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/bitops.h>
41 #include <linux/capability.h>
42 #include <linux/rcupdate.h>
43 #include <linux/completion.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 completion ctx_restart_done; /* 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 DEFINE_SPINLOCK(pfm_alt_install_check);
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);
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);
599 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data,
600 struct vfsmount *mnt)
602 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
605 static struct file_system_type pfm_fs_type = {
607 .get_sb = pfmfs_get_sb,
608 .kill_sb = kill_anon_super,
611 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
612 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
613 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
614 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
615 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
618 /* forward declaration */
619 static struct file_operations pfm_file_ops;
622 * forward declarations
625 static void pfm_lazy_save_regs (struct task_struct *ta);
628 void dump_pmu_state(const char *);
629 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
631 #include "perfmon_itanium.h"
632 #include "perfmon_mckinley.h"
633 #include "perfmon_montecito.h"
634 #include "perfmon_generic.h"
636 static pmu_config_t *pmu_confs[]={
640 &pmu_conf_gen, /* must be last */
645 static int pfm_end_notify_user(pfm_context_t *ctx);
648 pfm_clear_psr_pp(void)
650 ia64_rsm(IA64_PSR_PP);
657 ia64_ssm(IA64_PSR_PP);
662 pfm_clear_psr_up(void)
664 ia64_rsm(IA64_PSR_UP);
671 ia64_ssm(IA64_PSR_UP);
675 static inline unsigned long
679 tmp = ia64_getreg(_IA64_REG_PSR);
685 pfm_set_psr_l(unsigned long val)
687 ia64_setreg(_IA64_REG_PSR_L, val);
699 pfm_unfreeze_pmu(void)
706 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
710 for (i=0; i < nibrs; i++) {
711 ia64_set_ibr(i, ibrs[i]);
712 ia64_dv_serialize_instruction();
718 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
722 for (i=0; i < ndbrs; i++) {
723 ia64_set_dbr(i, dbrs[i]);
724 ia64_dv_serialize_data();
730 * PMD[i] must be a counter. no check is made
732 static inline unsigned long
733 pfm_read_soft_counter(pfm_context_t *ctx, int i)
735 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
739 * PMD[i] must be a counter. no check is made
742 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
744 unsigned long ovfl_val = pmu_conf->ovfl_val;
746 ctx->ctx_pmds[i].val = val & ~ovfl_val;
748 * writing to unimplemented part is ignore, so we do not need to
751 ia64_set_pmd(i, val & ovfl_val);
755 pfm_get_new_msg(pfm_context_t *ctx)
759 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
761 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
762 if (next == ctx->ctx_msgq_head) return NULL;
764 idx = ctx->ctx_msgq_tail;
765 ctx->ctx_msgq_tail = next;
767 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
769 return ctx->ctx_msgq+idx;
773 pfm_get_next_msg(pfm_context_t *ctx)
777 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
779 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
784 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
789 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
791 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));
797 pfm_reset_msgq(pfm_context_t *ctx)
799 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
800 DPRINT(("ctx=%p msgq reset\n", ctx));
804 pfm_rvmalloc(unsigned long size)
809 size = PAGE_ALIGN(size);
812 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
813 memset(mem, 0, size);
814 addr = (unsigned long)mem;
816 pfm_reserve_page(addr);
825 pfm_rvfree(void *mem, unsigned long size)
830 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
831 addr = (unsigned long) mem;
832 while ((long) size > 0) {
833 pfm_unreserve_page(addr);
842 static pfm_context_t *
843 pfm_context_alloc(void)
848 * allocate context descriptor
849 * must be able to free with interrupts disabled
851 ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
853 memset(ctx, 0, sizeof(pfm_context_t));
854 DPRINT(("alloc ctx @%p\n", ctx));
860 pfm_context_free(pfm_context_t *ctx)
863 DPRINT(("free ctx @%p\n", ctx));
869 pfm_mask_monitoring(struct task_struct *task)
871 pfm_context_t *ctx = PFM_GET_CTX(task);
872 struct thread_struct *th = &task->thread;
873 unsigned long mask, val, ovfl_mask;
876 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
878 ovfl_mask = pmu_conf->ovfl_val;
880 * monitoring can only be masked as a result of a valid
881 * counter overflow. In UP, it means that the PMU still
882 * has an owner. Note that the owner can be different
883 * from the current task. However the PMU state belongs
885 * In SMP, a valid overflow only happens when task is
886 * current. Therefore if we come here, we know that
887 * the PMU state belongs to the current task, therefore
888 * we can access the live registers.
890 * So in both cases, the live register contains the owner's
891 * state. We can ONLY touch the PMU registers and NOT the PSR.
893 * As a consequence to this call, the thread->pmds[] array
894 * contains stale information which must be ignored
895 * when context is reloaded AND monitoring is active (see
898 mask = ctx->ctx_used_pmds[0];
899 for (i = 0; mask; i++, mask>>=1) {
900 /* skip non used pmds */
901 if ((mask & 0x1) == 0) continue;
902 val = ia64_get_pmd(i);
904 if (PMD_IS_COUNTING(i)) {
906 * we rebuild the full 64 bit value of the counter
908 ctx->ctx_pmds[i].val += (val & ovfl_mask);
910 ctx->ctx_pmds[i].val = val;
912 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
914 ctx->ctx_pmds[i].val,
918 * mask monitoring by setting the privilege level to 0
919 * we cannot use psr.pp/psr.up for this, it is controlled by
922 * if task is current, modify actual registers, otherwise modify
923 * thread save state, i.e., what will be restored in pfm_load_regs()
925 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
926 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
927 if ((mask & 0x1) == 0UL) continue;
928 ia64_set_pmc(i, th->pmcs[i] & ~0xfUL);
929 th->pmcs[i] &= ~0xfUL;
930 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i]));
933 * make all of this visible
939 * must always be done with task == current
941 * context must be in MASKED state when calling
944 pfm_restore_monitoring(struct task_struct *task)
946 pfm_context_t *ctx = PFM_GET_CTX(task);
947 struct thread_struct *th = &task->thread;
948 unsigned long mask, ovfl_mask;
949 unsigned long psr, val;
952 is_system = ctx->ctx_fl_system;
953 ovfl_mask = pmu_conf->ovfl_val;
955 if (task != current) {
956 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
959 if (ctx->ctx_state != PFM_CTX_MASKED) {
960 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
961 task->pid, current->pid, ctx->ctx_state);
966 * monitoring is masked via the PMC.
967 * As we restore their value, we do not want each counter to
968 * restart right away. We stop monitoring using the PSR,
969 * restore the PMC (and PMD) and then re-establish the psr
970 * as it was. Note that there can be no pending overflow at
971 * this point, because monitoring was MASKED.
973 * system-wide session are pinned and self-monitoring
975 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
977 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
983 * first, we restore the PMD
985 mask = ctx->ctx_used_pmds[0];
986 for (i = 0; mask; i++, mask>>=1) {
987 /* skip non used pmds */
988 if ((mask & 0x1) == 0) continue;
990 if (PMD_IS_COUNTING(i)) {
992 * we split the 64bit value according to
995 val = ctx->ctx_pmds[i].val & ovfl_mask;
996 ctx->ctx_pmds[i].val &= ~ovfl_mask;
998 val = ctx->ctx_pmds[i].val;
1000 ia64_set_pmd(i, val);
1002 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1004 ctx->ctx_pmds[i].val,
1010 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1011 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1012 if ((mask & 0x1) == 0UL) continue;
1013 th->pmcs[i] = ctx->ctx_pmcs[i];
1014 ia64_set_pmc(i, th->pmcs[i]);
1015 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
1020 * must restore DBR/IBR because could be modified while masked
1021 * XXX: need to optimize
1023 if (ctx->ctx_fl_using_dbreg) {
1024 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1025 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1031 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1033 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1040 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1046 for (i=0; mask; i++, mask>>=1) {
1047 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1052 * reload from thread state (used for ctxw only)
1055 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1058 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1060 for (i=0; mask; i++, mask>>=1) {
1061 if ((mask & 0x1) == 0) continue;
1062 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1063 ia64_set_pmd(i, val);
1069 * propagate PMD from context to thread-state
1072 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1074 struct thread_struct *thread = &task->thread;
1075 unsigned long ovfl_val = pmu_conf->ovfl_val;
1076 unsigned long mask = ctx->ctx_all_pmds[0];
1080 DPRINT(("mask=0x%lx\n", mask));
1082 for (i=0; mask; i++, mask>>=1) {
1084 val = ctx->ctx_pmds[i].val;
1087 * We break up the 64 bit value into 2 pieces
1088 * the lower bits go to the machine state in the
1089 * thread (will be reloaded on ctxsw in).
1090 * The upper part stays in the soft-counter.
1092 if (PMD_IS_COUNTING(i)) {
1093 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1096 thread->pmds[i] = val;
1098 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1101 ctx->ctx_pmds[i].val));
1106 * propagate PMC from context to thread-state
1109 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1111 struct thread_struct *thread = &task->thread;
1112 unsigned long mask = ctx->ctx_all_pmcs[0];
1115 DPRINT(("mask=0x%lx\n", mask));
1117 for (i=0; mask; i++, mask>>=1) {
1118 /* masking 0 with ovfl_val yields 0 */
1119 thread->pmcs[i] = ctx->ctx_pmcs[i];
1120 DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
1127 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1131 for (i=0; mask; i++, mask>>=1) {
1132 if ((mask & 0x1) == 0) continue;
1133 ia64_set_pmc(i, pmcs[i]);
1139 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1141 return memcmp(a, b, sizeof(pfm_uuid_t));
1145 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1148 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1153 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1156 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1162 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1166 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1171 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1175 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1180 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1183 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1188 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)
1191 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1195 static pfm_buffer_fmt_t *
1196 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1198 struct list_head * pos;
1199 pfm_buffer_fmt_t * entry;
1201 list_for_each(pos, &pfm_buffer_fmt_list) {
1202 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1203 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1210 * find a buffer format based on its uuid
1212 static pfm_buffer_fmt_t *
1213 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1215 pfm_buffer_fmt_t * fmt;
1216 spin_lock(&pfm_buffer_fmt_lock);
1217 fmt = __pfm_find_buffer_fmt(uuid);
1218 spin_unlock(&pfm_buffer_fmt_lock);
1223 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1227 /* some sanity checks */
1228 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1230 /* we need at least a handler */
1231 if (fmt->fmt_handler == NULL) return -EINVAL;
1234 * XXX: need check validity of fmt_arg_size
1237 spin_lock(&pfm_buffer_fmt_lock);
1239 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1240 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1244 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1245 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1248 spin_unlock(&pfm_buffer_fmt_lock);
1251 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1254 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1256 pfm_buffer_fmt_t *fmt;
1259 spin_lock(&pfm_buffer_fmt_lock);
1261 fmt = __pfm_find_buffer_fmt(uuid);
1263 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1267 list_del_init(&fmt->fmt_list);
1268 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1271 spin_unlock(&pfm_buffer_fmt_lock);
1275 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1277 extern void update_pal_halt_status(int);
1280 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1282 unsigned long flags;
1284 * validy checks on cpu_mask have been done upstream
1288 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1289 pfm_sessions.pfs_sys_sessions,
1290 pfm_sessions.pfs_task_sessions,
1291 pfm_sessions.pfs_sys_use_dbregs,
1297 * cannot mix system wide and per-task sessions
1299 if (pfm_sessions.pfs_task_sessions > 0UL) {
1300 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1301 pfm_sessions.pfs_task_sessions));
1305 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1307 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1309 pfm_sessions.pfs_sys_session[cpu] = task;
1311 pfm_sessions.pfs_sys_sessions++ ;
1314 if (pfm_sessions.pfs_sys_sessions) goto abort;
1315 pfm_sessions.pfs_task_sessions++;
1318 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1319 pfm_sessions.pfs_sys_sessions,
1320 pfm_sessions.pfs_task_sessions,
1321 pfm_sessions.pfs_sys_use_dbregs,
1326 * disable default_idle() to go to PAL_HALT
1328 update_pal_halt_status(0);
1335 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1336 pfm_sessions.pfs_sys_session[cpu]->pid,
1346 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1348 unsigned long flags;
1350 * validy checks on cpu_mask have been done upstream
1354 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1355 pfm_sessions.pfs_sys_sessions,
1356 pfm_sessions.pfs_task_sessions,
1357 pfm_sessions.pfs_sys_use_dbregs,
1363 pfm_sessions.pfs_sys_session[cpu] = NULL;
1365 * would not work with perfmon+more than one bit in cpu_mask
1367 if (ctx && ctx->ctx_fl_using_dbreg) {
1368 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1369 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1371 pfm_sessions.pfs_sys_use_dbregs--;
1374 pfm_sessions.pfs_sys_sessions--;
1376 pfm_sessions.pfs_task_sessions--;
1378 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1379 pfm_sessions.pfs_sys_sessions,
1380 pfm_sessions.pfs_task_sessions,
1381 pfm_sessions.pfs_sys_use_dbregs,
1386 * if possible, enable default_idle() to go into PAL_HALT
1388 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1389 update_pal_halt_status(1);
1397 * removes virtual mapping of the sampling buffer.
1398 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1399 * a PROTECT_CTX() section.
