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) 1999-2005 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/module.h>
23 #include <linux/kernel.h>
24 #include <linux/sched.h>
25 #include <linux/interrupt.h>
26 #include <linux/proc_fs.h>
27 #include <linux/seq_file.h>
28 #include <linux/init.h>
29 #include <linux/vmalloc.h>
31 #include <linux/sysctl.h>
32 #include <linux/list.h>
33 #include <linux/file.h>
34 #include <linux/poll.h>
35 #include <linux/vfs.h>
36 #include <linux/smp.h>
37 #include <linux/pagemap.h>
38 #include <linux/mount.h>
39 #include <linux/bitops.h>
40 #include <linux/capability.h>
41 #include <linux/rcupdate.h>
42 #include <linux/completion.h>
43 #include <linux/tracehook.h>
45 #include <asm/errno.h>
46 #include <asm/intrinsics.h>
48 #include <asm/perfmon.h>
49 #include <asm/processor.h>
50 #include <asm/signal.h>
51 #include <asm/system.h>
52 #include <asm/uaccess.h>
53 #include <asm/delay.h>
57 * perfmon context state
59 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
60 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
61 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
62 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
64 #define PFM_INVALID_ACTIVATION (~0UL)
66 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
67 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
70 * depth of message queue
72 #define PFM_MAX_MSGS 32
73 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
76 * type of a PMU register (bitmask).
78 * bit0 : register implemented
81 * bit4 : pmc has pmc.pm
82 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
83 * bit6-7 : register type
86 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
87 #define PFM_REG_IMPL 0x1 /* register implemented */
88 #define PFM_REG_END 0x2 /* end marker */
89 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
90 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
91 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
92 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
93 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
95 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
96 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
98 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
100 /* i assumed unsigned */
101 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
102 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
104 /* XXX: these assume that register i is implemented */
105 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
106 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
107 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
108 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
110 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
111 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
112 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
113 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
115 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
116 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
118 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
119 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
120 #define PFM_CTX_TASK(h) (h)->ctx_task
122 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
124 /* XXX: does not support more than 64 PMDs */
125 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
126 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
128 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
130 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
131 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
132 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
133 #define PFM_CODE_RR 0 /* requesting code range restriction */
134 #define PFM_DATA_RR 1 /* requestion data range restriction */
136 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
137 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
138 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
140 #define RDEP(x) (1UL<<(x))
143 * context protection macros
145 * - we need to protect against CPU concurrency (spin_lock)
146 * - we need to protect against PMU overflow interrupts (local_irq_disable)
148 * - we need to protect against PMU overflow interrupts (local_irq_disable)
150 * spin_lock_irqsave()/spin_unlock_irqrestore():
151 * in SMP: local_irq_disable + spin_lock
152 * in UP : local_irq_disable
154 * spin_lock()/spin_lock():
155 * in UP : removed automatically
156 * in SMP: protect against context accesses from other CPU. interrupts
157 * are not masked. This is useful for the PMU interrupt handler
158 * because we know we will not get PMU concurrency in that code.
160 #define PROTECT_CTX(c, f) \
162 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
163 spin_lock_irqsave(&(c)->ctx_lock, f); \
164 DPRINT(("spinlocked ctx %p by [%d]\n", c, task_pid_nr(current))); \
167 #define UNPROTECT_CTX(c, f) \
169 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
170 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
173 #define PROTECT_CTX_NOPRINT(c, f) \
175 spin_lock_irqsave(&(c)->ctx_lock, f); \
179 #define UNPROTECT_CTX_NOPRINT(c, f) \
181 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
185 #define PROTECT_CTX_NOIRQ(c) \
187 spin_lock(&(c)->ctx_lock); \
190 #define UNPROTECT_CTX_NOIRQ(c) \
192 spin_unlock(&(c)->ctx_lock); \
198 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
199 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
200 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
202 #else /* !CONFIG_SMP */
203 #define SET_ACTIVATION(t) do {} while(0)
204 #define GET_ACTIVATION(t) do {} while(0)
205 #define INC_ACTIVATION(t) do {} while(0)
206 #endif /* CONFIG_SMP */
208 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
209 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
210 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
212 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
213 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
215 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
218 * cmp0 must be the value of pmc0
220 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
222 #define PFMFS_MAGIC 0xa0b4d889
227 #define PFM_DEBUGGING 1
231 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
234 #define DPRINT_ovfl(a) \
236 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
241 * 64-bit software counter structure
243 * the next_reset_type is applied to the next call to pfm_reset_regs()
246 unsigned long val; /* virtual 64bit counter value */
247 unsigned long lval; /* last reset value */
248 unsigned long long_reset; /* reset value on sampling overflow */
249 unsigned long short_reset; /* reset value on overflow */
250 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
251 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
252 unsigned long seed; /* seed for random-number generator */
253 unsigned long mask; /* mask for random-number generator */
254 unsigned int flags; /* notify/do not notify */
255 unsigned long eventid; /* overflow event identifier */
262 unsigned int block:1; /* when 1, task will blocked on user notifications */
263 unsigned int system:1; /* do system wide monitoring */
264 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
265 unsigned int is_sampling:1; /* true if using a custom format */
266 unsigned int excl_idle:1; /* exclude idle task in system wide session */
267 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
268 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
269 unsigned int no_msg:1; /* no message sent on overflow */
270 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
271 unsigned int reserved:22;
272 } pfm_context_flags_t;
274 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
275 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
276 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
280 * perfmon context: encapsulates all the state of a monitoring session
283 typedef struct pfm_context {
284 spinlock_t ctx_lock; /* context protection */
286 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
287 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
289 struct task_struct *ctx_task; /* task to which context is attached */
291 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
293 struct completion ctx_restart_done; /* use for blocking notification mode */
295 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
296 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
297 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
299 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
300 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
301 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
303 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
305 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
306 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
307 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
308 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
310 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
312 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
313 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
315 u64 ctx_saved_psr_up; /* only contains psr.up value */
317 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
318 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
319 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
321 int ctx_fd; /* file descriptor used my this context */
322 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
324 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
325 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
326 unsigned long ctx_smpl_size; /* size of sampling buffer */
327 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
329 wait_queue_head_t ctx_msgq_wait;
330 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
333 struct fasync_struct *ctx_async_queue;
335 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
339 * magic number used to verify that structure is really
342 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
344 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
347 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
348 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
350 #define SET_LAST_CPU(ctx, v) do {} while(0)
351 #define GET_LAST_CPU(ctx) do {} while(0)
355 #define ctx_fl_block ctx_flags.block
356 #define ctx_fl_system ctx_flags.system
357 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
358 #define ctx_fl_is_sampling ctx_flags.is_sampling
359 #define ctx_fl_excl_idle ctx_flags.excl_idle
360 #define ctx_fl_going_zombie ctx_flags.going_zombie
361 #define ctx_fl_trap_reason ctx_flags.trap_reason
362 #define ctx_fl_no_msg ctx_flags.no_msg
363 #define ctx_fl_can_restart ctx_flags.can_restart
365 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
366 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
369 * global information about all sessions
370 * mostly used to synchronize between system wide and per-process
373 spinlock_t pfs_lock; /* lock the structure */
375 unsigned int pfs_task_sessions; /* number of per task sessions */
376 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
377 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
378 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
379 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
383 * information about a PMC or PMD.
384 * dep_pmd[]: a bitmask of dependent PMD registers
385 * dep_pmc[]: a bitmask of dependent PMC registers
387 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
391 unsigned long default_value; /* power-on default value */
392 unsigned long reserved_mask; /* bitmask of reserved bits */
393 pfm_reg_check_t read_check;
394 pfm_reg_check_t write_check;
395 unsigned long dep_pmd[4];
396 unsigned long dep_pmc[4];
399 /* assume cnum is a valid monitor */
400 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
403 * This structure is initialized at boot time and contains
404 * a description of the PMU main characteristics.
406 * If the probe function is defined, detection is based
407 * on its return value:
408 * - 0 means recognized PMU
409 * - anything else means not supported
410 * When the probe function is not defined, then the pmu_family field
411 * is used and it must match the host CPU family such that:
412 * - cpu->family & config->pmu_family != 0
415 unsigned long ovfl_val; /* overflow value for counters */
417 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
418 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
420 unsigned int num_pmcs; /* number of PMCS: computed at init time */
421 unsigned int num_pmds; /* number of PMDS: computed at init time */
422 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
423 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
425 char *pmu_name; /* PMU family name */
426 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
427 unsigned int flags; /* pmu specific flags */
428 unsigned int num_ibrs; /* number of IBRS: computed at init time */
429 unsigned int num_dbrs; /* number of DBRS: computed at init time */
430 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
431 int (*probe)(void); /* customized probe routine */
432 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
437 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
440 * debug register related type definitions
443 unsigned long ibr_mask:56;
444 unsigned long ibr_plm:4;
445 unsigned long ibr_ig:3;
446 unsigned long ibr_x:1;
450 unsigned long dbr_mask:56;
451 unsigned long dbr_plm:4;
452 unsigned long dbr_ig:2;
453 unsigned long dbr_w:1;
454 unsigned long dbr_r:1;
465 * perfmon command descriptions
468 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
471 unsigned int cmd_narg;
473 int (*cmd_getsize)(void *arg, size_t *sz);
476 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
477 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
478 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
479 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
482 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
483 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
484 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
485 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
486 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
488 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
491 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
492 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
493 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
496 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
497 unsigned long pfm_smpl_handler_calls;
498 unsigned long pfm_smpl_handler_cycles;
499 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
503 * perfmon internal variables
505 static pfm_stats_t pfm_stats[NR_CPUS];
506 static pfm_session_t pfm_sessions; /* global sessions information */
508 static DEFINE_SPINLOCK(pfm_alt_install_check);
509 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
511 static struct proc_dir_entry *perfmon_dir;
512 static pfm_uuid_t pfm_null_uuid = {0,};
514 static spinlock_t pfm_buffer_fmt_lock;
515 static LIST_HEAD(pfm_buffer_fmt_list);
517 static pmu_config_t *pmu_conf;
519 /* sysctl() controls */
520 pfm_sysctl_t pfm_sysctl;
521 EXPORT_SYMBOL(pfm_sysctl);
523 static ctl_table pfm_ctl_table[]={
525 .ctl_name = CTL_UNNUMBERED,
527 .data = &pfm_sysctl.debug,
528 .maxlen = sizeof(int),
530 .proc_handler = &proc_dointvec,
533 .ctl_name = CTL_UNNUMBERED,
534 .procname = "debug_ovfl",
535 .data = &pfm_sysctl.debug_ovfl,
536 .maxlen = sizeof(int),
538 .proc_handler = &proc_dointvec,
541 .ctl_name = CTL_UNNUMBERED,
542 .procname = "fastctxsw",
543 .data = &pfm_sysctl.fastctxsw,
544 .maxlen = sizeof(int),
546 .proc_handler = &proc_dointvec,
549 .ctl_name = CTL_UNNUMBERED,
550 .procname = "expert_mode",
551 .data = &pfm_sysctl.expert_mode,
552 .maxlen = sizeof(int),
554 .proc_handler = &proc_dointvec,
558 static ctl_table pfm_sysctl_dir[] = {
560 .ctl_name = CTL_UNNUMBERED,
561 .procname = "perfmon",
563 .child = pfm_ctl_table,
567 static ctl_table pfm_sysctl_root[] = {
569 .ctl_name = CTL_KERN,
570 .procname = "kernel",
572 .child = pfm_sysctl_dir,
576 static struct ctl_table_header *pfm_sysctl_header;
578 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
580 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
581 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
584 pfm_put_task(struct task_struct *task)
586 if (task != current) put_task_struct(task);
590 pfm_reserve_page(unsigned long a)
592 SetPageReserved(vmalloc_to_page((void *)a));
595 pfm_unreserve_page(unsigned long a)
597 ClearPageReserved(vmalloc_to_page((void*)a));
600 static inline unsigned long
601 pfm_protect_ctx_ctxsw(pfm_context_t *x)
603 spin_lock(&(x)->ctx_lock);
608 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
610 spin_unlock(&(x)->ctx_lock);
613 static inline unsigned int
614 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
616 return do_munmap(mm, addr, len);
619 static inline unsigned long
620 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
622 return get_unmapped_area(file, addr, len, pgoff, flags);
627 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data,
628 struct vfsmount *mnt)
630 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
633 static struct file_system_type pfm_fs_type = {
635 .get_sb = pfmfs_get_sb,
636 .kill_sb = kill_anon_super,
639 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
640 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
641 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
642 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
643 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
646 /* forward declaration */
647 static const struct file_operations pfm_file_ops;
650 * forward declarations
653 static void pfm_lazy_save_regs (struct task_struct *ta);
656 void dump_pmu_state(const char *);
657 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
659 #include "perfmon_itanium.h"
660 #include "perfmon_mckinley.h"
661 #include "perfmon_montecito.h"
662 #include "perfmon_generic.h"
664 static pmu_config_t *pmu_confs[]={
668 &pmu_conf_gen, /* must be last */
673 static int pfm_end_notify_user(pfm_context_t *ctx);
676 pfm_clear_psr_pp(void)
678 ia64_rsm(IA64_PSR_PP);
685 ia64_ssm(IA64_PSR_PP);
690 pfm_clear_psr_up(void)
692 ia64_rsm(IA64_PSR_UP);
699 ia64_ssm(IA64_PSR_UP);
703 static inline unsigned long
707 tmp = ia64_getreg(_IA64_REG_PSR);
713 pfm_set_psr_l(unsigned long val)
715 ia64_setreg(_IA64_REG_PSR_L, val);
727 pfm_unfreeze_pmu(void)
734 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
738 for (i=0; i < nibrs; i++) {
739 ia64_set_ibr(i, ibrs[i]);
740 ia64_dv_serialize_instruction();
746 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
750 for (i=0; i < ndbrs; i++) {
751 ia64_set_dbr(i, dbrs[i]);
752 ia64_dv_serialize_data();
758 * PMD[i] must be a counter. no check is made
760 static inline unsigned long
761 pfm_read_soft_counter(pfm_context_t *ctx, int i)
763 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
767 * PMD[i] must be a counter. no check is made
770 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
772 unsigned long ovfl_val = pmu_conf->ovfl_val;
774 ctx->ctx_pmds[i].val = val & ~ovfl_val;
776 * writing to unimplemented part is ignore, so we do not need to
779 ia64_set_pmd(i, val & ovfl_val);
783 pfm_get_new_msg(pfm_context_t *ctx)
787 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
789 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
790 if (next == ctx->ctx_msgq_head) return NULL;
792 idx = ctx->ctx_msgq_tail;
793 ctx->ctx_msgq_tail = next;
795 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
797 return ctx->ctx_msgq+idx;
801 pfm_get_next_msg(pfm_context_t *ctx)
805 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
807 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
812 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
817 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
819 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));
825 pfm_reset_msgq(pfm_context_t *ctx)
827 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
828 DPRINT(("ctx=%p msgq reset\n", ctx));
832 pfm_rvmalloc(unsigned long size)
837 size = PAGE_ALIGN(size);
840 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
841 memset(mem, 0, size);
842 addr = (unsigned long)mem;
844 pfm_reserve_page(addr);
853 pfm_rvfree(void *mem, unsigned long size)
858 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
859 addr = (unsigned long) mem;
860 while ((long) size > 0) {
861 pfm_unreserve_page(addr);
870 static pfm_context_t *
871 pfm_context_alloc(int ctx_flags)
876 * allocate context descriptor
877 * must be able to free with interrupts disabled
879 ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
881 DPRINT(("alloc ctx @%p\n", ctx));
884 * init context protection lock
886 spin_lock_init(&ctx->ctx_lock);
889 * context is unloaded
891 ctx->ctx_state = PFM_CTX_UNLOADED;
894 * initialization of context's flags
896 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
897 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
898 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
900 * will move to set properties
901 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
905 * init restart semaphore to locked
907 init_completion(&ctx->ctx_restart_done);
910 * activation is used in SMP only
912 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
913 SET_LAST_CPU(ctx, -1);
916 * initialize notification message queue
918 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
919 init_waitqueue_head(&ctx->ctx_msgq_wait);
920 init_waitqueue_head(&ctx->ctx_zombieq);
927 pfm_context_free(pfm_context_t *ctx)
930 DPRINT(("free ctx @%p\n", ctx));
936 pfm_mask_monitoring(struct task_struct *task)
938 pfm_context_t *ctx = PFM_GET_CTX(task);
939 unsigned long mask, val, ovfl_mask;
942 DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
944 ovfl_mask = pmu_conf->ovfl_val;
946 * monitoring can only be masked as a result of a valid
947 * counter overflow. In UP, it means that the PMU still
948 * has an owner. Note that the owner can be different
949 * from the current task. However the PMU state belongs
951 * In SMP, a valid overflow only happens when task is
952 * current. Therefore if we come here, we know that
953 * the PMU state belongs to the current task, therefore
954 * we can access the live registers.
