx, i, val);
2792         }
2793 
2794         /*
2795          * Now take care of resetting the other registers
2796          */
2797         for(i = 0; reset_others; i++, reset_others >>= 1) {
2798 
2799                 if ((reset_others & 0x1) == 0) continue;
2800 
2801                 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2802 
2803                 if (PMD_IS_COUNTING(i)) {
2804                         pfm_write_soft_counter(ctx, i, val);
2805                 } else {
2806                         ia64_set_pmd(i, val);
2807                 }
2808                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2809                           is_long_reset ? "long" : "short", i, val));
2810         }
2811         ia64_srlz_d();
2812 }
2813 
2814 static int
2815 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2816 {
2817         struct task_struct *task;
2818         pfarg_reg_t *req = (pfarg_reg_t *)arg;
2819         unsigned long value, pmc_pm;
2820         unsigned long smpl_pmds, reset_pmds, impl_pmds;
2821         unsigned int cnum, reg_flags, flags, pmc_type;
2822         int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2823         int is_monitor, is_counting, state;
2824         int ret = -EINVAL;
2825         pfm_reg_check_t wr_func;
2826 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2827 
2828         state     = ctx->ctx_state;
2829         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2830         is_system = ctx->ctx_fl_system;
2831         task      = ctx->ctx_task;
2832         impl_pmds = pmu_conf->impl_pmds[0];
2833 
2834         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2835 
2836         if (is_loaded) {
2837                 /*
2838                  * In system wide and when the context is loaded, access can only happen
2839                  * when the caller is running on the CPU being monitored by the session.
2840                  * It does not have to be the owner (ctx_task) of the context per se.
2841                  */
2842                 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2843                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2844                         return -EBUSY;
2845                 }
2846                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2847         }
2848         expert_mode = pfm_sysctl.expert_mode; 
2849 
2850         for (i = 0; i < count; i++, req++) {
2851 
2852                 cnum       = req->reg_num;
2853                 reg_flags  = req->reg_flags;
2854                 value      = req->reg_value;
2855                 smpl_pmds  = req->reg_smpl_pmds[0];
2856                 reset_pmds = req->reg_reset_pmds[0];
2857                 flags      = 0;
2858 
2859 
2860                 if (cnum >= PMU_MAX_PMCS) {
2861                         DPRINT(("pmc%u is invalid\n", cnum));
2862                         goto error;
2863                 }
2864 
2865                 pmc_type   = pmu_conf->pmc_desc[cnum].type;
2866                 pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2867                 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2868                 is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2869 
2870                 /*
2871                  * we reject all non implemented PMC as well
2872                  * as attempts to modify PMC[0-3] which are used
2873                  * as status registers by the PMU
2874                  */
2875                 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2876                         DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2877                         goto error;
2878                 }
2879                 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2880                 /*
2881                  * If the PMC is a monitor, then if the value is not the default:
2882                  *      - system-wide session: PMCx.pm=1 (privileged monitor)
2883                  *      - per-task           : PMCx.pm=0 (user monitor)
2884                  */
2885                 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2886                         DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2887                                 cnum,
2888                                 pmc_pm,
2889                                 is_system));
2890                         goto error;
2891                 }
2892 
2893                 if (is_counting) {
2894                         /*
2895                          * enforce generation of overflow interrupt. Necessary on all
2896                          * CPUs.
2897                          */
2898                         value |= 1 << PMU_PMC_OI;
2899 
2900                         if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2901                                 flags |= PFM_REGFL_OVFL_NOTIFY;
2902                         }
2903 
2904                         if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2905 
2906                         /* verify validity of smpl_pmds */
2907                         if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2908                                 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2909                                 goto error;
2910                         }
2911 
2912                         /* verify validity of reset_pmds */
2913                         if ((reset_pmds & impl_pmds) != reset_pmds) {
2914                                 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2915                                 goto error;
2916                         }
2917                 } else {
2918                         if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2919                                 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2920                                 goto error;
2921                         }
2922                         /* eventid on non-counting monitors are ignored */
2923                 }
2924 
2925                 /*
2926                  * execute write checker, if any
2927                  */
2928                 if (likely(expert_mode == 0 && wr_func)) {
2929                         ret = (*wr_func)(task, ctx, cnum, &value, regs);
2930                         if (ret) goto error;
2931                         ret = -EINVAL;
2932                 }
2933 
2934                 /*
2935                  * no error on this register
2936                  */
2937                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2938 
2939                 /*
2940                  * Now we commit the changes to the software state
2941                  */
2942 
2943                 /*
2944                  * update overflow information
2945                  */
2946                 if (is_counting) {
2947                         /*
2948                          * full flag update each time a register is programmed
2949                          */
2950                         ctx->ctx_pmds[cnum].flags = flags;
2951 
2952                         ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2953                         ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2954                         ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
2955 
2956                         /*
2957                          * Mark all PMDS to be accessed as used.
2958                          *
2959                          * We do not keep track of PMC because we have to
2960                          * systematically restore ALL of them.
2961                          *
2962                          * We do not update the used_monitors mask, because
2963                          * if we have not programmed them, then will be in
2964                          * a quiescent state, therefore we will not need to
2965                          * mask/restore then when context is MASKED.
2966                          */
2967                         CTX_USED_PMD(ctx, reset_pmds);
2968                         CTX_USED_PMD(ctx, smpl_pmds);
2969                         /*
2970                          * make sure we do not try to reset on
2971                          * restart because we have established new values
2972                          */
2973                         if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
2974                 }
2975                 /*
2976                  * Needed in case the user does not initialize the equivalent
2977                  * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2978                  * possible leak here.
2979                  */
2980                 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
2981 
2982                 /*
2983                  * keep track of the monitor PMC that we are using.
2984                  * we save the value of the pmc in ctx_pmcs[] and if
2985                  * the monitoring is not stopped for the context we also
2986                  * place it in the saved state area so that it will be
2987                  * picked up later by the context switch code.
2988                  *
2989                  * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
2990                  *
2991                  * The value in th_pmcs[] may be modified on overflow, i.e.,  when
2992                  * monitoring needs to be stopped.
2993                  */
2994                 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
2995 
2996                 /*
2997                  * update context state
2998                  */
2999                 ctx->ctx_pmcs[cnum] = value;
3000 
3001                 if (is_loaded) {
3002                         /*
3003                          * write thread state
3004                          */
3005                         if (is_system == 0) ctx->th_pmcs[cnum] = value;
3006 
3007                         /*
3008                          * write hardware register if we can
3009                          */
3010                         if (can_access_pmu) {
3011                                 ia64_set_pmc(cnum, value);
3012                         }
3013 #ifdef CONFIG_SMP
3014                         else {
3015                                 /*
3016                                  * per-task SMP only here
3017                                  *
3018                                  * we are guaranteed that the task is not running on the other CPU,
3019                                  * we indicate that this PMD will need to be reloaded if the task
3020                                  * is rescheduled on the CPU it ran last on.
3021                                  */
3022                                 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3023                         }
3024 #endif
3025                 }
3026 
3027                 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",
3028                           cnum,
3029                           value,
3030                           is_loaded,
3031                           can_access_pmu,
3032                           flags,
3033                           ctx->ctx_all_pmcs[0],
3034                           ctx->ctx_used_pmds[0],
3035                           ctx->ctx_pmds[cnum].eventid,
3036                           smpl_pmds,
3037                           reset_pmds,
3038                           ctx->ctx_reload_pmcs[0],
3039                           ctx->ctx_used_monitors[0],
3040                           ctx->ctx_ovfl_regs[0]));
3041         }
3042 
3043         /*
3044          * make sure the changes are visible
3045          */
3046         if (can_access_pmu) ia64_srlz_d();
3047 
3048         return 0;
3049 error:
3050         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3051         return ret;
3052 }
3053 
3054 static int
3055 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3056 {
3057         struct task_struct *task;
3058         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3059         unsigned long value, hw_value, ovfl_mask;
3060         unsigned int cnum;
3061         int i, can_access_pmu = 0, state;
3062         int is_counting, is_loaded, is_system, expert_mode;
3063         int ret = -EINVAL;
3064         pfm_reg_check_t wr_func;
3065 
3066 
3067         state     = ctx->ctx_state;
3068         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3069         is_system = ctx->ctx_fl_system;
3070         ovfl_mask = pmu_conf->ovfl_val;
3071         task      = ctx->ctx_task;
3072 
3073         if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3074 
3075         /*
3076          * on both UP and SMP, we can only write to the PMC when the task is
3077          * the owner of the local PMU.
3078          */
3079         if (likely(is_loaded)) {
3080                 /*
3081                  * In system wide and when the context is loaded, access can only happen
3082                  * when the caller is running on the CPU being monitored by the session.
3083                  * It does not have to be the owner (ctx_task) of the context per se.
3084                  */
3085                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3086                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3087                         return -EBUSY;
3088                 }
3089                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3090         }
3091         expert_mode = pfm_sysctl.expert_mode; 
3092 
3093         for (i = 0; i < count; i++, req++) {
3094 
3095                 cnum  = req->reg_num;
3096                 value = req->reg_value;
3097 
3098                 if (!PMD_IS_IMPL(cnum)) {
3099                         DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3100                         goto abort_mission;
3101                 }
3102                 is_counting = PMD_IS_COUNTING(cnum);
3103                 wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3104 
3105                 /*
3106                  * execute write checker, if any
3107                  */
3108                 if (unlikely(expert_mode == 0 && wr_func)) {
3109                         unsigned long v = value;
3110 
3111                         ret = (*wr_func)(task, ctx, cnum, &v, regs);
3112                         if (ret) goto abort_mission;
3113 
3114                         value = v;
3115                         ret   = -EINVAL;
3116                 }
3117 
3118                 /*
3119                  * no error on this register
3120                  */
3121                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3122 
3123                 /*
3124                  * now commit changes to software state
3125                  */
3126                 hw_value = value;
3127 
3128                 /*
3129                  * update virtualized (64bits) counter
3130                  */
3131                 if (is_counting) {
3132                         /*
3133                          * write context state
3134                          */
3135                         ctx->ctx_pmds[cnum].lval = value;
3136 
3137                         /*
3138                          * when context is load we use the split value
3139                          */
3140                         if (is_loaded) {
3141                                 hw_value = value &  ovfl_mask;
3142                                 value    = value & ~ovfl_mask;
3143                         }
3144                 }
3145                 /*
3146                  * update reset values (not just for counters)
3147                  */
3148                 ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3149                 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3150 
3151                 /*
3152                  * update randomization parameters (not just for counters)
3153                  */
3154                 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3155                 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3156 
3157                 /*
3158                  * update context value
3159                  */
3160                 ctx->ctx_pmds[cnum].val  = value;
3161 
3162                 /*
3163                  * Keep track of what we use
3164                  *
3165                  * We do not keep track of PMC because we have to
3166                  * systematically restore ALL of them.
3167                  */
3168                 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3169 
3170                 /*
3171                  * mark this PMD register used as well
3172                  */
3173                 CTX_USED_PMD(ctx, RDEP(cnum));
3174 
3175                 /*
3176                  * make sure we do not try to reset on
3177                  * restart because we have established new values
3178                  */
3179                 if (is_counting && state == PFM_CTX_MASKED) {
3180                         ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3181                 }
3182 
3183                 if (is_loaded) {
3184                         /*
3185                          * write thread state
3186                          */
3187                         if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3188 
3189                         /*
3190                          * write hardware register if we can
3191                          */
3192                         if (can_access_pmu) {
3193                                 ia64_set_pmd(cnum, hw_value);
3194                         } else {
3195 #ifdef CONFIG_SMP
3196                                 /*
3197                                  * we are guaranteed that the task is not running on the other CPU,
3198                                  * we indicate that this PMD will need to be reloaded if the task
3199                                  * is rescheduled on the CPU it ran last on.
3200                                  */
3201                                 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3202 #endif
3203                         }
3204                 }
3205 
3206                 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3207                           "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",
3208                         cnum,
3209                         value,
3210                         is_loaded,
3211                         can_access_pmu,
3212                         hw_value,
3213                         ctx->ctx_pmds[cnum].val,
3214                         ctx->ctx_pmds[cnum].short_reset,
3215                         ctx->ctx_pmds[cnum].long_reset,
3216                         PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3217                         ctx->ctx_pmds[cnum].seed,
3218                         ctx->ctx_pmds[cnum].mask,
3219                         ctx->ctx_used_pmds[0],
3220                         ctx->ctx_pmds[cnum].reset_pmds[0],
3221                         ctx->ctx_reload_pmds[0],
3222                         ctx->ctx_all_pmds[0],
3223                         ctx->ctx_ovfl_regs[0]));
3224         }
3225 
3226         /*
3227          * make changes visible
3228          */
3229         if (can_access_pmu) ia64_srlz_d();
3230 
3231         return 0;
3232 
3233 abort_mission:
3234         /*
3235          * for now, we have only one possibility for error
3236          */
3237         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3238         return ret;
3239 }
3240 
3241 /*
3242  * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3243  * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3244  * interrupt is delivered during the call, it will be kept pending until we leave, making
3245  * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3246  * guaranteed to return consistent data to the user, it may simply be old. It is not
3247  * trivial to treat the overflow while inside the call because you may end up in
3248  * some module sampling buffer code causing deadlocks.
3249  */
3250 static int
3251 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3252 {
3253         struct task_struct *task;
3254         unsigned long val = 0UL, lval, ovfl_mask, sval;
3255         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3256         unsigned int cnum, reg_flags = 0;
3257         int i, can_access_pmu = 0, state;
3258         int is_loaded, is_system, is_counting, expert_mode;
3259         int ret = -EINVAL;
3260         pfm_reg_check_t rd_func;
3261 
3262         /*
3263          * access is possible when loaded only for
3264          * self-monitoring tasks or in UP mode
3265          */
3266 
3267         state     = ctx->ctx_state;
3268         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3269         is_system = ctx->ctx_fl_system;
3270         ovfl_mask = pmu_conf->ovfl_val;
3271         task      = ctx->ctx_task;
3272 
3273         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3274 
3275         if (likely(is_loaded)) {
3276                 /*
3277                  * In system wide and when the context is loaded, access can only happen
3278                  * when the caller is running on the CPU being monitored by the session.
