2 * This file implements the perfmon-2 subsystem which is used
3 * to program the IA-64 Performance Monitoring Unit (PMU).
5 * The initial version of perfmon.c was written by
6 * Ganesh Venkitachalam, IBM Corp.
8 * Then it was modified for perfmon-1.x by Stephane Eranian and
9 * David Mosberger, Hewlett Packard Co.
11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
12 * by Stephane Eranian, Hewlett Packard Co.
14 * Copyright (C) 1999-2003 Hewlett Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 * David Mosberger-Tang <davidm@hpl.hp.com>
18 * More information about perfmon available at:
19 * http://www.hpl.hp.com/research/linux/perfmon
22 #include <linux/config.h>
23 #include <linux/module.h>
24 #include <linux/kernel.h>
25 #include <linux/sched.h>
26 #include <linux/interrupt.h>
27 #include <linux/smp_lock.h>
28 #include <linux/proc_fs.h>
29 #include <linux/seq_file.h>
30 #include <linux/init.h>
31 #include <linux/vmalloc.h>
33 #include <linux/sysctl.h>
34 #include <linux/list.h>
35 #include <linux/file.h>
36 #include <linux/poll.h>
37 #include <linux/vfs.h>
38 #include <linux/pagemap.h>
39 #include <linux/mount.h>
40 #include <linux/version.h>
42 #include <asm/bitops.h>
43 #include <asm/errno.h>
44 #include <asm/intrinsics.h>
46 #include <asm/perfmon.h>
47 #include <asm/processor.h>
48 #include <asm/signal.h>
49 #include <asm/system.h>
50 #include <asm/uaccess.h>
51 #include <asm/delay.h>
55 * perfmon context state
57 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
58 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
59 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
60 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
62 #define PFM_INVALID_ACTIVATION (~0UL)
65 * depth of message queue
67 #define PFM_MAX_MSGS 32
68 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
71 * type of a PMU register (bitmask).
73 * bit0 : register implemented
76 * bit4 : pmc has pmc.pm
77 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
78 * bit6-7 : register type
81 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
82 #define PFM_REG_IMPL 0x1 /* register implemented */
83 #define PFM_REG_END 0x2 /* end marker */
84 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
85 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
86 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
87 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
88 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
90 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
91 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
93 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
95 /* i assumed unsigned */
96 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
97 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
99 /* XXX: these assume that register i is implemented */
100 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
101 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
102 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
103 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
105 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
106 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
107 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
108 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
110 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
111 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
113 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
114 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
115 #define PFM_CTX_TASK(h) (h)->ctx_task
117 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
119 /* XXX: does not support more than 64 PMDs */
120 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
121 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
123 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
125 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
126 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
127 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
128 #define PFM_CODE_RR 0 /* requesting code range restriction */
129 #define PFM_DATA_RR 1 /* requestion data range restriction */
131 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
132 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
133 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
135 #define RDEP(x) (1UL<<(x))
138 * context protection macros
140 * - we need to protect against CPU concurrency (spin_lock)
141 * - we need to protect against PMU overflow interrupts (local_irq_disable)
143 * - we need to protect against PMU overflow interrupts (local_irq_disable)
145 * spin_lock_irqsave()/spin_lock_irqrestore():
146 * in SMP: local_irq_disable + spin_lock
147 * in UP : local_irq_disable
149 * spin_lock()/spin_lock():
150 * in UP : removed automatically
151 * in SMP: protect against context accesses from other CPU. interrupts
152 * are not masked. This is useful for the PMU interrupt handler
153 * because we know we will not get PMU concurrency in that code.
155 #define PROTECT_CTX(c, f) \
157 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \
158 spin_lock_irqsave(&(c)->ctx_lock, f); \
159 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
162 #define UNPROTECT_CTX(c, f) \
164 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
165 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
168 #define PROTECT_CTX_NOPRINT(c, f) \
170 spin_lock_irqsave(&(c)->ctx_lock, f); \
174 #define UNPROTECT_CTX_NOPRINT(c, f) \
176 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
180 #define PROTECT_CTX_NOIRQ(c) \
182 spin_lock(&(c)->ctx_lock); \
185 #define UNPROTECT_CTX_NOIRQ(c) \
187 spin_unlock(&(c)->ctx_lock); \
193 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
194 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
195 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
197 #else /* !CONFIG_SMP */
198 #define SET_ACTIVATION(t) do {} while(0)
199 #define GET_ACTIVATION(t) do {} while(0)
200 #define INC_ACTIVATION(t) do {} while(0)
201 #endif /* CONFIG_SMP */
203 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
204 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
205 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
207 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
208 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
210 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
213 * cmp0 must be the value of pmc0
215 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
217 #define PFMFS_MAGIC 0xa0b4d889
222 #define PFM_DEBUGGING 1
226 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
229 #define DPRINT_ovfl(a) \
231 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
236 * 64-bit software counter structure
238 * the next_reset_type is applied to the next call to pfm_reset_regs()
241 unsigned long val; /* virtual 64bit counter value */
242 unsigned long lval; /* last reset value */
243 unsigned long long_reset; /* reset value on sampling overflow */
244 unsigned long short_reset; /* reset value on overflow */
245 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
246 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
247 unsigned long seed; /* seed for random-number generator */
248 unsigned long mask; /* mask for random-number generator */
249 unsigned int flags; /* notify/do not notify */
250 unsigned long eventid; /* overflow event identifier */
257 unsigned int block:1; /* when 1, task will blocked on user notifications */
258 unsigned int system:1; /* do system wide monitoring */
259 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
260 unsigned int is_sampling:1; /* true if using a custom format */
261 unsigned int excl_idle:1; /* exclude idle task in system wide session */
262 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
263 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
264 unsigned int no_msg:1; /* no message sent on overflow */
265 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
266 unsigned int reserved:22;
267 } pfm_context_flags_t;
269 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
270 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
271 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
275 * perfmon context: encapsulates all the state of a monitoring session
278 typedef struct pfm_context {
279 spinlock_t ctx_lock; /* context protection */
281 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
282 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
284 struct task_struct *ctx_task; /* task to which context is attached */
286 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
288 struct semaphore ctx_restart_sem; /* use for blocking notification mode */
290 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
291 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
292 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
294 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
295 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
296 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
298 unsigned long ctx_pmcs[IA64_NUM_PMC_REGS]; /* saved copies of PMC values */
300 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
301 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
302 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
303 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
305 pfm_counter_t ctx_pmds[IA64_NUM_PMD_REGS]; /* software state for PMDS */
307 u64 ctx_saved_psr_up; /* only contains psr.up value */
309 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
310 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
311 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
313 int ctx_fd; /* file descriptor used my this context */
314 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
316 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
317 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
318 unsigned long ctx_smpl_size; /* size of sampling buffer */
319 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
321 wait_queue_head_t ctx_msgq_wait;
322 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
325 struct fasync_struct *ctx_async_queue;
327 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
331 * magic number used to verify that structure is really
334 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
336 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
339 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
340 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
342 #define SET_LAST_CPU(ctx, v) do {} while(0)
343 #define GET_LAST_CPU(ctx) do {} while(0)
347 #define ctx_fl_block ctx_flags.block
348 #define ctx_fl_system ctx_flags.system
349 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
350 #define ctx_fl_is_sampling ctx_flags.is_sampling
351 #define ctx_fl_excl_idle ctx_flags.excl_idle
352 #define ctx_fl_going_zombie ctx_flags.going_zombie
353 #define ctx_fl_trap_reason ctx_flags.trap_reason
354 #define ctx_fl_no_msg ctx_flags.no_msg
355 #define ctx_fl_can_restart ctx_flags.can_restart
357 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
358 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
361 * global information about all sessions
362 * mostly used to synchronize between system wide and per-process
365 spinlock_t pfs_lock; /* lock the structure */
367 unsigned int pfs_task_sessions; /* number of per task sessions */
368 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
369 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
370 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
371 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
375 * information about a PMC or PMD.
376 * dep_pmd[]: a bitmask of dependent PMD registers
377 * dep_pmc[]: a bitmask of dependent PMC registers
379 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
383 unsigned long default_value; /* power-on default value */
384 unsigned long reserved_mask; /* bitmask of reserved bits */
385 pfm_reg_check_t read_check;
386 pfm_reg_check_t write_check;
387 unsigned long dep_pmd[4];
388 unsigned long dep_pmc[4];
391 /* assume cnum is a valid monitor */
392 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
395 * This structure is initialized at boot time and contains
396 * a description of the PMU main characteristics.
398 * If the probe function is defined, detection is based
399 * on its return value:
400 * - 0 means recognized PMU
401 * - anything else means not supported
402 * When the probe function is not defined, then the pmu_family field
403 * is used and it must match the host CPU family such that:
404 * - cpu->family & config->pmu_family != 0
407 unsigned long ovfl_val; /* overflow value for counters */
409 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
410 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
412 unsigned int num_pmcs; /* number of PMCS: computed at init time */
413 unsigned int num_pmds; /* number of PMDS: computed at init time */
414 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
415 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
417 char *pmu_name; /* PMU family name */
418 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
419 unsigned int flags; /* pmu specific flags */
420 unsigned int num_ibrs; /* number of IBRS: computed at init time */
421 unsigned int num_dbrs; /* number of DBRS: computed at init time */
422 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
423 int (*probe)(void); /* customized probe routine */
424 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
429 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
432 * debug register related type definitions
435 unsigned long ibr_mask:56;
436 unsigned long ibr_plm:4;
437 unsigned long ibr_ig:3;
438 unsigned long ibr_x:1;
442 unsigned long dbr_mask:56;
443 unsigned long dbr_plm:4;
444 unsigned long dbr_ig:2;
445 unsigned long dbr_w:1;
446 unsigned long dbr_r:1;
457 * perfmon command descriptions
460 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
463 unsigned int cmd_narg;
465 int (*cmd_getsize)(void *arg, size_t *sz);
468 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
469 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
470 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
471 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
474 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
475 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
476 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
477 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
478 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
480 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
483 int debug; /* turn on/off debugging via syslog */
484 int debug_ovfl; /* turn on/off debug printk in overflow handler */
485 int fastctxsw; /* turn on/off fast (unsecure) ctxsw */
486 int expert_mode; /* turn on/off value checking */
491 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
492 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
493 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
496 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
497 unsigned long pfm_smpl_handler_calls;
498 unsigned long pfm_smpl_handler_cycles;
499 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
503 * perfmon internal variables
505 static pfm_stats_t pfm_stats[NR_CPUS];
506 static pfm_session_t pfm_sessions; /* global sessions information */
508 static struct proc_dir_entry *perfmon_dir;
509 static pfm_uuid_t pfm_null_uuid = {0,};
511 static spinlock_t pfm_buffer_fmt_lock;
512 static LIST_HEAD(pfm_buffer_fmt_list);
514 static pmu_config_t *pmu_conf;
516 /* sysctl() controls */
517 static pfm_sysctl_t pfm_sysctl;
520 static ctl_table pfm_ctl_table[]={
521 {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
522 {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
523 {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
524 {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
527 static ctl_table pfm_sysctl_dir[] = {
528 {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
531 static ctl_table pfm_sysctl_root[] = {
532 {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
535 static struct ctl_table_header *pfm_sysctl_header;
537 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
538 static int pfm_flush(struct file *filp);
540 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
541 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
544 pfm_put_task(struct task_struct *task)
546 if (task != current) put_task_struct(task);
550 pfm_set_task_notify(struct task_struct *task)
552 struct thread_info *info;
554 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
555 set_bit(TIF_NOTIFY_RESUME, &info->flags);
559 pfm_clear_task_notify(void)
561 clear_thread_flag(TIF_NOTIFY_RESUME);
565 pfm_reserve_page(unsigned long a)
567 SetPageReserved(vmalloc_to_page((void *)a));
570 pfm_unreserve_page(unsigned long a)
572 ClearPageReserved(vmalloc_to_page((void*)a));
576 pfm_remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
578 return remap_page_range(vma, from, phys_addr, size, prot);
581 static inline unsigned long
582 pfm_protect_ctx_ctxsw(pfm_context_t *x)
584 spin_lock(&(x)->ctx_lock);
588 static inline unsigned long
589 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
591 spin_unlock(&(x)->ctx_lock);
594 static inline unsigned int
595 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
597 return do_munmap(mm, addr, len);
600 static inline unsigned long
601 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
603 return get_unmapped_area(file, addr, len, pgoff, flags);
607 static struct super_block *
608 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
610 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
613 static struct file_system_type pfm_fs_type = {
615 .get_sb = pfmfs_get_sb,
616 .kill_sb = kill_anon_super,
619 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
620 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
621 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
622 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
625 /* forward declaration */
626 static struct file_operations pfm_file_ops;
629 * forward declarations
632 static void pfm_lazy_save_regs (struct task_struct *ta);
635 void dump_pmu_state(const char *);
636 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
638 #include "perfmon_itanium.h"
639 #include "perfmon_mckinley.h"
640 #include "perfmon_generic.h"
642 static pmu_config_t *pmu_confs[]={
645 &pmu_conf_gen, /* must be last */
650 static int pfm_end_notify_user(pfm_context_t *ctx);
653 pfm_clear_psr_pp(void)
655 ia64_rsm(IA64_PSR_PP);
662 ia64_ssm(IA64_PSR_PP);
667 pfm_clear_psr_up(void)
669 ia64_rsm(IA64_PSR_UP);
676 ia64_ssm(IA64_PSR_UP);
680 static inline unsigned long
684 tmp = ia64_getreg(_IA64_REG_PSR);
690 pfm_set_psr_l(unsigned long val)
692 ia64_setreg(_IA64_REG_PSR_L, val);
704 pfm_unfreeze_pmu(void)
711 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
715 for (i=0; i < nibrs; i++) {
716 ia64_set_ibr(i, ibrs[i]);
717 ia64_dv_serialize_instruction();
723 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
727 for (i=0; i < ndbrs; i++) {
728 ia64_set_dbr(i, dbrs[i]);
729 ia64_dv_serialize_data();
735 * PMD[i] must be a counter. no check is made
737 static inline unsigned long
738 pfm_read_soft_counter(pfm_context_t *ctx, int i)
740 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
744 * PMD[i] must be a counter. no check is made
747 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
749 unsigned long ovfl_val = pmu_conf->ovfl_val;
751 ctx->ctx_pmds[i].val = val & ~ovfl_val;
753 * writing to unimplemented part is ignore, so we do not need to
756 ia64_set_pmd(i, val & ovfl_val);
760 pfm_get_new_msg(pfm_context_t *ctx)
764 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
766 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
767 if (next == ctx->ctx_msgq_head) return NULL;
769 idx = ctx->ctx_msgq_tail;
770 ctx->ctx_msgq_tail = next;
772 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
774 return ctx->ctx_msgq+idx;
778 pfm_get_next_msg(pfm_context_t *ctx)
782 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
784 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
789 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
794 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
796 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));
802 pfm_reset_msgq(pfm_context_t *ctx)
804 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
805 DPRINT(("ctx=%p msgq reset\n", ctx));
809 /* Here we want the physical address of the memory.
810 * This is used when initializing the contents of the
811 * area and marking the pages as reserved.