1402 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1407 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1408 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1412 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1415 * does the actual unmapping
1417 down_write(&task->mm->mmap_sem);
1419 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1421 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1423 up_write(&task->mm->mmap_sem);
1425 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1428 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1434 * free actual physical storage used by sampling buffer
1438 pfm_free_smpl_buffer(pfm_context_t *ctx)
1440 pfm_buffer_fmt_t *fmt;
1442 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1445 * we won't use the buffer format anymore
1447 fmt = ctx->ctx_buf_fmt;
1449 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1452 ctx->ctx_smpl_vaddr));
1454 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1459 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1461 ctx->ctx_smpl_hdr = NULL;
1462 ctx->ctx_smpl_size = 0UL;
1467 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1473 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1475 if (fmt == NULL) return;
1477 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1482 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1483 * no real gain from having the whole whorehouse mounted. So we don't need
1484 * any operations on the root directory. However, we need a non-trivial
1485 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1487 static struct vfsmount *pfmfs_mnt;
1492 int err = register_filesystem(&pfm_fs_type);
1494 pfmfs_mnt = kern_mount(&pfm_fs_type);
1495 err = PTR_ERR(pfmfs_mnt);
1496 if (IS_ERR(pfmfs_mnt))
1497 unregister_filesystem(&pfm_fs_type);
1507 unregister_filesystem(&pfm_fs_type);
1512 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1517 unsigned long flags;
1518 DECLARE_WAITQUEUE(wait, current);
1519 if (PFM_IS_FILE(filp) == 0) {
1520 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1524 ctx = (pfm_context_t *)filp->private_data;
1526 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1531 * check even when there is no message
1533 if (size < sizeof(pfm_msg_t)) {
1534 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1538 PROTECT_CTX(ctx, flags);
1541 * put ourselves on the wait queue
1543 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1551 set_current_state(TASK_INTERRUPTIBLE);
1553 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1556 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1558 UNPROTECT_CTX(ctx, flags);
1561 * check non-blocking read
1564 if(filp->f_flags & O_NONBLOCK) break;
1567 * check pending signals
1569 if(signal_pending(current)) {
1574 * no message, so wait
1578 PROTECT_CTX(ctx, flags);
1580 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1581 set_current_state(TASK_RUNNING);
1582 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1584 if (ret < 0) goto abort;
1587 msg = pfm_get_next_msg(ctx);
1589 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1593 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1596 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1599 UNPROTECT_CTX(ctx, flags);
1605 pfm_write(struct file *file, const char __user *ubuf,
1606 size_t size, loff_t *ppos)
1608 DPRINT(("pfm_write called\n"));
1613 pfm_poll(struct file *filp, poll_table * wait)
1616 unsigned long flags;
1617 unsigned int mask = 0;
1619 if (PFM_IS_FILE(filp) == 0) {
1620 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1624 ctx = (pfm_context_t *)filp->private_data;
1626 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1631 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1633 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1635 PROTECT_CTX(ctx, flags);
1637 if (PFM_CTXQ_EMPTY(ctx) == 0)
1638 mask = POLLIN | POLLRDNORM;
1640 UNPROTECT_CTX(ctx, flags);
1642 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1648 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1650 DPRINT(("pfm_ioctl called\n"));
1655 * interrupt cannot be masked when coming here
1658 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1662 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1664 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1668 ctx->ctx_async_queue, ret));
1674 pfm_fasync(int fd, struct file *filp, int on)
1679 if (PFM_IS_FILE(filp) == 0) {
1680 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1684 ctx = (pfm_context_t *)filp->private_data;
1686 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1690 * we cannot mask interrupts during this call because this may
1691 * may go to sleep if memory is not readily avalaible.
1693 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1694 * done in caller. Serialization of this function is ensured by caller.
1696 ret = pfm_do_fasync(fd, filp, ctx, on);
1699 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1702 ctx->ctx_async_queue, ret));
1709 * this function is exclusively called from pfm_close().
1710 * The context is not protected at that time, nor are interrupts
1711 * on the remote CPU. That's necessary to avoid deadlocks.
1714 pfm_syswide_force_stop(void *info)
1716 pfm_context_t *ctx = (pfm_context_t *)info;
1717 struct pt_regs *regs = task_pt_regs(current);
1718 struct task_struct *owner;
1719 unsigned long flags;
1722 if (ctx->ctx_cpu != smp_processor_id()) {
1723 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1725 smp_processor_id());
1728 owner = GET_PMU_OWNER();
1729 if (owner != ctx->ctx_task) {
1730 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1732 owner->pid, ctx->ctx_task->pid);
1735 if (GET_PMU_CTX() != ctx) {
1736 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1738 GET_PMU_CTX(), ctx);
1742 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1744 * the context is already protected in pfm_close(), we simply
1745 * need to mask interrupts to avoid a PMU interrupt race on
1748 local_irq_save(flags);
1750 ret = pfm_context_unload(ctx, NULL, 0, regs);
1752 DPRINT(("context_unload returned %d\n", ret));
1756 * unmask interrupts, PMU interrupts are now spurious here
1758 local_irq_restore(flags);
1762 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1766 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1767 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1768 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1770 #endif /* CONFIG_SMP */
1773 * called for each close(). Partially free resources.
1774 * When caller is self-monitoring, the context is unloaded.
1777 pfm_flush(struct file *filp)
1780 struct task_struct *task;
1781 struct pt_regs *regs;
1782 unsigned long flags;
1783 unsigned long smpl_buf_size = 0UL;
1784 void *smpl_buf_vaddr = NULL;
1785 int state, is_system;
1787 if (PFM_IS_FILE(filp) == 0) {
1788 DPRINT(("bad magic for\n"));
1792 ctx = (pfm_context_t *)filp->private_data;
1794 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1799 * remove our file from the async queue, if we use this mode.
1800 * This can be done without the context being protected. We come
1801 * here when the context has become unreacheable by other tasks.
1803 * We may still have active monitoring at this point and we may
1804 * end up in pfm_overflow_handler(). However, fasync_helper()
1805 * operates with interrupts disabled and it cleans up the
1806 * queue. If the PMU handler is called prior to entering
1807 * fasync_helper() then it will send a signal. If it is
1808 * invoked after, it will find an empty queue and no
1809 * signal will be sent. In both case, we are safe
1811 if (filp->f_flags & FASYNC) {
1812 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1813 pfm_do_fasync (-1, filp, ctx, 0);
1816 PROTECT_CTX(ctx, flags);
1818 state = ctx->ctx_state;
1819 is_system = ctx->ctx_fl_system;
1821 task = PFM_CTX_TASK(ctx);
1822 regs = task_pt_regs(task);
1824 DPRINT(("ctx_state=%d is_current=%d\n",
1826 task == current ? 1 : 0));
1829 * if state == UNLOADED, then task is NULL
1833 * we must stop and unload because we are losing access to the context.
1835 if (task == current) {
1838 * the task IS the owner but it migrated to another CPU: that's bad
1839 * but we must handle this cleanly. Unfortunately, the kernel does
1840 * not provide a mechanism to block migration (while the context is loaded).
1842 * We need to release the resource on the ORIGINAL cpu.
1844 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1846 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1848 * keep context protected but unmask interrupt for IPI
1850 local_irq_restore(flags);
1852 pfm_syswide_cleanup_other_cpu(ctx);
1855 * restore interrupt masking
1857 local_irq_save(flags);
1860 * context is unloaded at this point
1863 #endif /* CONFIG_SMP */
1866 DPRINT(("forcing unload\n"));
1868 * stop and unload, returning with state UNLOADED
1869 * and session unreserved.
1871 pfm_context_unload(ctx, NULL, 0, regs);
1873 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1878 * remove virtual mapping, if any, for the calling task.
1879 * cannot reset ctx field until last user is calling close().
1881 * ctx_smpl_vaddr must never be cleared because it is needed
1882 * by every task with access to the context
1884 * When called from do_exit(), the mm context is gone already, therefore
1885 * mm is NULL, i.e., the VMA is already gone and we do not have to
1888 if (ctx->ctx_smpl_vaddr && current->mm) {
1889 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1890 smpl_buf_size = ctx->ctx_smpl_size;
1893 UNPROTECT_CTX(ctx, flags);
1896 * if there was a mapping, then we systematically remove it
1897 * at this point. Cannot be done inside critical section
1898 * because some VM function reenables interrupts.
1901 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1906 * called either on explicit close() or from exit_files().
1907 * Only the LAST user of the file gets to this point, i.e., it is
1910 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1911 * (fput()),i.e, last task to access the file. Nobody else can access the
1912 * file at this point.
1914 * When called from exit_files(), the VMA has been freed because exit_mm()
1915 * is executed before exit_files().
1917 * When called from exit_files(), the current task is not yet ZOMBIE but we
1918 * flush the PMU state to the context.
1921 pfm_close(struct inode *inode, struct file *filp)
1924 struct task_struct *task;
1925 struct pt_regs *regs;
1926 DECLARE_WAITQUEUE(wait, current);
1927 unsigned long flags;
1928 unsigned long smpl_buf_size = 0UL;
1929 void *smpl_buf_addr = NULL;
1930 int free_possible = 1;
1931 int state, is_system;
1933 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1935 if (PFM_IS_FILE(filp) == 0) {
1936 DPRINT(("bad magic\n"));
1940 ctx = (pfm_context_t *)filp->private_data;
1942 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1946 PROTECT_CTX(ctx, flags);
1948 state = ctx->ctx_state;
1949 is_system = ctx->ctx_fl_system;
1951 task = PFM_CTX_TASK(ctx);
1952 regs = task_pt_regs(task);
1954 DPRINT(("ctx_state=%d is_current=%d\n",
1956 task == current ? 1 : 0));
1959 * if task == current, then pfm_flush() unloaded the context
1961 if (state == PFM_CTX_UNLOADED) goto doit;
1964 * context is loaded/masked and task != current, we need to
1965 * either force an unload or go zombie
1969 * The task is currently blocked or will block after an overflow.
1970 * we must force it to wakeup to get out of the
1971 * MASKED state and transition to the unloaded state by itself.
1973 * This situation is only possible for per-task mode
1975 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1978 * set a "partial" zombie state to be checked
1979 * upon return from down() in pfm_handle_work().
1981 * We cannot use the ZOMBIE state, because it is checked
1982 * by pfm_load_regs() which is called upon wakeup from down().
1983 * In such case, it would free the context and then we would
1984 * return to pfm_handle_work() which would access the
1985 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1986 * but visible to pfm_handle_work().
1988 * For some window of time, we have a zombie context with
1989 * ctx_state = MASKED and not ZOMBIE
1991 ctx->ctx_fl_going_zombie = 1;
1994 * force task to wake up from MASKED state
1996 complete(&ctx->ctx_restart_done);
1998 DPRINT(("waking up ctx_state=%d\n", state));
2001 * put ourself to sleep waiting for the other
2002 * task to report completion
2004 * the context is protected by mutex, therefore there
2005 * is no risk of being notified of completion before
2006 * begin actually on the waitq.
2008 set_current_state(TASK_INTERRUPTIBLE);
2009 add_wait_queue(&ctx->ctx_zombieq, &wait);
2011 UNPROTECT_CTX(ctx, flags);
2014 * XXX: check for signals :
2015 * - ok for explicit close
2016 * - not ok when coming from exit_files()
2021 PROTECT_CTX(ctx, flags);
2024 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2025 set_current_state(TASK_RUNNING);
2028 * context is unloaded at this point
2030 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2032 else if (task != current) {
2035 * switch context to zombie state
2037 ctx->ctx_state = PFM_CTX_ZOMBIE;
2039 DPRINT(("zombie ctx for [%d]\n", task->pid));
2041 * cannot free the context on the spot. deferred until
2042 * the task notices the ZOMBIE state
2046 pfm_context_unload(ctx, NULL, 0, regs);
2051 /* reload state, may have changed during opening of critical section */
2052 state = ctx->ctx_state;
2055 * the context is still attached to a task (possibly current)
2056 * we cannot destroy it right now
2060 * we must free the sampling buffer right here because
2061 * we cannot rely on it being cleaned up later by the
2062 * monitored task. It is not possible to free vmalloc'ed
2063 * memory in pfm_load_regs(). Instead, we remove the buffer
2064 * now. should there be subsequent PMU overflow originally
2065 * meant for sampling, the will be converted to spurious
2066 * and that's fine because the monitoring tools is gone anyway.
2068 if (ctx->ctx_smpl_hdr) {
2069 smpl_buf_addr = ctx->ctx_smpl_hdr;
2070 smpl_buf_size = ctx->ctx_smpl_size;
2071 /* no more sampling */
2072 ctx->ctx_smpl_hdr = NULL;
2073 ctx->ctx_fl_is_sampling = 0;
2076 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2082 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2085 * UNLOADED that the session has already been unreserved.
2087 if (state == PFM_CTX_ZOMBIE) {
2088 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2092 * disconnect file descriptor from context must be done
2095 filp->private_data = NULL;
2098 * if we free on the spot, the context is now completely unreacheable
2099 * from the callers side. The monitored task side is also cut, so we
2102 * If we have a deferred free, only the caller side is disconnected.
2104 UNPROTECT_CTX(ctx, flags);
2107 * All memory free operations (especially for vmalloc'ed memory)
2108 * MUST be done with interrupts ENABLED.
2110 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2113 * return the memory used by the context
2115 if (free_possible) pfm_context_free(ctx);
2121 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2123 DPRINT(("pfm_no_open called\n"));
2129 static struct file_operations pfm_file_ops = {
2130 .llseek = no_llseek,
2135 .open = pfm_no_open, /* special open code to disallow open via /proc */
2136 .fasync = pfm_fasync,
2137 .release = pfm_close,
2142 pfmfs_delete_dentry(struct dentry *dentry)
2147 static struct dentry_operations pfmfs_dentry_operations = {
2148 .d_delete = pfmfs_delete_dentry,
2153 pfm_alloc_fd(struct file **cfile)
2156 struct file *file = NULL;
2157 struct inode * inode;
2161 fd = get_unused_fd();
2162 if (fd < 0) return -ENFILE;
2166 file = get_empty_filp();
2167 if (!file) goto out;
2170 * allocate a new inode
2172 inode = new_inode(pfmfs_mnt->mnt_sb);
2173 if (!inode) goto out;
2175 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2177 inode->i_mode = S_IFCHR|S_IRUGO;
2178 inode->i_uid = current->fsuid;
2179 inode->i_gid = current->fsgid;
2181 sprintf(name, "[%lu]", inode->i_ino);
2183 this.len = strlen(name);
2184 this.hash = inode->i_ino;
2189 * allocate a new dcache entry
2191 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2192 if (!file->f_dentry) goto out;
2194 file->f_dentry->d_op = &pfmfs_dentry_operations;
2196 d_add(file->f_dentry, inode);
2197 file->f_vfsmnt = mntget(pfmfs_mnt);
2198 file->f_mapping = inode->i_mapping;
2200 file->f_op = &pfm_file_ops;
2201 file->f_mode = FMODE_READ;
2202 file->f_flags = O_RDONLY;
2206 * may have to delay until context is attached?
