2 * This file implements the perfmon-2 subsystem which is used
3 * to program the IA-64 Performance Monitoring Unit (PMU).
5 * The initial version of perfmon.c was written by
6 * Ganesh Venkitachalam, IBM Corp.
8 * Then it was modified for perfmon-1.x by Stephane Eranian and
9 * David Mosberger, Hewlett Packard Co.
11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
12 * by Stephane Eranian, Hewlett Packard Co.
14 * Copyright (C) 1999-2003 Hewlett Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 * David Mosberger-Tang <davidm@hpl.hp.com>
18 * More information about perfmon available at:
19 * http://www.hpl.hp.com/research/linux/perfmon
22 #include <linux/config.h>
23 #include <linux/module.h>
24 #include <linux/kernel.h>
25 #include <linux/sched.h>
26 #include <linux/interrupt.h>
27 #include <linux/smp_lock.h>
28 #include <linux/proc_fs.h>
29 #include <linux/seq_file.h>
30 #include <linux/init.h>
31 #include <linux/vmalloc.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/pagemap.h>
39 #include <linux/mount.h>
40 #include <linux/version.h>
42 #include <asm/bitops.h>
43 #include <asm/errno.h>
44 #include <asm/intrinsics.h>
46 #include <asm/perfmon.h>
47 #include <asm/processor.h>
48 #include <asm/signal.h>
49 #include <asm/system.h>
50 #include <asm/uaccess.h>
51 #include <asm/delay.h>
55 * perfmon context state
57 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
58 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
59 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
60 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
62 #define PFM_INVALID_ACTIVATION (~0UL)
65 * depth of message queue
67 #define PFM_MAX_MSGS 32
68 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
71 * type of a PMU register (bitmask).
73 * bit0 : register implemented
76 * bit4 : pmc has pmc.pm
77 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
78 * bit6-7 : register type
81 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
82 #define PFM_REG_IMPL 0x1 /* register implemented */
83 #define PFM_REG_END 0x2 /* end marker */
84 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
85 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
86 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
87 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
88 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
90 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
91 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
93 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
95 /* i assumed unsigned */
96 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
97 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
99 /* XXX: these assume that register i is implemented */
100 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
101 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
102 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
103 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
105 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
106 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
107 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
108 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
110 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
111 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
113 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
114 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
115 #define PFM_CTX_TASK(h) (h)->ctx_task
117 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
119 /* XXX: does not support more than 64 PMDs */
120 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
121 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
123 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
125 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
126 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
127 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
128 #define PFM_CODE_RR 0 /* requesting code range restriction */
129 #define PFM_DATA_RR 1 /* requestion data range restriction */
131 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
132 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
133 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
135 #define RDEP(x) (1UL<<(x))
138 * context protection macros
140 * - we need to protect against CPU concurrency (spin_lock)
141 * - we need to protect against PMU overflow interrupts (local_irq_disable)
143 * - we need to protect against PMU overflow interrupts (local_irq_disable)
145 * spin_lock_irqsave()/spin_lock_irqrestore():
146 * in SMP: local_irq_disable + spin_lock
147 * in UP : local_irq_disable
149 * spin_lock()/spin_lock():
150 * in UP : removed automatically
151 * in SMP: protect against context accesses from other CPU. interrupts
152 * are not masked. This is useful for the PMU interrupt handler
153 * because we know we will not get PMU concurrency in that code.
155 #define PROTECT_CTX(c, f) \
157 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \
158 spin_lock_irqsave(&(c)->ctx_lock, f); \
159 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
162 #define UNPROTECT_CTX(c, f) \
164 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
165 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
168 #define PROTECT_CTX_NOPRINT(c, f) \
170 spin_lock_irqsave(&(c)->ctx_lock, f); \
174 #define UNPROTECT_CTX_NOPRINT(c, f) \
176 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
180 #define PROTECT_CTX_NOIRQ(c) \
182 spin_lock(&(c)->ctx_lock); \
185 #define UNPROTECT_CTX_NOIRQ(c) \
187 spin_unlock(&(c)->ctx_lock); \
193 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
194 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
195 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
197 #else /* !CONFIG_SMP */
198 #define SET_ACTIVATION(t) do {} while(0)
199 #define GET_ACTIVATION(t) do {} while(0)
200 #define INC_ACTIVATION(t) do {} while(0)
201 #endif /* CONFIG_SMP */
203 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
204 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
205 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
207 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
208 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
210 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
213 * cmp0 must be the value of pmc0
215 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
217 #define PFMFS_MAGIC 0xa0b4d889
222 #define PFM_DEBUGGING 1
226 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
229 #define DPRINT_ovfl(a) \
231 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
236 * 64-bit software counter structure
238 * the next_reset_type is applied to the next call to pfm_reset_regs()
241 unsigned long val; /* virtual 64bit counter value */
242 unsigned long lval; /* last reset value */
243 unsigned long long_reset; /* reset value on sampling overflow */
244 unsigned long short_reset; /* reset value on overflow */
245 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
246 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
247 unsigned long seed; /* seed for random-number generator */
248 unsigned long mask; /* mask for random-number generator */
249 unsigned int flags; /* notify/do not notify */
250 unsigned long eventid; /* overflow event identifier */
257 unsigned int block:1; /* when 1, task will blocked on user notifications */
258 unsigned int system:1; /* do system wide monitoring */
259 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
260 unsigned int is_sampling:1; /* true if using a custom format */
261 unsigned int excl_idle:1; /* exclude idle task in system wide session */
262 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
263 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
264 unsigned int no_msg:1; /* no message sent on overflow */
265 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
266 unsigned int reserved:22;
267 } pfm_context_flags_t;
269 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
270 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
271 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
275 * perfmon context: encapsulates all the state of a monitoring session
278 typedef struct pfm_context {
279 spinlock_t ctx_lock; /* context protection */
281 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
282 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
284 struct task_struct *ctx_task; /* task to which context is attached */
286 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
288 struct semaphore ctx_restart_sem; /* use for blocking notification mode */
290 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
291 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
292 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
294 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
295 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
296 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
298 unsigned long ctx_pmcs[IA64_NUM_PMC_REGS]; /* saved copies of PMC values */
300 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
301 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
302 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
303 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
305 pfm_counter_t ctx_pmds[IA64_NUM_PMD_REGS]; /* software state for PMDS */
307 u64 ctx_saved_psr_up; /* only contains psr.up value */
309 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
310 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
311 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
313 int ctx_fd; /* file descriptor used my this context */
315 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
316 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
317 unsigned long ctx_smpl_size; /* size of sampling buffer */
318 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
320 wait_queue_head_t ctx_msgq_wait;
321 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
324 struct fasync_struct *ctx_async_queue;
326 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
330 * magic number used to verify that structure is really
333 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
335 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
338 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
339 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
341 #define SET_LAST_CPU(ctx, v) do {} while(0)
342 #define GET_LAST_CPU(ctx) do {} while(0)
346 #define ctx_fl_block ctx_flags.block
347 #define ctx_fl_system ctx_flags.system
348 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
349 #define ctx_fl_is_sampling ctx_flags.is_sampling
350 #define ctx_fl_excl_idle ctx_flags.excl_idle
351 #define ctx_fl_going_zombie ctx_flags.going_zombie
352 #define ctx_fl_trap_reason ctx_flags.trap_reason
353 #define ctx_fl_no_msg ctx_flags.no_msg
354 #define ctx_fl_can_restart ctx_flags.can_restart
356 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
357 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
360 * global information about all sessions
361 * mostly used to synchronize between system wide and per-process
364 spinlock_t pfs_lock; /* lock the structure */
366 unsigned int pfs_task_sessions; /* number of per task sessions */
367 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
368 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
369 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
370 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
374 * information about a PMC or PMD.
375 * dep_pmd[]: a bitmask of dependent PMD registers
376 * dep_pmc[]: a bitmask of dependent PMC registers
378 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
382 unsigned long default_value; /* power-on default value */
383 unsigned long reserved_mask; /* bitmask of reserved bits */
384 pfm_reg_check_t read_check;
385 pfm_reg_check_t write_check;
386 unsigned long dep_pmd[4];
387 unsigned long dep_pmc[4];
390 /* assume cnum is a valid monitor */
391 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
394 * This structure is initialized at boot time and contains
395 * a description of the PMU main characteristics.
397 * If the probe function is defined, detection is based
398 * on its return value:
399 * - 0 means recognized PMU
400 * - anything else means not supported
401 * When the probe function is not defined, then the pmu_family field
402 * is used and it must match the host CPU family such that:
403 * - cpu->family & config->pmu_family != 0
406 unsigned long ovfl_val; /* overflow value for counters */
408 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
409 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
411 unsigned int num_pmcs; /* number of PMCS: computed at init time */
412 unsigned int num_pmds; /* number of PMDS: computed at init time */
413 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
414 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
416 char *pmu_name; /* PMU family name */
417 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
418 unsigned int flags; /* pmu specific flags */
419 unsigned int num_ibrs; /* number of IBRS: computed at init time */
420 unsigned int num_dbrs; /* number of DBRS: computed at init time */
421 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
422 int (*probe)(void); /* customized probe routine */
423 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
428 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
431 * debug register related type definitions
434 unsigned long ibr_mask:56;
435 unsigned long ibr_plm:4;
436 unsigned long ibr_ig:3;
437 unsigned long ibr_x:1;
441 unsigned long dbr_mask:56;
442 unsigned long dbr_plm:4;
443 unsigned long dbr_ig:2;
444 unsigned long dbr_w:1;
445 unsigned long dbr_r:1;
456 * perfmon command descriptions
459 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
462 unsigned int cmd_narg;
464 int (*cmd_getsize)(void *arg, size_t *sz);
467 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
468 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
469 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
470 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
473 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
474 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
475 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
476 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
477 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
479 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
482 int debug; /* turn on/off debugging via syslog */
483 int debug_ovfl; /* turn on/off debug printk in overflow handler */
484 int fastctxsw; /* turn on/off fast (unsecure) ctxsw */
485 int expert_mode; /* turn on/off value checking */
490 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
491 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
492 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
493 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
496 unsigned long pfm_smpl_handler_calls;
497 unsigned long pfm_smpl_handler_cycles;
498 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
502 * perfmon internal variables
504 static pfm_stats_t pfm_stats[NR_CPUS];
505 static pfm_session_t pfm_sessions; /* global sessions information */
507 static struct proc_dir_entry *perfmon_dir;
508 static pfm_uuid_t pfm_null_uuid = {0,};
510 static spinlock_t pfm_buffer_fmt_lock;
511 static LIST_HEAD(pfm_buffer_fmt_list);
513 static pmu_config_t *pmu_conf;
515 /* sysctl() controls */
516 static pfm_sysctl_t pfm_sysctl;
519 static ctl_table pfm_ctl_table[]={
520 {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
521 {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
522 {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
523 {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
526 static ctl_table pfm_sysctl_dir[] = {
527 {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
530 static ctl_table pfm_sysctl_root[] = {
531 {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
534 static struct ctl_table_header *pfm_sysctl_header;
536 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
537 static int pfm_flush(struct file *filp);
539 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
540 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
543 pfm_put_task(struct task_struct *task)
545 if (task != current) put_task_struct(task);
549 pfm_set_task_notify(struct task_struct *task)
551 struct thread_info *info;
553 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
554 set_bit(TIF_NOTIFY_RESUME, &info->flags);
558 pfm_clear_task_notify(void)
560 clear_thread_flag(TIF_NOTIFY_RESUME);
564 pfm_reserve_page(unsigned long a)
566 SetPageReserved(vmalloc_to_page((void *)a));
569 pfm_unreserve_page(unsigned long a)
571 ClearPageReserved(vmalloc_to_page((void*)a));
575 pfm_remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
577 return remap_page_range(vma, from, phys_addr, size, prot);
580 static inline unsigned long
581 pfm_protect_ctx_ctxsw(pfm_context_t *x)
583 spin_lock(&(x)->ctx_lock);
587 static inline unsigned long
588 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
590 spin_unlock(&(x)->ctx_lock);
593 static inline unsigned int
594 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
596 return do_munmap(mm, addr, len);
599 static inline unsigned long
600 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
602 return get_unmapped_area(file, addr, len, pgoff, flags);
606 static struct super_block *
607 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
609 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
612 static struct file_system_type pfm_fs_type = {
614 .get_sb = pfmfs_get_sb,
615 .kill_sb = kill_anon_super,
618 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
619 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
620 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
621 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
624 /* forward declaration */
625 static struct file_operations pfm_file_ops;
628 * forward declarations
631 static void pfm_lazy_save_regs (struct task_struct *ta);
634 void dump_pmu_state(const char *);
635 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
637 #include "perfmon_itanium.h"
638 #include "perfmon_mckinley.h"
639 #include "perfmon_generic.h"
641 static pmu_config_t *pmu_confs[]={
644 &pmu_conf_gen, /* must be last */
649 static int pfm_end_notify_user(pfm_context_t *ctx);
652 pfm_clear_psr_pp(void)
654 ia64_rsm(IA64_PSR_PP);
661 ia64_ssm(IA64_PSR_PP);
666 pfm_clear_psr_up(void)
668 ia64_rsm(IA64_PSR_UP);
675 ia64_ssm(IA64_PSR_UP);
679 static inline unsigned long
683 tmp = ia64_getreg(_IA64_REG_PSR);
689 pfm_set_psr_l(unsigned long val)
691 ia64_setreg(_IA64_REG_PSR_L, val);
703 pfm_unfreeze_pmu(void)
710 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
714 for (i=0; i < nibrs; i++) {
715 ia64_set_ibr(i, ibrs[i]);
716 ia64_dv_serialize_instruction();
722 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
726 for (i=0; i < ndbrs; i++) {
727 ia64_set_dbr(i, dbrs[i]);
728 ia64_dv_serialize_data();
734 * PMD[i] must be a counter. no check is made
736 static inline unsigned long
737 pfm_read_soft_counter(pfm_context_t *ctx, int i)
739 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
743 * PMD[i] must be a counter. no check is made
746 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
748 unsigned long ovfl_val = pmu_conf->ovfl_val;
750 ctx->ctx_pmds[i].val = val & ~ovfl_val;
752 * writing to unimplemented part is ignore, so we do not need to
755 ia64_set_pmd(i, val & ovfl_val);
759 pfm_get_new_msg(pfm_context_t *ctx)
763 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
765 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
766 if (next == ctx->ctx_msgq_head) return NULL;
768 idx = ctx->ctx_msgq_tail;
769 ctx->ctx_msgq_tail = next;
771 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
773 return ctx->ctx_msgq+idx;
777 pfm_get_next_msg(pfm_context_t *ctx)
781 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
783 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
788 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
793 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
795 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));
801 pfm_reset_msgq(pfm_context_t *ctx)
803 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
804 DPRINT(("ctx=%p msgq reset\n", ctx));
808 /* Here we want the physical address of the memory.
809 * This is used when initializing the contents of the
810 * area and marking the pages as reserved.
