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

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

Version: ~ [ linux-5.11 ] ~ [ linux-5.10.17 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.99 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.176 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.221 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.257 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.257 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ 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-2.6.32.71 ] ~ [ 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 = vm_area_alloc(mm);
2282         if (!vma) {
2283                 DPRINT(("Cannot allocate vma\n"));
2284                 goto error_kmem;
2285         }
2286 
2287         /*
2288          * partially initialize the vma for the sampling buffer
2289          */
2290         vma->vm_file         = get_file(filp);
2291         vma->vm_flags        = VM_READ|VM_MAYREAD|VM_DONTEXPAND|VM_DONTDUMP;
2292         vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */
2293 
2294         /*
2295          * Now we have everything we need and we can initialize
2296          * and connect all the data structures
2297          */
2298 
2299         ctx->ctx_smpl_hdr   = smpl_buf;
2300         ctx->ctx_smpl_size  = size; /* aligned size */
2301 
2302         /*
2303          * Let's do the difficult operations next.
2304          *
2305          * now we atomically find some area in the address space and
2306          * remap the buffer in it.
2307          */
2308         down_write(&task->mm->mmap_sem);
2309 
2310         /* find some free area in address space, must have mmap sem held */
2311         vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2312         if (IS_ERR_VALUE(vma->vm_start)) {
2313                 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2314                 up_write(&task->mm->mmap_sem);
2315                 goto error;
2316         }
2317         vma->vm_end = vma->vm_start + size;
2318         vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2319 
2320         DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2321 
2322         /* can only be applied to current task, need to have the mm semaphore held when called */
2323         if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2324                 DPRINT(("Can't remap buffer\n"));
2325                 up_write(&task->mm->mmap_sem);
2326                 goto error;
2327         }
2328 
2329         /*
2330          * now insert the vma in the vm list for the process, must be
2331          * done with mmap lock held
2332          */
2333         insert_vm_struct(mm, vma);
2334 
2335         vm_stat_account(vma->vm_mm, vma->vm_flags, vma_pages(vma));
2336         up_write(&task->mm->mmap_sem);
2337 
2338         /*
2339          * keep track of user level virtual address
2340          */
2341         ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2342         *(unsigned long *)user_vaddr = vma->vm_start;
2343 
2344         return 0;
2345 
2346 error:
2347         vm_area_free(vma);
2348 error_kmem:
2349         pfm_rvfree(smpl_buf, size);
2350 
2351         return -ENOMEM;
2352 }
2353 
2354 /*
2355  * XXX: do something better here
2356  */
2357 static int
2358 pfm_bad_permissions(struct task_struct *task)
2359 {
2360         const struct cred *tcred;
2361         kuid_t uid = current_uid();
2362         kgid_t gid = current_gid();
2363         int ret;
2364 
2365         rcu_read_lock();
2366         tcred = __task_cred(task);
2367 
2368         /* inspired by ptrace_attach() */
2369         DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2370                 from_kuid(&init_user_ns, uid),
2371                 from_kgid(&init_user_ns, gid),
2372                 from_kuid(&init_user_ns, tcred->euid),
2373                 from_kuid(&init_user_ns, tcred->suid),
2374                 from_kuid(&init_user_ns, tcred->uid),
2375                 from_kgid(&init_user_ns, tcred->egid),
2376                 from_kgid(&init_user_ns, tcred->sgid)));
2377 
2378         ret = ((!uid_eq(uid, tcred->euid))
2379                || (!uid_eq(uid, tcred->suid))
2380                || (!uid_eq(uid, tcred->uid))
2381                || (!gid_eq(gid, tcred->egid))
2382                || (!gid_eq(gid, tcred->sgid))
2383                || (!gid_eq(gid, tcred->gid))) && !capable(CAP_SYS_PTRACE);
2384 
2385         rcu_read_unlock();
2386         return ret;
2387 }
2388 
2389 static int
2390 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2391 {
2392         int ctx_flags;
2393 
2394         /* valid signal */
2395 
2396         ctx_flags = pfx->ctx_flags;
2397 
2398         if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2399 
2400                 /*
2401                  * cannot block in this mode
2402                  */
2403                 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2404                         DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2405                         return -EINVAL;
2406                 }
2407         } else {
2408         }
2409         /* probably more to add here */
2410 
2411         return 0;
2412 }
2413 
2414 static int
2415 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2416                      unsigned int cpu, pfarg_context_t *arg)
2417 {
2418         pfm_buffer_fmt_t *fmt = NULL;
2419         unsigned long size = 0UL;
2420         void *uaddr = NULL;
2421         void *fmt_arg = NULL;
2422         int ret = 0;
2423 #define PFM_CTXARG_BUF_ARG(a)   (pfm_buffer_fmt_t *)(a+1)
2424 
2425         /* invoke and lock buffer format, if found */
2426         fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2427         if (fmt == NULL) {
2428                 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2429                 return -EINVAL;
2430         }
2431 
2432         /*
2433          * buffer argument MUST be contiguous to pfarg_context_t
2434          */
2435         if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2436 
2437         ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2438 
2439         DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2440 
2441         if (ret) goto error;
2442 
2443         /* link buffer format and context */
2444         ctx->ctx_buf_fmt = fmt;
2445         ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2446 
2447         /*
2448          * check if buffer format wants to use perfmon buffer allocation/mapping service
2449          */
2450         ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2451         if (ret) goto error;
2452 
2453         if (size) {
2454                 /*
2455                  * buffer is always remapped into the caller's address space
2456                  */
2457                 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2458                 if (ret) goto error;
2459 
2460                 /* keep track of user address of buffer */
2461                 arg->ctx_smpl_vaddr = uaddr;
2462         }
2463         ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2464 
2465 error:
2466         return ret;
2467 }
2468 
2469 static void
2470 pfm_reset_pmu_state(pfm_context_t *ctx)
2471 {
2472         int i;
2473 
2474         /*
2475          * install reset values for PMC.
2476          */
2477         for (i=1; PMC_IS_LAST(i) == 0; i++) {
2478                 if (PMC_IS_IMPL(i) == 0) continue;
2479                 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2480                 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2481         }
2482         /*
2483          * PMD registers are set to 0UL when the context in memset()
2484          */
2485 
2486         /*
2487          * On context switched restore, we must restore ALL pmc and ALL pmd even
2488          * when they are not actively used by the task. In UP, the incoming process
2489          * may otherwise pick up left over PMC, PMD state from the previous process.
2490          * As opposed to PMD, stale PMC can cause harm to the incoming
2491          * process because they may change what is being measured.
2492          * Therefore, we must systematically reinstall the entire
2493          * PMC state. In SMP, the same thing is possible on the
2494          * same CPU but also on between 2 CPUs.
2495          *
2496          * The problem with PMD is information leaking especially
2497          * to user level when psr.sp=0
2498          *
2499          * There is unfortunately no easy way to avoid this problem
2500          * on either UP or SMP. This definitively slows down the
2501          * pfm_load_regs() function.
2502          */
2503 
2504          /*
2505           * bitmask of all PMCs accessible to this context
2506           *
2507           * PMC0 is treated differently.
2508           */
2509         ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2510 
2511         /*
2512          * bitmask of all PMDs that are accessible to this context
2513          */
2514         ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2515 
2516         DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2517 
2518         /*
2519          * useful in case of re-enable after disable
2520          */
2521         ctx->ctx_used_ibrs[0] = 0UL;
2522         ctx->ctx_used_dbrs[0] = 0UL;
2523 }
2524 
2525 static int
2526 pfm_ctx_getsize(void *arg, size_t *sz)
2527 {
2528         pfarg_context_t *req = (pfarg_context_t *)arg;
2529         pfm_buffer_fmt_t *fmt;
2530 
2531         *sz = 0;
2532 
2533         if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2534 
2535         fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2536         if (fmt == NULL) {
2537                 DPRINT(("cannot find buffer format\n"));
2538                 return -EINVAL;
2539         }
2540         /* get just enough to copy in user parameters */
2541         *sz = fmt->fmt_arg_size;
2542         DPRINT(("arg_size=%lu\n", *sz));
2543 
2544         return 0;
2545 }
2546 
2547 
2548 
2549 /*
2550  * cannot attach if :
2551  *      - kernel task
2552  *      - task not owned by caller
2553  *      - task incompatible with context mode
2554  */
2555 static int
2556 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2557 {
2558         /*
2559          * no kernel task or task not owner by caller
2560          */
2561         if (task->mm == NULL) {
2562                 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2563                 return -EPERM;
2564         }
2565         if (pfm_bad_permissions(task)) {
2566                 DPRINT(("no permission to attach to  [%d]\n", task_pid_nr(task)));
2567                 return -EPERM;
2568         }
2569         /*
2570          * cannot block in self-monitoring mode
2571          */
2572         if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2573                 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2574                 return -EINVAL;
2575         }
2576 
2577         if (task->exit_state == EXIT_ZOMBIE) {
2578                 DPRINT(("cannot attach to  zombie task [%d]\n", task_pid_nr(task)));
2579                 return -EBUSY;
2580         }
2581 
2582         /*
2583          * always ok for self
2584          */
2585         if (task == current) return 0;
2586 
2587         if (!task_is_stopped_or_traced(task)) {
2588                 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2589                 return -EBUSY;
2590         }
2591         /*
2592          * make sure the task is off any CPU
2593          */
2594         wait_task_inactive(task, 0);
2595 
2596         /* more to come... */
2597 
2598         return 0;
2599 }
2600 
2601 static int
2602 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2603 {
2604         struct task_struct *p = current;
2605         int ret;
2606 
2607         /* XXX: need to add more checks here */
2608         if (pid < 2) return -EPERM;
2609 
2610         if (pid != task_pid_vnr(current)) {
2611                 /* make sure task cannot go away while we operate on it */
2612                 p = find_get_task_by_vpid(pid);
2613                 if (!p)
2614                         return -ESRCH;
2615         }
2616 
2617         ret = pfm_task_incompatible(ctx, p);
2618         if (ret == 0) {
2619                 *task = p;
2620         } else if (p != current) {
2621                 pfm_put_task(p);
2622         }
2623         return ret;
2624 }
2625 
2626 
2627 
2628 static int
2629 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2630 {
2631         pfarg_context_t *req = (pfarg_context_t *)arg;
2632         struct file *filp;
2633         struct path path;
2634         int ctx_flags;
2635         int fd;
2636         int ret;
2637 
2638         /* let's check the arguments first */
2639         ret = pfarg_is_sane(current, req);
2640         if (ret < 0)
2641                 return ret;
2642 
2643         ctx_flags = req->ctx_flags;
2644 
2645         ret = -ENOMEM;
2646 
2647         fd = get_unused_fd_flags(0);
2648         if (fd < 0)
2649                 return fd;
2650 
2651         ctx = pfm_context_alloc(ctx_flags);
2652         if (!ctx)
2653                 goto error;
2654 
2655         filp = pfm_alloc_file(ctx);
2656         if (IS_ERR(filp)) {
2657                 ret = PTR_ERR(filp);
2658                 goto error_file;
2659         }
2660 
2661         req->ctx_fd = ctx->ctx_fd = fd;
2662 
2663         /*
2664          * does the user want to sample?
2665          */
2666         if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2667                 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2668                 if (ret)
2669                         goto buffer_error;
2670         }
2671 
2672         DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2673                 ctx,
2674                 ctx_flags,
2675                 ctx->ctx_fl_system,
2676                 ctx->ctx_fl_block,
2677                 ctx->ctx_fl_excl_idle,
2678                 ctx->ctx_fl_no_msg,
2679                 ctx->ctx_fd));
2680 
2681         /*
2682          * initialize soft PMU state
2683          */
2684         pfm_reset_pmu_state(ctx);
2685 
2686         fd_install(fd, filp);
2687 
2688         return 0;
2689 
2690 buffer_error:
2691         path = filp->f_path;
2692         put_filp(filp);
2693         path_put(&path);
2694 
2695         if (ctx->ctx_buf_fmt) {
2696                 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2697         }
2698 error_file:
2699         pfm_context_free(ctx);
2700 
2701 error:
2702         put_unused_fd(fd);
2703         return ret;
2704 }
2705 
2706 static inline unsigned long
2707 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2708 {
2709         unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2710         unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2711         extern unsigned long carta_random32 (unsigned long seed);
2712 
2713         if (reg->flags & PFM_REGFL_RANDOM) {
2714                 new_seed = carta_random32(old_seed);
2715                 val -= (old_seed & mask);       /* counter values are negative numbers! */
2716                 if ((mask >> 32) != 0)
2717                         /* construct a full 64-bit random value: */
2718                         new_seed |= carta_random32(old_seed >> 32) << 32;
2719                 reg->seed = new_seed;
2720         }
2721         reg->lval = val;
2722         return val;
2723 }
2724 
2725 static void
2726 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2727 {
2728         unsigned long mask = ovfl_regs[0];
2729         unsigned long reset_others = 0UL;
2730         unsigned long val;
2731         int i;
2732 
2733         /*
2734          * now restore reset value on sampling overflowed counters
2735          */
2736         mask >>= PMU_FIRST_COUNTER;
2737         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2738 
2739                 if ((mask & 0x1UL) == 0UL) continue;
2740 
2741                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2742                 reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2743 
2744                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2745         }
2746 
2747         /*
2748          * Now take care of resetting the other registers
2749          */
2750         for(i = 0; reset_others; i++, reset_others >>= 1) {
2751 
2752                 if ((reset_others & 0x1) == 0) continue;
2753 
2754                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2755 
2756                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2757                           is_long_reset ? "long" : "short", i, val));
2758         }
2759 }
2760 
2761 static void
2762 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2763 {
2764         unsigned long mask = ovfl_regs[0];
2765         unsigned long reset_others = 0UL;
2766         unsigned long val;
2767         int i;
2768 
2769         DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2770 
2771         if (ctx->ctx_state == PFM_CTX_MASKED) {
2772                 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2773                 return;
2774         }
2775 
2776         /*
2777          * now restore reset value on sampling overflowed counters
2778          */
2779         mask >>= PMU_FIRST_COUNTER;
2780         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2781 
2782                 if ((mask & 0x1UL) == 0UL) continue;
2783 
2784                 val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2785                 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2786 
2787                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2788 
2789                 pfm_write_soft_counter(ctx, i, val);
2790         }
2791 
2792         /*
2793          * Now take care of resetting the other registers
2794          */
2795         for(i = 0; reset_others; i++, reset_others >>= 1) {
2796 
2797                 if ((reset_others & 0x1) == 0) continue;
2798 
2799                 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2800 
2801                 if (PMD_IS_COUNTING(i)) {
2802                         pfm_write_soft_counter(ctx, i, val);
2803                 } else {
2804                         ia64_set_pmd(i, val);
2805                 }
2806                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2807                           is_long_reset ? "long" : "short", i, val));
2808         }
2809         ia64_srlz_d();
2810 }
2811 
2812 static int
2813 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2814 {
2815         struct task_struct *task;
2816         pfarg_reg_t *req = (pfarg_reg_t *)arg;
2817         unsigned long value, pmc_pm;
2818         unsigned long smpl_pmds, reset_pmds, impl_pmds;
2819         unsigned int cnum, reg_flags, flags, pmc_type;
2820         int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2821         int is_monitor, is_counting, state;
2822         int ret = -EINVAL;
2823         pfm_reg_check_t wr_func;
2824 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2825 
2826         state     = ctx->ctx_state;
2827         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2828         is_system = ctx->ctx_fl_system;
2829         task      = ctx->ctx_task;
2830         impl_pmds = pmu_conf->impl_pmds[0];
2831 
2832         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2833 
2834         if (is_loaded) {
2835                 /*
2836                  * In system wide and when the context is loaded, access can only happen
2837                  * when the caller is running on the CPU being monitored by the session.