813 static inline unsigned long
814 pfm_kvirt_to_pa(unsigned long adr)
816 __u64 pa = ia64_tpa(adr);
821 pfm_rvmalloc(unsigned long size)
826 size = PAGE_ALIGN(size);
829 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
830 memset(mem, 0, size);
831 addr = (unsigned long)mem;
833 pfm_reserve_page(addr);
842 pfm_rvfree(void *mem, unsigned long size)
847 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
848 addr = (unsigned long) mem;
849 while ((long) size > 0) {
850 pfm_unreserve_page(addr);
859 static pfm_context_t *
860 pfm_context_alloc(void)
865 * allocate context descriptor
866 * must be able to free with interrupts disabled
868 ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
870 memset(ctx, 0, sizeof(pfm_context_t));
871 DPRINT(("alloc ctx @%p\n", ctx));
877 pfm_context_free(pfm_context_t *ctx)
880 DPRINT(("free ctx @%p\n", ctx));
886 pfm_mask_monitoring(struct task_struct *task)
888 pfm_context_t *ctx = PFM_GET_CTX(task);
889 struct thread_struct *th = &task->thread;
890 unsigned long mask, val, ovfl_mask;
893 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
895 ovfl_mask = pmu_conf->ovfl_val;
897 * monitoring can only be masked as a result of a valid
898 * counter overflow. In UP, it means that the PMU still
899 * has an owner. Note that the owner can be different
900 * from the current task. However the PMU state belongs
902 * In SMP, a valid overflow only happens when task is
903 * current. Therefore if we come here, we know that
904 * the PMU state belongs to the current task, therefore
905 * we can access the live registers.
907 * So in both cases, the live register contains the owner's
908 * state. We can ONLY touch the PMU registers and NOT the PSR.
910 * As a consequence to this call, the thread->pmds[] array
911 * contains stale information which must be ignored
912 * when context is reloaded AND monitoring is active (see
915 mask = ctx->ctx_used_pmds[0];
916 for (i = 0; mask; i++, mask>>=1) {
917 /* skip non used pmds */
918 if ((mask & 0x1) == 0) continue;
919 val = ia64_get_pmd(i);
921 if (PMD_IS_COUNTING(i)) {
923 * we rebuild the full 64 bit value of the counter
925 ctx->ctx_pmds[i].val += (val & ovfl_mask);
927 ctx->ctx_pmds[i].val = val;
929 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
931 ctx->ctx_pmds[i].val,
935 * mask monitoring by setting the privilege level to 0
936 * we cannot use psr.pp/psr.up for this, it is controlled by
939 * if task is current, modify actual registers, otherwise modify
940 * thread save state, i.e., what will be restored in pfm_load_regs()
942 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
943 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
944 if ((mask & 0x1) == 0UL) continue;
945 ia64_set_pmc(i, th->pmcs[i] & ~0xfUL);
946 th->pmcs[i] &= ~0xfUL;
947 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i]));
950 * make all of this visible
956 * must always be done with task == current
958 * context must be in MASKED state when calling
961 pfm_restore_monitoring(struct task_struct *task)
963 pfm_context_t *ctx = PFM_GET_CTX(task);
964 struct thread_struct *th = &task->thread;
965 unsigned long mask, ovfl_mask;
966 unsigned long psr, val;
969 is_system = ctx->ctx_fl_system;
970 ovfl_mask = pmu_conf->ovfl_val;
972 if (task != current) {
973 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
976 if (ctx->ctx_state != PFM_CTX_MASKED) {
977 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
978 task->pid, current->pid, ctx->ctx_state);
983 * monitoring is masked via the PMC.
984 * As we restore their value, we do not want each counter to
985 * restart right away. We stop monitoring using the PSR,
986 * restore the PMC (and PMD) and then re-establish the psr
987 * as it was. Note that there can be no pending overflow at
988 * this point, because monitoring was MASKED.
990 * system-wide session are pinned and self-monitoring
992 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
994 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1000 * first, we restore the PMD
1002 mask = ctx->ctx_used_pmds[0];
1003 for (i = 0; mask; i++, mask>>=1) {
1004 /* skip non used pmds */
1005 if ((mask & 0x1) == 0) continue;
1007 if (PMD_IS_COUNTING(i)) {
1009 * we split the 64bit value according to
1012 val = ctx->ctx_pmds[i].val & ovfl_mask;
1013 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1015 val = ctx->ctx_pmds[i].val;
1017 ia64_set_pmd(i, val);
1019 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1021 ctx->ctx_pmds[i].val,
1027 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1028 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1029 if ((mask & 0x1) == 0UL) continue;
1030 th->pmcs[i] = ctx->ctx_pmcs[i];
1031 ia64_set_pmc(i, th->pmcs[i]);
1032 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
1037 * must restore DBR/IBR because could be modified while masked
1038 * XXX: need to optimize
1040 if (ctx->ctx_fl_using_dbreg) {
1041 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1042 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1048 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1050 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1057 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1063 for (i=0; mask; i++, mask>>=1) {
1064 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1069 * reload from thread state (used for ctxw only)
1072 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1075 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1077 for (i=0; mask; i++, mask>>=1) {
1078 if ((mask & 0x1) == 0) continue;
1079 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1080 ia64_set_pmd(i, val);
1086 * propagate PMD from context to thread-state
1089 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1091 struct thread_struct *thread = &task->thread;
1092 unsigned long ovfl_val = pmu_conf->ovfl_val;
1093 unsigned long mask = ctx->ctx_all_pmds[0];
1097 DPRINT(("mask=0x%lx\n", mask));
1099 for (i=0; mask; i++, mask>>=1) {
1101 val = ctx->ctx_pmds[i].val;
1104 * We break up the 64 bit value into 2 pieces
1105 * the lower bits go to the machine state in the
1106 * thread (will be reloaded on ctxsw in).
1107 * The upper part stays in the soft-counter.
1109 if (PMD_IS_COUNTING(i)) {
1110 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1113 thread->pmds[i] = val;
1115 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1118 ctx->ctx_pmds[i].val));
1123 * propagate PMC from context to thread-state
1126 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1128 struct thread_struct *thread = &task->thread;
1129 unsigned long mask = ctx->ctx_all_pmcs[0];
1132 DPRINT(("mask=0x%lx\n", mask));
1134 for (i=0; mask; i++, mask>>=1) {
1135 /* masking 0 with ovfl_val yields 0 */
1136 thread->pmcs[i] = ctx->ctx_pmcs[i];
1137 DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
1144 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1148 for (i=0; mask; i++, mask>>=1) {
1149 if ((mask & 0x1) == 0) continue;
1150 ia64_set_pmc(i, pmcs[i]);
1156 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1158 return memcmp(a, b, sizeof(pfm_uuid_t));
1162 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1165 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1170 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1173 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1179 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1183 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1188 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1192 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1197 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1200 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1205 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)
1208 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1212 static pfm_buffer_fmt_t *
1213 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1215 struct list_head * pos;
1216 pfm_buffer_fmt_t * entry;
1218 list_for_each(pos, &pfm_buffer_fmt_list) {
1219 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1220 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1227 * find a buffer format based on its uuid
1229 static pfm_buffer_fmt_t *
1230 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1232 pfm_buffer_fmt_t * fmt;
1233 spin_lock(&pfm_buffer_fmt_lock);
1234 fmt = __pfm_find_buffer_fmt(uuid);
1235 spin_unlock(&pfm_buffer_fmt_lock);
1240 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1244 /* some sanity checks */
1245 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1247 /* we need at least a handler */
1248 if (fmt->fmt_handler == NULL) return -EINVAL;
1251 * XXX: need check validity of fmt_arg_size
1254 spin_lock(&pfm_buffer_fmt_lock);
1256 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1257 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1261 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1262 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1265 spin_unlock(&pfm_buffer_fmt_lock);
1268 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1271 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1273 pfm_buffer_fmt_t *fmt;
1276 spin_lock(&pfm_buffer_fmt_lock);
1278 fmt = __pfm_find_buffer_fmt(uuid);
1280 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1284 list_del_init(&fmt->fmt_list);
1285 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1288 spin_unlock(&pfm_buffer_fmt_lock);
1292 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1295 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1297 unsigned long flags;
1299 * validy checks on cpu_mask have been done upstream
1303 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1304 pfm_sessions.pfs_sys_sessions,
1305 pfm_sessions.pfs_task_sessions,
1306 pfm_sessions.pfs_sys_use_dbregs,
1312 * cannot mix system wide and per-task sessions
1314 if (pfm_sessions.pfs_task_sessions > 0UL) {
1315 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1316 pfm_sessions.pfs_task_sessions));
1320 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1322 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1324 pfm_sessions.pfs_sys_session[cpu] = task;
1326 pfm_sessions.pfs_sys_sessions++ ;
1329 if (pfm_sessions.pfs_sys_sessions) goto abort;
1330 pfm_sessions.pfs_task_sessions++;
1333 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1334 pfm_sessions.pfs_sys_sessions,
1335 pfm_sessions.pfs_task_sessions,
1336 pfm_sessions.pfs_sys_use_dbregs,
1345 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1346 pfm_sessions.pfs_sys_session[cpu]->pid,
1347 smp_processor_id()));
1356 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1358 unsigned long flags;
1360 * validy checks on cpu_mask have been done upstream
1364 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1365 pfm_sessions.pfs_sys_sessions,
1366 pfm_sessions.pfs_task_sessions,
1367 pfm_sessions.pfs_sys_use_dbregs,
1373 pfm_sessions.pfs_sys_session[cpu] = NULL;
1375 * would not work with perfmon+more than one bit in cpu_mask
1377 if (ctx && ctx->ctx_fl_using_dbreg) {
1378 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1379 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1381 pfm_sessions.pfs_sys_use_dbregs--;
1384 pfm_sessions.pfs_sys_sessions--;
1386 pfm_sessions.pfs_task_sessions--;
1388 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1389 pfm_sessions.pfs_sys_sessions,
1390 pfm_sessions.pfs_task_sessions,
1391 pfm_sessions.pfs_sys_use_dbregs,
1401 * removes virtual mapping of the sampling buffer.
1402 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1403 * a PROTECT_CTX() section.
1406 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1411 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1412 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1416 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1419 * does the actual unmapping
1421 down_write(&task->mm->mmap_sem);
1423 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1425 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1427 up_write(&task->mm->mmap_sem);
1429 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1432 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1438 * free actual physical storage used by sampling buffer
1442 pfm_free_smpl_buffer(pfm_context_t *ctx)
1444 pfm_buffer_fmt_t *fmt;
1446 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1449 * we won't use the buffer format anymore
1451 fmt = ctx->ctx_buf_fmt;
1453 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1456 ctx->ctx_smpl_vaddr));
1458 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1463 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1465 ctx->ctx_smpl_hdr = NULL;
1466 ctx->ctx_smpl_size = 0UL;
1471 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1477 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1479 if (fmt == NULL) return;
1481 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1486 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1487 * no real gain from having the whole whorehouse mounted. So we don't need
1488 * any operations on the root directory. However, we need a non-trivial
1489 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1491 static struct vfsmount *pfmfs_mnt;
1496 int err = register_filesystem(&pfm_fs_type);
1498 pfmfs_mnt = kern_mount(&pfm_fs_type);
1499 err = PTR_ERR(pfmfs_mnt);
1500 if (IS_ERR(pfmfs_mnt))
1501 unregister_filesystem(&pfm_fs_type);
1511 unregister_filesystem(&pfm_fs_type);
1516 pfm_read(struct file *filp, char *buf, size_t size, loff_t *ppos)
1521 unsigned long flags;
1522 DECLARE_WAITQUEUE(wait, current);
1523 if (PFM_IS_FILE(filp) == 0) {
1524 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1528 ctx = (pfm_context_t *)filp->private_data;
1530 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1535 * check even when there is no message
1537 if (size < sizeof(pfm_msg_t)) {
1538 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1542 PROTECT_CTX(ctx, flags);
1545 * put ourselves on the wait queue
1547 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1555 set_current_state(TASK_INTERRUPTIBLE);
1557 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1560 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1562 UNPROTECT_CTX(ctx, flags);
1565 * check non-blocking read
1568 if(filp->f_flags & O_NONBLOCK) break;
1571 * check pending signals
1573 if(signal_pending(current)) {
1578 * no message, so wait
1582 PROTECT_CTX(ctx, flags);
1584 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1585 set_current_state(TASK_RUNNING);
1586 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1588 if (ret < 0) goto abort;
1591 msg = pfm_get_next_msg(ctx);
1593 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1597 DPRINT(("[%d] fd=%d type=%d\n", current->pid, msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1600 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1603 UNPROTECT_CTX(ctx, flags);
1609 pfm_write(struct file *file, const char *ubuf,
1610 size_t size, loff_t *ppos)
1612 DPRINT(("pfm_write called\n"));
1617 pfm_poll(struct file *filp, poll_table * wait)
1620 unsigned long flags;
1621 unsigned int mask = 0;
1623 if (PFM_IS_FILE(filp) == 0) {
1624 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1628 ctx = (pfm_context_t *)filp->private_data;
1630 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1635 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1637 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1639 PROTECT_CTX(ctx, flags);
1641 if (PFM_CTXQ_EMPTY(ctx) == 0)
1642 mask = POLLIN | POLLRDNORM;
1644 UNPROTECT_CTX(ctx, flags);
1646 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1652 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1654 DPRINT(("pfm_ioctl called\n"));
1659 * context is locked when coming here and interrupts are disabled
1662 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1666 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1668 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1672 ctx->ctx_async_queue, ret));
1678 pfm_fasync(int fd, struct file *filp, int on)
1681 unsigned long flags;
1684 if (PFM_IS_FILE(filp) == 0) {
1685 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1689 ctx = (pfm_context_t *)filp->private_data;
1691 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1696 PROTECT_CTX(ctx, flags);
1698 ret = pfm_do_fasync(fd, filp, ctx, on);
1700 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1703 ctx->ctx_async_queue, ret));
1705 UNPROTECT_CTX(ctx, flags);
1712 * this function is exclusively called from pfm_close().
1713 * The context is not protected at that time, nor are interrupts
1714 * on the remote CPU. That's necessary to avoid deadlocks.
1717 pfm_syswide_force_stop(void *info)
1719 pfm_context_t *ctx = (pfm_context_t *)info;
1720 struct pt_regs *regs = ia64_task_regs(current);
1721 struct task_struct *owner;
1722 unsigned long flags;
1725 if (ctx->ctx_cpu != smp_processor_id()) {
1726 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1728 smp_processor_id());
1731 owner = GET_PMU_OWNER();
1732 if (owner != ctx->ctx_task) {
1733 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1735 owner->pid, ctx->ctx_task->pid);
1738 if (GET_PMU_CTX() != ctx) {
1739 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1741 GET_PMU_CTX(), ctx);
1745 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1747 * the context is already protected in pfm_close(), we simply
1748 * need to mask interrupts to avoid a PMU interrupt race on
1751 local_irq_save(flags);
1753 ret = pfm_context_unload(ctx, NULL, 0, regs);
1755 DPRINT(("context_unload returned %d\n", ret));
1759 * unmask interrupts, PMU interrupts are now spurious here
1761 local_irq_restore(flags);
1765 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1769 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1770 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1771 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1773 #endif /* CONFIG_SMP */
1776 * called for each close(). Partially free resources.
1777 * When caller is self-monitoring, the context is unloaded.
1780 pfm_flush(struct file *filp)
1783 struct task_struct *task;
1784 struct pt_regs *regs;
1785 unsigned long flags;
1786 unsigned long smpl_buf_size = 0UL;
1787 void *smpl_buf_vaddr = NULL;
1788 int state, is_system;
1790 if (PFM_IS_FILE(filp) == 0) {
1791 DPRINT(("bad magic for\n"));
1795 ctx = (pfm_context_t *)filp->private_data;
1797 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1802 * remove our file from the async queue, if we use this mode.
