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/config.h>
23 #include <linux/module.h>
24 #include <linux/kernel.h>
25 #include <linux/sched.h>
26 #include <linux/interrupt.h>
27 #include <linux/smp_lock.h>
28 #include <linux/proc_fs.h>
29 #include <linux/seq_file.h>
30 #include <linux/init.h>
31 #include <linux/vmalloc.h>
33 #include <linux/sysctl.h>
34 #include <linux/list.h>
35 #include <linux/file.h>
36 #include <linux/poll.h>
37 #include <linux/vfs.h>
38 #include <linux/pagemap.h>
39 #include <linux/mount.h>
40 #include <linux/version.h>
41 #include <linux/bitops.h>
42 #include <linux/vs_memory.h>
43 #include <linux/vs_cvirt.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)
67 * depth of message queue
69 #define PFM_MAX_MSGS 32
70 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
73 * type of a PMU register (bitmask).
75 * bit0 : register implemented
78 * bit4 : pmc has pmc.pm
79 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
80 * bit6-7 : register type
83 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
84 #define PFM_REG_IMPL 0x1 /* register implemented */
85 #define PFM_REG_END 0x2 /* end marker */
86 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
87 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
88 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
89 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
90 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
92 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
93 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
95 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
97 /* i assumed unsigned */
98 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
99 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
101 /* XXX: these assume that register i is implemented */
102 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
103 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
104 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
105 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
107 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
108 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
109 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
110 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
112 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
113 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
115 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
116 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
117 #define PFM_CTX_TASK(h) (h)->ctx_task
119 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
121 /* XXX: does not support more than 64 PMDs */
122 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
123 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
125 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
127 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
128 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
129 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
130 #define PFM_CODE_RR 0 /* requesting code range restriction */
131 #define PFM_DATA_RR 1 /* requestion data range restriction */
133 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
134 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
135 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
137 #define RDEP(x) (1UL<<(x))
140 * context protection macros
142 * - we need to protect against CPU concurrency (spin_lock)
143 * - we need to protect against PMU overflow interrupts (local_irq_disable)
145 * - we need to protect against PMU overflow interrupts (local_irq_disable)
147 * spin_lock_irqsave()/spin_lock_irqrestore():
148 * in SMP: local_irq_disable + spin_lock
149 * in UP : local_irq_disable
151 * spin_lock()/spin_lock():
152 * in UP : removed automatically
153 * in SMP: protect against context accesses from other CPU. interrupts
154 * are not masked. This is useful for the PMU interrupt handler
155 * because we know we will not get PMU concurrency in that code.
157 #define PROTECT_CTX(c, f) \
159 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \
160 spin_lock_irqsave(&(c)->ctx_lock, f); \
161 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
164 #define UNPROTECT_CTX(c, f) \
166 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
167 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
170 #define PROTECT_CTX_NOPRINT(c, f) \
172 spin_lock_irqsave(&(c)->ctx_lock, f); \
176 #define UNPROTECT_CTX_NOPRINT(c, f) \
178 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
182 #define PROTECT_CTX_NOIRQ(c) \
184 spin_lock(&(c)->ctx_lock); \
187 #define UNPROTECT_CTX_NOIRQ(c) \
189 spin_unlock(&(c)->ctx_lock); \
195 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
196 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
197 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
199 #else /* !CONFIG_SMP */
200 #define SET_ACTIVATION(t) do {} while(0)
201 #define GET_ACTIVATION(t) do {} while(0)
202 #define INC_ACTIVATION(t) do {} while(0)
203 #endif /* CONFIG_SMP */
205 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
206 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
207 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
209 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
210 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
212 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
215 * cmp0 must be the value of pmc0
217 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
219 #define PFMFS_MAGIC 0xa0b4d889
224 #define PFM_DEBUGGING 1
228 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
231 #define DPRINT_ovfl(a) \
233 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
238 * 64-bit software counter structure
240 * the next_reset_type is applied to the next call to pfm_reset_regs()
243 unsigned long val; /* virtual 64bit counter value */
244 unsigned long lval; /* last reset value */
245 unsigned long long_reset; /* reset value on sampling overflow */
246 unsigned long short_reset; /* reset value on overflow */
247 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
248 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
249 unsigned long seed; /* seed for random-number generator */
250 unsigned long mask; /* mask for random-number generator */
251 unsigned int flags; /* notify/do not notify */
252 unsigned long eventid; /* overflow event identifier */
259 unsigned int block:1; /* when 1, task will blocked on user notifications */
260 unsigned int system:1; /* do system wide monitoring */
261 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
262 unsigned int is_sampling:1; /* true if using a custom format */
263 unsigned int excl_idle:1; /* exclude idle task in system wide session */
264 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
265 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
266 unsigned int no_msg:1; /* no message sent on overflow */
267 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
268 unsigned int reserved:22;
269 } pfm_context_flags_t;
271 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
272 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
273 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
277 * perfmon context: encapsulates all the state of a monitoring session
280 typedef struct pfm_context {
281 spinlock_t ctx_lock; /* context protection */
283 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
284 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
286 struct task_struct *ctx_task; /* task to which context is attached */
288 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
290 struct semaphore ctx_restart_sem; /* use for blocking notification mode */
292 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
293 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
294 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
296 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
297 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
298 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
300 unsigned long ctx_pmcs[IA64_NUM_PMC_REGS]; /* saved copies of PMC values */
302 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
303 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
304 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
305 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
307 pfm_counter_t ctx_pmds[IA64_NUM_PMD_REGS]; /* software state for PMDS */
309 u64 ctx_saved_psr_up; /* only contains psr.up value */
311 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
312 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
313 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
315 int ctx_fd; /* file descriptor used my this context */
316 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
318 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
319 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
320 unsigned long ctx_smpl_size; /* size of sampling buffer */
321 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
323 wait_queue_head_t ctx_msgq_wait;
324 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
327 struct fasync_struct *ctx_async_queue;
329 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
333 * magic number used to verify that structure is really
336 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
338 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
341 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
342 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
344 #define SET_LAST_CPU(ctx, v) do {} while(0)
345 #define GET_LAST_CPU(ctx) do {} while(0)
349 #define ctx_fl_block ctx_flags.block
350 #define ctx_fl_system ctx_flags.system
351 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
352 #define ctx_fl_is_sampling ctx_flags.is_sampling
353 #define ctx_fl_excl_idle ctx_flags.excl_idle
354 #define ctx_fl_going_zombie ctx_flags.going_zombie
355 #define ctx_fl_trap_reason ctx_flags.trap_reason
356 #define ctx_fl_no_msg ctx_flags.no_msg
357 #define ctx_fl_can_restart ctx_flags.can_restart
359 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
360 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
363 * global information about all sessions
364 * mostly used to synchronize between system wide and per-process
367 spinlock_t pfs_lock; /* lock the structure */
369 unsigned int pfs_task_sessions; /* number of per task sessions */
370 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
371 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
372 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
373 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
377 * information about a PMC or PMD.
378 * dep_pmd[]: a bitmask of dependent PMD registers
379 * dep_pmc[]: a bitmask of dependent PMC registers
381 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
385 unsigned long default_value; /* power-on default value */
386 unsigned long reserved_mask; /* bitmask of reserved bits */
387 pfm_reg_check_t read_check;
388 pfm_reg_check_t write_check;
389 unsigned long dep_pmd[4];
390 unsigned long dep_pmc[4];
393 /* assume cnum is a valid monitor */
394 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
397 * This structure is initialized at boot time and contains
398 * a description of the PMU main characteristics.
400 * If the probe function is defined, detection is based
401 * on its return value:
402 * - 0 means recognized PMU
403 * - anything else means not supported
404 * When the probe function is not defined, then the pmu_family field
405 * is used and it must match the host CPU family such that:
406 * - cpu->family & config->pmu_family != 0
409 unsigned long ovfl_val; /* overflow value for counters */
411 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
412 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
414 unsigned int num_pmcs; /* number of PMCS: computed at init time */
415 unsigned int num_pmds; /* number of PMDS: computed at init time */
416 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
417 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
419 char *pmu_name; /* PMU family name */
420 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
421 unsigned int flags; /* pmu specific flags */
422 unsigned int num_ibrs; /* number of IBRS: computed at init time */
423 unsigned int num_dbrs; /* number of DBRS: computed at init time */
424 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
425 int (*probe)(void); /* customized probe routine */
426 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
431 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
434 * debug register related type definitions
437 unsigned long ibr_mask:56;
438 unsigned long ibr_plm:4;
439 unsigned long ibr_ig:3;
440 unsigned long ibr_x:1;
444 unsigned long dbr_mask:56;
445 unsigned long dbr_plm:4;
446 unsigned long dbr_ig:2;
447 unsigned long dbr_w:1;
448 unsigned long dbr_r:1;
459 * perfmon command descriptions
462 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
465 unsigned int cmd_narg;
467 int (*cmd_getsize)(void *arg, size_t *sz);
470 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
471 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
472 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
473 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
476 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
477 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
478 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
479 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
480 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
482 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
485 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
486 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
487 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
488 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
489 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
490 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
491 unsigned long pfm_smpl_handler_calls;
492 unsigned long pfm_smpl_handler_cycles;
493 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
497 * perfmon internal variables
499 static pfm_stats_t pfm_stats[NR_CPUS];
500 static pfm_session_t pfm_sessions; /* global sessions information */
502 static spinlock_t pfm_alt_install_check = SPIN_LOCK_UNLOCKED;
503 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
505 static struct proc_dir_entry *perfmon_dir;
506 static pfm_uuid_t pfm_null_uuid = {0,};
508 static spinlock_t pfm_buffer_fmt_lock;
509 static LIST_HEAD(pfm_buffer_fmt_list);
511 static pmu_config_t *pmu_conf;
513 /* sysctl() controls */
514 pfm_sysctl_t pfm_sysctl;
515 EXPORT_SYMBOL(pfm_sysctl);
517 static ctl_table pfm_ctl_table[]={
518 {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
519 {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
520 {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
521 {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
524 static ctl_table pfm_sysctl_dir[] = {
525 {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
528 static ctl_table pfm_sysctl_root[] = {
529 {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
532 static struct ctl_table_header *pfm_sysctl_header;
534 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
535 static int pfm_flush(struct file *filp);
537 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
538 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
541 pfm_put_task(struct task_struct *task)
543 if (task != current) put_task_struct(task);
547 pfm_set_task_notify(struct task_struct *task)
549 struct thread_info *info;
551 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
552 set_bit(TIF_NOTIFY_RESUME, &info->flags);
556 pfm_clear_task_notify(void)
558 clear_thread_flag(TIF_NOTIFY_RESUME);
562 pfm_reserve_page(unsigned long a)
564 SetPageReserved(vmalloc_to_page((void *)a));
567 pfm_unreserve_page(unsigned long a)
569 ClearPageReserved(vmalloc_to_page((void*)a));
572 static inline unsigned long
573 pfm_protect_ctx_ctxsw(pfm_context_t *x)
575 spin_lock(&(x)->ctx_lock);
579 static inline unsigned long
580 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
582 spin_unlock(&(x)->ctx_lock);
585 static inline unsigned int
586 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
588 return do_munmap(mm, addr, len);
591 static inline unsigned long
592 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
594 return get_unmapped_area(file, addr, len, pgoff, flags);
598 static struct super_block *
599 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
601 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
604 static struct file_system_type pfm_fs_type = {
606 .get_sb = pfmfs_get_sb,
607 .kill_sb = kill_anon_super,
610 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
611 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
612 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
613 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
614 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
617 /* forward declaration */
618 static struct file_operations pfm_file_ops;
621 * forward declarations
624 static void pfm_lazy_save_regs (struct task_struct *ta);
627 void dump_pmu_state(const char *);
628 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
630 #include "perfmon_itanium.h"
631 #include "perfmon_mckinley.h"
632 #include "perfmon_generic.h"
634 static pmu_config_t *pmu_confs[]={
637 &pmu_conf_gen, /* must be last */
642 static int pfm_end_notify_user(pfm_context_t *ctx);
645 pfm_clear_psr_pp(void)
647 ia64_rsm(IA64_PSR_PP);
654 ia64_ssm(IA64_PSR_PP);
659 pfm_clear_psr_up(void)
661 ia64_rsm(IA64_PSR_UP);
668 ia64_ssm(IA64_PSR_UP);
672 static inline unsigned long
676 tmp = ia64_getreg(_IA64_REG_PSR);
682 pfm_set_psr_l(unsigned long val)
684 ia64_setreg(_IA64_REG_PSR_L, val);
696 pfm_unfreeze_pmu(void)
703 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
707 for (i=0; i < nibrs; i++) {
708 ia64_set_ibr(i, ibrs[i]);
709 ia64_dv_serialize_instruction();
715 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
719 for (i=0; i < ndbrs; i++) {
720 ia64_set_dbr(i, dbrs[i]);
721 ia64_dv_serialize_data();
727 * PMD[i] must be a counter. no check is made
729 static inline unsigned long
730 pfm_read_soft_counter(pfm_context_t *ctx, int i)
732 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
736 * PMD[i] must be a counter. no check is made
739 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
741 unsigned long ovfl_val = pmu_conf->ovfl_val;
743 ctx->ctx_pmds[i].val = val & ~ovfl_val;
745 * writing to unimplemented part is ignore, so we do not need to
748 ia64_set_pmd(i, val & ovfl_val);
752 pfm_get_new_msg(pfm_context_t *ctx)
756 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
758 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
759 if (next == ctx->ctx_msgq_head) return NULL;
761 idx = ctx->ctx_msgq_tail;
762 ctx->ctx_msgq_tail = next;
764 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
766 return ctx->ctx_msgq+idx;
770 pfm_get_next_msg(pfm_context_t *ctx)
774 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
776 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
781 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
786 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
788 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));
794 pfm_reset_msgq(pfm_context_t *ctx)
796 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
797 DPRINT(("ctx=%p msgq reset\n", ctx));
801 pfm_rvmalloc(unsigned long size)
806 size = PAGE_ALIGN(size);
809 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
810 memset(mem, 0, size);
811 addr = (unsigned long)mem;
813 pfm_reserve_page(addr);
822 pfm_rvfree(void *mem, unsigned long size)
827 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
828 addr = (unsigned long) mem;
829 while ((long) size > 0) {
830 pfm_unreserve_page(addr);
839 static pfm_context_t *
840 pfm_context_alloc(void)
845 * allocate context descriptor
846 * must be able to free with interrupts disabled
848 ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
850 memset(ctx, 0, sizeof(pfm_context_t));
851 DPRINT(("alloc ctx @%p\n", ctx));
857 pfm_context_free(pfm_context_t *ctx)
860 DPRINT(("free ctx @%p\n", ctx));
866 pfm_mask_monitoring(struct task_struct *task)
868 pfm_context_t *ctx = PFM_GET_CTX(task);
869 struct thread_struct *th = &task->thread;
870 unsigned long mask, val, ovfl_mask;
873 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
875 ovfl_mask = pmu_conf->ovfl_val;
877 * monitoring can only be masked as a result of a valid
878 * counter overflow. In UP, it means that the PMU still
879 * has an owner. Note that the owner can be different
880 * from the current task. However the PMU state belongs
882 * In SMP, a valid overflow only happens when task is
883 * current. Therefore if we come here, we know that
884 * the PMU state belongs to the current task, therefore
885 * we can access the live registers.
887 * So in both cases, the live register contains the owner's
888 * state. We can ONLY touch the PMU registers and NOT the PSR.