2838                  * It does not have to be the owner (ctx_task) of the context per se.
2839                  */
2840                 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2841                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2842                         return -EBUSY;
2843                 }
2844                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2845         }
2846         expert_mode = pfm_sysctl.expert_mode; 
2847 
2848         for (i = 0; i < count; i++, req++) {
2849 
2850                 cnum       = req->reg_num;
2851                 reg_flags  = req->reg_flags;
2852                 value      = req->reg_value;
2853                 smpl_pmds  = req->reg_smpl_pmds[0];
2854                 reset_pmds = req->reg_reset_pmds[0];
2855                 flags      = 0;
2856 
2857 
2858                 if (cnum >= PMU_MAX_PMCS) {
2859                         DPRINT(("pmc%u is invalid\n", cnum));
2860                         goto error;
2861                 }
2862 
2863                 pmc_type   = pmu_conf->pmc_desc[cnum].type;
2864                 pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2865                 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2866                 is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2867 
2868                 /*
2869                  * we reject all non implemented PMC as well
2870                  * as attempts to modify PMC[0-3] which are used
2871                  * as status registers by the PMU
2872                  */
2873                 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2874                         DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2875                         goto error;
2876                 }
2877                 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2878                 /*
2879                  * If the PMC is a monitor, then if the value is not the default:
2880                  *      - system-wide session: PMCx.pm=1 (privileged monitor)
2881                  *      - per-task           : PMCx.pm=0 (user monitor)
2882                  */
2883                 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2884                         DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2885                                 cnum,
2886                                 pmc_pm,
2887                                 is_system));
2888                         goto error;
2889                 }
2890 
2891                 if (is_counting) {
2892                         /*
2893                          * enforce generation of overflow interrupt. Necessary on all
2894                          * CPUs.
2895                          */
2896                         value |= 1 << PMU_PMC_OI;
2897 
2898                         if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2899                                 flags |= PFM_REGFL_OVFL_NOTIFY;
2900                         }
2901 
2902                         if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2903 
2904                         /* verify validity of smpl_pmds */
2905                         if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2906                                 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2907                                 goto error;
2908                         }
2909 
2910                         /* verify validity of reset_pmds */
2911                         if ((reset_pmds & impl_pmds) != reset_pmds) {
2912                                 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2913                                 goto error;
2914                         }
2915                 } else {
2916                         if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2917                                 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2918                                 goto error;
2919                         }
2920                         /* eventid on non-counting monitors are ignored */
2921                 }
2922 
2923                 /*
2924                  * execute write checker, if any
2925                  */
2926                 if (likely(expert_mode == 0 && wr_func)) {
2927                         ret = (*wr_func)(task, ctx, cnum, &value, regs);
2928                         if (ret) goto error;
2929                         ret = -EINVAL;
2930                 }
2931 
2932                 /*
2933                  * no error on this register
2934                  */
2935                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2936 
2937                 /*
2938                  * Now we commit the changes to the software state
2939                  */
2940 
2941                 /*
2942                  * update overflow information
2943                  */
2944                 if (is_counting) {
2945                         /*
2946                          * full flag update each time a register is programmed
2947                          */
2948                         ctx->ctx_pmds[cnum].flags = flags;
2949 
2950                         ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2951                         ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2952                         ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
2953 
2954                         /*
2955                          * Mark all PMDS to be accessed as used.
2956                          *
2957                          * We do not keep track of PMC because we have to
2958                          * systematically restore ALL of them.
2959                          *
2960                          * We do not update the used_monitors mask, because
2961                          * if we have not programmed them, then will be in
2962                          * a quiescent state, therefore we will not need to
2963                          * mask/restore then when context is MASKED.
2964                          */
2965                         CTX_USED_PMD(ctx, reset_pmds);
2966                         CTX_USED_PMD(ctx, smpl_pmds);
2967                         /*
2968                          * make sure we do not try to reset on
2969                          * restart because we have established new values
2970                          */
2971                         if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
2972                 }
2973                 /*
2974                  * Needed in case the user does not initialize the equivalent
2975                  * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2976                  * possible leak here.
2977                  */
2978                 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
2979 
2980                 /*
2981                  * keep track of the monitor PMC that we are using.
2982                  * we save the value of the pmc in ctx_pmcs[] and if
2983                  * the monitoring is not stopped for the context we also
2984                  * place it in the saved state area so that it will be
2985                  * picked up later by the context switch code.
2986                  *
2987                  * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
2988                  *
2989                  * The value in th_pmcs[] may be modified on overflow, i.e.,  when
2990                  * monitoring needs to be stopped.
2991                  */
2992                 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
2993 
2994                 /*
2995                  * update context state
2996                  */
2997                 ctx->ctx_pmcs[cnum] = value;
2998 
2999                 if (is_loaded) {
3000                         /*
3001                          * write thread state
3002                          */
3003                         if (is_system == 0) ctx->th_pmcs[cnum] = value;
3004 
3005                         /*
3006                          * write hardware register if we can
3007                          */
3008                         if (can_access_pmu) {
3009                                 ia64_set_pmc(cnum, value);
3010                         }
3011 #ifdef CONFIG_SMP
3012                         else {
3013                                 /*
3014                                  * per-task SMP only here
3015                                  *
3016                                  * we are guaranteed that the task is not running on the other CPU,
3017                                  * we indicate that this PMD will need to be reloaded if the task
3018                                  * is rescheduled on the CPU it ran last on.
3019                                  */
3020                                 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3021                         }
3022 #endif
3023                 }
3024 
3025                 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",
3026                           cnum,
3027                           value,
3028                           is_loaded,
3029                           can_access_pmu,
3030                           flags,
3031                           ctx->ctx_all_pmcs[0],
3032                           ctx->ctx_used_pmds[0],
3033                           ctx->ctx_pmds[cnum].eventid,
3034                           smpl_pmds,
3035                           reset_pmds,
3036                           ctx->ctx_reload_pmcs[0],
3037                           ctx->ctx_used_monitors[0],
3038                           ctx->ctx_ovfl_regs[0]));
3039         }
3040 
3041         /*
3042          * make sure the changes are visible
3043          */
3044         if (can_access_pmu) ia64_srlz_d();
3045 
3046         return 0;
3047 error:
3048         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3049         return ret;
3050 }
3051 
3052 static int
3053 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3054 {
3055         struct task_struct *task;
3056         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3057         unsigned long value, hw_value, ovfl_mask;
3058         unsigned int cnum;
3059         int i, can_access_pmu = 0, state;
3060         int is_counting, is_loaded, is_system, expert_mode;
3061         int ret = -EINVAL;
3062         pfm_reg_check_t wr_func;
3063 
3064 
3065         state     = ctx->ctx_state;
3066         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3067         is_system = ctx->ctx_fl_system;
3068         ovfl_mask = pmu_conf->ovfl_val;
3069         task      = ctx->ctx_task;
3070 
3071         if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3072 
3073         /*
3074          * on both UP and SMP, we can only write to the PMC when the task is
3075          * the owner of the local PMU.
3076          */
3077         if (likely(is_loaded)) {
3078                 /*
3079                  * In system wide and when the context is loaded, access can only happen
3080                  * when the caller is running on the CPU being monitored by the session.
3081                  * It does not have to be the owner (ctx_task) of the context per se.
3082                  */
3083                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3084                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3085                         return -EBUSY;
3086                 }
3087                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3088         }
3089         expert_mode = pfm_sysctl.expert_mode; 
3090 
3091         for (i = 0; i < count; i++, req++) {
3092 
3093                 cnum  = req->reg_num;
3094                 value = req->reg_value;
3095 
3096                 if (!PMD_IS_IMPL(cnum)) {
3097                         DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3098                         goto abort_mission;
3099                 }
3100                 is_counting = PMD_IS_COUNTING(cnum);
3101                 wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3102 
3103                 /*
3104                  * execute write checker, if any
3105                  */
3106                 if (unlikely(expert_mode == 0 && wr_func)) {
3107                         unsigned long v = value;
3108 
3109                         ret = (*wr_func)(task, ctx, cnum, &v, regs);
3110                         if (ret) goto abort_mission;
3111 
3112                         value = v;
3113                         ret   = -EINVAL;
3114                 }
3115 
3116                 /*
3117                  * no error on this register
3118                  */
3119                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3120 
3121                 /*
3122                  * now commit changes to software state
3123                  */
3124                 hw_value = value;
3125 
3126                 /*
3127                  * update virtualized (64bits) counter
3128                  */
3129                 if (is_counting) {
3130                         /*
3131                          * write context state
3132                          */
3133                         ctx->ctx_pmds[cnum].lval = value;
3134 
3135                         /*
3136                          * when context is load we use the split value
3137                          */
3138                         if (is_loaded) {
3139                                 hw_value = value &  ovfl_mask;
3140                                 value    = value & ~ovfl_mask;
3141                         }
3142                 }
3143                 /*
3144                  * update reset values (not just for counters)
3145                  */
3146                 ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3147                 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3148 
3149                 /*
3150                  * update randomization parameters (not just for counters)
3151                  */
3152                 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3153                 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3154 
3155                 /*
3156                  * update context value
3157                  */
3158                 ctx->ctx_pmds[cnum].val  = value;
3159 
3160                 /*
3161                  * Keep track of what we use
3162                  *
3163                  * We do not keep track of PMC because we have to
3164                  * systematically restore ALL of them.
3165                  */
3166                 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3167 
3168                 /*
3169                  * mark this PMD register used as well
3170                  */
3171                 CTX_USED_PMD(ctx, RDEP(cnum));
3172 
3173                 /*
3174                  * make sure we do not try to reset on
3175                  * restart because we have established new values
3176                  */
3177                 if (is_counting && state == PFM_CTX_MASKED) {
3178                         ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3179                 }
3180 
3181                 if (is_loaded) {
3182                         /*
3183                          * write thread state
3184                          */
3185                         if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3186 
3187                         /*
3188                          * write hardware register if we can
3189                          */
3190                         if (can_access_pmu) {
3191                                 ia64_set_pmd(cnum, hw_value);
3192                         } else {
3193 #ifdef CONFIG_SMP
3194                                 /*
3195                                  * we are guaranteed that the task is not running on the other CPU,
3196                                  * we indicate that this PMD will need to be reloaded if the task
3197                                  * is rescheduled on the CPU it ran last on.
3198                                  */
3199                                 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3200 #endif
3201                         }
3202                 }
3203 
3204                 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3205                           "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",
3206                         cnum,
3207                         value,
3208                         is_loaded,
3209                         can_access_pmu,
3210                         hw_value,
3211                         ctx->ctx_pmds[cnum].val,
3212                         ctx->ctx_pmds[cnum].short_reset,
3213                         ctx->ctx_pmds[cnum].long_reset,
3214                         PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3215                         ctx->ctx_pmds[cnum].seed,
3216                         ctx->ctx_pmds[cnum].mask,
3217                         ctx->ctx_used_pmds[0],
3218                         ctx->ctx_pmds[cnum].reset_pmds[0],
3219                         ctx->ctx_reload_pmds[0],
3220                         ctx->ctx_all_pmds[0],
3221                         ctx->ctx_ovfl_regs[0]));
3222         }
3223 
3224         /*
3225          * make changes visible
3226          */
3227         if (can_access_pmu) ia64_srlz_d();
3228 
3229         return 0;
3230 
3231 abort_mission:
3232         /*
3233          * for now, we have only one possibility for error
3234          */
3235         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3236         return ret;
3237 }
3238 
3239 /*
3240  * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3241  * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3242  * interrupt is delivered during the call, it will be kept pending until we leave, making
3243  * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3244  * guaranteed to return consistent data to the user, it may simply be old. It is not
3245  * trivial to treat the overflow while inside the call because you may end up in
3246  * some module sampling buffer code causing deadlocks.