3279                  * It does not have to be the owner (ctx_task) of the context per se.
3280                  */
3281                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3282                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3283                         return -EBUSY;
3284                 }
3285                 /*
3286                  * this can be true when not self-monitoring only in UP
3287                  */
3288                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3289 
3290                 if (can_access_pmu) ia64_srlz_d();
3291         }
3292         expert_mode = pfm_sysctl.expert_mode; 
3293 
3294         DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3295                 is_loaded,
3296                 can_access_pmu,
3297                 state));
3298 
3299         /*
3300          * on both UP and SMP, we can only read the PMD from the hardware register when
3301          * the task is the owner of the local PMU.
3302          */
3303 
3304         for (i = 0; i < count; i++, req++) {
3305 
3306                 cnum        = req->reg_num;
3307                 reg_flags   = req->reg_flags;
3308 
3309                 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3310                 /*
3311                  * we can only read the register that we use. That includes
3312                  * the one we explicitly initialize AND the one we want included
3313                  * in the sampling buffer (smpl_regs).
3314                  *
3315                  * Having this restriction allows optimization in the ctxsw routine
3316                  * without compromising security (leaks)
3317                  */
3318                 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3319 
3320                 sval        = ctx->ctx_pmds[cnum].val;
3321                 lval        = ctx->ctx_pmds[cnum].lval;
3322                 is_counting = PMD_IS_COUNTING(cnum);
3323 
3324                 /*
3325                  * If the task is not the current one, then we check if the
3326                  * PMU state is still in the local live register due to lazy ctxsw.
3327                  * If true, then we read directly from the registers.
3328                  */
3329                 if (can_access_pmu){
3330                         val = ia64_get_pmd(cnum);
3331                 } else {
3332                         /*
3333                          * context has been saved
3334                          * if context is zombie, then task does not exist anymore.
3335                          * In this case, we use the full value saved in the context (pfm_flush_regs()).
3336                          */
3337                         val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3338                 }
3339                 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3340 
3341                 if (is_counting) {
3342                         /*
3343                          * XXX: need to check for overflow when loaded
3344                          */
3345                         val &= ovfl_mask;
3346                         val += sval;
3347                 }
3348 
3349                 /*
3350                  * execute read checker, if any
3351                  */
3352                 if (unlikely(expert_mode == 0 && rd_func)) {
3353                         unsigned long v = val;
3354                         ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3355                         if (ret) goto error;
3356                         val = v;
3357                         ret = -EINVAL;
3358                 }
3359 
3360                 PFM_REG_RETFLAG_SET(reg_flags, 0);
3361 
3362                 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3363 
3364                 /*
3365                  * update register return value, abort all if problem during copy.
3366                  * we only modify the reg_flags field. no check mode is fine because
3367                  * access has been verified upfront in sys_perfmonctl().
3368                  */
3369                 req->reg_value            = val;
3370                 req->reg_flags            = reg_flags;
3371                 req->reg_last_reset_val   = lval;
3372         }
3373 
3374         return 0;
3375 
3376 error:
3377         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3378         return ret;
3379 }
3380 
3381 int
3382 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3383 {
3384         pfm_context_t *ctx;
3385 
3386         if (req == NULL) return -EINVAL;
3387 
3388         ctx = GET_PMU_CTX();
3389 
3390         if (ctx == NULL) return -EINVAL;
3391 
3392         /*
3393          * for now limit to current task, which is enough when calling
3394          * from overflow handler
3395          */
3396         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3397 
3398         return pfm_write_pmcs(ctx, req, nreq, regs);
3399 }
3400 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3401 
3402 int
3403 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3404 {
3405         pfm_context_t *ctx;
3406 
3407         if (req == NULL) return -EINVAL;
3408 
3409         ctx = GET_PMU_CTX();
3410 
3411         if (ctx == NULL) return -EINVAL;
3412 
3413         /*
3414          * for now limit to current task, which is enough when calling
3415          * from overflow handler
3416          */
3417         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3418 
3419         return pfm_read_pmds(ctx, req, nreq, regs);
3420 }
3421 EXPORT_SYMBOL(pfm_mod_read_pmds);
3422 
3423 /*
3424  * Only call this function when a process it trying to
3425  * write the debug registers (reading is always allowed)
3426  */
3427 int
3428 pfm_use_debug_registers(struct task_struct *task)
3429 {
3430         pfm_context_t *ctx = task->thread.pfm_context;
3431         unsigned long flags;
3432         int ret = 0;
3433 
3434         if (pmu_conf->use_rr_dbregs == 0) return 0;
3435 
3436         DPRINT(("called for [%d]\n", task_pid_nr(task)));
3437 
3438         /*
3439          * do it only once
3440          */
3441         if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3442 
3443         /*
3444          * Even on SMP, we do not need to use an atomic here because
3445          * the only way in is via ptrace() and this is possible only when the
3446          * process is stopped. Even in the case where the ctxsw out is not totally
3447          * completed by the time we come here, there is no way the 'stopped' process
3448          * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3449          * So this is always safe.
3450          */
3451         if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3452 
3453         LOCK_PFS(flags);
3454 
3455         /*
3456          * We cannot allow setting breakpoints when system wide monitoring
3457          * sessions are using the debug registers.
3458          */
3459         if (pfm_sessions.pfs_sys_use_dbregs> 0)
3460                 ret = -1;
3461         else
3462                 pfm_sessions.pfs_ptrace_use_dbregs++;
3463 
3464         DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3465                   pfm_sessions.pfs_ptrace_use_dbregs,
3466                   pfm_sessions.pfs_sys_use_dbregs,
3467                   task_pid_nr(task), ret));
3468 
3469         UNLOCK_PFS(flags);
3470 
3471         return ret;
3472 }
3473 
3474 /*
3475  * This function is called for every task that exits with the
3476  * IA64_THREAD_DBG_VALID set. This indicates a task which was
3477  * able to use the debug registers for debugging purposes via
3478  * ptrace(). Therefore we know it was not using them for
3479  * performance monitoring, so we only decrement the number
3480  * of "ptraced" debug register users to keep the count up to date
3481  */
3482 int
3483 pfm_release_debug_registers(struct task_struct *task)
3484 {
3485         unsigned long flags;
3486         int ret;
3487 
3488         if (pmu_conf->use_rr_dbregs == 0) return 0;
3489 
3490         LOCK_PFS(flags);
3491         if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3492                 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3493                 ret = -1;
3494         }  else {
3495                 pfm_sessions.pfs_ptrace_use_dbregs--;
3496                 ret = 0;
3497         }
3498         UNLOCK_PFS(flags);
3499 
3500         return ret;
3501 }
3502 
3503 static int
3504 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3505 {
3506         struct task_struct *task;
3507         pfm_buffer_fmt_t *fmt;
3508         pfm_ovfl_ctrl_t rst_ctrl;
3509         int state, is_system;
3510         int ret = 0;
3511 
3512         state     = ctx->ctx_state;
3513         fmt       = ctx->ctx_buf_fmt;
3514         is_system = ctx->ctx_fl_system;
3515         task      = PFM_CTX_TASK(ctx);
3516 
3517         switch(state) {
3518                 case PFM_CTX_MASKED:
3519                         break;
3520                 case PFM_CTX_LOADED: 
3521                         if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3522                         /* fall through */
3523                 case PFM_CTX_UNLOADED:
3524                 case PFM_CTX_ZOMBIE:
3525                         DPRINT(("invalid state=%d\n", state));
3526                         return -EBUSY;
3527                 default:
3528                         DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3529                         return -EINVAL;
3530         }
3531 
3532         /*
3533          * In system wide and when the context is loaded, access can only happen
3534          * when the caller is running on the CPU being monitored by the session.
3535          * It does not have to be the owner (ctx_task) of the context per se.
3536          */
3537         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3538                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3539                 return -EBUSY;
3540         }
3541 
3542         /* sanity check */
3543         if (unlikely(task == NULL)) {
3544                 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3545                 return -EINVAL;
3546         }
3547 
3548         if (task == current || is_system) {
3549 
3550                 fmt = ctx->ctx_buf_fmt;
3551 
3552                 DPRINT(("restarting self %d ovfl=0x%lx\n",
3553                         task_pid_nr(task),
3554                         ctx->ctx_ovfl_regs[0]));
3555 
3556                 if (CTX_HAS_SMPL(ctx)) {
3557 
3558                         prefetch(ctx->ctx_smpl_hdr);
3559 
3560                         rst_ctrl.bits.mask_monitoring = 0;
3561                         rst_ctrl.bits.reset_ovfl_pmds = 0;
3562 
3563                         if (state == PFM_CTX_LOADED)
3564                                 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3565                         else
3566                                 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3567                 } else {
3568                         rst_ctrl.bits.mask_monitoring = 0;
3569                         rst_ctrl.bits.reset_ovfl_pmds = 1;
3570                 }
3571 
3572                 if (ret == 0) {
3573                         if (rst_ctrl.bits.reset_ovfl_pmds)
3574                                 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3575 
3576                         if (rst_ctrl.bits.mask_monitoring == 0) {
3577                                 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3578 
3579                                 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3580                         } else {
3581                                 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3582 
3583                                 // cannot use pfm_stop_monitoring(task, regs);
3584                         }
3585                 }
3586                 /*
3587                  * clear overflowed PMD mask to remove any stale information
3588                  */
3589                 ctx->ctx_ovfl_regs[0] = 0UL;
3590 
3591                 /*
3592                  * back to LOADED state
3593                  */
3594                 ctx->ctx_state = PFM_CTX_LOADED;
3595 
3596                 /*
3597                  * XXX: not really useful for self monitoring
3598                  */
3599                 ctx->ctx_fl_can_restart = 0;
3600 
3601                 return 0;
3602         }
3603 
3604         /* 
3605          * restart another task
3606          */
3607 
3608         /*
3609          * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3610          * one is seen by the task.
3611          */
3612         if (state == PFM_CTX_MASKED) {
3613                 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3614                 /*
3615                  * will prevent subsequent restart before this one is
3616                  * seen by other task
3617                  */
3618                 ctx->ctx_fl_can_restart = 0;
3619         }
3620 
3621         /*
3622          * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3623          * the task is blocked or on its way to block. That's the normal
3624          * restart path. If the monitoring is not masked, then the task
3625          * can be actively monitoring and we cannot directly intervene.
3626          * Therefore we use the trap mechanism to catch the task and
3627          * force it to reset the buffer/reset PMDs.
3628          *
3629          * if non-blocking, then we ensure that the task will go into
3630          * pfm_handle_work() before returning to user mode.
3631          *
3632          * We cannot explicitly reset another task, it MUST always
3633          * be done by the task itself. This works for system wide because
3634          * the tool that is controlling the session is logically doing 
3635          * "self-monitoring".
3636          */
3637         if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3638                 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3639                 complete(&ctx->ctx_restart_done);
3640         } else {
3641                 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3642 
3643                 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3644 
3645                 PFM_SET_WORK_PENDING(task, 1);
3646 
3647                 set_notify_resume(task);
3648 
3649                 /*
3650                  * XXX: send reschedule if task runs on another CPU
3651                  */
3652         }
3653         return 0;
3654 }
3655 
3656 static int
3657 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3658 {
3659         unsigned int m = *(unsigned int *)arg;
3660 
3661         pfm_sysctl.debug = m == 0 ? 0 : 1;
3662 
3663         printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3664 
3665         if (m == 0) {
3666                 memset(pfm_stats, 0, sizeof(pfm_stats));
3667                 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3668         }
3669         return 0;
3670 }
3671 
3672 /*
3673  * arg can be NULL and count can be zero for this function
3674  */
3675 static int
3676 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3677 {
3678         struct thread_struct *thread = NULL;
3679         struct task_struct *task;
3680         pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3681         unsigned long flags;
3682         dbreg_t dbreg;
3683         unsigned int rnum;
3684         int first_time;
3685         int ret = 0, state;
3686         int i, can_access_pmu = 0;
3687         int is_system, is_loaded;
3688 
3689         if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3690 
3691         state     = ctx->ctx_state;
3692         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3693         is_system = ctx->ctx_fl_system;
3694         task      = ctx->ctx_task;
3695 
3696         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3697 
3698         /*
3699          * on both UP and SMP, we can only write to the PMC when the task is
3700          * the owner of the local PMU.
3701          */
3702         if (is_loaded) {
3703                 thread = &task->thread;
3704                 /*
3705                  * In system wide and when the context is loaded, access can only happen
3706                  * when the caller is running on the CPU being monitored by the session.
3707                  * It does not have to be the owner (ctx_task) of the context per se.
3708                  */
3709                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3710                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3711                         return -EBUSY;
3712                 }
3713                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3714         }
3715 
3716         /*
3717          * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3718          * ensuring that no real breakpoint can be installed via this call.
3719          *
3720          * IMPORTANT: regs can be NULL in this function
3721          */
3722 
3723         first_time = ctx->ctx_fl_using_dbreg == 0;
3724 
3725         /*
3726          * don't bother if we are loaded and task is being debugged
3727          */
3728         if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3729                 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3730                 return -EBUSY;
3731         }
3732 
3733         /*
3734          * check for debug registers in system wide mode
3735          *
3736          * If though a check is done in pfm_context_load(),
3737          * we must repeat it here, in case the registers are
3738          * written after the context is loaded
3739          */
3740         if (is_loaded) {
3741                 LOCK_PFS(flags);
3742 
3743                 if (first_time && is_system) {
3744                         if (pfm_sessions.pfs_ptrace_use_dbregs)
3745                                 ret = -EBUSY;
3746                         else
3747                                 pfm_sessions.pfs_sys_use_dbregs++;
3748                 }
3749                 UNLOCK_PFS(flags);
3750         }
3751 
3752         if (ret != 0) return ret;
3753 
3754         /*
3755          * mark ourself as user of the debug registers for
3756          * perfmon purposes.
3757          */
3758         ctx->ctx_fl_using_dbreg = 1;
3759 
3760         /*
3761          * clear hardware registers to make sure we don't
3762          * pick up stale state.