890 * As a consequence to this call, the thread->pmds[] array
891 * contains stale information which must be ignored
892 * when context is reloaded AND monitoring is active (see
895 mask = ctx->ctx_used_pmds[0];
896 for (i = 0; mask; i++, mask>>=1) {
897 /* skip non used pmds */
898 if ((mask & 0x1) == 0) continue;
899 val = ia64_get_pmd(i);
901 if (PMD_IS_COUNTING(i)) {
903 * we rebuild the full 64 bit value of the counter
905 ctx->ctx_pmds[i].val += (val & ovfl_mask);
907 ctx->ctx_pmds[i].val = val;
909 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
911 ctx->ctx_pmds[i].val,
915 * mask monitoring by setting the privilege level to 0
916 * we cannot use psr.pp/psr.up for this, it is controlled by
919 * if task is current, modify actual registers, otherwise modify
920 * thread save state, i.e., what will be restored in pfm_load_regs()
922 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
923 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
924 if ((mask & 0x1) == 0UL) continue;
925 ia64_set_pmc(i, th->pmcs[i] & ~0xfUL);
926 th->pmcs[i] &= ~0xfUL;
927 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i]));
930 * make all of this visible
936 * must always be done with task == current
938 * context must be in MASKED state when calling
941 pfm_restore_monitoring(struct task_struct *task)
943 pfm_context_t *ctx = PFM_GET_CTX(task);
944 struct thread_struct *th = &task->thread;
945 unsigned long mask, ovfl_mask;
946 unsigned long psr, val;
949 is_system = ctx->ctx_fl_system;
950 ovfl_mask = pmu_conf->ovfl_val;
952 if (task != current) {
953 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
956 if (ctx->ctx_state != PFM_CTX_MASKED) {
957 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
958 task->pid, current->pid, ctx->ctx_state);
963 * monitoring is masked via the PMC.
964 * As we restore their value, we do not want each counter to
965 * restart right away. We stop monitoring using the PSR,
966 * restore the PMC (and PMD) and then re-establish the psr
967 * as it was. Note that there can be no pending overflow at
968 * this point, because monitoring was MASKED.
970 * system-wide session are pinned and self-monitoring
972 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
974 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
980 * first, we restore the PMD
982 mask = ctx->ctx_used_pmds[0];
983 for (i = 0; mask; i++, mask>>=1) {
984 /* skip non used pmds */
985 if ((mask & 0x1) == 0) continue;
987 if (PMD_IS_COUNTING(i)) {
989 * we split the 64bit value according to
992 val = ctx->ctx_pmds[i].val & ovfl_mask;
993 ctx->ctx_pmds[i].val &= ~ovfl_mask;
995 val = ctx->ctx_pmds[i].val;
997 ia64_set_pmd(i, val);
999 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1001 ctx->ctx_pmds[i].val,
1007 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1008 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1009 if ((mask & 0x1) == 0UL) continue;
1010 th->pmcs[i] = ctx->ctx_pmcs[i];
1011 ia64_set_pmc(i, th->pmcs[i]);
1012 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
1017 * must restore DBR/IBR because could be modified while masked
1018 * XXX: need to optimize
1020 if (ctx->ctx_fl_using_dbreg) {
1021 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1022 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1028 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1030 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1037 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1043 for (i=0; mask; i++, mask>>=1) {
1044 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1049 * reload from thread state (used for ctxw only)
1052 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1055 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1057 for (i=0; mask; i++, mask>>=1) {
1058 if ((mask & 0x1) == 0) continue;
1059 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1060 ia64_set_pmd(i, val);
1066 * propagate PMD from context to thread-state
1069 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1071 struct thread_struct *thread = &task->thread;
1072 unsigned long ovfl_val = pmu_conf->ovfl_val;
1073 unsigned long mask = ctx->ctx_all_pmds[0];
1077 DPRINT(("mask=0x%lx\n", mask));
1079 for (i=0; mask; i++, mask>>=1) {
1081 val = ctx->ctx_pmds[i].val;
1084 * We break up the 64 bit value into 2 pieces
1085 * the lower bits go to the machine state in the
1086 * thread (will be reloaded on ctxsw in).
1087 * The upper part stays in the soft-counter.
1089 if (PMD_IS_COUNTING(i)) {
1090 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1093 thread->pmds[i] = val;
1095 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1098 ctx->ctx_pmds[i].val));
1103 * propagate PMC from context to thread-state
1106 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1108 struct thread_struct *thread = &task->thread;
1109 unsigned long mask = ctx->ctx_all_pmcs[0];
1112 DPRINT(("mask=0x%lx\n", mask));
1114 for (i=0; mask; i++, mask>>=1) {
1115 /* masking 0 with ovfl_val yields 0 */
1116 thread->pmcs[i] = ctx->ctx_pmcs[i];
1117 DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
1124 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1128 for (i=0; mask; i++, mask>>=1) {
1129 if ((mask & 0x1) == 0) continue;
1130 ia64_set_pmc(i, pmcs[i]);
1136 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1138 return memcmp(a, b, sizeof(pfm_uuid_t));
1142 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1145 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1150 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1153 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1159 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1163 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1168 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1172 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1177 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1180 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1185 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)
1188 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1192 static pfm_buffer_fmt_t *
1193 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1195 struct list_head * pos;
1196 pfm_buffer_fmt_t * entry;
1198 list_for_each(pos, &pfm_buffer_fmt_list) {
1199 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1200 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1207 * find a buffer format based on its uuid
1209 static pfm_buffer_fmt_t *
1210 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1212 pfm_buffer_fmt_t * fmt;
1213 spin_lock(&pfm_buffer_fmt_lock);
1214 fmt = __pfm_find_buffer_fmt(uuid);
1215 spin_unlock(&pfm_buffer_fmt_lock);
1220 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1224 /* some sanity checks */
1225 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1227 /* we need at least a handler */
1228 if (fmt->fmt_handler == NULL) return -EINVAL;
1231 * XXX: need check validity of fmt_arg_size
1234 spin_lock(&pfm_buffer_fmt_lock);
1236 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1237 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1241 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1242 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1245 spin_unlock(&pfm_buffer_fmt_lock);
1248 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1251 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1253 pfm_buffer_fmt_t *fmt;
1256 spin_lock(&pfm_buffer_fmt_lock);
1258 fmt = __pfm_find_buffer_fmt(uuid);
1260 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1264 list_del_init(&fmt->fmt_list);
1265 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1268 spin_unlock(&pfm_buffer_fmt_lock);
1272 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1274 extern void update_pal_halt_status(int);
1277 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1279 unsigned long flags;
1281 * validy checks on cpu_mask have been done upstream
1285 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1286 pfm_sessions.pfs_sys_sessions,
1287 pfm_sessions.pfs_task_sessions,
1288 pfm_sessions.pfs_sys_use_dbregs,
1294 * cannot mix system wide and per-task sessions
1296 if (pfm_sessions.pfs_task_sessions > 0UL) {
1297 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1298 pfm_sessions.pfs_task_sessions));
1302 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1304 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1306 pfm_sessions.pfs_sys_session[cpu] = task;
1308 pfm_sessions.pfs_sys_sessions++ ;
1311 if (pfm_sessions.pfs_sys_sessions) goto abort;
1312 pfm_sessions.pfs_task_sessions++;
1315 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1316 pfm_sessions.pfs_sys_sessions,
1317 pfm_sessions.pfs_task_sessions,
1318 pfm_sessions.pfs_sys_use_dbregs,
1323 * disable default_idle() to go to PAL_HALT
1325 update_pal_halt_status(0);
1332 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1333 pfm_sessions.pfs_sys_session[cpu]->pid,
1343 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1345 unsigned long flags;
1347 * validy checks on cpu_mask have been done upstream
1351 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1352 pfm_sessions.pfs_sys_sessions,
1353 pfm_sessions.pfs_task_sessions,
1354 pfm_sessions.pfs_sys_use_dbregs,
1360 pfm_sessions.pfs_sys_session[cpu] = NULL;
1362 * would not work with perfmon+more than one bit in cpu_mask
1364 if (ctx && ctx->ctx_fl_using_dbreg) {
1365 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1366 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1368 pfm_sessions.pfs_sys_use_dbregs--;
1371 pfm_sessions.pfs_sys_sessions--;
1373 pfm_sessions.pfs_task_sessions--;
1375 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1376 pfm_sessions.pfs_sys_sessions,
1377 pfm_sessions.pfs_task_sessions,
1378 pfm_sessions.pfs_sys_use_dbregs,
1383 * if possible, enable default_idle() to go into PAL_HALT
1385 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1386 update_pal_halt_status(1);
1394 * removes virtual mapping of the sampling buffer.
1395 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1396 * a PROTECT_CTX() section.
1399 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1404 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1405 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1409 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1412 * does the actual unmapping
1414 down_write(&task->mm->mmap_sem);
1416 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1418 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1420 up_write(&task->mm->mmap_sem);
1422 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1425 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1431 * free actual physical storage used by sampling buffer
1435 pfm_free_smpl_buffer(pfm_context_t *ctx)
1437 pfm_buffer_fmt_t *fmt;
1439 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1442 * we won't use the buffer format anymore
1444 fmt = ctx->ctx_buf_fmt;
1446 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1449 ctx->ctx_smpl_vaddr));
1451 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1456 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1458 ctx->ctx_smpl_hdr = NULL;
1459 ctx->ctx_smpl_size = 0UL;
1464 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1470 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1472 if (fmt == NULL) return;
1474 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1479 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1480 * no real gain from having the whole whorehouse mounted. So we don't need
1481 * any operations on the root directory. However, we need a non-trivial
1482 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1484 static struct vfsmount *pfmfs_mnt;
1489 int err = register_filesystem(&pfm_fs_type);
1491 pfmfs_mnt = kern_mount(&pfm_fs_type);
1492 err = PTR_ERR(pfmfs_mnt);
1493 if (IS_ERR(pfmfs_mnt))
1494 unregister_filesystem(&pfm_fs_type);
1504 unregister_filesystem(&pfm_fs_type);
1509 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1514 unsigned long flags;
1515 DECLARE_WAITQUEUE(wait, current);
1516 if (PFM_IS_FILE(filp) == 0) {
1517 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1521 ctx = (pfm_context_t *)filp->private_data;
1523 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1528 * check even when there is no message
1530 if (size < sizeof(pfm_msg_t)) {
1531 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1535 PROTECT_CTX(ctx, flags);
1538 * put ourselves on the wait queue
1540 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1548 set_current_state(TASK_INTERRUPTIBLE);
1550 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1553 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1555 UNPROTECT_CTX(ctx, flags);
1558 * check non-blocking read
1561 if(filp->f_flags & O_NONBLOCK) break;
1564 * check pending signals
1566 if(signal_pending(current)) {
1571 * no message, so wait
1575 PROTECT_CTX(ctx, flags);
1577 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1578 set_current_state(TASK_RUNNING);
1579 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1581 if (ret < 0) goto abort;
1584 msg = pfm_get_next_msg(ctx);
1586 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1590 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1593 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1596 UNPROTECT_CTX(ctx, flags);
1602 pfm_write(struct file *file, const char __user *ubuf,
1603 size_t size, loff_t *ppos)
1605 DPRINT(("pfm_write called\n"));
1610 pfm_poll(struct file *filp, poll_table * wait)
1613 unsigned long flags;
1614 unsigned int mask = 0;
1616 if (PFM_IS_FILE(filp) == 0) {
1617 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1621 ctx = (pfm_context_t *)filp->private_data;
1623 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1628 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1630 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1632 PROTECT_CTX(ctx, flags);
1634 if (PFM_CTXQ_EMPTY(ctx) == 0)
1635 mask = POLLIN | POLLRDNORM;
1637 UNPROTECT_CTX(ctx, flags);
1639 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1645 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1647 DPRINT(("pfm_ioctl called\n"));
1652 * interrupt cannot be masked when coming here
1655 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1659 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1661 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1665 ctx->ctx_async_queue, ret));
1671 pfm_fasync(int fd, struct file *filp, int on)
1676 if (PFM_IS_FILE(filp) == 0) {
1677 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1681 ctx = (pfm_context_t *)filp->private_data;
1683 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1687 * we cannot mask interrupts during this call because this may
1688 * may go to sleep if memory is not readily avalaible.
1690 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1691 * done in caller. Serialization of this function is ensured by caller.
1693 ret = pfm_do_fasync(fd, filp, ctx, on);
1696 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1699 ctx->ctx_async_queue, ret));
1706 * this function is exclusively called from pfm_close().
1707 * The context is not protected at that time, nor are interrupts
1708 * on the remote CPU. That's necessary to avoid deadlocks.
1711 pfm_syswide_force_stop(void *info)
1713 pfm_context_t *ctx = (pfm_context_t *)info;
1714 struct pt_regs *regs = ia64_task_regs(current);
1715 struct task_struct *owner;
1716 unsigned long flags;
1719 if (ctx->ctx_cpu != smp_processor_id()) {
1720 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1722 smp_processor_id());
1725 owner = GET_PMU_OWNER();
1726 if (owner != ctx->ctx_task) {
1727 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1729 owner->pid, ctx->ctx_task->pid);
1732 if (GET_PMU_CTX() != ctx) {
1733 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1735 GET_PMU_CTX(), ctx);
1739 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1741 * the context is already protected in pfm_close(), we simply
1742 * need to mask interrupts to avoid a PMU interrupt race on
1745 local_irq_save(flags);
1747 ret = pfm_context_unload(ctx, NULL, 0, regs);
1749 DPRINT(("context_unload returned %d\n", ret));
1753 * unmask interrupts, PMU interrupts are now spurious here
1755 local_irq_restore(flags);
1759 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1763 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1764 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1765 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1767 #endif /* CONFIG_SMP */
1770 * called for each close(). Partially free resources.
1771 * When caller is self-monitoring, the context is unloaded.
1774 pfm_flush(struct file *filp)
1777 struct task_struct *task;
1778 struct pt_regs *regs;
1779 unsigned long flags;
1780 unsigned long smpl_buf_size = 0UL;
1781 void *smpl_buf_vaddr = NULL;
1782 int state, is_system;
1784 if (PFM_IS_FILE(filp) == 0) {
1785 DPRINT(("bad magic for\n"));
1789 ctx = (pfm_context_t *)filp->private_data;
1791 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1796 * remove our file from the async queue, if we use this mode.
1797 * This can be done without the context being protected. We come
1798 * here when the context has become unreacheable by other tasks.
1800 * We may still have active monitoring at this point and we may
1801 * end up in pfm_overflow_handler(). However, fasync_helper()
1802 * operates with interrupts disabled and it cleans up the
1803 * queue. If the PMU handler is called prior to entering
1804 * fasync_helper() then it will send a signal. If it is
1805 * invoked after, it will find an empty queue and no
1806 * signal will be sent. In both case, we are safe
1808 if (filp->f_flags & FASYNC) {
1809 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1810 pfm_do_fasync (-1, filp, ctx, 0);
1813 PROTECT_CTX(ctx, flags);
1815 state = ctx->ctx_state;
1816 is_system = ctx->ctx_fl_system;
1818 task = PFM_CTX_TASK(ctx);
1819 regs = ia64_task_regs(task);
1821 DPRINT(("ctx_state=%d is_current=%d\n",
1823 task == current ? 1 : 0));
1826 * if state == UNLOADED, then task is NULL
1830 * we must stop and unload because we are losing access to the context.
1832 if (task == current) {
1835 * the task IS the owner but it migrated to another CPU: that's bad
1836 * but we must handle this cleanly. Unfortunately, the kernel does
1837 * not provide a mechanism to block migration (while the context is loaded).
1839 * We need to release the resource on the ORIGINAL cpu.