812 static inline unsigned long
813 pfm_kvirt_to_pa(unsigned long adr)
815 __u64 pa = ia64_tpa(adr);
820 pfm_rvmalloc(unsigned long size)
825 size = PAGE_ALIGN(size);
828 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
829 memset(mem, 0, size);
830 addr = (unsigned long)mem;
832 pfm_reserve_page(addr);
841 pfm_rvfree(void *mem, unsigned long size)
846 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
847 addr = (unsigned long) mem;
848 while ((long) size > 0) {
849 pfm_unreserve_page(addr);
858 static pfm_context_t *
859 pfm_context_alloc(void)
864 * allocate context descriptor
865 * must be able to free with interrupts disabled
867 ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
869 memset(ctx, 0, sizeof(pfm_context_t));
870 DPRINT(("alloc ctx @%p\n", ctx));
876 pfm_context_free(pfm_context_t *ctx)
879 DPRINT(("free ctx @%p\n", ctx));
885 pfm_mask_monitoring(struct task_struct *task)
887 pfm_context_t *ctx = PFM_GET_CTX(task);
888 struct thread_struct *th = &task->thread;
889 unsigned long mask, val, ovfl_mask;
892 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
894 ovfl_mask = pmu_conf->ovfl_val;
896 * monitoring can only be masked as a result of a valid
897 * counter overflow. In UP, it means that the PMU still
898 * has an owner. Note that the owner can be different
899 * from the current task. However the PMU state belongs
901 * In SMP, a valid overflow only happens when task is
902 * current. Therefore if we come here, we know that
903 * the PMU state belongs to the current task, therefore
904 * we can access the live registers.
906 * So in both cases, the live register contains the owner's
907 * state. We can ONLY touch the PMU registers and NOT the PSR.
909 * As a consequence to this call, the thread->pmds[] array
910 * contains stale information which must be ignored
911 * when context is reloaded AND monitoring is active (see
914 mask = ctx->ctx_used_pmds[0];
915 for (i = 0; mask; i++, mask>>=1) {
916 /* skip non used pmds */
917 if ((mask & 0x1) == 0) continue;
918 val = ia64_get_pmd(i);
920 if (PMD_IS_COUNTING(i)) {
922 * we rebuild the full 64 bit value of the counter
924 ctx->ctx_pmds[i].val += (val & ovfl_mask);
926 ctx->ctx_pmds[i].val = val;
928 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
930 ctx->ctx_pmds[i].val,
934 * mask monitoring by setting the privilege level to 0
935 * we cannot use psr.pp/psr.up for this, it is controlled by
938 * if task is current, modify actual registers, otherwise modify
939 * thread save state, i.e., what will be restored in pfm_load_regs()
941 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
942 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
943 if ((mask & 0x1) == 0UL) continue;
944 ia64_set_pmc(i, th->pmcs[i] & ~0xfUL);
945 th->pmcs[i] &= ~0xfUL;
946 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i]));
949 * make all of this visible
955 * must always be done with task == current
957 * context must be in MASKED state when calling
960 pfm_restore_monitoring(struct task_struct *task)
962 pfm_context_t *ctx = PFM_GET_CTX(task);
963 struct thread_struct *th = &task->thread;
964 unsigned long mask, ovfl_mask;
965 unsigned long psr, val;
968 is_system = ctx->ctx_fl_system;
969 ovfl_mask = pmu_conf->ovfl_val;
971 if (task != current) {
972 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
975 if (ctx->ctx_state != PFM_CTX_MASKED) {
976 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
977 task->pid, current->pid, ctx->ctx_state);
982 * monitoring is masked via the PMC.
983 * As we restore their value, we do not want each counter to
984 * restart right away. We stop monitoring using the PSR,
985 * restore the PMC (and PMD) and then re-establish the psr
986 * as it was. Note that there can be no pending overflow at
987 * this point, because monitoring was MASKED.
989 * system-wide session are pinned and self-monitoring
991 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
993 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
999 * first, we restore the PMD
1001 mask = ctx->ctx_used_pmds[0];
1002 for (i = 0; mask; i++, mask>>=1) {
1003 /* skip non used pmds */
1004 if ((mask & 0x1) == 0) continue;
1006 if (PMD_IS_COUNTING(i)) {
1008 * we split the 64bit value according to
1011 val = ctx->ctx_pmds[i].val & ovfl_mask;
1012 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1014 val = ctx->ctx_pmds[i].val;
1016 ia64_set_pmd(i, val);
1018 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1020 ctx->ctx_pmds[i].val,
1026 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1027 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1028 if ((mask & 0x1) == 0UL) continue;
1029 th->pmcs[i] = ctx->ctx_pmcs[i];
1030 ia64_set_pmc(i, th->pmcs[i]);
1031 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
1036 * must restore DBR/IBR because could be modified while masked
1037 * XXX: need to optimize
1039 if (ctx->ctx_fl_using_dbreg) {
1040 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1041 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1047 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1049 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1056 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1062 for (i=0; mask; i++, mask>>=1) {
1063 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1068 * reload from thread state (used for ctxw only)
1071 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1074 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1076 for (i=0; mask; i++, mask>>=1) {
1077 if ((mask & 0x1) == 0) continue;
1078 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1079 ia64_set_pmd(i, val);
1085 * propagate PMD from context to thread-state
1088 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1090 struct thread_struct *thread = &task->thread;
1091 unsigned long ovfl_val = pmu_conf->ovfl_val;
1092 unsigned long mask = ctx->ctx_all_pmds[0];
1096 DPRINT(("mask=0x%lx\n", mask));
1098 for (i=0; mask; i++, mask>>=1) {
1100 val = ctx->ctx_pmds[i].val;
1103 * We break up the 64 bit value into 2 pieces
1104 * the lower bits go to the machine state in the
1105 * thread (will be reloaded on ctxsw in).
1106 * The upper part stays in the soft-counter.
1108 if (PMD_IS_COUNTING(i)) {
1109 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1112 thread->pmds[i] = val;
1114 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1117 ctx->ctx_pmds[i].val));
1122 * propagate PMC from context to thread-state
1125 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1127 struct thread_struct *thread = &task->thread;
1128 unsigned long mask = ctx->ctx_all_pmcs[0];
1131 DPRINT(("mask=0x%lx\n", mask));
1133 for (i=0; mask; i++, mask>>=1) {
1134 /* masking 0 with ovfl_val yields 0 */
1135 thread->pmcs[i] = ctx->ctx_pmcs[i];
1136 DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
1143 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1147 for (i=0; mask; i++, mask>>=1) {
1148 if ((mask & 0x1) == 0) continue;
1149 ia64_set_pmc(i, pmcs[i]);
1155 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1157 return memcmp(a, b, sizeof(pfm_uuid_t));
1161 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1164 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1169 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1172 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1178 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1182 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1187 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1191 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1196 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1199 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1204 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)
1207 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1211 static pfm_buffer_fmt_t *
1212 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1214 struct list_head * pos;
1215 pfm_buffer_fmt_t * entry;
1217 list_for_each(pos, &pfm_buffer_fmt_list) {
1218 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1219 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1226 * find a buffer format based on its uuid
1228 static pfm_buffer_fmt_t *
1229 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1231 pfm_buffer_fmt_t * fmt;
1232 spin_lock(&pfm_buffer_fmt_lock);
1233 fmt = __pfm_find_buffer_fmt(uuid);
1234 spin_unlock(&pfm_buffer_fmt_lock);
1239 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1243 /* some sanity checks */
1244 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1246 /* we need at least a handler */
1247 if (fmt->fmt_handler == NULL) return -EINVAL;
1250 * XXX: need check validity of fmt_arg_size
1253 spin_lock(&pfm_buffer_fmt_lock);
1255 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1256 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1260 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1261 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1264 spin_unlock(&pfm_buffer_fmt_lock);
1267 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1270 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1272 pfm_buffer_fmt_t *fmt;
1275 spin_lock(&pfm_buffer_fmt_lock);
1277 fmt = __pfm_find_buffer_fmt(uuid);
1279 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1283 list_del_init(&fmt->fmt_list);
1284 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1287 spin_unlock(&pfm_buffer_fmt_lock);
1291 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1294 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1296 unsigned long flags;
1298 * validy checks on cpu_mask have been done upstream
1302 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1303 pfm_sessions.pfs_sys_sessions,
1304 pfm_sessions.pfs_task_sessions,
1305 pfm_sessions.pfs_sys_use_dbregs,
1311 * cannot mix system wide and per-task sessions
1313 if (pfm_sessions.pfs_task_sessions > 0UL) {
1314 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1315 pfm_sessions.pfs_task_sessions));
1319 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1321 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1323 pfm_sessions.pfs_sys_session[cpu] = task;
1325 pfm_sessions.pfs_sys_sessions++ ;
1328 if (pfm_sessions.pfs_sys_sessions) goto abort;
1329 pfm_sessions.pfs_task_sessions++;
1332 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1333 pfm_sessions.pfs_sys_sessions,
1334 pfm_sessions.pfs_task_sessions,
1335 pfm_sessions.pfs_sys_use_dbregs,
1344 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1345 pfm_sessions.pfs_sys_session[cpu]->pid,
1346 smp_processor_id()));
1355 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1357 unsigned long flags;
1359 * validy checks on cpu_mask have been done upstream
1363 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1364 pfm_sessions.pfs_sys_sessions,
1365 pfm_sessions.pfs_task_sessions,
1366 pfm_sessions.pfs_sys_use_dbregs,
1372 pfm_sessions.pfs_sys_session[cpu] = NULL;
1374 * would not work with perfmon+more than one bit in cpu_mask
1376 if (ctx && ctx->ctx_fl_using_dbreg) {
1377 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1378 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1380 pfm_sessions.pfs_sys_use_dbregs--;
1383 pfm_sessions.pfs_sys_sessions--;
1385 pfm_sessions.pfs_task_sessions--;
1387 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1388 pfm_sessions.pfs_sys_sessions,
1389 pfm_sessions.pfs_task_sessions,
1390 pfm_sessions.pfs_sys_use_dbregs,
1400 * removes virtual mapping of the sampling buffer.
1401 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1402 * a PROTECT_CTX() section.
1405 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1410 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1411 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1415 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1418 * does the actual unmapping
1420 down_write(&task->mm->mmap_sem);
1422 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1424 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1426 up_write(&task->mm->mmap_sem);
1428 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1431 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1437 * free actual physical storage used by sampling buffer
1441 pfm_free_smpl_buffer(pfm_context_t *ctx)
1443 pfm_buffer_fmt_t *fmt;
1445 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1448 * we won't use the buffer format anymore
1450 fmt = ctx->ctx_buf_fmt;
1452 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1455 ctx->ctx_smpl_vaddr));
1457 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1462 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1464 ctx->ctx_smpl_hdr = NULL;
1465 ctx->ctx_smpl_size = 0UL;
1470 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1476 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1478 if (fmt == NULL) return;
1480 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1485 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1486 * no real gain from having the whole whorehouse mounted. So we don't need
1487 * any operations on the root directory. However, we need a non-trivial
1488 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1490 static struct vfsmount *pfmfs_mnt;
1495 int err = register_filesystem(&pfm_fs_type);
1497 pfmfs_mnt = kern_mount(&pfm_fs_type);
1498 err = PTR_ERR(pfmfs_mnt);
1499 if (IS_ERR(pfmfs_mnt))
1500 unregister_filesystem(&pfm_fs_type);
1510 unregister_filesystem(&pfm_fs_type);
1515 pfm_lseek(struct file *file, loff_t offset, int whence)
1517 DPRINT(("pfm_lseek called\n"));
1522 pfm_read(struct file *filp, char *buf, size_t size, loff_t *ppos)
1527 unsigned long flags;
1528 DECLARE_WAITQUEUE(wait, current);
1529 if (PFM_IS_FILE(filp) == 0) {
1530 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1534 ctx = (pfm_context_t *)filp->private_data;
1536 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1541 * check even when there is no message
1543 if (size < sizeof(pfm_msg_t)) {
1544 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1548 * seeks are not allowed on message queues
1550 if (ppos != &filp->f_pos) return -ESPIPE;
1552 PROTECT_CTX(ctx, flags);
1555 * put ourselves on the wait queue
1557 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1565 set_current_state(TASK_INTERRUPTIBLE);
1567 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1570 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1572 UNPROTECT_CTX(ctx, flags);
1575 * check non-blocking read
1578 if(filp->f_flags & O_NONBLOCK) break;
1581 * check pending signals
1583 if(signal_pending(current)) {
1588 * no message, so wait
1592 PROTECT_CTX(ctx, flags);
1594 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1595 set_current_state(TASK_RUNNING);
1596 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1598 if (ret < 0) goto abort;
1601 msg = pfm_get_next_msg(ctx);
1603 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1607 DPRINT(("[%d] fd=%d type=%d\n", current->pid, msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1610 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1613 UNPROTECT_CTX(ctx, flags);
1619 pfm_write(struct file *file, const char *ubuf,
1620 size_t size, loff_t *ppos)
1622 DPRINT(("pfm_write called\n"));
1627 pfm_poll(struct file *filp, poll_table * wait)
1630 unsigned long flags;
1631 unsigned int mask = 0;
1633 if (PFM_IS_FILE(filp) == 0) {
1634 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1638 ctx = (pfm_context_t *)filp->private_data;
1640 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1645 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1647 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1649 PROTECT_CTX(ctx, flags);
1651 if (PFM_CTXQ_EMPTY(ctx) == 0)
1652 mask = POLLIN | POLLRDNORM;
1654 UNPROTECT_CTX(ctx, flags);
1656 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1662 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1664 DPRINT(("pfm_ioctl called\n"));
1669 * context is locked when coming here and interrupts are disabled
1672 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1676 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1678 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1682 ctx->ctx_async_queue, ret));
1688 pfm_fasync(int fd, struct file *filp, int on)
1691 unsigned long flags;
1694 if (PFM_IS_FILE(filp) == 0) {
1695 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1699 ctx = (pfm_context_t *)filp->private_data;
1701 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1706 PROTECT_CTX(ctx, flags);
1708 ret = pfm_do_fasync(fd, filp, ctx, on);
1710 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1713 ctx->ctx_async_queue, ret));
1715 UNPROTECT_CTX(ctx, flags);
1722 * this function is exclusively called from pfm_close().
1723 * The context is not protected at that time, nor are interrupts
1724 * on the remote CPU. That's necessary to avoid deadlocks.
1727 pfm_syswide_force_stop(void *info)
1729 pfm_context_t *ctx = (pfm_context_t *)info;
1730 struct pt_regs *regs = ia64_task_regs(current);
1731 struct task_struct *owner;
1732 unsigned long flags;
1735 if (ctx->ctx_cpu != smp_processor_id()) {
1736 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1738 smp_processor_id());
1741 owner = GET_PMU_OWNER();
1742 if (owner != ctx->ctx_task) {
1743 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1745 owner->pid, ctx->ctx_task->pid);
1748 if (GET_PMU_CTX() != ctx) {
1749 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1751 GET_PMU_CTX(), ctx);
1755 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1757 * the context is already protected in pfm_close(), we simply
1758 * need to mask interrupts to avoid a PMU interrupt race on
1761 local_irq_save(flags);
1763 ret = pfm_context_unload(ctx, NULL, 0, regs);
1765 DPRINT(("context_unload returned %d\n", ret));
1769 * unmask interrupts, PMU interrupts are now spurious here
1771 local_irq_restore(flags);
1775 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1779 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1780 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1781 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1783 #endif /* CONFIG_SMP */
1786 * called for each close(). Partially free resources.
1787 * When caller is self-monitoring, the context is unloaded.
1790 pfm_flush(struct file *filp)
1793 struct task_struct *task;
1794 struct pt_regs *regs;
1795 unsigned long flags;
1796 unsigned long smpl_buf_size = 0UL;
1797 void *smpl_buf_vaddr = NULL;
1798 int state, is_system;
1800 if (PFM_IS_FILE(filp) == 0) {
1801 DPRINT(("bad magic for\n"));
1805 ctx = (pfm_context_t *)filp->private_data;
1807 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1812 * remove our file from the async queue, if we use this mode.