3763          *
3764          * for a system wide session, we do not use
3765          * thread.dbr, thread.ibr because this process
3766          * never leaves the current CPU and the state
3767          * is shared by all processes running on it
3768          */
3769         if (first_time && can_access_pmu) {
3770                 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3771                 for (i=0; i < pmu_conf->num_ibrs; i++) {
3772                         ia64_set_ibr(i, 0UL);
3773                         ia64_dv_serialize_instruction();
3774                 }
3775                 ia64_srlz_i();
3776                 for (i=0; i < pmu_conf->num_dbrs; i++) {
3777                         ia64_set_dbr(i, 0UL);
3778                         ia64_dv_serialize_data();
3779                 }
3780                 ia64_srlz_d();
3781         }
3782 
3783         /*
3784          * Now install the values into the registers
3785          */
3786         for (i = 0; i < count; i++, req++) {
3787 
3788                 rnum      = req->dbreg_num;
3789                 dbreg.val = req->dbreg_value;
3790 
3791                 ret = -EINVAL;
3792 
3793                 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3794                         DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3795                                   rnum, dbreg.val, mode, i, count));
3796 
3797                         goto abort_mission;
3798                 }
3799 
3800                 /*
3801                  * make sure we do not install enabled breakpoint
3802                  */
3803                 if (rnum & 0x1) {
3804                         if (mode == PFM_CODE_RR)
3805                                 dbreg.ibr.ibr_x = 0;
3806                         else
3807                                 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3808                 }
3809 
3810                 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3811 
3812                 /*
3813                  * Debug registers, just like PMC, can only be modified
3814                  * by a kernel call. Moreover, perfmon() access to those
3815                  * registers are centralized in this routine. The hardware
3816                  * does not modify the value of these registers, therefore,
3817                  * if we save them as they are written, we can avoid having
3818                  * to save them on context switch out. This is made possible
3819                  * by the fact that when perfmon uses debug registers, ptrace()
3820                  * won't be able to modify them concurrently.
3821                  */
3822                 if (mode == PFM_CODE_RR) {
3823                         CTX_USED_IBR(ctx, rnum);
3824 
3825                         if (can_access_pmu) {
3826                                 ia64_set_ibr(rnum, dbreg.val);
3827                                 ia64_dv_serialize_instruction();
3828                         }
3829 
3830                         ctx->ctx_ibrs[rnum] = dbreg.val;
3831 
3832                         DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3833                                 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3834                 } else {
3835                         CTX_USED_DBR(ctx, rnum);
3836 
3837                         if (can_access_pmu) {
3838                                 ia64_set_dbr(rnum, dbreg.val);
3839                                 ia64_dv_serialize_data();
3840                         }
3841                         ctx->ctx_dbrs[rnum] = dbreg.val;
3842 
3843                         DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3844                                 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3845                 }
3846         }
3847 
3848         return 0;
3849 
3850 abort_mission:
3851         /*
3852          * in case it was our first attempt, we undo the global modifications
3853          */
3854         if (first_time) {
3855                 LOCK_PFS(flags);
3856                 if (ctx->ctx_fl_system) {
3857                         pfm_sessions.pfs_sys_use_dbregs--;
3858                 }
3859                 UNLOCK_PFS(flags);
3860                 ctx->ctx_fl_using_dbreg = 0;
3861         }
3862         /*
3863          * install error return flag
3864          */
3865         PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3866 
3867         return ret;
3868 }
3869 
3870 static int
3871 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3872 {
3873         return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3874 }
3875 
3876 static int
3877 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3878 {
3879         return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3880 }
3881 
3882 int
3883 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3884 {
3885         pfm_context_t *ctx;
3886 
3887         if (req == NULL) return -EINVAL;
3888 
3889         ctx = GET_PMU_CTX();
3890 
3891         if (ctx == NULL) return -EINVAL;
3892 
3893         /*
3894          * for now limit to current task, which is enough when calling
3895          * from overflow handler
3896          */
3897         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3898 
3899         return pfm_write_ibrs(ctx, req, nreq, regs);
3900 }
3901 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3902 
3903 int
3904 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3905 {
3906         pfm_context_t *ctx;
3907 
3908         if (req == NULL) return -EINVAL;
3909 
3910         ctx = GET_PMU_CTX();
3911 
3912         if (ctx == NULL) return -EINVAL;
3913 
3914         /*
3915          * for now limit to current task, which is enough when calling
3916          * from overflow handler
3917          */
3918         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3919 
3920         return pfm_write_dbrs(ctx, req, nreq, regs);
3921 }
3922 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3923 
3924 
3925 static int
3926 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3927 {
3928         pfarg_features_t *req = (pfarg_features_t *)arg;
3929 
3930         req->ft_version = PFM_VERSION;
3931         return 0;
3932 }
3933 
3934 static int
3935 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3936 {
3937         struct pt_regs *tregs;
3938         struct task_struct *task = PFM_CTX_TASK(ctx);
3939         int state, is_system;
3940 
3941         state     = ctx->ctx_state;
3942         is_system = ctx->ctx_fl_system;
3943 
3944         /*
3945          * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3946          */
3947         if (state == PFM_CTX_UNLOADED) return -EINVAL;
3948 
3949         /*
3950          * In system wide and when the context is loaded, access can only happen
3951          * when the caller is running on the CPU being monitored by the session.
3952          * It does not have to be the owner (ctx_task) of the context per se.
3953          */
3954         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3955                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3956                 return -EBUSY;
3957         }
3958         DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3959                 task_pid_nr(PFM_CTX_TASK(ctx)),
3960                 state,
3961                 is_system));
3962         /*
3963          * in system mode, we need to update the PMU directly
3964          * and the user level state of the caller, which may not
3965          * necessarily be the creator of the context.
3966          */
3967         if (is_system) {
3968                 /*
3969                  * Update local PMU first
3970                  *
3971                  * disable dcr pp
3972                  */
3973                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
3974                 ia64_srlz_i();
3975 
3976                 /*
3977                  * update local cpuinfo
3978                  */
3979                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
3980 
3981                 /*
3982                  * stop monitoring, does srlz.i
3983                  */
3984                 pfm_clear_psr_pp();
3985 
3986                 /*
3987                  * stop monitoring in the caller
3988                  */
3989                 ia64_psr(regs)->pp = 0;
3990 
3991                 return 0;
3992         }
3993         /*
3994          * per-task mode
3995          */
3996 
3997         if (task == current) {
3998                 /* stop monitoring  at kernel level */
3999                 pfm_clear_psr_up();
4000 
4001                 /*
4002                  * stop monitoring at the user level
4003                  */
4004                 ia64_psr(regs)->up = 0;
4005         } else {
4006                 tregs = task_pt_regs(task);
4007 
4008                 /*
4009                  * stop monitoring at the user level
4010                  */
4011                 ia64_psr(tregs)->up = 0;
4012 
4013                 /*
4014                  * monitoring disabled in kernel at next reschedule
4015                  */
4016                 ctx->ctx_saved_psr_up = 0;
4017                 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4018         }
4019         return 0;
4020 }
4021 
4022 
4023 static int
4024 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4025 {
4026         struct pt_regs *tregs;
4027         int state, is_system;
4028 
4029         state     = ctx->ctx_state;
4030         is_system = ctx->ctx_fl_system;
4031 
4032         if (state != PFM_CTX_LOADED) return -EINVAL;
4033 
4034         /*
4035          * In system wide and when the context is loaded, access can only happen
4036          * when the caller is running on the CPU being monitored by the session.
4037          * It does not have to be the owner (ctx_task) of the context per se.
4038          */
4039         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4040                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4041                 return -EBUSY;
4042         }
4043 
4044         /*
4045          * in system mode, we need to update the PMU directly
4046          * and the user level state of the caller, which may not
4047          * necessarily be the creator of the context.
4048          */
4049         if (is_system) {
4050 
4051                 /*
4052                  * set user level psr.pp for the caller
4053                  */
4054                 ia64_psr(regs)->pp = 1;
4055 
4056                 /*
4057                  * now update the local PMU and cpuinfo
4058                  */
4059                 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4060 
4061                 /*
4062                  * start monitoring at kernel level
4063                  */
4064                 pfm_set_psr_pp();
4065 
4066                 /* enable dcr pp */
4067                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4068                 ia64_srlz_i();
4069 
4070                 return 0;
4071         }
4072 
4073         /*
4074          * per-process mode
4075          */
4076 
4077         if (ctx->ctx_task == current) {
4078 
4079                 /* start monitoring at kernel level */
4080                 pfm_set_psr_up();
4081 
4082                 /*
4083                  * activate monitoring at user level
4084                  */
4085                 ia64_psr(regs)->up = 1;
4086 
4087         } else {
4088                 tregs = task_pt_regs(ctx->ctx_task);
4089 
4090                 /*
4091                  * start monitoring at the kernel level the next
4092                  * time the task is scheduled
4093                  */
4094                 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4095 
4096                 /*
4097                  * activate monitoring at user level
4098                  */
4099                 ia64_psr(tregs)->up = 1;
4100         }
4101         return 0;
4102 }
4103 
4104 static int
4105 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4106 {
4107         pfarg_reg_t *req = (pfarg_reg_t *)arg;
4108         unsigned int cnum;
4109         int i;
4110         int ret = -EINVAL;
4111 
4112         for (i = 0; i < count; i++, req++) {
4113 
4114                 cnum = req->reg_num;
4115 
4116                 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4117 
4118                 req->reg_value = PMC_DFL_VAL(cnum);
4119 
4120                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4121 
4122                 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4123         }
4124         return 0;
4125 
4126 abort_mission:
4127         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4128         return ret;
4129 }
4130 
4131 static int
4132 pfm_check_task_exist(pfm_context_t *ctx)
4133 {
4134         struct task_struct *g, *t;
4135         int ret = -ESRCH;
4136 
4137         read_lock(&tasklist_lock);
4138 
4139         do_each_thread (g, t) {
4140                 if (t->thread.pfm_context == ctx) {
4141                         ret = 0;
4142                         goto out;
4143                 }
4144         } while_each_thread (g, t);
4145 out:
4146         read_unlock(&tasklist_lock);
4147 
4148         DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4149 
4150         return ret;
4151 }
4152 
4153 static int
4154 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4155 {
4156         struct task_struct *task;
4157         struct thread_struct *thread;
4158         struct pfm_context_t *old;
4159         unsigned long flags;
4160 #ifndef CONFIG_SMP
4161         struct task_struct *owner_task = NULL;
4162 #endif
4163         pfarg_load_t *req = (pfarg_load_t *)arg;
4164         unsigned long *pmcs_source, *pmds_source;
4165         int the_cpu;
4166         int ret = 0;
4167         int state, is_system, set_dbregs = 0;
4168 
4169         state     = ctx->ctx_state;
4170         is_system = ctx->ctx_fl_system;
4171         /*
4172          * can only load from unloaded or terminated state
4173          */
4174         if (state != PFM_CTX_UNLOADED) {
4175                 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4176                         req->load_pid,
4177                         ctx->ctx_state));
4178                 return -EBUSY;
4179         }
4180 
4181         DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4182 
4183         if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4184                 DPRINT(("cannot use blocking mode on self\n"));
4185                 return -EINVAL;
4186         }
4187 
4188         ret = pfm_get_task(ctx, req->load_pid, &task);
4189         if (ret) {
4190                 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4191                 return ret;
4192         }
4193 
4194         ret = -EINVAL;
4195 
4196         /*
4197          * system wide is self monitoring only
4198          */
4199         if (is_system && task != current) {
4200                 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4201                         req->load_pid));
4202                 goto error;
4203         }
4204 
4205         thread = &task->thread;
4206 
4207         ret = 0;
4208         /*
4209          * cannot load a context which is using range restrictions,
4210          * into a task that is being debugged.
4211          */
4212         if (ctx->ctx_fl_using_dbreg) {
4213                 if (thread->flags & IA64_THREAD_DBG_VALID) {
4214                         ret = -EBUSY;
4215                         DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4216                         goto error;
4217                 }
4218                 LOCK_PFS(flags);
4219 
4220                 if (is_system) {
4221                         if (pfm_sessions.pfs_ptrace_use_dbregs) {
4222                                 DPRINT(("cannot load [%d] dbregs in use\n",
4223                                                         task_pid_nr(task)));
4224                                 ret = -EBUSY;
4225                         } else {
4226                                 pfm_sessions.pfs_sys_use_dbregs++;
4227                                 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4228                                 set_dbregs = 1;
4229                         }
4230                 }
4231 
4232                 UNLOCK_PFS(flags);
4233 
4234                 if (ret) goto error;
4235         }
4236 
4237         /*
4238          * SMP system-wide monitoring implies self-monitoring.
4239          *
4240          * The programming model expects the task to
4241          * be pinned on a CPU throughout the session.
4242          * Here we take note of the current CPU at the
4243          * time the context is loaded. No call from
4244          * another CPU will be allowed.
4245          *
4246          * The pinning via shed_setaffinity()
4247          * must be done by the calling task prior
4248          * to this call.
4249          *
4250          * systemwide: keep track of CPU this session is supposed to run on
4251          */
4252         the_cpu = ctx->ctx_cpu = smp_processor_id();
4253 
4254         ret = -EBUSY;
4255         /*
4256          * now reserve the session
4257          */
4258         ret = pfm_reserve_session(current, is_system, the_cpu);
4259         if (ret) goto error;
4260 
4261         /*
4262          * task is necessarily stopped at this point.
4263          *
4264          * If the previous context was zombie, then it got removed in
4265          * pfm_save_regs(). Therefore we should not see it here.