3247  */
3248 static int
3249 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3250 {
3251         struct task_struct *task;
3252         unsigned long val = 0UL, lval, ovfl_mask, sval;
3253         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3254         unsigned int cnum, reg_flags = 0;
3255         int i, can_access_pmu = 0, state;
3256         int is_loaded, is_system, is_counting, expert_mode;
3257         int ret = -EINVAL;
3258         pfm_reg_check_t rd_func;
3259 
3260         /*
3261          * access is possible when loaded only for
3262          * self-monitoring tasks or in UP mode
3263          */
3264 
3265         state     = ctx->ctx_state;
3266         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3267         is_system = ctx->ctx_fl_system;
3268         ovfl_mask = pmu_conf->ovfl_val;
3269         task      = ctx->ctx_task;
3270 
3271         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3272 
3273         if (likely(is_loaded)) {
3274                 /*
3275                  * In system wide and when the context is loaded, access can only happen
3276                  * when the caller is running on the CPU being monitored by the session.
3277                  * It does not have to be the owner (ctx_task) of the context per se.
3278                  */
3279                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3280                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3281                         return -EBUSY;
3282                 }
3283                 /*
3284                  * this can be true when not self-monitoring only in UP
3285                  */
3286                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3287 
3288                 if (can_access_pmu) ia64_srlz_d();
3289         }
3290         expert_mode = pfm_sysctl.expert_mode; 
3291 
3292         DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3293                 is_loaded,
3294                 can_access_pmu,
3295                 state));
3296 
3297         /*
3298          * on both UP and SMP, we can only read the PMD from the hardware register when
3299          * the task is the owner of the local PMU.
3300          */
3301 
3302         for (i = 0; i < count; i++, req++) {
3303 
3304                 cnum        = req->reg_num;
3305                 reg_flags   = req->reg_flags;
3306 
3307                 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3308                 /*
3309                  * we can only read the register that we use. That includes
3310                  * the one we explicitly initialize AND the one we want included
3311                  * in the sampling buffer (smpl_regs).
3312                  *
3313                  * Having this restriction allows optimization in the ctxsw routine
3314                  * without compromising security (leaks)
3315                  */
3316                 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3317 
3318                 sval        = ctx->ctx_pmds[cnum].val;
3319                 lval        = ctx->ctx_pmds[cnum].lval;
3320                 is_counting = PMD_IS_COUNTING(cnum);
3321 
3322                 /*
3323                  * If the task is not the current one, then we check if the
3324                  * PMU state is still in the local live register due to lazy ctxsw.
3325                  * If true, then we read directly from the registers.
3326                  */
3327                 if (can_access_pmu){
3328                         val = ia64_get_pmd(cnum);
3329                 } else {
3330                         /*
3331                          * context has been saved
3332                          * if context is zombie, then task does not exist anymore.
3333                          * In this case, we use the full value saved in the context (pfm_flush_regs()).
3334                          */
3335                         val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3336                 }
3337                 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3338 
3339                 if (is_counting) {
3340                         /*
3341                          * XXX: need to check for overflow when loaded
3342                          */
3343                         val &= ovfl_mask;
3344                         val += sval;
3345                 }
3346 
3347                 /*
3348                  * execute read checker, if any
3349                  */
3350                 if (unlikely(expert_mode == 0 && rd_func)) {
3351                         unsigned long v = val;
3352                         ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3353                         if (ret) goto error;
3354                         val = v;
3355                         ret = -EINVAL;
3356                 }
3357 
3358                 PFM_REG_RETFLAG_SET(reg_flags, 0);
3359 
3360                 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3361 
3362                 /*
3363                  * update register return value, abort all if problem during copy.
3364                  * we only modify the reg_flags field. no check mode is fine because
3365                  * access has been verified upfront in sys_perfmonctl().
3366                  */
3367                 req->reg_value            = val;
3368                 req->reg_flags            = reg_flags;
3369                 req->reg_last_reset_val   = lval;
3370         }
3371 
3372         return 0;
3373 
3374 error:
3375         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3376         return ret;
3377 }
3378 
3379 int
3380 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3381 {
3382         pfm_context_t *ctx;
3383 
3384         if (req == NULL) return -EINVAL;
3385 
3386         ctx = GET_PMU_CTX();
3387 
3388         if (ctx == NULL) return -EINVAL;
3389 
3390         /*
3391          * for now limit to current task, which is enough when calling
3392          * from overflow handler
3393          */
3394         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3395 
3396         return pfm_write_pmcs(ctx, req, nreq, regs);
3397 }
3398 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3399 
3400 int
3401 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3402 {
3403         pfm_context_t *ctx;
3404 
3405         if (req == NULL) return -EINVAL;
3406 
3407         ctx = GET_PMU_CTX();
3408 
3409         if (ctx == NULL) return -EINVAL;
3410 
3411         /*
3412          * for now limit to current task, which is enough when calling
3413          * from overflow handler
3414          */
3415         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3416 
3417         return pfm_read_pmds(ctx, req, nreq, regs);
3418 }
3419 EXPORT_SYMBOL(pfm_mod_read_pmds);
3420 
3421 /*
3422  * Only call this function when a process it trying to
3423  * write the debug registers (reading is always allowed)
3424  */
3425 int
3426 pfm_use_debug_registers(struct task_struct *task)
3427 {
3428         pfm_context_t *ctx = task->thread.pfm_context;
3429         unsigned long flags;
3430         int ret = 0;
3431 
3432         if (pmu_conf->use_rr_dbregs == 0) return 0;
3433 
3434         DPRINT(("called for [%d]\n", task_pid_nr(task)));
3435 
3436         /*
3437          * do it only once
3438          */
3439         if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3440 
3441         /*
3442          * Even on SMP, we do not need to use an atomic here because
3443          * the only way in is via ptrace() and this is possible only when the
3444          * process is stopped. Even in the case where the ctxsw out is not totally
3445          * completed by the time we come here, there is no way the 'stopped' process
3446          * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3447          * So this is always safe.
3448          */
3449         if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3450 
3451         LOCK_PFS(flags);
3452 
3453         /*
3454          * We cannot allow setting breakpoints when system wide monitoring
3455          * sessions are using the debug registers.
3456          */
3457         if (pfm_sessions.pfs_sys_use_dbregs> 0)
3458                 ret = -1;
3459         else
3460                 pfm_sessions.pfs_ptrace_use_dbregs++;
3461 
3462         DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3463                   pfm_sessions.pfs_ptrace_use_dbregs,
3464                   pfm_sessions.pfs_sys_use_dbregs,
3465                   task_pid_nr(task), ret));
3466 
3467         UNLOCK_PFS(flags);
3468 
3469         return ret;
3470 }
3471 
3472 /*
3473  * This function is called for every task that exits with the
3474  * IA64_THREAD_DBG_VALID set. This indicates a task which was
3475  * able to use the debug registers for debugging purposes via
3476  * ptrace(). Therefore we know it was not using them for
3477  * performance monitoring, so we only decrement the number
3478  * of "ptraced" debug register users to keep the count up to date
3479  */
3480 int
3481 pfm_release_debug_registers(struct task_struct *task)
3482 {
3483         unsigned long flags;
3484         int ret;
3485 
3486         if (pmu_conf->use_rr_dbregs == 0) return 0;
3487 
3488         LOCK_PFS(flags);
3489         if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3490                 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3491                 ret = -1;
3492         }  else {
3493                 pfm_sessions.pfs_ptrace_use_dbregs--;
3494                 ret = 0;
3495         }
3496         UNLOCK_PFS(flags);
3497 
3498         return ret;
3499 }
3500 
3501 static int
3502 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3503 {
3504         struct task_struct *task;
3505         pfm_buffer_fmt_t *fmt;
3506         pfm_ovfl_ctrl_t rst_ctrl;
3507         int state, is_system;
3508         int ret = 0;
3509 
3510         state     = ctx->ctx_state;
3511         fmt       = ctx->ctx_buf_fmt;
3512         is_system = ctx->ctx_fl_system;
3513         task      = PFM_CTX_TASK(ctx);
3514 
3515         switch(state) {
3516                 case PFM_CTX_MASKED:
3517                         break;
3518                 case PFM_CTX_LOADED: 
3519                         if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3520                         /* fall through */
3521                 case PFM_CTX_UNLOADED:
3522                 case PFM_CTX_ZOMBIE:
3523                         DPRINT(("invalid state=%d\n", state));
3524                         return -EBUSY;
3525                 default:
3526                         DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3527                         return -EINVAL;
3528         }
3529 
3530         /*
3531          * In system wide and when the context is loaded, access can only happen
3532          * when the caller is running on the CPU being monitored by the session.
3533          * It does not have to be the owner (ctx_task) of the context per se.
3534          */
3535         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3536                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3537                 return -EBUSY;
3538         }
3539 
3540         /* sanity check */
3541         if (unlikely(task == NULL)) {
3542                 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3543                 return -EINVAL;
3544         }
3545 
3546         if (task == current || is_system) {
3547 
3548                 fmt = ctx->ctx_buf_fmt;
3549 
3550                 DPRINT(("restarting self %d ovfl=0x%lx\n",
3551                         task_pid_nr(task),
3552                         ctx->ctx_ovfl_regs[0]));
3553 
3554                 if (CTX_HAS_SMPL(ctx)) {
3555 
3556                         prefetch(ctx->ctx_smpl_hdr);
3557 
3558                         rst_ctrl.bits.mask_monitoring = 0;
3559                         rst_ctrl.bits.reset_ovfl_pmds = 0;
3560 
3561                         if (state == PFM_CTX_LOADED)
3562                                 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3563                         else
3564                                 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3565                 } else {
3566                         rst_ctrl.bits.mask_monitoring = 0;
3567                         rst_ctrl.bits.reset_ovfl_pmds = 1;
3568                 }
3569 
3570                 if (ret == 0) {
3571                         if (rst_ctrl.bits.reset_ovfl_pmds)
3572                                 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3573 
3574                         if (rst_ctrl.bits.mask_monitoring == 0) {
3575                                 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3576 
3577                                 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3578                         } else {
3579                                 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3580 
3581                                 // cannot use pfm_stop_monitoring(task, regs);
3582                         }
3583                 }
3584                 /*
3585                  * clear overflowed PMD mask to remove any stale information
3586                  */
3587                 ctx->ctx_ovfl_regs[0] = 0UL;
3588 
3589                 /*
3590                  * back to LOADED state
3591                  */
3592                 ctx->ctx_state = PFM_CTX_LOADED;
3593 
3594                 /*
3595                  * XXX: not really useful for self monitoring
3596                  */
3597                 ctx->ctx_fl_can_restart = 0;
3598 
3599                 return 0;
3600         }
3601 
3602         /* 
3603          * restart another task
3604          */
3605 
3606         /*
3607          * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3608          * one is seen by the task.
3609          */
3610         if (state == PFM_CTX_MASKED) {
3611                 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3612                 /*
3613                  * will prevent subsequent restart before this one is
3614                  * seen by other task
3615                  */
3616                 ctx->ctx_fl_can_restart = 0;
3617         }
3618 
3619         /*
3620          * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3621          * the task is blocked or on its way to block. That's the normal
3622          * restart path. If the monitoring is not masked, then the task
3623          * can be actively monitoring and we cannot directly intervene.
3624          * Therefore we use the trap mechanism to catch the task and
3625          * force it to reset the buffer/reset PMDs.
3626          *
3627          * if non-blocking, then we ensure that the task will go into
3628          * pfm_handle_work() before returning to user mode.
3629          *
3630          * We cannot explicitly reset another task, it MUST always
3631          * be done by the task itself. This works for system wide because
3632          * the tool that is controlling the session is logically doing 
3633          * "self-monitoring".
3634          */
3635         if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3636                 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3637                 complete(&ctx->ctx_restart_done);
3638         } else {
3639                 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3640 
3641                 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3642 
3643                 PFM_SET_WORK_PENDING(task, 1);
3644 
3645                 set_notify_resume(task);
3646 
3647                 /*
3648                  * XXX: send reschedule if task runs on another CPU
3649                  */
3650         }
3651         return 0;
3652 }
3653 
3654 static int
3655 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3656 {
3657         unsigned int m = *(unsigned int *)arg;
3658 
3659         pfm_sysctl.debug = m == 0 ? 0 : 1;
3660 
3661         printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3662 
3663         if (m == 0) {
3664                 memset(pfm_stats, 0, sizeof(pfm_stats));
3665                 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3666         }
3667         return 0;
3668 }
3669 
3670 /*
3671  * arg can be NULL and count can be zero for this function
3672  */
3673 static int
3674 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3675 {
3676         struct thread_struct *thread = NULL;
3677         struct task_struct *task;
3678         pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3679         unsigned long flags;
3680         dbreg_t dbreg;
3681         unsigned int rnum;
3682         int first_time;
3683         int ret = 0, state;
3684         int i, can_access_pmu = 0;
3685         int is_system, is_loaded;
3686 
3687         if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3688 
3689         state     = ctx->ctx_state;
3690         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3691         is_system = ctx->ctx_fl_system;
3692         task      = ctx->ctx_task;
3693 
3694         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3695 
3696         /*
3697          * on both UP and SMP, we can only write to the PMC when the task is
3698          * the owner of the local PMU.