1813 * This can be done without the context being protected. We come
1814 * here when the context has become unreacheable by other tasks.
1816 * We may still have active monitoring at this point and we may
1817 * end up in pfm_overflow_handler(). However, fasync_helper()
1818 * operates with interrupts disabled and it cleans up the
1819 * queue. If the PMU handler is called prior to entering
1820 * fasync_helper() then it will send a signal. If it is
1821 * invoked after, it will find an empty queue and no
1822 * signal will be sent. In both case, we are safe
1824 if (filp->f_flags & FASYNC) {
1825 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1826 pfm_do_fasync (-1, filp, ctx, 0);
1829 PROTECT_CTX(ctx, flags);
1831 state = ctx->ctx_state;
1832 is_system = ctx->ctx_fl_system;
1834 task = PFM_CTX_TASK(ctx);
1835 regs = ia64_task_regs(task);
1837 DPRINT(("ctx_state=%d is_current=%d\n",
1839 task == current ? 1 : 0));
1842 * if state == UNLOADED, then task is NULL
1846 * we must stop and unload because we are losing access to the context.
1848 if (task == current) {
1851 * the task IS the owner but it migrated to another CPU: that's bad
1852 * but we must handle this cleanly. Unfortunately, the kernel does
1853 * not provide a mechanism to block migration (while the context is loaded).
1855 * We need to release the resource on the ORIGINAL cpu.
1857 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1859 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1861 * keep context protected but unmask interrupt for IPI
1863 local_irq_restore(flags);
1865 pfm_syswide_cleanup_other_cpu(ctx);
1868 * restore interrupt masking
1870 local_irq_save(flags);
1873 * context is unloaded at this point
1876 #endif /* CONFIG_SMP */
1879 DPRINT(("forcing unload\n"));
1881 * stop and unload, returning with state UNLOADED
1882 * and session unreserved.
1884 pfm_context_unload(ctx, NULL, 0, regs);
1886 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1891 * remove virtual mapping, if any, for the calling task.
1892 * cannot reset ctx field until last user is calling close().
1894 * ctx_smpl_vaddr must never be cleared because it is needed
1895 * by every task with access to the context
1897 * When called from do_exit(), the mm context is gone already, therefore
1898 * mm is NULL, i.e., the VMA is already gone and we do not have to
1901 if (ctx->ctx_smpl_vaddr && current->mm) {
1902 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1903 smpl_buf_size = ctx->ctx_smpl_size;
1906 UNPROTECT_CTX(ctx, flags);
1909 * if there was a mapping, then we systematically remove it
1910 * at this point. Cannot be done inside critical section
1911 * because some VM function reenables interrupts.
1914 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1919 * called either on explicit close() or from exit_files().
1920 * Only the LAST user of the file gets to this point, i.e., it is
1923 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1924 * (fput()),i.e, last task to access the file. Nobody else can access the
1925 * file at this point.
1927 * When called from exit_files(), the VMA has been freed because exit_mm()
1928 * is executed before exit_files().
1930 * When called from exit_files(), the current task is not yet ZOMBIE but we
1931 * flush the PMU state to the context.
1934 pfm_close(struct inode *inode, struct file *filp)
1937 struct task_struct *task;
1938 struct pt_regs *regs;
1939 DECLARE_WAITQUEUE(wait, current);
1940 unsigned long flags;
1941 unsigned long smpl_buf_size = 0UL;
1942 void *smpl_buf_addr = NULL;
1943 int free_possible = 1;
1944 int state, is_system;
1946 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1948 if (PFM_IS_FILE(filp) == 0) {
1949 DPRINT(("bad magic\n"));
1953 ctx = (pfm_context_t *)filp->private_data;
1955 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1959 PROTECT_CTX(ctx, flags);
1961 state = ctx->ctx_state;
1962 is_system = ctx->ctx_fl_system;
1964 task = PFM_CTX_TASK(ctx);
1965 regs = ia64_task_regs(task);
1967 DPRINT(("ctx_state=%d is_current=%d\n",
1969 task == current ? 1 : 0));
1972 * if task == current, then pfm_flush() unloaded the context
1974 if (state == PFM_CTX_UNLOADED) goto doit;
1977 * context is loaded/masked and task != current, we need to
1978 * either force an unload or go zombie
1982 * The task is currently blocked or will block after an overflow.
1983 * we must force it to wakeup to get out of the
1984 * MASKED state and transition to the unloaded state by itself.
1986 * This situation is only possible for per-task mode
1988 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1991 * set a "partial" zombie state to be checked
1992 * upon return from down() in pfm_handle_work().
1994 * We cannot use the ZOMBIE state, because it is checked
1995 * by pfm_load_regs() which is called upon wakeup from down().
1996 * In such case, it would free the context and then we would
1997 * return to pfm_handle_work() which would access the
1998 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1999 * but visible to pfm_handle_work().
2001 * For some window of time, we have a zombie context with
2002 * ctx_state = MASKED and not ZOMBIE
2004 ctx->ctx_fl_going_zombie = 1;
2007 * force task to wake up from MASKED state
2009 up(&ctx->ctx_restart_sem);
2011 DPRINT(("waking up ctx_state=%d\n", state));
2014 * put ourself to sleep waiting for the other
2015 * task to report completion
2017 * the context is protected by mutex, therefore there
2018 * is no risk of being notified of completion before
2019 * begin actually on the waitq.
2021 set_current_state(TASK_INTERRUPTIBLE);
2022 add_wait_queue(&ctx->ctx_zombieq, &wait);
2024 UNPROTECT_CTX(ctx, flags);
2027 * XXX: check for signals :
2028 * - ok of explicit close
2029 * - not ok when coming from exit_files()
2034 PROTECT_CTX(ctx, flags);
2037 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2038 set_current_state(TASK_RUNNING);
2041 * context is unloaded at this point
2043 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2045 else if (task != current) {
2048 * switch context to zombie state
2050 ctx->ctx_state = PFM_CTX_ZOMBIE;
2052 DPRINT(("zombie ctx for [%d]\n", task->pid));
2054 * cannot free the context on the spot. deferred until
2055 * the task notices the ZOMBIE state
2059 pfm_context_unload(ctx, NULL, 0, regs);
2064 /* reload state, may have changed during opening of critical section */
2065 state = ctx->ctx_state;
2068 * the context is still attached to a task (possibly current)
2069 * we cannot destroy it right now
2073 * we must free the sampling buffer right here because
2074 * we cannot rely on it being cleaned up later by the
2075 * monitored task. It is not possible to free vmalloc'ed
2076 * memory in pfm_load_regs(). Instead, we remove the buffer
2077 * now. should there be subsequent PMU overflow originally
2078 * meant for sampling, the will be converted to spurious
2079 * and that's fine because the monitoring tools is gone anyway.
2081 if (ctx->ctx_smpl_hdr) {
2082 smpl_buf_addr = ctx->ctx_smpl_hdr;
2083 smpl_buf_size = ctx->ctx_smpl_size;
2084 /* no more sampling */
2085 ctx->ctx_smpl_hdr = NULL;
2086 ctx->ctx_fl_is_sampling = 0;
2089 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2095 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2098 * UNLOADED that the session has already been unreserved.
2100 if (state == PFM_CTX_ZOMBIE) {
2101 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2105 * disconnect file descriptor from context must be done
2108 filp->private_data = NULL;
2111 * if we free on the spot, the context is now completely unreacheable
2112 * from the callers side. The monitored task side is also cut, so we
2115 * If we have a deferred free, only the caller side is disconnected.
2117 UNPROTECT_CTX(ctx, flags);
2120 * All memory free operations (especially for vmalloc'ed memory)
2121 * MUST be done with interrupts ENABLED.
2123 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2126 * return the memory used by the context
2128 if (free_possible) pfm_context_free(ctx);
2134 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2136 DPRINT(("pfm_no_open called\n"));
2142 static struct file_operations pfm_file_ops = {
2143 .llseek = pfm_lseek,
2148 .open = pfm_no_open, /* special open code to disallow open via /proc */
2149 .fasync = pfm_fasync,
2150 .release = pfm_close,
2155 pfmfs_delete_dentry(struct dentry *dentry)
2160 static struct dentry_operations pfmfs_dentry_operations = {
2161 .d_delete = pfmfs_delete_dentry,
2166 pfm_alloc_fd(struct file **cfile)
2169 struct file *file = NULL;
2170 struct inode * inode;
2174 fd = get_unused_fd();
2175 if (fd < 0) return -ENFILE;
2179 file = get_empty_filp();
2180 if (!file) goto out;
2183 * allocate a new inode
2185 inode = new_inode(pfmfs_mnt->mnt_sb);
2186 if (!inode) goto out;
2188 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2190 inode->i_sb = pfmfs_mnt->mnt_sb;
2191 inode->i_mode = S_IFCHR|S_IRUGO;
2193 inode->i_uid = current->fsuid;
2194 inode->i_gid = current->fsgid;
2196 sprintf(name, "[%lu]", inode->i_ino);
2198 this.len = strlen(name);
2199 this.hash = inode->i_ino;
2204 * allocate a new dcache entry
2206 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2207 if (!file->f_dentry) goto out;
2209 file->f_dentry->d_op = &pfmfs_dentry_operations;
2211 d_add(file->f_dentry, inode);
2212 file->f_vfsmnt = mntget(pfmfs_mnt);
2213 file->f_mapping = inode->i_mapping;
2215 file->f_op = &pfm_file_ops;
2216 file->f_mode = FMODE_READ;
2217 file->f_flags = O_RDONLY;
2221 * may have to delay until context is attached?
2223 fd_install(fd, file);
2226 * the file structure we will use
2232 if (file) put_filp(file);
2238 pfm_free_fd(int fd, struct file *file)
2240 if (file) put_filp(file);
2245 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2249 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2252 page = pfm_kvirt_to_pa(buf);
2254 if (pfm_remap_page_range(vma, addr, page, PAGE_SIZE, PAGE_READONLY)) return -ENOMEM;
2264 * allocate a sampling buffer and remaps it into the user address space of the task
2267 pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2269 struct mm_struct *mm = task->mm;
2270 struct vm_area_struct *vma = NULL;
2276 * the fixed header + requested size and align to page boundary
2278 size = PAGE_ALIGN(rsize);
2280 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2283 * check requested size to avoid Denial-of-service attacks
2284 * XXX: may have to refine this test
2285 * Check against address space limit.
2287 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2290 if (size > task->rlim[RLIMIT_MEMLOCK].rlim_cur) return -EAGAIN;
2293 * We do the easy to undo allocations first.
2295 * pfm_rvmalloc(), clears the buffer, so there is no leak
2297 smpl_buf = pfm_rvmalloc(size);
2298 if (smpl_buf == NULL) {
2299 DPRINT(("Can't allocate sampling buffer\n"));
2303 DPRINT(("smpl_buf @%p\n", smpl_buf));
2306 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2308 DPRINT(("Cannot allocate vma\n"));
2311 memset(vma, 0, sizeof(*vma));
2314 * partially initialize the vma for the sampling buffer
2316 * The VM_DONTCOPY flag is very important as it ensures that the mapping
2317 * will never be inherited for any child process (via fork()) which is always
2321 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2322 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2325 * Now we have everything we need and we can initialize
2326 * and connect all the data structures
2329 ctx->ctx_smpl_hdr = smpl_buf;
2330 ctx->ctx_smpl_size = size; /* aligned size */
2333 * Let's do the difficult operations next.
2335 * now we atomically find some area in the address space and
2336 * remap the buffer in it.
2338 down_write(&task->mm->mmap_sem);
2340 /* find some free area in address space, must have mmap sem held */
2341 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2342 if (vma->vm_start == 0UL) {
2343 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2344 up_write(&task->mm->mmap_sem);
2347 vma->vm_end = vma->vm_start + size;
2349 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2351 /* can only be applied to current task, need to have the mm semaphore held when called */
2352 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2353 DPRINT(("Can't remap buffer\n"));
2354 up_write(&task->mm->mmap_sem);
2359 * now insert the vma in the vm list for the process, must be
2360 * done with mmap lock held
2362 insert_vm_struct(mm, vma);
2364 // mm->total_vm += size >> PAGE_SHIFT;
2365 vx_vmpages_add(mm, size >> PAGE_SHIFT);
2367 up_write(&task->mm->mmap_sem);
2370 * keep track of user level virtual address
2372 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2373 *(unsigned long *)user_vaddr = vma->vm_start;
2378 kmem_cache_free(vm_area_cachep, vma);
2380 pfm_rvfree(smpl_buf, size);
2386 * XXX: do something better here
2389 pfm_bad_permissions(struct task_struct *task)
2391 /* inspired by ptrace_attach() */
2392 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2401 return ((current->uid != task->euid)
2402 || (current->uid != task->suid)
2403 || (current->uid != task->uid)
2404 || (current->gid != task->egid)
2405 || (current->gid != task->sgid)
2406 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2410 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2416 ctx_flags = pfx->ctx_flags;
2418 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2421 * cannot block in this mode
2423 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2424 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2429 /* probably more to add here */
2435 pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2436 unsigned int cpu, pfarg_context_t *arg)
2438 pfm_buffer_fmt_t *fmt = NULL;
2439 unsigned long size = 0UL;
2441 void *fmt_arg = NULL;
2443 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2445 /* invoke and lock buffer format, if found */
2446 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2448 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2453 * buffer argument MUST be contiguous to pfarg_context_t
2455 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2457 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2459 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2461 if (ret) goto error;
2463 /* link buffer format and context */
2464 ctx->ctx_buf_fmt = fmt;
2467 * check if buffer format wants to use perfmon buffer allocation/mapping service
2469 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2470 if (ret) goto error;
2474 * buffer is always remapped into the caller's address space
2476 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2477 if (ret) goto error;
2479 /* keep track of user address of buffer */
2480 arg->ctx_smpl_vaddr = uaddr;
2482 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2489 pfm_reset_pmu_state(pfm_context_t *ctx)
2494 * install reset values for PMC.
2496 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2497 if (PMC_IS_IMPL(i) == 0) continue;
2498 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2499 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2502 * PMD registers are set to 0UL when the context in memset()
2506 * On context switched restore, we must restore ALL pmc and ALL pmd even
2507 * when they are not actively used by the task. In UP, the incoming process
2508 * may otherwise pick up left over PMC, PMD state from the previous process.
2509 * As opposed to PMD, stale PMC can cause harm to the incoming
2510 * process because they may change what is being measured.
2511 * Therefore, we must systematically reinstall the entire
2512 * PMC state. In SMP, the same thing is possible on the
2513 * same CPU but also on between 2 CPUs.
2515 * The problem with PMD is information leaking especially
2516 * to user level when psr.sp=0
2518 * There is unfortunately no easy way to avoid this problem
2519 * on either UP or SMP. This definitively slows down the
2520 * pfm_load_regs() function.
2524 * bitmask of all PMCs accessible to this context
2526 * PMC0 is treated differently.