956 * So in both cases, the live register contains the owner's
957 * state. We can ONLY touch the PMU registers and NOT the PSR.
959 * As a consequence to this call, the ctx->th_pmds[] array
960 * contains stale information which must be ignored
961 * when context is reloaded AND monitoring is active (see
964 mask = ctx->ctx_used_pmds[0];
965 for (i = 0; mask; i++, mask>>=1) {
966 /* skip non used pmds */
967 if ((mask & 0x1) == 0) continue;
968 val = ia64_get_pmd(i);
970 if (PMD_IS_COUNTING(i)) {
972 * we rebuild the full 64 bit value of the counter
974 ctx->ctx_pmds[i].val += (val & ovfl_mask);
976 ctx->ctx_pmds[i].val = val;
978 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
980 ctx->ctx_pmds[i].val,
984 * mask monitoring by setting the privilege level to 0
985 * we cannot use psr.pp/psr.up for this, it is controlled by
988 * if task is current, modify actual registers, otherwise modify
989 * thread save state, i.e., what will be restored in pfm_load_regs()
991 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
992 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
993 if ((mask & 0x1) == 0UL) continue;
994 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
995 ctx->th_pmcs[i] &= ~0xfUL;
996 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
999 * make all of this visible
1005 * must always be done with task == current
1007 * context must be in MASKED state when calling
1010 pfm_restore_monitoring(struct task_struct *task)
1012 pfm_context_t *ctx = PFM_GET_CTX(task);
1013 unsigned long mask, ovfl_mask;
1014 unsigned long psr, val;
1017 is_system = ctx->ctx_fl_system;
1018 ovfl_mask = pmu_conf->ovfl_val;
1020 if (task != current) {
1021 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1024 if (ctx->ctx_state != PFM_CTX_MASKED) {
1025 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1026 task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1029 psr = pfm_get_psr();
1031 * monitoring is masked via the PMC.
1032 * As we restore their value, we do not want each counter to
1033 * restart right away. We stop monitoring using the PSR,
1034 * restore the PMC (and PMD) and then re-establish the psr
1035 * as it was. Note that there can be no pending overflow at
1036 * this point, because monitoring was MASKED.
1038 * system-wide session are pinned and self-monitoring
1040 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1041 /* disable dcr pp */
1042 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1048 * first, we restore the PMD
1050 mask = ctx->ctx_used_pmds[0];
1051 for (i = 0; mask; i++, mask>>=1) {
1052 /* skip non used pmds */
1053 if ((mask & 0x1) == 0) continue;
1055 if (PMD_IS_COUNTING(i)) {
1057 * we split the 64bit value according to
1060 val = ctx->ctx_pmds[i].val & ovfl_mask;
1061 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1063 val = ctx->ctx_pmds[i].val;
1065 ia64_set_pmd(i, val);
1067 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1069 ctx->ctx_pmds[i].val,
1075 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1076 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1077 if ((mask & 0x1) == 0UL) continue;
1078 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1079 ia64_set_pmc(i, ctx->th_pmcs[i]);
1080 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1081 task_pid_nr(task), i, ctx->th_pmcs[i]));
1086 * must restore DBR/IBR because could be modified while masked
1087 * XXX: need to optimize
1089 if (ctx->ctx_fl_using_dbreg) {
1090 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1091 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1097 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1099 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1106 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1112 for (i=0; mask; i++, mask>>=1) {
1113 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1118 * reload from thread state (used for ctxw only)
1121 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1124 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1126 for (i=0; mask; i++, mask>>=1) {
1127 if ((mask & 0x1) == 0) continue;
1128 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1129 ia64_set_pmd(i, val);
1135 * propagate PMD from context to thread-state
1138 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1140 unsigned long ovfl_val = pmu_conf->ovfl_val;
1141 unsigned long mask = ctx->ctx_all_pmds[0];
1145 DPRINT(("mask=0x%lx\n", mask));
1147 for (i=0; mask; i++, mask>>=1) {
1149 val = ctx->ctx_pmds[i].val;
1152 * We break up the 64 bit value into 2 pieces
1153 * the lower bits go to the machine state in the
1154 * thread (will be reloaded on ctxsw in).
1155 * The upper part stays in the soft-counter.
1157 if (PMD_IS_COUNTING(i)) {
1158 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1161 ctx->th_pmds[i] = val;
1163 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1166 ctx->ctx_pmds[i].val));
1171 * propagate PMC from context to thread-state
1174 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1176 unsigned long mask = ctx->ctx_all_pmcs[0];
1179 DPRINT(("mask=0x%lx\n", mask));
1181 for (i=0; mask; i++, mask>>=1) {
1182 /* masking 0 with ovfl_val yields 0 */
1183 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1184 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1191 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1195 for (i=0; mask; i++, mask>>=1) {
1196 if ((mask & 0x1) == 0) continue;
1197 ia64_set_pmc(i, pmcs[i]);
1203 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1205 return memcmp(a, b, sizeof(pfm_uuid_t));
1209 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1212 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1217 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1220 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1226 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1230 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1235 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1239 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1244 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1247 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1252 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)
1255 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1259 static pfm_buffer_fmt_t *
1260 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1262 struct list_head * pos;
1263 pfm_buffer_fmt_t * entry;
1265 list_for_each(pos, &pfm_buffer_fmt_list) {
1266 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1267 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1274 * find a buffer format based on its uuid
1276 static pfm_buffer_fmt_t *
1277 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1279 pfm_buffer_fmt_t * fmt;
1280 spin_lock(&pfm_buffer_fmt_lock);
1281 fmt = __pfm_find_buffer_fmt(uuid);
1282 spin_unlock(&pfm_buffer_fmt_lock);
1287 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1291 /* some sanity checks */
1292 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1294 /* we need at least a handler */
1295 if (fmt->fmt_handler == NULL) return -EINVAL;
1298 * XXX: need check validity of fmt_arg_size
1301 spin_lock(&pfm_buffer_fmt_lock);
1303 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1304 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1308 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1309 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1312 spin_unlock(&pfm_buffer_fmt_lock);
1315 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1318 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1320 pfm_buffer_fmt_t *fmt;
1323 spin_lock(&pfm_buffer_fmt_lock);
1325 fmt = __pfm_find_buffer_fmt(uuid);
1327 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1331 list_del_init(&fmt->fmt_list);
1332 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1335 spin_unlock(&pfm_buffer_fmt_lock);
1339 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1341 extern void update_pal_halt_status(int);
1344 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1346 unsigned long flags;
1348 * validity checks on cpu_mask have been done upstream
1352 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1353 pfm_sessions.pfs_sys_sessions,
1354 pfm_sessions.pfs_task_sessions,
1355 pfm_sessions.pfs_sys_use_dbregs,
1361 * cannot mix system wide and per-task sessions
1363 if (pfm_sessions.pfs_task_sessions > 0UL) {
1364 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1365 pfm_sessions.pfs_task_sessions));
1369 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1371 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1373 pfm_sessions.pfs_sys_session[cpu] = task;
1375 pfm_sessions.pfs_sys_sessions++ ;
1378 if (pfm_sessions.pfs_sys_sessions) goto abort;
1379 pfm_sessions.pfs_task_sessions++;
1382 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1383 pfm_sessions.pfs_sys_sessions,
1384 pfm_sessions.pfs_task_sessions,
1385 pfm_sessions.pfs_sys_use_dbregs,
1390 * disable default_idle() to go to PAL_HALT
1392 update_pal_halt_status(0);
1399 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1400 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1410 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1412 unsigned long flags;
1414 * validity checks on cpu_mask have been done upstream
1418 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1419 pfm_sessions.pfs_sys_sessions,
1420 pfm_sessions.pfs_task_sessions,
1421 pfm_sessions.pfs_sys_use_dbregs,
1427 pfm_sessions.pfs_sys_session[cpu] = NULL;
1429 * would not work with perfmon+more than one bit in cpu_mask
1431 if (ctx && ctx->ctx_fl_using_dbreg) {
1432 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1433 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1435 pfm_sessions.pfs_sys_use_dbregs--;
1438 pfm_sessions.pfs_sys_sessions--;
1440 pfm_sessions.pfs_task_sessions--;
1442 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1443 pfm_sessions.pfs_sys_sessions,
1444 pfm_sessions.pfs_task_sessions,
1445 pfm_sessions.pfs_sys_use_dbregs,
1450 * if possible, enable default_idle() to go into PAL_HALT
1452 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1453 update_pal_halt_status(1);
1461 * removes virtual mapping of the sampling buffer.
1462 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1463 * a PROTECT_CTX() section.
1466 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1471 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1472 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1476 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1479 * does the actual unmapping
1481 down_write(&task->mm->mmap_sem);
1483 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1485 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1487 up_write(&task->mm->mmap_sem);
1489 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1492 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1498 * free actual physical storage used by sampling buffer
1502 pfm_free_smpl_buffer(pfm_context_t *ctx)
1504 pfm_buffer_fmt_t *fmt;
1506 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1509 * we won't use the buffer format anymore
1511 fmt = ctx->ctx_buf_fmt;
1513 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1516 ctx->ctx_smpl_vaddr));
1518 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1523 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1525 ctx->ctx_smpl_hdr = NULL;
1526 ctx->ctx_smpl_size = 0UL;
1531 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1537 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1539 if (fmt == NULL) return;
1541 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1546 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1547 * no real gain from having the whole whorehouse mounted. So we don't need
1548 * any operations on the root directory. However, we need a non-trivial
1549 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1551 static struct vfsmount *pfmfs_mnt;
1556 int err = register_filesystem(&pfm_fs_type);
1558 pfmfs_mnt = kern_mount(&pfm_fs_type);
1559 err = PTR_ERR(pfmfs_mnt);
1560 if (IS_ERR(pfmfs_mnt))
1561 unregister_filesystem(&pfm_fs_type);
1569 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1574 unsigned long flags;
1575 DECLARE_WAITQUEUE(wait, current);
1576 if (PFM_IS_FILE(filp) == 0) {
1577 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1581 ctx = (pfm_context_t *)filp->private_data;
1583 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1588 * check even when there is no message
1590 if (size < sizeof(pfm_msg_t)) {
1591 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1595 PROTECT_CTX(ctx, flags);
1598 * put ourselves on the wait queue
1600 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1608 set_current_state(TASK_INTERRUPTIBLE);
1610 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1613 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1615 UNPROTECT_CTX(ctx, flags);
1618 * check non-blocking read
1621 if(filp->f_flags & O_NONBLOCK) break;
1624 * check pending signals
1626 if(signal_pending(current)) {
1631 * no message, so wait
1635 PROTECT_CTX(ctx, flags);
1637 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1638 set_current_state(TASK_RUNNING);
1639 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1641 if (ret < 0) goto abort;
1644 msg = pfm_get_next_msg(ctx);
1646 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1650 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1653 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1656 UNPROTECT_CTX(ctx, flags);
1662 pfm_write(struct file *file, const char __user *ubuf,
1663 size_t size, loff_t *ppos)
1665 DPRINT(("pfm_write called\n"));
1670 pfm_poll(struct file *filp, poll_table * wait)
1673 unsigned long flags;
1674 unsigned int mask = 0;
1676 if (PFM_IS_FILE(filp) == 0) {
1677 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1681 ctx = (pfm_context_t *)filp->private_data;
1683 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1688 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1690 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1692 PROTECT_CTX(ctx, flags);
1694 if (PFM_CTXQ_EMPTY(ctx) == 0)
1695 mask = POLLIN | POLLRDNORM;
1697 UNPROTECT_CTX(ctx, flags);
1699 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1705 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1707 DPRINT(("pfm_ioctl called\n"));
1712 * interrupt cannot be masked when coming here
1715 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1719 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1721 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1722 task_pid_nr(current),
1725 ctx->ctx_async_queue, ret));
1731 pfm_fasync(int fd, struct file *filp, int on)
1736 if (PFM_IS_FILE(filp) == 0) {
1737 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1741 ctx = (pfm_context_t *)filp->private_data;
1743 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1747 * we cannot mask interrupts during this call because this may
1748 * may go to sleep if memory is not readily avalaible.
1750 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1751 * done in caller. Serialization of this function is ensured by caller.
1753 ret = pfm_do_fasync(fd, filp, ctx, on);
1756 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1759 ctx->ctx_async_queue, ret));
1766 * this function is exclusively called from pfm_close().
1767 * The context is not protected at that time, nor are interrupts
1768 * on the remote CPU. That's necessary to avoid deadlocks.
1771 pfm_syswide_force_stop(void *info)
1773 pfm_context_t *ctx = (pfm_context_t *)info;
1774 struct pt_regs *regs = task_pt_regs(current);
1775 struct task_struct *owner;
1776 unsigned long flags;
1779 if (ctx->ctx_cpu != smp_processor_id()) {
1780 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1782 smp_processor_id());
1785 owner = GET_PMU_OWNER();
1786 if (owner != ctx->ctx_task) {
1787 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1789 task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1792 if (GET_PMU_CTX() != ctx) {
1793 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1795 GET_PMU_CTX(), ctx);
1799 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1801 * the context is already protected in pfm_close(), we simply
1802 * need to mask interrupts to avoid a PMU interrupt race on
1805 local_irq_save(flags);
1807 ret = pfm_context_unload(ctx, NULL, 0, regs);
1809 DPRINT(("context_unload returned %d\n", ret));
1813 * unmask interrupts, PMU interrupts are now spurious here
1815 local_irq_restore(flags);
1819 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1823 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1824 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1825 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1827 #endif /* CONFIG_SMP */
1830 * called for each close(). Partially free resources.
1831 * When caller is self-monitoring, the context is unloaded.
1834 pfm_flush(struct file *filp, fl_owner_t id)
1837 struct task_struct *task;
1838 struct pt_regs *regs;
1839 unsigned long flags;
1840 unsigned long smpl_buf_size = 0UL;
1841 void *smpl_buf_vaddr = NULL;
1842 int state, is_system;
1844 if (PFM_IS_FILE(filp) == 0) {
1845 DPRINT(("bad magic for\n"));
1849 ctx = (pfm_context_t *)filp->private_data;
1851 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1856 * remove our file from the async queue, if we use this mode.