4266          * If we see a context, then this is an active context
4267          *
4268          * XXX: needs to be atomic
4269          */
4270         DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4271                 thread->pfm_context, ctx));
4272 
4273         ret = -EBUSY;
4274         old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4275         if (old != NULL) {
4276                 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4277                 goto error_unres;
4278         }
4279 
4280         pfm_reset_msgq(ctx);
4281 
4282         ctx->ctx_state = PFM_CTX_LOADED;
4283 
4284         /*
4285          * link context to task
4286          */
4287         ctx->ctx_task = task;
4288 
4289         if (is_system) {
4290                 /*
4291                  * we load as stopped
4292                  */
4293                 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4294                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4295 
4296                 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4297         } else {
4298                 thread->flags |= IA64_THREAD_PM_VALID;
4299         }
4300 
4301         /*
4302          * propagate into thread-state
4303          */
4304         pfm_copy_pmds(task, ctx);
4305         pfm_copy_pmcs(task, ctx);
4306 
4307         pmcs_source = ctx->th_pmcs;
4308         pmds_source = ctx->th_pmds;
4309 
4310         /*
4311          * always the case for system-wide
4312          */
4313         if (task == current) {
4314 
4315                 if (is_system == 0) {
4316 
4317                         /* allow user level control */
4318                         ia64_psr(regs)->sp = 0;
4319                         DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4320 
4321                         SET_LAST_CPU(ctx, smp_processor_id());
4322                         INC_ACTIVATION();
4323                         SET_ACTIVATION(ctx);
4324 #ifndef CONFIG_SMP
4325                         /*
4326                          * push the other task out, if any
4327                          */
4328                         owner_task = GET_PMU_OWNER();
4329                         if (owner_task) pfm_lazy_save_regs(owner_task);
4330 #endif
4331                 }
4332                 /*
4333                  * load all PMD from ctx to PMU (as opposed to thread state)
4334                  * restore all PMC from ctx to PMU
4335                  */
4336                 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4337                 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4338 
4339                 ctx->ctx_reload_pmcs[0] = 0UL;
4340                 ctx->ctx_reload_pmds[0] = 0UL;
4341 
4342                 /*
4343                  * guaranteed safe by earlier check against DBG_VALID
4344                  */
4345                 if (ctx->ctx_fl_using_dbreg) {
4346                         pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4347                         pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4348                 }
4349                 /*
4350                  * set new ownership
4351                  */
4352                 SET_PMU_OWNER(task, ctx);
4353 
4354                 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4355         } else {
4356                 /*
4357                  * when not current, task MUST be stopped, so this is safe
4358                  */
4359                 regs = task_pt_regs(task);
4360 
4361                 /* force a full reload */
4362                 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4363                 SET_LAST_CPU(ctx, -1);
4364 
4365                 /* initial saved psr (stopped) */
4366                 ctx->ctx_saved_psr_up = 0UL;
4367                 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4368         }
4369 
4370         ret = 0;
4371 
4372 error_unres:
4373         if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4374 error:
4375         /*
4376          * we must undo the dbregs setting (for system-wide)
4377          */
4378         if (ret && set_dbregs) {
4379                 LOCK_PFS(flags);
4380                 pfm_sessions.pfs_sys_use_dbregs--;
4381                 UNLOCK_PFS(flags);
4382         }
4383         /*
4384          * release task, there is now a link with the context
4385          */
4386         if (is_system == 0 && task != current) {
4387                 pfm_put_task(task);
4388 
4389                 if (ret == 0) {
4390                         ret = pfm_check_task_exist(ctx);
4391                         if (ret) {
4392                                 ctx->ctx_state = PFM_CTX_UNLOADED;
4393                                 ctx->ctx_task  = NULL;
4394                         }
4395                 }
4396         }
4397         return ret;
4398 }
4399 
4400 /*
4401  * in this function, we do not need to increase the use count
4402  * for the task via get_task_struct(), because we hold the
4403  * context lock. If the task were to disappear while having
4404  * a context attached, it would go through pfm_exit_thread()
4405  * which also grabs the context lock  and would therefore be blocked
4406  * until we are here.
4407  */
4408 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4409 
4410 static int
4411 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4412 {
4413         struct task_struct *task = PFM_CTX_TASK(ctx);
4414         struct pt_regs *tregs;
4415         int prev_state, is_system;
4416         int ret;
4417 
4418         DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4419 
4420         prev_state = ctx->ctx_state;
4421         is_system  = ctx->ctx_fl_system;
4422 
4423         /*
4424          * unload only when necessary
4425          */
4426         if (prev_state == PFM_CTX_UNLOADED) {
4427                 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4428                 return 0;
4429         }
4430 
4431         /*
4432          * clear psr and dcr bits
4433          */
4434         ret = pfm_stop(ctx, NULL, 0, regs);
4435         if (ret) return ret;
4436 
4437         ctx->ctx_state = PFM_CTX_UNLOADED;
4438 
4439         /*
4440          * in system mode, we need to update the PMU directly
4441          * and the user level state of the caller, which may not
4442          * necessarily be the creator of the context.
4443          */
4444         if (is_system) {
4445 
4446                 /*
4447                  * Update cpuinfo
4448                  *
4449                  * local PMU is taken care of in pfm_stop()
4450                  */
4451                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4452                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4453 
4454                 /*
4455                  * save PMDs in context
4456                  * release ownership
4457                  */
4458                 pfm_flush_pmds(current, ctx);
4459 
4460                 /*
4461                  * at this point we are done with the PMU
4462                  * so we can unreserve the resource.
4463                  */
4464                 if (prev_state != PFM_CTX_ZOMBIE) 
4465                         pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4466 
4467                 /*
4468                  * disconnect context from task
4469                  */
4470                 task->thread.pfm_context = NULL;
4471                 /*
4472                  * disconnect task from context
4473                  */
4474                 ctx->ctx_task = NULL;
4475 
4476                 /*
4477                  * There is nothing more to cleanup here.
4478                  */
4479                 return 0;
4480         }
4481 
4482         /*
4483          * per-task mode
4484          */
4485         tregs = task == current ? regs : task_pt_regs(task);
4486 
4487         if (task == current) {
4488                 /*
4489                  * cancel user level control
4490                  */
4491                 ia64_psr(regs)->sp = 1;
4492 
4493                 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4494         }
4495         /*
4496          * save PMDs to context
4497          * release ownership
4498          */
4499         pfm_flush_pmds(task, ctx);
4500 
4501         /*
4502          * at this point we are done with the PMU
4503          * so we can unreserve the resource.
4504          *
4505          * when state was ZOMBIE, we have already unreserved.
4506          */
4507         if (prev_state != PFM_CTX_ZOMBIE) 
4508                 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4509 
4510         /*
4511          * reset activation counter and psr
4512          */
4513         ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4514         SET_LAST_CPU(ctx, -1);
4515 
4516         /*
4517          * PMU state will not be restored
4518          */
4519         task->thread.flags &= ~IA64_THREAD_PM_VALID;
4520 
4521         /*
4522          * break links between context and task
4523          */
4524         task->thread.pfm_context  = NULL;
4525         ctx->ctx_task             = NULL;
4526 
4527         PFM_SET_WORK_PENDING(task, 0);
4528 
4529         ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4530         ctx->ctx_fl_can_restart  = 0;
4531         ctx->ctx_fl_going_zombie = 0;
4532 
4533         DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4534 
4535         return 0;
4536 }
4537 
4538 
4539 /*
4540  * called only from exit_thread()
4541  * we come here only if the task has a context attached (loaded or masked)
4542  */
4543 void
4544 pfm_exit_thread(struct task_struct *task)
4545 {
4546         pfm_context_t *ctx;
4547         unsigned long flags;
4548         struct pt_regs *regs = task_pt_regs(task);
4549         int ret, state;
4550         int free_ok = 0;
4551 
4552         ctx = PFM_GET_CTX(task);
4553 
4554         PROTECT_CTX(ctx, flags);
4555 
4556         DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4557 
4558         state = ctx->ctx_state;
4559         switch(state) {
4560                 case PFM_CTX_UNLOADED:
4561                         /*
4562                          * only comes to this function if pfm_context is not NULL, i.e., cannot
4563                          * be in unloaded state
4564                          */
4565                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4566                         break;
4567                 case PFM_CTX_LOADED:
4568                 case PFM_CTX_MASKED:
4569                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4570                         if (ret) {
4571                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4572                         }
4573                         DPRINT(("ctx unloaded for current state was %d\n", state));
4574 
4575                         pfm_end_notify_user(ctx);
4576                         break;
4577                 case PFM_CTX_ZOMBIE:
4578                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4579                         if (ret) {
4580                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4581                         }
4582                         free_ok = 1;
4583                         break;
4584                 default:
4585                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4586                         break;
4587         }
4588         UNPROTECT_CTX(ctx, flags);
4589 
4590         { u64 psr = pfm_get_psr();
4591           BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4592           BUG_ON(GET_PMU_OWNER());
4593           BUG_ON(ia64_psr(regs)->up);
4594           BUG_ON(ia64_psr(regs)->pp);
4595         }
4596 
4597         /*
4598          * All memory free operations (especially for vmalloc'ed memory)
4599          * MUST be done with interrupts ENABLED.
4600          */
4601         if (free_ok) pfm_context_free(ctx);
4602 }
4603 
4604 /*
4605  * functions MUST be listed in the increasing order of their index (see permfon.h)
4606  */
4607 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4608 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4609 #define PFM_CMD_PCLRWS  (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4610 #define PFM_CMD_PCLRW   (PFM_CMD_FD|PFM_CMD_ARG_RW)
4611 #define PFM_CMD_NONE    { NULL, "no-cmd", 0, 0, 0, NULL}
4612 
4613 static pfm_cmd_desc_t pfm_cmd_tab[]={
4614 /* 0  */PFM_CMD_NONE,
4615 /* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4616 /* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4617 /* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4618 /* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4619 /* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4620 /* 6  */PFM_CMD_NONE,
4621 /* 7  */PFM_CMD_NONE,
4622 /* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4623 /* 9  */PFM_CMD_NONE,
4624 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4625 /* 11 */PFM_CMD_NONE,
4626 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4627 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4628 /* 14 */PFM_CMD_NONE,
4629 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4630 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4631 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4632 /* 18 */PFM_CMD_NONE,
4633 /* 19 */PFM_CMD_NONE,
4634 /* 20 */PFM_CMD_NONE,
4635 /* 21 */PFM_CMD_NONE,
4636 /* 22 */PFM_CMD_NONE,
4637 /* 23 */PFM_CMD_NONE,
4638 /* 24 */PFM_CMD_NONE,
4639 /* 25 */PFM_CMD_NONE,
4640 /* 26 */PFM_CMD_NONE,
4641 /* 27 */PFM_CMD_NONE,
4642 /* 28 */PFM_CMD_NONE,
4643 /* 29 */PFM_CMD_NONE,
4644 /* 30 */PFM_CMD_NONE,
4645 /* 31 */PFM_CMD_NONE,
4646 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4647 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4648 };
4649 #define PFM_CMD_COUNT   (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4650 
4651 static int
4652 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4653 {
4654         struct task_struct *task;
4655         int state, old_state;
4656 
4657 recheck:
4658         state = ctx->ctx_state;
4659         task  = ctx->ctx_task;
4660 
4661         if (task == NULL) {
4662                 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4663                 return 0;
4664         }
4665 
4666         DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4667                 ctx->ctx_fd,
4668                 state,
4669                 task_pid_nr(task),
4670                 task->state, PFM_CMD_STOPPED(cmd)));
4671 
4672         /*
4673          * self-monitoring always ok.
4674          *
4675          * for system-wide the caller can either be the creator of the
4676          * context (to one to which the context is attached to) OR
4677          * a task running on the same CPU as the session.
4678          */
4679         if (task == current || ctx->ctx_fl_system) return 0;
4680 
4681         /*
4682          * we are monitoring another thread
4683          */
4684         switch(state) {
4685                 case PFM_CTX_UNLOADED:
4686                         /*
4687                          * if context is UNLOADED we are safe to go
4688                          */
4689                         return 0;
4690                 case PFM_CTX_ZOMBIE:
4691                         /*
4692                          * no command can operate on a zombie context
4693                          */
4694                         DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4695                         return -EINVAL;
4696                 case PFM_CTX_MASKED:
4697                         /*
4698                          * PMU state has been saved to software even though
4699                          * the thread may still be running.
4700                          */
4701                         if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4702         }
4703 
4704         /*
4705          * context is LOADED or MASKED. Some commands may need to have 
4706          * the task stopped.
4707          *
4708          * We could lift this restriction for UP but it would mean that
4709          * the user has no guarantee the task would not run between
4710          * two successive calls to perfmonctl(). That's probably OK.
4711          * If this user wants to ensure the task does not run, then
4712          * the task must be stopped.
4713          */
4714         if (PFM_CMD_STOPPED(cmd)) {
4715                 if (!task_is_stopped_or_traced(task)) {
4716                         DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4717                         return -EBUSY;
4718                 }
4719                 /*
4720                  * task is now stopped, wait for ctxsw out
4721                  *
4722                  * This is an interesting point in the code.
4723                  * We need to unprotect the context because
4724                  * the pfm_save_regs() routines needs to grab
4725                  * the same lock. There are danger in doing
4726                  * this because it leaves a window open for
4727                  * another task to get access to the context
4728                  * and possibly change its state. The one thing
4729                  * that is not possible is for the context to disappear
4730                  * because we are protected by the VFS layer, i.e.,
4731                  * get_fd()/put_fd().