2528 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2531 * bitmask of all PMDs that are accesible to this context
2533 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2535 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2538 * useful in case of re-enable after disable
2540 ctx->ctx_used_ibrs[0] = 0UL;
2541 ctx->ctx_used_dbrs[0] = 0UL;
2545 pfm_ctx_getsize(void *arg, size_t *sz)
2547 pfarg_context_t *req = (pfarg_context_t *)arg;
2548 pfm_buffer_fmt_t *fmt;
2552 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2554 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2556 DPRINT(("cannot find buffer format\n"));
2559 /* get just enough to copy in user parameters */
2560 *sz = fmt->fmt_arg_size;
2561 DPRINT(("arg_size=%lu\n", *sz));
2569 * cannot attach if :
2571 * - task not owned by caller
2572 * - task incompatible with context mode
2575 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2578 * no kernel task or task not owner by caller
2580 if (task->mm == NULL) {
2581 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2584 if (pfm_bad_permissions(task)) {
2585 DPRINT(("no permission to attach to [%d]\n", task->pid));
2589 * cannot block in self-monitoring mode
2591 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2592 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2596 if (task->state == TASK_ZOMBIE) {
2597 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2602 * always ok for self
2604 if (task == current) return 0;
2606 if (task->state != TASK_STOPPED) {
2607 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2611 * make sure the task is off any CPU
2613 wait_task_inactive(task);
2615 /* more to come... */
2621 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2623 struct task_struct *p = current;
2626 /* XXX: need to add more checks here */
2627 if (pid < 2) return -EPERM;
2629 if (pid != current->pid) {
2631 read_lock(&tasklist_lock);
2633 p = find_task_by_pid(pid);
2635 /* make sure task cannot go away while we operate on it */
2636 if (p) get_task_struct(p);
2638 read_unlock(&tasklist_lock);
2640 if (p == NULL) return -ESRCH;
2643 ret = pfm_task_incompatible(ctx, p);
2646 } else if (p != current) {
2655 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2657 pfarg_context_t *req = (pfarg_context_t *)arg;
2662 /* let's check the arguments first */
2663 ret = pfarg_is_sane(current, req);
2664 if (ret < 0) return ret;
2666 ctx_flags = req->ctx_flags;
2670 ctx = pfm_context_alloc();
2671 if (!ctx) goto error;
2673 req->ctx_fd = ctx->ctx_fd = pfm_alloc_fd(&filp);
2674 if (req->ctx_fd < 0) goto error_file;
2677 * attach context to file
2679 filp->private_data = ctx;
2682 * does the user want to sample?
2684 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2685 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2686 if (ret) goto buffer_error;
2690 * init context protection lock
2692 spin_lock_init(&ctx->ctx_lock);
2695 * context is unloaded
2697 ctx->ctx_state = PFM_CTX_UNLOADED;
2700 * initialization of context's flags
2702 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2703 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2704 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2705 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2707 * will move to set properties
2708 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2712 * init restart semaphore to locked
2714 sema_init(&ctx->ctx_restart_sem, 0);
2717 * activation is used in SMP only
2719 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2720 SET_LAST_CPU(ctx, -1);
2723 * initialize notification message queue
2725 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2726 init_waitqueue_head(&ctx->ctx_msgq_wait);
2727 init_waitqueue_head(&ctx->ctx_zombieq);
2729 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2734 ctx->ctx_fl_excl_idle,
2739 * initialize soft PMU state
2741 pfm_reset_pmu_state(ctx);
2746 pfm_free_fd(ctx->ctx_fd, filp);
2748 if (ctx->ctx_buf_fmt) {
2749 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2752 pfm_context_free(ctx);
2758 static inline unsigned long
2759 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2761 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2762 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2763 extern unsigned long carta_random32 (unsigned long seed);
2765 if (reg->flags & PFM_REGFL_RANDOM) {
2766 new_seed = carta_random32(old_seed);
2767 val -= (old_seed & mask); /* counter values are negative numbers! */
2768 if ((mask >> 32) != 0)
2769 /* construct a full 64-bit random value: */
2770 new_seed |= carta_random32(old_seed >> 32) << 32;
2771 reg->seed = new_seed;
2778 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2780 unsigned long mask = ovfl_regs[0];
2781 unsigned long reset_others = 0UL;
2786 * now restore reset value on sampling overflowed counters
2788 mask >>= PMU_FIRST_COUNTER;
2789 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2791 if ((mask & 0x1UL) == 0UL) continue;
2793 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2794 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2796 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2800 * Now take care of resetting the other registers
2802 for(i = 0; reset_others; i++, reset_others >>= 1) {
2804 if ((reset_others & 0x1) == 0) continue;
2806 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2808 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2809 is_long_reset ? "long" : "short", i, val));
2814 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2816 unsigned long mask = ovfl_regs[0];
2817 unsigned long reset_others = 0UL;
2821 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2823 if (ctx->ctx_state == PFM_CTX_MASKED) {
2824 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2829 * now restore reset value on sampling overflowed counters
2831 mask >>= PMU_FIRST_COUNTER;
2832 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2834 if ((mask & 0x1UL) == 0UL) continue;
2836 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2837 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2839 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2841 pfm_write_soft_counter(ctx, i, val);
2845 * Now take care of resetting the other registers
2847 for(i = 0; reset_others; i++, reset_others >>= 1) {
2849 if ((reset_others & 0x1) == 0) continue;
2851 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2853 if (PMD_IS_COUNTING(i)) {
2854 pfm_write_soft_counter(ctx, i, val);
2856 ia64_set_pmd(i, val);
2858 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2859 is_long_reset ? "long" : "short", i, val));
2865 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2867 struct thread_struct *thread = NULL;
2868 struct task_struct *task;
2869 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2870 unsigned long value, pmc_pm;
2871 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2872 unsigned int cnum, reg_flags, flags, pmc_type;
2873 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2874 int is_monitor, is_counting, state;
2876 pfm_reg_check_t wr_func;
2877 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2879 state = ctx->ctx_state;
2880 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2881 is_system = ctx->ctx_fl_system;
2882 task = ctx->ctx_task;
2883 impl_pmds = pmu_conf->impl_pmds[0];
2885 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2888 thread = &task->thread;
2890 * In system wide and when the context is loaded, access can only happen
2891 * when the caller is running on the CPU being monitored by the session.
2892 * It does not have to be the owner (ctx_task) of the context per se.
2894 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2895 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2898 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2900 expert_mode = pfm_sysctl.expert_mode;
2902 for (i = 0; i < count; i++, req++) {
2904 cnum = req->reg_num;
2905 reg_flags = req->reg_flags;
2906 value = req->reg_value;
2907 smpl_pmds = req->reg_smpl_pmds[0];
2908 reset_pmds = req->reg_reset_pmds[0];
2912 if (cnum >= PMU_MAX_PMCS) {
2913 DPRINT(("pmc%u is invalid\n", cnum));
2917 pmc_type = pmu_conf->pmc_desc[cnum].type;
2918 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2919 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2920 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2923 * we reject all non implemented PMC as well
2924 * as attempts to modify PMC[0-3] which are used
2925 * as status registers by the PMU
2927 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2928 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2931 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2933 * If the PMC is a monitor, then if the value is not the default:
2934 * - system-wide session: PMCx.pm=1 (privileged monitor)
2935 * - per-task : PMCx.pm=0 (user monitor)
2937 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2938 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2947 * enforce generation of overflow interrupt. Necessary on all
2950 value |= 1 << PMU_PMC_OI;
2952 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2953 flags |= PFM_REGFL_OVFL_NOTIFY;
2956 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2958 /* verify validity of smpl_pmds */
2959 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2960 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2964 /* verify validity of reset_pmds */
2965 if ((reset_pmds & impl_pmds) != reset_pmds) {
2966 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2970 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2971 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2974 /* eventid on non-counting monitors are ignored */
2978 * execute write checker, if any
2980 if (likely(expert_mode == 0 && wr_func)) {
2981 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2982 if (ret) goto error;
2987 * no error on this register
2989 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2992 * Now we commit the changes to the software state
2996 * update overflow information
3000 * full flag update each time a register is programmed
3002 ctx->ctx_pmds[cnum].flags = flags;
3004 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
3005 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
3006 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
3009 * Mark all PMDS to be accessed as used.
3011 * We do not keep track of PMC because we have to
3012 * systematically restore ALL of them.
3014 * We do not update the used_monitors mask, because
3015 * if we have not programmed them, then will be in
3016 * a quiescent state, therefore we will not need to
3017 * mask/restore then when context is MASKED.
3019 CTX_USED_PMD(ctx, reset_pmds);
3020 CTX_USED_PMD(ctx, smpl_pmds);
3022 * make sure we do not try to reset on
3023 * restart because we have established new values
3025 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3028 * Needed in case the user does not initialize the equivalent
3029 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3030 * possible leak here.
3032 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3035 * keep track of the monitor PMC that we are using.
3036 * we save the value of the pmc in ctx_pmcs[] and if
3037 * the monitoring is not stopped for the context we also
3038 * place it in the saved state area so that it will be
3039 * picked up later by the context switch code.
3041 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3043 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3044 * monitoring needs to be stopped.
3046 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3049 * update context state
3051 ctx->ctx_pmcs[cnum] = value;
3055 * write thread state
3057 if (is_system == 0) thread->pmcs[cnum] = value;
3060 * write hardware register if we can
3062 if (can_access_pmu) {
3063 ia64_set_pmc(cnum, value);
3068 * per-task SMP only here
3070 * we are guaranteed that the task is not running on the other CPU,
3071 * we indicate that this PMD will need to be reloaded if the task
3072 * is rescheduled on the CPU it ran last on.
3074 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3079 DPRINT(("pmc[%u]=0x%lx loaded=%d access_pmu=%d 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",
3084 ctx->ctx_all_pmcs[0],
3085 ctx->ctx_used_pmds[0],
3086 ctx->ctx_pmds[cnum].eventid,
3089 ctx->ctx_reload_pmcs[0],
3090 ctx->ctx_used_monitors[0],
3091 ctx->ctx_ovfl_regs[0]));
3095 * make sure the changes are visible
3097 if (can_access_pmu) ia64_srlz_d();
3101 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3106 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3108 struct thread_struct *thread = NULL;
3109 struct task_struct *task;
3110 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3111 unsigned long value, hw_value, ovfl_mask;
3113 int i, can_access_pmu = 0, state;
3114 int is_counting, is_loaded, is_system, expert_mode;
3116 pfm_reg_check_t wr_func;
3119 state = ctx->ctx_state;
3120 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3121 is_system = ctx->ctx_fl_system;
3122 ovfl_mask = pmu_conf->ovfl_val;
3123 task = ctx->ctx_task;
3125 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3128 * on both UP and SMP, we can only write to the PMC when the task is
3129 * the owner of the local PMU.
3131 if (likely(is_loaded)) {
3132 thread = &task->thread;
3134 * In system wide and when the context is loaded, access can only happen
3135 * when the caller is running on the CPU being monitored by the session.
3136 * It does not have to be the owner (ctx_task) of the context per se.
3138 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3139 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3142 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3144 expert_mode = pfm_sysctl.expert_mode;
3146 for (i = 0; i < count; i++, req++) {
3148 cnum = req->reg_num;
3149 value = req->reg_value;
3151 if (!PMD_IS_IMPL(cnum)) {
3152 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3155 is_counting = PMD_IS_COUNTING(cnum);
3156 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3159 * execute write checker, if any
3161 if (unlikely(expert_mode == 0 && wr_func)) {
3162 unsigned long v = value;
3164 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3165 if (ret) goto abort_mission;
3172 * no error on this register
3174 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3177 * now commit changes to software state
3182 * update virtualized (64bits) counter
3186 * write context state
3188 ctx->ctx_pmds[cnum].lval = value;
3191 * when context is load we use the split value
3194 hw_value = value & ovfl_mask;
3195 value = value & ~ovfl_mask;
3199 * update reset values (not just for counters)
3201 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3202 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3205 * update randomization parameters (not just for counters)
3207 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3208 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3211 * update context value
3213 ctx->ctx_pmds[cnum].val = value;
3216 * Keep track of what we use
3218 * We do not keep track of PMC because we have to
3219 * systematically restore ALL of them.
3221 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3224 * mark this PMD register used as well
3226 CTX_USED_PMD(ctx, RDEP(cnum));
3229 * make sure we do not try to reset on
3230 * restart because we have established new values
3232 if (is_counting && state == PFM_CTX_MASKED) {
3233 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3238 * write thread state
3240 if (is_system == 0) thread->pmds[cnum] = hw_value;
3243 * write hardware register if we can
3245 if (can_access_pmu) {
3246 ia64_set_pmd(cnum, hw_value);
3250 * we are guaranteed that the task is not running on the other CPU,
3251 * we indicate that this PMD will need to be reloaded if the task
3252 * is rescheduled on the CPU it ran last on.
3254 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3259 DPRINT(("pmd[%u]=0x%lx loaded=%d access_pmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3260 "long_reset=0x%lx notify=%c used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3266 ctx->ctx_pmds[cnum].val,
3267 ctx->ctx_pmds[cnum].short_reset,
3268 ctx->ctx_pmds[cnum].long_reset,
3269 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3270 ctx->ctx_used_pmds[0],
3271 ctx->ctx_pmds[cnum].reset_pmds[0],
3272 ctx->ctx_reload_pmds[0],
3273 ctx->ctx_all_pmds[0],
3274 ctx->ctx_ovfl_regs[0]));
3278 * make changes visible
3280 if (can_access_pmu) ia64_srlz_d();
3286 * for now, we have only one possibility for error
3288 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3293 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3294 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3295 * interrupt is delivered during the call, it will be kept pending until we leave, making
3296 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3297 * guaranteed to return consistent data to the user, it may simply be old. It is not
3298 * trivial to treat the overflow while inside the call because you may end up in
3299 * some module sampling buffer code causing deadlocks.
3302 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3304 struct thread_struct *thread = NULL;
3305 struct task_struct *task;
3306 unsigned long val = 0UL, lval, ovfl_mask, sval;
3307 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3308 unsigned int cnum, reg_flags = 0;
3309 int i, can_access_pmu = 0, state;
3310 int is_loaded, is_system, is_counting, expert_mode;
3312 pfm_reg_check_t rd_func;
3315 * access is possible when loaded only for
3316 * self-monitoring tasks or in UP mode
3319 state = ctx->ctx_state;
3320 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3321 is_system = ctx->ctx_fl_system;
3322 ovfl_mask = pmu_conf->ovfl_val;
3323 task = ctx->ctx_task;
3325 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3327 if (likely(is_loaded)) {
3328 thread = &task->thread;
3330 * In system wide and when the context is loaded, access can only happen
3331 * when the caller is running on the CPU being monitored by the session.
3332 * It does not have to be the owner (ctx_task) of the context per se.
3334 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3335 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3339 * this can be true when not self-monitoring only in UP
3341 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3343 if (can_access_pmu) ia64_srlz_d();
3345 expert_mode = pfm_sysctl.expert_mode;
3347 DPRINT(("loaded=%d access_pmu=%d ctx_state=%d\n",
3353 * on both UP and SMP, we can only read the PMD from the hardware register when
3354 * the task is the owner of the local PMU.
3357 for (i = 0; i < count; i++, req++) {
3359 cnum = req->reg_num;
3360 reg_flags = req->reg_flags;
3362 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3364 * we can only read the register that we use. That includes
3365 * the one we explicitely initialize AND the one we want included
3366 * in the sampling buffer (smpl_regs).
3368 * Having this restriction allows optimization in the ctxsw routine
3369 * without compromising security (leaks)
3371 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3373 sval = ctx->ctx_pmds[cnum].val;
3374 lval = ctx->ctx_pmds[cnum].lval;
3375 is_counting = PMD_IS_COUNTING(cnum);
3378 * If the task is not the current one, then we check if the
3379 * PMU state is still in the local live register due to lazy ctxsw.
3380 * If true, then we read directly from the registers.
3382 if (can_access_pmu){
3383 val = ia64_get_pmd(cnum);
3386 * context has been saved
3387 * if context is zombie, then task does not exist anymore.