1857 * This can be done without the context being protected. We come
1858 * here when the context has become unreachable by other tasks.
1860 * We may still have active monitoring at this point and we may
1861 * end up in pfm_overflow_handler(). However, fasync_helper()
1862 * operates with interrupts disabled and it cleans up the
1863 * queue. If the PMU handler is called prior to entering
1864 * fasync_helper() then it will send a signal. If it is
1865 * invoked after, it will find an empty queue and no
1866 * signal will be sent. In both case, we are safe
1868 PROTECT_CTX(ctx, flags);
1870 state = ctx->ctx_state;
1871 is_system = ctx->ctx_fl_system;
1873 task = PFM_CTX_TASK(ctx);
1874 regs = task_pt_regs(task);
1876 DPRINT(("ctx_state=%d is_current=%d\n",
1878 task == current ? 1 : 0));
1881 * if state == UNLOADED, then task is NULL
1885 * we must stop and unload because we are losing access to the context.
1887 if (task == current) {
1890 * the task IS the owner but it migrated to another CPU: that's bad
1891 * but we must handle this cleanly. Unfortunately, the kernel does
1892 * not provide a mechanism to block migration (while the context is loaded).
1894 * We need to release the resource on the ORIGINAL cpu.
1896 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1898 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1900 * keep context protected but unmask interrupt for IPI
1902 local_irq_restore(flags);
1904 pfm_syswide_cleanup_other_cpu(ctx);
1907 * restore interrupt masking
1909 local_irq_save(flags);
1912 * context is unloaded at this point
1915 #endif /* CONFIG_SMP */
1918 DPRINT(("forcing unload\n"));
1920 * stop and unload, returning with state UNLOADED
1921 * and session unreserved.
1923 pfm_context_unload(ctx, NULL, 0, regs);
1925 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1930 * remove virtual mapping, if any, for the calling task.
1931 * cannot reset ctx field until last user is calling close().
1933 * ctx_smpl_vaddr must never be cleared because it is needed
1934 * by every task with access to the context
1936 * When called from do_exit(), the mm context is gone already, therefore
1937 * mm is NULL, i.e., the VMA is already gone and we do not have to
1940 if (ctx->ctx_smpl_vaddr && current->mm) {
1941 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1942 smpl_buf_size = ctx->ctx_smpl_size;
1945 UNPROTECT_CTX(ctx, flags);
1948 * if there was a mapping, then we systematically remove it
1949 * at this point. Cannot be done inside critical section
1950 * because some VM function reenables interrupts.
1953 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1958 * called either on explicit close() or from exit_files().
1959 * Only the LAST user of the file gets to this point, i.e., it is
1962 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1963 * (fput()),i.e, last task to access the file. Nobody else can access the
1964 * file at this point.
1966 * When called from exit_files(), the VMA has been freed because exit_mm()
1967 * is executed before exit_files().
1969 * When called from exit_files(), the current task is not yet ZOMBIE but we
1970 * flush the PMU state to the context.
1973 pfm_close(struct inode *inode, struct file *filp)
1976 struct task_struct *task;
1977 struct pt_regs *regs;
1978 DECLARE_WAITQUEUE(wait, current);
1979 unsigned long flags;
1980 unsigned long smpl_buf_size = 0UL;
1981 void *smpl_buf_addr = NULL;
1982 int free_possible = 1;
1983 int state, is_system;
1985 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1987 if (PFM_IS_FILE(filp) == 0) {
1988 DPRINT(("bad magic\n"));
1992 ctx = (pfm_context_t *)filp->private_data;
1994 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1998 PROTECT_CTX(ctx, flags);
2000 state = ctx->ctx_state;
2001 is_system = ctx->ctx_fl_system;
2003 task = PFM_CTX_TASK(ctx);
2004 regs = task_pt_regs(task);
2006 DPRINT(("ctx_state=%d is_current=%d\n",
2008 task == current ? 1 : 0));
2011 * if task == current, then pfm_flush() unloaded the context
2013 if (state == PFM_CTX_UNLOADED) goto doit;
2016 * context is loaded/masked and task != current, we need to
2017 * either force an unload or go zombie
2021 * The task is currently blocked or will block after an overflow.
2022 * we must force it to wakeup to get out of the
2023 * MASKED state and transition to the unloaded state by itself.
2025 * This situation is only possible for per-task mode
2027 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2030 * set a "partial" zombie state to be checked
2031 * upon return from down() in pfm_handle_work().
2033 * We cannot use the ZOMBIE state, because it is checked
2034 * by pfm_load_regs() which is called upon wakeup from down().
2035 * In such case, it would free the context and then we would
2036 * return to pfm_handle_work() which would access the
2037 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2038 * but visible to pfm_handle_work().
2040 * For some window of time, we have a zombie context with
2041 * ctx_state = MASKED and not ZOMBIE
2043 ctx->ctx_fl_going_zombie = 1;
2046 * force task to wake up from MASKED state
2048 complete(&ctx->ctx_restart_done);
2050 DPRINT(("waking up ctx_state=%d\n", state));
2053 * put ourself to sleep waiting for the other
2054 * task to report completion
2056 * the context is protected by mutex, therefore there
2057 * is no risk of being notified of completion before
2058 * begin actually on the waitq.
2060 set_current_state(TASK_INTERRUPTIBLE);
2061 add_wait_queue(&ctx->ctx_zombieq, &wait);
2063 UNPROTECT_CTX(ctx, flags);
2066 * XXX: check for signals :
2067 * - ok for explicit close
2068 * - not ok when coming from exit_files()
2073 PROTECT_CTX(ctx, flags);
2076 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2077 set_current_state(TASK_RUNNING);
2080 * context is unloaded at this point
2082 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2084 else if (task != current) {
2087 * switch context to zombie state
2089 ctx->ctx_state = PFM_CTX_ZOMBIE;
2091 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2093 * cannot free the context on the spot. deferred until
2094 * the task notices the ZOMBIE state
2098 pfm_context_unload(ctx, NULL, 0, regs);
2103 /* reload state, may have changed during opening of critical section */
2104 state = ctx->ctx_state;
2107 * the context is still attached to a task (possibly current)
2108 * we cannot destroy it right now
2112 * we must free the sampling buffer right here because
2113 * we cannot rely on it being cleaned up later by the
2114 * monitored task. It is not possible to free vmalloc'ed
2115 * memory in pfm_load_regs(). Instead, we remove the buffer
2116 * now. should there be subsequent PMU overflow originally
2117 * meant for sampling, the will be converted to spurious
2118 * and that's fine because the monitoring tools is gone anyway.
2120 if (ctx->ctx_smpl_hdr) {
2121 smpl_buf_addr = ctx->ctx_smpl_hdr;
2122 smpl_buf_size = ctx->ctx_smpl_size;
2123 /* no more sampling */
2124 ctx->ctx_smpl_hdr = NULL;
2125 ctx->ctx_fl_is_sampling = 0;
2128 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2134 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2137 * UNLOADED that the session has already been unreserved.
2139 if (state == PFM_CTX_ZOMBIE) {
2140 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2144 * disconnect file descriptor from context must be done
2147 filp->private_data = NULL;
2150 * if we free on the spot, the context is now completely unreachable
2151 * from the callers side. The monitored task side is also cut, so we
2154 * If we have a deferred free, only the caller side is disconnected.
2156 UNPROTECT_CTX(ctx, flags);
2159 * All memory free operations (especially for vmalloc'ed memory)
2160 * MUST be done with interrupts ENABLED.
2162 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2165 * return the memory used by the context
2167 if (free_possible) pfm_context_free(ctx);
2173 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2175 DPRINT(("pfm_no_open called\n"));
2181 static const struct file_operations pfm_file_ops = {
2182 .llseek = no_llseek,
2187 .open = pfm_no_open, /* special open code to disallow open via /proc */
2188 .fasync = pfm_fasync,
2189 .release = pfm_close,
2194 pfmfs_delete_dentry(struct dentry *dentry)
2199 static struct dentry_operations pfmfs_dentry_operations = {
2200 .d_delete = pfmfs_delete_dentry,
2204 static struct file *
2205 pfm_alloc_file(pfm_context_t *ctx)
2208 struct inode *inode;
2209 struct dentry *dentry;
2214 * allocate a new inode
2216 inode = new_inode(pfmfs_mnt->mnt_sb);
2218 return ERR_PTR(-ENOMEM);
2220 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2222 inode->i_mode = S_IFCHR|S_IRUGO;
2223 inode->i_uid = current->fsuid;
2224 inode->i_gid = current->fsgid;
2226 sprintf(name, "[%lu]", inode->i_ino);
2228 this.len = strlen(name);
2229 this.hash = inode->i_ino;
2232 * allocate a new dcache entry
2234 dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2237 return ERR_PTR(-ENOMEM);
2240 dentry->d_op = &pfmfs_dentry_operations;
2241 d_add(dentry, inode);
2243 file = alloc_file(pfmfs_mnt, dentry, FMODE_READ, &pfm_file_ops);
2246 return ERR_PTR(-ENFILE);
2249 file->f_flags = O_RDONLY;
2250 file->private_data = ctx;
2256 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2258 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2261 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2264 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2275 * allocate a sampling buffer and remaps it into the user address space of the task
2278 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2280 struct mm_struct *mm = task->mm;
2281 struct vm_area_struct *vma = NULL;
2287 * the fixed header + requested size and align to page boundary
2289 size = PAGE_ALIGN(rsize);
2291 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2294 * check requested size to avoid Denial-of-service attacks
2295 * XXX: may have to refine this test
2296 * Check against address space limit.
2298 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2301 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2305 * We do the easy to undo allocations first.
2307 * pfm_rvmalloc(), clears the buffer, so there is no leak
2309 smpl_buf = pfm_rvmalloc(size);
2310 if (smpl_buf == NULL) {
2311 DPRINT(("Can't allocate sampling buffer\n"));
2315 DPRINT(("smpl_buf @%p\n", smpl_buf));
2318 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2320 DPRINT(("Cannot allocate vma\n"));
2325 * partially initialize the vma for the sampling buffer
2328 vma->vm_file = filp;
2329 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2330 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2333 * Now we have everything we need and we can initialize
2334 * and connect all the data structures
2337 ctx->ctx_smpl_hdr = smpl_buf;
2338 ctx->ctx_smpl_size = size; /* aligned size */
2341 * Let's do the difficult operations next.
2343 * now we atomically find some area in the address space and
2344 * remap the buffer in it.
2346 down_write(&task->mm->mmap_sem);
2348 /* find some free area in address space, must have mmap sem held */
2349 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2350 if (vma->vm_start == 0UL) {
2351 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2352 up_write(&task->mm->mmap_sem);
2355 vma->vm_end = vma->vm_start + size;
2356 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2358 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2360 /* can only be applied to current task, need to have the mm semaphore held when called */
2361 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2362 DPRINT(("Can't remap buffer\n"));
2363 up_write(&task->mm->mmap_sem);
2370 * now insert the vma in the vm list for the process, must be
2371 * done with mmap lock held
2373 insert_vm_struct(mm, vma);
2375 mm->total_vm += size >> PAGE_SHIFT;
2376 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2378 up_write(&task->mm->mmap_sem);
2381 * keep track of user level virtual address
2383 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2384 *(unsigned long *)user_vaddr = vma->vm_start;
2389 kmem_cache_free(vm_area_cachep, vma);
2391 pfm_rvfree(smpl_buf, size);
2397 * XXX: do something better here
2400 pfm_bad_permissions(struct task_struct *task)
2402 /* inspired by ptrace_attach() */
2403 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2412 return ((current->uid != task->euid)
2413 || (current->uid != task->suid)
2414 || (current->uid != task->uid)
2415 || (current->gid != task->egid)
2416 || (current->gid != task->sgid)
2417 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2421 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2427 ctx_flags = pfx->ctx_flags;
2429 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2432 * cannot block in this mode
2434 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2435 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2440 /* probably more to add here */
2446 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2447 unsigned int cpu, pfarg_context_t *arg)
2449 pfm_buffer_fmt_t *fmt = NULL;
2450 unsigned long size = 0UL;
2452 void *fmt_arg = NULL;
2454 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2456 /* invoke and lock buffer format, if found */
2457 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2459 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2464 * buffer argument MUST be contiguous to pfarg_context_t
2466 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2468 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2470 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2472 if (ret) goto error;
2474 /* link buffer format and context */
2475 ctx->ctx_buf_fmt = fmt;
2476 ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2479 * check if buffer format wants to use perfmon buffer allocation/mapping service
2481 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2482 if (ret) goto error;
2486 * buffer is always remapped into the caller's address space
2488 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2489 if (ret) goto error;
2491 /* keep track of user address of buffer */
2492 arg->ctx_smpl_vaddr = uaddr;
2494 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2501 pfm_reset_pmu_state(pfm_context_t *ctx)
2506 * install reset values for PMC.
2508 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2509 if (PMC_IS_IMPL(i) == 0) continue;
2510 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2511 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2514 * PMD registers are set to 0UL when the context in memset()
2518 * On context switched restore, we must restore ALL pmc and ALL pmd even
2519 * when they are not actively used by the task. In UP, the incoming process
2520 * may otherwise pick up left over PMC, PMD state from the previous process.
2521 * As opposed to PMD, stale PMC can cause harm to the incoming
2522 * process because they may change what is being measured.
2523 * Therefore, we must systematically reinstall the entire
2524 * PMC state. In SMP, the same thing is possible on the
2525 * same CPU but also on between 2 CPUs.
2527 * The problem with PMD is information leaking especially
2528 * to user level when psr.sp=0
2530 * There is unfortunately no easy way to avoid this problem
2531 * on either UP or SMP. This definitively slows down the
2532 * pfm_load_regs() function.
2536 * bitmask of all PMCs accessible to this context
2538 * PMC0 is treated differently.
2540 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2543 * bitmask of all PMDs that are accessible to this context
2545 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2547 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2550 * useful in case of re-enable after disable
2552 ctx->ctx_used_ibrs[0] = 0UL;
2553 ctx->ctx_used_dbrs[0] = 0UL;
2557 pfm_ctx_getsize(void *arg, size_t *sz)
2559 pfarg_context_t *req = (pfarg_context_t *)arg;
2560 pfm_buffer_fmt_t *fmt;
2564 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2566 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2568 DPRINT(("cannot find buffer format\n"));
2571 /* get just enough to copy in user parameters */
2572 *sz = fmt->fmt_arg_size;
2573 DPRINT(("arg_size=%lu\n", *sz));
2581 * cannot attach if :
2583 * - task not owned by caller
2584 * - task incompatible with context mode
2587 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2590 * no kernel task or task not owner by caller
2592 if (task->mm == NULL) {
2593 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2596 if (pfm_bad_permissions(task)) {
2597 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task)));
2601 * cannot block in self-monitoring mode
2603 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2604 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2608 if (task->exit_state == EXIT_ZOMBIE) {
2609 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task)));
2614 * always ok for self
2616 if (task == current) return 0;
2618 if (!task_is_stopped_or_traced(task)) {
2619 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2623 * make sure the task is off any CPU
2625 wait_task_inactive(task, 0);
2627 /* more to come... */
2633 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2635 struct task_struct *p = current;
2638 /* XXX: need to add more checks here */
2639 if (pid < 2) return -EPERM;
2641 if (pid != task_pid_vnr(current)) {
2643 read_lock(&tasklist_lock);
2645 p = find_task_by_vpid(pid);
2647 /* make sure task cannot go away while we operate on it */
2648 if (p) get_task_struct(p);
2650 read_unlock(&tasklist_lock);
2652 if (p == NULL) return -ESRCH;
2655 ret = pfm_task_incompatible(ctx, p);
2658 } else if (p != current) {
2667 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2669 pfarg_context_t *req = (pfarg_context_t *)arg;
2676 /* let's check the arguments first */
2677 ret = pfarg_is_sane(current, req);
2681 ctx_flags = req->ctx_flags;
2685 fd = get_unused_fd();
2689 ctx = pfm_context_alloc(ctx_flags);
2693 filp = pfm_alloc_file(ctx);
2695 ret = PTR_ERR(filp);
2699 req->ctx_fd = ctx->ctx_fd = fd;
2702 * does the user want to sample?