1803 * This can be done without the context being protected. We come
1804 * here when the context has become unreacheable by other tasks.
1806 * We may still have active monitoring at this point and we may
1807 * end up in pfm_overflow_handler(). However, fasync_helper()
1808 * operates with interrupts disabled and it cleans up the
1809 * queue. If the PMU handler is called prior to entering
1810 * fasync_helper() then it will send a signal. If it is
1811 * invoked after, it will find an empty queue and no
1812 * signal will be sent. In both case, we are safe
1814 if (filp->f_flags & FASYNC) {
1815 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1816 pfm_do_fasync (-1, filp, ctx, 0);
1819 PROTECT_CTX(ctx, flags);
1821 state = ctx->ctx_state;
1822 is_system = ctx->ctx_fl_system;
1824 task = PFM_CTX_TASK(ctx);
1825 regs = ia64_task_regs(task);
1827 DPRINT(("ctx_state=%d is_current=%d\n",
1829 task == current ? 1 : 0));
1832 * if state == UNLOADED, then task is NULL
1836 * we must stop and unload because we are losing access to the context.
1838 if (task == current) {
1841 * the task IS the owner but it migrated to another CPU: that's bad
1842 * but we must handle this cleanly. Unfortunately, the kernel does
1843 * not provide a mechanism to block migration (while the context is loaded).
1845 * We need to release the resource on the ORIGINAL cpu.
1847 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1849 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1851 * keep context protected but unmask interrupt for IPI
1853 local_irq_restore(flags);
1855 pfm_syswide_cleanup_other_cpu(ctx);
1858 * restore interrupt masking
1860 local_irq_save(flags);
1863 * context is unloaded at this point
1866 #endif /* CONFIG_SMP */
1869 DPRINT(("forcing unload\n"));
1871 * stop and unload, returning with state UNLOADED
1872 * and session unreserved.
1874 pfm_context_unload(ctx, NULL, 0, regs);
1876 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1881 * remove virtual mapping, if any, for the calling task.
1882 * cannot reset ctx field until last user is calling close().
1884 * ctx_smpl_vaddr must never be cleared because it is needed
1885 * by every task with access to the context
1887 * When called from do_exit(), the mm context is gone already, therefore
1888 * mm is NULL, i.e., the VMA is already gone and we do not have to
1891 if (ctx->ctx_smpl_vaddr && current->mm) {
1892 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1893 smpl_buf_size = ctx->ctx_smpl_size;
1896 UNPROTECT_CTX(ctx, flags);
1899 * if there was a mapping, then we systematically remove it
1900 * at this point. Cannot be done inside critical section
1901 * because some VM function reenables interrupts.
1904 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1909 * called either on explicit close() or from exit_files().
1910 * Only the LAST user of the file gets to this point, i.e., it is
1913 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1914 * (fput()),i.e, last task to access the file. Nobody else can access the
1915 * file at this point.
1917 * When called from exit_files(), the VMA has been freed because exit_mm()
1918 * is executed before exit_files().
1920 * When called from exit_files(), the current task is not yet ZOMBIE but we
1921 * flush the PMU state to the context.
1924 pfm_close(struct inode *inode, struct file *filp)
1927 struct task_struct *task;
1928 struct pt_regs *regs;
1929 DECLARE_WAITQUEUE(wait, current);
1930 unsigned long flags;
1931 unsigned long smpl_buf_size = 0UL;
1932 void *smpl_buf_addr = NULL;
1933 int free_possible = 1;
1934 int state, is_system;
1936 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1938 if (PFM_IS_FILE(filp) == 0) {
1939 DPRINT(("bad magic\n"));
1943 ctx = (pfm_context_t *)filp->private_data;
1945 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1949 PROTECT_CTX(ctx, flags);
1951 state = ctx->ctx_state;
1952 is_system = ctx->ctx_fl_system;
1954 task = PFM_CTX_TASK(ctx);
1955 regs = ia64_task_regs(task);
1957 DPRINT(("ctx_state=%d is_current=%d\n",
1959 task == current ? 1 : 0));
1962 * if task == current, then pfm_flush() unloaded the context
1964 if (state == PFM_CTX_UNLOADED) goto doit;
1967 * context is loaded/masked and task != current, we need to
1968 * either force an unload or go zombie
1972 * The task is currently blocked or will block after an overflow.
1973 * we must force it to wakeup to get out of the
1974 * MASKED state and transition to the unloaded state by itself.
1976 * This situation is only possible for per-task mode
1978 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1981 * set a "partial" zombie state to be checked
1982 * upon return from down() in pfm_handle_work().
1984 * We cannot use the ZOMBIE state, because it is checked
1985 * by pfm_load_regs() which is called upon wakeup from down().
1986 * In such case, it would free the context and then we would
1987 * return to pfm_handle_work() which would access the
1988 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1989 * but visible to pfm_handle_work().
1991 * For some window of time, we have a zombie context with
1992 * ctx_state = MASKED and not ZOMBIE
1994 ctx->ctx_fl_going_zombie = 1;
1997 * force task to wake up from MASKED state
1999 up(&ctx->ctx_restart_sem);
2001 DPRINT(("waking up ctx_state=%d\n", state));
2004 * put ourself to sleep waiting for the other
2005 * task to report completion
2007 * the context is protected by mutex, therefore there
2008 * is no risk of being notified of completion before
2009 * begin actually on the waitq.
2011 set_current_state(TASK_INTERRUPTIBLE);
2012 add_wait_queue(&ctx->ctx_zombieq, &wait);
2014 UNPROTECT_CTX(ctx, flags);
2017 * XXX: check for signals :
2018 * - ok of explicit close
2019 * - not ok when coming from exit_files()
2024 PROTECT_CTX(ctx, flags);
2027 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2028 set_current_state(TASK_RUNNING);
2031 * context is unloaded at this point
2033 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2035 else if (task != current) {
2038 * switch context to zombie state
2040 ctx->ctx_state = PFM_CTX_ZOMBIE;
2042 DPRINT(("zombie ctx for [%d]\n", task->pid));
2044 * cannot free the context on the spot. deferred until
2045 * the task notices the ZOMBIE state
2049 pfm_context_unload(ctx, NULL, 0, regs);
2054 /* reload state, may have changed during opening of critical section */
2055 state = ctx->ctx_state;
2058 * the context is still attached to a task (possibly current)
2059 * we cannot destroy it right now
2063 * we must free the sampling buffer right here because
2064 * we cannot rely on it being cleaned up later by the
2065 * monitored task. It is not possible to free vmalloc'ed
2066 * memory in pfm_load_regs(). Instead, we remove the buffer
2067 * now. should there be subsequent PMU overflow originally
2068 * meant for sampling, the will be converted to spurious
2069 * and that's fine because the monitoring tools is gone anyway.
2071 if (ctx->ctx_smpl_hdr) {
2072 smpl_buf_addr = ctx->ctx_smpl_hdr;
2073 smpl_buf_size = ctx->ctx_smpl_size;
2074 /* no more sampling */
2075 ctx->ctx_smpl_hdr = NULL;
2076 ctx->ctx_fl_is_sampling = 0;
2079 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2085 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2088 * UNLOADED that the session has already been unreserved.
2090 if (state == PFM_CTX_ZOMBIE) {
2091 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2095 * disconnect file descriptor from context must be done
2098 filp->private_data = NULL;
2101 * if we free on the spot, the context is now completely unreacheable
2102 * from the callers side. The monitored task side is also cut, so we
2105 * If we have a deferred free, only the caller side is disconnected.
2107 UNPROTECT_CTX(ctx, flags);
2110 * All memory free operations (especially for vmalloc'ed memory)
2111 * MUST be done with interrupts ENABLED.
2113 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2116 * return the memory used by the context
2118 if (free_possible) pfm_context_free(ctx);
2124 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2126 DPRINT(("pfm_no_open called\n"));
2132 static struct file_operations pfm_file_ops = {
2133 .llseek = no_llseek,
2138 .open = pfm_no_open, /* special open code to disallow open via /proc */
2139 .fasync = pfm_fasync,
2140 .release = pfm_close,
2145 pfmfs_delete_dentry(struct dentry *dentry)
2150 static struct dentry_operations pfmfs_dentry_operations = {
2151 .d_delete = pfmfs_delete_dentry,
2156 pfm_alloc_fd(struct file **cfile)
2159 struct file *file = NULL;
2160 struct inode * inode;
2164 fd = get_unused_fd();
2165 if (fd < 0) return -ENFILE;
2169 file = get_empty_filp();
2170 if (!file) goto out;
2173 * allocate a new inode
2175 inode = new_inode(pfmfs_mnt->mnt_sb);
2176 if (!inode) goto out;
2178 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2180 inode->i_sb = pfmfs_mnt->mnt_sb;
2181 inode->i_mode = S_IFCHR|S_IRUGO;
2183 inode->i_uid = current->fsuid;
2184 inode->i_gid = current->fsgid;
2186 sprintf(name, "[%lu]", inode->i_ino);
2188 this.len = strlen(name);
2189 this.hash = inode->i_ino;
2194 * allocate a new dcache entry
2196 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2197 if (!file->f_dentry) goto out;
2199 file->f_dentry->d_op = &pfmfs_dentry_operations;
2201 d_add(file->f_dentry, inode);
2202 file->f_vfsmnt = mntget(pfmfs_mnt);
2203 file->f_mapping = inode->i_mapping;
2205 file->f_op = &pfm_file_ops;
2206 file->f_mode = FMODE_READ;
2207 file->f_flags = O_RDONLY;
2211 * may have to delay until context is attached?
2213 fd_install(fd, file);
2216 * the file structure we will use
2222 if (file) put_filp(file);
2228 pfm_free_fd(int fd, struct file *file)
2230 if (file) put_filp(file);
2235 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2239 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2242 page = pfm_kvirt_to_pa(buf);
2244 if (pfm_remap_page_range(vma, addr, page, PAGE_SIZE, PAGE_READONLY)) return -ENOMEM;
2254 * allocate a sampling buffer and remaps it into the user address space of the task
2257 pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2259 struct mm_struct *mm = task->mm;
2260 struct vm_area_struct *vma = NULL;
2266 * the fixed header + requested size and align to page boundary
2268 size = PAGE_ALIGN(rsize);
2270 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2273 * check requested size to avoid Denial-of-service attacks
2274 * XXX: may have to refine this test
2275 * Check against address space limit.
2277 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2280 if (size > task->rlim[RLIMIT_MEMLOCK].rlim_cur) return -EAGAIN;
2283 * We do the easy to undo allocations first.
2285 * pfm_rvmalloc(), clears the buffer, so there is no leak
2287 smpl_buf = pfm_rvmalloc(size);
2288 if (smpl_buf == NULL) {
2289 DPRINT(("Can't allocate sampling buffer\n"));
2293 DPRINT(("smpl_buf @%p\n", smpl_buf));
2296 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2298 DPRINT(("Cannot allocate vma\n"));
2301 memset(vma, 0, sizeof(*vma));
2304 * partially initialize the vma for the sampling buffer
2306 * The VM_DONTCOPY flag is very important as it ensures that the mapping
2307 * will never be inherited for any child process (via fork()) which is always
2311 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2312 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2315 * Now we have everything we need and we can initialize
2316 * and connect all the data structures
2319 ctx->ctx_smpl_hdr = smpl_buf;
2320 ctx->ctx_smpl_size = size; /* aligned size */
2323 * Let's do the difficult operations next.
2325 * now we atomically find some area in the address space and
2326 * remap the buffer in it.
2328 down_write(&task->mm->mmap_sem);
2330 /* find some free area in address space, must have mmap sem held */
2331 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2332 if (vma->vm_start == 0UL) {
2333 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2334 up_write(&task->mm->mmap_sem);
2337 vma->vm_end = vma->vm_start + size;
2339 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2341 /* can only be applied to current task, need to have the mm semaphore held when called */
2342 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2343 DPRINT(("Can't remap buffer\n"));
2344 up_write(&task->mm->mmap_sem);
2349 * now insert the vma in the vm list for the process, must be
2350 * done with mmap lock held
2352 insert_vm_struct(mm, vma);
2354 mm->total_vm += size >> PAGE_SHIFT;
2356 up_write(&task->mm->mmap_sem);
2359 * keep track of user level virtual address
2361 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2362 *(unsigned long *)user_vaddr = vma->vm_start;
2367 kmem_cache_free(vm_area_cachep, vma);
2369 pfm_rvfree(smpl_buf, size);
2375 * XXX: do something better here
2378 pfm_bad_permissions(struct task_struct *task)
2380 /* inspired by ptrace_attach() */
2381 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2390 return ((current->uid != task->euid)
2391 || (current->uid != task->suid)
2392 || (current->uid != task->uid)
2393 || (current->gid != task->egid)
2394 || (current->gid != task->sgid)
2395 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2399 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2405 ctx_flags = pfx->ctx_flags;
2407 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2410 * cannot block in this mode
2412 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2413 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2418 /* probably more to add here */
2424 pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2425 unsigned int cpu, pfarg_context_t *arg)
2427 pfm_buffer_fmt_t *fmt = NULL;
2428 unsigned long size = 0UL;
2430 void *fmt_arg = NULL;
2432 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2434 /* invoke and lock buffer format, if found */
2435 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2437 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2442 * buffer argument MUST be contiguous to pfarg_context_t
2444 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2446 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2448 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2450 if (ret) goto error;
2452 /* link buffer format and context */
2453 ctx->ctx_buf_fmt = fmt;
2456 * check if buffer format wants to use perfmon buffer allocation/mapping service
2458 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2459 if (ret) goto error;
2463 * buffer is always remapped into the caller's address space
2465 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2466 if (ret) goto error;
2468 /* keep track of user address of buffer */
2469 arg->ctx_smpl_vaddr = uaddr;
2471 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2478 pfm_reset_pmu_state(pfm_context_t *ctx)
2483 * install reset values for PMC.
2485 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2486 if (PMC_IS_IMPL(i) == 0) continue;
2487 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2488 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2491 * PMD registers are set to 0UL when the context in memset()
2495 * On context switched restore, we must restore ALL pmc and ALL pmd even
2496 * when they are not actively used by the task. In UP, the incoming process
2497 * may otherwise pick up left over PMC, PMD state from the previous process.
2498 * As opposed to PMD, stale PMC can cause harm to the incoming
2499 * process because they may change what is being measured.
2500 * Therefore, we must systematically reinstall the entire
2501 * PMC state. In SMP, the same thing is possible on the
2502 * same CPU but also on between 2 CPUs.
2504 * The problem with PMD is information leaking especially
2505 * to user level when psr.sp=0
2507 * There is unfortunately no easy way to avoid this problem
2508 * on either UP or SMP. This definitively slows down the
2509 * pfm_load_regs() function.
2513 * bitmask of all PMCs accessible to this context
2515 * PMC0 is treated differently.