3699          */
3700         if (is_loaded) {
3701                 thread = &task->thread;
3702                 /*
3703                  * In system wide and when the context is loaded, access can only happen
3704                  * when the caller is running on the CPU being monitored by the session.
3705                  * It does not have to be the owner (ctx_task) of the context per se.
3706                  */
3707                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3708                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3709                         return -EBUSY;
3710                 }
3711                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3712         }
3713 
3714         /*
3715          * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3716          * ensuring that no real breakpoint can be installed via this call.
3717          *
3718          * IMPORTANT: regs can be NULL in this function
3719          */
3720 
3721         first_time = ctx->ctx_fl_using_dbreg == 0;
3722 
3723         /*
3724          * don't bother if we are loaded and task is being debugged
3725          */
3726         if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3727                 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3728                 return -EBUSY;
3729         }
3730 
3731         /*
3732          * check for debug registers in system wide mode
3733          *
3734          * If though a check is done in pfm_context_load(),
3735          * we must repeat it here, in case the registers are
3736          * written after the context is loaded
3737          */
3738         if (is_loaded) {
3739                 LOCK_PFS(flags);
3740 
3741                 if (first_time && is_system) {
3742                         if (pfm_sessions.pfs_ptrace_use_dbregs)
3743                                 ret = -EBUSY;
3744                         else
3745                                 pfm_sessions.pfs_sys_use_dbregs++;
3746                 }
3747                 UNLOCK_PFS(flags);
3748         }
3749 
3750         if (ret != 0) return ret;
3751 
3752         /*
3753          * mark ourself as user of the debug registers for
3754          * perfmon purposes.
3755          */
3756         ctx->ctx_fl_using_dbreg = 1;
3757 
3758         /*
3759          * clear hardware registers to make sure we don't
3760          * pick up stale state.
3761          *
3762          * for a system wide session, we do not use
3763          * thread.dbr, thread.ibr because this process
3764          * never leaves the current CPU and the state
3765          * is shared by all processes running on it
3766          */
3767         if (first_time && can_access_pmu) {
3768                 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3769                 for (i=0; i < pmu_conf->num_ibrs; i++) {
3770                         ia64_set_ibr(i, 0UL);
3771                         ia64_dv_serialize_instruction();
3772                 }
3773                 ia64_srlz_i();
3774                 for (i=0; i < pmu_conf->num_dbrs; i++) {
3775                         ia64_set_dbr(i, 0UL);
3776                         ia64_dv_serialize_data();
3777                 }
3778                 ia64_srlz_d();
3779         }
3780 
3781         /*
3782          * Now install the values into the registers
3783          */
3784         for (i = 0; i < count; i++, req++) {
3785 
3786                 rnum      = req->dbreg_num;
3787                 dbreg.val = req->dbreg_value;
3788 
3789                 ret = -EINVAL;
3790 
3791                 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3792                         DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3793                                   rnum, dbreg.val, mode, i, count));
3794 
3795                         goto abort_mission;
3796                 }
3797 
3798                 /*
3799                  * make sure we do not install enabled breakpoint
3800                  */
3801                 if (rnum & 0x1) {
3802                         if (mode == PFM_CODE_RR)
3803                                 dbreg.ibr.ibr_x = 0;
3804                         else
3805                                 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3806                 }
3807 
3808                 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3809 
3810                 /*
3811                  * Debug registers, just like PMC, can only be modified
3812                  * by a kernel call. Moreover, perfmon() access to those
3813                  * registers are centralized in this routine. The hardware
3814                  * does not modify the value of these registers, therefore,
3815                  * if we save them as they are written, we can avoid having
3816                  * to save them on context switch out. This is made possible
3817                  * by the fact that when perfmon uses debug registers, ptrace()
3818                  * won't be able to modify them concurrently.
3819                  */
3820                 if (mode == PFM_CODE_RR) {
3821                         CTX_USED_IBR(ctx, rnum);
3822 
3823                         if (can_access_pmu) {
3824                                 ia64_set_ibr(rnum, dbreg.val);
3825                                 ia64_dv_serialize_instruction();
3826                         }
3827 
3828                         ctx->ctx_ibrs[rnum] = dbreg.val;
3829 
3830                         DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3831                                 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3832                 } else {
3833                         CTX_USED_DBR(ctx, rnum);
3834 
3835                         if (can_access_pmu) {
3836                                 ia64_set_dbr(rnum, dbreg.val);
3837                                 ia64_dv_serialize_data();
3838                         }
3839                         ctx->ctx_dbrs[rnum] = dbreg.val;
3840 
3841                         DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3842                                 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3843                 }
3844         }
3845 
3846         return 0;
3847 
3848 abort_mission:
3849         /*
3850          * in case it was our first attempt, we undo the global modifications
3851          */
3852         if (first_time) {
3853                 LOCK_PFS(flags);
3854                 if (ctx->ctx_fl_system) {
3855                         pfm_sessions.pfs_sys_use_dbregs--;
3856                 }
3857                 UNLOCK_PFS(flags);
3858                 ctx->ctx_fl_using_dbreg = 0;
3859         }
3860         /*
3861          * install error return flag
3862          */
3863         PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3864 
3865         return ret;
3866 }
3867 
3868 static int
3869 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3870 {
3871         return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3872 }
3873 
3874 static int
3875 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3876 {
3877         return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3878 }
3879 
3880 int
3881 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3882 {
3883         pfm_context_t *ctx;
3884 
3885         if (req == NULL) return -EINVAL;
3886 
3887         ctx = GET_PMU_CTX();
3888 
3889         if (ctx == NULL) return -EINVAL;
3890 
3891         /*
3892          * for now limit to current task, which is enough when calling
3893          * from overflow handler
3894          */
3895         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3896 
3897         return pfm_write_ibrs(ctx, req, nreq, regs);
3898 }
3899 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3900 
3901 int
3902 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3903 {
3904         pfm_context_t *ctx;
3905 
3906         if (req == NULL) return -EINVAL;
3907 
3908         ctx = GET_PMU_CTX();
3909 
3910         if (ctx == NULL) return -EINVAL;
3911 
3912         /*
3913          * for now limit to current task, which is enough when calling
3914          * from overflow handler
3915          */
3916         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3917 
3918         return pfm_write_dbrs(ctx, req, nreq, regs);
3919 }
3920 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3921 
3922 
3923 static int
3924 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3925 {
3926         pfarg_features_t *req = (pfarg_features_t *)arg;
3927 
3928         req->ft_version = PFM_VERSION;
3929         return 0;
3930 }
3931 
3932 static int
3933 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3934 {
3935         struct pt_regs *tregs;
3936         struct task_struct *task = PFM_CTX_TASK(ctx);
3937         int state, is_system;
3938 
3939         state     = ctx->ctx_state;
3940         is_system = ctx->ctx_fl_system;
3941 
3942         /*
3943          * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3944          */
3945         if (state == PFM_CTX_UNLOADED) return -EINVAL;
3946 
3947         /*
3948          * In system wide and when the context is loaded, access can only happen
3949          * when the caller is running on the CPU being monitored by the session.
3950          * It does not have to be the owner (ctx_task) of the context per se.
3951          */
3952         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3953                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3954                 return -EBUSY;
3955         }
3956         DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3957                 task_pid_nr(PFM_CTX_TASK(ctx)),
3958                 state,
3959                 is_system));
3960         /*
3961          * in system mode, we need to update the PMU directly
3962          * and the user level state of the caller, which may not
3963          * necessarily be the creator of the context.
3964          */
3965         if (is_system) {
3966                 /*
3967                  * Update local PMU first
3968                  *
3969                  * disable dcr pp
3970                  */
3971                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
3972                 ia64_srlz_i();
3973 
3974                 /*
3975                  * update local cpuinfo
3976                  */
3977                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
3978 
3979                 /*
3980                  * stop monitoring, does srlz.i
3981                  */
3982                 pfm_clear_psr_pp();
3983 
3984                 /*
3985                  * stop monitoring in the caller
3986                  */
3987                 ia64_psr(regs)->pp = 0;
3988 
3989                 return 0;
3990         }
3991         /*
3992          * per-task mode
3993          */
3994 
3995         if (task == current) {
3996                 /* stop monitoring  at kernel level */
3997                 pfm_clear_psr_up();
3998 
3999                 /*
4000                  * stop monitoring at the user level
4001                  */
4002                 ia64_psr(regs)->up = 0;
4003         } else {
4004                 tregs = task_pt_regs(task);
4005 
4006                 /*
4007                  * stop monitoring at the user level
4008                  */
4009                 ia64_psr(tregs)->up = 0;
4010 
4011                 /*
4012                  * monitoring disabled in kernel at next reschedule
4013                  */
4014                 ctx->ctx_saved_psr_up = 0;
4015                 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4016         }
4017         return 0;
4018 }
4019 
4020 
4021 static int
4022 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4023 {
4024         struct pt_regs *tregs;
4025         int state, is_system;
4026 
4027         state     = ctx->ctx_state;
4028         is_system = ctx->ctx_fl_system;
4029 
4030         if (state != PFM_CTX_LOADED) return -EINVAL;
4031 
4032         /*
4033          * In system wide and when the context is loaded, access can only happen
4034          * when the caller is running on the CPU being monitored by the session.
4035          * It does not have to be the owner (ctx_task) of the context per se.
4036          */
4037         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4038                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4039                 return -EBUSY;
4040         }
4041 
4042         /*
4043          * in system mode, we need to update the PMU directly
4044          * and the user level state of the caller, which may not
4045          * necessarily be the creator of the context.
4046          */
4047         if (is_system) {
4048 
4049                 /*
4050                  * set user level psr.pp for the caller
4051                  */
4052                 ia64_psr(regs)->pp = 1;
4053 
4054                 /*
4055                  * now update the local PMU and cpuinfo
4056                  */
4057                 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4058 
4059                 /*
4060                  * start monitoring at kernel level
4061                  */
4062                 pfm_set_psr_pp();
4063 
4064                 /* enable dcr pp */
4065                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4066                 ia64_srlz_i();
4067 
4068                 return 0;
4069         }
4070 
4071         /*
4072          * per-process mode
4073          */
4074 
4075         if (ctx->ctx_task == current) {
4076 
4077                 /* start monitoring at kernel level */
4078                 pfm_set_psr_up();
4079 
4080                 /*
4081                  * activate monitoring at user level
4082                  */
4083                 ia64_psr(regs)->up = 1;
4084 
4085         } else {
4086                 tregs = task_pt_regs(ctx->ctx_task);
4087 
4088                 /*
4089                  * start monitoring at the kernel level the next
4090                  * time the task is scheduled
4091                  */
4092                 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4093 
4094                 /*
4095                  * activate monitoring at user level
4096                  */
4097                 ia64_psr(tregs)->up = 1;
4098         }
4099         return 0;
4100 }
4101 
4102 static int
4103 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4104 {
4105         pfarg_reg_t *req = (pfarg_reg_t *)arg;
4106         unsigned int cnum;
4107         int i;
4108         int ret = -EINVAL;
4109 
4110         for (i = 0; i < count; i++, req++) {
4111 
4112                 cnum = req->reg_num;
4113 
4114                 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4115 
4116                 req->reg_value = PMC_DFL_VAL(cnum);
4117 
4118                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4119 
4120                 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4121         }
4122         return 0;
4123 
4124 abort_mission:
4125         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4126         return ret;
4127 }
4128 
4129 static int
4130 pfm_check_task_exist(pfm_context_t *ctx)
4131 {
4132         struct task_struct *g, *t;
4133         int ret = -ESRCH;
4134 
4135         read_lock(&tasklist_lock);
4136 
4137         do_each_thread (g, t) {
4138                 if (t->thread.pfm_context == ctx) {
4139                         ret = 0;
4140                         goto out;
4141                 }
4142         } while_each_thread (g, t);
4143 out:
4144         read_unlock(&tasklist_lock);
4145 
4146         DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4147 
4148         return ret;
4149 }
4150 
4151 static int
4152 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4153 {
4154         struct task_struct *task;
4155         struct thread_struct *thread;
4156         struct pfm_context_t *old;
4157         unsigned long flags;
4158 #ifndef CONFIG_SMP
4159         struct task_struct *owner_task = NULL;
4160 #endif
4161         pfarg_load_t *req = (pfarg_load_t *)arg;
4162         unsigned long *pmcs_source, *pmds_source;
4163         int the_cpu;
4164         int ret = 0;
4165         int state, is_system, set_dbregs = 0;
4166 
4167         state     = ctx->ctx_state;
4168         is_system = ctx->ctx_fl_system;
4169         /*
4170          * can only load from unloaded or terminated state
4171          */
4172         if (state != PFM_CTX_UNLOADED) {
4173                 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4174                         req->load_pid,
4175                         ctx->ctx_state));
4176                 return -EBUSY;
4177         }
4178 
4179         DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4180 
4181         if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4182                 DPRINT(("cannot use blocking mode on self\n"));
4183                 return -EINVAL;
4184         }
4185 
4186         ret = pfm_get_task(ctx, req->load_pid, &task);
4187         if (ret) {
4188                 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4189                 return ret;
4190         }
4191 
4192         ret = -EINVAL;
4193 
4194         /*
4195          * system wide is self monitoring only
4196          */
4197         if (is_system && task != current) {
4198                 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4199                         req->load_pid));
4200                 goto error;
4201         }
4202 
4203         thread = &task->thread;
4204 
4205         ret = 0;
4206         /*
4207          * cannot load a context which is using range restrictions,
4208          * into a task that is being debugged.