2704 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2705 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2710 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2715 ctx->ctx_fl_excl_idle,
2720 * initialize soft PMU state
2722 pfm_reset_pmu_state(ctx);
2724 fd_install(fd, filp);
2729 path = filp->f_path;
2733 if (ctx->ctx_buf_fmt) {
2734 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2737 pfm_context_free(ctx);
2744 static inline unsigned long
2745 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2747 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2748 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2749 extern unsigned long carta_random32 (unsigned long seed);
2751 if (reg->flags & PFM_REGFL_RANDOM) {
2752 new_seed = carta_random32(old_seed);
2753 val -= (old_seed & mask); /* counter values are negative numbers! */
2754 if ((mask >> 32) != 0)
2755 /* construct a full 64-bit random value: */
2756 new_seed |= carta_random32(old_seed >> 32) << 32;
2757 reg->seed = new_seed;
2764 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2766 unsigned long mask = ovfl_regs[0];
2767 unsigned long reset_others = 0UL;
2772 * now restore reset value on sampling overflowed counters
2774 mask >>= PMU_FIRST_COUNTER;
2775 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2777 if ((mask & 0x1UL) == 0UL) continue;
2779 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2780 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2782 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2786 * Now take care of resetting the other registers
2788 for(i = 0; reset_others; i++, reset_others >>= 1) {
2790 if ((reset_others & 0x1) == 0) continue;
2792 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2794 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2795 is_long_reset ? "long" : "short", i, val));
2800 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2802 unsigned long mask = ovfl_regs[0];
2803 unsigned long reset_others = 0UL;
2807 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2809 if (ctx->ctx_state == PFM_CTX_MASKED) {
2810 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2815 * now restore reset value on sampling overflowed counters
2817 mask >>= PMU_FIRST_COUNTER;
2818 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2820 if ((mask & 0x1UL) == 0UL) continue;
2822 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2823 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2825 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2827 pfm_write_soft_counter(ctx, i, val);
2831 * Now take care of resetting the other registers
2833 for(i = 0; reset_others; i++, reset_others >>= 1) {
2835 if ((reset_others & 0x1) == 0) continue;
2837 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2839 if (PMD_IS_COUNTING(i)) {
2840 pfm_write_soft_counter(ctx, i, val);
2842 ia64_set_pmd(i, val);
2844 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2845 is_long_reset ? "long" : "short", i, val));
2851 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2853 struct task_struct *task;
2854 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2855 unsigned long value, pmc_pm;
2856 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2857 unsigned int cnum, reg_flags, flags, pmc_type;
2858 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2859 int is_monitor, is_counting, state;
2861 pfm_reg_check_t wr_func;
2862 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2864 state = ctx->ctx_state;
2865 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2866 is_system = ctx->ctx_fl_system;
2867 task = ctx->ctx_task;
2868 impl_pmds = pmu_conf->impl_pmds[0];
2870 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2874 * In system wide and when the context is loaded, access can only happen
2875 * when the caller is running on the CPU being monitored by the session.
2876 * It does not have to be the owner (ctx_task) of the context per se.
2878 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2879 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2882 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2884 expert_mode = pfm_sysctl.expert_mode;
2886 for (i = 0; i < count; i++, req++) {
2888 cnum = req->reg_num;
2889 reg_flags = req->reg_flags;
2890 value = req->reg_value;
2891 smpl_pmds = req->reg_smpl_pmds[0];
2892 reset_pmds = req->reg_reset_pmds[0];
2896 if (cnum >= PMU_MAX_PMCS) {
2897 DPRINT(("pmc%u is invalid\n", cnum));
2901 pmc_type = pmu_conf->pmc_desc[cnum].type;
2902 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2903 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2904 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2907 * we reject all non implemented PMC as well
2908 * as attempts to modify PMC[0-3] which are used
2909 * as status registers by the PMU
2911 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2912 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2915 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2917 * If the PMC is a monitor, then if the value is not the default:
2918 * - system-wide session: PMCx.pm=1 (privileged monitor)
2919 * - per-task : PMCx.pm=0 (user monitor)
2921 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2922 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2931 * enforce generation of overflow interrupt. Necessary on all
2934 value |= 1 << PMU_PMC_OI;
2936 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2937 flags |= PFM_REGFL_OVFL_NOTIFY;
2940 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2942 /* verify validity of smpl_pmds */
2943 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2944 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2948 /* verify validity of reset_pmds */
2949 if ((reset_pmds & impl_pmds) != reset_pmds) {
2950 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2954 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2955 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2958 /* eventid on non-counting monitors are ignored */
2962 * execute write checker, if any
2964 if (likely(expert_mode == 0 && wr_func)) {
2965 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2966 if (ret) goto error;
2971 * no error on this register
2973 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2976 * Now we commit the changes to the software state
2980 * update overflow information
2984 * full flag update each time a register is programmed
2986 ctx->ctx_pmds[cnum].flags = flags;
2988 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2989 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2990 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2993 * Mark all PMDS to be accessed as used.
2995 * We do not keep track of PMC because we have to
2996 * systematically restore ALL of them.
2998 * We do not update the used_monitors mask, because
2999 * if we have not programmed them, then will be in
3000 * a quiescent state, therefore we will not need to
3001 * mask/restore then when context is MASKED.
3003 CTX_USED_PMD(ctx, reset_pmds);
3004 CTX_USED_PMD(ctx, smpl_pmds);
3006 * make sure we do not try to reset on
3007 * restart because we have established new values
3009 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3012 * Needed in case the user does not initialize the equivalent
3013 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3014 * possible leak here.
3016 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3019 * keep track of the monitor PMC that we are using.
3020 * we save the value of the pmc in ctx_pmcs[] and if
3021 * the monitoring is not stopped for the context we also
3022 * place it in the saved state area so that it will be
3023 * picked up later by the context switch code.
3025 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3027 * The value in th_pmcs[] may be modified on overflow, i.e., when
3028 * monitoring needs to be stopped.
3030 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3033 * update context state
3035 ctx->ctx_pmcs[cnum] = value;
3039 * write thread state
3041 if (is_system == 0) ctx->th_pmcs[cnum] = value;
3044 * write hardware register if we can
3046 if (can_access_pmu) {
3047 ia64_set_pmc(cnum, value);
3052 * per-task SMP only here
3054 * we are guaranteed that the task is not running on the other CPU,
3055 * we indicate that this PMD will need to be reloaded if the task
3056 * is rescheduled on the CPU it ran last on.
3058 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3063 DPRINT(("pmc[%u]=0x%lx ld=%d apmu=%d flags=0x%x all_pmcs=0x%lx used_pmds=0x%lx eventid=%ld smpl_pmds=0x%lx reset_pmds=0x%lx reloads_pmcs=0x%lx used_monitors=0x%lx ovfl_regs=0x%lx\n",
3069 ctx->ctx_all_pmcs[0],
3070 ctx->ctx_used_pmds[0],
3071 ctx->ctx_pmds[cnum].eventid,
3074 ctx->ctx_reload_pmcs[0],
3075 ctx->ctx_used_monitors[0],
3076 ctx->ctx_ovfl_regs[0]));
3080 * make sure the changes are visible
3082 if (can_access_pmu) ia64_srlz_d();
3086 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3091 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3093 struct task_struct *task;
3094 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3095 unsigned long value, hw_value, ovfl_mask;
3097 int i, can_access_pmu = 0, state;
3098 int is_counting, is_loaded, is_system, expert_mode;
3100 pfm_reg_check_t wr_func;
3103 state = ctx->ctx_state;
3104 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3105 is_system = ctx->ctx_fl_system;
3106 ovfl_mask = pmu_conf->ovfl_val;
3107 task = ctx->ctx_task;
3109 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3112 * on both UP and SMP, we can only write to the PMC when the task is
3113 * the owner of the local PMU.
3115 if (likely(is_loaded)) {
3117 * In system wide and when the context is loaded, access can only happen
3118 * when the caller is running on the CPU being monitored by the session.
3119 * It does not have to be the owner (ctx_task) of the context per se.
3121 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3122 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3125 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3127 expert_mode = pfm_sysctl.expert_mode;
3129 for (i = 0; i < count; i++, req++) {
3131 cnum = req->reg_num;
3132 value = req->reg_value;
3134 if (!PMD_IS_IMPL(cnum)) {
3135 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3138 is_counting = PMD_IS_COUNTING(cnum);
3139 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3142 * execute write checker, if any
3144 if (unlikely(expert_mode == 0 && wr_func)) {
3145 unsigned long v = value;
3147 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3148 if (ret) goto abort_mission;
3155 * no error on this register
3157 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3160 * now commit changes to software state
3165 * update virtualized (64bits) counter
3169 * write context state
3171 ctx->ctx_pmds[cnum].lval = value;
3174 * when context is load we use the split value
3177 hw_value = value & ovfl_mask;
3178 value = value & ~ovfl_mask;
3182 * update reset values (not just for counters)
3184 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3185 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3188 * update randomization parameters (not just for counters)
3190 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3191 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3194 * update context value
3196 ctx->ctx_pmds[cnum].val = value;
3199 * Keep track of what we use
3201 * We do not keep track of PMC because we have to
3202 * systematically restore ALL of them.
3204 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3207 * mark this PMD register used as well
3209 CTX_USED_PMD(ctx, RDEP(cnum));
3212 * make sure we do not try to reset on
3213 * restart because we have established new values
3215 if (is_counting && state == PFM_CTX_MASKED) {
3216 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3221 * write thread state
3223 if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3226 * write hardware register if we can
3228 if (can_access_pmu) {
3229 ia64_set_pmd(cnum, hw_value);
3233 * we are guaranteed that the task is not running on the other CPU,
3234 * we indicate that this PMD will need to be reloaded if the task
3235 * is rescheduled on the CPU it ran last on.
3237 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3242 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3243 "long_reset=0x%lx notify=%c seed=0x%lx mask=0x%lx used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3249 ctx->ctx_pmds[cnum].val,
3250 ctx->ctx_pmds[cnum].short_reset,
3251 ctx->ctx_pmds[cnum].long_reset,
3252 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3253 ctx->ctx_pmds[cnum].seed,
3254 ctx->ctx_pmds[cnum].mask,
3255 ctx->ctx_used_pmds[0],
3256 ctx->ctx_pmds[cnum].reset_pmds[0],
3257 ctx->ctx_reload_pmds[0],
3258 ctx->ctx_all_pmds[0],
3259 ctx->ctx_ovfl_regs[0]));
3263 * make changes visible
3265 if (can_access_pmu) ia64_srlz_d();
3271 * for now, we have only one possibility for error
3273 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3278 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3279 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3280 * interrupt is delivered during the call, it will be kept pending until we leave, making
3281 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3282 * guaranteed to return consistent data to the user, it may simply be old. It is not
3283 * trivial to treat the overflow while inside the call because you may end up in
3284 * some module sampling buffer code causing deadlocks.
3287 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3289 struct task_struct *task;
3290 unsigned long val = 0UL, lval, ovfl_mask, sval;
3291 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3292 unsigned int cnum, reg_flags = 0;
3293 int i, can_access_pmu = 0, state;
3294 int is_loaded, is_system, is_counting, expert_mode;
3296 pfm_reg_check_t rd_func;
3299 * access is possible when loaded only for
3300 * self-monitoring tasks or in UP mode
3303 state = ctx->ctx_state;
3304 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3305 is_system = ctx->ctx_fl_system;
3306 ovfl_mask = pmu_conf->ovfl_val;
3307 task = ctx->ctx_task;
3309 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3311 if (likely(is_loaded)) {
3313 * In system wide and when the context is loaded, access can only happen
3314 * when the caller is running on the CPU being monitored by the session.
3315 * It does not have to be the owner (ctx_task) of the context per se.
3317 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3318 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3322 * this can be true when not self-monitoring only in UP
3324 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3326 if (can_access_pmu) ia64_srlz_d();
3328 expert_mode = pfm_sysctl.expert_mode;
3330 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3336 * on both UP and SMP, we can only read the PMD from the hardware register when
3337 * the task is the owner of the local PMU.
3340 for (i = 0; i < count; i++, req++) {
3342 cnum = req->reg_num;
3343 reg_flags = req->reg_flags;
3345 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3347 * we can only read the register that we use. That includes
3348 * the one we explicitly initialize AND the one we want included
3349 * in the sampling buffer (smpl_regs).
3351 * Having this restriction allows optimization in the ctxsw routine
3352 * without compromising security (leaks)
3354 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3356 sval = ctx->ctx_pmds[cnum].val;
3357 lval = ctx->ctx_pmds[cnum].lval;
3358 is_counting = PMD_IS_COUNTING(cnum);
3361 * If the task is not the current one, then we check if the
3362 * PMU state is still in the local live register due to lazy ctxsw.
3363 * If true, then we read directly from the registers.
3365 if (can_access_pmu){
3366 val = ia64_get_pmd(cnum);
3369 * context has been saved
3370 * if context is zombie, then task does not exist anymore.
3371 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3373 val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3375 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3379 * XXX: need to check for overflow when loaded
3386 * execute read checker, if any
3388 if (unlikely(expert_mode == 0 && rd_func)) {
3389 unsigned long v = val;
3390 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3391 if (ret) goto error;
3396 PFM_REG_RETFLAG_SET(reg_flags, 0);
3398 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3401 * update register return value, abort all if problem during copy.
3402 * we only modify the reg_flags field. no check mode is fine because
3403 * access has been verified upfront in sys_perfmonctl().
3405 req->reg_value = val;
3406 req->reg_flags = reg_flags;
3407 req->reg_last_reset_val = lval;
3413 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3418 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3422 if (req == NULL) return -EINVAL;
3424 ctx = GET_PMU_CTX();
3426 if (ctx == NULL) return -EINVAL;
3429 * for now limit to current task, which is enough when calling
3430 * from overflow handler
3432 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3434 return pfm_write_pmcs(ctx, req, nreq, regs);
3436 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3439 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3443 if (req == NULL) return -EINVAL;
3445 ctx = GET_PMU_CTX();
3447 if (ctx == NULL) return -EINVAL;
3450 * for now limit to current task, which is enough when calling
3451 * from overflow handler
3453 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3455 return pfm_read_pmds(ctx, req, nreq, regs);
3457 EXPORT_SYMBOL(pfm_mod_read_pmds);
3460 * Only call this function when a process it trying to
3461 * write the debug registers (reading is always allowed)
3464 pfm_use_debug_registers(struct task_struct *task)
3466 pfm_context_t *ctx = task->thread.pfm_context;
3467 unsigned long flags;
3470 if (pmu_conf->use_rr_dbregs == 0) return 0;
3472 DPRINT(("called for [%d]\n", task_pid_nr(task)));
3477 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3480 * Even on SMP, we do not need to use an atomic here because
3481 * the only way in is via ptrace() and this is possible only when the
3482 * process is stopped. Even in the case where the ctxsw out is not totally
3483 * completed by the time we come here, there is no way the 'stopped' process
3484 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3485 * So this is always safe.