1841 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1843 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1845 * keep context protected but unmask interrupt for IPI
1847 local_irq_restore(flags);
1849 pfm_syswide_cleanup_other_cpu(ctx);
1852 * restore interrupt masking
1854 local_irq_save(flags);
1857 * context is unloaded at this point
1860 #endif /* CONFIG_SMP */
1863 DPRINT(("forcing unload\n"));
1865 * stop and unload, returning with state UNLOADED
1866 * and session unreserved.
1868 pfm_context_unload(ctx, NULL, 0, regs);
1870 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1875 * remove virtual mapping, if any, for the calling task.
1876 * cannot reset ctx field until last user is calling close().
1878 * ctx_smpl_vaddr must never be cleared because it is needed
1879 * by every task with access to the context
1881 * When called from do_exit(), the mm context is gone already, therefore
1882 * mm is NULL, i.e., the VMA is already gone and we do not have to
1885 if (ctx->ctx_smpl_vaddr && current->mm) {
1886 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1887 smpl_buf_size = ctx->ctx_smpl_size;
1890 UNPROTECT_CTX(ctx, flags);
1893 * if there was a mapping, then we systematically remove it
1894 * at this point. Cannot be done inside critical section
1895 * because some VM function reenables interrupts.
1898 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1903 * called either on explicit close() or from exit_files().
1904 * Only the LAST user of the file gets to this point, i.e., it is
1907 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1908 * (fput()),i.e, last task to access the file. Nobody else can access the
1909 * file at this point.
1911 * When called from exit_files(), the VMA has been freed because exit_mm()
1912 * is executed before exit_files().
1914 * When called from exit_files(), the current task is not yet ZOMBIE but we
1915 * flush the PMU state to the context.
1918 pfm_close(struct inode *inode, struct file *filp)
1921 struct task_struct *task;
1922 struct pt_regs *regs;
1923 DECLARE_WAITQUEUE(wait, current);
1924 unsigned long flags;
1925 unsigned long smpl_buf_size = 0UL;
1926 void *smpl_buf_addr = NULL;
1927 int free_possible = 1;
1928 int state, is_system;
1930 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1932 if (PFM_IS_FILE(filp) == 0) {
1933 DPRINT(("bad magic\n"));
1937 ctx = (pfm_context_t *)filp->private_data;
1939 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1943 PROTECT_CTX(ctx, flags);
1945 state = ctx->ctx_state;
1946 is_system = ctx->ctx_fl_system;
1948 task = PFM_CTX_TASK(ctx);
1949 regs = ia64_task_regs(task);
1951 DPRINT(("ctx_state=%d is_current=%d\n",
1953 task == current ? 1 : 0));
1956 * if task == current, then pfm_flush() unloaded the context
1958 if (state == PFM_CTX_UNLOADED) goto doit;
1961 * context is loaded/masked and task != current, we need to
1962 * either force an unload or go zombie
1966 * The task is currently blocked or will block after an overflow.
1967 * we must force it to wakeup to get out of the
1968 * MASKED state and transition to the unloaded state by itself.
1970 * This situation is only possible for per-task mode
1972 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1975 * set a "partial" zombie state to be checked
1976 * upon return from down() in pfm_handle_work().
1978 * We cannot use the ZOMBIE state, because it is checked
1979 * by pfm_load_regs() which is called upon wakeup from down().
1980 * In such case, it would free the context and then we would
1981 * return to pfm_handle_work() which would access the
1982 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1983 * but visible to pfm_handle_work().
1985 * For some window of time, we have a zombie context with
1986 * ctx_state = MASKED and not ZOMBIE
1988 ctx->ctx_fl_going_zombie = 1;
1991 * force task to wake up from MASKED state
1993 up(&ctx->ctx_restart_sem);
1995 DPRINT(("waking up ctx_state=%d\n", state));
1998 * put ourself to sleep waiting for the other
1999 * task to report completion
2001 * the context is protected by mutex, therefore there
2002 * is no risk of being notified of completion before
2003 * begin actually on the waitq.
2005 set_current_state(TASK_INTERRUPTIBLE);
2006 add_wait_queue(&ctx->ctx_zombieq, &wait);
2008 UNPROTECT_CTX(ctx, flags);
2011 * XXX: check for signals :
2012 * - ok for explicit close
2013 * - not ok when coming from exit_files()
2018 PROTECT_CTX(ctx, flags);
2021 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2022 set_current_state(TASK_RUNNING);
2025 * context is unloaded at this point
2027 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2029 else if (task != current) {
2032 * switch context to zombie state
2034 ctx->ctx_state = PFM_CTX_ZOMBIE;
2036 DPRINT(("zombie ctx for [%d]\n", task->pid));
2038 * cannot free the context on the spot. deferred until
2039 * the task notices the ZOMBIE state
2043 pfm_context_unload(ctx, NULL, 0, regs);
2048 /* reload state, may have changed during opening of critical section */
2049 state = ctx->ctx_state;
2052 * the context is still attached to a task (possibly current)
2053 * we cannot destroy it right now
2057 * we must free the sampling buffer right here because
2058 * we cannot rely on it being cleaned up later by the
2059 * monitored task. It is not possible to free vmalloc'ed
2060 * memory in pfm_load_regs(). Instead, we remove the buffer
2061 * now. should there be subsequent PMU overflow originally
2062 * meant for sampling, the will be converted to spurious
2063 * and that's fine because the monitoring tools is gone anyway.
2065 if (ctx->ctx_smpl_hdr) {
2066 smpl_buf_addr = ctx->ctx_smpl_hdr;
2067 smpl_buf_size = ctx->ctx_smpl_size;
2068 /* no more sampling */
2069 ctx->ctx_smpl_hdr = NULL;
2070 ctx->ctx_fl_is_sampling = 0;
2073 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2079 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2082 * UNLOADED that the session has already been unreserved.
2084 if (state == PFM_CTX_ZOMBIE) {
2085 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2089 * disconnect file descriptor from context must be done
2092 filp->private_data = NULL;
2095 * if we free on the spot, the context is now completely unreacheable
2096 * from the callers side. The monitored task side is also cut, so we
2099 * If we have a deferred free, only the caller side is disconnected.
2101 UNPROTECT_CTX(ctx, flags);
2104 * All memory free operations (especially for vmalloc'ed memory)
2105 * MUST be done with interrupts ENABLED.
2107 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2110 * return the memory used by the context
2112 if (free_possible) pfm_context_free(ctx);
2118 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2120 DPRINT(("pfm_no_open called\n"));
2126 static struct file_operations pfm_file_ops = {
2127 .llseek = no_llseek,
2132 .open = pfm_no_open, /* special open code to disallow open via /proc */
2133 .fasync = pfm_fasync,
2134 .release = pfm_close,
2139 pfmfs_delete_dentry(struct dentry *dentry)
2144 static struct dentry_operations pfmfs_dentry_operations = {
2145 .d_delete = pfmfs_delete_dentry,
2150 pfm_alloc_fd(struct file **cfile)
2153 struct file *file = NULL;
2154 struct inode * inode;
2158 fd = get_unused_fd();
2159 if (fd < 0) return -ENFILE;
2163 file = get_empty_filp();
2164 if (!file) goto out;
2167 * allocate a new inode
2169 inode = new_inode(pfmfs_mnt->mnt_sb);
2170 if (!inode) goto out;
2172 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2174 inode->i_mode = S_IFCHR|S_IRUGO;
2175 inode->i_uid = current->fsuid;
2176 inode->i_gid = current->fsgid;
2178 sprintf(name, "[%lu]", inode->i_ino);
2180 this.len = strlen(name);
2181 this.hash = inode->i_ino;
2186 * allocate a new dcache entry
2188 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2189 if (!file->f_dentry) goto out;
2191 file->f_dentry->d_op = &pfmfs_dentry_operations;
2193 d_add(file->f_dentry, inode);
2194 file->f_vfsmnt = mntget(pfmfs_mnt);
2195 file->f_mapping = inode->i_mapping;
2197 file->f_op = &pfm_file_ops;
2198 file->f_mode = FMODE_READ;
2199 file->f_flags = O_RDONLY;
2203 * may have to delay until context is attached?
2205 fd_install(fd, file);
2208 * the file structure we will use
2214 if (file) put_filp(file);
2220 pfm_free_fd(int fd, struct file *file)
2222 struct files_struct *files = current->files;
2225 * there ie no fd_uninstall(), so we do it here
2227 spin_lock(&files->file_lock);
2228 files->fd[fd] = NULL;
2229 spin_unlock(&files->file_lock);
2231 if (file) put_filp(file);
2236 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2238 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2241 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2244 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2255 * allocate a sampling buffer and remaps it into the user address space of the task
2258 pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2260 struct mm_struct *mm = task->mm;
2261 struct vm_area_struct *vma = NULL;
2267 * the fixed header + requested size and align to page boundary
2269 size = PAGE_ALIGN(rsize);
2271 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2274 * check requested size to avoid Denial-of-service attacks
2275 * XXX: may have to refine this test
2276 * Check against address space limit.
2278 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2281 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2285 * We do the easy to undo allocations first.
2287 * pfm_rvmalloc(), clears the buffer, so there is no leak
2289 smpl_buf = pfm_rvmalloc(size);
2290 if (smpl_buf == NULL) {
2291 DPRINT(("Can't allocate sampling buffer\n"));
2295 DPRINT(("smpl_buf @%p\n", smpl_buf));
2298 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2300 DPRINT(("Cannot allocate vma\n"));
2303 memset(vma, 0, sizeof(*vma));
2306 * partially initialize the vma for the sampling buffer
2309 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2310 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2313 * Now we have everything we need and we can initialize
2314 * and connect all the data structures
2317 ctx->ctx_smpl_hdr = smpl_buf;
2318 ctx->ctx_smpl_size = size; /* aligned size */
2321 * Let's do the difficult operations next.
2323 * now we atomically find some area in the address space and
2324 * remap the buffer in it.
2326 down_write(&task->mm->mmap_sem);
2328 /* find some free area in address space, must have mmap sem held */
2329 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2330 if (vma->vm_start == 0UL) {
2331 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2332 up_write(&task->mm->mmap_sem);
2335 vma->vm_end = vma->vm_start + size;
2336 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2338 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2340 /* can only be applied to current task, need to have the mm semaphore held when called */
2341 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2342 DPRINT(("Can't remap buffer\n"));
2343 up_write(&task->mm->mmap_sem);
2348 * now insert the vma in the vm list for the process, must be
2349 * done with mmap lock held
2351 insert_vm_struct(mm, vma);
2353 vx_vmpages_add(mm, size >> PAGE_SHIFT);
2354 vm_stat_account(vma);
2355 up_write(&task->mm->mmap_sem);
2358 * keep track of user level virtual address
2360 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2361 *(unsigned long *)user_vaddr = vma->vm_start;
2366 kmem_cache_free(vm_area_cachep, vma);
2368 pfm_rvfree(smpl_buf, size);
2374 * XXX: do something better here
2377 pfm_bad_permissions(struct task_struct *task)
2379 /* inspired by ptrace_attach() */
2380 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2389 return ((current->uid != task->euid)
2390 || (current->uid != task->suid)
2391 || (current->uid != task->uid)
2392 || (current->gid != task->egid)
2393 || (current->gid != task->sgid)
2394 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2398 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2404 ctx_flags = pfx->ctx_flags;
2406 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2409 * cannot block in this mode
2411 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2412 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2417 /* probably more to add here */
2423 pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2424 unsigned int cpu, pfarg_context_t *arg)
2426 pfm_buffer_fmt_t *fmt = NULL;
2427 unsigned long size = 0UL;
2429 void *fmt_arg = NULL;
2431 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2433 /* invoke and lock buffer format, if found */
2434 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2436 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2441 * buffer argument MUST be contiguous to pfarg_context_t
2443 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2445 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2447 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2449 if (ret) goto error;
2451 /* link buffer format and context */
2452 ctx->ctx_buf_fmt = fmt;
2455 * check if buffer format wants to use perfmon buffer allocation/mapping service
2457 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2458 if (ret) goto error;
2462 * buffer is always remapped into the caller's address space
2464 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2465 if (ret) goto error;
2467 /* keep track of user address of buffer */
2468 arg->ctx_smpl_vaddr = uaddr;
2470 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2477 pfm_reset_pmu_state(pfm_context_t *ctx)
2482 * install reset values for PMC.
2484 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2485 if (PMC_IS_IMPL(i) == 0) continue;
2486 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2487 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2490 * PMD registers are set to 0UL when the context in memset()
2494 * On context switched restore, we must restore ALL pmc and ALL pmd even
2495 * when they are not actively used by the task. In UP, the incoming process
2496 * may otherwise pick up left over PMC, PMD state from the previous process.
2497 * As opposed to PMD, stale PMC can cause harm to the incoming
2498 * process because they may change what is being measured.
2499 * Therefore, we must systematically reinstall the entire
2500 * PMC state. In SMP, the same thing is possible on the
2501 * same CPU but also on between 2 CPUs.
2503 * The problem with PMD is information leaking especially
2504 * to user level when psr.sp=0
2506 * There is unfortunately no easy way to avoid this problem
2507 * on either UP or SMP. This definitively slows down the
2508 * pfm_load_regs() function.
2512 * bitmask of all PMCs accessible to this context
2514 * PMC0 is treated differently.
2516 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2519 * bitmask of all PMDs that are accesible to this context
2521 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2523 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2526 * useful in case of re-enable after disable
2528 ctx->ctx_used_ibrs[0] = 0UL;
2529 ctx->ctx_used_dbrs[0] = 0UL;
2533 pfm_ctx_getsize(void *arg, size_t *sz)
2535 pfarg_context_t *req = (pfarg_context_t *)arg;
2536 pfm_buffer_fmt_t *fmt;
2540 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2542 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2544 DPRINT(("cannot find buffer format\n"));
2547 /* get just enough to copy in user parameters */
2548 *sz = fmt->fmt_arg_size;
2549 DPRINT(("arg_size=%lu\n", *sz));
2557 * cannot attach if :
2559 * - task not owned by caller
2560 * - task incompatible with context mode
2563 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2566 * no kernel task or task not owner by caller
2568 if (task->mm == NULL) {
2569 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2572 if (pfm_bad_permissions(task)) {
2573 DPRINT(("no permission to attach to [%d]\n", task->pid));
2577 * cannot block in self-monitoring mode
2579 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2580 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2584 if (task->exit_state == EXIT_ZOMBIE) {
2585 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2590 * always ok for self
2592 if (task == current) return 0;
2594 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
2595 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2599 * make sure the task is off any CPU
2601 wait_task_inactive(task);
2603 /* more to come... */
2609 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2611 struct task_struct *p = current;
2614 /* XXX: need to add more checks here */
2615 if (pid < 2) return -EPERM;
2617 if (pid != current->pid) {
2619 read_lock(&tasklist_lock);
2621 p = find_task_by_pid(pid);
2623 /* make sure task cannot go away while we operate on it */
2624 if (p) get_task_struct(p);
2626 read_unlock(&tasklist_lock);
2628 if (p == NULL) return -ESRCH;
2631 ret = pfm_task_incompatible(ctx, p);
2634 } else if (p != current) {
2643 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2645 pfarg_context_t *req = (pfarg_context_t *)arg;
2650 /* let's check the arguments first */
2651 ret = pfarg_is_sane(current, req);
2652 if (ret < 0) return ret;
2654 ctx_flags = req->ctx_flags;
2658 ctx = pfm_context_alloc();
2659 if (!ctx) goto error;
2661 ret = pfm_alloc_fd(&filp);
2662 if (ret < 0) goto error_file;
2664 req->ctx_fd = ctx->ctx_fd = ret;
2667 * attach context to file
2669 filp->private_data = ctx;
2672 * does the user want to sample?