4209          */
4210         if (ctx->ctx_fl_using_dbreg) {
4211                 if (thread->flags & IA64_THREAD_DBG_VALID) {
4212                         ret = -EBUSY;
4213                         DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4214                         goto error;
4215                 }
4216                 LOCK_PFS(flags);
4217 
4218                 if (is_system) {
4219                         if (pfm_sessions.pfs_ptrace_use_dbregs) {
4220                                 DPRINT(("cannot load [%d] dbregs in use\n",
4221                                                         task_pid_nr(task)));
4222                                 ret = -EBUSY;
4223                         } else {
4224                                 pfm_sessions.pfs_sys_use_dbregs++;
4225                                 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4226                                 set_dbregs = 1;
4227                         }
4228                 }
4229 
4230                 UNLOCK_PFS(flags);
4231 
4232                 if (ret) goto error;
4233         }
4234 
4235         /*
4236          * SMP system-wide monitoring implies self-monitoring.
4237          *
4238          * The programming model expects the task to
4239          * be pinned on a CPU throughout the session.
4240          * Here we take note of the current CPU at the
4241          * time the context is loaded. No call from
4242          * another CPU will be allowed.
4243          *
4244          * The pinning via shed_setaffinity()
4245          * must be done by the calling task prior
4246          * to this call.
4247          *
4248          * systemwide: keep track of CPU this session is supposed to run on
4249          */
4250         the_cpu = ctx->ctx_cpu = smp_processor_id();
4251 
4252         ret = -EBUSY;
4253         /*
4254          * now reserve the session
4255          */
4256         ret = pfm_reserve_session(current, is_system, the_cpu);
4257         if (ret) goto error;
4258 
4259         /*
4260          * task is necessarily stopped at this point.
4261          *
4262          * If the previous context was zombie, then it got removed in
4263          * pfm_save_regs(). Therefore we should not see it here.
4264          * If we see a context, then this is an active context
4265          *
4266          * XXX: needs to be atomic
4267          */
4268         DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4269                 thread->pfm_context, ctx));
4270 
4271         ret = -EBUSY;
4272         old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4273         if (old != NULL) {
4274                 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4275                 goto error_unres;
4276         }
4277 
4278         pfm_reset_msgq(ctx);
4279 
4280         ctx->ctx_state = PFM_CTX_LOADED;
4281 
4282         /*
4283          * link context to task
4284          */
4285         ctx->ctx_task = task;
4286 
4287         if (is_system) {
4288                 /*
4289                  * we load as stopped
4290                  */
4291                 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4292                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4293 
4294                 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4295         } else {
4296                 thread->flags |= IA64_THREAD_PM_VALID;
4297         }
4298 
4299         /*
4300          * propagate into thread-state
4301          */
4302         pfm_copy_pmds(task, ctx);
4303         pfm_copy_pmcs(task, ctx);
4304 
4305         pmcs_source = ctx->th_pmcs;
4306         pmds_source = ctx->th_pmds;
4307 
4308         /*
4309          * always the case for system-wide
4310          */
4311         if (task == current) {
4312 
4313                 if (is_system == 0) {
4314 
4315                         /* allow user level control */
4316                         ia64_psr(regs)->sp = 0;
4317                         DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4318 
4319                         SET_LAST_CPU(ctx, smp_processor_id());
4320                         INC_ACTIVATION();
4321                         SET_ACTIVATION(ctx);
4322 #ifndef CONFIG_SMP
4323                         /*
4324                          * push the other task out, if any
4325                          */
4326                         owner_task = GET_PMU_OWNER();
4327                         if (owner_task) pfm_lazy_save_regs(owner_task);
4328 #endif
4329                 }
4330                 /*
4331                  * load all PMD from ctx to PMU (as opposed to thread state)
4332                  * restore all PMC from ctx to PMU
4333                  */
4334                 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4335                 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4336 
4337                 ctx->ctx_reload_pmcs[0] = 0UL;
4338                 ctx->ctx_reload_pmds[0] = 0UL;
4339 
4340                 /*
4341                  * guaranteed safe by earlier check against DBG_VALID
4342                  */
4343                 if (ctx->ctx_fl_using_dbreg) {
4344                         pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4345                         pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4346                 }
4347                 /*
4348                  * set new ownership
4349                  */
4350                 SET_PMU_OWNER(task, ctx);
4351 
4352                 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4353         } else {
4354                 /*
4355                  * when not current, task MUST be stopped, so this is safe
4356                  */
4357                 regs = task_pt_regs(task);
4358 
4359                 /* force a full reload */
4360                 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4361                 SET_LAST_CPU(ctx, -1);
4362 
4363                 /* initial saved psr (stopped) */
4364                 ctx->ctx_saved_psr_up = 0UL;
4365                 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4366         }
4367 
4368         ret = 0;
4369 
4370 error_unres:
4371         if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4372 error:
4373         /*
4374          * we must undo the dbregs setting (for system-wide)
4375          */
4376         if (ret && set_dbregs) {
4377                 LOCK_PFS(flags);
4378                 pfm_sessions.pfs_sys_use_dbregs--;
4379                 UNLOCK_PFS(flags);
4380         }
4381         /*
4382          * release task, there is now a link with the context
4383          */
4384         if (is_system == 0 && task != current) {
4385                 pfm_put_task(task);
4386 
4387                 if (ret == 0) {
4388                         ret = pfm_check_task_exist(ctx);
4389                         if (ret) {
4390                                 ctx->ctx_state = PFM_CTX_UNLOADED;
4391                                 ctx->ctx_task  = NULL;
4392                         }
4393                 }
4394         }
4395         return ret;
4396 }
4397 
4398 /*
4399  * in this function, we do not need to increase the use count
4400  * for the task via get_task_struct(), because we hold the
4401  * context lock. If the task were to disappear while having
4402  * a context attached, it would go through pfm_exit_thread()
4403  * which also grabs the context lock  and would therefore be blocked
4404  * until we are here.
4405  */
4406 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4407 
4408 static int
4409 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4410 {
4411         struct task_struct *task = PFM_CTX_TASK(ctx);
4412         struct pt_regs *tregs;
4413         int prev_state, is_system;
4414         int ret;
4415 
4416         DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4417 
4418         prev_state = ctx->ctx_state;
4419         is_system  = ctx->ctx_fl_system;
4420 
4421         /*
4422          * unload only when necessary
4423          */
4424         if (prev_state == PFM_CTX_UNLOADED) {
4425                 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4426                 return 0;
4427         }
4428 
4429         /*
4430          * clear psr and dcr bits
4431          */
4432         ret = pfm_stop(ctx, NULL, 0, regs);
4433         if (ret) return ret;
4434 
4435         ctx->ctx_state = PFM_CTX_UNLOADED;
4436 
4437         /*
4438          * in system mode, we need to update the PMU directly
4439          * and the user level state of the caller, which may not
4440          * necessarily be the creator of the context.
4441          */
4442         if (is_system) {
4443 
4444                 /*
4445                  * Update cpuinfo
4446                  *
4447                  * local PMU is taken care of in pfm_stop()
4448                  */
4449                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4450                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4451 
4452                 /*
4453                  * save PMDs in context
4454                  * release ownership
4455                  */
4456                 pfm_flush_pmds(current, ctx);
4457 
4458                 /*
4459                  * at this point we are done with the PMU
4460                  * so we can unreserve the resource.
4461                  */
4462                 if (prev_state != PFM_CTX_ZOMBIE) 
4463                         pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4464 
4465                 /*
4466                  * disconnect context from task
4467                  */
4468                 task->thread.pfm_context = NULL;
4469                 /*
4470                  * disconnect task from context
4471                  */
4472                 ctx->ctx_task = NULL;
4473 
4474                 /*
4475                  * There is nothing more to cleanup here.
4476                  */
4477                 return 0;
4478         }
4479 
4480         /*
4481          * per-task mode
4482          */
4483         tregs = task == current ? regs : task_pt_regs(task);
4484 
4485         if (task == current) {
4486                 /*
4487                  * cancel user level control
4488                  */
4489                 ia64_psr(regs)->sp = 1;
4490 
4491                 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4492         }
4493         /*
4494          * save PMDs to context
4495          * release ownership
4496          */
4497         pfm_flush_pmds(task, ctx);
4498 
4499         /*
4500          * at this point we are done with the PMU
4501          * so we can unreserve the resource.
4502          *
4503          * when state was ZOMBIE, we have already unreserved.
4504          */
4505         if (prev_state != PFM_CTX_ZOMBIE) 
4506                 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4507 
4508         /*
4509          * reset activation counter and psr
4510          */
4511         ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4512         SET_LAST_CPU(ctx, -1);
4513 
4514         /*
4515          * PMU state will not be restored
4516          */
4517         task->thread.flags &= ~IA64_THREAD_PM_VALID;
4518 
4519         /*
4520          * break links between context and task
4521          */
4522         task->thread.pfm_context  = NULL;
4523         ctx->ctx_task             = NULL;
4524 
4525         PFM_SET_WORK_PENDING(task, 0);
4526 
4527         ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4528         ctx->ctx_fl_can_restart  = 0;
4529         ctx->ctx_fl_going_zombie = 0;
4530 
4531         DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4532 
4533         return 0;
4534 }
4535 
4536 
4537 /*
4538  * called only from exit_thread()
4539  * we come here only if the task has a context attached (loaded or masked)
4540  */
4541 void
4542 pfm_exit_thread(struct task_struct *task)
4543 {
4544         pfm_context_t *ctx;
4545         unsigned long flags;
4546         struct pt_regs *regs = task_pt_regs(task);
4547         int ret, state;
4548         int free_ok = 0;
4549 
4550         ctx = PFM_GET_CTX(task);
4551 
4552         PROTECT_CTX(ctx, flags);
4553 
4554         DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4555 
4556         state = ctx->ctx_state;
4557         switch(state) {
4558                 case PFM_CTX_UNLOADED:
4559                         /*
4560                          * only comes to this function if pfm_context is not NULL, i.e., cannot
4561                          * be in unloaded state
4562                          */
4563                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4564                         break;
4565                 case PFM_CTX_LOADED:
4566                 case PFM_CTX_MASKED:
4567                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4568                         if (ret) {
4569                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4570                         }
4571                         DPRINT(("ctx unloaded for current state was %d\n", state));
4572 
4573                         pfm_end_notify_user(ctx);
4574                         break;
4575                 case PFM_CTX_ZOMBIE:
4576                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4577                         if (ret) {
4578                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4579                         }
4580                         free_ok = 1;
4581                         break;
4582                 default:
4583                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4584                         break;
4585         }
4586         UNPROTECT_CTX(ctx, flags);
4587 
4588         { u64 psr = pfm_get_psr();
4589           BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4590           BUG_ON(GET_PMU_OWNER());
4591           BUG_ON(ia64_psr(regs)->up);
4592           BUG_ON(ia64_psr(regs)->pp);
4593         }
4594 
4595         /*
4596          * All memory free operations (especially for vmalloc'ed memory)
4597          * MUST be done with interrupts ENABLED.
4598          */
4599         if (free_ok) pfm_context_free(ctx);
4600 }
4601 
4602 /*
4603  * functions MUST be listed in the increasing order of their index (see permfon.h)
4604  */
4605 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4606 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4607 #define PFM_CMD_PCLRWS  (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4608 #define PFM_CMD_PCLRW   (PFM_CMD_FD|PFM_CMD_ARG_RW)
4609 #define PFM_CMD_NONE    { NULL, "no-cmd", 0, 0, 0, NULL}
4610 
4611 static pfm_cmd_desc_t pfm_cmd_tab[]={
4612 /* 0  */PFM_CMD_NONE,
4613 /* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4614 /* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4615 /* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4616 /* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4617 /* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4618 /* 6  */PFM_CMD_NONE,
4619 /* 7  */PFM_CMD_NONE,
4620 /* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4621 /* 9  */PFM_CMD_NONE,
4622 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4623 /* 11 */PFM_CMD_NONE,
4624 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4625 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4626 /* 14 */PFM_CMD_NONE,
4627 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4628 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4629 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4630 /* 18 */PFM_CMD_NONE,
4631 /* 19 */PFM_CMD_NONE,
4632 /* 20 */PFM_CMD_NONE,
4633 /* 21 */PFM_CMD_NONE,
4634 /* 22 */PFM_CMD_NONE,
4635 /* 23 */PFM_CMD_NONE,
4636 /* 24 */PFM_CMD_NONE,
4637 /* 25 */PFM_CMD_NONE,
4638 /* 26 */PFM_CMD_NONE,
4639 /* 27 */PFM_CMD_NONE,
4640 /* 28 */PFM_CMD_NONE,
4641 /* 29 */PFM_CMD_NONE,
4642 /* 30 */PFM_CMD_NONE,
4643 /* 31 */PFM_CMD_NONE,
4644 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4645 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4646 };
4647 #define PFM_CMD_COUNT   (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4648 
4649 static int
4650 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4651 {
4652         struct task_struct *task;
4653         int state, old_state;
4654 
4655 recheck:
4656         state = ctx->ctx_state;
4657         task  = ctx->ctx_task;
4658 
4659         if (task == NULL) {
4660                 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4661                 return 0;
4662         }
4663 
4664         DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4665                 ctx->ctx_fd,
4666                 state,
4667                 task_pid_nr(task),
4668                 task->state, PFM_CMD_STOPPED(cmd)));
4669 
4670         /*
4671          * self-monitoring always ok.