3487 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3492 * We cannot allow setting breakpoints when system wide monitoring
3493 * sessions are using the debug registers.
3495 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3498 pfm_sessions.pfs_ptrace_use_dbregs++;
3500 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3501 pfm_sessions.pfs_ptrace_use_dbregs,
3502 pfm_sessions.pfs_sys_use_dbregs,
3503 task_pid_nr(task), ret));
3511 * This function is called for every task that exits with the
3512 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3513 * able to use the debug registers for debugging purposes via
3514 * ptrace(). Therefore we know it was not using them for
3515 * perfmormance monitoring, so we only decrement the number
3516 * of "ptraced" debug register users to keep the count up to date
3519 pfm_release_debug_registers(struct task_struct *task)
3521 unsigned long flags;
3524 if (pmu_conf->use_rr_dbregs == 0) return 0;
3527 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3528 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3531 pfm_sessions.pfs_ptrace_use_dbregs--;
3540 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3542 struct task_struct *task;
3543 pfm_buffer_fmt_t *fmt;
3544 pfm_ovfl_ctrl_t rst_ctrl;
3545 int state, is_system;
3548 state = ctx->ctx_state;
3549 fmt = ctx->ctx_buf_fmt;
3550 is_system = ctx->ctx_fl_system;
3551 task = PFM_CTX_TASK(ctx);
3554 case PFM_CTX_MASKED:
3556 case PFM_CTX_LOADED:
3557 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3559 case PFM_CTX_UNLOADED:
3560 case PFM_CTX_ZOMBIE:
3561 DPRINT(("invalid state=%d\n", state));
3564 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3569 * In system wide and when the context is loaded, access can only happen
3570 * when the caller is running on the CPU being monitored by the session.
3571 * It does not have to be the owner (ctx_task) of the context per se.
3573 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3574 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3579 if (unlikely(task == NULL)) {
3580 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3584 if (task == current || is_system) {
3586 fmt = ctx->ctx_buf_fmt;
3588 DPRINT(("restarting self %d ovfl=0x%lx\n",
3590 ctx->ctx_ovfl_regs[0]));
3592 if (CTX_HAS_SMPL(ctx)) {
3594 prefetch(ctx->ctx_smpl_hdr);
3596 rst_ctrl.bits.mask_monitoring = 0;
3597 rst_ctrl.bits.reset_ovfl_pmds = 0;
3599 if (state == PFM_CTX_LOADED)
3600 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3602 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3604 rst_ctrl.bits.mask_monitoring = 0;
3605 rst_ctrl.bits.reset_ovfl_pmds = 1;
3609 if (rst_ctrl.bits.reset_ovfl_pmds)
3610 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3612 if (rst_ctrl.bits.mask_monitoring == 0) {
3613 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3615 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3617 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3619 // cannot use pfm_stop_monitoring(task, regs);
3623 * clear overflowed PMD mask to remove any stale information
3625 ctx->ctx_ovfl_regs[0] = 0UL;
3628 * back to LOADED state
3630 ctx->ctx_state = PFM_CTX_LOADED;
3633 * XXX: not really useful for self monitoring
3635 ctx->ctx_fl_can_restart = 0;
3641 * restart another task
3645 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3646 * one is seen by the task.
3648 if (state == PFM_CTX_MASKED) {
3649 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3651 * will prevent subsequent restart before this one is
3652 * seen by other task
3654 ctx->ctx_fl_can_restart = 0;
3658 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3659 * the task is blocked or on its way to block. That's the normal
3660 * restart path. If the monitoring is not masked, then the task
3661 * can be actively monitoring and we cannot directly intervene.
3662 * Therefore we use the trap mechanism to catch the task and
3663 * force it to reset the buffer/reset PMDs.
3665 * if non-blocking, then we ensure that the task will go into
3666 * pfm_handle_work() before returning to user mode.
3668 * We cannot explicitly reset another task, it MUST always
3669 * be done by the task itself. This works for system wide because
3670 * the tool that is controlling the session is logically doing
3671 * "self-monitoring".
3673 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3674 DPRINT(("unblocking [%d] \n", task_pid_nr(task)));
3675 complete(&ctx->ctx_restart_done);
3677 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3679 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3681 PFM_SET_WORK_PENDING(task, 1);
3683 set_notify_resume(task);
3686 * XXX: send reschedule if task runs on another CPU
3693 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3695 unsigned int m = *(unsigned int *)arg;
3697 pfm_sysctl.debug = m == 0 ? 0 : 1;
3699 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3702 memset(pfm_stats, 0, sizeof(pfm_stats));
3703 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3709 * arg can be NULL and count can be zero for this function
3712 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3714 struct thread_struct *thread = NULL;
3715 struct task_struct *task;
3716 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3717 unsigned long flags;
3722 int i, can_access_pmu = 0;
3723 int is_system, is_loaded;
3725 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3727 state = ctx->ctx_state;
3728 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3729 is_system = ctx->ctx_fl_system;
3730 task = ctx->ctx_task;
3732 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3735 * on both UP and SMP, we can only write to the PMC when the task is
3736 * the owner of the local PMU.
3739 thread = &task->thread;
3741 * In system wide and when the context is loaded, access can only happen
3742 * when the caller is running on the CPU being monitored by the session.
3743 * It does not have to be the owner (ctx_task) of the context per se.
3745 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3746 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3749 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3753 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3754 * ensuring that no real breakpoint can be installed via this call.
3756 * IMPORTANT: regs can be NULL in this function
3759 first_time = ctx->ctx_fl_using_dbreg == 0;
3762 * don't bother if we are loaded and task is being debugged
3764 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3765 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3770 * check for debug registers in system wide mode
3772 * If though a check is done in pfm_context_load(),
3773 * we must repeat it here, in case the registers are
3774 * written after the context is loaded
3779 if (first_time && is_system) {
3780 if (pfm_sessions.pfs_ptrace_use_dbregs)
3783 pfm_sessions.pfs_sys_use_dbregs++;
3788 if (ret != 0) return ret;
3791 * mark ourself as user of the debug registers for
3794 ctx->ctx_fl_using_dbreg = 1;
3797 * clear hardware registers to make sure we don't
3798 * pick up stale state.
3800 * for a system wide session, we do not use
3801 * thread.dbr, thread.ibr because this process
3802 * never leaves the current CPU and the state
3803 * is shared by all processes running on it
3805 if (first_time && can_access_pmu) {
3806 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3807 for (i=0; i < pmu_conf->num_ibrs; i++) {
3808 ia64_set_ibr(i, 0UL);
3809 ia64_dv_serialize_instruction();
3812 for (i=0; i < pmu_conf->num_dbrs; i++) {
3813 ia64_set_dbr(i, 0UL);
3814 ia64_dv_serialize_data();
3820 * Now install the values into the registers
3822 for (i = 0; i < count; i++, req++) {
3824 rnum = req->dbreg_num;
3825 dbreg.val = req->dbreg_value;
3829 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3830 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3831 rnum, dbreg.val, mode, i, count));
3837 * make sure we do not install enabled breakpoint
3840 if (mode == PFM_CODE_RR)
3841 dbreg.ibr.ibr_x = 0;
3843 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3846 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3849 * Debug registers, just like PMC, can only be modified
3850 * by a kernel call. Moreover, perfmon() access to those
3851 * registers are centralized in this routine. The hardware
3852 * does not modify the value of these registers, therefore,
3853 * if we save them as they are written, we can avoid having
3854 * to save them on context switch out. This is made possible
3855 * by the fact that when perfmon uses debug registers, ptrace()
3856 * won't be able to modify them concurrently.
3858 if (mode == PFM_CODE_RR) {
3859 CTX_USED_IBR(ctx, rnum);
3861 if (can_access_pmu) {
3862 ia64_set_ibr(rnum, dbreg.val);
3863 ia64_dv_serialize_instruction();
3866 ctx->ctx_ibrs[rnum] = dbreg.val;
3868 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3869 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3871 CTX_USED_DBR(ctx, rnum);
3873 if (can_access_pmu) {
3874 ia64_set_dbr(rnum, dbreg.val);
3875 ia64_dv_serialize_data();
3877 ctx->ctx_dbrs[rnum] = dbreg.val;
3879 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3880 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3888 * in case it was our first attempt, we undo the global modifications
3892 if (ctx->ctx_fl_system) {
3893 pfm_sessions.pfs_sys_use_dbregs--;
3896 ctx->ctx_fl_using_dbreg = 0;
3899 * install error return flag
3901 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3907 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3909 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3913 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3915 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3919 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3923 if (req == NULL) return -EINVAL;
3925 ctx = GET_PMU_CTX();
3927 if (ctx == NULL) return -EINVAL;
3930 * for now limit to current task, which is enough when calling
3931 * from overflow handler
3933 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3935 return pfm_write_ibrs(ctx, req, nreq, regs);
3937 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3940 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3944 if (req == NULL) return -EINVAL;
3946 ctx = GET_PMU_CTX();
3948 if (ctx == NULL) return -EINVAL;
3951 * for now limit to current task, which is enough when calling
3952 * from overflow handler
3954 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3956 return pfm_write_dbrs(ctx, req, nreq, regs);
3958 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3962 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3964 pfarg_features_t *req = (pfarg_features_t *)arg;
3966 req->ft_version = PFM_VERSION;
3971 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3973 struct pt_regs *tregs;
3974 struct task_struct *task = PFM_CTX_TASK(ctx);
3975 int state, is_system;
3977 state = ctx->ctx_state;
3978 is_system = ctx->ctx_fl_system;
3981 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3983 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3986 * In system wide and when the context is loaded, access can only happen
3987 * when the caller is running on the CPU being monitored by the session.
3988 * It does not have to be the owner (ctx_task) of the context per se.
3990 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3991 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3994 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3995 task_pid_nr(PFM_CTX_TASK(ctx)),
3999 * in system mode, we need to update the PMU directly
4000 * and the user level state of the caller, which may not
4001 * necessarily be the creator of the context.
4005 * Update local PMU first
4009 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4013 * update local cpuinfo
4015 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4018 * stop monitoring, does srlz.i
4023 * stop monitoring in the caller
4025 ia64_psr(regs)->pp = 0;
4033 if (task == current) {
4034 /* stop monitoring at kernel level */
4038 * stop monitoring at the user level
4040 ia64_psr(regs)->up = 0;
4042 tregs = task_pt_regs(task);
4045 * stop monitoring at the user level
4047 ia64_psr(tregs)->up = 0;
4050 * monitoring disabled in kernel at next reschedule
4052 ctx->ctx_saved_psr_up = 0;
4053 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4060 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4062 struct pt_regs *tregs;
4063 int state, is_system;
4065 state = ctx->ctx_state;
4066 is_system = ctx->ctx_fl_system;
4068 if (state != PFM_CTX_LOADED) return -EINVAL;
4071 * In system wide and when the context is loaded, access can only happen
4072 * when the caller is running on the CPU being monitored by the session.
4073 * It does not have to be the owner (ctx_task) of the context per se.
4075 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4076 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4081 * in system mode, we need to update the PMU directly
4082 * and the user level state of the caller, which may not
4083 * necessarily be the creator of the context.
4088 * set user level psr.pp for the caller
4090 ia64_psr(regs)->pp = 1;
4093 * now update the local PMU and cpuinfo
4095 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4098 * start monitoring at kernel level
4103 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4113 if (ctx->ctx_task == current) {
4115 /* start monitoring at kernel level */
4119 * activate monitoring at user level
4121 ia64_psr(regs)->up = 1;
4124 tregs = task_pt_regs(ctx->ctx_task);
4127 * start monitoring at the kernel level the next
4128 * time the task is scheduled
4130 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4133 * activate monitoring at user level
4135 ia64_psr(tregs)->up = 1;
4141 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4143 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4148 for (i = 0; i < count; i++, req++) {
4150 cnum = req->reg_num;
4152 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4154 req->reg_value = PMC_DFL_VAL(cnum);
4156 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4158 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4163 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4168 pfm_check_task_exist(pfm_context_t *ctx)
4170 struct task_struct *g, *t;
4173 read_lock(&tasklist_lock);
4175 do_each_thread (g, t) {
4176 if (t->thread.pfm_context == ctx) {
4180 } while_each_thread (g, t);
4182 read_unlock(&tasklist_lock);
4184 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4190 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4192 struct task_struct *task;
4193 struct thread_struct *thread;
4194 struct pfm_context_t *old;
4195 unsigned long flags;
4197 struct task_struct *owner_task = NULL;
4199 pfarg_load_t *req = (pfarg_load_t *)arg;
4200 unsigned long *pmcs_source, *pmds_source;
4203 int state, is_system, set_dbregs = 0;
4205 state = ctx->ctx_state;
4206 is_system = ctx->ctx_fl_system;
4208 * can only load from unloaded or terminated state
4210 if (state != PFM_CTX_UNLOADED) {
4211 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4217 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4219 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4220 DPRINT(("cannot use blocking mode on self\n"));
4224 ret = pfm_get_task(ctx, req->load_pid, &task);
4226 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4233 * system wide is self monitoring only
4235 if (is_system && task != current) {
4236 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4241 thread = &task->thread;
4245 * cannot load a context which is using range restrictions,
4246 * into a task that is being debugged.
4248 if (ctx->ctx_fl_using_dbreg) {
4249 if (thread->flags & IA64_THREAD_DBG_VALID) {
4251 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4257 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4258 DPRINT(("cannot load [%d] dbregs in use\n",
4259 task_pid_nr(task)));
4262 pfm_sessions.pfs_sys_use_dbregs++;
4263 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4270 if (ret) goto error;
4274 * SMP system-wide monitoring implies self-monitoring.
4276 * The programming model expects the task to
4277 * be pinned on a CPU throughout the session.
4278 * Here we take note of the current CPU at the
4279 * time the context is loaded. No call from
4280 * another CPU will be allowed.
4282 * The pinning via shed_setaffinity()
4283 * must be done by the calling task prior
4286 * systemwide: keep track of CPU this session is supposed to run on
4288 the_cpu = ctx->ctx_cpu = smp_processor_id();
4292 * now reserve the session
4294 ret = pfm_reserve_session(current, is_system, the_cpu);
4295 if (ret) goto error;
4298 * task is necessarily stopped at this point.
4300 * If the previous context was zombie, then it got removed in
4301 * pfm_save_regs(). Therefore we should not see it here.