2674 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2675 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2676 if (ret) goto buffer_error;
2680 * init context protection lock
2682 spin_lock_init(&ctx->ctx_lock);
2685 * context is unloaded
2687 ctx->ctx_state = PFM_CTX_UNLOADED;
2690 * initialization of context's flags
2692 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2693 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2694 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2695 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2697 * will move to set properties
2698 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2702 * init restart semaphore to locked
2704 sema_init(&ctx->ctx_restart_sem, 0);
2707 * activation is used in SMP only
2709 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2710 SET_LAST_CPU(ctx, -1);
2713 * initialize notification message queue
2715 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2716 init_waitqueue_head(&ctx->ctx_msgq_wait);
2717 init_waitqueue_head(&ctx->ctx_zombieq);
2719 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2724 ctx->ctx_fl_excl_idle,
2729 * initialize soft PMU state
2731 pfm_reset_pmu_state(ctx);
2736 pfm_free_fd(ctx->ctx_fd, filp);
2738 if (ctx->ctx_buf_fmt) {
2739 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2742 pfm_context_free(ctx);
2748 static inline unsigned long
2749 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2751 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2752 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2753 extern unsigned long carta_random32 (unsigned long seed);
2755 if (reg->flags & PFM_REGFL_RANDOM) {
2756 new_seed = carta_random32(old_seed);
2757 val -= (old_seed & mask); /* counter values are negative numbers! */
2758 if ((mask >> 32) != 0)
2759 /* construct a full 64-bit random value: */
2760 new_seed |= carta_random32(old_seed >> 32) << 32;
2761 reg->seed = new_seed;
2768 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2770 unsigned long mask = ovfl_regs[0];
2771 unsigned long reset_others = 0UL;
2776 * now restore reset value on sampling overflowed counters
2778 mask >>= PMU_FIRST_COUNTER;
2779 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2781 if ((mask & 0x1UL) == 0UL) continue;
2783 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2784 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2786 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2790 * Now take care of resetting the other registers
2792 for(i = 0; reset_others; i++, reset_others >>= 1) {
2794 if ((reset_others & 0x1) == 0) continue;
2796 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2798 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2799 is_long_reset ? "long" : "short", i, val));
2804 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2806 unsigned long mask = ovfl_regs[0];
2807 unsigned long reset_others = 0UL;
2811 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2813 if (ctx->ctx_state == PFM_CTX_MASKED) {
2814 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2819 * now restore reset value on sampling overflowed counters
2821 mask >>= PMU_FIRST_COUNTER;
2822 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2824 if ((mask & 0x1UL) == 0UL) continue;
2826 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2827 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2829 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2831 pfm_write_soft_counter(ctx, i, val);
2835 * Now take care of resetting the other registers
2837 for(i = 0; reset_others; i++, reset_others >>= 1) {
2839 if ((reset_others & 0x1) == 0) continue;
2841 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2843 if (PMD_IS_COUNTING(i)) {
2844 pfm_write_soft_counter(ctx, i, val);
2846 ia64_set_pmd(i, val);
2848 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2849 is_long_reset ? "long" : "short", i, val));
2855 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2857 struct thread_struct *thread = NULL;
2858 struct task_struct *task;
2859 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2860 unsigned long value, pmc_pm;
2861 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2862 unsigned int cnum, reg_flags, flags, pmc_type;
2863 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2864 int is_monitor, is_counting, state;
2866 pfm_reg_check_t wr_func;
2867 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2869 state = ctx->ctx_state;
2870 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2871 is_system = ctx->ctx_fl_system;
2872 task = ctx->ctx_task;
2873 impl_pmds = pmu_conf->impl_pmds[0];
2875 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2878 thread = &task->thread;
2880 * In system wide and when the context is loaded, access can only happen
2881 * when the caller is running on the CPU being monitored by the session.
2882 * It does not have to be the owner (ctx_task) of the context per se.
2884 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2885 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2888 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2890 expert_mode = pfm_sysctl.expert_mode;
2892 for (i = 0; i < count; i++, req++) {
2894 cnum = req->reg_num;
2895 reg_flags = req->reg_flags;
2896 value = req->reg_value;
2897 smpl_pmds = req->reg_smpl_pmds[0];
2898 reset_pmds = req->reg_reset_pmds[0];
2902 if (cnum >= PMU_MAX_PMCS) {
2903 DPRINT(("pmc%u is invalid\n", cnum));
2907 pmc_type = pmu_conf->pmc_desc[cnum].type;
2908 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2909 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2910 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2913 * we reject all non implemented PMC as well
2914 * as attempts to modify PMC[0-3] which are used
2915 * as status registers by the PMU
2917 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2918 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2921 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2923 * If the PMC is a monitor, then if the value is not the default:
2924 * - system-wide session: PMCx.pm=1 (privileged monitor)
2925 * - per-task : PMCx.pm=0 (user monitor)
2927 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2928 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2937 * enforce generation of overflow interrupt. Necessary on all
2940 value |= 1 << PMU_PMC_OI;
2942 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2943 flags |= PFM_REGFL_OVFL_NOTIFY;
2946 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2948 /* verify validity of smpl_pmds */
2949 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2950 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2954 /* verify validity of reset_pmds */
2955 if ((reset_pmds & impl_pmds) != reset_pmds) {
2956 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2960 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2961 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2964 /* eventid on non-counting monitors are ignored */
2968 * execute write checker, if any
2970 if (likely(expert_mode == 0 && wr_func)) {
2971 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2972 if (ret) goto error;
2977 * no error on this register
2979 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2982 * Now we commit the changes to the software state
2986 * update overflow information
2990 * full flag update each time a register is programmed
2992 ctx->ctx_pmds[cnum].flags = flags;
2994 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2995 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2996 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2999 * Mark all PMDS to be accessed as used.
3001 * We do not keep track of PMC because we have to
3002 * systematically restore ALL of them.
3004 * We do not update the used_monitors mask, because
3005 * if we have not programmed them, then will be in
3006 * a quiescent state, therefore we will not need to
3007 * mask/restore then when context is MASKED.
3009 CTX_USED_PMD(ctx, reset_pmds);
3010 CTX_USED_PMD(ctx, smpl_pmds);
3012 * make sure we do not try to reset on
3013 * restart because we have established new values
3015 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3018 * Needed in case the user does not initialize the equivalent
3019 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3020 * possible leak here.
3022 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3025 * keep track of the monitor PMC that we are using.
3026 * we save the value of the pmc in ctx_pmcs[] and if
3027 * the monitoring is not stopped for the context we also
3028 * place it in the saved state area so that it will be
3029 * picked up later by the context switch code.
3031 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3033 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3034 * monitoring needs to be stopped.
3036 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3039 * update context state
3041 ctx->ctx_pmcs[cnum] = value;
3045 * write thread state
3047 if (is_system == 0) thread->pmcs[cnum] = value;
3050 * write hardware register if we can
3052 if (can_access_pmu) {
3053 ia64_set_pmc(cnum, value);
3058 * per-task SMP only here
3060 * we are guaranteed that the task is not running on the other CPU,
3061 * we indicate that this PMD will need to be reloaded if the task
3062 * is rescheduled on the CPU it ran last on.
3064 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3069 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",
3075 ctx->ctx_all_pmcs[0],
3076 ctx->ctx_used_pmds[0],
3077 ctx->ctx_pmds[cnum].eventid,
3080 ctx->ctx_reload_pmcs[0],
3081 ctx->ctx_used_monitors[0],
3082 ctx->ctx_ovfl_regs[0]));
3086 * make sure the changes are visible
3088 if (can_access_pmu) ia64_srlz_d();
3092 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3097 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3099 struct thread_struct *thread = NULL;
3100 struct task_struct *task;
3101 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3102 unsigned long value, hw_value, ovfl_mask;
3104 int i, can_access_pmu = 0, state;
3105 int is_counting, is_loaded, is_system, expert_mode;
3107 pfm_reg_check_t wr_func;
3110 state = ctx->ctx_state;
3111 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3112 is_system = ctx->ctx_fl_system;
3113 ovfl_mask = pmu_conf->ovfl_val;
3114 task = ctx->ctx_task;
3116 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3119 * on both UP and SMP, we can only write to the PMC when the task is
3120 * the owner of the local PMU.
3122 if (likely(is_loaded)) {
3123 thread = &task->thread;
3125 * In system wide and when the context is loaded, access can only happen
3126 * when the caller is running on the CPU being monitored by the session.
3127 * It does not have to be the owner (ctx_task) of the context per se.
3129 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3130 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3133 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3135 expert_mode = pfm_sysctl.expert_mode;
3137 for (i = 0; i < count; i++, req++) {
3139 cnum = req->reg_num;
3140 value = req->reg_value;
3142 if (!PMD_IS_IMPL(cnum)) {
3143 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3146 is_counting = PMD_IS_COUNTING(cnum);
3147 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3150 * execute write checker, if any
3152 if (unlikely(expert_mode == 0 && wr_func)) {
3153 unsigned long v = value;
3155 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3156 if (ret) goto abort_mission;
3163 * no error on this register
3165 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3168 * now commit changes to software state
3173 * update virtualized (64bits) counter
3177 * write context state
3179 ctx->ctx_pmds[cnum].lval = value;
3182 * when context is load we use the split value
3185 hw_value = value & ovfl_mask;
3186 value = value & ~ovfl_mask;
3190 * update reset values (not just for counters)
3192 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3193 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3196 * update randomization parameters (not just for counters)
3198 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3199 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3202 * update context value
3204 ctx->ctx_pmds[cnum].val = value;
3207 * Keep track of what we use
3209 * We do not keep track of PMC because we have to
3210 * systematically restore ALL of them.
3212 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3215 * mark this PMD register used as well
3217 CTX_USED_PMD(ctx, RDEP(cnum));
3220 * make sure we do not try to reset on
3221 * restart because we have established new values
3223 if (is_counting && state == PFM_CTX_MASKED) {
3224 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3229 * write thread state
3231 if (is_system == 0) thread->pmds[cnum] = hw_value;
3234 * write hardware register if we can
3236 if (can_access_pmu) {
3237 ia64_set_pmd(cnum, hw_value);
3241 * we are guaranteed that the task is not running on the other CPU,
3242 * we indicate that this PMD will need to be reloaded if the task
3243 * is rescheduled on the CPU it ran last on.
3245 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3250 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3251 "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",
3257 ctx->ctx_pmds[cnum].val,
3258 ctx->ctx_pmds[cnum].short_reset,
3259 ctx->ctx_pmds[cnum].long_reset,
3260 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3261 ctx->ctx_pmds[cnum].seed,
3262 ctx->ctx_pmds[cnum].mask,
3263 ctx->ctx_used_pmds[0],
3264 ctx->ctx_pmds[cnum].reset_pmds[0],
3265 ctx->ctx_reload_pmds[0],
3266 ctx->ctx_all_pmds[0],
3267 ctx->ctx_ovfl_regs[0]));
3271 * make changes visible
3273 if (can_access_pmu) ia64_srlz_d();
3279 * for now, we have only one possibility for error
3281 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3286 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3287 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3288 * interrupt is delivered during the call, it will be kept pending until we leave, making
3289 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3290 * guaranteed to return consistent data to the user, it may simply be old. It is not
3291 * trivial to treat the overflow while inside the call because you may end up in
3292 * some module sampling buffer code causing deadlocks.
3295 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3297 struct thread_struct *thread = NULL;
3298 struct task_struct *task;
3299 unsigned long val = 0UL, lval, ovfl_mask, sval;
3300 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3301 unsigned int cnum, reg_flags = 0;
3302 int i, can_access_pmu = 0, state;
3303 int is_loaded, is_system, is_counting, expert_mode;
3305 pfm_reg_check_t rd_func;
3308 * access is possible when loaded only for
3309 * self-monitoring tasks or in UP mode
3312 state = ctx->ctx_state;
3313 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3314 is_system = ctx->ctx_fl_system;
3315 ovfl_mask = pmu_conf->ovfl_val;
3316 task = ctx->ctx_task;
3318 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3320 if (likely(is_loaded)) {
3321 thread = &task->thread;
3323 * In system wide and when the context is loaded, access can only happen
3324 * when the caller is running on the CPU being monitored by the session.
3325 * It does not have to be the owner (ctx_task) of the context per se.
3327 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3328 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3332 * this can be true when not self-monitoring only in UP
3334 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3336 if (can_access_pmu) ia64_srlz_d();
3338 expert_mode = pfm_sysctl.expert_mode;
3340 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3346 * on both UP and SMP, we can only read the PMD from the hardware register when
3347 * the task is the owner of the local PMU.
3350 for (i = 0; i < count; i++, req++) {
3352 cnum = req->reg_num;
3353 reg_flags = req->reg_flags;
3355 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3357 * we can only read the register that we use. That includes
3358 * the one we explicitely initialize AND the one we want included
3359 * in the sampling buffer (smpl_regs).
3361 * Having this restriction allows optimization in the ctxsw routine
3362 * without compromising security (leaks)
3364 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3366 sval = ctx->ctx_pmds[cnum].val;
3367 lval = ctx->ctx_pmds[cnum].lval;
3368 is_counting = PMD_IS_COUNTING(cnum);
3371 * If the task is not the current one, then we check if the
3372 * PMU state is still in the local live register due to lazy ctxsw.
3373 * If true, then we read directly from the registers.
3375 if (can_access_pmu){
3376 val = ia64_get_pmd(cnum);
3379 * context has been saved
3380 * if context is zombie, then task does not exist anymore.
3381 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3383 val = is_loaded ? thread->pmds[cnum] : 0UL;
3385 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3389 * XXX: need to check for overflow when loaded
3396 * execute read checker, if any
3398 if (unlikely(expert_mode == 0 && rd_func)) {
3399 unsigned long v = val;
3400 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3401 if (ret) goto error;
3406 PFM_REG_RETFLAG_SET(reg_flags, 0);
3408 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3411 * update register return value, abort all if problem during copy.
3412 * we only modify the reg_flags field. no check mode is fine because
3413 * access has been verified upfront in sys_perfmonctl().
3415 req->reg_value = val;
3416 req->reg_flags = reg_flags;
3417 req->reg_last_reset_val = lval;
3423 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3428 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3432 if (req == NULL) return -EINVAL;
3434 ctx = GET_PMU_CTX();
3436 if (ctx == NULL) return -EINVAL;
3439 * for now limit to current task, which is enough when calling
3440 * from overflow handler
3442 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3444 return pfm_write_pmcs(ctx, req, nreq, regs);
3446 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3449 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3453 if (req == NULL) return -EINVAL;
3455 ctx = GET_PMU_CTX();
3457 if (ctx == NULL) return -EINVAL;
3460 * for now limit to current task, which is enough when calling
3461 * from overflow handler
3463 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3465 return pfm_read_pmds(ctx, req, nreq, regs);
3467 EXPORT_SYMBOL(pfm_mod_read_pmds);
3470 * Only call this function when a process it trying to
3471 * write the debug registers (reading is always allowed)
3474 pfm_use_debug_registers(struct task_struct *task)
3476 pfm_context_t *ctx = task->thread.pfm_context;
3477 unsigned long flags;
3480 if (pmu_conf->use_rr_dbregs == 0) return 0;
3482 DPRINT(("called for [%d]\n", task->pid));
3487 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3490 * Even on SMP, we do not need to use an atomic here because
3491 * the only way in is via ptrace() and this is possible only when the
3492 * process is stopped. Even in the case where the ctxsw out is not totally
3493 * completed by the time we come here, there is no way the 'stopped' process
3494 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3495 * So this is always safe.