2517 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2520 * bitmask of all PMDs that are accesible to this context
2522 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2524 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2527 * useful in case of re-enable after disable
2529 ctx->ctx_used_ibrs[0] = 0UL;
2530 ctx->ctx_used_dbrs[0] = 0UL;
2534 pfm_ctx_getsize(void *arg, size_t *sz)
2536 pfarg_context_t *req = (pfarg_context_t *)arg;
2537 pfm_buffer_fmt_t *fmt;
2541 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2543 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2545 DPRINT(("cannot find buffer format\n"));
2548 /* get just enough to copy in user parameters */
2549 *sz = fmt->fmt_arg_size;
2550 DPRINT(("arg_size=%lu\n", *sz));
2558 * cannot attach if :
2560 * - task not owned by caller
2561 * - task incompatible with context mode
2564 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2567 * no kernel task or task not owner by caller
2569 if (task->mm == NULL) {
2570 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2573 if (pfm_bad_permissions(task)) {
2574 DPRINT(("no permission to attach to [%d]\n", task->pid));
2578 * cannot block in self-monitoring mode
2580 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2581 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2585 if (task->state == TASK_ZOMBIE) {
2586 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2591 * always ok for self
2593 if (task == current) return 0;
2595 if (task->state != TASK_STOPPED) {
2596 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2600 * make sure the task is off any CPU
2602 wait_task_inactive(task);
2604 /* more to come... */
2610 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2612 struct task_struct *p = current;
2615 /* XXX: need to add more checks here */
2616 if (pid < 2) return -EPERM;
2618 if (pid != current->pid) {
2620 read_lock(&tasklist_lock);
2622 p = find_task_by_pid(pid);
2624 /* make sure task cannot go away while we operate on it */
2625 if (p) get_task_struct(p);
2627 read_unlock(&tasklist_lock);
2629 if (p == NULL) return -ESRCH;
2632 ret = pfm_task_incompatible(ctx, p);
2635 } else if (p != current) {
2644 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2646 pfarg_context_t *req = (pfarg_context_t *)arg;
2651 /* let's check the arguments first */
2652 ret = pfarg_is_sane(current, req);
2653 if (ret < 0) return ret;
2655 ctx_flags = req->ctx_flags;
2659 ctx = pfm_context_alloc();
2660 if (!ctx) goto error;
2662 req->ctx_fd = ctx->ctx_fd = pfm_alloc_fd(&filp);
2663 if (req->ctx_fd < 0) goto error_file;
2666 * attach context to file
2668 filp->private_data = ctx;
2671 * does the user want to sample?
2673 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2674 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2675 if (ret) goto buffer_error;
2679 * init context protection lock
2681 spin_lock_init(&ctx->ctx_lock);
2684 * context is unloaded
2686 ctx->ctx_state = PFM_CTX_UNLOADED;
2689 * initialization of context's flags
2691 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2692 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2693 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2694 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2696 * will move to set properties
2697 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2701 * init restart semaphore to locked
2703 sema_init(&ctx->ctx_restart_sem, 0);
2706 * activation is used in SMP only
2708 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2709 SET_LAST_CPU(ctx, -1);
2712 * initialize notification message queue
2714 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2715 init_waitqueue_head(&ctx->ctx_msgq_wait);
2716 init_waitqueue_head(&ctx->ctx_zombieq);
2718 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2723 ctx->ctx_fl_excl_idle,
2728 * initialize soft PMU state
2730 pfm_reset_pmu_state(ctx);
2735 pfm_free_fd(ctx->ctx_fd, filp);
2737 if (ctx->ctx_buf_fmt) {
2738 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2741 pfm_context_free(ctx);
2747 static inline unsigned long
2748 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2750 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2751 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2752 extern unsigned long carta_random32 (unsigned long seed);
2754 if (reg->flags & PFM_REGFL_RANDOM) {
2755 new_seed = carta_random32(old_seed);
2756 val -= (old_seed & mask); /* counter values are negative numbers! */
2757 if ((mask >> 32) != 0)
2758 /* construct a full 64-bit random value: */
2759 new_seed |= carta_random32(old_seed >> 32) << 32;
2760 reg->seed = new_seed;
2767 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2769 unsigned long mask = ovfl_regs[0];
2770 unsigned long reset_others = 0UL;
2775 * now restore reset value on sampling overflowed counters
2777 mask >>= PMU_FIRST_COUNTER;
2778 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2780 if ((mask & 0x1UL) == 0UL) continue;
2782 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2783 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2785 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2789 * Now take care of resetting the other registers
2791 for(i = 0; reset_others; i++, reset_others >>= 1) {
2793 if ((reset_others & 0x1) == 0) continue;
2795 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2797 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2798 is_long_reset ? "long" : "short", i, val));
2803 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2805 unsigned long mask = ovfl_regs[0];
2806 unsigned long reset_others = 0UL;
2810 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2812 if (ctx->ctx_state == PFM_CTX_MASKED) {
2813 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2818 * now restore reset value on sampling overflowed counters
2820 mask >>= PMU_FIRST_COUNTER;
2821 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2823 if ((mask & 0x1UL) == 0UL) continue;
2825 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2826 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2828 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2830 pfm_write_soft_counter(ctx, i, val);
2834 * Now take care of resetting the other registers
2836 for(i = 0; reset_others; i++, reset_others >>= 1) {
2838 if ((reset_others & 0x1) == 0) continue;
2840 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2842 if (PMD_IS_COUNTING(i)) {
2843 pfm_write_soft_counter(ctx, i, val);
2845 ia64_set_pmd(i, val);
2847 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2848 is_long_reset ? "long" : "short", i, val));
2854 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2856 struct thread_struct *thread = NULL;
2857 struct task_struct *task;
2858 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2859 unsigned long value, pmc_pm;
2860 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2861 unsigned int cnum, reg_flags, flags, pmc_type;
2862 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2863 int is_monitor, is_counting, state;
2865 pfm_reg_check_t wr_func;
2866 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2868 state = ctx->ctx_state;
2869 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2870 is_system = ctx->ctx_fl_system;
2871 task = ctx->ctx_task;
2872 impl_pmds = pmu_conf->impl_pmds[0];
2874 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2877 thread = &task->thread;
2879 * In system wide and when the context is loaded, access can only happen
2880 * when the caller is running on the CPU being monitored by the session.
2881 * It does not have to be the owner (ctx_task) of the context per se.
2883 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2884 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2887 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2889 expert_mode = pfm_sysctl.expert_mode;
2891 for (i = 0; i < count; i++, req++) {
2893 cnum = req->reg_num;
2894 reg_flags = req->reg_flags;
2895 value = req->reg_value;
2896 smpl_pmds = req->reg_smpl_pmds[0];
2897 reset_pmds = req->reg_reset_pmds[0];
2901 if (cnum >= PMU_MAX_PMCS) {
2902 DPRINT(("pmc%u is invalid\n", cnum));
2906 pmc_type = pmu_conf->pmc_desc[cnum].type;
2907 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2908 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2909 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2912 * we reject all non implemented PMC as well
2913 * as attempts to modify PMC[0-3] which are used
2914 * as status registers by the PMU
2916 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2917 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2920 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2922 * If the PMC is a monitor, then if the value is not the default:
2923 * - system-wide session: PMCx.pm=1 (privileged monitor)
2924 * - per-task : PMCx.pm=0 (user monitor)
2926 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2927 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2936 * enforce generation of overflow interrupt. Necessary on all
2939 value |= 1 << PMU_PMC_OI;
2941 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2942 flags |= PFM_REGFL_OVFL_NOTIFY;
2945 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2947 /* verify validity of smpl_pmds */
2948 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2949 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2953 /* verify validity of reset_pmds */
2954 if ((reset_pmds & impl_pmds) != reset_pmds) {
2955 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2959 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2960 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2963 /* eventid on non-counting monitors are ignored */
2967 * execute write checker, if any
2969 if (likely(expert_mode == 0 && wr_func)) {
2970 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2971 if (ret) goto error;
2976 * no error on this register
2978 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2981 * Now we commit the changes to the software state
2985 * update overflow information
2989 * full flag update each time a register is programmed
2991 ctx->ctx_pmds[cnum].flags = flags;
2993 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2994 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2995 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2998 * Mark all PMDS to be accessed as used.
3000 * We do not keep track of PMC because we have to
3001 * systematically restore ALL of them.
3003 * We do not update the used_monitors mask, because
3004 * if we have not programmed them, then will be in
3005 * a quiescent state, therefore we will not need to
3006 * mask/restore then when context is MASKED.
3008 CTX_USED_PMD(ctx, reset_pmds);
3009 CTX_USED_PMD(ctx, smpl_pmds);
3011 * make sure we do not try to reset on
3012 * restart because we have established new values
3014 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3017 * Needed in case the user does not initialize the equivalent
3018 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3019 * possible leak here.
3021 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3024 * keep track of the monitor PMC that we are using.
3025 * we save the value of the pmc in ctx_pmcs[] and if
3026 * the monitoring is not stopped for the context we also
3027 * place it in the saved state area so that it will be
3028 * picked up later by the context switch code.
3030 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3032 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3033 * monitoring needs to be stopped.
3035 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3038 * update context state
3040 ctx->ctx_pmcs[cnum] = value;
3044 * write thread state
3046 if (is_system == 0) thread->pmcs[cnum] = value;
3049 * write hardware register if we can
3051 if (can_access_pmu) {
3052 ia64_set_pmc(cnum, value);
3057 * per-task SMP only here
3059 * we are guaranteed that the task is not running on the other CPU,
3060 * we indicate that this PMD will need to be reloaded if the task
3061 * is rescheduled on the CPU it ran last on.
3063 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3068 DPRINT(("pmc[%u]=0x%lx loaded=%d access_pmu=%d all_pmcs=0x%lx used_pmds=0x%lx eventid=%ld smpl_pmds=0x%lx reset_pmds=0x%lx reloads_pmcs=0x%lx used_monitors=0x%lx ovfl_regs=0x%lx\n",
3073 ctx->ctx_all_pmcs[0],
3074 ctx->ctx_used_pmds[0],
3075 ctx->ctx_pmds[cnum].eventid,
3078 ctx->ctx_reload_pmcs[0],
3079 ctx->ctx_used_monitors[0],
3080 ctx->ctx_ovfl_regs[0]));
3084 * make sure the changes are visible
3086 if (can_access_pmu) ia64_srlz_d();
3090 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3095 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3097 struct thread_struct *thread = NULL;
3098 struct task_struct *task;
3099 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3100 unsigned long value, hw_value, ovfl_mask;
3102 int i, can_access_pmu = 0, state;
3103 int is_counting, is_loaded, is_system, expert_mode;
3105 pfm_reg_check_t wr_func;
3108 state = ctx->ctx_state;
3109 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3110 is_system = ctx->ctx_fl_system;
3111 ovfl_mask = pmu_conf->ovfl_val;
3112 task = ctx->ctx_task;
3114 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3117 * on both UP and SMP, we can only write to the PMC when the task is
3118 * the owner of the local PMU.
3120 if (likely(is_loaded)) {
3121 thread = &task->thread;
3123 * In system wide and when the context is loaded, access can only happen
3124 * when the caller is running on the CPU being monitored by the session.
3125 * It does not have to be the owner (ctx_task) of the context per se.
3127 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3128 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3131 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3133 expert_mode = pfm_sysctl.expert_mode;
3135 for (i = 0; i < count; i++, req++) {
3137 cnum = req->reg_num;
3138 value = req->reg_value;
3140 if (!PMD_IS_IMPL(cnum)) {
3141 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3144 is_counting = PMD_IS_COUNTING(cnum);
3145 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3148 * execute write checker, if any
3150 if (unlikely(expert_mode == 0 && wr_func)) {
3151 unsigned long v = value;
3153 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3154 if (ret) goto abort_mission;
3161 * no error on this register
3163 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3166 * now commit changes to software state
3171 * update virtualized (64bits) counter
3175 * write context state
3177 ctx->ctx_pmds[cnum].lval = value;
3180 * when context is load we use the split value
3183 hw_value = value & ovfl_mask;
3184 value = value & ~ovfl_mask;
3188 * update reset values (not just for counters)
3190 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3191 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3194 * update randomization parameters (not just for counters)
3196 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3197 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3200 * update context value
3202 ctx->ctx_pmds[cnum].val = value;
3205 * Keep track of what we use
3207 * We do not keep track of PMC because we have to
3208 * systematically restore ALL of them.
3210 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3213 * mark this PMD register used as well
3215 CTX_USED_PMD(ctx, RDEP(cnum));
3218 * make sure we do not try to reset on
3219 * restart because we have established new values
3221 if (is_counting && state == PFM_CTX_MASKED) {
3222 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3227 * write thread state
3229 if (is_system == 0) thread->pmds[cnum] = hw_value;
3232 * write hardware register if we can
3234 if (can_access_pmu) {
3235 ia64_set_pmd(cnum, hw_value);
3239 * we are guaranteed that the task is not running on the other CPU,
3240 * we indicate that this PMD will need to be reloaded if the task
3241 * is rescheduled on the CPU it ran last on.
3243 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3248 DPRINT(("pmd[%u]=0x%lx loaded=%d access_pmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3249 "long_reset=0x%lx notify=%c used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3255 ctx->ctx_pmds[cnum].val,
3256 ctx->ctx_pmds[cnum].short_reset,
3257 ctx->ctx_pmds[cnum].long_reset,
3258 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3259 ctx->ctx_used_pmds[0],
3260 ctx->ctx_pmds[cnum].reset_pmds[0],
3261 ctx->ctx_reload_pmds[0],
3262 ctx->ctx_all_pmds[0],
3263 ctx->ctx_ovfl_regs[0]));
3267 * make changes visible
3269 if (can_access_pmu) ia64_srlz_d();
3275 * for now, we have only one possibility for error
3277 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3282 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3283 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3284 * interrupt is delivered during the call, it will be kept pending until we leave, making
3285 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3286 * guaranteed to return consistent data to the user, it may simply be old. It is not
3287 * trivial to treat the overflow while inside the call because you may end up in
3288 * some module sampling buffer code causing deadlocks.
3291 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3293 struct thread_struct *thread = NULL;
3294 struct task_struct *task;
3295 unsigned long val = 0UL, lval, ovfl_mask, sval;
3296 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3297 unsigned int cnum, reg_flags = 0;
3298 int i, can_access_pmu = 0, state;
3299 int is_loaded, is_system, is_counting, expert_mode;
3301 pfm_reg_check_t rd_func;
3304 * access is possible when loaded only for
3305 * self-monitoring tasks or in UP mode
3308 state = ctx->ctx_state;
3309 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3310 is_system = ctx->ctx_fl_system;
3311 ovfl_mask = pmu_conf->ovfl_val;
3312 task = ctx->ctx_task;
3314 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3316 if (likely(is_loaded)) {
3317 thread = &task->thread;
3319 * In system wide and when the context is loaded, access can only happen
3320 * when the caller is running on the CPU being monitored by the session.
3321 * It does not have to be the owner (ctx_task) of the context per se.
3323 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3324 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3328 * this can be true when not self-monitoring only in UP
3330 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3332 if (can_access_pmu) ia64_srlz_d();
3334 expert_mode = pfm_sysctl.expert_mode;
3336 DPRINT(("loaded=%d access_pmu=%d ctx_state=%d\n",
3342 * on both UP and SMP, we can only read the PMD from the hardware register when
3343 * the task is the owner of the local PMU.
3346 for (i = 0; i < count; i++, req++) {
3348 cnum = req->reg_num;
3349 reg_flags = req->reg_flags;
3351 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3353 * we can only read the register that we use. That includes
3354 * the one we explicitely initialize AND the one we want included
3355 * in the sampling buffer (smpl_regs).
3357 * Having this restriction allows optimization in the ctxsw routine
3358 * without compromising security (leaks)
3360 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3362 sval = ctx->ctx_pmds[cnum].val;
3363 lval = ctx->ctx_pmds[cnum].lval;
3364 is_counting = PMD_IS_COUNTING(cnum);
3367 * If the task is not the current one, then we check if the
3368 * PMU state is still in the local live register due to lazy ctxsw.
3369 * If true, then we read directly from the registers.