4732                  */
4733                 old_state = state;
4734 
4735                 UNPROTECT_CTX(ctx, flags);
4736 
4737                 wait_task_inactive(task, 0);
4738 
4739                 PROTECT_CTX(ctx, flags);
4740 
4741                 /*
4742                  * we must recheck to verify if state has changed
4743                  */
4744                 if (ctx->ctx_state != old_state) {
4745                         DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4746                         goto recheck;
4747                 }
4748         }
4749         return 0;
4750 }
4751 
4752 /*
4753  * system-call entry point (must return long)
4754  */
4755 asmlinkage long
4756 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4757 {
4758         struct fd f = {NULL, 0};
4759         pfm_context_t *ctx = NULL;
4760         unsigned long flags = 0UL;
4761         void *args_k = NULL;
4762         long ret; /* will expand int return types */
4763         size_t base_sz, sz, xtra_sz = 0;
4764         int narg, completed_args = 0, call_made = 0, cmd_flags;
4765         int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4766         int (*getsize)(void *arg, size_t *sz);
4767 #define PFM_MAX_ARGSIZE 4096
4768 
4769         /*
4770          * reject any call if perfmon was disabled at initialization
4771          */
4772         if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4773 
4774         if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4775                 DPRINT(("invalid cmd=%d\n", cmd));
4776                 return -EINVAL;
4777         }
4778 
4779         func      = pfm_cmd_tab[cmd].cmd_func;
4780         narg      = pfm_cmd_tab[cmd].cmd_narg;
4781         base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4782         getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4783         cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4784 
4785         if (unlikely(func == NULL)) {
4786                 DPRINT(("invalid cmd=%d\n", cmd));
4787                 return -EINVAL;
4788         }
4789 
4790         DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4791                 PFM_CMD_NAME(cmd),
4792                 cmd,
4793                 narg,
4794                 base_sz,
4795                 count));
4796 
4797         /*
4798          * check if number of arguments matches what the command expects
4799          */
4800         if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4801                 return -EINVAL;
4802 
4803 restart_args:
4804         sz = xtra_sz + base_sz*count;
4805         /*
4806          * limit abuse to min page size
4807          */
4808         if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4809                 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4810                 return -E2BIG;
4811         }
4812 
4813         /*
4814          * allocate default-sized argument buffer
4815          */
4816         if (likely(count && args_k == NULL)) {
4817                 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4818                 if (args_k == NULL) return -ENOMEM;
4819         }
4820 
4821         ret = -EFAULT;
4822 
4823         /*
4824          * copy arguments
4825          *
4826          * assume sz = 0 for command without parameters
4827          */
4828         if (sz && copy_from_user(args_k, arg, sz)) {
4829                 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4830                 goto error_args;
4831         }
4832 
4833         /*
4834          * check if command supports extra parameters
4835          */
4836         if (completed_args == 0 && getsize) {
4837                 /*
4838                  * get extra parameters size (based on main argument)
4839                  */
4840                 ret = (*getsize)(args_k, &xtra_sz);
4841                 if (ret) goto error_args;
4842 
4843                 completed_args = 1;
4844 
4845                 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4846 
4847                 /* retry if necessary */
4848                 if (likely(xtra_sz)) goto restart_args;
4849         }
4850 
4851         if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4852 
4853         ret = -EBADF;
4854 
4855         f = fdget(fd);
4856         if (unlikely(f.file == NULL)) {
4857                 DPRINT(("invalid fd %d\n", fd));
4858                 goto error_args;
4859         }
4860         if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4861                 DPRINT(("fd %d not related to perfmon\n", fd));
4862                 goto error_args;
4863         }
4864 
4865         ctx = f.file->private_data;
4866         if (unlikely(ctx == NULL)) {
4867                 DPRINT(("no context for fd %d\n", fd));
4868                 goto error_args;
4869         }
4870         prefetch(&ctx->ctx_state);
4871 
4872         PROTECT_CTX(ctx, flags);
4873 
4874         /*
4875          * check task is stopped
4876          */
4877         ret = pfm_check_task_state(ctx, cmd, flags);
4878         if (unlikely(ret)) goto abort_locked;
4879 
4880 skip_fd:
4881         ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4882 
4883         call_made = 1;
4884 
4885 abort_locked:
4886         if (likely(ctx)) {
4887                 DPRINT(("context unlocked\n"));
4888                 UNPROTECT_CTX(ctx, flags);
4889         }
4890 
4891         /* copy argument back to user, if needed */
4892         if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4893 
4894 error_args:
4895         if (f.file)
4896                 fdput(f);
4897 
4898         kfree(args_k);
4899 
4900         DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4901 
4902         return ret;
4903 }
4904 
4905 static void
4906 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4907 {
4908         pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4909         pfm_ovfl_ctrl_t rst_ctrl;
4910         int state;
4911         int ret = 0;
4912 
4913         state = ctx->ctx_state;
4914         /*
4915          * Unlock sampling buffer and reset index atomically
4916          * XXX: not really needed when blocking
4917          */
4918         if (CTX_HAS_SMPL(ctx)) {
4919 
4920                 rst_ctrl.bits.mask_monitoring = 0;
4921                 rst_ctrl.bits.reset_ovfl_pmds = 0;
4922 
4923                 if (state == PFM_CTX_LOADED)
4924                         ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4925                 else
4926                         ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4927         } else {
4928                 rst_ctrl.bits.mask_monitoring = 0;
4929                 rst_ctrl.bits.reset_ovfl_pmds = 1;
4930         }
4931 
4932         if (ret == 0) {
4933                 if (rst_ctrl.bits.reset_ovfl_pmds) {
4934                         pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4935                 }
4936                 if (rst_ctrl.bits.mask_monitoring == 0) {
4937                         DPRINT(("resuming monitoring\n"));
4938                         if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4939                 } else {
4940                         DPRINT(("stopping monitoring\n"));
4941                         //pfm_stop_monitoring(current, regs);
4942                 }
4943                 ctx->ctx_state = PFM_CTX_LOADED;
4944         }
4945 }
4946 
4947 /*
4948  * context MUST BE LOCKED when calling
4949  * can only be called for current
4950  */
4951 static void
4952 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4953 {
4954         int ret;
4955 
4956         DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4957 
4958         ret = pfm_context_unload(ctx, NULL, 0, regs);
4959         if (ret) {
4960                 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4961         }
4962 
4963         /*
4964          * and wakeup controlling task, indicating we are now disconnected
4965          */
4966         wake_up_interruptible(&ctx->ctx_zombieq);
4967 
4968         /*
4969          * given that context is still locked, the controlling
4970          * task will only get access when we return from
4971          * pfm_handle_work().
4972          */
4973 }
4974 
4975 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4976 
4977  /*
4978   * pfm_handle_work() can be called with interrupts enabled
4979   * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4980   * call may sleep, therefore we must re-enable interrupts
4981   * to avoid deadlocks. It is safe to do so because this function
4982   * is called ONLY when returning to user level (pUStk=1), in which case
4983   * there is no risk of kernel stack overflow due to deep
4984   * interrupt nesting.
4985   */
4986 void
4987 pfm_handle_work(void)
4988 {
4989         pfm_context_t *ctx;
4990         struct pt_regs *regs;
4991         unsigned long flags, dummy_flags;
4992         unsigned long ovfl_regs;
4993         unsigned int reason;
4994         int ret;
4995 
4996         ctx = PFM_GET_CTX(current);
4997         if (ctx == NULL) {
4998                 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
4999                         task_pid_nr(current));
5000                 return;
5001         }
5002 
5003         PROTECT_CTX(ctx, flags);
5004 
5005         PFM_SET_WORK_PENDING(current, 0);
5006 
5007         regs = task_pt_regs(current);
5008 
5009         /*
5010          * extract reason for being here and clear
5011          */
5012         reason = ctx->ctx_fl_trap_reason;
5013         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5014         ovfl_regs = ctx->ctx_ovfl_regs[0];
5015 
5016         DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5017 
5018         /*
5019          * must be done before we check for simple-reset mode
5020          */
5021         if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5022                 goto do_zombie;
5023 
5024         //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5025         if (reason == PFM_TRAP_REASON_RESET)
5026                 goto skip_blocking;
5027 
5028         /*
5029          * restore interrupt mask to what it was on entry.
5030          * Could be enabled/diasbled.
5031          */
5032         UNPROTECT_CTX(ctx, flags);
5033 
5034         /*
5035          * force interrupt enable because of down_interruptible()
5036          */
5037         local_irq_enable();
5038 
5039         DPRINT(("before block sleeping\n"));
5040 
5041         /*
5042          * may go through without blocking on SMP systems
5043          * if restart has been received already by the time we call down()
5044          */
5045         ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5046 
5047         DPRINT(("after block sleeping ret=%d\n", ret));
5048 
5049         /*
5050          * lock context and mask interrupts again
5051          * We save flags into a dummy because we may have
5052          * altered interrupts mask compared to entry in this
5053          * function.
5054          */
5055         PROTECT_CTX(ctx, dummy_flags);
5056 
5057         /*
5058          * we need to read the ovfl_regs only after wake-up
5059          * because we may have had pfm_write_pmds() in between
5060          * and that can changed PMD values and therefore 
5061          * ovfl_regs is reset for these new PMD values.
5062          */
5063         ovfl_regs = ctx->ctx_ovfl_regs[0];
5064 
5065         if (ctx->ctx_fl_going_zombie) {
5066 do_zombie:
5067                 DPRINT(("context is zombie, bailing out\n"));
5068                 pfm_context_force_terminate(ctx, regs);
5069                 goto nothing_to_do;
5070         }
5071         /*
5072          * in case of interruption of down() we don't restart anything
5073          */
5074         if (ret < 0)
5075                 goto nothing_to_do;
5076 
5077 skip_blocking:
5078         pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5079         ctx->ctx_ovfl_regs[0] = 0UL;
5080 
5081 nothing_to_do:
5082         /*
5083          * restore flags as they were upon entry
5084          */
5085         UNPROTECT_CTX(ctx, flags);
5086 }
5087 
5088 static int
5089 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5090 {
5091         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5092                 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5093                 return 0;
5094         }
5095 
5096         DPRINT(("waking up somebody\n"));
5097 
5098         if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5099 
5100         /*
5101          * safe, we are not in intr handler, nor in ctxsw when
5102          * we come here
5103          */
5104         kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5105 
5106         return 0;
5107 }
5108 
5109 static int
5110 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5111 {
5112         pfm_msg_t *msg = NULL;
5113 
5114         if (ctx->ctx_fl_no_msg == 0) {
5115                 msg = pfm_get_new_msg(ctx);
5116                 if (msg == NULL) {
5117                         printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5118                         return -1;
5119                 }
5120 
5121                 msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5122                 msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5123                 msg->pfm_ovfl_msg.msg_active_set   = 0;
5124                 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5125                 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5126                 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5127                 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5128                 msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5129         }
5130 
5131         DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5132                 msg,
5133                 ctx->ctx_fl_no_msg,
5134                 ctx->ctx_fd,
5135                 ovfl_pmds));
5136 
5137         return pfm_notify_user(ctx, msg);
5138 }
5139 
5140 static int
5141 pfm_end_notify_user(pfm_context_t *ctx)
5142 {
5143         pfm_msg_t *msg;
5144 
5145         msg = pfm_get_new_msg(ctx);
5146         if (msg == NULL) {
5147                 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5148                 return -1;
5149         }
5150         /* no leak */
5151         memset(msg, 0, sizeof(*msg));
5152 
5153         msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5154         msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5155         msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5156 
5157         DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5158                 msg,
5159                 ctx->ctx_fl_no_msg,
5160                 ctx->ctx_fd));
5161 
5162         return pfm_notify_user(ctx, msg);
5163 }
5164 
5165 /*
5166  * main overflow processing routine.
5167  * it can be called from the interrupt path or explicitly during the context switch code
5168  */
5169 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5170                                 unsigned long pmc0, struct pt_regs *regs)
5171 {
5172         pfm_ovfl_arg_t *ovfl_arg;
5173         unsigned long mask;
5174         unsigned long old_val, ovfl_val, new_val;
5175         unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5176         unsigned long tstamp;
5177         pfm_ovfl_ctrl_t ovfl_ctrl;
5178         unsigned int i, has_smpl;
5179         int must_notify = 0;
5180 
5181         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5182 
5183         /*
5184          * sanity test. Should never happen
5185          */
5186         if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5187 
5188         tstamp   = ia64_get_itc();
5189         mask     = pmc0 >> PMU_FIRST_COUNTER;
5190         ovfl_val = pmu_conf->ovfl_val;
5191         has_smpl = CTX_HAS_SMPL(ctx);
5192 
5193         DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5194                      "used_pmds=0x%lx\n",
5195                         pmc0,
5196                         task ? task_pid_nr(task): -1,
5197                         (regs ? regs->cr_iip : 0),
5198                         CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5199                         ctx->ctx_used_pmds[0]));
5200 
5201 
5202         /*
5203          * first we update the virtual counters
5204          * assume there was a prior ia64_srlz_d() issued
5205          */
5206         for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5207 
5208                 /* skip pmd which did not overflow */
5209                 if ((mask & 0x1) == 0) continue;
5210 
5211                 /*
5212                  * Note that the pmd is not necessarily 0 at this point as qualified events
5213                  * may have happened before the PMU was frozen. The residual count is not
5214                  * taken into consideration here but will be with any read of the pmd via
5215                  * pfm_read_pmds().
5216                  */
5217                 old_val              = new_val = ctx->ctx_pmds[i].val;
5218                 new_val             += 1 + ovfl_val;
5219                 ctx->ctx_pmds[i].val = new_val;
5220 
5221                 /*
5222                  * check for overflow condition
5223                  */
5224                 if (likely(old_val > new_val)) {
5225                         ovfl_pmds |= 1UL << i;
5226                         if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5227                 }
5228 
5229                 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5230                         i,
5231                         new_val,
5232                         old_val,
5233                         ia64_get_pmd(i) & ovfl_val,
5234                         ovfl_pmds,
5235                         ovfl_notify));
5236         }
5237 
5238         /*
5239          * there was no 64-bit overflow, nothing else to do
5240          */
5241         if (ovfl_pmds == 0UL) return;
5242 
5243         /* 
5244          * reset all control bits
5245          */
5246         ovfl_ctrl.val = 0;
5247         reset_pmds    = 0UL;
5248 
5249         /*
5250          * if a sampling format module exists, then we "cache" the overflow by 
5251          * calling the module's handler() routine.
5252          */
5253         if (has_smpl) {
5254                 unsigned long start_cycles, end_cycles;
5255                 unsigned long pmd_mask;
5256                 int j, k, ret = 0;
5257                 int this_cpu = smp_processor_id();
5258 
5259                 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5260                 ovfl_arg = &ctx->ctx_ovfl_arg;
5261 
5262                 prefetch(ctx->ctx_smpl_hdr);
5263 
5264                 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5265 
5266                         mask = 1UL << i;
5267 
5268                         if ((pmd_mask & 0x1) == 0) continue;
5269 
5270                         ovfl_arg->ovfl_pmd      = (unsigned char )i;
5271                         ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5272                         ovfl_arg->active_set    = 0;
5273                         ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5274                         ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5275 
5276                         ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5277                         ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5278                         ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5279 
5280                         /*
5281                          * copy values of pmds of interest. Sampling format may copy them
5282                          * into sampling buffer.