4672          *
4673          * for system-wide the caller can either be the creator of the
4674          * context (to one to which the context is attached to) OR
4675          * a task running on the same CPU as the session.
4676          */
4677         if (task == current || ctx->ctx_fl_system) return 0;
4678 
4679         /*
4680          * we are monitoring another thread
4681          */
4682         switch(state) {
4683                 case PFM_CTX_UNLOADED:
4684                         /*
4685                          * if context is UNLOADED we are safe to go
4686                          */
4687                         return 0;
4688                 case PFM_CTX_ZOMBIE:
4689                         /*
4690                          * no command can operate on a zombie context
4691                          */
4692                         DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4693                         return -EINVAL;
4694                 case PFM_CTX_MASKED:
4695                         /*
4696                          * PMU state has been saved to software even though
4697                          * the thread may still be running.
4698                          */
4699                         if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4700         }
4701 
4702         /*
4703          * context is LOADED or MASKED. Some commands may need to have 
4704          * the task stopped.
4705          *
4706          * We could lift this restriction for UP but it would mean that
4707          * the user has no guarantee the task would not run between
4708          * two successive calls to perfmonctl(). That's probably OK.
4709          * If this user wants to ensure the task does not run, then
4710          * the task must be stopped.
4711          */
4712         if (PFM_CMD_STOPPED(cmd)) {
4713                 if (!task_is_stopped_or_traced(task)) {
4714                         DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4715                         return -EBUSY;
4716                 }
4717                 /*
4718                  * task is now stopped, wait for ctxsw out
4719                  *
4720                  * This is an interesting point in the code.
4721                  * We need to unprotect the context because
4722                  * the pfm_save_regs() routines needs to grab
4723                  * the same lock. There are danger in doing
4724                  * this because it leaves a window open for
4725                  * another task to get access to the context
4726                  * and possibly change its state. The one thing
4727                  * that is not possible is for the context to disappear
4728                  * because we are protected by the VFS layer, i.e.,
4729                  * get_fd()/put_fd().
4730                  */
4731                 old_state = state;
4732 
4733                 UNPROTECT_CTX(ctx, flags);
4734 
4735                 wait_task_inactive(task, 0);
4736 
4737                 PROTECT_CTX(ctx, flags);
4738 
4739                 /*
4740                  * we must recheck to verify if state has changed
4741                  */
4742                 if (ctx->ctx_state != old_state) {
4743                         DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4744                         goto recheck;
4745                 }
4746         }
4747         return 0;
4748 }
4749 
4750 /*
4751  * system-call entry point (must return long)
4752  */
4753 asmlinkage long
4754 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4755 {
4756         struct fd f = {NULL, 0};
4757         pfm_context_t *ctx = NULL;
4758         unsigned long flags = 0UL;
4759         void *args_k = NULL;
4760         long ret; /* will expand int return types */
4761         size_t base_sz, sz, xtra_sz = 0;
4762         int narg, completed_args = 0, call_made = 0, cmd_flags;
4763         int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4764         int (*getsize)(void *arg, size_t *sz);
4765 #define PFM_MAX_ARGSIZE 4096
4766 
4767         /*
4768          * reject any call if perfmon was disabled at initialization
4769          */
4770         if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4771 
4772         if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4773                 DPRINT(("invalid cmd=%d\n", cmd));
4774                 return -EINVAL;
4775         }
4776 
4777         func      = pfm_cmd_tab[cmd].cmd_func;
4778         narg      = pfm_cmd_tab[cmd].cmd_narg;
4779         base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4780         getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4781         cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4782 
4783         if (unlikely(func == NULL)) {
4784                 DPRINT(("invalid cmd=%d\n", cmd));
4785                 return -EINVAL;
4786         }
4787 
4788         DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4789                 PFM_CMD_NAME(cmd),
4790                 cmd,
4791                 narg,
4792                 base_sz,
4793                 count));
4794 
4795         /*
4796          * check if number of arguments matches what the command expects
4797          */
4798         if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4799                 return -EINVAL;
4800 
4801 restart_args:
4802         sz = xtra_sz + base_sz*count;
4803         /*
4804          * limit abuse to min page size
4805          */
4806         if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4807                 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4808                 return -E2BIG;
4809         }
4810 
4811         /*
4812          * allocate default-sized argument buffer
4813          */
4814         if (likely(count && args_k == NULL)) {
4815                 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4816                 if (args_k == NULL) return -ENOMEM;
4817         }
4818 
4819         ret = -EFAULT;
4820 
4821         /*
4822          * copy arguments
4823          *
4824          * assume sz = 0 for command without parameters
4825          */
4826         if (sz && copy_from_user(args_k, arg, sz)) {
4827                 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4828                 goto error_args;
4829         }
4830 
4831         /*
4832          * check if command supports extra parameters
4833          */
4834         if (completed_args == 0 && getsize) {
4835                 /*
4836                  * get extra parameters size (based on main argument)
4837                  */
4838                 ret = (*getsize)(args_k, &xtra_sz);
4839                 if (ret) goto error_args;
4840 
4841                 completed_args = 1;
4842 
4843                 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4844 
4845                 /* retry if necessary */
4846                 if (likely(xtra_sz)) goto restart_args;
4847         }
4848 
4849         if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4850 
4851         ret = -EBADF;
4852 
4853         f = fdget(fd);
4854         if (unlikely(f.file == NULL)) {
4855                 DPRINT(("invalid fd %d\n", fd));
4856                 goto error_args;
4857         }
4858         if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4859                 DPRINT(("fd %d not related to perfmon\n", fd));
4860                 goto error_args;
4861         }
4862 
4863         ctx = f.file->private_data;
4864         if (unlikely(ctx == NULL)) {
4865                 DPRINT(("no context for fd %d\n", fd));
4866                 goto error_args;
4867         }
4868         prefetch(&ctx->ctx_state);
4869 
4870         PROTECT_CTX(ctx, flags);
4871 
4872         /*
4873          * check task is stopped
4874          */
4875         ret = pfm_check_task_state(ctx, cmd, flags);
4876         if (unlikely(ret)) goto abort_locked;
4877 
4878 skip_fd:
4879         ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4880 
4881         call_made = 1;
4882 
4883 abort_locked:
4884         if (likely(ctx)) {
4885                 DPRINT(("context unlocked\n"));
4886                 UNPROTECT_CTX(ctx, flags);
4887         }
4888 
4889         /* copy argument back to user, if needed */
4890         if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4891 
4892 error_args:
4893         if (f.file)
4894                 fdput(f);
4895 
4896         kfree(args_k);
4897 
4898         DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4899 
4900         return ret;
4901 }
4902 
4903 static void
4904 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4905 {
4906         pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4907         pfm_ovfl_ctrl_t rst_ctrl;
4908         int state;
4909         int ret = 0;
4910 
4911         state = ctx->ctx_state;
4912         /*
4913          * Unlock sampling buffer and reset index atomically
4914          * XXX: not really needed when blocking
4915          */
4916         if (CTX_HAS_SMPL(ctx)) {
4917 
4918                 rst_ctrl.bits.mask_monitoring = 0;
4919                 rst_ctrl.bits.reset_ovfl_pmds = 0;
4920 
4921                 if (state == PFM_CTX_LOADED)
4922                         ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4923                 else
4924                         ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4925         } else {
4926                 rst_ctrl.bits.mask_monitoring = 0;
4927                 rst_ctrl.bits.reset_ovfl_pmds = 1;
4928         }
4929 
4930         if (ret == 0) {
4931                 if (rst_ctrl.bits.reset_ovfl_pmds) {
4932                         pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4933                 }
4934                 if (rst_ctrl.bits.mask_monitoring == 0) {
4935                         DPRINT(("resuming monitoring\n"));
4936                         if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4937                 } else {
4938                         DPRINT(("stopping monitoring\n"));
4939                         //pfm_stop_monitoring(current, regs);
4940                 }
4941                 ctx->ctx_state = PFM_CTX_LOADED;
4942         }
4943 }
4944 
4945 /*
4946  * context MUST BE LOCKED when calling
4947  * can only be called for current
4948  */
4949 static void
4950 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4951 {
4952         int ret;
4953 
4954         DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4955 
4956         ret = pfm_context_unload(ctx, NULL, 0, regs);
4957         if (ret) {
4958                 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4959         }
4960 
4961         /*
4962          * and wakeup controlling task, indicating we are now disconnected
4963          */
4964         wake_up_interruptible(&ctx->ctx_zombieq);
4965 
4966         /*
4967          * given that context is still locked, the controlling
4968          * task will only get access when we return from
4969          * pfm_handle_work().
4970          */
4971 }
4972 
4973 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4974 
4975  /*
4976   * pfm_handle_work() can be called with interrupts enabled
4977   * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4978   * call may sleep, therefore we must re-enable interrupts
4979   * to avoid deadlocks. It is safe to do so because this function
4980   * is called ONLY when returning to user level (pUStk=1), in which case
4981   * there is no risk of kernel stack overflow due to deep
4982   * interrupt nesting.
4983   */
4984 void
4985 pfm_handle_work(void)
4986 {
4987         pfm_context_t *ctx;
4988         struct pt_regs *regs;
4989         unsigned long flags, dummy_flags;
4990         unsigned long ovfl_regs;
4991         unsigned int reason;
4992         int ret;
4993 
4994         ctx = PFM_GET_CTX(current);
4995         if (ctx == NULL) {
4996                 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
4997                         task_pid_nr(current));
4998                 return;
4999         }
5000 
5001         PROTECT_CTX(ctx, flags);
5002 
5003         PFM_SET_WORK_PENDING(current, 0);
5004 
5005         regs = task_pt_regs(current);
5006 
5007         /*
5008          * extract reason for being here and clear
5009          */
5010         reason = ctx->ctx_fl_trap_reason;
5011         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5012         ovfl_regs = ctx->ctx_ovfl_regs[0];
5013 
5014         DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5015 
5016         /*
5017          * must be done before we check for simple-reset mode
5018          */
5019         if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5020                 goto do_zombie;
5021 
5022         //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5023         if (reason == PFM_TRAP_REASON_RESET)
5024                 goto skip_blocking;
5025 
5026         /*
5027          * restore interrupt mask to what it was on entry.
5028          * Could be enabled/diasbled.
5029          */
5030         UNPROTECT_CTX(ctx, flags);
5031 
5032         /*
5033          * force interrupt enable because of down_interruptible()
5034          */
5035         local_irq_enable();
5036 
5037         DPRINT(("before block sleeping\n"));
5038 
5039         /*
5040          * may go through without blocking on SMP systems
5041          * if restart has been received already by the time we call down()
5042          */
5043         ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5044 
5045         DPRINT(("after block sleeping ret=%d\n", ret));
5046 
5047         /*
5048          * lock context and mask interrupts again
5049          * We save flags into a dummy because we may have
5050          * altered interrupts mask compared to entry in this
5051          * function.
5052          */
5053         PROTECT_CTX(ctx, dummy_flags);
5054 
5055         /*
5056          * we need to read the ovfl_regs only after wake-up
5057          * because we may have had pfm_write_pmds() in between
5058          * and that can changed PMD values and therefore 
5059          * ovfl_regs is reset for these new PMD values.
5060          */
5061         ovfl_regs = ctx->ctx_ovfl_regs[0];
5062 
5063         if (ctx->ctx_fl_going_zombie) {
5064 do_zombie:
5065                 DPRINT(("context is zombie, bailing out\n"));
5066                 pfm_context_force_terminate(ctx, regs);
5067                 goto nothing_to_do;
5068         }
5069         /*
5070          * in case of interruption of down() we don't restart anything
5071          */
5072         if (ret < 0)
5073                 goto nothing_to_do;
5074 
5075 skip_blocking:
5076         pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5077         ctx->ctx_ovfl_regs[0] = 0UL;
5078 
5079 nothing_to_do:
5080         /*
5081          * restore flags as they were upon entry
5082          */
5083         UNPROTECT_CTX(ctx, flags);
5084 }
5085 
5086 static int
5087 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5088 {
5089         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5090                 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5091                 return 0;
5092         }
5093 
5094         DPRINT(("waking up somebody\n"));
5095 
5096         if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5097 
5098         /*
5099          * safe, we are not in intr handler, nor in ctxsw when
5100          * we come here
5101          */
5102         kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5103 
5104         return 0;
5105 }
5106 
5107 static int
5108 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5109 {
5110         pfm_msg_t *msg = NULL;
5111 
5112         if (ctx->ctx_fl_no_msg == 0) {
5113                 msg = pfm_get_new_msg(ctx);
5114                 if (msg == NULL) {
5115                         printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5116                         return -1;
5117                 }
5118 
5119                 msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5120                 msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5121                 msg->pfm_ovfl_msg.msg_active_set   = 0;
5122                 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5123                 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5124                 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5125                 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5126                 msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5127         }
5128 
5129         DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5130                 msg,
5131                 ctx->ctx_fl_no_msg,
5132                 ctx->ctx_fd,
5133                 ovfl_pmds));
5134 
5135         return pfm_notify_user(ctx, msg);
5136 }
5137 
5138 static int
5139 pfm_end_notify_user(pfm_context_t *ctx)
5140 {
5141         pfm_msg_t *msg;
5142 
5143         msg = pfm_get_new_msg(ctx);
5144         if (msg == NULL) {
5145                 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5146                 return -1;
5147         }
5148         /* no leak */
5149         memset(msg, 0, sizeof(*msg));
5150 
5151         msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5152         msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5153         msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5154 
5155         DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5156                 msg,
5157                 ctx->ctx_fl_no_msg,
5158                 ctx->ctx_fd));
5159 
5160         return pfm_notify_user(ctx, msg);
5161 }
5162 
5163 /*
5164  * main overflow processing routine.