3388 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3390 val = is_loaded ? thread->pmds[cnum] : 0UL;
3392 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3396 * XXX: need to check for overflow when loaded
3403 * execute read checker, if any
3405 if (unlikely(expert_mode == 0 && rd_func)) {
3406 unsigned long v = val;
3407 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3408 if (ret) goto error;
3413 PFM_REG_RETFLAG_SET(reg_flags, 0);
3415 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3418 * update register return value, abort all if problem during copy.
3419 * we only modify the reg_flags field. no check mode is fine because
3420 * access has been verified upfront in sys_perfmonctl().
3422 req->reg_value = val;
3423 req->reg_flags = reg_flags;
3424 req->reg_last_reset_val = lval;
3430 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3435 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3439 if (req == NULL) return -EINVAL;
3441 ctx = GET_PMU_CTX();
3443 if (ctx == NULL) return -EINVAL;
3446 * for now limit to current task, which is enough when calling
3447 * from overflow handler
3449 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3451 return pfm_write_pmcs(ctx, req, nreq, regs);
3453 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3456 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3460 if (req == NULL) return -EINVAL;
3462 ctx = GET_PMU_CTX();
3464 if (ctx == NULL) return -EINVAL;
3467 * for now limit to current task, which is enough when calling
3468 * from overflow handler
3470 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3472 return pfm_read_pmds(ctx, req, nreq, regs);
3474 EXPORT_SYMBOL(pfm_mod_read_pmds);
3477 * Only call this function when a process it trying to
3478 * write the debug registers (reading is always allowed)
3481 pfm_use_debug_registers(struct task_struct *task)
3483 pfm_context_t *ctx = task->thread.pfm_context;
3484 unsigned long flags;
3487 if (pmu_conf->use_rr_dbregs == 0) return 0;
3489 DPRINT(("called for [%d]\n", task->pid));
3494 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3497 * Even on SMP, we do not need to use an atomic here because
3498 * the only way in is via ptrace() and this is possible only when the
3499 * process is stopped. Even in the case where the ctxsw out is not totally
3500 * completed by the time we come here, there is no way the 'stopped' process
3501 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3502 * So this is always safe.
3504 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3509 * We cannot allow setting breakpoints when system wide monitoring
3510 * sessions are using the debug registers.
3512 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3515 pfm_sessions.pfs_ptrace_use_dbregs++;
3517 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3518 pfm_sessions.pfs_ptrace_use_dbregs,
3519 pfm_sessions.pfs_sys_use_dbregs,
3528 * This function is called for every task that exits with the
3529 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3530 * able to use the debug registers for debugging purposes via
3531 * ptrace(). Therefore we know it was not using them for
3532 * perfmormance monitoring, so we only decrement the number
3533 * of "ptraced" debug register users to keep the count up to date
3536 pfm_release_debug_registers(struct task_struct *task)
3538 unsigned long flags;
3541 if (pmu_conf->use_rr_dbregs == 0) return 0;
3544 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3545 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3548 pfm_sessions.pfs_ptrace_use_dbregs--;
3557 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3559 struct task_struct *task;
3560 pfm_buffer_fmt_t *fmt;
3561 pfm_ovfl_ctrl_t rst_ctrl;
3562 int state, is_system;
3565 state = ctx->ctx_state;
3566 fmt = ctx->ctx_buf_fmt;
3567 is_system = ctx->ctx_fl_system;
3568 task = PFM_CTX_TASK(ctx);
3571 case PFM_CTX_MASKED:
3573 case PFM_CTX_LOADED:
3574 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3576 case PFM_CTX_UNLOADED:
3577 case PFM_CTX_ZOMBIE:
3578 DPRINT(("invalid state=%d\n", state));
3581 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3586 * In system wide and when the context is loaded, access can only happen
3587 * when the caller is running on the CPU being monitored by the session.
3588 * It does not have to be the owner (ctx_task) of the context per se.
3590 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3591 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3596 if (unlikely(task == NULL)) {
3597 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3601 if (task == current || is_system) {
3603 fmt = ctx->ctx_buf_fmt;
3605 DPRINT(("restarting self %d ovfl=0x%lx\n",
3607 ctx->ctx_ovfl_regs[0]));
3609 if (CTX_HAS_SMPL(ctx)) {
3611 prefetch(ctx->ctx_smpl_hdr);
3613 rst_ctrl.bits.mask_monitoring = 0;
3614 rst_ctrl.bits.reset_ovfl_pmds = 0;
3616 if (state == PFM_CTX_LOADED)
3617 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3619 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3621 rst_ctrl.bits.mask_monitoring = 0;
3622 rst_ctrl.bits.reset_ovfl_pmds = 1;
3626 if (rst_ctrl.bits.reset_ovfl_pmds)
3627 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3629 if (rst_ctrl.bits.mask_monitoring == 0) {
3630 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3632 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3634 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3636 // cannot use pfm_stop_monitoring(task, regs);
3640 * clear overflowed PMD mask to remove any stale information
3642 ctx->ctx_ovfl_regs[0] = 0UL;
3645 * back to LOADED state
3647 ctx->ctx_state = PFM_CTX_LOADED;
3650 * XXX: not really useful for self monitoring
3652 ctx->ctx_fl_can_restart = 0;
3658 * restart another task
3662 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3663 * one is seen by the task.
3665 if (state == PFM_CTX_MASKED) {
3666 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3668 * will prevent subsequent restart before this one is
3669 * seen by other task
3671 ctx->ctx_fl_can_restart = 0;
3675 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3676 * the task is blocked or on its way to block. That's the normal
3677 * restart path. If the monitoring is not masked, then the task
3678 * can be actively monitoring and we cannot directly intervene.
3679 * Therefore we use the trap mechanism to catch the task and
3680 * force it to reset the buffer/reset PMDs.
3682 * if non-blocking, then we ensure that the task will go into
3683 * pfm_handle_work() before returning to user mode.
3685 * We cannot explicitely reset another task, it MUST always
3686 * be done by the task itself. This works for system wide because
3687 * the tool that is controlling the session is logically doing
3688 * "self-monitoring".
3690 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3691 DPRINT(("unblocking [%d] \n", task->pid));
3692 up(&ctx->ctx_restart_sem);
3694 DPRINT(("[%d] armed exit trap\n", task->pid));
3696 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3698 PFM_SET_WORK_PENDING(task, 1);
3700 pfm_set_task_notify(task);
3703 * XXX: send reschedule if task runs on another CPU
3710 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3712 unsigned int m = *(unsigned int *)arg;
3714 pfm_sysctl.debug = m == 0 ? 0 : 1;
3716 pfm_debug_var = pfm_sysctl.debug;
3718 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3721 memset(pfm_stats, 0, sizeof(pfm_stats));
3722 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3728 * arg can be NULL and count can be zero for this function
3731 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3733 struct thread_struct *thread = NULL;
3734 struct task_struct *task;
3735 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3736 unsigned long flags;
3741 int i, can_access_pmu = 0;
3742 int is_system, is_loaded;
3744 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3746 state = ctx->ctx_state;
3747 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3748 is_system = ctx->ctx_fl_system;
3749 task = ctx->ctx_task;
3751 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3754 * on both UP and SMP, we can only write to the PMC when the task is
3755 * the owner of the local PMU.
3758 thread = &task->thread;
3760 * In system wide and when the context is loaded, access can only happen
3761 * when the caller is running on the CPU being monitored by the session.
3762 * It does not have to be the owner (ctx_task) of the context per se.
3764 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3765 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3768 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3772 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3773 * ensuring that no real breakpoint can be installed via this call.
3775 * IMPORTANT: regs can be NULL in this function
3778 first_time = ctx->ctx_fl_using_dbreg == 0;
3781 * don't bother if we are loaded and task is being debugged
3783 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3784 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3789 * check for debug registers in system wide mode
3791 * If though a check is done in pfm_context_load(),
3792 * we must repeat it here, in case the registers are
3793 * written after the context is loaded
3798 if (first_time && is_system) {
3799 if (pfm_sessions.pfs_ptrace_use_dbregs)
3802 pfm_sessions.pfs_sys_use_dbregs++;
3807 if (ret != 0) return ret;
3810 * mark ourself as user of the debug registers for
3813 ctx->ctx_fl_using_dbreg = 1;
3816 * clear hardware registers to make sure we don't
3817 * pick up stale state.
3819 * for a system wide session, we do not use
3820 * thread.dbr, thread.ibr because this process
3821 * never leaves the current CPU and the state
3822 * is shared by all processes running on it
3824 if (first_time && can_access_pmu) {
3825 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3826 for (i=0; i < pmu_conf->num_ibrs; i++) {
3827 ia64_set_ibr(i, 0UL);
3828 ia64_dv_serialize_instruction();
3831 for (i=0; i < pmu_conf->num_dbrs; i++) {
3832 ia64_set_dbr(i, 0UL);
3833 ia64_dv_serialize_data();
3839 * Now install the values into the registers
3841 for (i = 0; i < count; i++, req++) {
3843 rnum = req->dbreg_num;
3844 dbreg.val = req->dbreg_value;
3848 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3849 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3850 rnum, dbreg.val, mode, i, count));
3856 * make sure we do not install enabled breakpoint
3859 if (mode == PFM_CODE_RR)
3860 dbreg.ibr.ibr_x = 0;
3862 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3865 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3868 * Debug registers, just like PMC, can only be modified
3869 * by a kernel call. Moreover, perfmon() access to those
3870 * registers are centralized in this routine. The hardware
3871 * does not modify the value of these registers, therefore,
3872 * if we save them as they are written, we can avoid having
3873 * to save them on context switch out. This is made possible
3874 * by the fact that when perfmon uses debug registers, ptrace()
3875 * won't be able to modify them concurrently.
3877 if (mode == PFM_CODE_RR) {
3878 CTX_USED_IBR(ctx, rnum);
3880 if (can_access_pmu) {
3881 ia64_set_ibr(rnum, dbreg.val);
3882 ia64_dv_serialize_instruction();
3885 ctx->ctx_ibrs[rnum] = dbreg.val;
3887 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x is_loaded=%d access_pmu=%d\n",
3888 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3890 CTX_USED_DBR(ctx, rnum);
3892 if (can_access_pmu) {
3893 ia64_set_dbr(rnum, dbreg.val);
3894 ia64_dv_serialize_data();
3896 ctx->ctx_dbrs[rnum] = dbreg.val;
3898 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x is_loaded=%d access_pmu=%d\n",
3899 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3907 * in case it was our first attempt, we undo the global modifications
3911 if (ctx->ctx_fl_system) {
3912 pfm_sessions.pfs_sys_use_dbregs--;
3915 ctx->ctx_fl_using_dbreg = 0;
3918 * install error return flag
3920 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3926 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3928 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3932 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3934 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3938 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3942 if (req == NULL) return -EINVAL;
3944 ctx = GET_PMU_CTX();
3946 if (ctx == NULL) return -EINVAL;
3949 * for now limit to current task, which is enough when calling
3950 * from overflow handler
3952 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3954 return pfm_write_ibrs(ctx, req, nreq, regs);
3956 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3959 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3963 if (req == NULL) return -EINVAL;
3965 ctx = GET_PMU_CTX();
3967 if (ctx == NULL) return -EINVAL;
3970 * for now limit to current task, which is enough when calling
3971 * from overflow handler
3973 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3975 return pfm_write_dbrs(ctx, req, nreq, regs);
3977 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3981 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3983 pfarg_features_t *req = (pfarg_features_t *)arg;
3985 req->ft_version = PFM_VERSION;
3990 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3992 struct pt_regs *tregs;
3993 struct task_struct *task = PFM_CTX_TASK(ctx);
3994 int state, is_system;
3996 state = ctx->ctx_state;
3997 is_system = ctx->ctx_fl_system;
3999 if (state != PFM_CTX_LOADED && state != PFM_CTX_MASKED) return -EINVAL;
4002 * In system wide and when the context is loaded, access can only happen
4003 * when the caller is running on the CPU being monitored by the session.
4004 * It does not have to be the owner (ctx_task) of the context per se.
4006 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4007 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4010 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4011 PFM_CTX_TASK(ctx)->pid,
4015 * in system mode, we need to update the PMU directly
4016 * and the user level state of the caller, which may not
4017 * necessarily be the creator of the context.
4021 * Update local PMU first
4025 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4029 * update local cpuinfo
4031 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4034 * stop monitoring, does srlz.i
4039 * stop monitoring in the caller
4041 ia64_psr(regs)->pp = 0;
4049 if (task == current) {
4050 /* stop monitoring at kernel level */
4054 * stop monitoring at the user level
4056 ia64_psr(regs)->up = 0;
4058 tregs = ia64_task_regs(task);
4061 * stop monitoring at the user level
4063 ia64_psr(tregs)->up = 0;
4066 * monitoring disabled in kernel at next reschedule
4068 ctx->ctx_saved_psr_up = 0;
4069 DPRINT(("task=[%d]\n", task->pid));
4076 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4078 struct pt_regs *tregs;
4079 int state, is_system;
4081 state = ctx->ctx_state;
4082 is_system = ctx->ctx_fl_system;
4084 if (state != PFM_CTX_LOADED) return -EINVAL;
4087 * In system wide and when the context is loaded, access can only happen
4088 * when the caller is running on the CPU being monitored by the session.
4089 * It does not have to be the owner (ctx_task) of the context per se.
4091 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4092 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4097 * in system mode, we need to update the PMU directly
4098 * and the user level state of the caller, which may not
4099 * necessarily be the creator of the context.
4104 * set user level psr.pp for the caller
4106 ia64_psr(regs)->pp = 1;
4109 * now update the local PMU and cpuinfo
4111 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4114 * start monitoring at kernel level
4119 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4129 if (ctx->ctx_task == current) {
4131 /* start monitoring at kernel level */
4135 * activate monitoring at user level
4137 ia64_psr(regs)->up = 1;
4140 tregs = ia64_task_regs(ctx->ctx_task);
4143 * start monitoring at the kernel level the next
4144 * time the task is scheduled
4146 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4149 * activate monitoring at user level
4151 ia64_psr(tregs)->up = 1;
4157 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4159 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4164 for (i = 0; i < count; i++, req++) {
4166 cnum = req->reg_num;
4168 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4170 req->reg_value = PMC_DFL_VAL(cnum);
4172 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4174 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4179 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4184 pfm_check_task_exist(pfm_context_t *ctx)
4186 struct task_struct *g, *t;
4189 read_lock(&tasklist_lock);
4191 do_each_thread (g, t) {
4192 if (t->thread.pfm_context == ctx) {
4196 } while_each_thread (g, t);
4198 read_unlock(&tasklist_lock);
4200 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4206 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4208 struct task_struct *task;
4209 struct thread_struct *thread;
4210 struct pfm_context_t *old;
4211 unsigned long flags;
4213 struct task_struct *owner_task = NULL;
4215 pfarg_load_t *req = (pfarg_load_t *)arg;
4216 unsigned long *pmcs_source, *pmds_source;
4219 int state, is_system, set_dbregs = 0;
4221 state = ctx->ctx_state;
4222 is_system = ctx->ctx_fl_system;
4224 * can only load from unloaded or terminated state
4226 if (state != PFM_CTX_UNLOADED) {
4227 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4233 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4235 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4236 DPRINT(("cannot use blocking mode on self\n"));
4240 ret = pfm_get_task(ctx, req->load_pid, &task);
4242 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4249 * system wide is self monitoring only
4251 if (is_system && task != current) {
4252 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4257 thread = &task->thread;
4261 * cannot load a context which is using range restrictions,
4262 * into a task that is being debugged.