2208 fd_install(fd, file);
2211 * the file structure we will use
2217 if (file) put_filp(file);
2223 pfm_free_fd(int fd, struct file *file)
2225 struct files_struct *files = current->files;
2226 struct fdtable *fdt;
2229 * there ie no fd_uninstall(), so we do it here
2231 spin_lock(&files->file_lock);
2232 fdt = files_fdtable(files);
2233 rcu_assign_pointer(fdt->fd[fd], NULL);
2234 spin_unlock(&files->file_lock);
2242 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2244 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2247 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2250 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2261 * allocate a sampling buffer and remaps it into the user address space of the task
2264 pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2266 struct mm_struct *mm = task->mm;
2267 struct vm_area_struct *vma = NULL;
2273 * the fixed header + requested size and align to page boundary
2275 size = PAGE_ALIGN(rsize);
2277 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2280 * check requested size to avoid Denial-of-service attacks
2281 * XXX: may have to refine this test
2282 * Check against address space limit.
2284 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2287 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2291 * We do the easy to undo allocations first.
2293 * pfm_rvmalloc(), clears the buffer, so there is no leak
2295 smpl_buf = pfm_rvmalloc(size);
2296 if (smpl_buf == NULL) {
2297 DPRINT(("Can't allocate sampling buffer\n"));
2301 DPRINT(("smpl_buf @%p\n", smpl_buf));
2304 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2306 DPRINT(("Cannot allocate vma\n"));
2309 memset(vma, 0, sizeof(*vma));
2312 * partially initialize the vma for the sampling buffer
2315 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2316 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2319 * Now we have everything we need and we can initialize
2320 * and connect all the data structures
2323 ctx->ctx_smpl_hdr = smpl_buf;
2324 ctx->ctx_smpl_size = size; /* aligned size */
2327 * Let's do the difficult operations next.
2329 * now we atomically find some area in the address space and
2330 * remap the buffer in it.
2332 down_write(&task->mm->mmap_sem);
2334 /* find some free area in address space, must have mmap sem held */
2335 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2336 if (vma->vm_start == 0UL) {
2337 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2338 up_write(&task->mm->mmap_sem);
2341 vma->vm_end = vma->vm_start + size;
2342 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2344 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2346 /* can only be applied to current task, need to have the mm semaphore held when called */
2347 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2348 DPRINT(("Can't remap buffer\n"));
2349 up_write(&task->mm->mmap_sem);
2354 * now insert the vma in the vm list for the process, must be
2355 * done with mmap lock held
2357 insert_vm_struct(mm, vma);
2359 mm->total_vm += size >> PAGE_SHIFT;
2360 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2362 up_write(&task->mm->mmap_sem);
2365 * keep track of user level virtual address
2367 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2368 *(unsigned long *)user_vaddr = vma->vm_start;
2373 kmem_cache_free(vm_area_cachep, vma);
2375 pfm_rvfree(smpl_buf, size);
2381 * XXX: do something better here
2384 pfm_bad_permissions(struct task_struct *task)
2386 /* inspired by ptrace_attach() */
2387 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2396 return ((current->uid != task->euid)
2397 || (current->uid != task->suid)
2398 || (current->uid != task->uid)
2399 || (current->gid != task->egid)
2400 || (current->gid != task->sgid)
2401 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2405 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2411 ctx_flags = pfx->ctx_flags;
2413 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2416 * cannot block in this mode
2418 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2419 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2424 /* probably more to add here */
2430 pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2431 unsigned int cpu, pfarg_context_t *arg)
2433 pfm_buffer_fmt_t *fmt = NULL;
2434 unsigned long size = 0UL;
2436 void *fmt_arg = NULL;
2438 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2440 /* invoke and lock buffer format, if found */
2441 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2443 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2448 * buffer argument MUST be contiguous to pfarg_context_t
2450 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2452 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2454 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2456 if (ret) goto error;
2458 /* link buffer format and context */
2459 ctx->ctx_buf_fmt = fmt;
2462 * check if buffer format wants to use perfmon buffer allocation/mapping service
2464 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2465 if (ret) goto error;
2469 * buffer is always remapped into the caller's address space
2471 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2472 if (ret) goto error;
2474 /* keep track of user address of buffer */
2475 arg->ctx_smpl_vaddr = uaddr;
2477 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2484 pfm_reset_pmu_state(pfm_context_t *ctx)
2489 * install reset values for PMC.
2491 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2492 if (PMC_IS_IMPL(i) == 0) continue;
2493 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2494 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2497 * PMD registers are set to 0UL when the context in memset()
2501 * On context switched restore, we must restore ALL pmc and ALL pmd even
2502 * when they are not actively used by the task. In UP, the incoming process
2503 * may otherwise pick up left over PMC, PMD state from the previous process.
2504 * As opposed to PMD, stale PMC can cause harm to the incoming
2505 * process because they may change what is being measured.
2506 * Therefore, we must systematically reinstall the entire
2507 * PMC state. In SMP, the same thing is possible on the
2508 * same CPU but also on between 2 CPUs.
2510 * The problem with PMD is information leaking especially
2511 * to user level when psr.sp=0
2513 * There is unfortunately no easy way to avoid this problem
2514 * on either UP or SMP. This definitively slows down the
2515 * pfm_load_regs() function.
2519 * bitmask of all PMCs accessible to this context
2521 * PMC0 is treated differently.
2523 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2526 * bitmask of all PMDs that are accesible to this context
2528 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2530 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2533 * useful in case of re-enable after disable
2535 ctx->ctx_used_ibrs[0] = 0UL;
2536 ctx->ctx_used_dbrs[0] = 0UL;
2540 pfm_ctx_getsize(void *arg, size_t *sz)
2542 pfarg_context_t *req = (pfarg_context_t *)arg;
2543 pfm_buffer_fmt_t *fmt;
2547 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2549 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2551 DPRINT(("cannot find buffer format\n"));
2554 /* get just enough to copy in user parameters */
2555 *sz = fmt->fmt_arg_size;
2556 DPRINT(("arg_size=%lu\n", *sz));
2564 * cannot attach if :
2566 * - task not owned by caller
2567 * - task incompatible with context mode
2570 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2573 * no kernel task or task not owner by caller
2575 if (task->mm == NULL) {
2576 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2579 if (pfm_bad_permissions(task)) {
2580 DPRINT(("no permission to attach to [%d]\n", task->pid));
2584 * cannot block in self-monitoring mode
2586 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2587 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2591 if (task->exit_state == EXIT_ZOMBIE) {
2592 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2597 * always ok for self
2599 if (task == current) return 0;
2601 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
2602 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2606 * make sure the task is off any CPU
2608 wait_task_inactive(task);
2610 /* more to come... */
2616 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2618 struct task_struct *p = current;
2621 /* XXX: need to add more checks here */
2622 if (pid < 2) return -EPERM;
2624 if (pid != current->pid) {
2626 read_lock(&tasklist_lock);
2628 p = find_task_by_pid(pid);
2630 /* make sure task cannot go away while we operate on it */
2631 if (p) get_task_struct(p);
2633 read_unlock(&tasklist_lock);
2635 if (p == NULL) return -ESRCH;
2638 ret = pfm_task_incompatible(ctx, p);
2641 } else if (p != current) {
2650 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2652 pfarg_context_t *req = (pfarg_context_t *)arg;
2657 /* let's check the arguments first */
2658 ret = pfarg_is_sane(current, req);
2659 if (ret < 0) return ret;
2661 ctx_flags = req->ctx_flags;
2665 ctx = pfm_context_alloc();
2666 if (!ctx) goto error;
2668 ret = pfm_alloc_fd(&filp);
2669 if (ret < 0) goto error_file;
2671 req->ctx_fd = ctx->ctx_fd = ret;
2674 * attach context to file
2676 filp->private_data = ctx;
2679 * does the user want to sample?
2681 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2682 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2683 if (ret) goto buffer_error;
2687 * init context protection lock
2689 spin_lock_init(&ctx->ctx_lock);
2692 * context is unloaded
2694 ctx->ctx_state = PFM_CTX_UNLOADED;
2697 * initialization of context's flags
2699 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2700 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2701 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2702 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2704 * will move to set properties
2705 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2709 * init restart semaphore to locked
2711 init_completion(&ctx->ctx_restart_done);
2714 * activation is used in SMP only
2716 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2717 SET_LAST_CPU(ctx, -1);
2720 * initialize notification message queue
2722 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2723 init_waitqueue_head(&ctx->ctx_msgq_wait);
2724 init_waitqueue_head(&ctx->ctx_zombieq);
2726 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2731 ctx->ctx_fl_excl_idle,
2736 * initialize soft PMU state
2738 pfm_reset_pmu_state(ctx);
2743 pfm_free_fd(ctx->ctx_fd, filp);
2745 if (ctx->ctx_buf_fmt) {
2746 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2749 pfm_context_free(ctx);
2755 static inline unsigned long
2756 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2758 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2759 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2760 extern unsigned long carta_random32 (unsigned long seed);
2762 if (reg->flags & PFM_REGFL_RANDOM) {
2763 new_seed = carta_random32(old_seed);
2764 val -= (old_seed & mask); /* counter values are negative numbers! */
2765 if ((mask >> 32) != 0)
2766 /* construct a full 64-bit random value: */
2767 new_seed |= carta_random32(old_seed >> 32) << 32;
2768 reg->seed = new_seed;
2775 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2777 unsigned long mask = ovfl_regs[0];
2778 unsigned long reset_others = 0UL;
2783 * now restore reset value on sampling overflowed counters
2785 mask >>= PMU_FIRST_COUNTER;
2786 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2788 if ((mask & 0x1UL) == 0UL) continue;
2790 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2791 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2793 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2797 * Now take care of resetting the other registers
2799 for(i = 0; reset_others; i++, reset_others >>= 1) {
2801 if ((reset_others & 0x1) == 0) continue;
2803 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2805 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2806 is_long_reset ? "long" : "short", i, val));
2811 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2813 unsigned long mask = ovfl_regs[0];
2814 unsigned long reset_others = 0UL;
2818 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2820 if (ctx->ctx_state == PFM_CTX_MASKED) {
2821 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2826 * now restore reset value on sampling overflowed counters
2828 mask >>= PMU_FIRST_COUNTER;
2829 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2831 if ((mask & 0x1UL) == 0UL) continue;
2833 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2834 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2836 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2838 pfm_write_soft_counter(ctx, i, val);
2842 * Now take care of resetting the other registers
2844 for(i = 0; reset_others; i++, reset_others >>= 1) {
2846 if ((reset_others & 0x1) == 0) continue;
2848 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2850 if (PMD_IS_COUNTING(i)) {
2851 pfm_write_soft_counter(ctx, i, val);
2853 ia64_set_pmd(i, val);
2855 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2856 is_long_reset ? "long" : "short", i, val));
2862 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2864 struct thread_struct *thread = NULL;
2865 struct task_struct *task;
2866 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2867 unsigned long value, pmc_pm;
2868 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2869 unsigned int cnum, reg_flags, flags, pmc_type;
2870 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2871 int is_monitor, is_counting, state;
2873 pfm_reg_check_t wr_func;
2874 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2876 state = ctx->ctx_state;
2877 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2878 is_system = ctx->ctx_fl_system;
2879 task = ctx->ctx_task;
2880 impl_pmds = pmu_conf->impl_pmds[0];
2882 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2885 thread = &task->thread;
2887 * In system wide and when the context is loaded, access can only happen
2888 * when the caller is running on the CPU being monitored by the session.
2889 * It does not have to be the owner (ctx_task) of the context per se.
2891 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2892 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2895 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2897 expert_mode = pfm_sysctl.expert_mode;
2899 for (i = 0; i < count; i++, req++) {
2901 cnum = req->reg_num;
2902 reg_flags = req->reg_flags;
2903 value = req->reg_value;
2904 smpl_pmds = req->reg_smpl_pmds[0];
2905 reset_pmds = req->reg_reset_pmds[0];
2909 if (cnum >= PMU_MAX_PMCS) {
2910 DPRINT(("pmc%u is invalid\n", cnum));
2914 pmc_type = pmu_conf->pmc_desc[cnum].type;
2915 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2916 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2917 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2920 * we reject all non implemented PMC as well
2921 * as attempts to modify PMC[0-3] which are used
2922 * as status registers by the PMU
2924 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2925 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2928 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2930 * If the PMC is a monitor, then if the value is not the default:
2931 * - system-wide session: PMCx.pm=1 (privileged monitor)
2932 * - per-task : PMCx.pm=0 (user monitor)
2934 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2935 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2944 * enforce generation of overflow interrupt. Necessary on all
2947 value |= 1 << PMU_PMC_OI;
2949 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2950 flags |= PFM_REGFL_OVFL_NOTIFY;
2953 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2955 /* verify validity of smpl_pmds */
2956 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2957 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2961 /* verify validity of reset_pmds */
2962 if ((reset_pmds & impl_pmds) != reset_pmds) {
2963 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2967 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2968 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2971 /* eventid on non-counting monitors are ignored */
2975 * execute write checker, if any
2977 if (likely(expert_mode == 0 && wr_func)) {
2978 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2979 if (ret) goto error;
2984 * no error on this register
2986 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2989 * Now we commit the changes to the software state
2993 * update overflow information
2997 * full flag update each time a register is programmed
2999 ctx->ctx_pmds[cnum].flags = flags;
3001 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
3002 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
3003 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
3006 * Mark all PMDS to be accessed as used.
3008 * We do not keep track of PMC because we have to
3009 * systematically restore ALL of them.
3011 * We do not update the used_monitors mask, because
3012 * if we have not programmed them, then will be in
3013 * a quiescent state, therefore we will not need to
3014 * mask/restore then when context is MASKED.
3016 CTX_USED_PMD(ctx, reset_pmds);
3017 CTX_USED_PMD(ctx, smpl_pmds);
3019 * make sure we do not try to reset on
3020 * restart because we have established new values
3022 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3025 * Needed in case the user does not initialize the equivalent
3026 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3027 * possible leak here.
3029 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3032 * keep track of the monitor PMC that we are using.
3033 * we save the value of the pmc in ctx_pmcs[] and if
3034 * the monitoring is not stopped for the context we also
3035 * place it in the saved state area so that it will be
3036 * picked up later by the context switch code.