5165  * it can be called from the interrupt path or explicitly during the context switch code
5166  */
5167 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5168                                 unsigned long pmc0, struct pt_regs *regs)
5169 {
5170         pfm_ovfl_arg_t *ovfl_arg;
5171         unsigned long mask;
5172         unsigned long old_val, ovfl_val, new_val;
5173         unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5174         unsigned long tstamp;
5175         pfm_ovfl_ctrl_t ovfl_ctrl;
5176         unsigned int i, has_smpl;
5177         int must_notify = 0;
5178 
5179         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5180 
5181         /*
5182          * sanity test. Should never happen
5183          */
5184         if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5185 
5186         tstamp   = ia64_get_itc();
5187         mask     = pmc0 >> PMU_FIRST_COUNTER;
5188         ovfl_val = pmu_conf->ovfl_val;
5189         has_smpl = CTX_HAS_SMPL(ctx);
5190 
5191         DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5192                      "used_pmds=0x%lx\n",
5193                         pmc0,
5194                         task ? task_pid_nr(task): -1,
5195                         (regs ? regs->cr_iip : 0),
5196                         CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5197                         ctx->ctx_used_pmds[0]));
5198 
5199 
5200         /*
5201          * first we update the virtual counters
5202          * assume there was a prior ia64_srlz_d() issued
5203          */
5204         for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5205 
5206                 /* skip pmd which did not overflow */
5207                 if ((mask & 0x1) == 0) continue;
5208 
5209                 /*
5210                  * Note that the pmd is not necessarily 0 at this point as qualified events
5211                  * may have happened before the PMU was frozen. The residual count is not
5212                  * taken into consideration here but will be with any read of the pmd via
5213                  * pfm_read_pmds().
5214                  */
5215                 old_val              = new_val = ctx->ctx_pmds[i].val;
5216                 new_val             += 1 + ovfl_val;
5217                 ctx->ctx_pmds[i].val = new_val;
5218 
5219                 /*
5220                  * check for overflow condition
5221                  */
5222                 if (likely(old_val > new_val)) {
5223                         ovfl_pmds |= 1UL << i;
5224                         if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5225                 }
5226 
5227                 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5228                         i,
5229                         new_val,
5230                         old_val,
5231                         ia64_get_pmd(i) & ovfl_val,
5232                         ovfl_pmds,
5233                         ovfl_notify));
5234         }
5235 
5236         /*
5237          * there was no 64-bit overflow, nothing else to do
5238          */
5239         if (ovfl_pmds == 0UL) return;
5240 
5241         /* 
5242          * reset all control bits
5243          */
5244         ovfl_ctrl.val = 0;
5245         reset_pmds    = 0UL;
5246 
5247         /*
5248          * if a sampling format module exists, then we "cache" the overflow by 
5249          * calling the module's handler() routine.
5250          */
5251         if (has_smpl) {
5252                 unsigned long start_cycles, end_cycles;
5253                 unsigned long pmd_mask;
5254                 int j, k, ret = 0;
5255                 int this_cpu = smp_processor_id();
5256 
5257                 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5258                 ovfl_arg = &ctx->ctx_ovfl_arg;
5259 
5260                 prefetch(ctx->ctx_smpl_hdr);
5261 
5262                 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5263 
5264                         mask = 1UL << i;
5265 
5266                         if ((pmd_mask & 0x1) == 0) continue;
5267 
5268                         ovfl_arg->ovfl_pmd      = (unsigned char )i;
5269                         ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5270                         ovfl_arg->active_set    = 0;
5271                         ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5272                         ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5273 
5274                         ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5275                         ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5276                         ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5277 
5278                         /*
5279                          * copy values of pmds of interest. Sampling format may copy them
5280                          * into sampling buffer.
5281                          */
5282                         if (smpl_pmds) {
5283                                 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5284                                         if ((smpl_pmds & 0x1) == 0) continue;
5285                                         ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5286                                         DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5287                                 }
5288                         }
5289 
5290                         pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5291 
5292                         start_cycles = ia64_get_itc();
5293 
5294                         /*
5295                          * call custom buffer format record (handler) routine
5296                          */
5297                         ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5298 
5299                         end_cycles = ia64_get_itc();
5300 
5301                         /*
5302                          * For those controls, we take the union because they have
5303                          * an all or nothing behavior.
5304                          */
5305                         ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5306                         ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5307                         ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5308                         /*
5309                          * build the bitmask of pmds to reset now
5310                          */
5311                         if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5312 
5313                         pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5314                 }
5315                 /*
5316                  * when the module cannot handle the rest of the overflows, we abort right here
5317                  */
5318                 if (ret && pmd_mask) {
5319                         DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5320                                 pmd_mask<<PMU_FIRST_COUNTER));
5321                 }
5322                 /*
5323                  * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5324                  */
5325                 ovfl_pmds &= ~reset_pmds;
5326         } else {
5327                 /*
5328                  * when no sampling module is used, then the default
5329                  * is to notify on overflow if requested by user
5330                  */
5331                 ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5332                 ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5333                 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5334                 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5335                 /*
5336                  * if needed, we reset all overflowed pmds
5337                  */
5338                 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5339         }
5340 
5341         DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5342 
5343         /*
5344          * reset the requested PMD registers using the short reset values
5345          */
5346         if (reset_pmds) {
5347                 unsigned long bm = reset_pmds;
5348                 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5349         }
5350 
5351         if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5352                 /*
5353                  * keep track of what to reset when unblocking
5354                  */
5355                 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5356 
5357                 /*
5358                  * check for blocking context 
5359                  */
5360                 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5361 
5362                         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5363 
5364                         /*
5365                          * set the perfmon specific checking pending work for the task
5366                          */
5367                         PFM_SET_WORK_PENDING(task, 1);
5368 
5369                         /*
5370                          * when coming from ctxsw, current still points to the
5371                          * previous task, therefore we must work with task and not current.
5372                          */
5373                         set_notify_resume(task);
5374                 }
5375                 /*
5376                  * defer until state is changed (shorten spin window). the context is locked
5377                  * anyway, so the signal receiver would come spin for nothing.
5378                  */
5379                 must_notify = 1;
5380         }
5381 
5382         DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5383                         GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5384                         PFM_GET_WORK_PENDING(task),
5385                         ctx->ctx_fl_trap_reason,
5386                         ovfl_pmds,
5387                         ovfl_notify,
5388                         ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5389         /*
5390          * in case monitoring must be stopped, we toggle the psr bits
5391          */
5392         if (ovfl_ctrl.bits.mask_monitoring) {
5393                 pfm_mask_monitoring(task);
5394                 ctx->ctx_state = PFM_CTX_MASKED;
5395                 ctx->ctx_fl_can_restart = 1;
5396         }
5397 
5398         /*
5399          * send notification now
5400          */
5401         if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5402 
5403         return;
5404 
5405 sanity_check:
5406         printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5407                         smp_processor_id(),
5408                         task ? task_pid_nr(task) : -1,
5409                         pmc0);
5410         return;
5411 
5412 stop_monitoring:
5413         /*
5414          * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5415          * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5416          * come here as zombie only if the task is the current task. In which case, we
5417          * can access the PMU  hardware directly.
5418          *
5419          * Note that zombies do have PM_VALID set. So here we do the minimal.
5420          *
5421          * In case the context was zombified it could not be reclaimed at the time
5422          * the monitoring program exited. At this point, the PMU reservation has been
5423          * returned, the sampiing buffer has been freed. We must convert this call
5424          * into a spurious interrupt. However, we must also avoid infinite overflows
5425          * by stopping monitoring for this task. We can only come here for a per-task
5426          * context. All we need to do is to stop monitoring using the psr bits which
5427          * are always task private. By re-enabling secure montioring, we ensure that
5428          * the monitored task will not be able to re-activate monitoring.
5429          * The task will eventually be context switched out, at which point the context
5430          * will be reclaimed (that includes releasing ownership of the PMU).
5431          *
5432          * So there might be a window of time where the number of per-task session is zero
5433          * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5434          * context. This is safe because if a per-task session comes in, it will push this one
5435          * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5436          * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5437          * also push our zombie context out.
5438          *
5439          * Overall pretty hairy stuff....
5440          */
5441         DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5442         pfm_clear_psr_up();
5443         ia64_psr(regs)->up = 0;
5444         ia64_psr(regs)->sp = 1;
5445         return;
5446 }
5447 
5448 static int
5449 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5450 {
5451         struct task_struct *task;
5452         pfm_context_t *ctx;
5453         unsigned long flags;
5454         u64 pmc0;
5455         int this_cpu = smp_processor_id();
5456         int retval = 0;
5457 
5458         pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5459 
5460         /*
5461          * srlz.d done before arriving here
5462          */
5463         pmc0 = ia64_get_pmc(0);
5464 
5465         task = GET_PMU_OWNER();
5466         ctx  = GET_PMU_CTX();
5467 
5468         /*
5469          * if we have some pending bits set
5470          * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5471          */
5472         if (PMC0_HAS_OVFL(pmc0) && task) {
5473                 /*
5474                  * we assume that pmc0.fr is always set here
5475                  */
5476 
5477                 /* sanity check */
5478                 if (!ctx) goto report_spurious1;
5479 
5480                 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5481                         goto report_spurious2;
5482 
5483                 PROTECT_CTX_NOPRINT(ctx, flags);
5484 
5485                 pfm_overflow_handler(task, ctx, pmc0, regs);
5486 
5487                 UNPROTECT_CTX_NOPRINT(ctx, flags);
5488 
5489         } else {
5490                 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5491                 retval = -1;
5492         }
5493         /*
5494          * keep it unfrozen at all times
5495          */
5496         pfm_unfreeze_pmu();
5497 
5498         return retval;
5499 
5500 report_spurious1:
5501         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5502                 this_cpu, task_pid_nr(task));
5503         pfm_unfreeze_pmu();
5504         return -1;
5505 report_spurious2:
5506         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5507                 this_cpu, 
5508                 task_pid_nr(task));
5509         pfm_unfreeze_pmu();
5510         return -1;
5511 }
5512 
5513 static irqreturn_t
5514 pfm_interrupt_handler(int irq, void *arg)
5515 {
5516         unsigned long start_cycles, total_cycles;
5517         unsigned long min, max;
5518         int this_cpu;
5519         int ret;
5520         struct pt_regs *regs = get_irq_regs();
5521 
5522         this_cpu = get_cpu();
5523         if (likely(!pfm_alt_intr_handler)) {
5524                 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5525                 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5526 
5527                 start_cycles = ia64_get_itc();
5528 
5529                 ret = pfm_do_interrupt_handler(arg, regs);
5530 
5531                 total_cycles = ia64_get_itc();
5532 
5533                 /*
5534                  * don't measure spurious interrupts
5535                  */
5536                 if (likely(ret == 0)) {
5537                         total_cycles -= start_cycles;
5538 
5539                         if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5540                         if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5541 
5542                         pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5543                 }
5544         }
5545         else {
5546                 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5547         }
5548 
5549         put_cpu();
5550         return IRQ_HANDLED;
5551 }
5552 
5553 /*
5554  * /proc/perfmon interface, for debug only
5555  */
5556 
5557 #define PFM_PROC_SHOW_HEADER    ((void *)(long)nr_cpu_ids+1)
5558 
5559 static void *
5560 pfm_proc_start(struct seq_file *m, loff_t *pos)
5561 {
5562         if (*pos == 0) {
5563                 return PFM_PROC_SHOW_HEADER;
5564         }
5565 
5566         while (*pos <= nr_cpu_ids) {
5567                 if (cpu_online(*pos - 1)) {
5568                         return (void *)*pos;
5569                 }
5570                 ++*pos;
5571         }
5572         return NULL;
5573 }
5574 
5575 static void *
5576 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5577 {
5578         ++*pos;
5579         return pfm_proc_start(m, pos);
5580 }
5581 
5582 static void
5583 pfm_proc_stop(struct seq_file *m, void *v)
5584 {
5585 }
5586 
5587 static void
5588 pfm_proc_show_header(struct seq_file *m)
5589 {
5590         struct list_head * pos;
5591         pfm_buffer_fmt_t * entry;
5592         unsigned long flags;
5593 
5594         seq_printf(m,
5595                 "perfmon version           : %u.%u\n"
5596                 "model                     : %s\n"
5597                 "fastctxsw                 : %s\n"
5598                 "expert mode               : %s\n"
5599                 "ovfl_mask                 : 0x%lx\n"
5600                 "PMU flags                 : 0x%x\n",
5601                 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5602                 pmu_conf->pmu_name,
5603                 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5604                 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5605                 pmu_conf->ovfl_val,
5606                 pmu_conf->flags);
5607 
5608         LOCK_PFS(flags);
5609 
5610         seq_printf(m,
5611                 "proc_sessions             : %u\n"
5612                 "sys_sessions              : %u\n"
5613                 "sys_use_dbregs            : %u\n"
5614                 "ptrace_use_dbregs         : %u\n",
5615                 pfm_sessions.pfs_task_sessions,
5616                 pfm_sessions.pfs_sys_sessions,
5617                 pfm_sessions.pfs_sys_use_dbregs,
5618                 pfm_sessions.pfs_ptrace_use_dbregs);
5619 
5620         UNLOCK_PFS(flags);
5621 
5622         spin_lock(&pfm_buffer_fmt_lock);
5623 
5624         list_for_each(pos, &pfm_buffer_fmt_list) {
5625                 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5626                 seq_printf(m, "format                    : %16phD %s\n",
5627                            entry->fmt_uuid, entry->fmt_name);
5628         }
5629         spin_unlock(&pfm_buffer_fmt_lock);
5630 
5631 }
5632 
5633 static int
5634 pfm_proc_show(struct seq_file *m, void *v)
5635 {
5636         unsigned long psr;
5637         unsigned int i;
5638         int cpu;
5639 
5640         if (v == PFM_PROC_SHOW_HEADER) {
5641                 pfm_proc_show_header(m);
5642                 return 0;
5643         }
5644 
5645         /* show info for CPU (v - 1) */
5646 
5647         cpu = (long)v - 1;
5648         seq_printf(m,
5649                 "CPU%-2d overflow intrs      : %lu\n"
5650                 "CPU%-2d overflow cycles     : %lu\n"
5651                 "CPU%-2d overflow min        : %lu\n"
5652                 "CPU%-2d overflow max        : %lu\n"
5653                 "CPU%-2d smpl handler calls  : %lu\n"
5654                 "CPU%-2d smpl handler cycles : %lu\n"
5655                 "CPU%-2d spurious intrs      : %lu\n"
5656                 "CPU%-2d replay   intrs      : %lu\n"
5657                 "CPU%-2d syst_wide           : %d\n"
5658                 "CPU%-2d dcr_pp              : %d\n"
5659                 "CPU%-2d exclude idle        : %d\n"
5660                 "CPU%-2d owner               : %d\n"
5661                 "CPU%-2d context             : %p\n"
5662                 "CPU%-2d activations         : %lu\n",
5663                 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5664                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5665                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5666                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5667                 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5668                 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5669                 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5670                 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5671                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5672                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5673                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5674                 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5675                 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5676                 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5677 
5678         if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5679 
5680                 psr = pfm_get_psr();
5681 
5682                 ia64_srlz_d();
5683 
5684                 seq_printf(m, 
5685                         "CPU%-2d psr                 : 0x%lx\n"
5686                         "CPU%-2d pmc0                : 0x%lx\n", 
5687                         cpu, psr,
5688                         cpu, ia64_get_pmc(0));
5689 
5690                 for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5691                         if (PMC_IS_COUNTING(i) == 0) continue;
5692                         seq_printf(m, 
5693                                 "CPU%-2d pmc%u                : 0x%lx\n"
5694                                 "CPU%-2d pmd%u                : 0x%lx\n", 
5695                                 cpu, i, ia64_get_pmc(i),
5696                                 cpu, i, ia64_get_pmd(i));
5697                 }
5698         }
5699         return 0;
5700 }
5701 
5702 const struct seq_operations pfm_seq_ops = {
5703         .start =        pfm_proc_start,
5704         .next =         pfm_proc_next,
5705         .stop =         pfm_proc_stop,
5706         .show =         pfm_proc_show
5707 };
5708 
5709 /*
5710  * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5711  * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5712  * is active or inactive based on mode. We must rely on the value in
5713  * local_cpu_data->pfm_syst_info
5714  */
5715 void
5716 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5717 {
5718         struct pt_regs *regs;
5719         unsigned long dcr;
5720         unsigned long dcr_pp;
5721 
5722         dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5723 
5724         /*
5725          * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5726          * on every CPU, so we can rely on the pid to identify the idle task.