5283                          */
5284                         if (smpl_pmds) {
5285                                 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5286                                         if ((smpl_pmds & 0x1) == 0) continue;
5287                                         ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5288                                         DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5289                                 }
5290                         }
5291 
5292                         pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5293 
5294                         start_cycles = ia64_get_itc();
5295 
5296                         /*
5297                          * call custom buffer format record (handler) routine
5298                          */
5299                         ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5300 
5301                         end_cycles = ia64_get_itc();
5302 
5303                         /*
5304                          * For those controls, we take the union because they have
5305                          * an all or nothing behavior.
5306                          */
5307                         ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5308                         ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5309                         ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5310                         /*
5311                          * build the bitmask of pmds to reset now
5312                          */
5313                         if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5314 
5315                         pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5316                 }
5317                 /*
5318                  * when the module cannot handle the rest of the overflows, we abort right here
5319                  */
5320                 if (ret && pmd_mask) {
5321                         DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5322                                 pmd_mask<<PMU_FIRST_COUNTER));
5323                 }
5324                 /*
5325                  * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5326                  */
5327                 ovfl_pmds &= ~reset_pmds;
5328         } else {
5329                 /*
5330                  * when no sampling module is used, then the default
5331                  * is to notify on overflow if requested by user
5332                  */
5333                 ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5334                 ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5335                 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5336                 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5337                 /*
5338                  * if needed, we reset all overflowed pmds
5339                  */
5340                 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5341         }
5342 
5343         DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5344 
5345         /*
5346          * reset the requested PMD registers using the short reset values
5347          */
5348         if (reset_pmds) {
5349                 unsigned long bm = reset_pmds;
5350                 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5351         }
5352 
5353         if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5354                 /*
5355                  * keep track of what to reset when unblocking
5356                  */
5357                 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5358 
5359                 /*
5360                  * check for blocking context 
5361                  */
5362                 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5363 
5364                         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5365 
5366                         /*
5367                          * set the perfmon specific checking pending work for the task
5368                          */
5369                         PFM_SET_WORK_PENDING(task, 1);
5370 
5371                         /*
5372                          * when coming from ctxsw, current still points to the
5373                          * previous task, therefore we must work with task and not current.
5374                          */
5375                         set_notify_resume(task);
5376                 }
5377                 /*
5378                  * defer until state is changed (shorten spin window). the context is locked
5379                  * anyway, so the signal receiver would come spin for nothing.
5380                  */
5381                 must_notify = 1;
5382         }
5383 
5384         DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5385                         GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5386                         PFM_GET_WORK_PENDING(task),
5387                         ctx->ctx_fl_trap_reason,
5388                         ovfl_pmds,
5389                         ovfl_notify,
5390                         ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5391         /*
5392          * in case monitoring must be stopped, we toggle the psr bits
5393          */
5394         if (ovfl_ctrl.bits.mask_monitoring) {
5395                 pfm_mask_monitoring(task);
5396                 ctx->ctx_state = PFM_CTX_MASKED;
5397                 ctx->ctx_fl_can_restart = 1;
5398         }
5399 
5400         /*
5401          * send notification now
5402          */
5403         if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5404 
5405         return;
5406 
5407 sanity_check:
5408         printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5409                         smp_processor_id(),
5410                         task ? task_pid_nr(task) : -1,
5411                         pmc0);
5412         return;
5413 
5414 stop_monitoring:
5415         /*
5416          * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5417          * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5418          * come here as zombie only if the task is the current task. In which case, we
5419          * can access the PMU  hardware directly.
5420          *
5421          * Note that zombies do have PM_VALID set. So here we do the minimal.
5422          *
5423          * In case the context was zombified it could not be reclaimed at the time
5424          * the monitoring program exited. At this point, the PMU reservation has been
5425          * returned, the sampiing buffer has been freed. We must convert this call
5426          * into a spurious interrupt. However, we must also avoid infinite overflows
5427          * by stopping monitoring for this task. We can only come here for a per-task
5428          * context. All we need to do is to stop monitoring using the psr bits which
5429          * are always task private. By re-enabling secure montioring, we ensure that
5430          * the monitored task will not be able to re-activate monitoring.
5431          * The task will eventually be context switched out, at which point the context
5432          * will be reclaimed (that includes releasing ownership of the PMU).
5433          *
5434          * So there might be a window of time where the number of per-task session is zero
5435          * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5436          * context. This is safe because if a per-task session comes in, it will push this one
5437          * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5438          * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5439          * also push our zombie context out.
5440          *
5441          * Overall pretty hairy stuff....
5442          */
5443         DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5444         pfm_clear_psr_up();
5445         ia64_psr(regs)->up = 0;
5446         ia64_psr(regs)->sp = 1;
5447         return;
5448 }
5449 
5450 static int
5451 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5452 {
5453         struct task_struct *task;
5454         pfm_context_t *ctx;
5455         unsigned long flags;
5456         u64 pmc0;
5457         int this_cpu = smp_processor_id();
5458         int retval = 0;
5459 
5460         pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5461 
5462         /*
5463          * srlz.d done before arriving here
5464          */
5465         pmc0 = ia64_get_pmc(0);
5466 
5467         task = GET_PMU_OWNER();
5468         ctx  = GET_PMU_CTX();
5469 
5470         /*
5471          * if we have some pending bits set
5472          * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5473          */
5474         if (PMC0_HAS_OVFL(pmc0) && task) {
5475                 /*
5476                  * we assume that pmc0.fr is always set here
5477                  */
5478 
5479                 /* sanity check */
5480                 if (!ctx) goto report_spurious1;
5481 
5482                 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5483                         goto report_spurious2;
5484 
5485                 PROTECT_CTX_NOPRINT(ctx, flags);
5486 
5487                 pfm_overflow_handler(task, ctx, pmc0, regs);
5488 
5489                 UNPROTECT_CTX_NOPRINT(ctx, flags);
5490 
5491         } else {
5492                 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5493                 retval = -1;
5494         }
5495         /*
5496          * keep it unfrozen at all times
5497          */
5498         pfm_unfreeze_pmu();
5499 
5500         return retval;
5501 
5502 report_spurious1:
5503         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5504                 this_cpu, task_pid_nr(task));
5505         pfm_unfreeze_pmu();
5506         return -1;
5507 report_spurious2:
5508         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5509                 this_cpu, 
5510                 task_pid_nr(task));
5511         pfm_unfreeze_pmu();
5512         return -1;
5513 }
5514 
5515 static irqreturn_t
5516 pfm_interrupt_handler(int irq, void *arg)
5517 {
5518         unsigned long start_cycles, total_cycles;
5519         unsigned long min, max;
5520         int this_cpu;
5521         int ret;
5522         struct pt_regs *regs = get_irq_regs();
5523 
5524         this_cpu = get_cpu();
5525         if (likely(!pfm_alt_intr_handler)) {
5526                 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5527                 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5528 
5529                 start_cycles = ia64_get_itc();
5530 
5531                 ret = pfm_do_interrupt_handler(arg, regs);
5532 
5533                 total_cycles = ia64_get_itc();
5534 
5535                 /*
5536                  * don't measure spurious interrupts
5537                  */
5538                 if (likely(ret == 0)) {
5539                         total_cycles -= start_cycles;
5540 
5541                         if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5542                         if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5543 
5544                         pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5545                 }
5546         }
5547         else {
5548                 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5549         }
5550 
5551         put_cpu();
5552         return IRQ_HANDLED;
5553 }
5554 
5555 /*
5556  * /proc/perfmon interface, for debug only
5557  */
5558 
5559 #define PFM_PROC_SHOW_HEADER    ((void *)(long)nr_cpu_ids+1)
5560 
5561 static void *
5562 pfm_proc_start(struct seq_file *m, loff_t *pos)
5563 {
5564         if (*pos == 0) {
5565                 return PFM_PROC_SHOW_HEADER;
5566         }
5567 
5568         while (*pos <= nr_cpu_ids) {
5569                 if (cpu_online(*pos - 1)) {
5570                         return (void *)*pos;
5571                 }
5572                 ++*pos;
5573         }
5574         return NULL;
5575 }
5576 
5577 static void *
5578 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5579 {
5580         ++*pos;
5581         return pfm_proc_start(m, pos);
5582 }
5583 
5584 static void
5585 pfm_proc_stop(struct seq_file *m, void *v)
5586 {
5587 }
5588 
5589 static void
5590 pfm_proc_show_header(struct seq_file *m)
5591 {
5592         struct list_head * pos;
5593         pfm_buffer_fmt_t * entry;
5594         unsigned long flags;
5595 
5596         seq_printf(m,
5597                 "perfmon version           : %u.%u\n"
5598                 "model                     : %s\n"
5599                 "fastctxsw                 : %s\n"
5600                 "expert mode               : %s\n"
5601                 "ovfl_mask                 : 0x%lx\n"
5602                 "PMU flags                 : 0x%x\n",
5603                 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5604                 pmu_conf->pmu_name,
5605                 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5606                 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5607                 pmu_conf->ovfl_val,
5608                 pmu_conf->flags);
5609 
5610         LOCK_PFS(flags);
5611 
5612         seq_printf(m,
5613                 "proc_sessions             : %u\n"
5614                 "sys_sessions              : %u\n"
5615                 "sys_use_dbregs            : %u\n"
5616                 "ptrace_use_dbregs         : %u\n",
5617                 pfm_sessions.pfs_task_sessions,
5618                 pfm_sessions.pfs_sys_sessions,
5619                 pfm_sessions.pfs_sys_use_dbregs,
5620                 pfm_sessions.pfs_ptrace_use_dbregs);
5621 
5622         UNLOCK_PFS(flags);
5623 
5624         spin_lock(&pfm_buffer_fmt_lock);
5625 
5626         list_for_each(pos, &pfm_buffer_fmt_list) {
5627                 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5628                 seq_printf(m, "format                    : %16phD %s\n",
5629                            entry->fmt_uuid, entry->fmt_name);
5630         }
5631         spin_unlock(&pfm_buffer_fmt_lock);
5632 
5633 }
5634 
5635 static int
5636 pfm_proc_show(struct seq_file *m, void *v)
5637 {
5638         unsigned long psr;
5639         unsigned int i;
5640         int cpu;
5641 
5642         if (v == PFM_PROC_SHOW_HEADER) {
5643                 pfm_proc_show_header(m);
5644                 return 0;
5645         }
5646 
5647         /* show info for CPU (v - 1) */
5648 
5649         cpu = (long)v - 1;
5650         seq_printf(m,
5651                 "CPU%-2d overflow intrs      : %lu\n"
5652                 "CPU%-2d overflow cycles     : %lu\n"
5653                 "CPU%-2d overflow min        : %lu\n"
5654                 "CPU%-2d overflow max        : %lu\n"
5655                 "CPU%-2d smpl handler calls  : %lu\n"
5656                 "CPU%-2d smpl handler cycles : %lu\n"
5657                 "CPU%-2d spurious intrs      : %lu\n"
5658                 "CPU%-2d replay   intrs      : %lu\n"
5659                 "CPU%-2d syst_wide           : %d\n"
5660                 "CPU%-2d dcr_pp              : %d\n"
5661                 "CPU%-2d exclude idle        : %d\n"
5662                 "CPU%-2d owner               : %d\n"
5663                 "CPU%-2d context             : %p\n"
5664                 "CPU%-2d activations         : %lu\n",
5665                 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5666                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5667                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5668                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5669                 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5670                 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5671                 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5672                 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5673                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5674                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5675                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5676                 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5677                 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5678                 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5679 
5680         if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5681 
5682                 psr = pfm_get_psr();
5683 
5684                 ia64_srlz_d();
5685 
5686                 seq_printf(m, 
5687                         "CPU%-2d psr                 : 0x%lx\n"
5688                         "CPU%-2d pmc0                : 0x%lx\n", 
5689                         cpu, psr,
5690                         cpu, ia64_get_pmc(0));
5691 
5692                 for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5693                         if (PMC_IS_COUNTING(i) == 0) continue;
5694                         seq_printf(m, 
5695                                 "CPU%-2d pmc%u                : 0x%lx\n"
5696                                 "CPU%-2d pmd%u                : 0x%lx\n", 
5697                                 cpu, i, ia64_get_pmc(i),
5698                                 cpu, i, ia64_get_pmd(i));
5699                 }
5700         }
5701         return 0;
5702 }
5703 
5704 const struct seq_operations pfm_seq_ops = {
5705         .start =        pfm_proc_start,
5706         .next =         pfm_proc_next,
5707         .stop =         pfm_proc_stop,
5708         .show =         pfm_proc_show
5709 };
5710 
5711 /*
5712  * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5713  * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5714  * is active or inactive based on mode. We must rely on the value in
5715  * local_cpu_data->pfm_syst_info
5716  */
5717 void
5718 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5719 {
5720         struct pt_regs *regs;
5721         unsigned long dcr;
5722         unsigned long dcr_pp;
5723 
5724         dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5725 
5726         /*
5727          * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5728          * on every CPU, so we can rely on the pid to identify the idle task.
5729          */
5730         if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5731                 regs = task_pt_regs(task);
5732                 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5733                 return;
5734         }
5735         /*
5736          * if monitoring has started
5737          */
5738         if (dcr_pp) {
5739                 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5740                 /*
5741                  * context switching in?
5742                  */
5743                 if (is_ctxswin) {
5744                         /* mask monitoring for the idle task */
5745                         ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5746                         pfm_clear_psr_pp();
5747                         ia64_srlz_i();
5748                         return;
5749                 }
5750                 /*
5751                  * context switching out
5752                  * restore monitoring for next task
5753                  *
5754                  * Due to inlining this odd if-then-else construction generates
5755                  * better code.
5756                  */
5757                 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5758                 pfm_set_psr_pp();
5759                 ia64_srlz_i();
5760         }
5761 }
5762 
5763 #ifdef CONFIG_SMP
5764 
5765 static void
5766 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5767 {
5768         struct task_struct *task = ctx->ctx_task;
5769 
5770         ia64_psr(regs)->up = 0;
5771         ia64_psr(regs)->sp = 1;
5772 
5773         if (GET_PMU_OWNER() == task) {
5774                 DPRINT(("cleared ownership for [%d]\n",
5775                                         task_pid_nr(ctx->ctx_task)));
5776                 SET_PMU_OWNER(NULL, NULL);
5777         }
5778 
5779         /*
5780          * disconnect the task from the context and vice-versa
5781          */
5782         PFM_SET_WORK_PENDING(task, 0);
5783 
5784         task->thread.pfm_context  = NULL;
5785         task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5786 
5787         DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5788 }
5789 
5790 
5791 /*
5792  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5793  */
5794 void
5795 pfm_save_regs(struct task_struct *task)
5796 {
5797         pfm_context_t *ctx;
5798         unsigned long flags;
5799         u64 psr;
5800 
5801 
5802         ctx = PFM_GET_CTX(task);
5803         if (ctx == NULL) return;
5804 
5805         /*
5806          * we always come here with interrupts ALREADY disabled by
5807          * the scheduler. So we simply need to protect against concurrent
5808          * access, not CPU concurrency.