3038 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3040 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3041 * monitoring needs to be stopped.
3043 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3046 * update context state
3048 ctx->ctx_pmcs[cnum] = value;
3052 * write thread state
3054 if (is_system == 0) thread->pmcs[cnum] = value;
3057 * write hardware register if we can
3059 if (can_access_pmu) {
3060 ia64_set_pmc(cnum, value);
3065 * per-task SMP only here
3067 * we are guaranteed that the task is not running on the other CPU,
3068 * we indicate that this PMD will need to be reloaded if the task
3069 * is rescheduled on the CPU it ran last on.
3071 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3076 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",
3082 ctx->ctx_all_pmcs[0],
3083 ctx->ctx_used_pmds[0],
3084 ctx->ctx_pmds[cnum].eventid,
3087 ctx->ctx_reload_pmcs[0],
3088 ctx->ctx_used_monitors[0],
3089 ctx->ctx_ovfl_regs[0]));
3093 * make sure the changes are visible
3095 if (can_access_pmu) ia64_srlz_d();
3099 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3104 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3106 struct thread_struct *thread = NULL;
3107 struct task_struct *task;
3108 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3109 unsigned long value, hw_value, ovfl_mask;
3111 int i, can_access_pmu = 0, state;
3112 int is_counting, is_loaded, is_system, expert_mode;
3114 pfm_reg_check_t wr_func;
3117 state = ctx->ctx_state;
3118 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3119 is_system = ctx->ctx_fl_system;
3120 ovfl_mask = pmu_conf->ovfl_val;
3121 task = ctx->ctx_task;
3123 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3126 * on both UP and SMP, we can only write to the PMC when the task is
3127 * the owner of the local PMU.
3129 if (likely(is_loaded)) {
3130 thread = &task->thread;
3132 * In system wide and when the context is loaded, access can only happen
3133 * when the caller is running on the CPU being monitored by the session.
3134 * It does not have to be the owner (ctx_task) of the context per se.
3136 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3137 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3140 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3142 expert_mode = pfm_sysctl.expert_mode;
3144 for (i = 0; i < count; i++, req++) {
3146 cnum = req->reg_num;
3147 value = req->reg_value;
3149 if (!PMD_IS_IMPL(cnum)) {
3150 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3153 is_counting = PMD_IS_COUNTING(cnum);
3154 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3157 * execute write checker, if any
3159 if (unlikely(expert_mode == 0 && wr_func)) {
3160 unsigned long v = value;
3162 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3163 if (ret) goto abort_mission;
3170 * no error on this register
3172 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3175 * now commit changes to software state
3180 * update virtualized (64bits) counter
3184 * write context state
3186 ctx->ctx_pmds[cnum].lval = value;
3189 * when context is load we use the split value
3192 hw_value = value & ovfl_mask;
3193 value = value & ~ovfl_mask;
3197 * update reset values (not just for counters)
3199 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3200 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3203 * update randomization parameters (not just for counters)
3205 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3206 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3209 * update context value
3211 ctx->ctx_pmds[cnum].val = value;
3214 * Keep track of what we use
3216 * We do not keep track of PMC because we have to
3217 * systematically restore ALL of them.
3219 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3222 * mark this PMD register used as well
3224 CTX_USED_PMD(ctx, RDEP(cnum));
3227 * make sure we do not try to reset on
3228 * restart because we have established new values
3230 if (is_counting && state == PFM_CTX_MASKED) {
3231 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3236 * write thread state
3238 if (is_system == 0) thread->pmds[cnum] = hw_value;
3241 * write hardware register if we can
3243 if (can_access_pmu) {
3244 ia64_set_pmd(cnum, hw_value);
3248 * we are guaranteed that the task is not running on the other CPU,
3249 * we indicate that this PMD will need to be reloaded if the task
3250 * is rescheduled on the CPU it ran last on.
3252 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3257 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3258 "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",
3264 ctx->ctx_pmds[cnum].val,
3265 ctx->ctx_pmds[cnum].short_reset,
3266 ctx->ctx_pmds[cnum].long_reset,
3267 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3268 ctx->ctx_pmds[cnum].seed,
3269 ctx->ctx_pmds[cnum].mask,
3270 ctx->ctx_used_pmds[0],
3271 ctx->ctx_pmds[cnum].reset_pmds[0],
3272 ctx->ctx_reload_pmds[0],
3273 ctx->ctx_all_pmds[0],
3274 ctx->ctx_ovfl_regs[0]));
3278 * make changes visible
3280 if (can_access_pmu) ia64_srlz_d();
3286 * for now, we have only one possibility for error
3288 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3293 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3294 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3295 * interrupt is delivered during the call, it will be kept pending until we leave, making
3296 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3297 * guaranteed to return consistent data to the user, it may simply be old. It is not
3298 * trivial to treat the overflow while inside the call because you may end up in
3299 * some module sampling buffer code causing deadlocks.
3302 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3304 struct thread_struct *thread = NULL;
3305 struct task_struct *task;
3306 unsigned long val = 0UL, lval, ovfl_mask, sval;
3307 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3308 unsigned int cnum, reg_flags = 0;
3309 int i, can_access_pmu = 0, state;
3310 int is_loaded, is_system, is_counting, expert_mode;
3312 pfm_reg_check_t rd_func;
3315 * access is possible when loaded only for
3316 * self-monitoring tasks or in UP mode
3319 state = ctx->ctx_state;
3320 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3321 is_system = ctx->ctx_fl_system;
3322 ovfl_mask = pmu_conf->ovfl_val;
3323 task = ctx->ctx_task;
3325 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3327 if (likely(is_loaded)) {
3328 thread = &task->thread;
3330 * In system wide and when the context is loaded, access can only happen
3331 * when the caller is running on the CPU being monitored by the session.
3332 * It does not have to be the owner (ctx_task) of the context per se.
3334 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3335 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3339 * this can be true when not self-monitoring only in UP
3341 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3343 if (can_access_pmu) ia64_srlz_d();
3345 expert_mode = pfm_sysctl.expert_mode;
3347 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3353 * on both UP and SMP, we can only read the PMD from the hardware register when
3354 * the task is the owner of the local PMU.
3357 for (i = 0; i < count; i++, req++) {
3359 cnum = req->reg_num;
3360 reg_flags = req->reg_flags;
3362 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3364 * we can only read the register that we use. That includes
3365 * the one we explicitely initialize AND the one we want included
3366 * in the sampling buffer (smpl_regs).
3368 * Having this restriction allows optimization in the ctxsw routine
3369 * without compromising security (leaks)
3371 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3373 sval = ctx->ctx_pmds[cnum].val;
3374 lval = ctx->ctx_pmds[cnum].lval;
3375 is_counting = PMD_IS_COUNTING(cnum);
3378 * If the task is not the current one, then we check if the
3379 * PMU state is still in the local live register due to lazy ctxsw.
3380 * If true, then we read directly from the registers.
3382 if (can_access_pmu){
3383 val = ia64_get_pmd(cnum);
3386 * context has been saved
3387 * if context is zombie, then task does not exist anymore.
3388 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3390 val = is_loaded ? thread->pmds[cnum] : 0UL;
3392 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3396 * XXX: need to check for overflow when loaded
3403 * execute read checker, if any
3405 if (unlikely(expert_mode == 0 && rd_func)) {
3406 unsigned long v = val;
3407 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3408 if (ret) goto error;
3413 PFM_REG_RETFLAG_SET(reg_flags, 0);
3415 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3418 * update register return value, abort all if problem during copy.
3419 * we only modify the reg_flags field. no check mode is fine because
3420 * access has been verified upfront in sys_perfmonctl().
3422 req->reg_value = val;
3423 req->reg_flags = reg_flags;
3424 req->reg_last_reset_val = lval;
3430 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3435 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3439 if (req == NULL) return -EINVAL;
3441 ctx = GET_PMU_CTX();
3443 if (ctx == NULL) return -EINVAL;
3446 * for now limit to current task, which is enough when calling
3447 * from overflow handler
3449 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3451 return pfm_write_pmcs(ctx, req, nreq, regs);
3453 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3456 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3460 if (req == NULL) return -EINVAL;
3462 ctx = GET_PMU_CTX();
3464 if (ctx == NULL) return -EINVAL;
3467 * for now limit to current task, which is enough when calling
3468 * from overflow handler
3470 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3472 return pfm_read_pmds(ctx, req, nreq, regs);
3474 EXPORT_SYMBOL(pfm_mod_read_pmds);
3477 * Only call this function when a process it trying to
3478 * write the debug registers (reading is always allowed)
3481 pfm_use_debug_registers(struct task_struct *task)
3483 pfm_context_t *ctx = task->thread.pfm_context;
3484 unsigned long flags;
3487 if (pmu_conf->use_rr_dbregs == 0) return 0;
3489 DPRINT(("called for [%d]\n", task->pid));
3494 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3497 * Even on SMP, we do not need to use an atomic here because
3498 * the only way in is via ptrace() and this is possible only when the
3499 * process is stopped. Even in the case where the ctxsw out is not totally
3500 * completed by the time we come here, there is no way the 'stopped' process
3501 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3502 * So this is always safe.
3504 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3509 * We cannot allow setting breakpoints when system wide monitoring
3510 * sessions are using the debug registers.
3512 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3515 pfm_sessions.pfs_ptrace_use_dbregs++;
3517 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3518 pfm_sessions.pfs_ptrace_use_dbregs,
3519 pfm_sessions.pfs_sys_use_dbregs,
3528 * This function is called for every task that exits with the
3529 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3530 * able to use the debug registers for debugging purposes via
3531 * ptrace(). Therefore we know it was not using them for
3532 * perfmormance monitoring, so we only decrement the number
3533 * of "ptraced" debug register users to keep the count up to date
3536 pfm_release_debug_registers(struct task_struct *task)
3538 unsigned long flags;
3541 if (pmu_conf->use_rr_dbregs == 0) return 0;
3544 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3545 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3548 pfm_sessions.pfs_ptrace_use_dbregs--;
3557 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3559 struct task_struct *task;
3560 pfm_buffer_fmt_t *fmt;
3561 pfm_ovfl_ctrl_t rst_ctrl;
3562 int state, is_system;
3565 state = ctx->ctx_state;
3566 fmt = ctx->ctx_buf_fmt;
3567 is_system = ctx->ctx_fl_system;
3568 task = PFM_CTX_TASK(ctx);
3571 case PFM_CTX_MASKED:
3573 case PFM_CTX_LOADED:
3574 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3576 case PFM_CTX_UNLOADED:
3577 case PFM_CTX_ZOMBIE:
3578 DPRINT(("invalid state=%d\n", state));
3581 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3586 * In system wide and when the context is loaded, access can only happen
3587 * when the caller is running on the CPU being monitored by the session.
3588 * It does not have to be the owner (ctx_task) of the context per se.
3590 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3591 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3596 if (unlikely(task == NULL)) {
3597 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3601 if (task == current || is_system) {
3603 fmt = ctx->ctx_buf_fmt;
3605 DPRINT(("restarting self %d ovfl=0x%lx\n",
3607 ctx->ctx_ovfl_regs[0]));
3609 if (CTX_HAS_SMPL(ctx)) {
3611 prefetch(ctx->ctx_smpl_hdr);
3613 rst_ctrl.bits.mask_monitoring = 0;
3614 rst_ctrl.bits.reset_ovfl_pmds = 0;
3616 if (state == PFM_CTX_LOADED)
3617 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3619 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3621 rst_ctrl.bits.mask_monitoring = 0;
3622 rst_ctrl.bits.reset_ovfl_pmds = 1;
3626 if (rst_ctrl.bits.reset_ovfl_pmds)
3627 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3629 if (rst_ctrl.bits.mask_monitoring == 0) {
3630 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3632 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3634 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3636 // cannot use pfm_stop_monitoring(task, regs);
3640 * clear overflowed PMD mask to remove any stale information
3642 ctx->ctx_ovfl_regs[0] = 0UL;
3645 * back to LOADED state
3647 ctx->ctx_state = PFM_CTX_LOADED;
3650 * XXX: not really useful for self monitoring
3652 ctx->ctx_fl_can_restart = 0;
3658 * restart another task
3662 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3663 * one is seen by the task.
3665 if (state == PFM_CTX_MASKED) {
3666 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3668 * will prevent subsequent restart before this one is
3669 * seen by other task
3671 ctx->ctx_fl_can_restart = 0;
3675 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3676 * the task is blocked or on its way to block. That's the normal
3677 * restart path. If the monitoring is not masked, then the task
3678 * can be actively monitoring and we cannot directly intervene.
3679 * Therefore we use the trap mechanism to catch the task and
3680 * force it to reset the buffer/reset PMDs.
3682 * if non-blocking, then we ensure that the task will go into
3683 * pfm_handle_work() before returning to user mode.
3685 * We cannot explicitely reset another task, it MUST always
3686 * be done by the task itself. This works for system wide because
3687 * the tool that is controlling the session is logically doing
3688 * "self-monitoring".
3690 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3691 DPRINT(("unblocking [%d] \n", task->pid));
3692 complete(&ctx->ctx_restart_done);
3694 DPRINT(("[%d] armed exit trap\n", task->pid));
3696 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3698 PFM_SET_WORK_PENDING(task, 1);
3700 pfm_set_task_notify(task);
3703 * XXX: send reschedule if task runs on another CPU
3710 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3712 unsigned int m = *(unsigned int *)arg;
3714 pfm_sysctl.debug = m == 0 ? 0 : 1;
3716 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3719 memset(pfm_stats, 0, sizeof(pfm_stats));
3720 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3726 * arg can be NULL and count can be zero for this function
3729 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3731 struct thread_struct *thread = NULL;
3732 struct task_struct *task;
3733 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3734 unsigned long flags;
3739 int i, can_access_pmu = 0;
3740 int is_system, is_loaded;
3742 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3744 state = ctx->ctx_state;
3745 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3746 is_system = ctx->ctx_fl_system;
3747 task = ctx->ctx_task;
3749 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3752 * on both UP and SMP, we can only write to the PMC when the task is
3753 * the owner of the local PMU.
3756 thread = &task->thread;
3758 * In system wide and when the context is loaded, access can only happen
3759 * when the caller is running on the CPU being monitored by the session.