3371 if (can_access_pmu){
3372 val = ia64_get_pmd(cnum);
3375 * context has been saved
3376 * if context is zombie, then task does not exist anymore.
3377 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3379 val = is_loaded ? thread->pmds[cnum] : 0UL;
3381 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3385 * XXX: need to check for overflow when loaded
3392 * execute read checker, if any
3394 if (unlikely(expert_mode == 0 && rd_func)) {
3395 unsigned long v = val;
3396 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3397 if (ret) goto error;
3402 PFM_REG_RETFLAG_SET(reg_flags, 0);
3404 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3407 * update register return value, abort all if problem during copy.
3408 * we only modify the reg_flags field. no check mode is fine because
3409 * access has been verified upfront in sys_perfmonctl().
3411 req->reg_value = val;
3412 req->reg_flags = reg_flags;
3413 req->reg_last_reset_val = lval;
3419 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3424 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3428 if (req == NULL) return -EINVAL;
3430 ctx = GET_PMU_CTX();
3432 if (ctx == NULL) return -EINVAL;
3435 * for now limit to current task, which is enough when calling
3436 * from overflow handler
3438 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3440 return pfm_write_pmcs(ctx, req, nreq, regs);
3442 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3445 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3449 if (req == NULL) return -EINVAL;
3451 ctx = GET_PMU_CTX();
3453 if (ctx == NULL) return -EINVAL;
3456 * for now limit to current task, which is enough when calling
3457 * from overflow handler
3459 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3461 return pfm_read_pmds(ctx, req, nreq, regs);
3463 EXPORT_SYMBOL(pfm_mod_read_pmds);
3466 * Only call this function when a process it trying to
3467 * write the debug registers (reading is always allowed)
3470 pfm_use_debug_registers(struct task_struct *task)
3472 pfm_context_t *ctx = task->thread.pfm_context;
3473 unsigned long flags;
3476 if (pmu_conf->use_rr_dbregs == 0) return 0;
3478 DPRINT(("called for [%d]\n", task->pid));
3483 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3486 * Even on SMP, we do not need to use an atomic here because
3487 * the only way in is via ptrace() and this is possible only when the
3488 * process is stopped. Even in the case where the ctxsw out is not totally
3489 * completed by the time we come here, there is no way the 'stopped' process
3490 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3491 * So this is always safe.
3493 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3498 * We cannot allow setting breakpoints when system wide monitoring
3499 * sessions are using the debug registers.
3501 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3504 pfm_sessions.pfs_ptrace_use_dbregs++;
3506 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3507 pfm_sessions.pfs_ptrace_use_dbregs,
3508 pfm_sessions.pfs_sys_use_dbregs,
3517 * This function is called for every task that exits with the
3518 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3519 * able to use the debug registers for debugging purposes via
3520 * ptrace(). Therefore we know it was not using them for
3521 * perfmormance monitoring, so we only decrement the number
3522 * of "ptraced" debug register users to keep the count up to date
3525 pfm_release_debug_registers(struct task_struct *task)
3527 unsigned long flags;
3530 if (pmu_conf->use_rr_dbregs == 0) return 0;
3533 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3534 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3537 pfm_sessions.pfs_ptrace_use_dbregs--;
3546 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3548 struct task_struct *task;
3549 pfm_buffer_fmt_t *fmt;
3550 pfm_ovfl_ctrl_t rst_ctrl;
3551 int state, is_system;
3554 state = ctx->ctx_state;
3555 fmt = ctx->ctx_buf_fmt;
3556 is_system = ctx->ctx_fl_system;
3557 task = PFM_CTX_TASK(ctx);
3560 case PFM_CTX_MASKED:
3562 case PFM_CTX_LOADED:
3563 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3565 case PFM_CTX_UNLOADED:
3566 case PFM_CTX_ZOMBIE:
3567 DPRINT(("invalid state=%d\n", state));
3570 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3575 * In system wide and when the context is loaded, access can only happen
3576 * when the caller is running on the CPU being monitored by the session.
3577 * It does not have to be the owner (ctx_task) of the context per se.
3579 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3580 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3585 if (unlikely(task == NULL)) {
3586 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3590 if (task == current || is_system) {
3592 fmt = ctx->ctx_buf_fmt;
3594 DPRINT(("restarting self %d ovfl=0x%lx\n",
3596 ctx->ctx_ovfl_regs[0]));
3598 if (CTX_HAS_SMPL(ctx)) {
3600 prefetch(ctx->ctx_smpl_hdr);
3602 rst_ctrl.bits.mask_monitoring = 0;
3603 rst_ctrl.bits.reset_ovfl_pmds = 0;
3605 if (state == PFM_CTX_LOADED)
3606 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3608 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3610 rst_ctrl.bits.mask_monitoring = 0;
3611 rst_ctrl.bits.reset_ovfl_pmds = 1;
3615 if (rst_ctrl.bits.reset_ovfl_pmds)
3616 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3618 if (rst_ctrl.bits.mask_monitoring == 0) {
3619 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3621 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3623 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3625 // cannot use pfm_stop_monitoring(task, regs);
3629 * clear overflowed PMD mask to remove any stale information
3631 ctx->ctx_ovfl_regs[0] = 0UL;
3634 * back to LOADED state
3636 ctx->ctx_state = PFM_CTX_LOADED;
3639 * XXX: not really useful for self monitoring
3641 ctx->ctx_fl_can_restart = 0;
3647 * restart another task
3651 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3652 * one is seen by the task.
3654 if (state == PFM_CTX_MASKED) {
3655 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3657 * will prevent subsequent restart before this one is
3658 * seen by other task
3660 ctx->ctx_fl_can_restart = 0;
3664 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3665 * the task is blocked or on its way to block. That's the normal
3666 * restart path. If the monitoring is not masked, then the task
3667 * can be actively monitoring and we cannot directly intervene.
3668 * Therefore we use the trap mechanism to catch the task and
3669 * force it to reset the buffer/reset PMDs.
3671 * if non-blocking, then we ensure that the task will go into
3672 * pfm_handle_work() before returning to user mode.
3674 * We cannot explicitely reset another task, it MUST always
3675 * be done by the task itself. This works for system wide because
3676 * the tool that is controlling the session is logically doing
3677 * "self-monitoring".
3679 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3680 DPRINT(("unblocking [%d] \n", task->pid));
3681 up(&ctx->ctx_restart_sem);
3683 DPRINT(("[%d] armed exit trap\n", task->pid));
3685 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3687 PFM_SET_WORK_PENDING(task, 1);
3689 pfm_set_task_notify(task);
3692 * XXX: send reschedule if task runs on another CPU
3699 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3701 unsigned int m = *(unsigned int *)arg;
3703 pfm_sysctl.debug = m == 0 ? 0 : 1;
3705 pfm_debug_var = pfm_sysctl.debug;
3707 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3710 memset(pfm_stats, 0, sizeof(pfm_stats));
3711 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3717 * arg can be NULL and count can be zero for this function
3720 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3722 struct thread_struct *thread = NULL;
3723 struct task_struct *task;
3724 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3725 unsigned long flags;
3730 int i, can_access_pmu = 0;
3731 int is_system, is_loaded;
3733 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3735 state = ctx->ctx_state;
3736 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3737 is_system = ctx->ctx_fl_system;
3738 task = ctx->ctx_task;
3740 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3743 * on both UP and SMP, we can only write to the PMC when the task is
3744 * the owner of the local PMU.
3747 thread = &task->thread;
3749 * In system wide and when the context is loaded, access can only happen
3750 * when the caller is running on the CPU being monitored by the session.
3751 * It does not have to be the owner (ctx_task) of the context per se.
3753 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3754 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3757 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3761 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3762 * ensuring that no real breakpoint can be installed via this call.
3764 * IMPORTANT: regs can be NULL in this function
3767 first_time = ctx->ctx_fl_using_dbreg == 0;
3770 * don't bother if we are loaded and task is being debugged
3772 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3773 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3778 * check for debug registers in system wide mode
3780 * If though a check is done in pfm_context_load(),
3781 * we must repeat it here, in case the registers are
3782 * written after the context is loaded
3787 if (first_time && is_system) {
3788 if (pfm_sessions.pfs_ptrace_use_dbregs)
3791 pfm_sessions.pfs_sys_use_dbregs++;
3796 if (ret != 0) return ret;
3799 * mark ourself as user of the debug registers for
3802 ctx->ctx_fl_using_dbreg = 1;
3805 * clear hardware registers to make sure we don't
3806 * pick up stale state.
3808 * for a system wide session, we do not use
3809 * thread.dbr, thread.ibr because this process
3810 * never leaves the current CPU and the state
3811 * is shared by all processes running on it
3813 if (first_time && can_access_pmu) {
3814 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3815 for (i=0; i < pmu_conf->num_ibrs; i++) {
3816 ia64_set_ibr(i, 0UL);
3817 ia64_dv_serialize_instruction();
3820 for (i=0; i < pmu_conf->num_dbrs; i++) {
3821 ia64_set_dbr(i, 0UL);
3822 ia64_dv_serialize_data();
3828 * Now install the values into the registers
3830 for (i = 0; i < count; i++, req++) {
3832 rnum = req->dbreg_num;
3833 dbreg.val = req->dbreg_value;
3837 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3838 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3839 rnum, dbreg.val, mode, i, count));
3845 * make sure we do not install enabled breakpoint
3848 if (mode == PFM_CODE_RR)
3849 dbreg.ibr.ibr_x = 0;
3851 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3854 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3857 * Debug registers, just like PMC, can only be modified
3858 * by a kernel call. Moreover, perfmon() access to those
3859 * registers are centralized in this routine. The hardware
3860 * does not modify the value of these registers, therefore,
3861 * if we save them as they are written, we can avoid having
3862 * to save them on context switch out. This is made possible
3863 * by the fact that when perfmon uses debug registers, ptrace()
3864 * won't be able to modify them concurrently.
3866 if (mode == PFM_CODE_RR) {
3867 CTX_USED_IBR(ctx, rnum);
3869 if (can_access_pmu) {
3870 ia64_set_ibr(rnum, dbreg.val);
3871 ia64_dv_serialize_instruction();
3874 ctx->ctx_ibrs[rnum] = dbreg.val;
3876 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x is_loaded=%d access_pmu=%d\n",
3877 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3879 CTX_USED_DBR(ctx, rnum);
3881 if (can_access_pmu) {
3882 ia64_set_dbr(rnum, dbreg.val);
3883 ia64_dv_serialize_data();
3885 ctx->ctx_dbrs[rnum] = dbreg.val;
3887 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x is_loaded=%d access_pmu=%d\n",
3888 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3896 * in case it was our first attempt, we undo the global modifications
3900 if (ctx->ctx_fl_system) {
3901 pfm_sessions.pfs_sys_use_dbregs--;
3904 ctx->ctx_fl_using_dbreg = 0;
3907 * install error return flag
3909 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3915 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3917 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3921 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3923 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3927 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3931 if (req == NULL) return -EINVAL;
3933 ctx = GET_PMU_CTX();
3935 if (ctx == NULL) return -EINVAL;
3938 * for now limit to current task, which is enough when calling
3939 * from overflow handler
3941 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3943 return pfm_write_ibrs(ctx, req, nreq, regs);
3945 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3948 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3952 if (req == NULL) return -EINVAL;
3954 ctx = GET_PMU_CTX();
3956 if (ctx == NULL) return -EINVAL;
3959 * for now limit to current task, which is enough when calling
3960 * from overflow handler
3962 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3964 return pfm_write_dbrs(ctx, req, nreq, regs);
3966 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3970 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3972 pfarg_features_t *req = (pfarg_features_t *)arg;
3974 req->ft_version = PFM_VERSION;
3979 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3981 struct pt_regs *tregs;
3982 struct task_struct *task = PFM_CTX_TASK(ctx);
3983 int state, is_system;
3985 state = ctx->ctx_state;
3986 is_system = ctx->ctx_fl_system;
3988 if (state != PFM_CTX_LOADED && state != PFM_CTX_MASKED) return -EINVAL;
3991 * In system wide and when the context is loaded, access can only happen
3992 * when the caller is running on the CPU being monitored by the session.
3993 * It does not have to be the owner (ctx_task) of the context per se.
3995 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3996 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3999 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4000 PFM_CTX_TASK(ctx)->pid,
4004 * in system mode, we need to update the PMU directly
4005 * and the user level state of the caller, which may not
4006 * necessarily be the creator of the context.
4010 * Update local PMU first
4014 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4018 * update local cpuinfo
4020 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4023 * stop monitoring, does srlz.i
4028 * stop monitoring in the caller
4030 ia64_psr(regs)->pp = 0;
4038 if (task == current) {
4039 /* stop monitoring at kernel level */
4043 * stop monitoring at the user level
4045 ia64_psr(regs)->up = 0;
4047 tregs = ia64_task_regs(task);
4050 * stop monitoring at the user level
4052 ia64_psr(tregs)->up = 0;
4055 * monitoring disabled in kernel at next reschedule
4057 ctx->ctx_saved_psr_up = 0;
4058 DPRINT(("task=[%d]\n", task->pid));
4065 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4067 struct pt_regs *tregs;
4068 int state, is_system;
4070 state = ctx->ctx_state;
4071 is_system = ctx->ctx_fl_system;
4073 if (state != PFM_CTX_LOADED) return -EINVAL;
4076 * In system wide and when the context is loaded, access can only happen
4077 * when the caller is running on the CPU being monitored by the session.
4078 * It does not have to be the owner (ctx_task) of the context per se.
4080 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4081 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4086 * in system mode, we need to update the PMU directly
4087 * and the user level state of the caller, which may not
4088 * necessarily be the creator of the context.
4093 * set user level psr.pp for the caller
4095 ia64_psr(regs)->pp = 1;
4098 * now update the local PMU and cpuinfo
4100 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4103 * start monitoring at kernel level
4108 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4118 if (ctx->ctx_task == current) {
4120 /* start monitoring at kernel level */
4124 * activate monitoring at user level
4126 ia64_psr(regs)->up = 1;
4129 tregs = ia64_task_regs(ctx->ctx_task);
4132 * start monitoring at the kernel level the next
4133 * time the task is scheduled
4135 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4138 * activate monitoring at user level
4140 ia64_psr(tregs)->up = 1;
4146 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4148 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4153 for (i = 0; i < count; i++, req++) {
4155 cnum = req->reg_num;
4157 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4159 req->reg_value = PMC_DFL_VAL(cnum);
4161 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4163 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4168 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4173 pfm_check_task_exist(pfm_context_t *ctx)
4175 struct task_struct *g, *t;
4178 read_lock(&tasklist_lock);
4180 do_each_thread (g, t) {
4181 if (t->thread.pfm_context == ctx) {
4185 } while_each_thread (g, t);
4187 read_unlock(&tasklist_lock);
4189 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4195 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4197 struct task_struct *task;
4198 struct thread_struct *thread;
4199 struct pfm_context_t *old;
4200 unsigned long flags;
4202 struct task_struct *owner_task = NULL;
4204 pfarg_load_t *req = (pfarg_load_t *)arg;
4205 unsigned long *pmcs_source, *pmds_source;
4208 int state, is_system, set_dbregs = 0;
4210 state = ctx->ctx_state;
4211 is_system = ctx->ctx_fl_system;
4213 * can only load from unloaded or terminated state
4215 if (state != PFM_CTX_UNLOADED) {
4216 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4222 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4224 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4225 DPRINT(("cannot use blocking mode on self\n"));
4229 ret = pfm_get_task(ctx, req->load_pid, &task);
4231 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4238 * system wide is self monitoring only
4240 if (is_system && task != current) {
4241 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4246 thread = &task->thread;
4250 * cannot load a context which is using range restrictions,
4251 * into a task that is being debugged.