3497 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3502 * We cannot allow setting breakpoints when system wide monitoring
3503 * sessions are using the debug registers.
3505 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3508 pfm_sessions.pfs_ptrace_use_dbregs++;
3510 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3511 pfm_sessions.pfs_ptrace_use_dbregs,
3512 pfm_sessions.pfs_sys_use_dbregs,
3521 * This function is called for every task that exits with the
3522 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3523 * able to use the debug registers for debugging purposes via
3524 * ptrace(). Therefore we know it was not using them for
3525 * perfmormance monitoring, so we only decrement the number
3526 * of "ptraced" debug register users to keep the count up to date
3529 pfm_release_debug_registers(struct task_struct *task)
3531 unsigned long flags;
3534 if (pmu_conf->use_rr_dbregs == 0) return 0;
3537 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3538 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3541 pfm_sessions.pfs_ptrace_use_dbregs--;
3550 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3552 struct task_struct *task;
3553 pfm_buffer_fmt_t *fmt;
3554 pfm_ovfl_ctrl_t rst_ctrl;
3555 int state, is_system;
3558 state = ctx->ctx_state;
3559 fmt = ctx->ctx_buf_fmt;
3560 is_system = ctx->ctx_fl_system;
3561 task = PFM_CTX_TASK(ctx);
3564 case PFM_CTX_MASKED:
3566 case PFM_CTX_LOADED:
3567 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3569 case PFM_CTX_UNLOADED:
3570 case PFM_CTX_ZOMBIE:
3571 DPRINT(("invalid state=%d\n", state));
3574 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3579 * In system wide and when the context is loaded, access can only happen
3580 * when the caller is running on the CPU being monitored by the session.
3581 * It does not have to be the owner (ctx_task) of the context per se.
3583 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3584 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3589 if (unlikely(task == NULL)) {
3590 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3594 if (task == current || is_system) {
3596 fmt = ctx->ctx_buf_fmt;
3598 DPRINT(("restarting self %d ovfl=0x%lx\n",
3600 ctx->ctx_ovfl_regs[0]));
3602 if (CTX_HAS_SMPL(ctx)) {
3604 prefetch(ctx->ctx_smpl_hdr);
3606 rst_ctrl.bits.mask_monitoring = 0;
3607 rst_ctrl.bits.reset_ovfl_pmds = 0;
3609 if (state == PFM_CTX_LOADED)
3610 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3612 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3614 rst_ctrl.bits.mask_monitoring = 0;
3615 rst_ctrl.bits.reset_ovfl_pmds = 1;
3619 if (rst_ctrl.bits.reset_ovfl_pmds)
3620 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3622 if (rst_ctrl.bits.mask_monitoring == 0) {
3623 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3625 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3627 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3629 // cannot use pfm_stop_monitoring(task, regs);
3633 * clear overflowed PMD mask to remove any stale information
3635 ctx->ctx_ovfl_regs[0] = 0UL;
3638 * back to LOADED state
3640 ctx->ctx_state = PFM_CTX_LOADED;
3643 * XXX: not really useful for self monitoring
3645 ctx->ctx_fl_can_restart = 0;
3651 * restart another task
3655 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3656 * one is seen by the task.
3658 if (state == PFM_CTX_MASKED) {
3659 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3661 * will prevent subsequent restart before this one is
3662 * seen by other task
3664 ctx->ctx_fl_can_restart = 0;
3668 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3669 * the task is blocked or on its way to block. That's the normal
3670 * restart path. If the monitoring is not masked, then the task
3671 * can be actively monitoring and we cannot directly intervene.
3672 * Therefore we use the trap mechanism to catch the task and
3673 * force it to reset the buffer/reset PMDs.
3675 * if non-blocking, then we ensure that the task will go into
3676 * pfm_handle_work() before returning to user mode.
3678 * We cannot explicitely reset another task, it MUST always
3679 * be done by the task itself. This works for system wide because
3680 * the tool that is controlling the session is logically doing
3681 * "self-monitoring".
3683 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3684 DPRINT(("unblocking [%d] \n", task->pid));
3685 up(&ctx->ctx_restart_sem);
3687 DPRINT(("[%d] armed exit trap\n", task->pid));
3689 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3691 PFM_SET_WORK_PENDING(task, 1);
3693 pfm_set_task_notify(task);
3696 * XXX: send reschedule if task runs on another CPU
3703 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3705 unsigned int m = *(unsigned int *)arg;
3707 pfm_sysctl.debug = m == 0 ? 0 : 1;
3709 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3712 memset(pfm_stats, 0, sizeof(pfm_stats));
3713 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3719 * arg can be NULL and count can be zero for this function
3722 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3724 struct thread_struct *thread = NULL;
3725 struct task_struct *task;
3726 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3727 unsigned long flags;
3732 int i, can_access_pmu = 0;
3733 int is_system, is_loaded;
3735 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3737 state = ctx->ctx_state;
3738 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3739 is_system = ctx->ctx_fl_system;
3740 task = ctx->ctx_task;
3742 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3745 * on both UP and SMP, we can only write to the PMC when the task is
3746 * the owner of the local PMU.
3749 thread = &task->thread;
3751 * In system wide and when the context is loaded, access can only happen
3752 * when the caller is running on the CPU being monitored by the session.
3753 * It does not have to be the owner (ctx_task) of the context per se.
3755 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3756 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3759 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3763 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3764 * ensuring that no real breakpoint can be installed via this call.
3766 * IMPORTANT: regs can be NULL in this function
3769 first_time = ctx->ctx_fl_using_dbreg == 0;
3772 * don't bother if we are loaded and task is being debugged
3774 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3775 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3780 * check for debug registers in system wide mode
3782 * If though a check is done in pfm_context_load(),
3783 * we must repeat it here, in case the registers are
3784 * written after the context is loaded
3789 if (first_time && is_system) {
3790 if (pfm_sessions.pfs_ptrace_use_dbregs)
3793 pfm_sessions.pfs_sys_use_dbregs++;
3798 if (ret != 0) return ret;
3801 * mark ourself as user of the debug registers for
3804 ctx->ctx_fl_using_dbreg = 1;
3807 * clear hardware registers to make sure we don't
3808 * pick up stale state.
3810 * for a system wide session, we do not use
3811 * thread.dbr, thread.ibr because this process
3812 * never leaves the current CPU and the state
3813 * is shared by all processes running on it
3815 if (first_time && can_access_pmu) {
3816 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3817 for (i=0; i < pmu_conf->num_ibrs; i++) {
3818 ia64_set_ibr(i, 0UL);
3819 ia64_dv_serialize_instruction();
3822 for (i=0; i < pmu_conf->num_dbrs; i++) {
3823 ia64_set_dbr(i, 0UL);
3824 ia64_dv_serialize_data();
3830 * Now install the values into the registers
3832 for (i = 0; i < count; i++, req++) {
3834 rnum = req->dbreg_num;
3835 dbreg.val = req->dbreg_value;
3839 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3840 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3841 rnum, dbreg.val, mode, i, count));
3847 * make sure we do not install enabled breakpoint
3850 if (mode == PFM_CODE_RR)
3851 dbreg.ibr.ibr_x = 0;
3853 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3856 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3859 * Debug registers, just like PMC, can only be modified
3860 * by a kernel call. Moreover, perfmon() access to those
3861 * registers are centralized in this routine. The hardware
3862 * does not modify the value of these registers, therefore,
3863 * if we save them as they are written, we can avoid having
3864 * to save them on context switch out. This is made possible
3865 * by the fact that when perfmon uses debug registers, ptrace()
3866 * won't be able to modify them concurrently.
3868 if (mode == PFM_CODE_RR) {
3869 CTX_USED_IBR(ctx, rnum);
3871 if (can_access_pmu) {
3872 ia64_set_ibr(rnum, dbreg.val);
3873 ia64_dv_serialize_instruction();
3876 ctx->ctx_ibrs[rnum] = dbreg.val;
3878 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3879 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3881 CTX_USED_DBR(ctx, rnum);
3883 if (can_access_pmu) {
3884 ia64_set_dbr(rnum, dbreg.val);
3885 ia64_dv_serialize_data();
3887 ctx->ctx_dbrs[rnum] = dbreg.val;
3889 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3890 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3898 * in case it was our first attempt, we undo the global modifications
3902 if (ctx->ctx_fl_system) {
3903 pfm_sessions.pfs_sys_use_dbregs--;
3906 ctx->ctx_fl_using_dbreg = 0;
3909 * install error return flag
3911 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3917 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3919 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3923 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3925 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3929 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3933 if (req == NULL) return -EINVAL;
3935 ctx = GET_PMU_CTX();
3937 if (ctx == NULL) return -EINVAL;
3940 * for now limit to current task, which is enough when calling
3941 * from overflow handler
3943 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3945 return pfm_write_ibrs(ctx, req, nreq, regs);
3947 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3950 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3954 if (req == NULL) return -EINVAL;
3956 ctx = GET_PMU_CTX();
3958 if (ctx == NULL) return -EINVAL;
3961 * for now limit to current task, which is enough when calling
3962 * from overflow handler
3964 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3966 return pfm_write_dbrs(ctx, req, nreq, regs);
3968 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3972 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3974 pfarg_features_t *req = (pfarg_features_t *)arg;
3976 req->ft_version = PFM_VERSION;
3981 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3983 struct pt_regs *tregs;
3984 struct task_struct *task = PFM_CTX_TASK(ctx);
3985 int state, is_system;
3987 state = ctx->ctx_state;
3988 is_system = ctx->ctx_fl_system;
3991 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3993 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3996 * In system wide and when the context is loaded, access can only happen
3997 * when the caller is running on the CPU being monitored by the session.
3998 * It does not have to be the owner (ctx_task) of the context per se.
4000 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4001 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4004 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4005 PFM_CTX_TASK(ctx)->pid,
4009 * in system mode, we need to update the PMU directly
4010 * and the user level state of the caller, which may not
4011 * necessarily be the creator of the context.
4015 * Update local PMU first
4019 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4023 * update local cpuinfo
4025 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4028 * stop monitoring, does srlz.i
4033 * stop monitoring in the caller
4035 ia64_psr(regs)->pp = 0;
4043 if (task == current) {
4044 /* stop monitoring at kernel level */
4048 * stop monitoring at the user level
4050 ia64_psr(regs)->up = 0;
4052 tregs = ia64_task_regs(task);
4055 * stop monitoring at the user level
4057 ia64_psr(tregs)->up = 0;
4060 * monitoring disabled in kernel at next reschedule
4062 ctx->ctx_saved_psr_up = 0;
4063 DPRINT(("task=[%d]\n", task->pid));
4070 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4072 struct pt_regs *tregs;
4073 int state, is_system;
4075 state = ctx->ctx_state;
4076 is_system = ctx->ctx_fl_system;
4078 if (state != PFM_CTX_LOADED) return -EINVAL;
4081 * In system wide and when the context is loaded, access can only happen
4082 * when the caller is running on the CPU being monitored by the session.
4083 * It does not have to be the owner (ctx_task) of the context per se.
4085 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4086 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4091 * in system mode, we need to update the PMU directly
4092 * and the user level state of the caller, which may not
4093 * necessarily be the creator of the context.
4098 * set user level psr.pp for the caller
4100 ia64_psr(regs)->pp = 1;
4103 * now update the local PMU and cpuinfo
4105 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4108 * start monitoring at kernel level
4113 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4123 if (ctx->ctx_task == current) {
4125 /* start monitoring at kernel level */
4129 * activate monitoring at user level
4131 ia64_psr(regs)->up = 1;
4134 tregs = ia64_task_regs(ctx->ctx_task);
4137 * start monitoring at the kernel level the next
4138 * time the task is scheduled
4140 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4143 * activate monitoring at user level
4145 ia64_psr(tregs)->up = 1;
4151 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4153 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4158 for (i = 0; i < count; i++, req++) {
4160 cnum = req->reg_num;
4162 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4164 req->reg_value = PMC_DFL_VAL(cnum);
4166 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4168 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4173 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4178 pfm_check_task_exist(pfm_context_t *ctx)
4180 struct task_struct *g, *t;
4183 read_lock(&tasklist_lock);
4185 do_each_thread (g, t) {
4186 if (t->thread.pfm_context == ctx) {
4190 } while_each_thread (g, t);
4192 read_unlock(&tasklist_lock);
4194 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4200 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4202 struct task_struct *task;
4203 struct thread_struct *thread;
4204 struct pfm_context_t *old;
4205 unsigned long flags;
4207 struct task_struct *owner_task = NULL;
4209 pfarg_load_t *req = (pfarg_load_t *)arg;
4210 unsigned long *pmcs_source, *pmds_source;
4213 int state, is_system, set_dbregs = 0;
4215 state = ctx->ctx_state;
4216 is_system = ctx->ctx_fl_system;
4218 * can only load from unloaded or terminated state
4220 if (state != PFM_CTX_UNLOADED) {
4221 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4227 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4229 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4230 DPRINT(("cannot use blocking mode on self\n"));
4234 ret = pfm_get_task(ctx, req->load_pid, &task);
4236 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4243 * system wide is self monitoring only
4245 if (is_system && task != current) {
4246 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4251 thread = &task->thread;
4255 * cannot load a context which is using range restrictions,
4256 * into a task that is being debugged.
4258 if (ctx->ctx_fl_using_dbreg) {
4259 if (thread->flags & IA64_THREAD_DBG_VALID) {
4261 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4267 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4268 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4271 pfm_sessions.pfs_sys_use_dbregs++;
4272 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4279 if (ret) goto error;
4283 * SMP system-wide monitoring implies self-monitoring.
4285 * The programming model expects the task to
4286 * be pinned on a CPU throughout the session.
4287 * Here we take note of the current CPU at the
4288 * time the context is loaded. No call from
4289 * another CPU will be allowed.
4291 * The pinning via shed_setaffinity()
4292 * must be done by the calling task prior
4295 * systemwide: keep track of CPU this session is supposed to run on
4297 the_cpu = ctx->ctx_cpu = smp_processor_id();
4301 * now reserve the session
4303 ret = pfm_reserve_session(current, is_system, the_cpu);
4304 if (ret) goto error;
4307 * task is necessarily stopped at this point.
4309 * If the previous context was zombie, then it got removed in
4310 * pfm_save_regs(). Therefore we should not see it here.