5727          */
5728         if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5729                 regs = task_pt_regs(task);
5730                 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5731                 return;
5732         }
5733         /*
5734          * if monitoring has started
5735          */
5736         if (dcr_pp) {
5737                 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5738                 /*
5739                  * context switching in?
5740                  */
5741                 if (is_ctxswin) {
5742                         /* mask monitoring for the idle task */
5743                         ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5744                         pfm_clear_psr_pp();
5745                         ia64_srlz_i();
5746                         return;
5747                 }
5748                 /*
5749                  * context switching out
5750                  * restore monitoring for next task
5751                  *
5752                  * Due to inlining this odd if-then-else construction generates
5753                  * better code.
5754                  */
5755                 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5756                 pfm_set_psr_pp();
5757                 ia64_srlz_i();
5758         }
5759 }
5760 
5761 #ifdef CONFIG_SMP
5762 
5763 static void
5764 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5765 {
5766         struct task_struct *task = ctx->ctx_task;
5767 
5768         ia64_psr(regs)->up = 0;
5769         ia64_psr(regs)->sp = 1;
5770 
5771         if (GET_PMU_OWNER() == task) {
5772                 DPRINT(("cleared ownership for [%d]\n",
5773                                         task_pid_nr(ctx->ctx_task)));
5774                 SET_PMU_OWNER(NULL, NULL);
5775         }
5776 
5777         /*
5778          * disconnect the task from the context and vice-versa
5779          */
5780         PFM_SET_WORK_PENDING(task, 0);
5781 
5782         task->thread.pfm_context  = NULL;
5783         task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5784 
5785         DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5786 }
5787 
5788 
5789 /*
5790  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5791  */
5792 void
5793 pfm_save_regs(struct task_struct *task)
5794 {
5795         pfm_context_t *ctx;
5796         unsigned long flags;
5797         u64 psr;
5798 
5799 
5800         ctx = PFM_GET_CTX(task);
5801         if (ctx == NULL) return;
5802 
5803         /*
5804          * we always come here with interrupts ALREADY disabled by
5805          * the scheduler. So we simply need to protect against concurrent
5806          * access, not CPU concurrency.
5807          */
5808         flags = pfm_protect_ctx_ctxsw(ctx);
5809 
5810         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5811                 struct pt_regs *regs = task_pt_regs(task);
5812 
5813                 pfm_clear_psr_up();
5814 
5815                 pfm_force_cleanup(ctx, regs);
5816 
5817                 BUG_ON(ctx->ctx_smpl_hdr);
5818 
5819                 pfm_unprotect_ctx_ctxsw(ctx, flags);
5820 
5821                 pfm_context_free(ctx);
5822                 return;
5823         }
5824 
5825         /*
5826          * save current PSR: needed because we modify it
5827          */
5828         ia64_srlz_d();
5829         psr = pfm_get_psr();
5830 
5831         BUG_ON(psr & (IA64_PSR_I));
5832 
5833         /*
5834          * stop monitoring:
5835          * This is the last instruction which may generate an overflow
5836          *
5837          * We do not need to set psr.sp because, it is irrelevant in kernel.
5838          * It will be restored from ipsr when going back to user level
5839          */
5840         pfm_clear_psr_up();
5841 
5842         /*
5843          * keep a copy of psr.up (for reload)
5844          */
5845         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5846 
5847         /*
5848          * release ownership of this PMU.
5849          * PM interrupts are masked, so nothing
5850          * can happen.
5851          */
5852         SET_PMU_OWNER(NULL, NULL);
5853 
5854         /*
5855          * we systematically save the PMD as we have no
5856          * guarantee we will be schedule at that same
5857          * CPU again.
5858          */
5859         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5860 
5861         /*
5862          * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5863          * we will need it on the restore path to check
5864          * for pending overflow.
5865          */
5866         ctx->th_pmcs[0] = ia64_get_pmc(0);
5867 
5868         /*
5869          * unfreeze PMU if had pending overflows
5870          */
5871         if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5872 
5873         /*
5874          * finally, allow context access.
5875          * interrupts will still be masked after this call.
5876          */
5877         pfm_unprotect_ctx_ctxsw(ctx, flags);
5878 }
5879 
5880 #else /* !CONFIG_SMP */
5881 void
5882 pfm_save_regs(struct task_struct *task)
5883 {
5884         pfm_context_t *ctx;
5885         u64 psr;
5886 
5887         ctx = PFM_GET_CTX(task);
5888         if (ctx == NULL) return;
5889 
5890         /*
5891          * save current PSR: needed because we modify it
5892          */
5893         psr = pfm_get_psr();
5894 
5895         BUG_ON(psr & (IA64_PSR_I));
5896 
5897         /*
5898          * stop monitoring:
5899          * This is the last instruction which may generate an overflow
5900          *
5901          * We do not need to set psr.sp because, it is irrelevant in kernel.
5902          * It will be restored from ipsr when going back to user level
5903          */
5904         pfm_clear_psr_up();
5905 
5906         /*
5907          * keep a copy of psr.up (for reload)
5908          */
5909         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5910 }
5911 
5912 static void
5913 pfm_lazy_save_regs (struct task_struct *task)
5914 {
5915         pfm_context_t *ctx;
5916         unsigned long flags;
5917 
5918         { u64 psr  = pfm_get_psr();
5919           BUG_ON(psr & IA64_PSR_UP);
5920         }
5921 
5922         ctx = PFM_GET_CTX(task);
5923 
5924         /*
5925          * we need to mask PMU overflow here to
5926          * make sure that we maintain pmc0 until
5927          * we save it. overflow interrupts are
5928          * treated as spurious if there is no
5929          * owner.
5930          *
5931          * XXX: I don't think this is necessary
5932          */
5933         PROTECT_CTX(ctx,flags);
5934 
5935         /*
5936          * release ownership of this PMU.
5937          * must be done before we save the registers.
5938          *
5939          * after this call any PMU interrupt is treated
5940          * as spurious.
5941          */
5942         SET_PMU_OWNER(NULL, NULL);
5943 
5944         /*
5945          * save all the pmds we use
5946          */
5947         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5948 
5949         /*
5950          * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5951          * it is needed to check for pended overflow
5952          * on the restore path
5953          */
5954         ctx->th_pmcs[0] = ia64_get_pmc(0);
5955 
5956         /*
5957          * unfreeze PMU if had pending overflows
5958          */
5959         if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5960 
5961         /*
5962          * now get can unmask PMU interrupts, they will
5963          * be treated as purely spurious and we will not
5964          * lose any information
5965          */
5966         UNPROTECT_CTX(ctx,flags);
5967 }
5968 #endif /* CONFIG_SMP */
5969 
5970 #ifdef CONFIG_SMP
5971 /*
5972  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5973  */
5974 void
5975 pfm_load_regs (struct task_struct *task)
5976 {
5977         pfm_context_t *ctx;
5978         unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
5979         unsigned long flags;
5980         u64 psr, psr_up;
5981         int need_irq_resend;
5982 
5983         ctx = PFM_GET_CTX(task);
5984         if (unlikely(ctx == NULL)) return;
5985 
5986         BUG_ON(GET_PMU_OWNER());
5987 
5988         /*
5989          * possible on unload
5990          */
5991         if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
5992 
5993         /*
5994          * we always come here with interrupts ALREADY disabled by
5995          * the scheduler. So we simply need to protect against concurrent
5996          * access, not CPU concurrency.
5997          */
5998         flags = pfm_protect_ctx_ctxsw(ctx);
5999         psr   = pfm_get_psr();
6000 
6001         need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6002 
6003         BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6004         BUG_ON(psr & IA64_PSR_I);
6005 
6006         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6007                 struct pt_regs *regs = task_pt_regs(task);
6008 
6009                 BUG_ON(ctx->ctx_smpl_hdr);
6010 
6011                 pfm_force_cleanup(ctx, regs);
6012 
6013                 pfm_unprotect_ctx_ctxsw(ctx, flags);
6014 
6015                 /*
6016                  * this one (kmalloc'ed) is fine with interrupts disabled
6017                  */
6018                 pfm_context_free(ctx);
6019 
6020                 return;
6021         }
6022 
6023         /*
6024          * we restore ALL the debug registers to avoid picking up
6025          * stale state.
6026          */
6027         if (ctx->ctx_fl_using_dbreg) {
6028                 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6029                 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6030         }
6031         /*
6032          * retrieve saved psr.up
6033          */
6034         psr_up = ctx->ctx_saved_psr_up;
6035 
6036         /*
6037          * if we were the last user of the PMU on that CPU,
6038          * then nothing to do except restore psr
6039          */
6040         if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6041 
6042                 /*
6043                  * retrieve partial reload masks (due to user modifications)
6044                  */
6045                 pmc_mask = ctx->ctx_reload_pmcs[0];
6046                 pmd_mask = ctx->ctx_reload_pmds[0];
6047 
6048         } else {
6049                 /*
6050                  * To avoid leaking information to the user level when psr.sp=0,
6051                  * we must reload ALL implemented pmds (even the ones we don't use).
6052                  * In the kernel we only allow PFM_READ_PMDS on registers which
6053                  * we initialized or requested (sampling) so there is no risk there.
6054                  */
6055                 pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6056 
6057                 /*
6058                  * ALL accessible PMCs are systematically reloaded, unused registers
6059                  * get their default (from pfm_reset_pmu_state()) values to avoid picking
6060                  * up stale configuration.
6061                  *
6062                  * PMC0 is never in the mask. It is always restored separately.
6063                  */
6064                 pmc_mask = ctx->ctx_all_pmcs[0];
6065         }
6066         /*
6067          * when context is MASKED, we will restore PMC with plm=0
6068          * and PMD with stale information, but that's ok, nothing
6069          * will be captured.
6070          *
6071          * XXX: optimize here
6072          */
6073         if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6074         if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6075 
607