4253 if (ctx->ctx_fl_using_dbreg) {
4254 if (thread->flags & IA64_THREAD_DBG_VALID) {
4256 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4262 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4263 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4266 pfm_sessions.pfs_sys_use_dbregs++;
4267 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4274 if (ret) goto error;
4278 * SMP system-wide monitoring implies self-monitoring.
4280 * The programming model expects the task to
4281 * be pinned on a CPU throughout the session.
4282 * Here we take note of the current CPU at the
4283 * time the context is loaded. No call from
4284 * another CPU will be allowed.
4286 * The pinning via shed_setaffinity()
4287 * must be done by the calling task prior
4290 * systemwide: keep track of CPU this session is supposed to run on
4292 the_cpu = ctx->ctx_cpu = smp_processor_id();
4296 * now reserve the session
4298 ret = pfm_reserve_session(current, is_system, the_cpu);
4299 if (ret) goto error;
4302 * task is necessarily stopped at this point.
4304 * If the previous context was zombie, then it got removed in
4305 * pfm_save_regs(). Therefore we should not see it here.
4306 * If we see a context, then this is an active context
4308 * XXX: needs to be atomic
4310 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4311 thread->pfm_context, ctx));
4313 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4315 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4319 pfm_reset_msgq(ctx);
4321 ctx->ctx_state = PFM_CTX_LOADED;
4324 * link context to task
4326 ctx->ctx_task = task;
4330 * we load as stopped
4332 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4333 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4335 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4337 thread->flags |= IA64_THREAD_PM_VALID;
4341 * propagate into thread-state
4343 pfm_copy_pmds(task, ctx);
4344 pfm_copy_pmcs(task, ctx);
4346 pmcs_source = thread->pmcs;
4347 pmds_source = thread->pmds;
4350 * always the case for system-wide
4352 if (task == current) {
4354 if (is_system == 0) {
4356 /* allow user level control */
4357 ia64_psr(regs)->sp = 0;
4358 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4360 SET_LAST_CPU(ctx, smp_processor_id());
4362 SET_ACTIVATION(ctx);
4365 * push the other task out, if any
4367 owner_task = GET_PMU_OWNER();
4368 if (owner_task) pfm_lazy_save_regs(owner_task);
4372 * load all PMD from ctx to PMU (as opposed to thread state)
4373 * restore all PMC from ctx to PMU
4375 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4376 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4378 ctx->ctx_reload_pmcs[0] = 0UL;
4379 ctx->ctx_reload_pmds[0] = 0UL;
4382 * guaranteed safe by earlier check against DBG_VALID
4384 if (ctx->ctx_fl_using_dbreg) {
4385 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4386 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4391 SET_PMU_OWNER(task, ctx);
4393 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4396 * when not current, task MUST be stopped, so this is safe
4398 regs = ia64_task_regs(task);
4400 /* force a full reload */
4401 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4402 SET_LAST_CPU(ctx, -1);
4404 /* initial saved psr (stopped) */
4405 ctx->ctx_saved_psr_up = 0UL;
4406 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4412 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4415 * we must undo the dbregs setting (for system-wide)
4417 if (ret && set_dbregs) {
4419 pfm_sessions.pfs_sys_use_dbregs--;
4423 * release task, there is now a link with the context
4425 if (is_system == 0 && task != current) {
4429 ret = pfm_check_task_exist(ctx);
4431 ctx->ctx_state = PFM_CTX_UNLOADED;
4432 ctx->ctx_task = NULL;
4440 * in this function, we do not need to increase the use count
4441 * for the task via get_task_struct(), because we hold the
4442 * context lock. If the task were to disappear while having
4443 * a context attached, it would go through pfm_exit_thread()
4444 * which also grabs the context lock and would therefore be blocked
4445 * until we are here.
4447 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4450 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4452 struct task_struct *task = PFM_CTX_TASK(ctx);
4453 struct pt_regs *tregs;
4454 int prev_state, is_system;
4457 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4459 prev_state = ctx->ctx_state;
4460 is_system = ctx->ctx_fl_system;
4463 * unload only when necessary
4465 if (prev_state == PFM_CTX_UNLOADED) {
4466 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4471 * clear psr and dcr bits
4473 ret = pfm_stop(ctx, NULL, 0, regs);
4474 if (ret) return ret;
4476 ctx->ctx_state = PFM_CTX_UNLOADED;
4479 * in system mode, we need to update the PMU directly
4480 * and the user level state of the caller, which may not
4481 * necessarily be the creator of the context.
4488 * local PMU is taken care of in pfm_stop()
4490 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4491 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4494 * save PMDs in context
4497 pfm_flush_pmds(current, ctx);
4500 * at this point we are done with the PMU
4501 * so we can unreserve the resource.
4503 if (prev_state != PFM_CTX_ZOMBIE)
4504 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4507 * disconnect context from task
4509 task->thread.pfm_context = NULL;
4511 * disconnect task from context
4513 ctx->ctx_task = NULL;
4516 * There is nothing more to cleanup here.
4524 tregs = task == current ? regs : ia64_task_regs(task);
4526 if (task == current) {
4528 * cancel user level control
4530 ia64_psr(regs)->sp = 1;
4532 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4535 * save PMDs to context
4538 pfm_flush_pmds(task, ctx);
4541 * at this point we are done with the PMU
4542 * so we can unreserve the resource.
4544 * when state was ZOMBIE, we have already unreserved.
4546 if (prev_state != PFM_CTX_ZOMBIE)
4547 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4550 * reset activation counter and psr
4552 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4553 SET_LAST_CPU(ctx, -1);
4556 * PMU state will not be restored
4558 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4561 * break links between context and task
4563 task->thread.pfm_context = NULL;
4564 ctx->ctx_task = NULL;
4566 PFM_SET_WORK_PENDING(task, 0);
4568 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4569 ctx->ctx_fl_can_restart = 0;
4570 ctx->ctx_fl_going_zombie = 0;
4572 DPRINT(("disconnected [%d] from context\n", task->pid));
4579 * called only from exit_thread(): task == current
4580 * we come here only if current has a context attached (loaded or masked)
4583 pfm_exit_thread(struct task_struct *task)
4586 unsigned long flags;
4587 struct pt_regs *regs = ia64_task_regs(task);
4591 ctx = PFM_GET_CTX(task);
4593 PROTECT_CTX(ctx, flags);
4595 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4597 state = ctx->ctx_state;
4599 case PFM_CTX_UNLOADED:
4601 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4602 * be in unloaded state
4604 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4606 case PFM_CTX_LOADED:
4607 case PFM_CTX_MASKED:
4608 ret = pfm_context_unload(ctx, NULL, 0, regs);
4610 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4612 DPRINT(("ctx unloaded for current state was %d\n", state));
4614 pfm_end_notify_user(ctx);
4616 case PFM_CTX_ZOMBIE:
4617 ret = pfm_context_unload(ctx, NULL, 0, regs);
4619 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4624 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4627 UNPROTECT_CTX(ctx, flags);
4629 { u64 psr = pfm_get_psr();
4630 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4631 BUG_ON(GET_PMU_OWNER());
4632 BUG_ON(ia64_psr(regs)->up);
4633 BUG_ON(ia64_psr(regs)->pp);
4637 * All memory free operations (especially for vmalloc'ed memory)
4638 * MUST be done with interrupts ENABLED.
4640 if (free_ok) pfm_context_free(ctx);
4644 * functions MUST be listed in the increasing order of their index (see permfon.h)
4646 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4647 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4648 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4649 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4650 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4652 static pfm_cmd_desc_t pfm_cmd_tab[]={
4653 /* 0 */PFM_CMD_NONE,
4654 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4655 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4656 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4657 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4658 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4659 /* 6 */PFM_CMD_NONE,
4660 /* 7 */PFM_CMD_NONE,
4661 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4662 /* 9 */PFM_CMD_NONE,
4663 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4664 /* 11 */PFM_CMD_NONE,
4665 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4666 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4667 /* 14 */PFM_CMD_NONE,
4668 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4669 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4670 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4671 /* 18 */PFM_CMD_NONE,
4672 /* 19 */PFM_CMD_NONE,
4673 /* 20 */PFM_CMD_NONE,
4674 /* 21 */PFM_CMD_NONE,
4675 /* 22 */PFM_CMD_NONE,
4676 /* 23 */PFM_CMD_NONE,
4677 /* 24 */PFM_CMD_NONE,
4678 /* 25 */PFM_CMD_NONE,
4679 /* 26 */PFM_CMD_NONE,
4680 /* 27 */PFM_CMD_NONE,
4681 /* 28 */PFM_CMD_NONE,
4682 /* 29 */PFM_CMD_NONE,
4683 /* 30 */PFM_CMD_NONE,
4684 /* 31 */PFM_CMD_NONE,
4685 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4686 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4688 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4691 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4693 struct task_struct *task;
4694 int state, old_state;
4697 state = ctx->ctx_state;
4698 task = ctx->ctx_task;
4701 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4705 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4709 task->state, PFM_CMD_STOPPED(cmd)));
4712 * self-monitoring always ok.
4714 * for system-wide the caller can either be the creator of the
4715 * context (to one to which the context is attached to) OR
4716 * a task running on the same CPU as the session.
4718 if (task == current || ctx->ctx_fl_system) return 0;
4721 * if context is UNLOADED we are safe to go
4723 if (state == PFM_CTX_UNLOADED) return 0;
4726 * no command can operate on a zombie context
4728 if (state == PFM_CTX_ZOMBIE) {
4729 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4734 * context is LOADED or MASKED. Some commands may need to have
4737 * We could lift this restriction for UP but it would mean that
4738 * the user has no guarantee the task would not run between
4739 * two successive calls to perfmonctl(). That's probably OK.
4740 * If this user wants to ensure the task does not run, then
4741 * the task must be stopped.
4743 if (PFM_CMD_STOPPED(cmd)) {
4744 if (task->state != TASK_STOPPED) {
4745 DPRINT(("[%d] task not in stopped state\n", task->pid));
4749 * task is now stopped, wait for ctxsw out
4751 * This is an interesting point in the code.
4752 * We need to unprotect the context because
4753 * the pfm_save_regs() routines needs to grab
4754 * the same lock. There are danger in doing
4755 * this because it leaves a window open for
4756 * another task to get access to the context
4757 * and possibly change its state. The one thing
4758 * that is not possible is for the context to disappear
4759 * because we are protected by the VFS layer, i.e.,
4760 * get_fd()/put_fd().
4764 UNPROTECT_CTX(ctx, flags);
4766 wait_task_inactive(task);
4768 PROTECT_CTX(ctx, flags);
4771 * we must recheck to verify if state has changed
4773 if (ctx->ctx_state != old_state) {
4774 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4782 * system-call entry point (must return long)
4785 sys_perfmonctl (int fd, int cmd, void *arg, int count, long arg5, long arg6, long arg7,
4786 long arg8, long stack)
4788 struct pt_regs *regs = (struct pt_regs *)&stack;
4789 struct file *file = NULL;
4790 pfm_context_t *ctx = NULL;
4791 unsigned long flags = 0UL;
4792 void *args_k = NULL;
4793 long ret; /* will expand int return types */
4794 size_t base_sz, sz, xtra_sz = 0;
4795 int narg, completed_args = 0, call_made = 0, cmd_flags;
4796 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4797 int (*getsize)(void *arg, size_t *sz);
4798 #define PFM_MAX_ARGSIZE 4096
4801 * reject any call if perfmon was disabled at initialization
4803 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4805 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4806 DPRINT(("invalid cmd=%d\n", cmd));
4810 func = pfm_cmd_tab[cmd].cmd_func;
4811 narg = pfm_cmd_tab[cmd].cmd_narg;
4812 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4813 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4814 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4816 if (unlikely(func == NULL)) {
4817 DPRINT(("invalid cmd=%d\n", cmd));
4821 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4829 * check if number of arguments matches what the command expects
4831 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4835 sz = xtra_sz + base_sz*count;
4837 * limit abuse to min page size
4839 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4840 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4845 * allocate default-sized argument buffer
4847 if (likely(count && args_k == NULL)) {
4848 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4849 if (args_k == NULL) return -ENOMEM;
4857 * assume sz = 0 for command without parameters
4859 if (sz && copy_from_user(args_k, arg, sz)) {
4860 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4865 * check if command supports extra parameters
4867 if (completed_args == 0 && getsize) {
4869 * get extra parameters size (based on main argument)
4871 ret = (*getsize)(args_k, &xtra_sz);
4872 if (ret) goto error_args;
4876 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4878 /* retry if necessary */
4879 if (likely(xtra_sz)) goto restart_args;
4882 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4887 if (unlikely(file == NULL)) {
4888 DPRINT(("invalid fd %d\n", fd));
4891 if (unlikely(PFM_IS_FILE(file) == 0)) {
4892 DPRINT(("fd %d not related to perfmon\n", fd));
4896 ctx = (pfm_context_t *)file->private_data;
4897 if (unlikely(ctx == NULL)) {
4898 DPRINT(("no context for fd %d\n", fd));
4901 prefetch(&ctx->ctx_state);
4903 PROTECT_CTX(ctx, flags);
4906 * check task is stopped
4908 ret = pfm_check_task_state(ctx, cmd, flags);
4909 if (unlikely(ret)) goto abort_locked;
4912 ret = (*func)(ctx, args_k, count, regs);
4918 DPRINT(("context unlocked\n"));
4919 UNPROTECT_CTX(ctx, flags);
4923 /* copy argument back to user, if needed */
4924 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4927 if (args_k) kfree(args_k);
4929 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4935 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4937 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4938 pfm_ovfl_ctrl_t rst_ctrl;
4942 state = ctx->ctx_state;
4944 * Unlock sampling buffer and reset index atomically
4945 * XXX: not really needed when blocking
4947 if (CTX_HAS_SMPL(ctx)) {
4949 rst_ctrl.bits.mask_monitoring = 0;
4950 rst_ctrl.bits.reset_ovfl_pmds = 0;
4952 if (state == PFM_CTX_LOADED)
4953 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4955 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4957 rst_ctrl.bits.mask_monitoring = 0;
4958 rst_ctrl.bits.reset_ovfl_pmds = 1;
4962 if (rst_ctrl.bits.reset_ovfl_pmds) {
4963 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4965 if (rst_ctrl.bits.mask_monitoring == 0) {
4966 DPRINT(("resuming monitoring\n"));
4967 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4969 DPRINT(("stopping monitoring\n"));
4970 //pfm_stop_monitoring(current, regs);
4972 ctx->ctx_state = PFM_CTX_LOADED;
4977 * context MUST BE LOCKED when calling
4978 * can only be called for current
4981 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4983 if (ctx->ctx_fl_system) {
4984 printk(KERN_ERR "perfmon: pfm_context_force_terminate [%d] is system-wide\n", current->pid);
4988 * we stop the whole thing, we do no need to flush
4989 * we know we WERE masked
4992 ia64_psr(regs)->up = 0;
4993 ia64_psr(regs)->sp = 1;
4996 * disconnect the task from the context and vice-versa
4998 current->thread.pfm_context = NULL;
4999 current->thread.flags &= ~IA64_THREAD_PM_VALID;
5000 ctx->ctx_task = NULL;
5002 DPRINT(("context terminated\n"));
5005 * and wakeup controlling task, indicating we are now disconnected
5007 wake_up_interruptible(&ctx->ctx_zombieq);
5010 * given that context is still locked, the controlling
5011 * task will only get access when we return from
5012 * pfm_handle_work().