4302 * If we see a context, then this is an active context
4304 * XXX: needs to be atomic
4306 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4307 thread->pfm_context, ctx));
4310 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4312 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4316 pfm_reset_msgq(ctx);
4318 ctx->ctx_state = PFM_CTX_LOADED;
4321 * link context to task
4323 ctx->ctx_task = task;
4327 * we load as stopped
4329 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4330 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4332 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4334 thread->flags |= IA64_THREAD_PM_VALID;
4338 * propagate into thread-state
4340 pfm_copy_pmds(task, ctx);
4341 pfm_copy_pmcs(task, ctx);
4343 pmcs_source = ctx->th_pmcs;
4344 pmds_source = ctx->th_pmds;
4347 * always the case for system-wide
4349 if (task == current) {
4351 if (is_system == 0) {
4353 /* allow user level control */
4354 ia64_psr(regs)->sp = 0;
4355 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4357 SET_LAST_CPU(ctx, smp_processor_id());
4359 SET_ACTIVATION(ctx);
4362 * push the other task out, if any
4364 owner_task = GET_PMU_OWNER();
4365 if (owner_task) pfm_lazy_save_regs(owner_task);
4369 * load all PMD from ctx to PMU (as opposed to thread state)
4370 * restore all PMC from ctx to PMU
4372 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4373 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4375 ctx->ctx_reload_pmcs[0] = 0UL;
4376 ctx->ctx_reload_pmds[0] = 0UL;
4379 * guaranteed safe by earlier check against DBG_VALID
4381 if (ctx->ctx_fl_using_dbreg) {
4382 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4383 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4388 SET_PMU_OWNER(task, ctx);
4390 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4393 * when not current, task MUST be stopped, so this is safe
4395 regs = task_pt_regs(task);
4397 /* force a full reload */
4398 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4399 SET_LAST_CPU(ctx, -1);
4401 /* initial saved psr (stopped) */
4402 ctx->ctx_saved_psr_up = 0UL;
4403 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4409 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4412 * we must undo the dbregs setting (for system-wide)
4414 if (ret && set_dbregs) {
4416 pfm_sessions.pfs_sys_use_dbregs--;
4420 * release task, there is now a link with the context
4422 if (is_system == 0 && task != current) {
4426 ret = pfm_check_task_exist(ctx);
4428 ctx->ctx_state = PFM_CTX_UNLOADED;
4429 ctx->ctx_task = NULL;
4437 * in this function, we do not need to increase the use count
4438 * for the task via get_task_struct(), because we hold the
4439 * context lock. If the task were to disappear while having
4440 * a context attached, it would go through pfm_exit_thread()
4441 * which also grabs the context lock and would therefore be blocked
4442 * until we are here.
4444 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4447 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4449 struct task_struct *task = PFM_CTX_TASK(ctx);
4450 struct pt_regs *tregs;
4451 int prev_state, is_system;
4454 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4456 prev_state = ctx->ctx_state;
4457 is_system = ctx->ctx_fl_system;
4460 * unload only when necessary
4462 if (prev_state == PFM_CTX_UNLOADED) {
4463 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4468 * clear psr and dcr bits
4470 ret = pfm_stop(ctx, NULL, 0, regs);
4471 if (ret) return ret;
4473 ctx->ctx_state = PFM_CTX_UNLOADED;
4476 * in system mode, we need to update the PMU directly
4477 * and the user level state of the caller, which may not
4478 * necessarily be the creator of the context.
4485 * local PMU is taken care of in pfm_stop()
4487 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4488 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4491 * save PMDs in context
4494 pfm_flush_pmds(current, ctx);
4497 * at this point we are done with the PMU
4498 * so we can unreserve the resource.
4500 if (prev_state != PFM_CTX_ZOMBIE)
4501 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4504 * disconnect context from task
4506 task->thread.pfm_context = NULL;
4508 * disconnect task from context
4510 ctx->ctx_task = NULL;
4513 * There is nothing more to cleanup here.
4521 tregs = task == current ? regs : task_pt_regs(task);
4523 if (task == current) {
4525 * cancel user level control
4527 ia64_psr(regs)->sp = 1;
4529 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4532 * save PMDs to context
4535 pfm_flush_pmds(task, ctx);
4538 * at this point we are done with the PMU
4539 * so we can unreserve the resource.
4541 * when state was ZOMBIE, we have already unreserved.
4543 if (prev_state != PFM_CTX_ZOMBIE)
4544 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4547 * reset activation counter and psr
4549 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4550 SET_LAST_CPU(ctx, -1);
4553 * PMU state will not be restored
4555 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4558 * break links between context and task
4560 task->thread.pfm_context = NULL;
4561 ctx->ctx_task = NULL;
4563 PFM_SET_WORK_PENDING(task, 0);
4565 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4566 ctx->ctx_fl_can_restart = 0;
4567 ctx->ctx_fl_going_zombie = 0;
4569 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4576 * called only from exit_thread(): task == current
4577 * we come here only if current has a context attached (loaded or masked)
4580 pfm_exit_thread(struct task_struct *task)
4583 unsigned long flags;
4584 struct pt_regs *regs = task_pt_regs(task);
4588 ctx = PFM_GET_CTX(task);
4590 PROTECT_CTX(ctx, flags);
4592 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4594 state = ctx->ctx_state;
4596 case PFM_CTX_UNLOADED:
4598 * only comes to this function if pfm_context is not NULL, i.e., cannot
4599 * be in unloaded state
4601 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4603 case PFM_CTX_LOADED:
4604 case PFM_CTX_MASKED:
4605 ret = pfm_context_unload(ctx, NULL, 0, regs);
4607 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4609 DPRINT(("ctx unloaded for current state was %d\n", state));
4611 pfm_end_notify_user(ctx);
4613 case PFM_CTX_ZOMBIE:
4614 ret = pfm_context_unload(ctx, NULL, 0, regs);
4616 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4621 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4624 UNPROTECT_CTX(ctx, flags);
4626 { u64 psr = pfm_get_psr();
4627 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4628 BUG_ON(GET_PMU_OWNER());
4629 BUG_ON(ia64_psr(regs)->up);
4630 BUG_ON(ia64_psr(regs)->pp);
4634 * All memory free operations (especially for vmalloc'ed memory)
4635 * MUST be done with interrupts ENABLED.
4637 if (free_ok) pfm_context_free(ctx);
4641 * functions MUST be listed in the increasing order of their index (see permfon.h)
4643 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4644 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4645 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4646 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4647 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4649 static pfm_cmd_desc_t pfm_cmd_tab[]={
4650 /* 0 */PFM_CMD_NONE,
4651 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4652 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4653 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4654 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4655 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4656 /* 6 */PFM_CMD_NONE,
4657 /* 7 */PFM_CMD_NONE,
4658 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4659 /* 9 */PFM_CMD_NONE,
4660 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4661 /* 11 */PFM_CMD_NONE,
4662 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4663 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4664 /* 14 */PFM_CMD_NONE,
4665 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4666 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4667 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4668 /* 18 */PFM_CMD_NONE,
4669 /* 19 */PFM_CMD_NONE,
4670 /* 20 */PFM_CMD_NONE,
4671 /* 21 */PFM_CMD_NONE,
4672 /* 22 */PFM_CMD_NONE,
4673 /* 23 */PFM_CMD_NONE,
4674 /* 24 */PFM_CMD_NONE,
4675 /* 25 */PFM_CMD_NONE,
4676 /* 26 */PFM_CMD_NONE,
4677 /* 27 */PFM_CMD_NONE,
4678 /* 28 */PFM_CMD_NONE,
4679 /* 29 */PFM_CMD_NONE,
4680 /* 30 */PFM_CMD_NONE,
4681 /* 31 */PFM_CMD_NONE,
4682 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4683 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4685 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4688 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4690 struct task_struct *task;
4691 int state, old_state;
4694 state = ctx->ctx_state;
4695 task = ctx->ctx_task;
4698 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4702 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4706 task->state, PFM_CMD_STOPPED(cmd)));
4709 * self-monitoring always ok.
4711 * for system-wide the caller can either be the creator of the
4712 * context (to one to which the context is attached to) OR
4713 * a task running on the same CPU as the session.
4715 if (task == current || ctx->ctx_fl_system) return 0;
4718 * we are monitoring another thread
4721 case PFM_CTX_UNLOADED:
4723 * if context is UNLOADED we are safe to go
4726 case PFM_CTX_ZOMBIE:
4728 * no command can operate on a zombie context
4730 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4732 case PFM_CTX_MASKED:
4734 * PMU state has been saved to software even though
4735 * the thread may still be running.
4737 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4741 * context is LOADED or MASKED. Some commands may need to have
4744 * We could lift this restriction for UP but it would mean that
4745 * the user has no guarantee the task would not run between
4746 * two successive calls to perfmonctl(). That's probably OK.
4747 * If this user wants to ensure the task does not run, then
4748 * the task must be stopped.
4750 if (PFM_CMD_STOPPED(cmd)) {
4751 if (!task_is_stopped_or_traced(task)) {
4752 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4756 * task is now stopped, wait for ctxsw out
4758 * This is an interesting point in the code.
4759 * We need to unprotect the context because
4760 * the pfm_save_regs() routines needs to grab
4761 * the same lock. There are danger in doing
4762 * this because it leaves a window open for
4763 * another task to get access to the context
4764 * and possibly change its state. The one thing
4765 * that is not possible is for the context to disappear
4766 * because we are protected by the VFS layer, i.e.,
4767 * get_fd()/put_fd().
4771 UNPROTECT_CTX(ctx, flags);
4773 wait_task_inactive(task, 0);
4775 PROTECT_CTX(ctx, flags);
4778 * we must recheck to verify if state has changed
4780 if (ctx->ctx_state != old_state) {
4781 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4789 * system-call entry point (must return long)
4792 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4794 struct file *file = NULL;
4795 pfm_context_t *ctx = NULL;
4796 unsigned long flags = 0UL;
4797 void *args_k = NULL;
4798 long ret; /* will expand int return types */
4799 size_t base_sz, sz, xtra_sz = 0;
4800 int narg, completed_args = 0, call_made = 0, cmd_flags;
4801 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4802 int (*getsize)(void *arg, size_t *sz);
4803 #define PFM_MAX_ARGSIZE 4096
4806 * reject any call if perfmon was disabled at initialization
4808 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4810 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4811 DPRINT(("invalid cmd=%d\n", cmd));
4815 func = pfm_cmd_tab[cmd].cmd_func;
4816 narg = pfm_cmd_tab[cmd].cmd_narg;
4817 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4818 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4819 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4821 if (unlikely(func == NULL)) {
4822 DPRINT(("invalid cmd=%d\n", cmd));
4826 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4834 * check if number of arguments matches what the command expects
4836 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4840 sz = xtra_sz + base_sz*count;
4842 * limit abuse to min page size
4844 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4845 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4850 * allocate default-sized argument buffer
4852 if (likely(count && args_k == NULL)) {
4853 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4854 if (args_k == NULL) return -ENOMEM;
4862 * assume sz = 0 for command without parameters
4864 if (sz && copy_from_user(args_k, arg, sz)) {
4865 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4870 * check if command supports extra parameters
4872 if (completed_args == 0 && getsize) {
4874 * get extra parameters size (based on main argument)
4876 ret = (*getsize)(args_k, &xtra_sz);
4877 if (ret) goto error_args;
4881 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4883 /* retry if necessary */
4884 if (likely(xtra_sz)) goto restart_args;
4887 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4892 if (unlikely(file == NULL)) {
4893 DPRINT(("invalid fd %d\n", fd));
4896 if (unlikely(PFM_IS_FILE(file) == 0)) {
4897 DPRINT(("fd %d not related to perfmon\n", fd));
4901 ctx = (pfm_context_t *)file->private_data;
4902 if (unlikely(ctx == NULL)) {
4903 DPRINT(("no context for fd %d\n", fd));
4906 prefetch(&ctx->ctx_state);
4908 PROTECT_CTX(ctx, flags);
4911 * check task is stopped
4913 ret = pfm_check_task_state(ctx, cmd, flags);
4914 if (unlikely(ret)) goto abort_locked;
4917 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4923 DPRINT(("context unlocked\n"));
4924 UNPROTECT_CTX(ctx, flags);
4927 /* copy argument back to user, if needed */
4928 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4936 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4942 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4944 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4945 pfm_ovfl_ctrl_t rst_ctrl;
4949 state = ctx->ctx_state;
4951 * Unlock sampling buffer and reset index atomically
4952 * XXX: not really needed when blocking
4954 if (CTX_HAS_SMPL(ctx)) {
4956 rst_ctrl.bits.mask_monitoring = 0;
4957 rst_ctrl.bits.reset_ovfl_pmds = 0;
4959 if (state == PFM_CTX_LOADED)
4960 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4962 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4964 rst_ctrl.bits.mask_monitoring = 0;
4965 rst_ctrl.bits.reset_ovfl_pmds = 1;
4969 if (rst_ctrl.bits.reset_ovfl_pmds) {
4970 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4972 if (rst_ctrl.bits.mask_monitoring == 0) {
4973 DPRINT(("resuming monitoring\n"));
4974 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4976 DPRINT(("stopping monitoring\n"));
4977 //pfm_stop_monitoring(current, regs);
4979 ctx->ctx_state = PFM_CTX_LOADED;
4984 * context MUST BE LOCKED when calling
4985 * can only be called for current
4988 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4992 DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4994 ret = pfm_context_unload(ctx, NULL, 0, regs);
4996 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
5000 * and wakeup controlling task, indicating we are now disconnected
5002 wake_up_interruptible(&ctx->ctx_zombieq);
5005 * given that context is still locked, the controlling
5006 * task will only get access when we return from
5007 * pfm_handle_work().
5011 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5014 * pfm_handle_work() can be called with interrupts enabled
5015 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5016 * call may sleep, therefore we must re-enable interrupts
5017 * to avoid deadlocks. It is safe to do so because this function
5018 * is called ONLY when returning to user level (pUStk=1), in which case
5019 * there is no risk of kernel stack overflow due to deep
5020 * interrupt nesting.
5023 pfm_handle_work(void)
5026 struct pt_regs *regs;
5027 unsigned long flags, dummy_flags;
5028 unsigned long ovfl_regs;
5029 unsigned int reason;
5032 ctx = PFM_GET_CTX(current);
5034 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5035 task_pid_nr(current));
5039 PROTECT_CTX(ctx, flags);
5041 PFM_SET_WORK_PENDING(current, 0);
5043 regs = task_pt_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)
5060 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5061 if (reason == PFM_TRAP_REASON_RESET)
5065 * restore interrupt mask to what it was on entry.
5066 * Could be enabled/diasbled.
5068 UNPROTECT_CTX(ctx, flags);
5071 * force interrupt enable because of down_interruptible()
5075 DPRINT(("before block sleeping\n"));
5078 * may go through without blocking on SMP systems
5079 * if restart has been received already by the time we call down()
5081 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5083 DPRINT(("after block sleeping ret=%d\n", ret));
5086 * lock context and mask interrupts again
5087 * We save flags into a dummy because we may have
5088 * altered interrupts mask compared to entry in this
5091 PROTECT_CTX(ctx, dummy_flags);
5094 * we need to read the ovfl_regs only after wake-up
5095 * because we may have had pfm_write_pmds() in between
5096 * and that can changed PMD values and therefore
5097 * ovfl_regs is reset for these new PMD values.
5099 ovfl_regs = ctx->ctx_ovfl_regs[0];
5101 if (ctx->ctx_fl_going_zombie) {
5103 DPRINT(("context is zombie, bailing out\n"));
5104 pfm_context_force_terminate(ctx, regs);
5108 * in case of interruption of down() we don't restart anything
5114 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5115 ctx->ctx_ovfl_regs[0] = 0UL;
5119 * restore flags as they were upon entry
5121 UNPROTECT_CTX(ctx, flags);
5125 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5127 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5128 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5132 DPRINT(("waking up somebody\n"));
5134 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5137 * safe, we are not in intr handler, nor in ctxsw when
5140 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5146 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5148 pfm_msg_t *msg = NULL;
5150 if (ctx->ctx_fl_no_msg == 0) {
5151 msg = pfm_get_new_msg(ctx);
5153 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5157 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5158 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5159 msg->pfm_ovfl_msg.msg_active_set = 0;
5160 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5161 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5162 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5163 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5164 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5167 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5173 return pfm_notify_user(ctx, msg);
5177 pfm_end_notify_user(pfm_context_t *ctx)
5181 msg = pfm_get_new_msg(ctx);
5183 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5187 memset(msg, 0, sizeof(*msg));
5189 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5190 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5191 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5193 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5198 return pfm_notify_user(ctx, msg);
5202 * main overflow processing routine.