3760 * It does not have to be the owner (ctx_task) of the context per se.
3762 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3763 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3766 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3770 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3771 * ensuring that no real breakpoint can be installed via this call.
3773 * IMPORTANT: regs can be NULL in this function
3776 first_time = ctx->ctx_fl_using_dbreg == 0;
3779 * don't bother if we are loaded and task is being debugged
3781 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3782 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3787 * check for debug registers in system wide mode
3789 * If though a check is done in pfm_context_load(),
3790 * we must repeat it here, in case the registers are
3791 * written after the context is loaded
3796 if (first_time && is_system) {
3797 if (pfm_sessions.pfs_ptrace_use_dbregs)
3800 pfm_sessions.pfs_sys_use_dbregs++;
3805 if (ret != 0) return ret;
3808 * mark ourself as user of the debug registers for
3811 ctx->ctx_fl_using_dbreg = 1;
3814 * clear hardware registers to make sure we don't
3815 * pick up stale state.
3817 * for a system wide session, we do not use
3818 * thread.dbr, thread.ibr because this process
3819 * never leaves the current CPU and the state
3820 * is shared by all processes running on it
3822 if (first_time && can_access_pmu) {
3823 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3824 for (i=0; i < pmu_conf->num_ibrs; i++) {
3825 ia64_set_ibr(i, 0UL);
3826 ia64_dv_serialize_instruction();
3829 for (i=0; i < pmu_conf->num_dbrs; i++) {
3830 ia64_set_dbr(i, 0UL);
3831 ia64_dv_serialize_data();
3837 * Now install the values into the registers
3839 for (i = 0; i < count; i++, req++) {
3841 rnum = req->dbreg_num;
3842 dbreg.val = req->dbreg_value;
3846 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3847 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3848 rnum, dbreg.val, mode, i, count));
3854 * make sure we do not install enabled breakpoint
3857 if (mode == PFM_CODE_RR)
3858 dbreg.ibr.ibr_x = 0;
3860 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3863 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3866 * Debug registers, just like PMC, can only be modified
3867 * by a kernel call. Moreover, perfmon() access to those
3868 * registers are centralized in this routine. The hardware
3869 * does not modify the value of these registers, therefore,
3870 * if we save them as they are written, we can avoid having
3871 * to save them on context switch out. This is made possible
3872 * by the fact that when perfmon uses debug registers, ptrace()
3873 * won't be able to modify them concurrently.
3875 if (mode == PFM_CODE_RR) {
3876 CTX_USED_IBR(ctx, rnum);
3878 if (can_access_pmu) {
3879 ia64_set_ibr(rnum, dbreg.val);
3880 ia64_dv_serialize_instruction();
3883 ctx->ctx_ibrs[rnum] = dbreg.val;
3885 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3886 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3888 CTX_USED_DBR(ctx, rnum);
3890 if (can_access_pmu) {
3891 ia64_set_dbr(rnum, dbreg.val);
3892 ia64_dv_serialize_data();
3894 ctx->ctx_dbrs[rnum] = dbreg.val;
3896 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3897 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3905 * in case it was our first attempt, we undo the global modifications
3909 if (ctx->ctx_fl_system) {
3910 pfm_sessions.pfs_sys_use_dbregs--;
3913 ctx->ctx_fl_using_dbreg = 0;
3916 * install error return flag
3918 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3924 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3926 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3930 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3932 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3936 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3940 if (req == NULL) return -EINVAL;
3942 ctx = GET_PMU_CTX();
3944 if (ctx == NULL) return -EINVAL;
3947 * for now limit to current task, which is enough when calling
3948 * from overflow handler
3950 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3952 return pfm_write_ibrs(ctx, req, nreq, regs);
3954 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3957 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3961 if (req == NULL) return -EINVAL;
3963 ctx = GET_PMU_CTX();
3965 if (ctx == NULL) return -EINVAL;
3968 * for now limit to current task, which is enough when calling
3969 * from overflow handler
3971 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3973 return pfm_write_dbrs(ctx, req, nreq, regs);
3975 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3979 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3981 pfarg_features_t *req = (pfarg_features_t *)arg;
3983 req->ft_version = PFM_VERSION;
3988 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3990 struct pt_regs *tregs;
3991 struct task_struct *task = PFM_CTX_TASK(ctx);
3992 int state, is_system;
3994 state = ctx->ctx_state;
3995 is_system = ctx->ctx_fl_system;
3998 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
4000 if (state == PFM_CTX_UNLOADED) return -EINVAL;
4003 * In system wide and when the context is loaded, access can only happen
4004 * when the caller is running on the CPU being monitored by the session.
4005 * It does not have to be the owner (ctx_task) of the context per se.
4007 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4008 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4011 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4012 PFM_CTX_TASK(ctx)->pid,
4016 * in system mode, we need to update the PMU directly
4017 * and the user level state of the caller, which may not
4018 * necessarily be the creator of the context.
4022 * Update local PMU first
4026 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4030 * update local cpuinfo
4032 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4035 * stop monitoring, does srlz.i
4040 * stop monitoring in the caller
4042 ia64_psr(regs)->pp = 0;
4050 if (task == current) {
4051 /* stop monitoring at kernel level */
4055 * stop monitoring at the user level
4057 ia64_psr(regs)->up = 0;
4059 tregs = task_pt_regs(task);
4062 * stop monitoring at the user level
4064 ia64_psr(tregs)->up = 0;
4067 * monitoring disabled in kernel at next reschedule
4069 ctx->ctx_saved_psr_up = 0;
4070 DPRINT(("task=[%d]\n", task->pid));
4077 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4079 struct pt_regs *tregs;
4080 int state, is_system;
4082 state = ctx->ctx_state;
4083 is_system = ctx->ctx_fl_system;
4085 if (state != PFM_CTX_LOADED) return -EINVAL;
4088 * In system wide and when the context is loaded, access can only happen
4089 * when the caller is running on the CPU being monitored by the session.
4090 * It does not have to be the owner (ctx_task) of the context per se.
4092 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4093 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4098 * in system mode, we need to update the PMU directly
4099 * and the user level state of the caller, which may not
4100 * necessarily be the creator of the context.
4105 * set user level psr.pp for the caller
4107 ia64_psr(regs)->pp = 1;
4110 * now update the local PMU and cpuinfo
4112 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4115 * start monitoring at kernel level
4120 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4130 if (ctx->ctx_task == current) {
4132 /* start monitoring at kernel level */
4136 * activate monitoring at user level
4138 ia64_psr(regs)->up = 1;
4141 tregs = task_pt_regs(ctx->ctx_task);
4144 * start monitoring at the kernel level the next
4145 * time the task is scheduled
4147 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4150 * activate monitoring at user level
4152 ia64_psr(tregs)->up = 1;
4158 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4160 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4165 for (i = 0; i < count; i++, req++) {
4167 cnum = req->reg_num;
4169 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4171 req->reg_value = PMC_DFL_VAL(cnum);
4173 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4175 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4180 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4185 pfm_check_task_exist(pfm_context_t *ctx)
4187 struct task_struct *g, *t;
4190 read_lock(&tasklist_lock);
4192 do_each_thread (g, t) {
4193 if (t->thread.pfm_context == ctx) {
4197 } while_each_thread (g, t);
4199 read_unlock(&tasklist_lock);
4201 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4207 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4209 struct task_struct *task;
4210 struct thread_struct *thread;
4211 struct pfm_context_t *old;
4212 unsigned long flags;
4214 struct task_struct *owner_task = NULL;
4216 pfarg_load_t *req = (pfarg_load_t *)arg;
4217 unsigned long *pmcs_source, *pmds_source;
4220 int state, is_system, set_dbregs = 0;
4222 state = ctx->ctx_state;
4223 is_system = ctx->ctx_fl_system;
4225 * can only load from unloaded or terminated state
4227 if (state != PFM_CTX_UNLOADED) {
4228 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4234 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4236 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4237 DPRINT(("cannot use blocking mode on self\n"));
4241 ret = pfm_get_task(ctx, req->load_pid, &task);
4243 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4250 * system wide is self monitoring only
4252 if (is_system && task != current) {
4253 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4258 thread = &task->thread;
4262 * cannot load a context which is using range restrictions,
4263 * into a task that is being debugged.
4265 if (ctx->ctx_fl_using_dbreg) {
4266 if (thread->flags & IA64_THREAD_DBG_VALID) {
4268 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4274 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4275 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4278 pfm_sessions.pfs_sys_use_dbregs++;
4279 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4286 if (ret) goto error;
4290 * SMP system-wide monitoring implies self-monitoring.
4292 * The programming model expects the task to
4293 * be pinned on a CPU throughout the session.
4294 * Here we take note of the current CPU at the
4295 * time the context is loaded. No call from
4296 * another CPU will be allowed.
4298 * The pinning via shed_setaffinity()
4299 * must be done by the calling task prior
4302 * systemwide: keep track of CPU this session is supposed to run on
4304 the_cpu = ctx->ctx_cpu = smp_processor_id();
4308 * now reserve the session
4310 ret = pfm_reserve_session(current, is_system, the_cpu);
4311 if (ret) goto error;
4314 * task is necessarily stopped at this point.
4316 * If the previous context was zombie, then it got removed in
4317 * pfm_save_regs(). Therefore we should not see it here.
4318 * If we see a context, then this is an active context
4320 * XXX: needs to be atomic
4322 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4323 thread->pfm_context, ctx));
4326 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4328 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4332 pfm_reset_msgq(ctx);
4334 ctx->ctx_state = PFM_CTX_LOADED;
4337 * link context to task
4339 ctx->ctx_task = task;
4343 * we load as stopped
4345 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4346 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4348 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4350 thread->flags |= IA64_THREAD_PM_VALID;
4354 * propagate into thread-state
4356 pfm_copy_pmds(task, ctx);
4357 pfm_copy_pmcs(task, ctx);
4359 pmcs_source = thread->pmcs;
4360 pmds_source = thread->pmds;
4363 * always the case for system-wide
4365 if (task == current) {
4367 if (is_system == 0) {
4369 /* allow user level control */
4370 ia64_psr(regs)->sp = 0;
4371 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4373 SET_LAST_CPU(ctx, smp_processor_id());
4375 SET_ACTIVATION(ctx);
4378 * push the other task out, if any
4380 owner_task = GET_PMU_OWNER();
4381 if (owner_task) pfm_lazy_save_regs(owner_task);
4385 * load all PMD from ctx to PMU (as opposed to thread state)
4386 * restore all PMC from ctx to PMU
4388 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4389 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4391 ctx->ctx_reload_pmcs[0] = 0UL;
4392 ctx->ctx_reload_pmds[0] = 0UL;
4395 * guaranteed safe by earlier check against DBG_VALID
4397 if (ctx->ctx_fl_using_dbreg) {
4398 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4399 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4404 SET_PMU_OWNER(task, ctx);
4406 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4409 * when not current, task MUST be stopped, so this is safe
4411 regs = task_pt_regs(task);
4413 /* force a full reload */
4414 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4415 SET_LAST_CPU(ctx, -1);
4417 /* initial saved psr (stopped) */
4418 ctx->ctx_saved_psr_up = 0UL;
4419 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4425 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4428 * we must undo the dbregs setting (for system-wide)
4430 if (ret && set_dbregs) {
4432 pfm_sessions.pfs_sys_use_dbregs--;
4436 * release task, there is now a link with the context
4438 if (is_system == 0 && task != current) {
4442 ret = pfm_check_task_exist(ctx);
4444 ctx->ctx_state = PFM_CTX_UNLOADED;
4445 ctx->ctx_task = NULL;
4453 * in this function, we do not need to increase the use count
4454 * for the task via get_task_struct(), because we hold the
4455 * context lock. If the task were to disappear while having
4456 * a context attached, it would go through pfm_exit_thread()
4457 * which also grabs the context lock and would therefore be blocked
4458 * until we are here.
4460 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4463 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4465 struct task_struct *task = PFM_CTX_TASK(ctx);
4466 struct pt_regs *tregs;
4467 int prev_state, is_system;
4470 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4472 prev_state = ctx->ctx_state;
4473 is_system = ctx->ctx_fl_system;
4476 * unload only when necessary
4478 if (prev_state == PFM_CTX_UNLOADED) {
4479 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4484 * clear psr and dcr bits
4486 ret = pfm_stop(ctx, NULL, 0, regs);
4487 if (ret) return ret;
4489 ctx->ctx_state = PFM_CTX_UNLOADED;
4492 * in system mode, we need to update the PMU directly
4493 * and the user level state of the caller, which may not
4494 * necessarily be the creator of the context.
4501 * local PMU is taken care of in pfm_stop()
4503 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4504 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4507 * save PMDs in context
4510 pfm_flush_pmds(current, ctx);
4513 * at this point we are done with the PMU
4514 * so we can unreserve the resource.
4516 if (prev_state != PFM_CTX_ZOMBIE)
4517 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4520 * disconnect context from task
4522 task->thread.pfm_context = NULL;
4524 * disconnect task from context
4526 ctx->ctx_task = NULL;
4529 * There is nothing more to cleanup here.
4537 tregs = task == current ? regs : task_pt_regs(task);
4539 if (task == current) {
4541 * cancel user level control
4543 ia64_psr(regs)->sp = 1;
4545 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4548 * save PMDs to context
4551 pfm_flush_pmds(task, ctx);
4554 * at this point we are done with the PMU
4555 * so we can unreserve the resource.
4557 * when state was ZOMBIE, we have already unreserved.