5809          */
5810         flags = pfm_protect_ctx_ctxsw(ctx);
5811 
5812         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5813                 struct pt_regs *regs = task_pt_regs(task);
5814 
5815                 pfm_clear_psr_up();
5816 
5817                 pfm_force_cleanup(ctx, regs);
5818 
5819                 BUG_ON(ctx->ctx_smpl_hdr);
5820 
5821                 pfm_unprotect_ctx_ctxsw(ctx, flags);
5822 
5823                 pfm_context_free(ctx);
5824                 return;
5825         }
5826 
5827         /*
5828          * save current PSR: needed because we modify it
5829          */
5830         ia64_srlz_d();
5831         psr = pfm_get_psr();
5832 
5833         BUG_ON(psr & (IA64_PSR_I));
5834 
5835         /*
5836          * stop monitoring:
5837          * This is the last instruction which may generate an overflow
5838          *
5839          * We do not need to set psr.sp because, it is irrelevant in kernel.
5840          * It will be restored from ipsr when going back to user level
5841          */
5842         pfm_clear_psr_up();
5843 
5844         /*
5845          * keep a copy of psr.up (for reload)
5846          */
5847         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5848 
5849         /*
5850          * release ownership of this PMU.
5851          * PM interrupts are masked, so nothing
5852          * can happen.
5853          */
5854         SET_PMU_OWNER(NULL, NULL);
5855 
5856         /*
5857          * we systematically save the PMD as we have no
5858          * guarantee we will be schedule at that same
5859          * CPU again.
5860          */
5861         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5862 
5863         /*
5864          * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5865          * we will need it on the restore path to check
5866          * for pending overflow.
5867          */
5868         ctx->th_pmcs[0] = ia64_get_pmc(0);
5869 
5870         /*
5871          * unfreeze PMU if had pending overflows
5872          */
5873         if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5874 
5875         /*
5876          * finally, allow context access.
5877          * interrupts will still be masked after this call.
5878          */
5879         pfm_unprotect_ctx_ctxsw(ctx, flags);
5880 }
5881 
5882 #else /* !CONFIG_SMP */
5883 void
5884 pfm_save_regs(struct task_struct *task)
5885 {
5886         pfm_context_t *ctx;
5887         u64 psr;
5888 
5889         ctx = PFM_GET_CTX(task);
5890         if (ctx == NULL) return;
5891 
5892         /*
5893          * save current PSR: needed because we modify it
5894          */
5895         psr = pfm_get_psr();
5896 
5897         BUG_ON(psr & (IA64_PSR_I));
5898 
5899         /*
5900          * stop monitoring:
5901          * This is the last instruction which may generate an overflow
5902          *
5903          * We do not need to set psr.sp because, it is irrelevant in kernel.
5904          * It will be restored from ipsr when going back to user level
5905          */
5906         pfm_clear_psr_up();
5907 
5908         /*
5909          * keep a copy of psr.up (for reload)
5910          */
5911         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5912 }
5913 
5914 static void
5915 pfm_lazy_save_regs (struct task_struct *task)
5916 {
5917         pfm_context_t *ctx;
5918         unsigned long flags;
5919 
5920         { u64 psr  = pfm_get_psr();
5921           BUG_ON(psr & IA64_PSR_UP);
5922         }
5923 
5924         ctx = PFM_GET_CTX(task);
5925 
5926         /*
5927          * we need to mask PMU overflow here to
5928          * make sure that we maintain pmc0 until
5929          * we save it. overflow interrupts are
5930          * treated as spurious if there is no
5931          * owner.
5932          *
5933          * XXX: I don't think this is necessary
5934          */
5935         PROTECT_CTX(ctx,flags);
5936 
5937         /*
5938          * release ownership of this PMU.
5939          * must be done before we save the registers.
5940          *
5941          * after this call any PMU interrupt is treated
5942          * as spurious.
5943          */
5944         SET_PMU_OWNER(NULL, NULL);
5945 
5946         /*
5947          * save all the pmds we use
5948          */
5949         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5950 
5951         /*
5952          * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5953          * it is needed to check for pended overflow
5954          * on the restore path
5955          */
5956         ctx->th_pmcs[0] = ia64_get_pmc(0);
5957 
5958         /*
5959          * unfreeze PMU if had pending overflows
5960          */
5961         if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5962 
5963         /*
5964          * now get can unmask PMU interrupts, they will
5965          * be treated as purely spurious and we will not
5966          * lose any information
5967          */
5968         UNPROTECT_CTX(ctx,flags);
5969 }
5970 #endif /* CONFIG_SMP */
5971 
5972 #ifdef CONFIG_SMP
5973 /*
5974  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5975  */
5976 void
5977 pfm_load_regs (struct task_struct *task)
5978 {
5979         pfm_context_t *ctx;
5980         unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
5981         unsigned long flags;
5982         u64 psr, psr_up;
5983         int need_irq_resend;
5984 
5985         ctx = PFM_GET_CTX(task);
5986         if (unlikely(ctx == NULL)) return;
5987 
5988         BUG_ON(GET_PMU_OWNER());
5989 
5990         /*
5991          * possible on unload
5992          */
5993         if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
5994 
5995         /*
5996          * we always come here with interrupts ALREADY disabled by
5997          * the scheduler. So we simply need to protect against concurrent
5998          * access, not CPU concurrency.
5999          */
6000         flags = pfm_protect_ctx_ctxsw(ctx);
6001         psr   = pfm_get_psr();
6002 
6003         need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6004 
6005         BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6006         BUG_ON(psr & IA64_PSR_I);
6007 
6008         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6009                 struct pt_regs *regs = task_pt_regs(task);
6010 
6011                 BUG_ON(ctx->ctx_smpl_hdr);
6012 
6013                 pfm_force_cleanup(ctx, regs);
6014 
6015                 pfm_unprotect_ctx_ctxsw(ctx, flags);
6016 
6017                 /*
6018                  * this one (kmalloc'ed) is fine with interrupts disabled
6019                  */
6020                 pfm_context_free(ctx);
6021 
6022                 return;
6023         }
6024 
6025         /*
6026          * we restore ALL the debug registers to avoid picking up
6027          * stale state.
6028          */
6029         if (ctx->ctx_fl_using_dbreg) {
6030                 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6031                 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6032         }
6033         /*
6034          * retrieve saved psr.up
6035          */
6036         psr_up = ctx->ctx_saved_psr_up;
6037 
6038         /*
6039          * if we were the last user of the PMU on that CPU,
6040          * then nothing to do except restore psr
6041          */
6042         if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6043 
6044                 /*
6045                  * retrieve partial reload masks (due to user modifications)
6046                  */
6047                 pmc_mask = ctx->ctx_reload_pmcs[0];
6048                 pmd_mask = ctx->ctx_reload_pmds[0];
6049 
6050         } else {
6051                 /*
6052                  * To avoid leaking information to the user level when psr.sp=0,
6053                  * we must reload ALL implemented pmds (even the ones we don't use).
6054                  * In the kernel we only allow PFM_READ_PMDS on registers which
6055                  * we initialized or requested (sampling) so there is no risk there.
6056                  */
6057                 pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6058 
6059                 /*
6060                  * ALL accessible PMCs are systematically reloaded, unused registers
6061                  * get their default (from pfm_reset_pmu_state()) values to avoid picking
6062                  * up stale configuration.
6063                  *
6064                  * PMC0 is never in the mask. It is always restored separately.
6065                  */
6066                 pmc_mask = ctx->ctx_all_pmcs[0];
6067         }
6068         /*
6069          * when context is MASKED, we will restore PMC with plm=0
6070          * and PMD with stale information, but that's ok, nothing
6071          * will be captured.
6072          *
6073          * XXX: optimize here
6074          */
6075         if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6076         if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6077 
6078         /*
6079          * check for pending overflow at the time the state
6080          * was saved.
6081          */
6082         if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6083                 /*
6084                  * reload pmc0 with the overflow information
6085                  * On McKinley PMU, this will trigger a PMU interrupt
6086                  */
6087                 ia64_set_pmc(0, ctx->th_pmcs[0]);
6088                 ia64_srlz_d();
6089                 ctx->th_pmcs[0] = 0UL;
6090 
6091                 /*
6092                  * will replay the PMU interrupt
6093                  */
6094                 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6095 
6096                 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6097         }
6098 
6099         /*
6100          * we just did a reload, so we reset the partial reload fields
6101          */
6102         ctx->ctx_reload_pmcs[0] = 0UL;
6103         ctx->ctx_reload_pmds[0] = 0UL;
6104 
6105         SET_LAST_CPU(ctx, smp_processor_id());
6106 
6107         /*
6108          * dump activation value for this PMU
6109          */
6110         INC_ACTIVATION();
6111         /*
6112          * record current activation for this context
6113          */
6114         SET_ACTIVATION(ctx);
6115 
6116         /*
6117          * establish new ownership. 
6118          */
6119         SET_PMU_OWNER(task, ctx);
6120 
6121         /*
6122          * restore the psr.up bit. measurement
6123          * is active again.
6124          * no PMU interrupt can happen at this point
6125          * because we still have interrupts disabled.
6126          */
6127         if (likely(psr_up)) pfm_set_psr_up();
6128 
6129         /*
6130          * allow concurrent access to context
6131          */
6132         pfm_unprotect_ctx_ctxsw(ctx, flags);
6133 }
6134 #else /*  !CONFIG_SMP */
6135 /*
6136  * reload PMU state for UP kernels
6137  * in 2.5 we come here with interrupts disabled
6138  */
6139 void
6140 pfm_load_regs (struct task_struct *task)
6141 {
6142         pfm_context_t *ctx;
6143         struct task_struct *owner;
6144         unsigned long pmd_mask, pmc_mask;
6145         u64 psr, psr_up;
6146         int need_irq_resend;
6147 
6148         owner = GET_PMU_OWNER();
6149         ctx   = PFM_GET_CTX(task);
6150         psr   = pfm_get_psr();
6151 
6152         BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6153         BUG_ON(psr & IA64_PSR_I);
6154 
6155         /*
6156          * we restore ALL the debug registers to avoid picking up
6157          * stale state.
6158          *
6159          * This must be done even when the task is still the owner
6160          * as the registers may have been modified via ptrace()
6161          * (not perfmon) by the previous task.
6162          */
6163         if (ctx->ctx_fl_using_dbreg) {
6164                 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6165                 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6166         }
6167 
6168         /*
6169          * retrieved saved psr.up
6170          */
6171         psr_up = ctx->ctx_saved_psr_up;
6172         need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6173 
6174         /*
6175          * short path, our state is still there, just
6176          * need to restore psr and we go
6177          *
6178          * we do not touch either PMC nor PMD. the psr is not touched
6179          * by the overflow_handler. So we are safe w.r.t. to interrupt
6180          * concurrency even without interrupt masking.
6181          */
6182         if (likely(owner == task)) {
6183                 if (likely(psr_up)) pfm_set_psr_up();
6184                 return;
6185         }
6186 
6187         /*
6188          * someone else is still using the PMU, first push it out and
6189          * then we'll be able to install our stuff !
6190          *
6191          * Upon return, there will be no owner for the current PMU
6192          */
6193         if (owner) pfm_lazy_save_regs(owner);
6194 
6195         /*
6196          * To avoid leaking information to the user level when psr.sp=0,
6197          * we must reload ALL implemented pmds (even the ones we don't use).
6198          * In the kernel we only allow PFM_READ_PMDS on registers which
6199          * we initialized or requested (sampling) so there is no risk there.
6200          */
6201         pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6202 
6203         /*
6204          * ALL accessible PMCs are systematically reloaded, unused registers
6205          * get their default (from pfm_reset_pmu_state()) values to avoid picking
6206          * up stale configuration.
6207          *
6208          * PMC0 is never in the mask. It is always restored separately
6209          */
6210         pmc_mask = ctx->ctx_all_pmcs[0];
6211 
6212         pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6213         pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6214 
6215         /*
6216          * check for pending overflow at the time the state
6217          * was saved.
6218          */
6219         if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6220                 /*
6221                  * reload pmc0 with the overflow information
6222                  * On McKinley PMU, this will trigger a PMU interrupt
6223                  */
6224                 ia64_set_pmc(0, ctx->th_pmcs[0]);
6225                 ia64_srlz_d();
6226 
6227                 ctx->th_pmcs[0] = 0UL;
6228 
6229                 /*
6230                  * will replay the PMU interrupt
6231                  */
6232                 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6233 
6234                 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6235         }
6236 
6237         /*
6238          * establish new ownership. 
6239          */
6240         SET_PMU_OWNER(task, ctx);
6241 
6242         /*
6243          * restore the psr.up bit. measurement
6244          * is active again.
6245          * no PMU interrupt can happen at this point
6246          * because we still have interrupts disabled.
6247          */
6248         if (likely(psr_up)) pfm_set_psr_up();
6249 }
6250 #endif /* CONFIG_SMP */
6251 
6252 /*
6253  * this function assumes monitoring is stopped
6254  */
6255 static void
6256 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6257 {
6258         u64 pmc0;
6259         unsigned long mask2, val, pmd_val, ovfl_val;
6260         int i, can_access_pmu = 0;
6261         int is_self;
6262 
6263         /*
6264          * is the caller the task being monitored (or which initiated the
6265          * session for system wide measurements)
6266          */
6267         is_self = ctx->ctx_task == task ? 1 : 0;
6268 
6269         /*
6270          * can access PMU is task is the owner of the PMU state on the current CPU
6271          * or if we are running on the CPU bound to the context in system-wide mode
6272          * (that is not necessarily the task the context is attached to in this mode).