4311 * If we see a context, then this is an active context
4313 * XXX: needs to be atomic
4315 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4316 thread->pfm_context, ctx));
4318 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4320 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4324 pfm_reset_msgq(ctx);
4326 ctx->ctx_state = PFM_CTX_LOADED;
4329 * link context to task
4331 ctx->ctx_task = task;
4335 * we load as stopped
4337 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4338 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4340 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4342 thread->flags |= IA64_THREAD_PM_VALID;
4346 * propagate into thread-state
4348 pfm_copy_pmds(task, ctx);
4349 pfm_copy_pmcs(task, ctx);
4351 pmcs_source = thread->pmcs;
4352 pmds_source = thread->pmds;
4355 * always the case for system-wide
4357 if (task == current) {
4359 if (is_system == 0) {
4361 /* allow user level control */
4362 ia64_psr(regs)->sp = 0;
4363 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4365 SET_LAST_CPU(ctx, smp_processor_id());
4367 SET_ACTIVATION(ctx);
4370 * push the other task out, if any
4372 owner_task = GET_PMU_OWNER();
4373 if (owner_task) pfm_lazy_save_regs(owner_task);
4377 * load all PMD from ctx to PMU (as opposed to thread state)
4378 * restore all PMC from ctx to PMU
4380 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4381 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4383 ctx->ctx_reload_pmcs[0] = 0UL;
4384 ctx->ctx_reload_pmds[0] = 0UL;
4387 * guaranteed safe by earlier check against DBG_VALID
4389 if (ctx->ctx_fl_using_dbreg) {
4390 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4391 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4396 SET_PMU_OWNER(task, ctx);
4398 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4401 * when not current, task MUST be stopped, so this is safe
4403 regs = ia64_task_regs(task);
4405 /* force a full reload */
4406 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4407 SET_LAST_CPU(ctx, -1);
4409 /* initial saved psr (stopped) */
4410 ctx->ctx_saved_psr_up = 0UL;
4411 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4417 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4420 * we must undo the dbregs setting (for system-wide)
4422 if (ret && set_dbregs) {
4424 pfm_sessions.pfs_sys_use_dbregs--;
4428 * release task, there is now a link with the context
4430 if (is_system == 0 && task != current) {
4434 ret = pfm_check_task_exist(ctx);
4436 ctx->ctx_state = PFM_CTX_UNLOADED;
4437 ctx->ctx_task = NULL;
4445 * in this function, we do not need to increase the use count
4446 * for the task via get_task_struct(), because we hold the
4447 * context lock. If the task were to disappear while having
4448 * a context attached, it would go through pfm_exit_thread()
4449 * which also grabs the context lock and would therefore be blocked
4450 * until we are here.
4452 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4455 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4457 struct task_struct *task = PFM_CTX_TASK(ctx);
4458 struct pt_regs *tregs;
4459 int prev_state, is_system;
4462 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4464 prev_state = ctx->ctx_state;
4465 is_system = ctx->ctx_fl_system;
4468 * unload only when necessary
4470 if (prev_state == PFM_CTX_UNLOADED) {
4471 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4476 * clear psr and dcr bits
4478 ret = pfm_stop(ctx, NULL, 0, regs);
4479 if (ret) return ret;
4481 ctx->ctx_state = PFM_CTX_UNLOADED;
4484 * in system mode, we need to update the PMU directly
4485 * and the user level state of the caller, which may not
4486 * necessarily be the creator of the context.
4493 * local PMU is taken care of in pfm_stop()
4495 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4496 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4499 * save PMDs in context
4502 pfm_flush_pmds(current, ctx);
4505 * at this point we are done with the PMU
4506 * so we can unreserve the resource.
4508 if (prev_state != PFM_CTX_ZOMBIE)
4509 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4512 * disconnect context from task
4514 task->thread.pfm_context = NULL;
4516 * disconnect task from context
4518 ctx->ctx_task = NULL;
4521 * There is nothing more to cleanup here.
4529 tregs = task == current ? regs : ia64_task_regs(task);
4531 if (task == current) {
4533 * cancel user level control
4535 ia64_psr(regs)->sp = 1;
4537 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4540 * save PMDs to context
4543 pfm_flush_pmds(task, ctx);
4546 * at this point we are done with the PMU
4547 * so we can unreserve the resource.
4549 * when state was ZOMBIE, we have already unreserved.
4551 if (prev_state != PFM_CTX_ZOMBIE)
4552 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4555 * reset activation counter and psr
4557 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4558 SET_LAST_CPU(ctx, -1);
4561 * PMU state will not be restored
4563 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4566 * break links between context and task
4568 task->thread.pfm_context = NULL;
4569 ctx->ctx_task = NULL;
4571 PFM_SET_WORK_PENDING(task, 0);
4573 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4574 ctx->ctx_fl_can_restart = 0;
4575 ctx->ctx_fl_going_zombie = 0;
4577 DPRINT(("disconnected [%d] from context\n", task->pid));
4584 * called only from exit_thread(): task == current
4585 * we come here only if current has a context attached (loaded or masked)
4588 pfm_exit_thread(struct task_struct *task)
4591 unsigned long flags;
4592 struct pt_regs *regs = ia64_task_regs(task);
4596 ctx = PFM_GET_CTX(task);
4598 PROTECT_CTX(ctx, flags);
4600 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4602 state = ctx->ctx_state;
4604 case PFM_CTX_UNLOADED:
4606 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4607 * be in unloaded state
4609 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4611 case PFM_CTX_LOADED:
4612 case PFM_CTX_MASKED:
4613 ret = pfm_context_unload(ctx, NULL, 0, regs);
4615 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4617 DPRINT(("ctx unloaded for current state was %d\n", state));
4619 pfm_end_notify_user(ctx);
4621 case PFM_CTX_ZOMBIE:
4622 ret = pfm_context_unload(ctx, NULL, 0, regs);
4624 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4629 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4632 UNPROTECT_CTX(ctx, flags);
4634 { u64 psr = pfm_get_psr();
4635 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4636 BUG_ON(GET_PMU_OWNER());
4637 BUG_ON(ia64_psr(regs)->up);
4638 BUG_ON(ia64_psr(regs)->pp);
4642 * All memory free operations (especially for vmalloc'ed memory)
4643 * MUST be done with interrupts ENABLED.
4645 if (free_ok) pfm_context_free(ctx);
4649 * functions MUST be listed in the increasing order of their index (see permfon.h)
4651 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4652 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4653 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4654 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4655 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4657 static pfm_cmd_desc_t pfm_cmd_tab[]={
4658 /* 0 */PFM_CMD_NONE,
4659 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4660 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4661 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4662 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4663 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4664 /* 6 */PFM_CMD_NONE,
4665 /* 7 */PFM_CMD_NONE,
4666 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4667 /* 9 */PFM_CMD_NONE,
4668 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4669 /* 11 */PFM_CMD_NONE,
4670 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4671 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4672 /* 14 */PFM_CMD_NONE,
4673 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4674 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4675 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4676 /* 18 */PFM_CMD_NONE,
4677 /* 19 */PFM_CMD_NONE,
4678 /* 20 */PFM_CMD_NONE,
4679 /* 21 */PFM_CMD_NONE,
4680 /* 22 */PFM_CMD_NONE,
4681 /* 23 */PFM_CMD_NONE,
4682 /* 24 */PFM_CMD_NONE,
4683 /* 25 */PFM_CMD_NONE,
4684 /* 26 */PFM_CMD_NONE,
4685 /* 27 */PFM_CMD_NONE,
4686 /* 28 */PFM_CMD_NONE,
4687 /* 29 */PFM_CMD_NONE,
4688 /* 30 */PFM_CMD_NONE,
4689 /* 31 */PFM_CMD_NONE,
4690 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4691 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4693 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4696 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4698 struct task_struct *task;
4699 int state, old_state;
4702 state = ctx->ctx_state;
4703 task = ctx->ctx_task;
4706 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4710 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4714 task->state, PFM_CMD_STOPPED(cmd)));
4717 * self-monitoring always ok.
4719 * for system-wide the caller can either be the creator of the
4720 * context (to one to which the context is attached to) OR
4721 * a task running on the same CPU as the session.
4723 if (task == current || ctx->ctx_fl_system) return 0;
4726 * we are monitoring another thread
4729 case PFM_CTX_UNLOADED:
4731 * if context is UNLOADED we are safe to go
4734 case PFM_CTX_ZOMBIE:
4736 * no command can operate on a zombie context
4738 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4740 case PFM_CTX_MASKED:
4742 * PMU state has been saved to software even though
4743 * the thread may still be running.
4745 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4749 * context is LOADED or MASKED. Some commands may need to have
4752 * We could lift this restriction for UP but it would mean that
4753 * the user has no guarantee the task would not run between
4754 * two successive calls to perfmonctl(). That's probably OK.
4755 * If this user wants to ensure the task does not run, then
4756 * the task must be stopped.
4758 if (PFM_CMD_STOPPED(cmd)) {
4759 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
4760 DPRINT(("[%d] task not in stopped state\n", task->pid));
4764 * task is now stopped, wait for ctxsw out
4766 * This is an interesting point in the code.
4767 * We need to unprotect the context because
4768 * the pfm_save_regs() routines needs to grab
4769 * the same lock. There are danger in doing
4770 * this because it leaves a window open for
4771 * another task to get access to the context
4772 * and possibly change its state. The one thing
4773 * that is not possible is for the context to disappear
4774 * because we are protected by the VFS layer, i.e.,
4775 * get_fd()/put_fd().
4779 UNPROTECT_CTX(ctx, flags);
4781 wait_task_inactive(task);
4783 PROTECT_CTX(ctx, flags);
4786 * we must recheck to verify if state has changed
4788 if (ctx->ctx_state != old_state) {
4789 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4797 * system-call entry point (must return long)
4800 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4802 struct file *file = NULL;
4803 pfm_context_t *ctx = NULL;
4804 unsigned long flags = 0UL;
4805 void *args_k = NULL;
4806 long ret; /* will expand int return types */
4807 size_t base_sz, sz, xtra_sz = 0;
4808 int narg, completed_args = 0, call_made = 0, cmd_flags;
4809 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4810 int (*getsize)(void *arg, size_t *sz);
4811 #define PFM_MAX_ARGSIZE 4096
4814 * reject any call if perfmon was disabled at initialization
4816 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4818 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4819 DPRINT(("invalid cmd=%d\n", cmd));
4823 func = pfm_cmd_tab[cmd].cmd_func;
4824 narg = pfm_cmd_tab[cmd].cmd_narg;
4825 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4826 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4827 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4829 if (unlikely(func == NULL)) {
4830 DPRINT(("invalid cmd=%d\n", cmd));
4834 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4842 * check if number of arguments matches what the command expects
4844 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4848 sz = xtra_sz + base_sz*count;
4850 * limit abuse to min page size
4852 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4853 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4858 * allocate default-sized argument buffer
4860 if (likely(count && args_k == NULL)) {
4861 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4862 if (args_k == NULL) return -ENOMEM;
4870 * assume sz = 0 for command without parameters
4872 if (sz && copy_from_user(args_k, arg, sz)) {
4873 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4878 * check if command supports extra parameters
4880 if (completed_args == 0 && getsize) {
4882 * get extra parameters size (based on main argument)
4884 ret = (*getsize)(args_k, &xtra_sz);
4885 if (ret) goto error_args;
4889 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4891 /* retry if necessary */
4892 if (likely(xtra_sz)) goto restart_args;
4895 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4900 if (unlikely(file == NULL)) {
4901 DPRINT(("invalid fd %d\n", fd));
4904 if (unlikely(PFM_IS_FILE(file) == 0)) {
4905 DPRINT(("fd %d not related to perfmon\n", fd));
4909 ctx = (pfm_context_t *)file->private_data;
4910 if (unlikely(ctx == NULL)) {
4911 DPRINT(("no context for fd %d\n", fd));
4914 prefetch(&ctx->ctx_state);
4916 PROTECT_CTX(ctx, flags);
4919 * check task is stopped
4921 ret = pfm_check_task_state(ctx, cmd, flags);
4922 if (unlikely(ret)) goto abort_locked;
4925 ret = (*func)(ctx, args_k, count, ia64_task_regs(current));
4931 DPRINT(("context unlocked\n"));
4932 UNPROTECT_CTX(ctx, flags);
4936 /* copy argument back to user, if needed */
4937 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4940 if (args_k) kfree(args_k);
4942 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4948 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4950 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4951 pfm_ovfl_ctrl_t rst_ctrl;
4955 state = ctx->ctx_state;
4957 * Unlock sampling buffer and reset index atomically
4958 * XXX: not really needed when blocking
4960 if (CTX_HAS_SMPL(ctx)) {
4962 rst_ctrl.bits.mask_monitoring = 0;
4963 rst_ctrl.bits.reset_ovfl_pmds = 0;
4965 if (state == PFM_CTX_LOADED)
4966 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4968 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4970 rst_ctrl.bits.mask_monitoring = 0;
4971 rst_ctrl.bits.reset_ovfl_pmds = 1;
4975 if (rst_ctrl.bits.reset_ovfl_pmds) {
4976 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4978 if (rst_ctrl.bits.mask_monitoring == 0) {
4979 DPRINT(("resuming monitoring\n"));
4980 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4982 DPRINT(("stopping monitoring\n"));
4983 //pfm_stop_monitoring(current, regs);
4985 ctx->ctx_state = PFM_CTX_LOADED;
4990 * context MUST BE LOCKED when calling
4991 * can only be called for current
4994 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4998 DPRINT(("entering for [%d]\n", current->pid));
5000 ret = pfm_context_unload(ctx, NULL, 0, regs);
5002 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", current->pid, ret);
5006 * and wakeup controlling task, indicating we are now disconnected
5008 wake_up_interruptible(&ctx->ctx_zombieq);
5011 * given that context is still locked, the controlling
5012 * task will only get access when we return from
5013 * pfm_handle_work().
5017 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5019 * pfm_handle_work() can be called with interrupts enabled
5020 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5021 * call may sleep, therefore we must re-enable interrupts
5022 * to avoid deadlocks. It is safe to do so because this function
5023 * is called ONLY when returning to user level (PUStk=1), in which case
5024 * there is no risk of kernel stack overflow due to deep
5025 * interrupt nesting.
5028 pfm_handle_work(void)
5031 struct pt_regs *regs;
5032 unsigned long flags, dummy_flags;
5033 unsigned long ovfl_regs;
5034 unsigned int reason;
5037 ctx = PFM_GET_CTX(current);
5039 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5043 PROTECT_CTX(ctx, flags);
5045 PFM_SET_WORK_PENDING(current, 0);
5047 pfm_clear_task_notify();
5049 regs = ia64_task_regs(current);
5052 * extract reason for being here and clear
5054 reason = ctx->ctx_fl_trap_reason;
5055 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5056 ovfl_regs = ctx->ctx_ovfl_regs[0];
5058 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5061 * must be done before we check for simple-reset mode
5063 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5066 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5067 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5070 * restore interrupt mask to what it was on entry.
5071 * Could be enabled/diasbled.
5073 UNPROTECT_CTX(ctx, flags);
5076 * force interrupt enable because of down_interruptible()
5080 DPRINT(("before block sleeping\n"));
5083 * may go through without blocking on SMP systems
5084 * if restart has been received already by the time we call down()
5086 ret = down_interruptible(&ctx->ctx_restart_sem);
5088 DPRINT(("after block sleeping ret=%d\n", ret));
5091 * lock context and mask interrupts again
5092 * We save flags into a dummy because we may have
5093 * altered interrupts mask compared to entry in this
5096 PROTECT_CTX(ctx, dummy_flags);
5099 * we need to read the ovfl_regs only after wake-up
5100 * because we may have had pfm_write_pmds() in between
5101 * and that can changed PMD values and therefore
5102 * ovfl_regs is reset for these new PMD values.
5104 ovfl_regs = ctx->ctx_ovfl_regs[0];
5106 if (ctx->ctx_fl_going_zombie) {
5108 DPRINT(("context is zombie, bailing out\n"));
5109 pfm_context_force_terminate(ctx, regs);
5113 * in case of interruption of down() we don't restart anything
5115 if (ret < 0) goto nothing_to_do;
5118 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5119 ctx->ctx_ovfl_regs[0] = 0UL;
5123 * restore flags as they were upon entry
5125 UNPROTECT_CTX(ctx, flags);
5129 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5131 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5132 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5136 DPRINT(("waking up somebody\n"));
5138 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5141 * safe, we are not in intr handler, nor in ctxsw when
5144 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5150 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5152 pfm_msg_t *msg = NULL;
5154 if (ctx->ctx_fl_no_msg == 0) {
5155 msg = pfm_get_new_msg(ctx);
5157 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5161 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5162 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5163 msg->pfm_ovfl_msg.msg_active_set = 0;
5164 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5165 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5166 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5167 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5168 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5171 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5177 return pfm_notify_user(ctx, msg);
5181 pfm_end_notify_user(pfm_context_t *ctx)
5185 msg = pfm_get_new_msg(ctx);
5187 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5191 memset(msg, 0, sizeof(*msg));
5193 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5194 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5195 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5197 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5202 return pfm_notify_user(ctx, msg);
5206 * main overflow processing routine.