4559 if (prev_state != PFM_CTX_ZOMBIE)
4560 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4563 * reset activation counter and psr
4565 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4566 SET_LAST_CPU(ctx, -1);
4569 * PMU state will not be restored
4571 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4574 * break links between context and task
4576 task->thread.pfm_context = NULL;
4577 ctx->ctx_task = NULL;
4579 PFM_SET_WORK_PENDING(task, 0);
4581 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4582 ctx->ctx_fl_can_restart = 0;
4583 ctx->ctx_fl_going_zombie = 0;
4585 DPRINT(("disconnected [%d] from context\n", task->pid));
4592 * called only from exit_thread(): task == current
4593 * we come here only if current has a context attached (loaded or masked)
4596 pfm_exit_thread(struct task_struct *task)
4599 unsigned long flags;
4600 struct pt_regs *regs = task_pt_regs(task);
4604 ctx = PFM_GET_CTX(task);
4606 PROTECT_CTX(ctx, flags);
4608 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4610 state = ctx->ctx_state;
4612 case PFM_CTX_UNLOADED:
4614 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4615 * be in unloaded state
4617 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4619 case PFM_CTX_LOADED:
4620 case PFM_CTX_MASKED:
4621 ret = pfm_context_unload(ctx, NULL, 0, regs);
4623 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4625 DPRINT(("ctx unloaded for current state was %d\n", state));
4627 pfm_end_notify_user(ctx);
4629 case PFM_CTX_ZOMBIE:
4630 ret = pfm_context_unload(ctx, NULL, 0, regs);
4632 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4637 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4640 UNPROTECT_CTX(ctx, flags);
4642 { u64 psr = pfm_get_psr();
4643 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4644 BUG_ON(GET_PMU_OWNER());
4645 BUG_ON(ia64_psr(regs)->up);
4646 BUG_ON(ia64_psr(regs)->pp);
4650 * All memory free operations (especially for vmalloc'ed memory)
4651 * MUST be done with interrupts ENABLED.
4653 if (free_ok) pfm_context_free(ctx);
4657 * functions MUST be listed in the increasing order of their index (see permfon.h)
4659 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4660 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4661 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4662 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4663 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4665 static pfm_cmd_desc_t pfm_cmd_tab[]={
4666 /* 0 */PFM_CMD_NONE,
4667 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4668 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4669 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4670 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4671 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4672 /* 6 */PFM_CMD_NONE,
4673 /* 7 */PFM_CMD_NONE,
4674 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4675 /* 9 */PFM_CMD_NONE,
4676 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4677 /* 11 */PFM_CMD_NONE,
4678 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4679 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4680 /* 14 */PFM_CMD_NONE,
4681 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4682 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4683 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4684 /* 18 */PFM_CMD_NONE,
4685 /* 19 */PFM_CMD_NONE,
4686 /* 20 */PFM_CMD_NONE,
4687 /* 21 */PFM_CMD_NONE,
4688 /* 22 */PFM_CMD_NONE,
4689 /* 23 */PFM_CMD_NONE,
4690 /* 24 */PFM_CMD_NONE,
4691 /* 25 */PFM_CMD_NONE,
4692 /* 26 */PFM_CMD_NONE,
4693 /* 27 */PFM_CMD_NONE,
4694 /* 28 */PFM_CMD_NONE,
4695 /* 29 */PFM_CMD_NONE,
4696 /* 30 */PFM_CMD_NONE,
4697 /* 31 */PFM_CMD_NONE,
4698 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4699 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4701 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4704 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4706 struct task_struct *task;
4707 int state, old_state;
4710 state = ctx->ctx_state;
4711 task = ctx->ctx_task;
4714 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4718 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4722 task->state, PFM_CMD_STOPPED(cmd)));
4725 * self-monitoring always ok.
4727 * for system-wide the caller can either be the creator of the
4728 * context (to one to which the context is attached to) OR
4729 * a task running on the same CPU as the session.
4731 if (task == current || ctx->ctx_fl_system) return 0;
4734 * we are monitoring another thread
4737 case PFM_CTX_UNLOADED:
4739 * if context is UNLOADED we are safe to go
4742 case PFM_CTX_ZOMBIE:
4744 * no command can operate on a zombie context
4746 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4748 case PFM_CTX_MASKED:
4750 * PMU state has been saved to software even though
4751 * the thread may still be running.
4753 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4757 * context is LOADED or MASKED. Some commands may need to have
4760 * We could lift this restriction for UP but it would mean that
4761 * the user has no guarantee the task would not run between
4762 * two successive calls to perfmonctl(). That's probably OK.
4763 * If this user wants to ensure the task does not run, then
4764 * the task must be stopped.
4766 if (PFM_CMD_STOPPED(cmd)) {
4767 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
4768 DPRINT(("[%d] task not in stopped state\n", task->pid));
4772 * task is now stopped, wait for ctxsw out
4774 * This is an interesting point in the code.
4775 * We need to unprotect the context because
4776 * the pfm_save_regs() routines needs to grab
4777 * the same lock. There are danger in doing
4778 * this because it leaves a window open for
4779 * another task to get access to the context
4780 * and possibly change its state. The one thing
4781 * that is not possible is for the context to disappear
4782 * because we are protected by the VFS layer, i.e.,
4783 * get_fd()/put_fd().
4787 UNPROTECT_CTX(ctx, flags);
4789 wait_task_inactive(task);
4791 PROTECT_CTX(ctx, flags);
4794 * we must recheck to verify if state has changed
4796 if (ctx->ctx_state != old_state) {
4797 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4805 * system-call entry point (must return long)
4808 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4810 struct file *file = NULL;
4811 pfm_context_t *ctx = NULL;
4812 unsigned long flags = 0UL;
4813 void *args_k = NULL;
4814 long ret; /* will expand int return types */
4815 size_t base_sz, sz, xtra_sz = 0;
4816 int narg, completed_args = 0, call_made = 0, cmd_flags;
4817 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4818 int (*getsize)(void *arg, size_t *sz);
4819 #define PFM_MAX_ARGSIZE 4096
4822 * reject any call if perfmon was disabled at initialization
4824 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4826 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4827 DPRINT(("invalid cmd=%d\n", cmd));
4831 func = pfm_cmd_tab[cmd].cmd_func;
4832 narg = pfm_cmd_tab[cmd].cmd_narg;
4833 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4834 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4835 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4837 if (unlikely(func == NULL)) {
4838 DPRINT(("invalid cmd=%d\n", cmd));
4842 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4850 * check if number of arguments matches what the command expects
4852 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4856 sz = xtra_sz + base_sz*count;
4858 * limit abuse to min page size
4860 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4861 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4866 * allocate default-sized argument buffer
4868 if (likely(count && args_k == NULL)) {
4869 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4870 if (args_k == NULL) return -ENOMEM;
4878 * assume sz = 0 for command without parameters
4880 if (sz && copy_from_user(args_k, arg, sz)) {
4881 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4886 * check if command supports extra parameters
4888 if (completed_args == 0 && getsize) {
4890 * get extra parameters size (based on main argument)
4892 ret = (*getsize)(args_k, &xtra_sz);
4893 if (ret) goto error_args;
4897 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4899 /* retry if necessary */
4900 if (likely(xtra_sz)) goto restart_args;
4903 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4908 if (unlikely(file == NULL)) {
4909 DPRINT(("invalid fd %d\n", fd));
4912 if (unlikely(PFM_IS_FILE(file) == 0)) {
4913 DPRINT(("fd %d not related to perfmon\n", fd));
4917 ctx = (pfm_context_t *)file->private_data;
4918 if (unlikely(ctx == NULL)) {
4919 DPRINT(("no context for fd %d\n", fd));
4922 prefetch(&ctx->ctx_state);
4924 PROTECT_CTX(ctx, flags);
4927 * check task is stopped
4929 ret = pfm_check_task_state(ctx, cmd, flags);
4930 if (unlikely(ret)) goto abort_locked;
4933 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4939 DPRINT(("context unlocked\n"));
4940 UNPROTECT_CTX(ctx, flags);
4944 /* copy argument back to user, if needed */
4945 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4950 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4956 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4958 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4959 pfm_ovfl_ctrl_t rst_ctrl;
4963 state = ctx->ctx_state;
4965 * Unlock sampling buffer and reset index atomically
4966 * XXX: not really needed when blocking
4968 if (CTX_HAS_SMPL(ctx)) {
4970 rst_ctrl.bits.mask_monitoring = 0;
4971 rst_ctrl.bits.reset_ovfl_pmds = 0;
4973 if (state == PFM_CTX_LOADED)
4974 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4976 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4978 rst_ctrl.bits.mask_monitoring = 0;
4979 rst_ctrl.bits.reset_ovfl_pmds = 1;
4983 if (rst_ctrl.bits.reset_ovfl_pmds) {
4984 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4986 if (rst_ctrl.bits.mask_monitoring == 0) {
4987 DPRINT(("resuming monitoring\n"));
4988 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4990 DPRINT(("stopping monitoring\n"));
4991 //pfm_stop_monitoring(current, regs);
4993 ctx->ctx_state = PFM_CTX_LOADED;
4998 * context MUST BE LOCKED when calling
4999 * can only be called for current
5002 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
5006 DPRINT(("entering for [%d]\n", current->pid));
5008 ret = pfm_context_unload(ctx, NULL, 0, regs);
5010 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", current->pid, ret);
5014 * and wakeup controlling task, indicating we are now disconnected
5016 wake_up_interruptible(&ctx->ctx_zombieq);
5019 * given that context is still locked, the controlling
5020 * task will only get access when we return from
5021 * pfm_handle_work().
5025 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5027 * pfm_handle_work() can be called with interrupts enabled
5028 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5029 * call may sleep, therefore we must re-enable interrupts
5030 * to avoid deadlocks. It is safe to do so because this function
5031 * is called ONLY when returning to user level (PUStk=1), in which case
5032 * there is no risk of kernel stack overflow due to deep
5033 * interrupt nesting.
5036 pfm_handle_work(void)
5039 struct pt_regs *regs;
5040 unsigned long flags, dummy_flags;
5041 unsigned long ovfl_regs;
5042 unsigned int reason;
5045 ctx = PFM_GET_CTX(current);
5047 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5051 PROTECT_CTX(ctx, flags);
5053 PFM_SET_WORK_PENDING(current, 0);
5055 pfm_clear_task_notify();
5057 regs = task_pt_regs(current);
5060 * extract reason for being here and clear
5062 reason = ctx->ctx_fl_trap_reason;
5063 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5064 ovfl_regs = ctx->ctx_ovfl_regs[0];
5066 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5069 * must be done before we check for simple-reset mode
5071 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5074 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5075 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5078 * restore interrupt mask to what it was on entry.
5079 * Could be enabled/diasbled.
5081 UNPROTECT_CTX(ctx, flags);
5084 * force interrupt enable because of down_interruptible()
5088 DPRINT(("before block sleeping\n"));
5091 * may go through without blocking on SMP systems
5092 * if restart has been received already by the time we call down()
5094 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5096 DPRINT(("after block sleeping ret=%d\n", ret));
5099 * lock context and mask interrupts again
5100 * We save flags into a dummy because we may have
5101 * altered interrupts mask compared to entry in this
5104 PROTECT_CTX(ctx, dummy_flags);
5107 * we need to read the ovfl_regs only after wake-up
5108 * because we may have had pfm_write_pmds() in between
5109 * and that can changed PMD values and therefore
5110 * ovfl_regs is reset for these new PMD values.
5112 ovfl_regs = ctx->ctx_ovfl_regs[0];
5114 if (ctx->ctx_fl_going_zombie) {
5116 DPRINT(("context is zombie, bailing out\n"));
5117 pfm_context_force_terminate(ctx, regs);
5121 * in case of interruption of down() we don't restart anything
5123 if (ret < 0) goto nothing_to_do;
5126 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5127 ctx->ctx_ovfl_regs[0] = 0UL;
5131 * restore flags as they were upon entry
5133 UNPROTECT_CTX(ctx, flags);
5137 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5139 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5140 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5144 DPRINT(("waking up somebody\n"));
5146 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5149 * safe, we are not in intr handler, nor in ctxsw when
5152 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5158 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5160 pfm_msg_t *msg = NULL;
5162 if (ctx->ctx_fl_no_msg == 0) {
5163 msg = pfm_get_new_msg(ctx);
5165 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5169 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5170 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5171 msg->pfm_ovfl_msg.msg_active_set = 0;
5172 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5173 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5174 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5175 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5176 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5179 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5185 return pfm_notify_user(ctx, msg);
5189 pfm_end_notify_user(pfm_context_t *ctx)
5193 msg = pfm_get_new_msg(ctx);
5195 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5199 memset(msg, 0, sizeof(*msg));
5201 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5202 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5203 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5205 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5210 return pfm_notify_user(ctx, msg);
5214 * main overflow processing routine.
5215 * it can be called from the interrupt path or explicitely during the context switch code
5218 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5220 pfm_ovfl_arg_t *ovfl_arg;
5222 unsigned long old_val, ovfl_val, new_val;
5223 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5224 unsigned long tstamp;
5225 pfm_ovfl_ctrl_t ovfl_ctrl;
5226 unsigned int i, has_smpl;
5227 int must_notify = 0;
5229 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5232 * sanity test. Should never happen
5234 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5236 tstamp = ia64_get_itc();
5237 mask = pmc0 >> PMU_FIRST_COUNTER;
5238 ovfl_val = pmu_conf->ovfl_val;
5239 has_smpl = CTX_HAS_SMPL(ctx);
5241 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5242 "used_pmds=0x%lx\n",
5244 task ? task->pid: -1,
5245 (regs ? regs->cr_iip : 0),
5246 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5247 ctx->ctx_used_pmds[0]));
5251 * first we update the virtual counters
5252 * assume there was a prior ia64_srlz_d() issued
5254 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5256 /* skip pmd which did not overflow */
5257 if ((mask & 0x1) == 0) continue;
5260 * Note that the pmd is not necessarily 0 at this point as qualified events
5261 * may have happened before the PMU was frozen. The residual count is not
5262 * taken into consideration here but will be with any read of the pmd via
5265 old_val = new_val = ctx->ctx_pmds[i].val;
5266 new_val += 1 + ovfl_val;
5267 ctx->ctx_pmds[i].val = new_val;
5270 * check for overflow condition
5272 if (likely(old_val > new_val)) {
5273 ovfl_pmds |= 1UL << i;
5274 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5277 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5281 ia64_get_pmd(i) & ovfl_val,
5287 * there was no 64-bit overflow, nothing else to do
5289 if (ovfl_pmds == 0UL) return;
5292 * reset all control bits
5298 * if a sampling format module exists, then we "cache" the overflow by
5299 * calling the module's handler() routine.