4264 if (ctx->ctx_fl_using_dbreg) {
4265 if (thread->flags & IA64_THREAD_DBG_VALID) {
4267 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4273 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4274 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4277 pfm_sessions.pfs_sys_use_dbregs++;
4278 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4285 if (ret) goto error;
4289 * SMP system-wide monitoring implies self-monitoring.
4291 * The programming model expects the task to
4292 * be pinned on a CPU throughout the session.
4293 * Here we take note of the current CPU at the
4294 * time the context is loaded. No call from
4295 * another CPU will be allowed.
4297 * The pinning via shed_setaffinity()
4298 * must be done by the calling task prior
4301 * systemwide: keep track of CPU this session is supposed to run on
4303 the_cpu = ctx->ctx_cpu = smp_processor_id();
4307 * now reserve the session
4309 ret = pfm_reserve_session(current, is_system, the_cpu);
4310 if (ret) goto error;
4313 * task is necessarily stopped at this point.
4315 * If the previous context was zombie, then it got removed in
4316 * pfm_save_regs(). Therefore we should not see it here.
4317 * If we see a context, then this is an active context
4319 * XXX: needs to be atomic
4321 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4322 thread->pfm_context, ctx));
4324 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4326 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4330 pfm_reset_msgq(ctx);
4332 ctx->ctx_state = PFM_CTX_LOADED;
4335 * link context to task
4337 ctx->ctx_task = task;
4341 * we load as stopped
4343 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4344 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4346 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4348 thread->flags |= IA64_THREAD_PM_VALID;
4352 * propagate into thread-state
4354 pfm_copy_pmds(task, ctx);
4355 pfm_copy_pmcs(task, ctx);
4357 pmcs_source = thread->pmcs;
4358 pmds_source = thread->pmds;
4361 * always the case for system-wide
4363 if (task == current) {
4365 if (is_system == 0) {
4367 /* allow user level control */
4368 ia64_psr(regs)->sp = 0;
4369 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4371 SET_LAST_CPU(ctx, smp_processor_id());
4373 SET_ACTIVATION(ctx);
4376 * push the other task out, if any
4378 owner_task = GET_PMU_OWNER();
4379 if (owner_task) pfm_lazy_save_regs(owner_task);
4383 * load all PMD from ctx to PMU (as opposed to thread state)
4384 * restore all PMC from ctx to PMU
4386 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4387 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4389 ctx->ctx_reload_pmcs[0] = 0UL;
4390 ctx->ctx_reload_pmds[0] = 0UL;
4393 * guaranteed safe by earlier check against DBG_VALID
4395 if (ctx->ctx_fl_using_dbreg) {
4396 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4397 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4402 SET_PMU_OWNER(task, ctx);
4404 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4407 * when not current, task MUST be stopped, so this is safe
4409 regs = ia64_task_regs(task);
4411 /* force a full reload */
4412 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4413 SET_LAST_CPU(ctx, -1);
4415 /* initial saved psr (stopped) */
4416 ctx->ctx_saved_psr_up = 0UL;
4417 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4423 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4426 * we must undo the dbregs setting (for system-wide)
4428 if (ret && set_dbregs) {
4430 pfm_sessions.pfs_sys_use_dbregs--;
4434 * release task, there is now a link with the context
4436 if (is_system == 0 && task != current) {
4440 ret = pfm_check_task_exist(ctx);
4442 ctx->ctx_state = PFM_CTX_UNLOADED;
4443 ctx->ctx_task = NULL;
4451 * in this function, we do not need to increase the use count
4452 * for the task via get_task_struct(), because we hold the
4453 * context lock. If the task were to disappear while having
4454 * a context attached, it would go through pfm_exit_thread()
4455 * which also grabs the context lock and would therefore be blocked
4456 * until we are here.
4458 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4461 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4463 struct task_struct *task = PFM_CTX_TASK(ctx);
4464 struct pt_regs *tregs;
4465 int prev_state, is_system;
4468 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4470 prev_state = ctx->ctx_state;
4471 is_system = ctx->ctx_fl_system;
4474 * unload only when necessary
4476 if (prev_state == PFM_CTX_UNLOADED) {
4477 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4482 * clear psr and dcr bits
4484 ret = pfm_stop(ctx, NULL, 0, regs);
4485 if (ret) return ret;
4487 ctx->ctx_state = PFM_CTX_UNLOADED;
4490 * in system mode, we need to update the PMU directly
4491 * and the user level state of the caller, which may not
4492 * necessarily be the creator of the context.
4499 * local PMU is taken care of in pfm_stop()
4501 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4502 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4505 * save PMDs in context
4508 pfm_flush_pmds(current, ctx);
4511 * at this point we are done with the PMU
4512 * so we can unreserve the resource.
4514 if (prev_state != PFM_CTX_ZOMBIE)
4515 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4518 * disconnect context from task
4520 task->thread.pfm_context = NULL;
4522 * disconnect task from context
4524 ctx->ctx_task = NULL;
4527 * There is nothing more to cleanup here.
4535 tregs = task == current ? regs : ia64_task_regs(task);
4537 if (task == current) {
4539 * cancel user level control
4541 ia64_psr(regs)->sp = 1;
4543 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4546 * save PMDs to context
4549 pfm_flush_pmds(task, ctx);
4552 * at this point we are done with the PMU
4553 * so we can unreserve the resource.
4555 * when state was ZOMBIE, we have already unreserved.
4557 if (prev_state != PFM_CTX_ZOMBIE)
4558 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4561 * reset activation counter and psr
4563 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4564 SET_LAST_CPU(ctx, -1);
4567 * PMU state will not be restored
4569 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4572 * break links between context and task
4574 task->thread.pfm_context = NULL;
4575 ctx->ctx_task = NULL;
4577 PFM_SET_WORK_PENDING(task, 0);
4579 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4580 ctx->ctx_fl_can_restart = 0;
4581 ctx->ctx_fl_going_zombie = 0;
4583 DPRINT(("disconnected [%d] from context\n", task->pid));
4589 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
4591 struct task_struct *task = ctx->ctx_task;
4593 ia64_psr(regs)->up = 0;
4594 ia64_psr(regs)->sp = 1;
4596 if (GET_PMU_OWNER() == task) {
4597 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
4598 SET_PMU_OWNER(NULL, NULL);
4602 * disconnect the task from the context and vice-versa
4604 PFM_SET_WORK_PENDING(task, 0);
4606 task->thread.pfm_context = NULL;
4607 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4609 DPRINT(("force cleanupf for [%d]\n", task->pid));
4615 * called only from exit_thread(): task == current
4616 * we come here only if current has a context attached (loaded or masked)
4619 pfm_exit_thread(struct task_struct *task)
4622 unsigned long flags;
4623 struct pt_regs *regs = ia64_task_regs(task);
4627 ctx = PFM_GET_CTX(task);
4629 PROTECT_CTX(ctx, flags);
4631 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4633 state = ctx->ctx_state;
4635 case PFM_CTX_UNLOADED:
4637 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4638 * be in unloaded state
4640 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4642 case PFM_CTX_LOADED:
4643 case PFM_CTX_MASKED:
4644 ret = pfm_context_unload(ctx, NULL, 0, regs);
4646 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4648 DPRINT(("ctx unloaded for current state was %d\n", state));
4650 pfm_end_notify_user(ctx);
4652 case PFM_CTX_ZOMBIE:
4653 ret = pfm_context_unload(ctx, NULL, 0, regs);
4655 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4660 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4663 UNPROTECT_CTX(ctx, flags);
4665 { u64 psr = pfm_get_psr();
4666 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4667 BUG_ON(GET_PMU_OWNER());
4668 BUG_ON(ia64_psr(regs)->up);
4669 BUG_ON(ia64_psr(regs)->pp);
4673 * All memory free operations (especially for vmalloc'ed memory)
4674 * MUST be done with interrupts ENABLED.
4676 if (free_ok) pfm_context_free(ctx);
4680 * functions MUST be listed in the increasing order of their index (see permfon.h)
4682 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4683 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4684 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4685 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4686 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4688 static pfm_cmd_desc_t pfm_cmd_tab[]={
4689 /* 0 */PFM_CMD_NONE,
4690 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4691 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4692 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4693 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4694 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4695 /* 6 */PFM_CMD_NONE,
4696 /* 7 */PFM_CMD_NONE,
4697 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4698 /* 9 */PFM_CMD_NONE,
4699 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4700 /* 11 */PFM_CMD_NONE,
4701 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4702 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4703 /* 14 */PFM_CMD_NONE,
4704 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4705 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4706 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4707 /* 18 */PFM_CMD_NONE,
4708 /* 19 */PFM_CMD_NONE,
4709 /* 20 */PFM_CMD_NONE,
4710 /* 21 */PFM_CMD_NONE,
4711 /* 22 */PFM_CMD_NONE,
4712 /* 23 */PFM_CMD_NONE,
4713 /* 24 */PFM_CMD_NONE,
4714 /* 25 */PFM_CMD_NONE,
4715 /* 26 */PFM_CMD_NONE,
4716 /* 27 */PFM_CMD_NONE,
4717 /* 28 */PFM_CMD_NONE,
4718 /* 29 */PFM_CMD_NONE,
4719 /* 30 */PFM_CMD_NONE,
4720 /* 31 */PFM_CMD_NONE,
4721 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4722 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4724 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4727 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4729 struct task_struct *task;
4732 state = ctx->ctx_state;
4734 task = PFM_CTX_TASK(ctx);
4736 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4740 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4744 task->state, PFM_CMD_STOPPED(cmd)));
4747 * self-monitoring always ok.
4749 * for system-wide the caller can either be the creator of the
4750 * context (to one to which the context is attached to) OR
4751 * a task running on the same CPU as the session.
4753 if (task == current || ctx->ctx_fl_system) return 0;
4756 * context is UNLOADED, MASKED we are safe to go
4758 if (state != PFM_CTX_LOADED) return 0;
4760 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
4763 * context is loaded, we must make sure the task is stopped
4764 * We could lift this restriction for UP but it would mean that
4765 * the user has no guarantee the task would not run between
4766 * two successive calls to perfmonctl(). That's probably OK.
4767 * If this user wants to ensure the task does not run, then
4768 * the task must be stopped.
4770 if (PFM_CMD_STOPPED(cmd) && task->state != TASK_STOPPED) {
4771 DPRINT(("[%d] task not in stopped state\n", task->pid));
4775 UNPROTECT_CTX(ctx, flags);
4777 wait_task_inactive(task);
4779 PROTECT_CTX(ctx, flags);
4785 * system-call entry point (must return long)
4788 sys_perfmonctl (int fd, int cmd, void *arg, int count, long arg5, long arg6, long arg7,
4789 long arg8, long stack)
4791 struct pt_regs *regs = (struct pt_regs *)&stack;
4792 struct file *file = NULL;
4793 pfm_context_t *ctx = NULL;
4794 unsigned long flags = 0UL;
4795 void *args_k = NULL;
4796 long ret; /* will expand int return types */
4797 size_t base_sz, sz, xtra_sz = 0;
4798 int narg, completed_args = 0, call_made = 0, cmd_flags;
4799 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4800 int (*getsize)(void *arg, size_t *sz);
4801 #define PFM_MAX_ARGSIZE 4096
4804 * reject any call if perfmon was disabled at initialization
4806 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4808 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4809 DPRINT(("invalid cmd=%d\n", cmd));
4813 func = pfm_cmd_tab[cmd].cmd_func;
4814 narg = pfm_cmd_tab[cmd].cmd_narg;
4815 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4816 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4817 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4819 if (unlikely(func == NULL)) {
4820 DPRINT(("invalid cmd=%d\n", cmd));
4824 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4832 * check if number of arguments matches what the command expects
4834 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4838 sz = xtra_sz + base_sz*count;
4840 * limit abuse to min page size
4842 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4843 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4848 * allocate default-sized argument buffer
4850 if (likely(count && args_k == NULL)) {
4851 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4852 if (args_k == NULL) return -ENOMEM;
4860 * assume sz = 0 for command without parameters
4862 if (sz && copy_from_user(args_k, arg, sz)) {
4863 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4868 * check if command supports extra parameters
4870 if (completed_args == 0 && getsize) {
4872 * get extra parameters size (based on main argument)
4874 ret = (*getsize)(args_k, &xtra_sz);
4875 if (ret) goto error_args;
4879 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4881 /* retry if necessary */
4882 if (likely(xtra_sz)) goto restart_args;
4885 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4890 if (unlikely(file == NULL)) {
4891 DPRINT(("invalid fd %d\n", fd));
4894 if (unlikely(PFM_IS_FILE(file) == 0)) {
4895 DPRINT(("fd %d not related to perfmon\n", fd));
4899 ctx = (pfm_context_t *)file->private_data;
4900 if (unlikely(ctx == NULL)) {
4901 DPRINT(("no context for fd %d\n", fd));
4904 prefetch(&ctx->ctx_state);
4906 PROTECT_CTX(ctx, flags);
4909 * check task is stopped
4911 ret = pfm_check_task_state(ctx, cmd, flags);
4912 if (unlikely(ret)) goto abort_locked;
4915 ret = (*func)(ctx, args_k, count, regs);
4921 DPRINT(("context unlocked\n"));
4922 UNPROTECT_CTX(ctx, flags);
4926 /* copy argument back to user, if needed */
4927 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4930 if (args_k) kfree(args_k);
4932 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4938 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4940 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4941 pfm_ovfl_ctrl_t rst_ctrl;
4945 state = ctx->ctx_state;
4947 * Unlock sampling buffer and reset index atomically
4948 * XXX: not really needed when blocking
4950 if (CTX_HAS_SMPL(ctx)) {
4952 rst_ctrl.bits.mask_monitoring = 0;
4953 rst_ctrl.bits.reset_ovfl_pmds = 0;
4955 if (state == PFM_CTX_LOADED)
4956 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4958 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4960 rst_ctrl.bits.mask_monitoring = 0;
4961 rst_ctrl.bits.reset_ovfl_pmds = 1;
4965 if (rst_ctrl.bits.reset_ovfl_pmds) {
4966 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4968 if (rst_ctrl.bits.mask_monitoring == 0) {
4969 DPRINT(("resuming monitoring\n"));
4970 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4972 DPRINT(("stopping monitoring\n"));
4973 //pfm_stop_monitoring(current, regs);
4975 ctx->ctx_state = PFM_CTX_LOADED;
4980 * context MUST BE LOCKED when calling
4981 * can only be called for current
4984 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4986 if (ctx->ctx_fl_system) {
4987 printk(KERN_ERR "perfmon: pfm_context_force_terminate [%d] is system-wide\n", current->pid);
4991 * we stop the whole thing, we do no need to flush
4992 * we know we WERE masked
4995 ia64_psr(regs)->up = 0;
4996 ia64_psr(regs)->sp = 1;
4999 * disconnect the task from the context and vice-versa
5001 current->thread.pfm_context = NULL;
5002 current->thread.flags &= ~IA64_THREAD_PM_VALID;
5003 ctx->ctx_task = NULL;
5005 DPRINT(("context terminated\n"));
5008 * and wakeup controlling task, indicating we are now disconnected
5010 wake_up_interruptible(&ctx->ctx_zombieq);
5013 * given that context is still locked, the controlling
5014 * task will only get access when we return from
5015 * pfm_handle_work().