5207 * it can be called from the interrupt path or explicitely during the context switch code
5210 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5212 pfm_ovfl_arg_t *ovfl_arg;
5214 unsigned long old_val, ovfl_val, new_val;
5215 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5216 unsigned long tstamp;
5217 pfm_ovfl_ctrl_t ovfl_ctrl;
5218 unsigned int i, has_smpl;
5219 int must_notify = 0;
5221 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5224 * sanity test. Should never happen
5226 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5228 tstamp = ia64_get_itc();
5229 mask = pmc0 >> PMU_FIRST_COUNTER;
5230 ovfl_val = pmu_conf->ovfl_val;
5231 has_smpl = CTX_HAS_SMPL(ctx);
5233 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5234 "used_pmds=0x%lx\n",
5236 task ? task->pid: -1,
5237 (regs ? regs->cr_iip : 0),
5238 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5239 ctx->ctx_used_pmds[0]));
5243 * first we update the virtual counters
5244 * assume there was a prior ia64_srlz_d() issued
5246 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5248 /* skip pmd which did not overflow */
5249 if ((mask & 0x1) == 0) continue;
5252 * Note that the pmd is not necessarily 0 at this point as qualified events
5253 * may have happened before the PMU was frozen. The residual count is not
5254 * taken into consideration here but will be with any read of the pmd via
5257 old_val = new_val = ctx->ctx_pmds[i].val;
5258 new_val += 1 + ovfl_val;
5259 ctx->ctx_pmds[i].val = new_val;
5262 * check for overflow condition
5264 if (likely(old_val > new_val)) {
5265 ovfl_pmds |= 1UL << i;
5266 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5269 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5273 ia64_get_pmd(i) & ovfl_val,
5279 * there was no 64-bit overflow, nothing else to do
5281 if (ovfl_pmds == 0UL) return;
5284 * reset all control bits
5290 * if a sampling format module exists, then we "cache" the overflow by
5291 * calling the module's handler() routine.
5294 unsigned long start_cycles, end_cycles;
5295 unsigned long pmd_mask;
5297 int this_cpu = smp_processor_id();
5299 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5300 ovfl_arg = &ctx->ctx_ovfl_arg;
5302 prefetch(ctx->ctx_smpl_hdr);
5304 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5308 if ((pmd_mask & 0x1) == 0) continue;
5310 ovfl_arg->ovfl_pmd = (unsigned char )i;
5311 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5312 ovfl_arg->active_set = 0;
5313 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5314 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5316 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5317 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5318 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5321 * copy values of pmds of interest. Sampling format may copy them
5322 * into sampling buffer.
5325 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5326 if ((smpl_pmds & 0x1) == 0) continue;
5327 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5328 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5332 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5334 start_cycles = ia64_get_itc();
5337 * call custom buffer format record (handler) routine
5339 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5341 end_cycles = ia64_get_itc();
5344 * For those controls, we take the union because they have
5345 * an all or nothing behavior.
5347 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5348 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5349 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5351 * build the bitmask of pmds to reset now
5353 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5355 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5358 * when the module cannot handle the rest of the overflows, we abort right here
5360 if (ret && pmd_mask) {
5361 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5362 pmd_mask<<PMU_FIRST_COUNTER));
5365 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5367 ovfl_pmds &= ~reset_pmds;
5370 * when no sampling module is used, then the default
5371 * is to notify on overflow if requested by user
5373 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5374 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5375 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5376 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5378 * if needed, we reset all overflowed pmds
5380 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5383 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5386 * reset the requested PMD registers using the short reset values
5389 unsigned long bm = reset_pmds;
5390 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5393 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5395 * keep track of what to reset when unblocking
5397 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5400 * check for blocking context
5402 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5404 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5407 * set the perfmon specific checking pending work for the task
5409 PFM_SET_WORK_PENDING(task, 1);
5412 * when coming from ctxsw, current still points to the
5413 * previous task, therefore we must work with task and not current.
5415 pfm_set_task_notify(task);
5418 * defer until state is changed (shorten spin window). the context is locked
5419 * anyway, so the signal receiver would come spin for nothing.
5424 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5425 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5426 PFM_GET_WORK_PENDING(task),
5427 ctx->ctx_fl_trap_reason,
5430 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5432 * in case monitoring must be stopped, we toggle the psr bits
5434 if (ovfl_ctrl.bits.mask_monitoring) {
5435 pfm_mask_monitoring(task);
5436 ctx->ctx_state = PFM_CTX_MASKED;
5437 ctx->ctx_fl_can_restart = 1;
5441 * send notification now
5443 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5448 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5450 task ? task->pid : -1,
5456 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5457 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5458 * come here as zombie only if the task is the current task. In which case, we
5459 * can access the PMU hardware directly.
5461 * Note that zombies do have PM_VALID set. So here we do the minimal.
5463 * In case the context was zombified it could not be reclaimed at the time
5464 * the monitoring program exited. At this point, the PMU reservation has been
5465 * returned, the sampiing buffer has been freed. We must convert this call
5466 * into a spurious interrupt. However, we must also avoid infinite overflows
5467 * by stopping monitoring for this task. We can only come here for a per-task
5468 * context. All we need to do is to stop monitoring using the psr bits which
5469 * are always task private. By re-enabling secure montioring, we ensure that
5470 * the monitored task will not be able to re-activate monitoring.
5471 * The task will eventually be context switched out, at which point the context
5472 * will be reclaimed (that includes releasing ownership of the PMU).
5474 * So there might be a window of time where the number of per-task session is zero
5475 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5476 * context. This is safe because if a per-task session comes in, it will push this one
5477 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5478 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5479 * also push our zombie context out.
5481 * Overall pretty hairy stuff....
5483 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5485 ia64_psr(regs)->up = 0;
5486 ia64_psr(regs)->sp = 1;
5491 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5493 struct task_struct *task;
5495 unsigned long flags;
5497 int this_cpu = smp_processor_id();
5500 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5503 * srlz.d done before arriving here
5505 pmc0 = ia64_get_pmc(0);
5507 task = GET_PMU_OWNER();
5508 ctx = GET_PMU_CTX();
5511 * if we have some pending bits set
5512 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5514 if (PMC0_HAS_OVFL(pmc0) && task) {
5516 * we assume that pmc0.fr is always set here
5520 if (!ctx) goto report_spurious1;
5522 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5523 goto report_spurious2;
5525 PROTECT_CTX_NOPRINT(ctx, flags);
5527 pfm_overflow_handler(task, ctx, pmc0, regs);
5529 UNPROTECT_CTX_NOPRINT(ctx, flags);
5532 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5536 * keep it unfrozen at all times
5543 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5544 this_cpu, task->pid);
5548 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5556 pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5558 unsigned long start_cycles, total_cycles;
5559 unsigned long min, max;
5563 this_cpu = get_cpu();
5564 if (likely(!pfm_alt_intr_handler)) {
5565 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5566 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5568 start_cycles = ia64_get_itc();
5570 ret = pfm_do_interrupt_handler(irq, arg, regs);
5572 total_cycles = ia64_get_itc();
5575 * don't measure spurious interrupts
5577 if (likely(ret == 0)) {
5578 total_cycles -= start_cycles;
5580 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5581 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5583 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5587 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5590 put_cpu_no_resched();
5595 * /proc/perfmon interface, for debug only
5598 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5601 pfm_proc_start(struct seq_file *m, loff_t *pos)
5604 return PFM_PROC_SHOW_HEADER;
5607 while (*pos <= NR_CPUS) {
5608 if (cpu_online(*pos - 1)) {
5609 return (void *)*pos;
5617 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5620 return pfm_proc_start(m, pos);
5624 pfm_proc_stop(struct seq_file *m, void *v)
5629 pfm_proc_show_header(struct seq_file *m)
5631 struct list_head * pos;
5632 pfm_buffer_fmt_t * entry;
5633 unsigned long flags;
5636 "perfmon version : %u.%u\n"
5639 "expert mode : %s\n"
5640 "ovfl_mask : 0x%lx\n"
5641 "PMU flags : 0x%x\n",
5642 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5644 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5645 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5652 "proc_sessions : %u\n"
5653 "sys_sessions : %u\n"
5654 "sys_use_dbregs : %u\n"
5655 "ptrace_use_dbregs : %u\n",
5656 pfm_sessions.pfs_task_sessions,
5657 pfm_sessions.pfs_sys_sessions,
5658 pfm_sessions.pfs_sys_use_dbregs,
5659 pfm_sessions.pfs_ptrace_use_dbregs);
5663 spin_lock(&pfm_buffer_fmt_lock);
5665 list_for_each(pos, &pfm_buffer_fmt_list) {
5666 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5667 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5678 entry->fmt_uuid[10],
5679 entry->fmt_uuid[11],
5680 entry->fmt_uuid[12],
5681 entry->fmt_uuid[13],
5682 entry->fmt_uuid[14],
5683 entry->fmt_uuid[15],
5686 spin_unlock(&pfm_buffer_fmt_lock);
5691 pfm_proc_show(struct seq_file *m, void *v)
5697 if (v == PFM_PROC_SHOW_HEADER) {
5698 pfm_proc_show_header(m);
5702 /* show info for CPU (v - 1) */
5706 "CPU%-2d overflow intrs : %lu\n"
5707 "CPU%-2d overflow cycles : %lu\n"
5708 "CPU%-2d overflow min : %lu\n"
5709 "CPU%-2d overflow max : %lu\n"
5710 "CPU%-2d smpl handler calls : %lu\n"
5711 "CPU%-2d smpl handler cycles : %lu\n"
5712 "CPU%-2d spurious intrs : %lu\n"
5713 "CPU%-2d replay intrs : %lu\n"
5714 "CPU%-2d syst_wide : %d\n"
5715 "CPU%-2d dcr_pp : %d\n"
5716 "CPU%-2d exclude idle : %d\n"
5717 "CPU%-2d owner : %d\n"
5718 "CPU%-2d context : %p\n"
5719 "CPU%-2d activations : %lu\n",
5720 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5721 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5722 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5723 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5724 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5725 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5726 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5727 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5728 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5729 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5730 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5731 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5732 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5733 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5735 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5737 psr = pfm_get_psr();
5742 "CPU%-2d psr : 0x%lx\n"
5743 "CPU%-2d pmc0 : 0x%lx\n",
5745 cpu, ia64_get_pmc(0));
5747 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5748 if (PMC_IS_COUNTING(i) == 0) continue;
5750 "CPU%-2d pmc%u : 0x%lx\n"
5751 "CPU%-2d pmd%u : 0x%lx\n",
5752 cpu, i, ia64_get_pmc(i),
5753 cpu, i, ia64_get_pmd(i));
5759 struct seq_operations pfm_seq_ops = {
5760 .start = pfm_proc_start,
5761 .next = pfm_proc_next,
5762 .stop = pfm_proc_stop,
5763 .show = pfm_proc_show
5767 pfm_proc_open(struct inode *inode, struct file *file)
5769 return seq_open(file, &pfm_seq_ops);
5774 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5775 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5776 * is active or inactive based on mode. We must rely on the value in
5777 * local_cpu_data->pfm_syst_info
5780 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5782 struct pt_regs *regs;
5784 unsigned long dcr_pp;
5786 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5789 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5790 * on every CPU, so we can rely on the pid to identify the idle task.
5792 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5793 regs = ia64_task_regs(task);
5794 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5798 * if monitoring has started
5801 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5803 * context switching in?
5806 /* mask monitoring for the idle task */
5807 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5813 * context switching out
5814 * restore monitoring for next task
5816 * Due to inlining this odd if-then-else construction generates
5819 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5828 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5830 struct task_struct *task = ctx->ctx_task;
5832 ia64_psr(regs)->up = 0;
5833 ia64_psr(regs)->sp = 1;
5835 if (GET_PMU_OWNER() == task) {
5836 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5837 SET_PMU_OWNER(NULL, NULL);
5841 * disconnect the task from the context and vice-versa
5843 PFM_SET_WORK_PENDING(task, 0);
5845 task->thread.pfm_context = NULL;
5846 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5848 DPRINT(("force cleanup for [%d]\n", task->pid));
5853 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5856 pfm_save_regs(struct task_struct *task)
5859 struct thread_struct *t;
5860 unsigned long flags;
5864 ctx = PFM_GET_CTX(task);
5865 if (ctx == NULL) return;
5869 * we always come here with interrupts ALREADY disabled by
5870 * the scheduler. So we simply need to protect against concurrent
5871 * access, not CPU concurrency.
5873 flags = pfm_protect_ctx_ctxsw(ctx);
5875 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5876 struct pt_regs *regs = ia64_task_regs(task);
5880 pfm_force_cleanup(ctx, regs);
5882 BUG_ON(ctx->ctx_smpl_hdr);
5884 pfm_unprotect_ctx_ctxsw(ctx, flags);
5886 pfm_context_free(ctx);
5891 * save current PSR: needed because we modify it
5894 psr = pfm_get_psr();
5896 BUG_ON(psr & (IA64_PSR_I));
5900 * This is the last instruction which may generate an overflow
5902 * We do not need to set psr.sp because, it is irrelevant in kernel.
5903 * It will be restored from ipsr when going back to user level
5908 * keep a copy of psr.up (for reload)
5910 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5913 * release ownership of this PMU.
5914 * PM interrupts are masked, so nothing
5917 SET_PMU_OWNER(NULL, NULL);
5920 * we systematically save the PMD as we have no
5921 * guarantee we will be schedule at that same
5924 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5927 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5928 * we will need it on the restore path to check
5929 * for pending overflow.
5931 t->pmcs[0] = ia64_get_pmc(0);
5934 * unfreeze PMU if had pending overflows
5936 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5939 * finally, allow context access.
5940 * interrupts will still be masked after this call.
5942 pfm_unprotect_ctx_ctxsw(ctx, flags);
5945 #else /* !CONFIG_SMP */
5947 pfm_save_regs(struct task_struct *task)
5952 ctx = PFM_GET_CTX(task);
5953 if (ctx == NULL) return;
5956 * save current PSR: needed because we modify it
5958 psr = pfm_get_psr();
5960 BUG_ON(psr & (IA64_PSR_I));
5964 * This is the last instruction which may generate an overflow
5966 * We do not need to set psr.sp because, it is irrelevant in kernel.
5967 * It will be restored from ipsr when going back to user level
5972 * keep a copy of psr.up (for reload)
5974 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5978 pfm_lazy_save_regs (struct task_struct *task)
5981 struct thread_struct *t;
5982 unsigned long flags;
5984 { u64 psr = pfm_get_psr();
5985 BUG_ON(psr & IA64_PSR_UP);
5988 ctx = PFM_GET_CTX(task);
5992 * we need to mask PMU overflow here to
5993 * make sure that we maintain pmc0 until
5994 * we save it. overflow interrupts are
5995 * treated as spurious if there is no
5998 * XXX: I don't think this is necessary
6000 PROTECT_CTX(ctx,flags);
6003 * release ownership of this PMU.