5203 * it can be called from the interrupt path or explicitly during the context switch code
5206 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5208 pfm_ovfl_arg_t *ovfl_arg;
5210 unsigned long old_val, ovfl_val, new_val;
5211 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5212 unsigned long tstamp;
5213 pfm_ovfl_ctrl_t ovfl_ctrl;
5214 unsigned int i, has_smpl;
5215 int must_notify = 0;
5217 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5220 * sanity test. Should never happen
5222 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5224 tstamp = ia64_get_itc();
5225 mask = pmc0 >> PMU_FIRST_COUNTER;
5226 ovfl_val = pmu_conf->ovfl_val;
5227 has_smpl = CTX_HAS_SMPL(ctx);
5229 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5230 "used_pmds=0x%lx\n",
5232 task ? task_pid_nr(task): -1,
5233 (regs ? regs->cr_iip : 0),
5234 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5235 ctx->ctx_used_pmds[0]));
5239 * first we update the virtual counters
5240 * assume there was a prior ia64_srlz_d() issued
5242 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5244 /* skip pmd which did not overflow */
5245 if ((mask & 0x1) == 0) continue;
5248 * Note that the pmd is not necessarily 0 at this point as qualified events
5249 * may have happened before the PMU was frozen. The residual count is not
5250 * taken into consideration here but will be with any read of the pmd via
5253 old_val = new_val = ctx->ctx_pmds[i].val;
5254 new_val += 1 + ovfl_val;
5255 ctx->ctx_pmds[i].val = new_val;
5258 * check for overflow condition
5260 if (likely(old_val > new_val)) {
5261 ovfl_pmds |= 1UL << i;
5262 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5265 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5269 ia64_get_pmd(i) & ovfl_val,
5275 * there was no 64-bit overflow, nothing else to do
5277 if (ovfl_pmds == 0UL) return;
5280 * reset all control bits
5286 * if a sampling format module exists, then we "cache" the overflow by
5287 * calling the module's handler() routine.
5290 unsigned long start_cycles, end_cycles;
5291 unsigned long pmd_mask;
5293 int this_cpu = smp_processor_id();
5295 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5296 ovfl_arg = &ctx->ctx_ovfl_arg;
5298 prefetch(ctx->ctx_smpl_hdr);
5300 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5304 if ((pmd_mask & 0x1) == 0) continue;
5306 ovfl_arg->ovfl_pmd = (unsigned char )i;
5307 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5308 ovfl_arg->active_set = 0;
5309 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5310 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5312 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5313 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5314 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5317 * copy values of pmds of interest. Sampling format may copy them
5318 * into sampling buffer.
5321 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5322 if ((smpl_pmds & 0x1) == 0) continue;
5323 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5324 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5328 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5330 start_cycles = ia64_get_itc();
5333 * call custom buffer format record (handler) routine
5335 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5337 end_cycles = ia64_get_itc();
5340 * For those controls, we take the union because they have
5341 * an all or nothing behavior.
5343 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5344 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5345 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5347 * build the bitmask of pmds to reset now
5349 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5351 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5354 * when the module cannot handle the rest of the overflows, we abort right here
5356 if (ret && pmd_mask) {
5357 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5358 pmd_mask<<PMU_FIRST_COUNTER));
5361 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5363 ovfl_pmds &= ~reset_pmds;
5366 * when no sampling module is used, then the default
5367 * is to notify on overflow if requested by user
5369 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5370 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5371 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5372 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5374 * if needed, we reset all overflowed pmds
5376 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5379 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5382 * reset the requested PMD registers using the short reset values
5385 unsigned long bm = reset_pmds;
5386 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5389 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5391 * keep track of what to reset when unblocking
5393 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5396 * check for blocking context
5398 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5400 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5403 * set the perfmon specific checking pending work for the task
5405 PFM_SET_WORK_PENDING(task, 1);
5408 * when coming from ctxsw, current still points to the
5409 * previous task, therefore we must work with task and not current.
5411 set_notify_resume(task);
5414 * defer until state is changed (shorten spin window). the context is locked
5415 * anyway, so the signal receiver would come spin for nothing.
5420 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5421 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5422 PFM_GET_WORK_PENDING(task),
5423 ctx->ctx_fl_trap_reason,
5426 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5428 * in case monitoring must be stopped, we toggle the psr bits
5430 if (ovfl_ctrl.bits.mask_monitoring) {
5431 pfm_mask_monitoring(task);
5432 ctx->ctx_state = PFM_CTX_MASKED;
5433 ctx->ctx_fl_can_restart = 1;
5437 * send notification now
5439 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5444 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5446 task ? task_pid_nr(task) : -1,
5452 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5453 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5454 * come here as zombie only if the task is the current task. In which case, we
5455 * can access the PMU hardware directly.
5457 * Note that zombies do have PM_VALID set. So here we do the minimal.
5459 * In case the context was zombified it could not be reclaimed at the time
5460 * the monitoring program exited. At this point, the PMU reservation has been
5461 * returned, the sampiing buffer has been freed. We must convert this call
5462 * into a spurious interrupt. However, we must also avoid infinite overflows
5463 * by stopping monitoring for this task. We can only come here for a per-task
5464 * context. All we need to do is to stop monitoring using the psr bits which
5465 * are always task private. By re-enabling secure montioring, we ensure that
5466 * the monitored task will not be able to re-activate monitoring.
5467 * The task will eventually be context switched out, at which point the context
5468 * will be reclaimed (that includes releasing ownership of the PMU).
5470 * So there might be a window of time where the number of per-task session is zero
5471 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5472 * context. This is safe because if a per-task session comes in, it will push this one
5473 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5474 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5475 * also push our zombie context out.
5477 * Overall pretty hairy stuff....
5479 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5481 ia64_psr(regs)->up = 0;
5482 ia64_psr(regs)->sp = 1;
5487 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5489 struct task_struct *task;
5491 unsigned long flags;
5493 int this_cpu = smp_processor_id();
5496 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5499 * srlz.d done before arriving here
5501 pmc0 = ia64_get_pmc(0);
5503 task = GET_PMU_OWNER();
5504 ctx = GET_PMU_CTX();
5507 * if we have some pending bits set
5508 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5510 if (PMC0_HAS_OVFL(pmc0) && task) {
5512 * we assume that pmc0.fr is always set here
5516 if (!ctx) goto report_spurious1;
5518 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5519 goto report_spurious2;
5521 PROTECT_CTX_NOPRINT(ctx, flags);
5523 pfm_overflow_handler(task, ctx, pmc0, regs);
5525 UNPROTECT_CTX_NOPRINT(ctx, flags);
5528 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5532 * keep it unfrozen at all times
5539 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5540 this_cpu, task_pid_nr(task));
5544 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5552 pfm_interrupt_handler(int irq, void *arg)
5554 unsigned long start_cycles, total_cycles;
5555 unsigned long min, max;
5558 struct pt_regs *regs = get_irq_regs();
5560 this_cpu = get_cpu();
5561 if (likely(!pfm_alt_intr_handler)) {
5562 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5563 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5565 start_cycles = ia64_get_itc();
5567 ret = pfm_do_interrupt_handler(arg, regs);
5569 total_cycles = ia64_get_itc();
5572 * don't measure spurious interrupts
5574 if (likely(ret == 0)) {
5575 total_cycles -= start_cycles;
5577 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5578 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5580 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5584 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5587 put_cpu_no_resched();
5592 * /proc/perfmon interface, for debug only
5595 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5598 pfm_proc_start(struct seq_file *m, loff_t *pos)
5601 return PFM_PROC_SHOW_HEADER;
5604 while (*pos <= NR_CPUS) {
5605 if (cpu_online(*pos - 1)) {
5606 return (void *)*pos;
5614 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5617 return pfm_proc_start(m, pos);
5621 pfm_proc_stop(struct seq_file *m, void *v)
5626 pfm_proc_show_header(struct seq_file *m)
5628 struct list_head * pos;
5629 pfm_buffer_fmt_t * entry;
5630 unsigned long flags;
5633 "perfmon version : %u.%u\n"
5636 "expert mode : %s\n"
5637 "ovfl_mask : 0x%lx\n"
5638 "PMU flags : 0x%x\n",
5639 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5641 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5642 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5649 "proc_sessions : %u\n"
5650 "sys_sessions : %u\n"
5651 "sys_use_dbregs : %u\n"
5652 "ptrace_use_dbregs : %u\n",
5653 pfm_sessions.pfs_task_sessions,
5654 pfm_sessions.pfs_sys_sessions,
5655 pfm_sessions.pfs_sys_use_dbregs,
5656 pfm_sessions.pfs_ptrace_use_dbregs);
5660 spin_lock(&pfm_buffer_fmt_lock);
5662 list_for_each(pos, &pfm_buffer_fmt_list) {
5663 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5664 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5675 entry->fmt_uuid[10],
5676 entry->fmt_uuid[11],
5677 entry->fmt_uuid[12],
5678 entry->fmt_uuid[13],
5679 entry->fmt_uuid[14],
5680 entry->fmt_uuid[15],
5683 spin_unlock(&pfm_buffer_fmt_lock);
5688 pfm_proc_show(struct seq_file *m, void *v)
5694 if (v == PFM_PROC_SHOW_HEADER) {
5695 pfm_proc_show_header(m);
5699 /* show info for CPU (v - 1) */
5703 "CPU%-2d overflow intrs : %lu\n"
5704 "CPU%-2d overflow cycles : %lu\n"
5705 "CPU%-2d overflow min : %lu\n"
5706 "CPU%-2d overflow max : %lu\n"
5707 "CPU%-2d smpl handler calls : %lu\n"
5708 "CPU%-2d smpl handler cycles : %lu\n"
5709 "CPU%-2d spurious intrs : %lu\n"
5710 "CPU%-2d replay intrs : %lu\n"
5711 "CPU%-2d syst_wide : %d\n"
5712 "CPU%-2d dcr_pp : %d\n"
5713 "CPU%-2d exclude idle : %d\n"
5714 "CPU%-2d owner : %d\n"
5715 "CPU%-2d context : %p\n"
5716 "CPU%-2d activations : %lu\n",
5717 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5718 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5719 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5720 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5721 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5722 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5723 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5724 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5725 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5726 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5727 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5728 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5729 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5730 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5732 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5734 psr = pfm_get_psr();
5739 "CPU%-2d psr : 0x%lx\n"
5740 "CPU%-2d pmc0 : 0x%lx\n",
5742 cpu, ia64_get_pmc(0));
5744 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5745 if (PMC_IS_COUNTING(i) == 0) continue;
5747 "CPU%-2d pmc%u : 0x%lx\n"
5748 "CPU%-2d pmd%u : 0x%lx\n",
5749 cpu, i, ia64_get_pmc(i),
5750 cpu, i, ia64_get_pmd(i));
5756 const struct seq_operations pfm_seq_ops = {
5757 .start = pfm_proc_start,
5758 .next = pfm_proc_next,
5759 .stop = pfm_proc_stop,
5760 .show = pfm_proc_show
5764 pfm_proc_open(struct inode *inode, struct file *file)
5766 return seq_open(file, &pfm_seq_ops);
5771 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5772 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5773 * is active or inactive based on mode. We must rely on the value in
5774 * local_cpu_data->pfm_syst_info
5777 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5779 struct pt_regs *regs;
5781 unsigned long dcr_pp;
5783 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5786 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5787 * on every CPU, so we can rely on the pid to identify the idle task.
5789 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5790 regs = task_pt_regs(task);
5791 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5795 * if monitoring has started
5798 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5800 * context switching in?
5803 /* mask monitoring for the idle task */
5804 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5810 * context switching out
5811 * restore monitoring for next task
5813 * Due to inlining this odd if-then-else construction generates
5816 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5825 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5827 struct task_struct *task = ctx->ctx_task;
5829 ia64_psr(regs)->up = 0;
5830 ia64_psr(regs)->sp = 1;
5832 if (GET_PMU_OWNER() == task) {
5833 DPRINT(("cleared ownership for [%d]\n",
5834 task_pid_nr(ctx->ctx_task)));
5835 SET_PMU_OWNER(NULL, NULL);
5839 * disconnect the task from the context and vice-versa
5841 PFM_SET_WORK_PENDING(task, 0);
5843 task->thread.pfm_context = NULL;
5844 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5846 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task)));
5851 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5854 pfm_save_regs(struct task_struct *task)
5857 unsigned long flags;
5861 ctx = PFM_GET_CTX(task);
5862 if (ctx == NULL) return;
5865 * we always come here with interrupts ALREADY disabled by
5866 * the scheduler. So we simply need to protect against concurrent
5867 * access, not CPU concurrency.
5869 flags = pfm_protect_ctx_ctxsw(ctx);
5871 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5872 struct pt_regs *regs = task_pt_regs(task);
5876 pfm_force_cleanup(ctx, regs);
5878 BUG_ON(ctx->ctx_smpl_hdr);
5880 pfm_unprotect_ctx_ctxsw(ctx, flags);
5882 pfm_context_free(ctx);
5887 * save current PSR: needed because we modify it
5890 psr = pfm_get_psr();
5892 BUG_ON(psr & (IA64_PSR_I));
5896 * This is the last instruction which may generate an overflow
5898 * We do not need to set psr.sp because, it is irrelevant in kernel.
5899 * It will be restored from ipsr when going back to user level
5904 * keep a copy of psr.up (for reload)
5906 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5909 * release ownership of this PMU.
5910 * PM interrupts are masked, so nothing
5913 SET_PMU_OWNER(NULL, NULL);
5916 * we systematically save the PMD as we have no
5917 * guarantee we will be schedule at that same
5920 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5923 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5924 * we will need it on the restore path to check
5925 * for pending overflow.
5927 ctx->th_pmcs[0] = ia64_get_pmc(0);
5930 * unfreeze PMU if had pending overflows
5932 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5935 * finally, allow context access.
5936 * interrupts will still be masked after this call.
5938 pfm_unprotect_ctx_ctxsw(ctx, flags);
5941 #else /* !CONFIG_SMP */
5943 pfm_save_regs(struct task_struct *task)
5948 ctx = PFM_GET_CTX(task);
5949 if (ctx == NULL) return;
5952 * save current PSR: needed because we modify it
5954 psr = pfm_get_psr();
5956 BUG_ON(psr & (IA64_PSR_I));
5960 * This is the last instruction which may generate an overflow
5962 * We do not need to set psr.sp because, it is irrelevant in kernel.
5963 * It will be restored from ipsr when going back to user level
5968 * keep a copy of psr.up (for reload)
5970 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5974 pfm_lazy_save_regs (struct task_struct *task)
5977 unsigned long flags;
5979 { u64 psr = pfm_get_psr();
5980 BUG_ON(psr & IA64_PSR_UP);
5983 ctx = PFM_GET_CTX(task);
5986 * we need to mask PMU overflow here to
5987 * make sure that we maintain pmc0 until
5988 * we save it. overflow interrupts are
5989 * treated as spurious if there is no
5992 * XXX: I don't think this is necessary
5994 PROTECT_CTX(ctx,flags);
5997 * release ownership of this PMU.