5016 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5019 pfm_handle_work(void)
5022 struct pt_regs *regs;
5023 unsigned long flags;
5024 unsigned long ovfl_regs;
5025 unsigned int reason;
5028 ctx = PFM_GET_CTX(current);
5030 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5034 PROTECT_CTX(ctx, flags);
5036 PFM_SET_WORK_PENDING(current, 0);
5038 pfm_clear_task_notify();
5040 regs = ia64_task_regs(current);
5043 * extract reason for being here and clear
5045 reason = ctx->ctx_fl_trap_reason;
5046 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5047 ovfl_regs = ctx->ctx_ovfl_regs[0];
5049 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5052 * must be done before we check for simple-reset mode
5054 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5057 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5058 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5060 UNPROTECT_CTX(ctx, flags);
5062 DPRINT(("before block sleeping\n"));
5065 * may go through without blocking on SMP systems
5066 * if restart has been received already by the time we call down()
5068 ret = down_interruptible(&ctx->ctx_restart_sem);
5070 DPRINT(("after block sleeping ret=%d\n", ret));
5072 PROTECT_CTX(ctx, flags);
5075 * we need to read the ovfl_regs only after wake-up
5076 * because we may have had pfm_write_pmds() in between
5077 * and that can changed PMD values and therefore
5078 * ovfl_regs is reset for these new PMD values.
5080 ovfl_regs = ctx->ctx_ovfl_regs[0];
5082 if (ctx->ctx_fl_going_zombie) {
5084 DPRINT(("context is zombie, bailing out\n"));
5085 pfm_context_force_terminate(ctx, regs);
5089 * in case of interruption of down() we don't restart anything
5091 if (ret < 0) goto nothing_to_do;
5094 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5095 ctx->ctx_ovfl_regs[0] = 0UL;
5099 UNPROTECT_CTX(ctx, flags);
5103 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5105 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5106 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5110 DPRINT(("waking up somebody\n"));
5112 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5115 * safe, we are not in intr handler, nor in ctxsw when
5118 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5124 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5126 pfm_msg_t *msg = NULL;
5128 if (ctx->ctx_fl_no_msg == 0) {
5129 msg = pfm_get_new_msg(ctx);
5131 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5135 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5136 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5137 msg->pfm_ovfl_msg.msg_active_set = 0;
5138 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5139 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5140 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5141 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5142 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5145 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5151 return pfm_notify_user(ctx, msg);
5155 pfm_end_notify_user(pfm_context_t *ctx)
5159 msg = pfm_get_new_msg(ctx);
5161 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5165 memset(msg, 0, sizeof(*msg));
5167 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5168 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5169 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5171 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5176 return pfm_notify_user(ctx, msg);
5180 * main overflow processing routine.
5181 * it can be called from the interrupt path or explicitely during the context switch code
5184 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5186 pfm_ovfl_arg_t *ovfl_arg;
5188 unsigned long old_val, ovfl_val, new_val;
5189 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5190 unsigned long tstamp;
5191 pfm_ovfl_ctrl_t ovfl_ctrl;
5192 unsigned int i, has_smpl;
5193 int must_notify = 0;
5195 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5198 * sanity test. Should never happen
5200 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5202 tstamp = ia64_get_itc();
5203 mask = pmc0 >> PMU_FIRST_COUNTER;
5204 ovfl_val = pmu_conf->ovfl_val;
5205 has_smpl = CTX_HAS_SMPL(ctx);
5207 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5208 "used_pmds=0x%lx\n",
5210 task ? task->pid: -1,
5211 (regs ? regs->cr_iip : 0),
5212 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5213 ctx->ctx_used_pmds[0]));
5217 * first we update the virtual counters
5218 * assume there was a prior ia64_srlz_d() issued
5220 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5222 /* skip pmd which did not overflow */
5223 if ((mask & 0x1) == 0) continue;
5226 * Note that the pmd is not necessarily 0 at this point as qualified events
5227 * may have happened before the PMU was frozen. The residual count is not
5228 * taken into consideration here but will be with any read of the pmd via
5231 old_val = new_val = ctx->ctx_pmds[i].val;
5232 new_val += 1 + ovfl_val;
5233 ctx->ctx_pmds[i].val = new_val;
5236 * check for overflow condition
5238 if (likely(old_val > new_val)) {
5239 ovfl_pmds |= 1UL << i;
5240 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5243 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5247 ia64_get_pmd(i) & ovfl_val,
5253 * there was no 64-bit overflow, nothing else to do
5255 if (ovfl_pmds == 0UL) return;
5258 * reset all control bits
5264 * if a sampling format module exists, then we "cache" the overflow by
5265 * calling the module's handler() routine.
5268 unsigned long start_cycles, end_cycles;
5269 unsigned long pmd_mask;
5271 int this_cpu = smp_processor_id();
5273 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5274 ovfl_arg = &ctx->ctx_ovfl_arg;
5276 prefetch(ctx->ctx_smpl_hdr);
5278 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5282 if ((pmd_mask & 0x1) == 0) continue;
5284 ovfl_arg->ovfl_pmd = (unsigned char )i;
5285 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5286 ovfl_arg->active_set = 0;
5287 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5288 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5290 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5291 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5292 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5295 * copy values of pmds of interest. Sampling format may copy them
5296 * into sampling buffer.
5299 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5300 if ((smpl_pmds & 0x1) == 0) continue;
5301 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5302 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5306 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5308 start_cycles = ia64_get_itc();
5311 * call custom buffer format record (handler) routine
5313 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5315 end_cycles = ia64_get_itc();
5318 * For those controls, we take the union because they have
5319 * an all or nothing behavior.
5321 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5322 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5323 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5325 * build the bitmask of pmds to reset now
5327 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5329 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5332 * when the module cannot handle the rest of the overflows, we abort right here
5334 if (ret && pmd_mask) {
5335 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5336 pmd_mask<<PMU_FIRST_COUNTER));
5339 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5341 ovfl_pmds &= ~reset_pmds;
5344 * when no sampling module is used, then the default
5345 * is to notify on overflow if requested by user
5347 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5348 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5349 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5350 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5352 * if needed, we reset all overflowed pmds
5354 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5357 DPRINT(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n",
5361 * reset the requested PMD registers using the short reset values
5364 unsigned long bm = reset_pmds;
5365 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5368 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5370 * keep track of what to reset when unblocking
5372 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5375 * check for blocking context
5377 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5379 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5382 * set the perfmon specific checking pending work for the task
5384 PFM_SET_WORK_PENDING(task, 1);
5387 * when coming from ctxsw, current still points to the
5388 * previous task, therefore we must work with task and not current.
5390 pfm_set_task_notify(task);
5393 * defer until state is changed (shorten spin window). the context is locked
5394 * anyway, so the signal receiver would come spin for nothing.
5399 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5400 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5401 PFM_GET_WORK_PENDING(task),
5402 ctx->ctx_fl_trap_reason,
5405 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5407 * in case monitoring must be stopped, we toggle the psr bits
5409 if (ovfl_ctrl.bits.mask_monitoring) {
5410 pfm_mask_monitoring(task);
5411 ctx->ctx_state = PFM_CTX_MASKED;
5412 ctx->ctx_fl_can_restart = 1;
5416 * send notification now
5418 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5423 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5425 task ? task->pid : -1,
5431 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5432 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5433 * come here as zombie only if the task is the current task. In which case, we
5434 * can access the PMU hardware directly.
5436 * Note that zombies do have PM_VALID set. So here we do the minimal.
5438 * In case the context was zombified it could not be reclaimed at the time
5439 * the monitoring program exited. At this point, the PMU reservation has been
5440 * returned, the sampiing buffer has been freed. We must convert this call
5441 * into a spurious interrupt. However, we must also avoid infinite overflows
5442 * by stopping monitoring for this task. We can only come here for a per-task
5443 * context. All we need to do is to stop monitoring using the psr bits which
5444 * are always task private. By re-enabling secure montioring, we ensure that
5445 * the monitored task will not be able to re-activate monitoring.
5446 * The task will eventually be context switched out, at which point the context
5447 * will be reclaimed (that includes releasing ownership of the PMU).
5449 * So there might be a window of time where the number of per-task session is zero
5450 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5451 * context. This is safe because if a per-task session comes in, it will push this one
5452 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5453 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5454 * also push our zombie context out.
5456 * Overall pretty hairy stuff....
5458 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5460 ia64_psr(regs)->up = 0;
5461 ia64_psr(regs)->sp = 1;
5466 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5468 struct task_struct *task;
5470 unsigned long flags;
5472 int this_cpu = smp_processor_id();
5475 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5478 * srlz.d done before arriving here
5480 pmc0 = ia64_get_pmc(0);
5482 task = GET_PMU_OWNER();
5483 ctx = GET_PMU_CTX();
5486 * if we have some pending bits set
5487 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5489 if (PMC0_HAS_OVFL(pmc0) && task) {
5491 * we assume that pmc0.fr is always set here
5495 if (!ctx) goto report_spurious1;
5497 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5498 goto report_spurious2;
5500 PROTECT_CTX_NOPRINT(ctx, flags);
5502 pfm_overflow_handler(task, ctx, pmc0, regs);
5504 UNPROTECT_CTX_NOPRINT(ctx, flags);
5507 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5511 * keep it unfrozen at all times
5518 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5519 this_cpu, task->pid);
5523 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5531 pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5533 unsigned long start_cycles, total_cycles;
5534 unsigned long min, max;
5538 this_cpu = get_cpu();
5539 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5540 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5542 start_cycles = ia64_get_itc();
5544 ret = pfm_do_interrupt_handler(irq, arg, regs);
5546 total_cycles = ia64_get_itc();
5549 * don't measure spurious interrupts
5551 if (likely(ret == 0)) {
5552 total_cycles -= start_cycles;
5554 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5555 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5557 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5559 put_cpu_no_resched();
5564 * /proc/perfmon interface, for debug only
5567 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5570 pfm_proc_start(struct seq_file *m, loff_t *pos)
5573 return PFM_PROC_SHOW_HEADER;
5576 while (*pos <= NR_CPUS) {
5577 if (cpu_online(*pos - 1)) {
5578 return (void *)*pos;
5586 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5589 return pfm_proc_start(m, pos);
5593 pfm_proc_stop(struct seq_file *m, void *v)
5598 pfm_proc_show_header(struct seq_file *m)
5600 struct list_head * pos;
5601 pfm_buffer_fmt_t * entry;
5602 unsigned long flags;
5605 "perfmon version : %u.%u\n"
5608 "expert mode : %s\n"
5609 "ovfl_mask : 0x%lx\n"
5610 "PMU flags : 0x%x\n",
5611 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5613 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5614 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5621 "proc_sessions : %u\n"
5622 "sys_sessions : %u\n"
5623 "sys_use_dbregs : %u\n"
5624 "ptrace_use_dbregs : %u\n",
5625 pfm_sessions.pfs_task_sessions,
5626 pfm_sessions.pfs_sys_sessions,
5627 pfm_sessions.pfs_sys_use_dbregs,
5628 pfm_sessions.pfs_ptrace_use_dbregs);
5632 spin_lock(&pfm_buffer_fmt_lock);
5634 list_for_each(pos, &pfm_buffer_fmt_list) {
5635 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5636 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5647 entry->fmt_uuid[10],
5648 entry->fmt_uuid[11],
5649 entry->fmt_uuid[12],
5650 entry->fmt_uuid[13],
5651 entry->fmt_uuid[14],
5652 entry->fmt_uuid[15],
5655 spin_unlock(&pfm_buffer_fmt_lock);
5660 pfm_proc_show(struct seq_file *m, void *v)
5666 if (v == PFM_PROC_SHOW_HEADER) {
5667 pfm_proc_show_header(m);
5671 /* show info for CPU (v - 1) */
5675 "CPU%-2d overflow intrs : %lu\n"
5676 "CPU%-2d overflow cycles : %lu\n"
5677 "CPU%-2d overflow min : %lu\n"
5678 "CPU%-2d overflow max : %lu\n"
5679 "CPU%-2d smpl handler calls : %lu\n"
5680 "CPU%-2d smpl handler cycles : %lu\n"
5681 "CPU%-2d spurious intrs : %lu\n"
5682 "CPU%-2d replay intrs : %lu\n"
5683 "CPU%-2d syst_wide : %d\n"
5684 "CPU%-2d dcr_pp : %d\n"
5685 "CPU%-2d exclude idle : %d\n"
5686 "CPU%-2d owner : %d\n"
5687 "CPU%-2d context : %p\n"
5688 "CPU%-2d activations : %lu\n",
5689 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5690 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5691 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5692 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5693 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5694 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5695 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5696 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5697 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5698 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5699 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5700 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5701 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5702 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5704 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5706 psr = pfm_get_psr();
5711 "CPU%-2d psr : 0x%lx\n"
5712 "CPU%-2d pmc0 : 0x%lx\n",
5714 cpu, ia64_get_pmc(0));
5716 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5717 if (PMC_IS_COUNTING(i) == 0) continue;
5719 "CPU%-2d pmc%u : 0x%lx\n"
5720 "CPU%-2d pmd%u : 0x%lx\n",
5721 cpu, i, ia64_get_pmc(i),
5722 cpu, i, ia64_get_pmd(i));
5728 struct seq_operations pfm_seq_ops = {
5729 .start = pfm_proc_start,
5730 .next = pfm_proc_next,
5731 .stop = pfm_proc_stop,
5732 .show = pfm_proc_show
5736 pfm_proc_open(struct inode *inode, struct file *file)
5738 return seq_open(file, &pfm_seq_ops);
5743 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5744 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5745 * is active or inactive based on mode. We must rely on the value in
5746 * local_cpu_data->pfm_syst_info
5749 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5751 struct pt_regs *regs;
5753 unsigned long dcr_pp;
5755 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5758 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5759 * on every CPU, so we can rely on the pid to identify the idle task.
5761 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5762 regs = ia64_task_regs(task);
5763 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5767 * if monitoring has started
5770 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5772 * context switching in?
5775 /* mask monitoring for the idle task */
5776 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5782 * context switching out
5783 * restore monitoring for next task
5785 * Due to inlining this odd if-then-else construction generates
5788 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5797 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5799 struct task_struct *task = ctx->ctx_task;
5801 ia64_psr(regs)->up = 0;
5802 ia64_psr(regs)->sp = 1;
5804 if (GET_PMU_OWNER() == task) {
5805 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5806 SET_PMU_OWNER(NULL, NULL);
5810 * disconnect the task from the context and vice-versa
5812 PFM_SET_WORK_PENDING(task, 0);
5814 task->thread.pfm_context = NULL;
5815 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5817 DPRINT(("force cleanup for [%d]\n", task->pid));
5822 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5825 pfm_save_regs(struct task_struct *task)
5828 struct thread_struct *t;
5829 unsigned long flags;
5833 ctx = PFM_GET_CTX(task);
5834 if (ctx == NULL) return;
5838 * we always come here with interrupts ALREADY disabled by
5839 * the scheduler. So we simply need to protect against concurrent
5840 * access, not CPU concurrency.