6273          * In system-wide we always have can_access_pmu true because a task running on an
6274          * invalid processor is flagged earlier in the call stack (see pfm_stop).
6275          */
6276         can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6277         if (can_access_pmu) {
6278                 /*
6279                  * Mark the PMU as not owned
6280                  * This will cause the interrupt handler to do nothing in case an overflow
6281                  * interrupt was in-flight
6282                  * This also guarantees that pmc0 will contain the final state
6283                  * It virtually gives us full control on overflow processing from that point
6284                  * on.
6285                  */
6286                 SET_PMU_OWNER(NULL, NULL);
6287                 DPRINT(("releasing ownership\n"));
6288 
6289                 /*
6290                  * read current overflow status:
6291                  *
6292                  * we are guaranteed to read the final stable state
6293                  */
6294                 ia64_srlz_d();
6295                 pmc0 = ia64_get_pmc(0); /* slow */
6296 
6297                 /*
6298                  * reset freeze bit, overflow status information destroyed
6299                  */
6300                 pfm_unfreeze_pmu();
6301         } else {
6302                 pmc0 = ctx->th_pmcs[0];
6303                 /*
6304                  * clear whatever overflow status bits there were
6305                  */
6306                 ctx->th_pmcs[0] = 0;
6307         }
6308         ovfl_val = pmu_conf->ovfl_val;
6309         /*
6310          * we save all the used pmds
6311          * we take care of overflows for counting PMDs
6312          *
6313          * XXX: sampling situation is not taken into account here
6314          */
6315         mask2 = ctx->ctx_used_pmds[0];
6316 
6317         DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6318 
6319         for (i = 0; mask2; i++, mask2>>=1) {
6320 
6321                 /* skip non used pmds */
6322                 if ((mask2 & 0x1) == 0) continue;
6323 
6324                 /*
6325                  * can access PMU always true in system wide mode
6326                  */
6327                 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6328 
6329                 if (PMD_IS_COUNTING(i)) {
6330                         DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6331                                 task_pid_nr(task),
6332                                 i,
6333                                 ctx->ctx_pmds[i].val,
6334                                 val & ovfl_val));
6335 
6336                         /*
6337                          * we rebuild the full 64 bit value of the counter
6338                          */
6339                         val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6340 
6341                         /*
6342                          * now everything is in ctx_pmds[] and we need
6343                          * to clear the saved context from save_regs() such that
6344                          * pfm_read_pmds() gets the correct value
6345                          */
6346                         pmd_val = 0UL;
6347 
6348                         /*
6349                          * take care of overflow inline
6350                          */
6351                         if (pmc0 & (1UL << i)) {
6352                                 val += 1 + ovfl_val;
6353                                 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6354                         }
6355                 }
6356 
6357                 DPRINT(("[%d] ctx_pmd[%d]=0x%lx  pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6358 
6359                 if (is_self) ctx->th_pmds[i] = pmd_val;
6360 
6361                 ctx->ctx_pmds[i].val = val;
6362         }
6363 }
6364 
6365 static struct irqaction perfmon_irqaction = {
6366         .handler = pfm_interrupt_handler,
6367         .name    = "perfmon"
6368 };
6369 
6370 static void
6371 pfm_alt_save_pmu_state(void *data)
6372 {
6373         struct pt_regs *regs;
6374 
6375         regs = task_pt_regs(current);
6376 
6377         DPRINT(("called\n"));
6378 
6379         /*
6380          * should not be necessary but
6381          * let's take not risk
6382          */
6383         pfm_clear_psr_up();
6384         pfm_clear_psr_pp();
6385         ia64_psr(regs)->pp = 0;
6386 
6387         /*
6388          * This call is required
6389          * May cause a spurious interrupt on some processors
6390          */
6391         pfm_freeze_pmu();
6392 
6393         ia64_srlz_d();
6394 }
6395 
6396 void
6397 pfm_alt_restore_pmu_state(void *data)
6398 {
6399         struct pt_regs *regs;
6400 
6401         regs = task_pt_regs(current);
6402 
6403         DPRINT(("called\n"));
6404 
6405         /*
6406          * put PMU back in state expected
6407          * by perfmon
6408          */
6409         pfm_clear_psr_up();
6410         pfm_clear_psr_pp();
6411         ia64_psr(regs)->pp = 0;
6412 
6413         /*
6414          * perfmon runs with PMU unfrozen at all times
6415          */
6416         pfm_unfreeze_pmu();
6417 
6418         ia64_srlz_d();
6419 }
6420 
6421 int
6422 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6423 {
6424         int ret, i;
6425         int reserve_cpu;
6426 
6427         /* some sanity checks */
6428         if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6429 
6430         /* do the easy test first */
6431         if (pfm_alt_intr_handler) return -EBUSY;
6432 
6433         /* one at a time in the install or remove, just fail the others */
6434         if (!spin_trylock(&pfm_alt_install_check)) {
6435                 return -EBUSY;
6436         }
6437 
6438         /* reserve our session */
6439         for_each_online_cpu(reserve_cpu) {
6440                 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6441                 if (ret) goto cleanup_reserve;
6442         }
6443 
6444         /* save the current system wide pmu states */
6445         ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6446         if (ret) {
6447                 DPRINT(("on_each_cpu() failed: %d\n", ret));
6448                 goto cleanup_reserve;
6449         }
6450 
6451         /* officially change to the alternate interrupt handler */
6452         pfm_alt_intr_handler = hdl;
6453 
6454         spin_unlock(&pfm_alt_install_check);
6455 
6456         return 0;
6457 
6458 cleanup_reserve:
6459         for_each_online_cpu(i) {
6460                 /* don't unreserve more than we reserved */
6461                 if (i >= reserve_cpu) break;
6462 
6463                 pfm_unreserve_session(NULL, 1, i);
6464         }
6465 
6466         spin_unlock(&pfm_alt_install_check);
6467 
6468         return ret;
6469 }
6470 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6471 
6472 int
6473 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6474 {
6475         int i;
6476         int ret;
6477 
6478         if (hdl == NULL) return -EINVAL;
6479 
6480         /* cannot remove someone else's handler! */
6481         if (pfm_alt_intr_handler != hdl) return -EINVAL;
6482 
6483         /* one at a time in the install or remove, just fail the others */
6484         if (!spin_trylock(&pfm_alt_install_check)) {
6485                 return -EBUSY;
6486         }
6487 
6488         pfm_alt_intr_handler = NULL;
6489 
6490         ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6491         if (ret) {
6492                 DPRINT(("on_each_cpu() failed: %d\n", ret));
6493         }
6494 
6495         for_each_online_cpu(i) {
6496                 pfm_unreserve_session(NULL, 1, i);
6497         }
6498 
6499         spin_unlock(&pfm_alt_install_check);
6500 
6501         return 0;
6502 }
6503 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6504 
6505 /*
6506  * perfmon initialization routine, called from the initcall() table
6507  */
6508 static int init_pfm_fs(void);
6509 
6510 static int __init
6511 pfm_probe_pmu(void)
6512 {
6513         pmu_config_t **p;
6514         int family;
6515 
6516         family = local_cpu_data->family;
6517         p      = pmu_confs;
6518 
6519         while(*p) {
6520                 if ((*p)->probe) {
6521                         if ((*p)->probe() == 0) goto found;
6522                 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6523                         goto found;
6524                 }
6525                 p++;
6526         }
6527         return -1;
6528 found:
6529         pmu_conf = *p;
6530         return 0;
6531 }
6532 
6533 int __init
6534 pfm_init(void)
6535 {
6536         unsigned int n, n_counters, i;
6537 
6538         printk("perfmon: version %u.%u IRQ %u\n",
6539                 PFM_VERSION_MAJ,
6540                 PFM_VERSION_MIN,
6541                 IA64_PERFMON_VECTOR);
6542 
6543         if (pfm_probe_pmu()) {
6544                 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n", 
6545                                 local_cpu_data->family);
6546                 return -ENODEV;
6547         }
6548 
6549         /*
6550          * compute the number of implemented PMD/PMC from the
6551          * description tables
6552          */
6553         n = 0;
6554         for (i=0; PMC_IS_LAST(i) == 0;  i++) {
6555                 if (PMC_IS_IMPL(i) == 0) continue;
6556                 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6557                 n++;
6558         }
6559         pmu_conf->num_pmcs = n;
6560 
6561         n = 0; n_counters = 0;
6562         for (i=0; PMD_IS_LAST(i) == 0;  i++) {
6563                 if (PMD_IS_IMPL(i) == 0) continue;
6564                 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6565                 n++;
6566                 if (PMD_IS_COUNTING(i)) n_counters++;
6567         }
6568         pmu_conf->num_pmds      = n;
6569         pmu_conf->num_counters  = n_counters;
6570 
6571         /*
6572          * sanity checks on the number of debug registers
6573          */
6574         if (pmu_conf->use_rr_dbregs) {
6575                 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6576                         printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6577                         pmu_conf = NULL;
6578                         return -1;
6579                 }
6580                 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6581                         printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6582                         pmu_conf = NULL;
6583                         return -1;
6584                 }
6585         }
6586 
6587         printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6588                pmu_conf->pmu_name,
6589                pmu_conf->num_pmcs,
6590                pmu_conf->num_pmds,
6591                pmu_conf->num_counters,
6592                ffz(pmu_conf->ovfl_val));
6593 
6594         /* sanity check */
6595         if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6596                 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6597                 pmu_conf = NULL;
6598                 return -1;
6599         }
6600 
6601         /*
6602          * create /proc/perfmon (mostly for debugging purposes)
6603          */
6604         perfmon_dir = proc_create_seq("perfmon", S_IRUGO, NULL, &pfm_seq_ops);
6605         if (perfmon_dir == NULL) {
6606                 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6607                 pmu_conf = NULL;
6608                 return -1;
6609         }
6610 
6611         /*
6612          * create /proc/sys/kernel/perfmon (for debugging purposes)
6613          */
6614         pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6615 
6616         /*
6617          * initialize all our spinlocks
6618          */
6619         spin_lock_init(&pfm_sessions.pfs_lock);
6620         spin_lock_init(&pfm_buffer_fmt_lock);
6621 
6622         init_pfm_fs();
6623 
6624         for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6625 
6626         return 0;
6627 }
6628 
6629 __initcall(pfm_init);
6630 
6631 /*
6632  * this function is called before pfm_init()
6633  */
6634 void
6635 pfm_init_percpu (void)
6636 {
6637         static int first_time=1;
6638         /*
6639          * make sure no measurement is active
6640          * (may inherit programmed PMCs from EFI).
6641          */
6642         pfm_clear_psr_pp();
6643         pfm_clear_psr_up();
6644 
6645         /*
6646          * we run with the PMU not frozen at all times
6647          */
6648         pfm_unfreeze_pmu();
6649 
6650         if (first_time) {
6651                 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6652                 first_time=0;
6653         }
6654 
6655         ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6656         ia64_srlz_d();
6657 }
6658 
6659 /*
6660  * used for debug purposes only
6661  */
6662 void
6663 dump_pmu_state(const char *from)
6664 {
6665         struct task_struct *task;
6666         struct pt_regs *regs;
6667         pfm_context_t *ctx;
6668         unsigned long psr, dcr, info, flags;
6669         int i, this_cpu;
6670 
6671         local_irq_save(flags);
6672 
6673         this_cpu = smp_processor_id();
6674         regs     = task_pt_regs(current);
6675         info     = PFM_CPUINFO_GET();
6676         dcr      = ia64_getreg(_IA64_REG_CR_DCR);
6677 
6678         if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6679                 local_irq_restore(flags);
6680                 return;
6681         }
6682 
6683         printk("CPU%d from %s() current [%d] iip=0x%lx %s\n", 
6684                 this_cpu, 
6685                 from, 
6686                 task_pid_nr(current),
6687                 regs->cr_iip,
6688                 current->comm);
6689 
6690         task = GET_PMU_OWNER();
6691         ctx  = GET_PMU_CTX();
6692 
6693         printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6694 
6695         psr = pfm_get_psr();
6696 
6697         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", 
6698                 this_cpu,
6699                 ia64_get_pmc(0),
6700                 psr & IA64_PSR_PP ? 1 : 0,
6701                 psr & IA64_PSR_UP ? 1 : 0,
6702                 dcr & IA64_DCR_PP ? 1 : 0,
6703                 info,
6704                 ia64_psr(regs)->up,
6705                 ia64_psr(regs)->pp);
6706 
6707         ia64_psr(regs)->up = 0;
6708         ia64_psr(regs)->pp = 0;
6709 
6710         for (i=1; PMC_IS_LAST(i) == 0; i++) {
6711                 if (PMC_IS_IMPL(i) == 0) continue;
6712                 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]);
6713         }
6714 
6715         for (i=1; PMD_IS_LAST(i) == 0; i++) {
6716                 if (PMD_IS_IMPL(i) == 0) continue;
6717                 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]);
6718         }
6719 
6720         if (ctx) {
6721                 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6722                                 this_cpu,
6723                                 ctx->ctx_state,
6724                                 ctx->ctx_smpl_vaddr,
6725                                 ctx->ctx_smpl_hdr,
6726                                 ctx->ctx_msgq_head,
6727                                 ctx->ctx_msgq_tail,
6728                                 ctx->ctx_saved_psr_up);
6729         }
6730         local_irq_restore(flags);
6731 }
6732 
6733 /*
6734  * called from process.c:copy_thread(). task is new child.
6735  */
6736 void
6737 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6738 {
6739         struct thread_struct *thread;
6740 
6741         DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6742 
6743         thread = &task->thread;
6744 
6745         /*
6746          * cut links inherited from parent (current)
6747          */
6748         thread->pfm_context = NULL;
6749 
6750         PFM_SET_WORK_PENDING(task, 0);
6751 
6752         /*
6753          * the psr bits are already set properly in copy_threads()
6754          */
6755 }
6756 #else  /* !CONFIG_PERFMON */
6757 asmlinkage long
6758 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6759 {
6760         return -ENOSYS;
6761 }
6762 #endif /* CONFIG_PERFMON */
6763 

~ [ 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