5998 * must be done before we save the registers.
6000 * after this call any PMU interrupt is treated
6003 SET_PMU_OWNER(NULL, NULL);
6006 * save all the pmds we use
6008 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6011 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6012 * it is needed to check for pended overflow
6013 * on the restore path
6015 ctx->th_pmcs[0] = ia64_get_pmc(0);
6018 * unfreeze PMU if had pending overflows
6020 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6023 * now get can unmask PMU interrupts, they will
6024 * be treated as purely spurious and we will not
6025 * lose any information
6027 UNPROTECT_CTX(ctx,flags);
6029 #endif /* CONFIG_SMP */
6033 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6036 pfm_load_regs (struct task_struct *task)
6039 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6040 unsigned long flags;
6042 int need_irq_resend;
6044 ctx = PFM_GET_CTX(task);
6045 if (unlikely(ctx == NULL)) return;
6047 BUG_ON(GET_PMU_OWNER());
6050 * possible on unload
6052 if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6055 * we always come here with interrupts ALREADY disabled by
6056 * the scheduler. So we simply need to protect against concurrent
6057 * access, not CPU concurrency.
6059 flags = pfm_protect_ctx_ctxsw(ctx);
6060 psr = pfm_get_psr();
6062 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6064 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6065 BUG_ON(psr & IA64_PSR_I);
6067 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6068 struct pt_regs *regs = task_pt_regs(task);
6070 BUG_ON(ctx->ctx_smpl_hdr);
6072 pfm_force_cleanup(ctx, regs);
6074 pfm_unprotect_ctx_ctxsw(ctx, flags);
6077 * this one (kmalloc'ed) is fine with interrupts disabled
6079 pfm_context_free(ctx);
6085 * we restore ALL the debug registers to avoid picking up
6088 if (ctx->ctx_fl_using_dbreg) {
6089 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6090 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6093 * retrieve saved psr.up
6095 psr_up = ctx->ctx_saved_psr_up;
6098 * if we were the last user of the PMU on that CPU,
6099 * then nothing to do except restore psr
6101 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6104 * retrieve partial reload masks (due to user modifications)
6106 pmc_mask = ctx->ctx_reload_pmcs[0];
6107 pmd_mask = ctx->ctx_reload_pmds[0];
6111 * To avoid leaking information to the user level when psr.sp=0,
6112 * we must reload ALL implemented pmds (even the ones we don't use).
6113 * In the kernel we only allow PFM_READ_PMDS on registers which
6114 * we initialized or requested (sampling) so there is no risk there.
6116 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6119 * ALL accessible PMCs are systematically reloaded, unused registers
6120 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6121 * up stale configuration.
6123 * PMC0 is never in the mask. It is always restored separately.
6125 pmc_mask = ctx->ctx_all_pmcs[0];
6128 * when context is MASKED, we will restore PMC with plm=0
6129 * and PMD with stale information, but that's ok, nothing
6132 * XXX: optimize here
6134 if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6135 if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6138 * check for pending overflow at the time the state
6141 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6143 * reload pmc0 with the overflow information
6144 * On McKinley PMU, this will trigger a PMU interrupt
6146 ia64_set_pmc(0, ctx->th_pmcs[0]);
6148 ctx->th_pmcs[0] = 0UL;
6151 * will replay the PMU interrupt
6153 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6155 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6159 * we just did a reload, so we reset the partial reload fields
6161 ctx->ctx_reload_pmcs[0] = 0UL;
6162 ctx->ctx_reload_pmds[0] = 0UL;
6164 SET_LAST_CPU(ctx, smp_processor_id());
6167 * dump activation value for this PMU
6171 * record current activation for this context
6173 SET_ACTIVATION(ctx);
6176 * establish new ownership.
6178 SET_PMU_OWNER(task, ctx);
6181 * restore the psr.up bit. measurement
6183 * no PMU interrupt can happen at this point
6184 * because we still have interrupts disabled.
6186 if (likely(psr_up)) pfm_set_psr_up();
6189 * allow concurrent access to context
6191 pfm_unprotect_ctx_ctxsw(ctx, flags);
6193 #else /* !CONFIG_SMP */
6195 * reload PMU state for UP kernels
6196 * in 2.5 we come here with interrupts disabled
6199 pfm_load_regs (struct task_struct *task)
6202 struct task_struct *owner;
6203 unsigned long pmd_mask, pmc_mask;
6205 int need_irq_resend;
6207 owner = GET_PMU_OWNER();
6208 ctx = PFM_GET_CTX(task);
6209 psr = pfm_get_psr();
6211 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6212 BUG_ON(psr & IA64_PSR_I);
6215 * we restore ALL the debug registers to avoid picking up
6218 * This must be done even when the task is still the owner
6219 * as the registers may have been modified via ptrace()
6220 * (not perfmon) by the previous task.
6222 if (ctx->ctx_fl_using_dbreg) {
6223 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6224 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6228 * retrieved saved psr.up
6230 psr_up = ctx->ctx_saved_psr_up;
6231 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6234 * short path, our state is still there, just
6235 * need to restore psr and we go
6237 * we do not touch either PMC nor PMD. the psr is not touched
6238 * by the overflow_handler. So we are safe w.r.t. to interrupt
6239 * concurrency even without interrupt masking.
6241 if (likely(owner == task)) {
6242 if (likely(psr_up)) pfm_set_psr_up();
6247 * someone else is still using the PMU, first push it out and
6248 * then we'll be able to install our stuff !
6250 * Upon return, there will be no owner for the current PMU
6252 if (owner) pfm_lazy_save_regs(owner);
6255 * To avoid leaking information to the user level when psr.sp=0,
6256 * we must reload ALL implemented pmds (even the ones we don't use).
6257 * In the kernel we only allow PFM_READ_PMDS on registers which
6258 * we initialized or requested (sampling) so there is no risk there.
6260 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6263 * ALL accessible PMCs are systematically reloaded, unused registers
6264 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6265 * up stale configuration.
6267 * PMC0 is never in the mask. It is always restored separately
6269 pmc_mask = ctx->ctx_all_pmcs[0];
6271 pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6272 pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6275 * check for pending overflow at the time the state
6278 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6280 * reload pmc0 with the overflow information
6281 * On McKinley PMU, this will trigger a PMU interrupt
6283 ia64_set_pmc(0, ctx->th_pmcs[0]);
6286 ctx->th_pmcs[0] = 0UL;
6289 * will replay the PMU interrupt
6291 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6293 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6297 * establish new ownership.
6299 SET_PMU_OWNER(task, ctx);
6302 * restore the psr.up bit. measurement
6304 * no PMU interrupt can happen at this point
6305 * because we still have interrupts disabled.
6307 if (likely(psr_up)) pfm_set_psr_up();
6309 #endif /* CONFIG_SMP */
6312 * this function assumes monitoring is stopped
6315 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6318 unsigned long mask2, val, pmd_val, ovfl_val;
6319 int i, can_access_pmu = 0;
6323 * is the caller the task being monitored (or which initiated the
6324 * session for system wide measurements)
6326 is_self = ctx->ctx_task == task ? 1 : 0;
6329 * can access PMU is task is the owner of the PMU state on the current CPU
6330 * or if we are running on the CPU bound to the context in system-wide mode
6331 * (that is not necessarily the task the context is attached to in this mode).
6332 * In system-wide we always have can_access_pmu true because a task running on an
6333 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6335 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6336 if (can_access_pmu) {
6338 * Mark the PMU as not owned
6339 * This will cause the interrupt handler to do nothing in case an overflow
6340 * interrupt was in-flight
6341 * This also guarantees that pmc0 will contain the final state
6342 * It virtually gives us full control on overflow processing from that point
6345 SET_PMU_OWNER(NULL, NULL);
6346 DPRINT(("releasing ownership\n"));
6349 * read current overflow status:
6351 * we are guaranteed to read the final stable state
6354 pmc0 = ia64_get_pmc(0); /* slow */
6357 * reset freeze bit, overflow status information destroyed
6361 pmc0 = ctx->th_pmcs[0];
6363 * clear whatever overflow status bits there were
6365 ctx->th_pmcs[0] = 0;
6367 ovfl_val = pmu_conf->ovfl_val;
6369 * we save all the used pmds
6370 * we take care of overflows for counting PMDs
6372 * XXX: sampling situation is not taken into account here
6374 mask2 = ctx->ctx_used_pmds[0];
6376 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6378 for (i = 0; mask2; i++, mask2>>=1) {
6380 /* skip non used pmds */
6381 if ((mask2 & 0x1) == 0) continue;
6384 * can access PMU always true in system wide mode
6386 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6388 if (PMD_IS_COUNTING(i)) {
6389 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6392 ctx->ctx_pmds[i].val,
6396 * we rebuild the full 64 bit value of the counter
6398 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6401 * now everything is in ctx_pmds[] and we need
6402 * to clear the saved context from save_regs() such that
6403 * pfm_read_pmds() gets the correct value
6408 * take care of overflow inline
6410 if (pmc0 & (1UL << i)) {
6411 val += 1 + ovfl_val;
6412 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6416 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6418 if (is_self) ctx->th_pmds[i] = pmd_val;
6420 ctx->ctx_pmds[i].val = val;
6424 static struct irqaction perfmon_irqaction = {
6425 .handler = pfm_interrupt_handler,
6426 .flags = IRQF_DISABLED,
6431 pfm_alt_save_pmu_state(void *data)
6433 struct pt_regs *regs;
6435 regs = task_pt_regs(current);
6437 DPRINT(("called\n"));
6440 * should not be necessary but
6441 * let's take not risk
6445 ia64_psr(regs)->pp = 0;
6448 * This call is required
6449 * May cause a spurious interrupt on some processors
6457 pfm_alt_restore_pmu_state(void *data)
6459 struct pt_regs *regs;
6461 regs = task_pt_regs(current);
6463 DPRINT(("called\n"));
6466 * put PMU back in state expected
6471 ia64_psr(regs)->pp = 0;
6474 * perfmon runs with PMU unfrozen at all times
6482 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6487 /* some sanity checks */
6488 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6490 /* do the easy test first */
6491 if (pfm_alt_intr_handler) return -EBUSY;
6493 /* one at a time in the install or remove, just fail the others */
6494 if (!spin_trylock(&pfm_alt_install_check)) {
6498 /* reserve our session */
6499 for_each_online_cpu(reserve_cpu) {
6500 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6501 if (ret) goto cleanup_reserve;
6504 /* save the current system wide pmu states */
6505 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6507 DPRINT(("on_each_cpu() failed: %d\n", ret));
6508 goto cleanup_reserve;
6511 /* officially change to the alternate interrupt handler */
6512 pfm_alt_intr_handler = hdl;
6514 spin_unlock(&pfm_alt_install_check);
6519 for_each_online_cpu(i) {
6520 /* don't unreserve more than we reserved */
6521 if (i >= reserve_cpu) break;
6523 pfm_unreserve_session(NULL, 1, i);
6526 spin_unlock(&pfm_alt_install_check);
6530 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6533 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6538 if (hdl == NULL) return -EINVAL;
6540 /* cannot remove someone else's handler! */
6541 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6543 /* one at a time in the install or remove, just fail the others */
6544 if (!spin_trylock(&pfm_alt_install_check)) {
6548 pfm_alt_intr_handler = NULL;
6550 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6552 DPRINT(("on_each_cpu() failed: %d\n", ret));
6555 for_each_online_cpu(i) {
6556 pfm_unreserve_session(NULL, 1, i);
6559 spin_unlock(&pfm_alt_install_check);
6563 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6566 * perfmon initialization routine, called from the initcall() table
6568 static int init_pfm_fs(void);
6576 family = local_cpu_data->family;
6581 if ((*p)->probe() == 0) goto found;
6582 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6593 static const struct file_operations pfm_proc_fops = {
6594 .open = pfm_proc_open,
6596 .llseek = seq_lseek,
6597 .release = seq_release,
6603 unsigned int n, n_counters, i;
6605 printk("perfmon: version %u.%u IRQ %u\n",
6608 IA64_PERFMON_VECTOR);
6610 if (pfm_probe_pmu()) {
6611 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6612 local_cpu_data->family);
6617 * compute the number of implemented PMD/PMC from the
6618 * description tables
6621 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6622 if (PMC_IS_IMPL(i) == 0) continue;
6623 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6626 pmu_conf->num_pmcs = n;
6628 n = 0; n_counters = 0;
6629 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6630 if (PMD_IS_IMPL(i) == 0) continue;
6631 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6633 if (PMD_IS_COUNTING(i)) n_counters++;
6635 pmu_conf->num_pmds = n;
6636 pmu_conf->num_counters = n_counters;
6639 * sanity checks on the number of debug registers
6641 if (pmu_conf->use_rr_dbregs) {
6642 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6643 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6647 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6648 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6654 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6658 pmu_conf->num_counters,
6659 ffz(pmu_conf->ovfl_val));
6662 if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6663 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6669 * create /proc/perfmon (mostly for debugging purposes)
6671 perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6672 if (perfmon_dir == NULL) {
6673 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6679 * create /proc/sys/kernel/perfmon (for debugging purposes)
6681 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6684 * initialize all our spinlocks
6686 spin_lock_init(&pfm_sessions.pfs_lock);
6687 spin_lock_init(&pfm_buffer_fmt_lock);
6691 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6696 __initcall(pfm_init);
6699 * this function is called before pfm_init()
6702 pfm_init_percpu (void)
6704 static int first_time=1;
6706 * make sure no measurement is active
6707 * (may inherit programmed PMCs from EFI).
6713 * we run with the PMU not frozen at all times
6718 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6722 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6727 * used for debug purposes only
6730 dump_pmu_state(const char *from)
6732 struct task_struct *task;
6733 struct pt_regs *regs;
6735 unsigned long psr, dcr, info, flags;
6738 local_irq_save(flags);
6740 this_cpu = smp_processor_id();
6741 regs = task_pt_regs(current);
6742 info = PFM_CPUINFO_GET();
6743 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6745 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6746 local_irq_restore(flags);
6750 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6753 task_pid_nr(current),
6757 task = GET_PMU_OWNER();
6758 ctx = GET_PMU_CTX();
6760 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6762 psr = pfm_get_psr();
6764 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",
6767 psr & IA64_PSR_PP ? 1 : 0,
6768 psr & IA64_PSR_UP ? 1 : 0,
6769 dcr & IA64_DCR_PP ? 1 : 0,
6772 ia64_psr(regs)->pp);
6774 ia64_psr(regs)->up = 0;
6775 ia64_psr(regs)->pp = 0;
6777 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6778 if (PMC_IS_IMPL(i) == 0) continue;
6779 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, ctx->th_pmcs[i]);
6782 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6783 if (PMD_IS_IMPL(i) == 0) continue;
6784 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, ctx->th_pmds[i]);
6788 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6791 ctx->ctx_smpl_vaddr,
6795 ctx->ctx_saved_psr_up);
6797 local_irq_restore(flags);
6801 * called from process.c:copy_thread(). task is new child.
6804 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6806 struct thread_struct *thread;
6808 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6810 thread = &task->thread;
6813 * cut links inherited from parent (current)
6815 thread->pfm_context = NULL;
6817 PFM_SET_WORK_PENDING(task, 0);
6820 * the psr bits are already set properly in copy_threads()
6823 #else /* !CONFIG_PERFMON */
6825 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6829 #endif /* CONFIG_PERFMON */