5842 flags = pfm_protect_ctx_ctxsw(ctx);
5844 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5845 struct pt_regs *regs = ia64_task_regs(task);
5849 pfm_force_cleanup(ctx, regs);
5851 BUG_ON(ctx->ctx_smpl_hdr);
5853 pfm_unprotect_ctx_ctxsw(ctx, flags);
5855 pfm_context_free(ctx);
5862 if (ctx->ctx_last_activation != GET_ACTIVATION()) {
5863 pfm_unprotect_ctx_ctxsw(ctx, flags);
5868 * save current PSR: needed because we modify it
5871 psr = pfm_get_psr();
5873 BUG_ON(psr & (IA64_PSR_I));
5877 * This is the last instruction which may generate an overflow
5879 * We do not need to set psr.sp because, it is irrelevant in kernel.
5880 * It will be restored from ipsr when going back to user level
5885 * keep a copy of psr.up (for reload)
5887 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5890 * release ownership of this PMU.
5891 * PM interrupts are masked, so nothing
5894 SET_PMU_OWNER(NULL, NULL);
5897 * we systematically save the PMD as we have no
5898 * guarantee we will be schedule at that same
5901 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5904 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5905 * we will need it on the restore path to check
5906 * for pending overflow.
5908 t->pmcs[0] = ia64_get_pmc(0);
5911 * unfreeze PMU if had pending overflows
5913 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5916 * finally, allow context access.
5917 * interrupts will still be masked after this call.
5919 pfm_unprotect_ctx_ctxsw(ctx, flags);
5922 #else /* !CONFIG_SMP */
5924 pfm_save_regs(struct task_struct *task)
5929 ctx = PFM_GET_CTX(task);
5930 if (ctx == NULL) return;
5933 * save current PSR: needed because we modify it
5935 psr = pfm_get_psr();
5937 BUG_ON(psr & (IA64_PSR_I));
5941 * This is the last instruction which may generate an overflow
5943 * We do not need to set psr.sp because, it is irrelevant in kernel.
5944 * It will be restored from ipsr when going back to user level
5949 * keep a copy of psr.up (for reload)
5951 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5955 pfm_lazy_save_regs (struct task_struct *task)
5958 struct thread_struct *t;
5959 unsigned long flags;
5961 { u64 psr = pfm_get_psr();
5962 BUG_ON(psr & IA64_PSR_UP);
5965 ctx = PFM_GET_CTX(task);
5969 * we need to mask PMU overflow here to
5970 * make sure that we maintain pmc0 until
5971 * we save it. overflow interrupts are
5972 * treated as spurious if there is no
5975 * XXX: I don't think this is necessary
5977 PROTECT_CTX(ctx,flags);
5980 * release ownership of this PMU.
5981 * must be done before we save the registers.
5983 * after this call any PMU interrupt is treated
5986 SET_PMU_OWNER(NULL, NULL);
5989 * save all the pmds we use
5991 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5994 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5995 * it is needed to check for pended overflow
5996 * on the restore path
5998 t->pmcs[0] = ia64_get_pmc(0);
6001 * unfreeze PMU if had pending overflows
6003 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6006 * now get can unmask PMU interrupts, they will
6007 * be treated as purely spurious and we will not
6008 * lose any information
6010 UNPROTECT_CTX(ctx,flags);
6012 #endif /* CONFIG_SMP */
6016 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6019 pfm_load_regs (struct task_struct *task)
6022 struct thread_struct *t;
6023 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6024 unsigned long flags;
6026 int need_irq_resend;
6028 ctx = PFM_GET_CTX(task);
6029 if (unlikely(ctx == NULL)) return;
6031 BUG_ON(GET_PMU_OWNER());
6035 * possible on unload
6037 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6040 * we always come here with interrupts ALREADY disabled by
6041 * the scheduler. So we simply need to protect against concurrent
6042 * access, not CPU concurrency.
6044 flags = pfm_protect_ctx_ctxsw(ctx);
6045 psr = pfm_get_psr();
6047 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6049 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6050 BUG_ON(psr & IA64_PSR_I);
6052 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6053 struct pt_regs *regs = ia64_task_regs(task);
6055 BUG_ON(ctx->ctx_smpl_hdr);
6057 pfm_force_cleanup(ctx, regs);
6059 pfm_unprotect_ctx_ctxsw(ctx, flags);
6062 * this one (kmalloc'ed) is fine with interrupts disabled
6064 pfm_context_free(ctx);
6070 * we restore ALL the debug registers to avoid picking up
6073 if (ctx->ctx_fl_using_dbreg) {
6074 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6075 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6078 * retrieve saved psr.up
6080 psr_up = ctx->ctx_saved_psr_up;
6083 * if we were the last user of the PMU on that CPU,
6084 * then nothing to do except restore psr
6086 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6089 * retrieve partial reload masks (due to user modifications)
6091 pmc_mask = ctx->ctx_reload_pmcs[0];
6092 pmd_mask = ctx->ctx_reload_pmds[0];
6096 * To avoid leaking information to the user level when psr.sp=0,
6097 * we must reload ALL implemented pmds (even the ones we don't use).
6098 * In the kernel we only allow PFM_READ_PMDS on registers which
6099 * we initialized or requested (sampling) so there is no risk there.
6101 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6104 * ALL accessible PMCs are systematically reloaded, unused registers
6105 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6106 * up stale configuration.
6108 * PMC0 is never in the mask. It is always restored separately.
6110 pmc_mask = ctx->ctx_all_pmcs[0];
6113 * when context is MASKED, we will restore PMC with plm=0
6114 * and PMD with stale information, but that's ok, nothing
6117 * XXX: optimize here
6119 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6120 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6123 * check for pending overflow at the time the state
6126 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6128 * reload pmc0 with the overflow information
6129 * On McKinley PMU, this will trigger a PMU interrupt
6131 ia64_set_pmc(0, t->pmcs[0]);
6136 * will replay the PMU interrupt
6138 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6140 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6144 * we just did a reload, so we reset the partial reload fields
6146 ctx->ctx_reload_pmcs[0] = 0UL;
6147 ctx->ctx_reload_pmds[0] = 0UL;
6149 SET_LAST_CPU(ctx, smp_processor_id());
6152 * dump activation value for this PMU
6156 * record current activation for this context
6158 SET_ACTIVATION(ctx);
6161 * establish new ownership.
6163 SET_PMU_OWNER(task, ctx);
6166 * restore the psr.up bit. measurement
6168 * no PMU interrupt can happen at this point
6169 * because we still have interrupts disabled.
6171 if (likely(psr_up)) pfm_set_psr_up();
6174 * allow concurrent access to context
6176 pfm_unprotect_ctx_ctxsw(ctx, flags);
6178 #else /* !CONFIG_SMP */
6180 * reload PMU state for UP kernels
6181 * in 2.5 we come here with interrupts disabled
6184 pfm_load_regs (struct task_struct *task)
6186 struct thread_struct *t;
6188 struct task_struct *owner;
6189 unsigned long pmd_mask, pmc_mask;
6191 int need_irq_resend;
6193 owner = GET_PMU_OWNER();
6194 ctx = PFM_GET_CTX(task);
6196 psr = pfm_get_psr();
6198 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6199 BUG_ON(psr & IA64_PSR_I);
6202 * we restore ALL the debug registers to avoid picking up
6205 * This must be done even when the task is still the owner
6206 * as the registers may have been modified via ptrace()
6207 * (not perfmon) by the previous task.
6209 if (ctx->ctx_fl_using_dbreg) {
6210 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6211 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6215 * retrieved saved psr.up
6217 psr_up = ctx->ctx_saved_psr_up;
6218 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6221 * short path, our state is still there, just
6222 * need to restore psr and we go
6224 * we do not touch either PMC nor PMD. the psr is not touched
6225 * by the overflow_handler. So we are safe w.r.t. to interrupt
6226 * concurrency even without interrupt masking.
6228 if (likely(owner == task)) {
6229 if (likely(psr_up)) pfm_set_psr_up();
6234 * someone else is still using the PMU, first push it out and
6235 * then we'll be able to install our stuff !
6237 * Upon return, there will be no owner for the current PMU
6239 if (owner) pfm_lazy_save_regs(owner);
6242 * To avoid leaking information to the user level when psr.sp=0,
6243 * we must reload ALL implemented pmds (even the ones we don't use).
6244 * In the kernel we only allow PFM_READ_PMDS on registers which
6245 * we initialized or requested (sampling) so there is no risk there.
6247 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6250 * ALL accessible PMCs are systematically reloaded, unused registers
6251 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6252 * up stale configuration.
6254 * PMC0 is never in the mask. It is always restored separately
6256 pmc_mask = ctx->ctx_all_pmcs[0];
6258 pfm_restore_pmds(t->pmds, pmd_mask);
6259 pfm_restore_pmcs(t->pmcs, pmc_mask);
6262 * check for pending overflow at the time the state
6265 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6267 * reload pmc0 with the overflow information
6268 * On McKinley PMU, this will trigger a PMU interrupt
6270 ia64_set_pmc(0, t->pmcs[0]);
6276 * will replay the PMU interrupt
6278 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6280 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6284 * establish new ownership.
6286 SET_PMU_OWNER(task, ctx);
6289 * restore the psr.up bit. measurement
6291 * no PMU interrupt can happen at this point
6292 * because we still have interrupts disabled.
6294 if (likely(psr_up)) pfm_set_psr_up();
6296 #endif /* CONFIG_SMP */
6299 * this function assumes monitoring is stopped
6302 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6305 unsigned long mask2, val, pmd_val, ovfl_val;
6306 int i, can_access_pmu = 0;
6310 * is the caller the task being monitored (or which initiated the
6311 * session for system wide measurements)
6313 is_self = ctx->ctx_task == task ? 1 : 0;
6316 if (task == current) {
6319 * in UP, the state can still be in the registers
6321 if (task == current || GET_PMU_OWNER() == task) {
6325 * Mark the PMU as not owned
6326 * This will cause the interrupt handler to do nothing in case an overflow
6327 * interrupt was in-flight
6328 * This also guarantees that pmc0 will contain the final state
6329 * It virtually gives us full control on overflow processing from that point
6332 SET_PMU_OWNER(NULL, NULL);
6335 * read current overflow status:
6337 * we are guaranteed to read the final stable state
6340 pmc0 = ia64_get_pmc(0); /* slow */
6343 * reset freeze bit, overflow status information destroyed
6347 pmc0 = task->thread.pmcs[0];
6349 * clear whatever overflow status bits there were
6351 task->thread.pmcs[0] = 0;
6353 ovfl_val = pmu_conf->ovfl_val;
6355 * we save all the used pmds
6356 * we take care of overflows for counting PMDs
6358 * XXX: sampling situation is not taken into account here
6360 mask2 = ctx->ctx_used_pmds[0];
6361 for (i = 0; mask2; i++, mask2>>=1) {
6363 /* skip non used pmds */
6364 if ((mask2 & 0x1) == 0) continue;
6367 * can access PMU always true in system wide mode
6369 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6371 if (PMD_IS_COUNTING(i)) {
6372 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6375 ctx->ctx_pmds[i].val,
6379 * we rebuild the full 64 bit value of the counter
6381 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6384 * now everything is in ctx_pmds[] and we need
6385 * to clear the saved context from save_regs() such that
6386 * pfm_read_pmds() gets the correct value
6391 * take care of overflow inline
6393 if (pmc0 & (1UL << i)) {
6394 val += 1 + ovfl_val;
6395 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6399 DPRINT(("[%d] is_self=%d ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, is_self, i, val, pmd_val));
6401 if (is_self) task->thread.pmds[i] = pmd_val;
6403 ctx->ctx_pmds[i].val = val;
6407 static struct irqaction perfmon_irqaction = {
6408 .handler = pfm_interrupt_handler,
6409 .flags = SA_INTERRUPT,
6414 * perfmon initialization routine, called from the initcall() table
6416 static int init_pfm_fs(void);
6424 family = local_cpu_data->family;
6429 if ((*p)->probe() == 0) goto found;
6430 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6441 static struct file_operations pfm_proc_fops = {
6442 .open = pfm_proc_open,
6444 .llseek = seq_lseek,
6445 .release = seq_release,
6451 unsigned int n, n_counters, i;
6453 printk("perfmon: version %u.%u IRQ %u\n",
6456 IA64_PERFMON_VECTOR);
6458 if (pfm_probe_pmu()) {
6459 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6460 local_cpu_data->family);
6465 * compute the number of implemented PMD/PMC from the
6466 * description tables
6469 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6470 if (PMC_IS_IMPL(i) == 0) continue;
6471 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6474 pmu_conf->num_pmcs = n;
6476 n = 0; n_counters = 0;
6477 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6478 if (PMD_IS_IMPL(i) == 0) continue;
6479 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6481 if (PMD_IS_COUNTING(i)) n_counters++;
6483 pmu_conf->num_pmds = n;
6484 pmu_conf->num_counters = n_counters;
6487 * sanity checks on the number of debug registers
6489 if (pmu_conf->use_rr_dbregs) {
6490 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6491 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6495 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6496 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6502 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6506 pmu_conf->num_counters,
6507 ffz(pmu_conf->ovfl_val));
6510 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6511 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6517 * create /proc/perfmon (mostly for debugging purposes)
6519 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6520 if (perfmon_dir == NULL) {
6521 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6526 * install customized file operations for /proc/perfmon entry
6528 perfmon_dir->proc_fops = &pfm_proc_fops;
6531 * create /proc/sys/kernel/perfmon (for debugging purposes)
6533 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6536 * initialize all our spinlocks
6538 spin_lock_init(&pfm_sessions.pfs_lock);
6539 spin_lock_init(&pfm_buffer_fmt_lock);
6543 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6548 __initcall(pfm_init);
6551 * this function is called before pfm_init()
6554 pfm_init_percpu (void)
6557 * make sure no measurement is active
6558 * (may inherit programmed PMCs from EFI).
6564 * we run with the PMU not frozen at all times
6568 if (smp_processor_id() == 0)
6569 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6571 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6576 * used for debug purposes only
6579 dump_pmu_state(const char *from)
6581 struct task_struct *task;
6582 struct thread_struct *t;
6583 struct pt_regs *regs;
6585 unsigned long psr, dcr, info, flags;
6588 local_irq_save(flags);
6590 this_cpu = smp_processor_id();
6591 regs = ia64_task_regs(current);
6592 info = PFM_CPUINFO_GET();
6593 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6595 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6596 local_irq_restore(flags);
6600 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6607 task = GET_PMU_OWNER();
6608 ctx = GET_PMU_CTX();
6610 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6612 psr = pfm_get_psr();
6614 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",
6617 psr & IA64_PSR_PP ? 1 : 0,
6618 psr & IA64_PSR_UP ? 1 : 0,
6619 dcr & IA64_DCR_PP ? 1 : 0,
6622 ia64_psr(regs)->pp);
6624 ia64_psr(regs)->up = 0;
6625 ia64_psr(regs)->pp = 0;
6627 t = ¤t->thread;
6629 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6630 if (PMC_IS_IMPL(i) == 0) continue;
6631 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6634 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6635 if (PMD_IS_IMPL(i) == 0) continue;
6636 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6640 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6643 ctx->ctx_smpl_vaddr,
6647 ctx->ctx_saved_psr_up);
6649 local_irq_restore(flags);
6653 * called from process.c:copy_thread(). task is new child.
6656 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6658 struct thread_struct *thread;
6660 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6662 thread = &task->thread;
6665 * cut links inherited from parent (current)
6667 thread->pfm_context = NULL;
6669 PFM_SET_WORK_PENDING(task, 0);
6672 * the psr bits are already set properly in copy_threads()
6675 #else /* !CONFIG_PERFMON */
6677 sys_perfmonctl (int fd, int cmd, void *arg, int count, long arg5, long arg6, long arg7,
6678 long arg8, long stack)
6682 #endif /* CONFIG_PERFMON */