5302 unsigned long start_cycles, end_cycles;
5303 unsigned long pmd_mask;
5305 int this_cpu = smp_processor_id();
5307 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5308 ovfl_arg = &ctx->ctx_ovfl_arg;
5310 prefetch(ctx->ctx_smpl_hdr);
5312 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5316 if ((pmd_mask & 0x1) == 0) continue;
5318 ovfl_arg->ovfl_pmd = (unsigned char )i;
5319 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5320 ovfl_arg->active_set = 0;
5321 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5322 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5324 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5325 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5326 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5329 * copy values of pmds of interest. Sampling format may copy them
5330 * into sampling buffer.
5333 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5334 if ((smpl_pmds & 0x1) == 0) continue;
5335 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5336 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5340 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5342 start_cycles = ia64_get_itc();
5345 * call custom buffer format record (handler) routine
5347 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5349 end_cycles = ia64_get_itc();
5352 * For those controls, we take the union because they have
5353 * an all or nothing behavior.
5355 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5356 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5357 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5359 * build the bitmask of pmds to reset now
5361 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5363 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5366 * when the module cannot handle the rest of the overflows, we abort right here
5368 if (ret && pmd_mask) {
5369 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5370 pmd_mask<<PMU_FIRST_COUNTER));
5373 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5375 ovfl_pmds &= ~reset_pmds;
5378 * when no sampling module is used, then the default
5379 * is to notify on overflow if requested by user
5381 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5382 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5383 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5384 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5386 * if needed, we reset all overflowed pmds
5388 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5391 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5394 * reset the requested PMD registers using the short reset values
5397 unsigned long bm = reset_pmds;
5398 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5401 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5403 * keep track of what to reset when unblocking
5405 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5408 * check for blocking context
5410 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5412 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5415 * set the perfmon specific checking pending work for the task
5417 PFM_SET_WORK_PENDING(task, 1);
5420 * when coming from ctxsw, current still points to the
5421 * previous task, therefore we must work with task and not current.
5423 pfm_set_task_notify(task);
5426 * defer until state is changed (shorten spin window). the context is locked
5427 * anyway, so the signal receiver would come spin for nothing.
5432 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5433 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5434 PFM_GET_WORK_PENDING(task),
5435 ctx->ctx_fl_trap_reason,
5438 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5440 * in case monitoring must be stopped, we toggle the psr bits
5442 if (ovfl_ctrl.bits.mask_monitoring) {
5443 pfm_mask_monitoring(task);
5444 ctx->ctx_state = PFM_CTX_MASKED;
5445 ctx->ctx_fl_can_restart = 1;
5449 * send notification now
5451 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5456 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5458 task ? task->pid : -1,
5464 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5465 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5466 * come here as zombie only if the task is the current task. In which case, we
5467 * can access the PMU hardware directly.
5469 * Note that zombies do have PM_VALID set. So here we do the minimal.
5471 * In case the context was zombified it could not be reclaimed at the time
5472 * the monitoring program exited. At this point, the PMU reservation has been
5473 * returned, the sampiing buffer has been freed. We must convert this call
5474 * into a spurious interrupt. However, we must also avoid infinite overflows
5475 * by stopping monitoring for this task. We can only come here for a per-task
5476 * context. All we need to do is to stop monitoring using the psr bits which
5477 * are always task private. By re-enabling secure montioring, we ensure that
5478 * the monitored task will not be able to re-activate monitoring.
5479 * The task will eventually be context switched out, at which point the context
5480 * will be reclaimed (that includes releasing ownership of the PMU).
5482 * So there might be a window of time where the number of per-task session is zero
5483 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5484 * context. This is safe because if a per-task session comes in, it will push this one
5485 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5486 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5487 * also push our zombie context out.
5489 * Overall pretty hairy stuff....
5491 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5493 ia64_psr(regs)->up = 0;
5494 ia64_psr(regs)->sp = 1;
5499 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5501 struct task_struct *task;
5503 unsigned long flags;
5505 int this_cpu = smp_processor_id();
5508 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5511 * srlz.d done before arriving here
5513 pmc0 = ia64_get_pmc(0);
5515 task = GET_PMU_OWNER();
5516 ctx = GET_PMU_CTX();
5519 * if we have some pending bits set
5520 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5522 if (PMC0_HAS_OVFL(pmc0) && task) {
5524 * we assume that pmc0.fr is always set here
5528 if (!ctx) goto report_spurious1;
5530 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5531 goto report_spurious2;
5533 PROTECT_CTX_NOPRINT(ctx, flags);
5535 pfm_overflow_handler(task, ctx, pmc0, regs);
5537 UNPROTECT_CTX_NOPRINT(ctx, flags);
5540 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5544 * keep it unfrozen at all times
5551 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5552 this_cpu, task->pid);
5556 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5564 pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5566 unsigned long start_cycles, total_cycles;
5567 unsigned long min, max;
5571 this_cpu = get_cpu();
5572 if (likely(!pfm_alt_intr_handler)) {
5573 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5574 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5576 start_cycles = ia64_get_itc();
5578 ret = pfm_do_interrupt_handler(irq, arg, regs);
5580 total_cycles = ia64_get_itc();
5583 * don't measure spurious interrupts
5585 if (likely(ret == 0)) {
5586 total_cycles -= start_cycles;
5588 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5589 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5591 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5595 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5598 put_cpu_no_resched();
5603 * /proc/perfmon interface, for debug only
5606 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5609 pfm_proc_start(struct seq_file *m, loff_t *pos)
5612 return PFM_PROC_SHOW_HEADER;
5615 while (*pos <= NR_CPUS) {
5616 if (cpu_online(*pos - 1)) {
5617 return (void *)*pos;
5625 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5628 return pfm_proc_start(m, pos);
5632 pfm_proc_stop(struct seq_file *m, void *v)
5637 pfm_proc_show_header(struct seq_file *m)
5639 struct list_head * pos;
5640 pfm_buffer_fmt_t * entry;
5641 unsigned long flags;
5644 "perfmon version : %u.%u\n"
5647 "expert mode : %s\n"
5648 "ovfl_mask : 0x%lx\n"
5649 "PMU flags : 0x%x\n",
5650 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5652 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5653 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5660 "proc_sessions : %u\n"
5661 "sys_sessions : %u\n"
5662 "sys_use_dbregs : %u\n"
5663 "ptrace_use_dbregs : %u\n",
5664 pfm_sessions.pfs_task_sessions,
5665 pfm_sessions.pfs_sys_sessions,
5666 pfm_sessions.pfs_sys_use_dbregs,
5667 pfm_sessions.pfs_ptrace_use_dbregs);
5671 spin_lock(&pfm_buffer_fmt_lock);
5673 list_for_each(pos, &pfm_buffer_fmt_list) {
5674 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5675 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5686 entry->fmt_uuid[10],
5687 entry->fmt_uuid[11],
5688 entry->fmt_uuid[12],
5689 entry->fmt_uuid[13],
5690 entry->fmt_uuid[14],
5691 entry->fmt_uuid[15],
5694 spin_unlock(&pfm_buffer_fmt_lock);
5699 pfm_proc_show(struct seq_file *m, void *v)
5705 if (v == PFM_PROC_SHOW_HEADER) {
5706 pfm_proc_show_header(m);
5710 /* show info for CPU (v - 1) */
5714 "CPU%-2d overflow intrs : %lu\n"
5715 "CPU%-2d overflow cycles : %lu\n"
5716 "CPU%-2d overflow min : %lu\n"
5717 "CPU%-2d overflow max : %lu\n"
5718 "CPU%-2d smpl handler calls : %lu\n"
5719 "CPU%-2d smpl handler cycles : %lu\n"
5720 "CPU%-2d spurious intrs : %lu\n"
5721 "CPU%-2d replay intrs : %lu\n"
5722 "CPU%-2d syst_wide : %d\n"
5723 "CPU%-2d dcr_pp : %d\n"
5724 "CPU%-2d exclude idle : %d\n"
5725 "CPU%-2d owner : %d\n"
5726 "CPU%-2d context : %p\n"
5727 "CPU%-2d activations : %lu\n",
5728 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5729 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5730 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5731 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5732 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5733 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5734 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5735 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5736 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5737 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5738 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5739 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5740 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5741 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5743 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5745 psr = pfm_get_psr();
5750 "CPU%-2d psr : 0x%lx\n"
5751 "CPU%-2d pmc0 : 0x%lx\n",
5753 cpu, ia64_get_pmc(0));
5755 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5756 if (PMC_IS_COUNTING(i) == 0) continue;
5758 "CPU%-2d pmc%u : 0x%lx\n"
5759 "CPU%-2d pmd%u : 0x%lx\n",
5760 cpu, i, ia64_get_pmc(i),
5761 cpu, i, ia64_get_pmd(i));
5767 struct seq_operations pfm_seq_ops = {
5768 .start = pfm_proc_start,
5769 .next = pfm_proc_next,
5770 .stop = pfm_proc_stop,
5771 .show = pfm_proc_show
5775 pfm_proc_open(struct inode *inode, struct file *file)
5777 return seq_open(file, &pfm_seq_ops);
5782 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5783 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5784 * is active or inactive based on mode. We must rely on the value in
5785 * local_cpu_data->pfm_syst_info
5788 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5790 struct pt_regs *regs;
5792 unsigned long dcr_pp;
5794 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5797 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5798 * on every CPU, so we can rely on the pid to identify the idle task.
5800 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5801 regs = task_pt_regs(task);
5802 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5806 * if monitoring has started
5809 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5811 * context switching in?
5814 /* mask monitoring for the idle task */
5815 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5821 * context switching out
5822 * restore monitoring for next task
5824 * Due to inlining this odd if-then-else construction generates
5827 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5836 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5838 struct task_struct *task = ctx->ctx_task;
5840 ia64_psr(regs)->up = 0;
5841 ia64_psr(regs)->sp = 1;
5843 if (GET_PMU_OWNER() == task) {
5844 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5845 SET_PMU_OWNER(NULL, NULL);
5849 * disconnect the task from the context and vice-versa
5851 PFM_SET_WORK_PENDING(task, 0);
5853 task->thread.pfm_context = NULL;
5854 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5856 DPRINT(("force cleanup for [%d]\n", task->pid));
5861 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5864 pfm_save_regs(struct task_struct *task)
5867 struct thread_struct *t;
5868 unsigned long flags;
5872 ctx = PFM_GET_CTX(task);
5873 if (ctx == NULL) return;
5877 * we always come here with interrupts ALREADY disabled by
5878 * the scheduler. So we simply need to protect against concurrent
5879 * access, not CPU concurrency.
5881 flags = pfm_protect_ctx_ctxsw(ctx);
5883 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5884 struct pt_regs *regs = task_pt_regs(task);
5888 pfm_force_cleanup(ctx, regs);
5890 BUG_ON(ctx->ctx_smpl_hdr);
5892 pfm_unprotect_ctx_ctxsw(ctx, flags);
5894 pfm_context_free(ctx);
5899 * save current PSR: needed because we modify it
5902 psr = pfm_get_psr();
5904 BUG_ON(psr & (IA64_PSR_I));
5908 * This is the last instruction which may generate an overflow
5910 * We do not need to set psr.sp because, it is irrelevant in kernel.
5911 * It will be restored from ipsr when going back to user level
5916 * keep a copy of psr.up (for reload)
5918 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5921 * release ownership of this PMU.
5922 * PM interrupts are masked, so nothing
5925 SET_PMU_OWNER(NULL, NULL);
5928 * we systematically save the PMD as we have no
5929 * guarantee we will be schedule at that same
5932 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5935 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5936 * we will need it on the restore path to check
5937 * for pending overflow.
5939 t->pmcs[0] = ia64_get_pmc(0);
5942 * unfreeze PMU if had pending overflows
5944 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5947 * finally, allow context access.
5948 * interrupts will still be masked after this call.
5950 pfm_unprotect_ctx_ctxsw(ctx, flags);
5953 #else /* !CONFIG_SMP */
5955 pfm_save_regs(struct task_struct *task)
5960 ctx = PFM_GET_CTX(task);
5961 if (ctx == NULL) return;
5964 * save current PSR: needed because we modify it
5966 psr = pfm_get_psr();
5968 BUG_ON(psr & (IA64_PSR_I));
5972 * This is the last instruction which may generate an overflow
5974 * We do not need to set psr.sp because, it is irrelevant in kernel.
5975 * It will be restored from ipsr when going back to user level
5980 * keep a copy of psr.up (for reload)
5982 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5986 pfm_lazy_save_regs (struct task_struct *task)
5989 struct thread_struct *t;
5990 unsigned long flags;
5992 { u64 psr = pfm_get_psr();
5993 BUG_ON(psr & IA64_PSR_UP);
5996 ctx = PFM_GET_CTX(task);
6000 * we need to mask PMU overflow here to
6001 * make sure that we maintain pmc0 until
6002 * we save it. overflow interrupts are
6003 * treated as spurious if there is no
6006 * XXX: I don't think this is necessary
6008 PROTECT_CTX(ctx,flags);
6011 * release ownership of this PMU.
6012 * must be done before we save the registers.
6014 * after this call any PMU interrupt is treated
6017 SET_PMU_OWNER(NULL, NULL);
6020 * save all the pmds we use
6022 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
6025 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6026 * it is needed to check for pended overflow
6027 * on the restore path
6029 t->pmcs[0] = ia64_get_pmc(0);
6032 * unfreeze PMU if had pending overflows
6034 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6037 * now get can unmask PMU interrupts, they will
6038 * be treated as purely spurious and we will not
6039 * lose any information
6041 UNPROTECT_CTX(ctx,flags);
6043 #endif /* CONFIG_SMP */
6047 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6050 pfm_load_regs (struct task_struct *task)
6053 struct thread_struct *t;
6054 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6055 unsigned long flags;
6057 int need_irq_resend;
6059 ctx = PFM_GET_CTX(task);
6060 if (unlikely(ctx == NULL)) return;
6062 BUG_ON(GET_PMU_OWNER());
6066 * possible on unload
6068 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6071 * we always come here with interrupts ALREADY disabled by
6072 * the scheduler. So we simply need to protect against concurrent
6073 * access, not CPU concurrency.