5019 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5022 pfm_handle_work(void)
5025 struct pt_regs *regs;
5026 unsigned long flags;
5027 unsigned long ovfl_regs;
5028 unsigned int reason;
5031 ctx = PFM_GET_CTX(current);
5033 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5037 PROTECT_CTX(ctx, flags);
5039 PFM_SET_WORK_PENDING(current, 0);
5041 pfm_clear_task_notify();
5043 regs = ia64_task_regs(current);
5046 * extract reason for being here and clear
5048 reason = ctx->ctx_fl_trap_reason;
5049 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5050 ovfl_regs = ctx->ctx_ovfl_regs[0];
5052 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5055 * must be done before we check for simple-reset mode
5057 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5060 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5061 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5063 UNPROTECT_CTX(ctx, flags);
5065 DPRINT(("before block sleeping\n"));
5068 * may go through without blocking on SMP systems
5069 * if restart has been received already by the time we call down()
5071 ret = down_interruptible(&ctx->ctx_restart_sem);
5073 DPRINT(("after block sleeping ret=%d\n", ret));
5075 PROTECT_CTX(ctx, flags);
5078 * we need to read the ovfl_regs only after wake-up
5079 * because we may have had pfm_write_pmds() in between
5080 * and that can changed PMD values and therefore
5081 * ovfl_regs is reset for these new PMD values.
5083 ovfl_regs = ctx->ctx_ovfl_regs[0];
5085 if (ctx->ctx_fl_going_zombie) {
5087 DPRINT(("context is zombie, bailing out\n"));
5088 pfm_context_force_terminate(ctx, regs);
5092 * in case of interruption of down() we don't restart anything
5094 if (ret < 0) goto nothing_to_do;
5097 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5098 ctx->ctx_ovfl_regs[0] = 0UL;
5102 UNPROTECT_CTX(ctx, flags);
5106 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5108 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5109 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5113 DPRINT(("waking up somebody\n"));
5115 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5118 * safe, we are not in intr handler, nor in ctxsw when
5121 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5127 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5129 pfm_msg_t *msg = NULL;
5131 if (ctx->ctx_fl_no_msg == 0) {
5132 msg = pfm_get_new_msg(ctx);
5134 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5138 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5139 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5140 msg->pfm_ovfl_msg.msg_active_set = 0;
5141 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5142 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5143 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5144 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5145 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5148 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5154 return pfm_notify_user(ctx, msg);
5158 pfm_end_notify_user(pfm_context_t *ctx)
5162 msg = pfm_get_new_msg(ctx);
5164 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5168 memset(msg, 0, sizeof(*msg));
5170 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5171 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5172 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5174 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5179 return pfm_notify_user(ctx, msg);
5183 * main overflow processing routine.
5184 * it can be called from the interrupt path or explicitely during the context switch code
5187 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5189 pfm_ovfl_arg_t ovfl_arg;
5191 unsigned long old_val, ovfl_val, new_val;
5192 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5193 unsigned long tstamp;
5194 pfm_ovfl_ctrl_t ovfl_ctrl;
5195 unsigned int i, has_smpl;
5196 int must_notify = 0;
5198 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5201 * sanity test. Should never happen
5203 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5205 tstamp = ia64_get_itc();
5206 mask = pmc0 >> PMU_FIRST_COUNTER;
5207 ovfl_val = pmu_conf->ovfl_val;
5208 has_smpl = CTX_HAS_SMPL(ctx);
5210 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5211 "used_pmds=0x%lx\n",
5213 task ? task->pid: -1,
5214 (regs ? regs->cr_iip : 0),
5215 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5216 ctx->ctx_used_pmds[0]));
5220 * first we update the virtual counters
5221 * assume there was a prior ia64_srlz_d() issued
5223 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5225 /* skip pmd which did not overflow */
5226 if ((mask & 0x1) == 0) continue;
5229 * Note that the pmd is not necessarily 0 at this point as qualified events
5230 * may have happened before the PMU was frozen. The residual count is not
5231 * taken into consideration here but will be with any read of the pmd via
5234 old_val = new_val = ctx->ctx_pmds[i].val;
5235 new_val += 1 + ovfl_val;
5236 ctx->ctx_pmds[i].val = new_val;
5239 * check for overflow condition
5241 if (likely(old_val > new_val)) {
5242 ovfl_pmds |= 1UL << i;
5243 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5246 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5250 ia64_get_pmd(i) & ovfl_val,
5256 * there was no 64-bit overflow, nothing else to do
5258 if (ovfl_pmds == 0UL) return;
5261 * reset all control bits
5267 * if a sampling format module exists, then we "cache" the overflow by
5268 * calling the module's handler() routine.
5271 unsigned long start_cycles, end_cycles;
5272 unsigned long pmd_mask;
5274 int this_cpu = smp_processor_id();
5276 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5278 prefetch(ctx->ctx_smpl_hdr);
5280 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5284 if ((pmd_mask & 0x1) == 0) continue;
5286 ovfl_arg.ovfl_pmd = (unsigned char )i;
5287 ovfl_arg.ovfl_notify = ovfl_notify & mask ? 1 : 0;
5288 ovfl_arg.active_set = 0;
5289 ovfl_arg.ovfl_ctrl.val = 0; /* module must fill in all fields */
5290 ovfl_arg.smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5292 ovfl_arg.pmd_value = ctx->ctx_pmds[i].val;
5293 ovfl_arg.pmd_last_reset = ctx->ctx_pmds[i].lval;
5294 ovfl_arg.pmd_eventid = ctx->ctx_pmds[i].eventid;
5297 * copy values of pmds of interest. Sampling format may copy them
5298 * into sampling buffer.
5301 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5302 if ((smpl_pmds & 0x1) == 0) continue;
5303 ovfl_arg.smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5304 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg.smpl_pmds_values[k-1]));
5308 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5310 start_cycles = ia64_get_itc();
5313 * call custom buffer format record (handler) routine
5315 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, &ovfl_arg, regs, tstamp);
5317 end_cycles = ia64_get_itc();
5320 * For those controls, we take the union because they have
5321 * an all or nothing behavior.
5323 ovfl_ctrl.bits.notify_user |= ovfl_arg.ovfl_ctrl.bits.notify_user;
5324 ovfl_ctrl.bits.block_task |= ovfl_arg.ovfl_ctrl.bits.block_task;
5325 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg.ovfl_ctrl.bits.mask_monitoring;
5327 * build the bitmask of pmds to reset now
5329 if (ovfl_arg.ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5331 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5334 * when the module cannot handle the rest of the overflows, we abort right here
5336 if (ret && pmd_mask) {
5337 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5338 pmd_mask<<PMU_FIRST_COUNTER));
5341 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5343 ovfl_pmds &= ~reset_pmds;
5346 * when no sampling module is used, then the default
5347 * is to notify on overflow if requested by user
5349 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5350 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5351 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5352 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5354 * if needed, we reset all overflowed pmds
5356 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5359 DPRINT(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n",
5363 * reset the requested PMD registers using the short reset values
5366 unsigned long bm = reset_pmds;
5367 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5370 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5372 * keep track of what to reset when unblocking
5374 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5377 * check for blocking context
5379 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5381 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5384 * set the perfmon specific checking pending work for the task
5386 PFM_SET_WORK_PENDING(task, 1);
5389 * when coming from ctxsw, current still points to the
5390 * previous task, therefore we must work with task and not current.
5392 pfm_set_task_notify(task);
5395 * defer until state is changed (shorten spin window). the context is locked
5396 * anyway, so the signal receiver would come spin for nothing.
5401 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5402 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5403 PFM_GET_WORK_PENDING(task),
5404 ctx->ctx_fl_trap_reason,
5407 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5409 * in case monitoring must be stopped, we toggle the psr bits
5411 if (ovfl_ctrl.bits.mask_monitoring) {
5412 pfm_mask_monitoring(task);
5413 ctx->ctx_state = PFM_CTX_MASKED;
5414 ctx->ctx_fl_can_restart = 1;
5418 * send notification now
5420 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5425 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5427 task ? task->pid : -1,
5433 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5434 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5435 * come here as zombie only if the task is the current task. In which case, we
5436 * can access the PMU hardware directly.
5438 * Note that zombies do have PM_VALID set. So here we do the minimal.
5440 * In case the context was zombified it could not be reclaimed at the time
5441 * the monitoring program exited. At this point, the PMU reservation has been
5442 * returned, the sampiing buffer has been freed. We must convert this call
5443 * into a spurious interrupt. However, we must also avoid infinite overflows
5444 * by stopping monitoring for this task. We can only come here for a per-task
5445 * context. All we need to do is to stop monitoring using the psr bits which
5446 * are always task private. By re-enabling secure montioring, we ensure that
5447 * the monitored task will not be able to re-activate monitoring.
5448 * The task will eventually be context switched out, at which point the context
5449 * will be reclaimed (that includes releasing ownership of the PMU).
5451 * So there might be a window of time where the number of per-task session is zero
5452 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5453 * context. This is safe because if a per-task session comes in, it will push this one
5454 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5455 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5456 * also push our zombie context out.
5458 * Overall pretty hairy stuff....
5460 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5462 ia64_psr(regs)->up = 0;
5463 ia64_psr(regs)->sp = 1;
5468 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5470 struct task_struct *task;
5472 unsigned long flags;
5474 int this_cpu = smp_processor_id();
5477 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5480 * srlz.d done before arriving here
5482 pmc0 = ia64_get_pmc(0);
5484 task = GET_PMU_OWNER();
5485 ctx = GET_PMU_CTX();
5488 * if we have some pending bits set
5489 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5491 if (PMC0_HAS_OVFL(pmc0) && task) {
5493 * we assume that pmc0.fr is always set here
5497 if (!ctx) goto report_spurious1;
5499 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5500 goto report_spurious2;
5502 PROTECT_CTX_NOPRINT(ctx, flags);
5504 pfm_overflow_handler(task, ctx, pmc0, regs);
5506 UNPROTECT_CTX_NOPRINT(ctx, flags);
5509 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5513 * keep it unfrozen at all times
5520 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5521 this_cpu, task->pid);
5525 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5533 pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5535 unsigned long start_cycles, total_cycles;
5536 unsigned long min, max;
5540 this_cpu = get_cpu();
5541 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5542 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5544 start_cycles = ia64_get_itc();
5546 ret = pfm_do_interrupt_handler(irq, arg, regs);
5548 total_cycles = ia64_get_itc();
5551 * don't measure spurious interrupts
5553 if (likely(ret == 0)) {
5554 total_cycles -= start_cycles;
5556 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5557 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5559 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5561 put_cpu_no_resched();
5566 * /proc/perfmon interface, for debug only
5569 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5572 pfm_proc_start(struct seq_file *m, loff_t *pos)
5575 return PFM_PROC_SHOW_HEADER;
5578 while (*pos <= NR_CPUS) {
5579 if (cpu_online(*pos - 1)) {
5580 return (void *)*pos;
5588 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5591 return pfm_proc_start(m, pos);
5595 pfm_proc_stop(struct seq_file *m, void *v)
5600 pfm_proc_show_header(struct seq_file *m)
5602 struct list_head * pos;
5603 pfm_buffer_fmt_t * entry;
5604 unsigned long flags;
5607 "perfmon version : %u.%u\n"
5610 "expert mode : %s\n"
5611 "ovfl_mask : 0x%lx\n"
5612 "PMU flags : 0x%x\n",
5613 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5615 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5616 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5623 "proc_sessions : %u\n"
5624 "sys_sessions : %u\n"
5625 "sys_use_dbregs : %u\n"
5626 "ptrace_use_dbregs : %u\n",
5627 pfm_sessions.pfs_task_sessions,
5628 pfm_sessions.pfs_sys_sessions,
5629 pfm_sessions.pfs_sys_use_dbregs,
5630 pfm_sessions.pfs_ptrace_use_dbregs);
5634 spin_lock(&pfm_buffer_fmt_lock);
5636 list_for_each(pos, &pfm_buffer_fmt_list) {
5637 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5638 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5649 entry->fmt_uuid[10],
5650 entry->fmt_uuid[11],
5651 entry->fmt_uuid[12],
5652 entry->fmt_uuid[13],
5653 entry->fmt_uuid[14],
5654 entry->fmt_uuid[15],
5657 spin_unlock(&pfm_buffer_fmt_lock);
5662 pfm_proc_show(struct seq_file *m, void *v)
5668 if (v == PFM_PROC_SHOW_HEADER) {
5669 pfm_proc_show_header(m);
5673 /* show info for CPU (v - 1) */
5677 "CPU%-2d overflow intrs : %lu\n"
5678 "CPU%-2d overflow cycles : %lu\n"
5679 "CPU%-2d overflow min : %lu\n"
5680 "CPU%-2d overflow max : %lu\n"
5681 "CPU%-2d smpl handler calls : %lu\n"
5682 "CPU%-2d smpl handler cycles : %lu\n"
5683 "CPU%-2d spurious intrs : %lu\n"
5684 "CPU%-2d replay intrs : %lu\n"
5685 "CPU%-2d syst_wide : %d\n"
5686 "CPU%-2d dcr_pp : %d\n"
5687 "CPU%-2d exclude idle : %d\n"
5688 "CPU%-2d owner : %d\n"
5689 "CPU%-2d context : %p\n"
5690 "CPU%-2d activations : %lu\n",
5691 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5692 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5693 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5694 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5695 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5696 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5697 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5698 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5699 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5700 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5701 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5702 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5703 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5704 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5706 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5708 psr = pfm_get_psr();
5713 "CPU%-2d psr : 0x%lx\n"
5714 "CPU%-2d pmc0 : 0x%lx\n",
5716 cpu, ia64_get_pmc(0));
5718 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5719 if (PMC_IS_COUNTING(i) == 0) continue;
5721 "CPU%-2d pmc%u : 0x%lx\n"
5722 "CPU%-2d pmd%u : 0x%lx\n",
5723 cpu, i, ia64_get_pmc(i),
5724 cpu, i, ia64_get_pmd(i));
5730 struct seq_operations pfm_seq_ops = {
5731 .start = pfm_proc_start,
5732 .next = pfm_proc_next,
5733 .stop = pfm_proc_stop,
5734 .show = pfm_proc_show
5738 pfm_proc_open(struct inode *inode, struct file *file)
5740 return seq_open(file, &pfm_seq_ops);
5745 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5746 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5747 * is active or inactive based on mode. We must rely on the value in
5748 * local_cpu_data->pfm_syst_info
5751 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5753 struct pt_regs *regs;
5755 unsigned long dcr_pp;
5757 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5760 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5761 * on every CPU, so we can rely on the pid to identify the idle task.
5763 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5764 regs = ia64_task_regs(task);
5765 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5769 * if monitoring has started
5772 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5774 * context switching in?
5777 /* mask monitoring for the idle task */
5778 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5784 * context switching out
5785 * restore monitoring for next task
5787 * Due to inlining this odd if-then-else construction generates
5790 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5798 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5801 pfm_save_regs(struct task_struct *task)
5804 struct thread_struct *t;
5805 unsigned long flags;
5809 ctx = PFM_GET_CTX(task);
5810 if (ctx == NULL) return;
5814 * we always come here with interrupts ALREADY disabled by
5815 * the scheduler. So we simply need to protect against concurrent
5816 * access, not CPU concurrency.
5818 flags = pfm_protect_ctx_ctxsw(ctx);
5820 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5821 struct pt_regs *regs = ia64_task_regs(task);
5825 pfm_force_cleanup(ctx, regs);
5827 BUG_ON(ctx->ctx_smpl_hdr);
5829 pfm_unprotect_ctx_ctxsw(ctx, flags);
5831 pfm_context_free(ctx);
5838 if (ctx->ctx_last_activation != GET_ACTIVATION()) {
5839 pfm_unprotect_ctx_ctxsw(ctx, flags);
5844 * save current PSR: needed because we modify it
5847 psr = pfm_get_psr();
5849 BUG_ON(psr & (IA64_PSR_I));
5853 * This is the last instruction which may generate an overflow
5855 * We do not need to set psr.sp because, it is irrelevant in kernel.