6004 * must be done before we save the registers.
6006 * after this call any PMU interrupt is treated
6009 SET_PMU_OWNER(NULL, NULL);
6012 * save all the pmds we use
6014 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
6017 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6018 * it is needed to check for pended overflow
6019 * on the restore path
6021 t->pmcs[0] = ia64_get_pmc(0);
6024 * unfreeze PMU if had pending overflows
6026 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6029 * now get can unmask PMU interrupts, they will
6030 * be treated as purely spurious and we will not
6031 * lose any information
6033 UNPROTECT_CTX(ctx,flags);
6035 #endif /* CONFIG_SMP */
6039 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6042 pfm_load_regs (struct task_struct *task)
6045 struct thread_struct *t;
6046 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6047 unsigned long flags;
6049 int need_irq_resend;
6051 ctx = PFM_GET_CTX(task);
6052 if (unlikely(ctx == NULL)) return;
6054 BUG_ON(GET_PMU_OWNER());
6058 * possible on unload
6060 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6063 * we always come here with interrupts ALREADY disabled by
6064 * the scheduler. So we simply need to protect against concurrent
6065 * access, not CPU concurrency.
6067 flags = pfm_protect_ctx_ctxsw(ctx);
6068 psr = pfm_get_psr();
6070 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6072 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6073 BUG_ON(psr & IA64_PSR_I);
6075 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6076 struct pt_regs *regs = ia64_task_regs(task);
6078 BUG_ON(ctx->ctx_smpl_hdr);
6080 pfm_force_cleanup(ctx, regs);
6082 pfm_unprotect_ctx_ctxsw(ctx, flags);
6085 * this one (kmalloc'ed) is fine with interrupts disabled
6087 pfm_context_free(ctx);
6093 * we restore ALL the debug registers to avoid picking up
6096 if (ctx->ctx_fl_using_dbreg) {
6097 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6098 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6101 * retrieve saved psr.up
6103 psr_up = ctx->ctx_saved_psr_up;
6106 * if we were the last user of the PMU on that CPU,
6107 * then nothing to do except restore psr
6109 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6112 * retrieve partial reload masks (due to user modifications)
6114 pmc_mask = ctx->ctx_reload_pmcs[0];
6115 pmd_mask = ctx->ctx_reload_pmds[0];
6119 * To avoid leaking information to the user level when psr.sp=0,
6120 * we must reload ALL implemented pmds (even the ones we don't use).
6121 * In the kernel we only allow PFM_READ_PMDS on registers which
6122 * we initialized or requested (sampling) so there is no risk there.
6124 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6127 * ALL accessible PMCs are systematically reloaded, unused registers
6128 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6129 * up stale configuration.
6131 * PMC0 is never in the mask. It is always restored separately.
6133 pmc_mask = ctx->ctx_all_pmcs[0];
6136 * when context is MASKED, we will restore PMC with plm=0
6137 * and PMD with stale information, but that's ok, nothing
6140 * XXX: optimize here
6142 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6143 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6146 * check for pending overflow at the time the state
6149 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6151 * reload pmc0 with the overflow information
6152 * On McKinley PMU, this will trigger a PMU interrupt
6154 ia64_set_pmc(0, t->pmcs[0]);
6159 * will replay the PMU interrupt
6161 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6163 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6167 * we just did a reload, so we reset the partial reload fields
6169 ctx->ctx_reload_pmcs[0] = 0UL;
6170 ctx->ctx_reload_pmds[0] = 0UL;
6172 SET_LAST_CPU(ctx, smp_processor_id());
6175 * dump activation value for this PMU
6179 * record current activation for this context
6181 SET_ACTIVATION(ctx);
6184 * establish new ownership.
6186 SET_PMU_OWNER(task, ctx);
6189 * restore the psr.up bit. measurement
6191 * no PMU interrupt can happen at this point
6192 * because we still have interrupts disabled.
6194 if (likely(psr_up)) pfm_set_psr_up();
6197 * allow concurrent access to context
6199 pfm_unprotect_ctx_ctxsw(ctx, flags);
6201 #else /* !CONFIG_SMP */
6203 * reload PMU state for UP kernels
6204 * in 2.5 we come here with interrupts disabled
6207 pfm_load_regs (struct task_struct *task)
6209 struct thread_struct *t;
6211 struct task_struct *owner;
6212 unsigned long pmd_mask, pmc_mask;
6214 int need_irq_resend;
6216 owner = GET_PMU_OWNER();
6217 ctx = PFM_GET_CTX(task);
6219 psr = pfm_get_psr();
6221 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6222 BUG_ON(psr & IA64_PSR_I);
6225 * we restore ALL the debug registers to avoid picking up
6228 * This must be done even when the task is still the owner
6229 * as the registers may have been modified via ptrace()
6230 * (not perfmon) by the previous task.
6232 if (ctx->ctx_fl_using_dbreg) {
6233 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6234 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6238 * retrieved saved psr.up
6240 psr_up = ctx->ctx_saved_psr_up;
6241 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6244 * short path, our state is still there, just
6245 * need to restore psr and we go
6247 * we do not touch either PMC nor PMD. the psr is not touched
6248 * by the overflow_handler. So we are safe w.r.t. to interrupt
6249 * concurrency even without interrupt masking.
6251 if (likely(owner == task)) {
6252 if (likely(psr_up)) pfm_set_psr_up();
6257 * someone else is still using the PMU, first push it out and
6258 * then we'll be able to install our stuff !
6260 * Upon return, there will be no owner for the current PMU
6262 if (owner) pfm_lazy_save_regs(owner);
6265 * To avoid leaking information to the user level when psr.sp=0,
6266 * we must reload ALL implemented pmds (even the ones we don't use).
6267 * In the kernel we only allow PFM_READ_PMDS on registers which
6268 * we initialized or requested (sampling) so there is no risk there.
6270 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6273 * ALL accessible PMCs are systematically reloaded, unused registers
6274 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6275 * up stale configuration.
6277 * PMC0 is never in the mask. It is always restored separately
6279 pmc_mask = ctx->ctx_all_pmcs[0];
6281 pfm_restore_pmds(t->pmds, pmd_mask);
6282 pfm_restore_pmcs(t->pmcs, pmc_mask);
6285 * check for pending overflow at the time the state
6288 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6290 * reload pmc0 with the overflow information
6291 * On McKinley PMU, this will trigger a PMU interrupt
6293 ia64_set_pmc(0, t->pmcs[0]);
6299 * will replay the PMU interrupt
6301 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6303 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6307 * establish new ownership.
6309 SET_PMU_OWNER(task, ctx);
6312 * restore the psr.up bit. measurement
6314 * no PMU interrupt can happen at this point
6315 * because we still have interrupts disabled.
6317 if (likely(psr_up)) pfm_set_psr_up();
6319 #endif /* CONFIG_SMP */
6322 * this function assumes monitoring is stopped
6325 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6328 unsigned long mask2, val, pmd_val, ovfl_val;
6329 int i, can_access_pmu = 0;
6333 * is the caller the task being monitored (or which initiated the
6334 * session for system wide measurements)
6336 is_self = ctx->ctx_task == task ? 1 : 0;
6339 * can access PMU is task is the owner of the PMU state on the current CPU
6340 * or if we are running on the CPU bound to the context in system-wide mode
6341 * (that is not necessarily the task the context is attached to in this mode).
6342 * In system-wide we always have can_access_pmu true because a task running on an
6343 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6345 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6346 if (can_access_pmu) {
6348 * Mark the PMU as not owned
6349 * This will cause the interrupt handler to do nothing in case an overflow
6350 * interrupt was in-flight
6351 * This also guarantees that pmc0 will contain the final state
6352 * It virtually gives us full control on overflow processing from that point
6355 SET_PMU_OWNER(NULL, NULL);
6356 DPRINT(("releasing ownership\n"));
6359 * read current overflow status:
6361 * we are guaranteed to read the final stable state
6364 pmc0 = ia64_get_pmc(0); /* slow */
6367 * reset freeze bit, overflow status information destroyed
6371 pmc0 = task->thread.pmcs[0];
6373 * clear whatever overflow status bits there were
6375 task->thread.pmcs[0] = 0;
6377 ovfl_val = pmu_conf->ovfl_val;
6379 * we save all the used pmds
6380 * we take care of overflows for counting PMDs
6382 * XXX: sampling situation is not taken into account here
6384 mask2 = ctx->ctx_used_pmds[0];
6386 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6388 for (i = 0; mask2; i++, mask2>>=1) {
6390 /* skip non used pmds */
6391 if ((mask2 & 0x1) == 0) continue;
6394 * can access PMU always true in system wide mode
6396 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6398 if (PMD_IS_COUNTING(i)) {
6399 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6402 ctx->ctx_pmds[i].val,
6406 * we rebuild the full 64 bit value of the counter
6408 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6411 * now everything is in ctx_pmds[] and we need
6412 * to clear the saved context from save_regs() such that
6413 * pfm_read_pmds() gets the correct value
6418 * take care of overflow inline
6420 if (pmc0 & (1UL << i)) {
6421 val += 1 + ovfl_val;
6422 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6426 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, i, val, pmd_val));
6428 if (is_self) task->thread.pmds[i] = pmd_val;
6430 ctx->ctx_pmds[i].val = val;
6434 static struct irqaction perfmon_irqaction = {
6435 .handler = pfm_interrupt_handler,
6436 .flags = SA_INTERRUPT,
6441 pfm_alt_save_pmu_state(void *data)
6443 struct pt_regs *regs;
6445 regs = ia64_task_regs(current);
6447 DPRINT(("called\n"));
6450 * should not be necessary but
6451 * let's take not risk
6455 ia64_psr(regs)->pp = 0;
6458 * This call is required
6459 * May cause a spurious interrupt on some processors
6467 pfm_alt_restore_pmu_state(void *data)
6469 struct pt_regs *regs;
6471 regs = ia64_task_regs(current);
6473 DPRINT(("called\n"));
6476 * put PMU back in state expected
6481 ia64_psr(regs)->pp = 0;
6484 * perfmon runs with PMU unfrozen at all times
6492 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6497 /* some sanity checks */
6498 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6500 /* do the easy test first */
6501 if (pfm_alt_intr_handler) return -EBUSY;
6503 /* one at a time in the install or remove, just fail the others */
6504 if (!spin_trylock(&pfm_alt_install_check)) {
6508 /* reserve our session */
6509 for_each_online_cpu(reserve_cpu) {
6510 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6511 if (ret) goto cleanup_reserve;
6514 /* save the current system wide pmu states */
6515 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 0, 1);
6517 DPRINT(("on_each_cpu() failed: %d\n", ret));
6518 goto cleanup_reserve;
6521 /* officially change to the alternate interrupt handler */
6522 pfm_alt_intr_handler = hdl;
6524 spin_unlock(&pfm_alt_install_check);
6529 for_each_online_cpu(i) {
6530 /* don't unreserve more than we reserved */
6531 if (i >= reserve_cpu) break;
6533 pfm_unreserve_session(NULL, 1, i);
6536 spin_unlock(&pfm_alt_install_check);
6540 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6543 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6548 if (hdl == NULL) return -EINVAL;
6550 /* cannot remove someone else's handler! */
6551 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6553 /* one at a time in the install or remove, just fail the others */
6554 if (!spin_trylock(&pfm_alt_install_check)) {
6558 pfm_alt_intr_handler = NULL;
6560 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 0, 1);
6562 DPRINT(("on_each_cpu() failed: %d\n", ret));
6565 for_each_online_cpu(i) {
6566 pfm_unreserve_session(NULL, 1, i);
6569 spin_unlock(&pfm_alt_install_check);
6573 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6576 * perfmon initialization routine, called from the initcall() table
6578 static int init_pfm_fs(void);
6586 family = local_cpu_data->family;
6591 if ((*p)->probe() == 0) goto found;
6592 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6603 static struct file_operations pfm_proc_fops = {
6604 .open = pfm_proc_open,
6606 .llseek = seq_lseek,
6607 .release = seq_release,
6613 unsigned int n, n_counters, i;
6615 printk("perfmon: version %u.%u IRQ %u\n",
6618 IA64_PERFMON_VECTOR);
6620 if (pfm_probe_pmu()) {
6621 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6622 local_cpu_data->family);
6627 * compute the number of implemented PMD/PMC from the
6628 * description tables
6631 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6632 if (PMC_IS_IMPL(i) == 0) continue;
6633 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6636 pmu_conf->num_pmcs = n;
6638 n = 0; n_counters = 0;
6639 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6640 if (PMD_IS_IMPL(i) == 0) continue;
6641 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6643 if (PMD_IS_COUNTING(i)) n_counters++;
6645 pmu_conf->num_pmds = n;
6646 pmu_conf->num_counters = n_counters;
6649 * sanity checks on the number of debug registers
6651 if (pmu_conf->use_rr_dbregs) {
6652 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6653 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6657 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6658 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6664 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6668 pmu_conf->num_counters,
6669 ffz(pmu_conf->ovfl_val));
6672 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6673 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6679 * create /proc/perfmon (mostly for debugging purposes)
6681 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6682 if (perfmon_dir == NULL) {
6683 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6688 * install customized file operations for /proc/perfmon entry
6690 perfmon_dir->proc_fops = &pfm_proc_fops;
6693 * create /proc/sys/kernel/perfmon (for debugging purposes)
6695 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6698 * initialize all our spinlocks
6700 spin_lock_init(&pfm_sessions.pfs_lock);
6701 spin_lock_init(&pfm_buffer_fmt_lock);
6705 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6710 __initcall(pfm_init);
6713 * this function is called before pfm_init()
6716 pfm_init_percpu (void)
6719 * make sure no measurement is active
6720 * (may inherit programmed PMCs from EFI).
6726 * we run with the PMU not frozen at all times
6730 if (smp_processor_id() == 0)
6731 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6733 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6738 * used for debug purposes only
6741 dump_pmu_state(const char *from)
6743 struct task_struct *task;
6744 struct thread_struct *t;
6745 struct pt_regs *regs;
6747 unsigned long psr, dcr, info, flags;
6750 local_irq_save(flags);
6752 this_cpu = smp_processor_id();
6753 regs = ia64_task_regs(current);
6754 info = PFM_CPUINFO_GET();
6755 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6757 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6758 local_irq_restore(flags);
6762 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6769 task = GET_PMU_OWNER();
6770 ctx = GET_PMU_CTX();
6772 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6774 psr = pfm_get_psr();
6776 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",
6779 psr & IA64_PSR_PP ? 1 : 0,
6780 psr & IA64_PSR_UP ? 1 : 0,
6781 dcr & IA64_DCR_PP ? 1 : 0,
6784 ia64_psr(regs)->pp);
6786 ia64_psr(regs)->up = 0;
6787 ia64_psr(regs)->pp = 0;
6789 t = ¤t->thread;
6791 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6792 if (PMC_IS_IMPL(i) == 0) continue;
6793 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6796 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6797 if (PMD_IS_IMPL(i) == 0) continue;
6798 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6802 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6805 ctx->ctx_smpl_vaddr,
6809 ctx->ctx_saved_psr_up);
6811 local_irq_restore(flags);
6815 * called from process.c:copy_thread(). task is new child.
6818 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6820 struct thread_struct *thread;
6822 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6824 thread = &task->thread;
6827 * cut links inherited from parent (current)
6829 thread->pfm_context = NULL;
6831 PFM_SET_WORK_PENDING(task, 0);
6834 * the psr bits are already set properly in copy_threads()
6837 #else /* !CONFIG_PERFMON */
6839 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6843 #endif /* CONFIG_PERFMON */