6075 flags = pfm_protect_ctx_ctxsw(ctx);
6076 psr = pfm_get_psr();
6078 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6080 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6081 BUG_ON(psr & IA64_PSR_I);
6083 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6084 struct pt_regs *regs = task_pt_regs(task);
6086 BUG_ON(ctx->ctx_smpl_hdr);
6088 pfm_force_cleanup(ctx, regs);
6090 pfm_unprotect_ctx_ctxsw(ctx, flags);
6093 * this one (kmalloc'ed) is fine with interrupts disabled
6095 pfm_context_free(ctx);
6101 * we restore ALL the debug registers to avoid picking up
6104 if (ctx->ctx_fl_using_dbreg) {
6105 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6106 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6109 * retrieve saved psr.up
6111 psr_up = ctx->ctx_saved_psr_up;
6114 * if we were the last user of the PMU on that CPU,
6115 * then nothing to do except restore psr
6117 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6120 * retrieve partial reload masks (due to user modifications)
6122 pmc_mask = ctx->ctx_reload_pmcs[0];
6123 pmd_mask = ctx->ctx_reload_pmds[0];
6127 * To avoid leaking information to the user level when psr.sp=0,
6128 * we must reload ALL implemented pmds (even the ones we don't use).
6129 * In the kernel we only allow PFM_READ_PMDS on registers which
6130 * we initialized or requested (sampling) so there is no risk there.
6132 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6135 * ALL accessible PMCs are systematically reloaded, unused registers
6136 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6137 * up stale configuration.
6139 * PMC0 is never in the mask. It is always restored separately.
6141 pmc_mask = ctx->ctx_all_pmcs[0];
6144 * when context is MASKED, we will restore PMC with plm=0
6145 * and PMD with stale information, but that's ok, nothing
6148 * XXX: optimize here
6150 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6151 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6154 * check for pending overflow at the time the state
6157 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6159 * reload pmc0 with the overflow information
6160 * On McKinley PMU, this will trigger a PMU interrupt
6162 ia64_set_pmc(0, t->pmcs[0]);
6167 * will replay the PMU interrupt
6169 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6171 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6175 * we just did a reload, so we reset the partial reload fields
6177 ctx->ctx_reload_pmcs[0] = 0UL;
6178 ctx->ctx_reload_pmds[0] = 0UL;
6180 SET_LAST_CPU(ctx, smp_processor_id());
6183 * dump activation value for this PMU
6187 * record current activation for this context
6189 SET_ACTIVATION(ctx);
6192 * establish new ownership.
6194 SET_PMU_OWNER(task, ctx);
6197 * restore the psr.up bit. measurement
6199 * no PMU interrupt can happen at this point
6200 * because we still have interrupts disabled.
6202 if (likely(psr_up)) pfm_set_psr_up();
6205 * allow concurrent access to context
6207 pfm_unprotect_ctx_ctxsw(ctx, flags);
6209 #else /* !CONFIG_SMP */
6211 * reload PMU state for UP kernels
6212 * in 2.5 we come here with interrupts disabled
6215 pfm_load_regs (struct task_struct *task)
6217 struct thread_struct *t;
6219 struct task_struct *owner;
6220 unsigned long pmd_mask, pmc_mask;
6222 int need_irq_resend;
6224 owner = GET_PMU_OWNER();
6225 ctx = PFM_GET_CTX(task);
6227 psr = pfm_get_psr();
6229 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6230 BUG_ON(psr & IA64_PSR_I);
6233 * we restore ALL the debug registers to avoid picking up
6236 * This must be done even when the task is still the owner
6237 * as the registers may have been modified via ptrace()
6238 * (not perfmon) by the previous task.
6240 if (ctx->ctx_fl_using_dbreg) {
6241 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6242 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6246 * retrieved saved psr.up
6248 psr_up = ctx->ctx_saved_psr_up;
6249 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6252 * short path, our state is still there, just
6253 * need to restore psr and we go
6255 * we do not touch either PMC nor PMD. the psr is not touched
6256 * by the overflow_handler. So we are safe w.r.t. to interrupt
6257 * concurrency even without interrupt masking.
6259 if (likely(owner == task)) {
6260 if (likely(psr_up)) pfm_set_psr_up();
6265 * someone else is still using the PMU, first push it out and
6266 * then we'll be able to install our stuff !
6268 * Upon return, there will be no owner for the current PMU
6270 if (owner) pfm_lazy_save_regs(owner);
6273 * To avoid leaking information to the user level when psr.sp=0,
6274 * we must reload ALL implemented pmds (even the ones we don't use).
6275 * In the kernel we only allow PFM_READ_PMDS on registers which
6276 * we initialized or requested (sampling) so there is no risk there.
6278 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6281 * ALL accessible PMCs are systematically reloaded, unused registers
6282 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6283 * up stale configuration.
6285 * PMC0 is never in the mask. It is always restored separately
6287 pmc_mask = ctx->ctx_all_pmcs[0];
6289 pfm_restore_pmds(t->pmds, pmd_mask);
6290 pfm_restore_pmcs(t->pmcs, pmc_mask);
6293 * check for pending overflow at the time the state
6296 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6298 * reload pmc0 with the overflow information
6299 * On McKinley PMU, this will trigger a PMU interrupt
6301 ia64_set_pmc(0, t->pmcs[0]);
6307 * will replay the PMU interrupt
6309 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6311 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6315 * establish new ownership.
6317 SET_PMU_OWNER(task, ctx);
6320 * restore the psr.up bit. measurement
6322 * no PMU interrupt can happen at this point
6323 * because we still have interrupts disabled.
6325 if (likely(psr_up)) pfm_set_psr_up();
6327 #endif /* CONFIG_SMP */
6330 * this function assumes monitoring is stopped
6333 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6336 unsigned long mask2, val, pmd_val, ovfl_val;
6337 int i, can_access_pmu = 0;
6341 * is the caller the task being monitored (or which initiated the
6342 * session for system wide measurements)
6344 is_self = ctx->ctx_task == task ? 1 : 0;
6347 * can access PMU is task is the owner of the PMU state on the current CPU
6348 * or if we are running on the CPU bound to the context in system-wide mode
6349 * (that is not necessarily the task the context is attached to in this mode).
6350 * In system-wide we always have can_access_pmu true because a task running on an
6351 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6353 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6354 if (can_access_pmu) {
6356 * Mark the PMU as not owned
6357 * This will cause the interrupt handler to do nothing in case an overflow
6358 * interrupt was in-flight
6359 * This also guarantees that pmc0 will contain the final state
6360 * It virtually gives us full control on overflow processing from that point
6363 SET_PMU_OWNER(NULL, NULL);
6364 DPRINT(("releasing ownership\n"));
6367 * read current overflow status:
6369 * we are guaranteed to read the final stable state
6372 pmc0 = ia64_get_pmc(0); /* slow */
6375 * reset freeze bit, overflow status information destroyed
6379 pmc0 = task->thread.pmcs[0];
6381 * clear whatever overflow status bits there were
6383 task->thread.pmcs[0] = 0;
6385 ovfl_val = pmu_conf->ovfl_val;
6387 * we save all the used pmds
6388 * we take care of overflows for counting PMDs
6390 * XXX: sampling situation is not taken into account here
6392 mask2 = ctx->ctx_used_pmds[0];
6394 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6396 for (i = 0; mask2; i++, mask2>>=1) {
6398 /* skip non used pmds */
6399 if ((mask2 & 0x1) == 0) continue;
6402 * can access PMU always true in system wide mode
6404 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6406 if (PMD_IS_COUNTING(i)) {
6407 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6410 ctx->ctx_pmds[i].val,
6414 * we rebuild the full 64 bit value of the counter
6416 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6419 * now everything is in ctx_pmds[] and we need
6420 * to clear the saved context from save_regs() such that
6421 * pfm_read_pmds() gets the correct value
6426 * take care of overflow inline
6428 if (pmc0 & (1UL << i)) {
6429 val += 1 + ovfl_val;
6430 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6434 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, i, val, pmd_val));
6436 if (is_self) task->thread.pmds[i] = pmd_val;
6438 ctx->ctx_pmds[i].val = val;
6442 static struct irqaction perfmon_irqaction = {
6443 .handler = pfm_interrupt_handler,
6444 .flags = SA_INTERRUPT,
6449 pfm_alt_save_pmu_state(void *data)
6451 struct pt_regs *regs;
6453 regs = task_pt_regs(current);
6455 DPRINT(("called\n"));
6458 * should not be necessary but
6459 * let's take not risk
6463 ia64_psr(regs)->pp = 0;
6466 * This call is required
6467 * May cause a spurious interrupt on some processors
6475 pfm_alt_restore_pmu_state(void *data)
6477 struct pt_regs *regs;
6479 regs = task_pt_regs(current);
6481 DPRINT(("called\n"));
6484 * put PMU back in state expected
6489 ia64_psr(regs)->pp = 0;
6492 * perfmon runs with PMU unfrozen at all times
6500 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6505 /* some sanity checks */
6506 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6508 /* do the easy test first */
6509 if (pfm_alt_intr_handler) return -EBUSY;
6511 /* one at a time in the install or remove, just fail the others */
6512 if (!spin_trylock(&pfm_alt_install_check)) {
6516 /* reserve our session */
6517 for_each_online_cpu(reserve_cpu) {
6518 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6519 if (ret) goto cleanup_reserve;
6522 /* save the current system wide pmu states */
6523 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 0, 1);
6525 DPRINT(("on_each_cpu() failed: %d\n", ret));
6526 goto cleanup_reserve;
6529 /* officially change to the alternate interrupt handler */
6530 pfm_alt_intr_handler = hdl;
6532 spin_unlock(&pfm_alt_install_check);
6537 for_each_online_cpu(i) {
6538 /* don't unreserve more than we reserved */
6539 if (i >= reserve_cpu) break;
6541 pfm_unreserve_session(NULL, 1, i);
6544 spin_unlock(&pfm_alt_install_check);
6548 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6551 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6556 if (hdl == NULL) return -EINVAL;
6558 /* cannot remove someone else's handler! */
6559 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6561 /* one at a time in the install or remove, just fail the others */
6562 if (!spin_trylock(&pfm_alt_install_check)) {
6566 pfm_alt_intr_handler = NULL;
6568 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 0, 1);
6570 DPRINT(("on_each_cpu() failed: %d\n", ret));
6573 for_each_online_cpu(i) {
6574 pfm_unreserve_session(NULL, 1, i);
6577 spin_unlock(&pfm_alt_install_check);
6581 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6584 * perfmon initialization routine, called from the initcall() table
6586 static int init_pfm_fs(void);
6594 family = local_cpu_data->family;
6599 if ((*p)->probe() == 0) goto found;
6600 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6611 static struct file_operations pfm_proc_fops = {
6612 .open = pfm_proc_open,
6614 .llseek = seq_lseek,
6615 .release = seq_release,
6621 unsigned int n, n_counters, i;
6623 printk("perfmon: version %u.%u IRQ %u\n",
6626 IA64_PERFMON_VECTOR);
6628 if (pfm_probe_pmu()) {
6629 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6630 local_cpu_data->family);
6635 * compute the number of implemented PMD/PMC from the
6636 * description tables
6639 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6640 if (PMC_IS_IMPL(i) == 0) continue;
6641 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6644 pmu_conf->num_pmcs = n;
6646 n = 0; n_counters = 0;
6647 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6648 if (PMD_IS_IMPL(i) == 0) continue;
6649 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6651 if (PMD_IS_COUNTING(i)) n_counters++;
6653 pmu_conf->num_pmds = n;
6654 pmu_conf->num_counters = n_counters;
6657 * sanity checks on the number of debug registers
6659 if (pmu_conf->use_rr_dbregs) {
6660 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6661 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6665 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6666 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6672 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6676 pmu_conf->num_counters,
6677 ffz(pmu_conf->ovfl_val));
6680 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6681 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6687 * create /proc/perfmon (mostly for debugging purposes)
6689 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6690 if (perfmon_dir == NULL) {
6691 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6696 * install customized file operations for /proc/perfmon entry
6698 perfmon_dir->proc_fops = &pfm_proc_fops;
6701 * create /proc/sys/kernel/perfmon (for debugging purposes)
6703 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6706 * initialize all our spinlocks
6708 spin_lock_init(&pfm_sessions.pfs_lock);
6709 spin_lock_init(&pfm_buffer_fmt_lock);
6713 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6718 __initcall(pfm_init);
6721 * this function is called before pfm_init()
6724 pfm_init_percpu (void)
6726 static int first_time=1;
6728 * make sure no measurement is active
6729 * (may inherit programmed PMCs from EFI).
6735 * we run with the PMU not frozen at all times
6740 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6744 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6749 * used for debug purposes only
6752 dump_pmu_state(const char *from)
6754 struct task_struct *task;
6755 struct thread_struct *t;
6756 struct pt_regs *regs;
6758 unsigned long psr, dcr, info, flags;
6761 local_irq_save(flags);
6763 this_cpu = smp_processor_id();
6764 regs = task_pt_regs(current);
6765 info = PFM_CPUINFO_GET();
6766 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6768 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6769 local_irq_restore(flags);
6773 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6780 task = GET_PMU_OWNER();
6781 ctx = GET_PMU_CTX();
6783 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6785 psr = pfm_get_psr();
6787 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",
6790 psr & IA64_PSR_PP ? 1 : 0,
6791 psr & IA64_PSR_UP ? 1 : 0,
6792 dcr & IA64_DCR_PP ? 1 : 0,
6795 ia64_psr(regs)->pp);
6797 ia64_psr(regs)->up = 0;
6798 ia64_psr(regs)->pp = 0;
6800 t = ¤t->thread;
6802 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6803 if (PMC_IS_IMPL(i) == 0) continue;
6804 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6807 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6808 if (PMD_IS_IMPL(i) == 0) continue;
6809 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6813 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6816 ctx->ctx_smpl_vaddr,
6820 ctx->ctx_saved_psr_up);
6822 local_irq_restore(flags);
6826 * called from process.c:copy_thread(). task is new child.
6829 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6831 struct thread_struct *thread;
6833 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6835 thread = &task->thread;
6838 * cut links inherited from parent (current)
6840 thread->pfm_context = NULL;
6842 PFM_SET_WORK_PENDING(task, 0);
6845 * the psr bits are already set properly in copy_threads()
6848 #else /* !CONFIG_PERFMON */
6850 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6854 #endif /* CONFIG_PERFMON */