5856 * It will be restored from ipsr when going back to user level
5861 * keep a copy of psr.up (for reload)
5863 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5866 * release ownership of this PMU.
5867 * PM interrupts are masked, so nothing
5870 SET_PMU_OWNER(NULL, NULL);
5873 * we systematically save the PMD as we have no
5874 * guarantee we will be schedule at that same
5877 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5880 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5881 * we will need it on the restore path to check
5882 * for pending overflow.
5884 t->pmcs[0] = ia64_get_pmc(0);
5887 * unfreeze PMU if had pending overflows
5889 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5892 * finally, allow context access.
5893 * interrupts will still be masked after this call.
5895 pfm_unprotect_ctx_ctxsw(ctx, flags);
5898 #else /* !CONFIG_SMP */
5900 pfm_save_regs(struct task_struct *task)
5905 ctx = PFM_GET_CTX(task);
5906 if (ctx == NULL) return;
5909 * save current PSR: needed because we modify it
5911 psr = pfm_get_psr();
5913 BUG_ON(psr & (IA64_PSR_I));
5917 * This is the last instruction which may generate an overflow
5919 * We do not need to set psr.sp because, it is irrelevant in kernel.
5920 * It will be restored from ipsr when going back to user level
5925 * keep a copy of psr.up (for reload)
5927 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5931 pfm_lazy_save_regs (struct task_struct *task)
5934 struct thread_struct *t;
5935 unsigned long flags;
5937 { u64 psr = pfm_get_psr();
5938 BUG_ON(psr & IA64_PSR_UP);
5941 ctx = PFM_GET_CTX(task);
5945 * we need to mask PMU overflow here to
5946 * make sure that we maintain pmc0 until
5947 * we save it. overflow interrupts are
5948 * treated as spurious if there is no
5951 * XXX: I don't think this is necessary
5953 PROTECT_CTX(ctx,flags);
5956 * release ownership of this PMU.
5957 * must be done before we save the registers.
5959 * after this call any PMU interrupt is treated
5962 SET_PMU_OWNER(NULL, NULL);
5965 * save all the pmds we use
5967 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5970 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5971 * it is needed to check for pended overflow
5972 * on the restore path
5974 t->pmcs[0] = ia64_get_pmc(0);
5977 * unfreeze PMU if had pending overflows
5979 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5982 * now get can unmask PMU interrupts, they will
5983 * be treated as purely spurious and we will not
5984 * lose any information
5986 UNPROTECT_CTX(ctx,flags);
5988 #endif /* CONFIG_SMP */
5992 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5995 pfm_load_regs (struct task_struct *task)
5998 struct thread_struct *t;
5999 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6000 unsigned long flags;
6002 int need_irq_resend;
6004 ctx = PFM_GET_CTX(task);
6005 if (unlikely(ctx == NULL)) return;
6007 BUG_ON(GET_PMU_OWNER());
6011 * possible on unload
6013 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6016 * we always come here with interrupts ALREADY disabled by
6017 * the scheduler. So we simply need to protect against concurrent
6018 * access, not CPU concurrency.
6020 flags = pfm_protect_ctx_ctxsw(ctx);
6021 psr = pfm_get_psr();
6023 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6025 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6026 BUG_ON(psr & IA64_PSR_I);
6028 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6029 struct pt_regs *regs = ia64_task_regs(task);
6031 BUG_ON(ctx->ctx_smpl_hdr);
6033 pfm_force_cleanup(ctx, regs);
6035 pfm_unprotect_ctx_ctxsw(ctx, flags);
6038 * this one (kmalloc'ed) is fine with interrupts disabled
6040 pfm_context_free(ctx);
6046 * we restore ALL the debug registers to avoid picking up
6049 if (ctx->ctx_fl_using_dbreg) {
6050 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6051 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6054 * retrieve saved psr.up
6056 psr_up = ctx->ctx_saved_psr_up;
6059 * if we were the last user of the PMU on that CPU,
6060 * then nothing to do except restore psr
6062 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6065 * retrieve partial reload masks (due to user modifications)
6067 pmc_mask = ctx->ctx_reload_pmcs[0];
6068 pmd_mask = ctx->ctx_reload_pmds[0];
6072 * To avoid leaking information to the user level when psr.sp=0,
6073 * we must reload ALL implemented pmds (even the ones we don't use).
6074 * In the kernel we only allow PFM_READ_PMDS on registers which
6075 * we initialized or requested (sampling) so there is no risk there.
6077 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6080 * ALL accessible PMCs are systematically reloaded, unused registers
6081 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6082 * up stale configuration.
6084 * PMC0 is never in the mask. It is always restored separately.
6086 pmc_mask = ctx->ctx_all_pmcs[0];
6089 * when context is MASKED, we will restore PMC with plm=0
6090 * and PMD with stale information, but that's ok, nothing
6093 * XXX: optimize here
6095 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6096 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6099 * check for pending overflow at the time the state
6102 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6104 * reload pmc0 with the overflow information
6105 * On McKinley PMU, this will trigger a PMU interrupt
6107 ia64_set_pmc(0, t->pmcs[0]);
6112 * will replay the PMU interrupt
6114 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6116 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6120 * we just did a reload, so we reset the partial reload fields
6122 ctx->ctx_reload_pmcs[0] = 0UL;
6123 ctx->ctx_reload_pmds[0] = 0UL;
6125 SET_LAST_CPU(ctx, smp_processor_id());
6128 * dump activation value for this PMU
6132 * record current activation for this context
6134 SET_ACTIVATION(ctx);
6137 * establish new ownership.
6139 SET_PMU_OWNER(task, ctx);
6142 * restore the psr.up bit. measurement
6144 * no PMU interrupt can happen at this point
6145 * because we still have interrupts disabled.
6147 if (likely(psr_up)) pfm_set_psr_up();
6150 * allow concurrent access to context
6152 pfm_unprotect_ctx_ctxsw(ctx, flags);
6154 #else /* !CONFIG_SMP */
6156 * reload PMU state for UP kernels
6157 * in 2.5 we come here with interrupts disabled
6160 pfm_load_regs (struct task_struct *task)
6162 struct thread_struct *t;
6164 struct task_struct *owner;
6165 unsigned long pmd_mask, pmc_mask;
6167 int need_irq_resend;
6169 owner = GET_PMU_OWNER();
6170 ctx = PFM_GET_CTX(task);
6172 psr = pfm_get_psr();
6174 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6175 BUG_ON(psr & IA64_PSR_I);
6178 * we restore ALL the debug registers to avoid picking up
6181 * This must be done even when the task is still the owner
6182 * as the registers may have been modified via ptrace()
6183 * (not perfmon) by the previous task.
6185 if (ctx->ctx_fl_using_dbreg) {
6186 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6187 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6191 * retrieved saved psr.up
6193 psr_up = ctx->ctx_saved_psr_up;
6194 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6197 * short path, our state is still there, just
6198 * need to restore psr and we go
6200 * we do not touch either PMC nor PMD. the psr is not touched
6201 * by the overflow_handler. So we are safe w.r.t. to interrupt
6202 * concurrency even without interrupt masking.
6204 if (likely(owner == task)) {
6205 if (likely(psr_up)) pfm_set_psr_up();
6210 * someone else is still using the PMU, first push it out and
6211 * then we'll be able to install our stuff !
6213 * Upon return, there will be no owner for the current PMU
6215 if (owner) pfm_lazy_save_regs(owner);
6218 * To avoid leaking information to the user level when psr.sp=0,
6219 * we must reload ALL implemented pmds (even the ones we don't use).
6220 * In the kernel we only allow PFM_READ_PMDS on registers which
6221 * we initialized or requested (sampling) so there is no risk there.
6223 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6226 * ALL accessible PMCs are systematically reloaded, unused registers
6227 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6228 * up stale configuration.
6230 * PMC0 is never in the mask. It is always restored separately
6232 pmc_mask = ctx->ctx_all_pmcs[0];
6234 pfm_restore_pmds(t->pmds, pmd_mask);
6235 pfm_restore_pmcs(t->pmcs, pmc_mask);
6238 * check for pending overflow at the time the state
6241 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6243 * reload pmc0 with the overflow information
6244 * On McKinley PMU, this will trigger a PMU interrupt
6246 ia64_set_pmc(0, t->pmcs[0]);
6252 * will replay the PMU interrupt
6254 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6256 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6260 * establish new ownership.
6262 SET_PMU_OWNER(task, ctx);
6265 * restore the psr.up bit. measurement
6267 * no PMU interrupt can happen at this point
6268 * because we still have interrupts disabled.
6270 if (likely(psr_up)) pfm_set_psr_up();
6272 #endif /* CONFIG_SMP */
6275 * this function assumes monitoring is stopped
6278 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6281 unsigned long mask2, val, pmd_val, ovfl_val;
6282 int i, can_access_pmu = 0;
6286 * is the caller the task being monitored (or which initiated the
6287 * session for system wide measurements)
6289 is_self = ctx->ctx_task == task ? 1 : 0;
6292 if (task == current) {
6295 * in UP, the state can still be in the registers
6297 if (task == current || GET_PMU_OWNER() == task) {
6301 * Mark the PMU as not owned
6302 * This will cause the interrupt handler to do nothing in case an overflow
6303 * interrupt was in-flight
6304 * This also guarantees that pmc0 will contain the final state
6305 * It virtually gives us full control on overflow processing from that point
6308 SET_PMU_OWNER(NULL, NULL);
6311 * read current overflow status:
6313 * we are guaranteed to read the final stable state
6316 pmc0 = ia64_get_pmc(0); /* slow */
6319 * reset freeze bit, overflow status information destroyed
6323 pmc0 = task->thread.pmcs[0];
6325 * clear whatever overflow status bits there were
6327 task->thread.pmcs[0] = 0;
6329 ovfl_val = pmu_conf->ovfl_val;
6331 * we save all the used pmds
6332 * we take care of overflows for counting PMDs
6334 * XXX: sampling situation is not taken into account here
6336 mask2 = ctx->ctx_used_pmds[0];
6337 for (i = 0; mask2; i++, mask2>>=1) {
6339 /* skip non used pmds */
6340 if ((mask2 & 0x1) == 0) continue;
6343 * can access PMU always true in system wide mode
6345 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6347 if (PMD_IS_COUNTING(i)) {
6348 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6351 ctx->ctx_pmds[i].val,
6355 * we rebuild the full 64 bit value of the counter
6357 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6360 * now everything is in ctx_pmds[] and we need
6361 * to clear the saved context from save_regs() such that
6362 * pfm_read_pmds() gets the correct value
6367 * take care of overflow inline
6369 if (pmc0 & (1UL << i)) {
6370 val += 1 + ovfl_val;
6371 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6375 DPRINT(("[%d] is_self=%d ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, is_self, i, val, pmd_val));
6377 if (is_self) task->thread.pmds[i] = pmd_val;
6379 ctx->ctx_pmds[i].val = val;
6383 static struct irqaction perfmon_irqaction = {
6384 .handler = pfm_interrupt_handler,
6385 .flags = SA_INTERRUPT,
6390 * perfmon initialization routine, called from the initcall() table
6392 static int init_pfm_fs(void);
6400 family = local_cpu_data->family;
6405 if ((*p)->probe() == 0) goto found;
6406 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6417 static struct file_operations pfm_proc_fops = {
6418 .open = pfm_proc_open,
6420 .llseek = seq_lseek,
6421 .release = seq_release,
6427 unsigned int n, n_counters, i;
6429 printk("perfmon: version %u.%u IRQ %u\n",
6432 IA64_PERFMON_VECTOR);
6434 if (pfm_probe_pmu()) {
6435 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6436 local_cpu_data->family);
6441 * compute the number of implemented PMD/PMC from the
6442 * description tables
6445 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6446 if (PMC_IS_IMPL(i) == 0) continue;
6447 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6450 pmu_conf->num_pmcs = n;
6452 n = 0; n_counters = 0;
6453 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6454 if (PMD_IS_IMPL(i) == 0) continue;
6455 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6457 if (PMD_IS_COUNTING(i)) n_counters++;
6459 pmu_conf->num_pmds = n;
6460 pmu_conf->num_counters = n_counters;
6463 * sanity checks on the number of debug registers
6465 if (pmu_conf->use_rr_dbregs) {
6466 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6467 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6471 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6472 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6478 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6482 pmu_conf->num_counters,
6483 ffz(pmu_conf->ovfl_val));
6486 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6487 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6493 * create /proc/perfmon (mostly for debugging purposes)
6495 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6496 if (perfmon_dir == NULL) {
6497 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6502 * install customized file operations for /proc/perfmon entry
6504 perfmon_dir->proc_fops = &pfm_proc_fops;
6507 * create /proc/sys/kernel/perfmon (for debugging purposes)
6509 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6512 * initialize all our spinlocks
6514 spin_lock_init(&pfm_sessions.pfs_lock);
6515 spin_lock_init(&pfm_buffer_fmt_lock);
6519 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6524 __initcall(pfm_init);
6527 * this function is called before pfm_init()
6530 pfm_init_percpu (void)
6533 * make sure no measurement is active
6534 * (may inherit programmed PMCs from EFI).
6540 * we run with the PMU not frozen at all times
6544 if (smp_processor_id() == 0)
6545 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6547 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6552 * used for debug purposes only
6555 dump_pmu_state(const char *from)
6557 struct task_struct *task;
6558 struct thread_struct *t;
6559 struct pt_regs *regs;
6561 unsigned long psr, dcr, info, flags;
6564 local_irq_save(flags);
6566 this_cpu = smp_processor_id();
6567 regs = ia64_task_regs(current);
6568 info = PFM_CPUINFO_GET();
6569 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6571 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6572 local_irq_restore(flags);
6576 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6583 task = GET_PMU_OWNER();
6584 ctx = GET_PMU_CTX();
6586 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6588 psr = pfm_get_psr();
6590 printk("->CPU%d pmc0=0x%lx psr.pp=%d psr.up=%d dcr.pp=%d syst_info=0x%lx user_psr.up=%d user_psr.pp=%d\n",
6593 psr & IA64_PSR_PP ? 1 : 0,
6594 psr & IA64_PSR_UP ? 1 : 0,
6595 dcr & IA64_DCR_PP ? 1 : 0,
6598 ia64_psr(regs)->pp);
6600 ia64_psr(regs)->up = 0;
6601 ia64_psr(regs)->pp = 0;
6603 t = ¤t->thread;
6605 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6606 if (PMC_IS_IMPL(i) == 0) continue;
6607 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6610 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6611 if (PMD_IS_IMPL(i) == 0) continue;
6612 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6616 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6619 ctx->ctx_smpl_vaddr,
6623 ctx->ctx_saved_psr_up);
6625 local_irq_restore(flags);
6629 * called from process.c:copy_thread(). task is new child.
6632 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6634 struct thread_struct *thread;
6636 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6638 thread = &task->thread;
6641 * cut links inherited from parent (current)
6643 thread->pfm_context = NULL;
6645 PFM_SET_WORK_PENDING(task, 0);
6648 * the psr bits are already set properly in copy_threads()
6651 #else /* !CONFIG_PERFMON */
6653 sys_perfmonctl (int fd, int cmd, void *arg, int count, long arg5, long arg6, long arg7,
6654 long arg8, long stack)
6658 #endif /* CONFIG_PERFMON */