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, 0);
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_lseek(struct file *file, loff_t offset, int whence)
1518 DPRINT(("pfm_lseek called\n"));
1523 pfm_read(struct file *filp, char *buf, size_t size, loff_t *ppos)
1528 unsigned long flags;
1529 DECLARE_WAITQUEUE(wait, current);
1530 if (PFM_IS_FILE(filp) == 0) {
1531 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1535 ctx = (pfm_context_t *)filp->private_data;
1537 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1542 * check even when there is no message
1544 if (size < sizeof(pfm_msg_t)) {
1545 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1549 * seeks are not allowed on message queues
1551 if (ppos != &filp->f_pos) return -ESPIPE;
1553 PROTECT_CTX(ctx, flags);
1556 * put ourselves on the wait queue
1558 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1566 set_current_state(TASK_INTERRUPTIBLE);
1568 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1571 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1573 UNPROTECT_CTX(ctx, flags);
1576 * check non-blocking read
1579 if(filp->f_flags & O_NONBLOCK) break;
1582 * check pending signals
1584 if(signal_pending(current)) {
1589 * no message, so wait
1593 PROTECT_CTX(ctx, flags);
1595 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1596 set_current_state(TASK_RUNNING);
1597 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1599 if (ret < 0) goto abort;
1602 msg = pfm_get_next_msg(ctx);
1604 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1608 DPRINT(("[%d] fd=%d type=%d\n", current->pid, msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1611 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1614 UNPROTECT_CTX(ctx, flags);
1620 pfm_write(struct file *file, const char *ubuf,
1621 size_t size, loff_t *ppos)
1623 DPRINT(("pfm_write called\n"));
1628 pfm_poll(struct file *filp, poll_table * wait)
1631 unsigned long flags;
1632 unsigned int mask = 0;
1634 if (PFM_IS_FILE(filp) == 0) {
1635 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1639 ctx = (pfm_context_t *)filp->private_data;
1641 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1646 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1648 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1650 PROTECT_CTX(ctx, flags);
1652 if (PFM_CTXQ_EMPTY(ctx) == 0)
1653 mask = POLLIN | POLLRDNORM;
1655 UNPROTECT_CTX(ctx, flags);
1657 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1663 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1665 DPRINT(("pfm_ioctl called\n"));
1670 * context is locked when coming here and interrupts are disabled
1673 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1677 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1679 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1683 ctx->ctx_async_queue, ret));
1689 pfm_fasync(int fd, struct file *filp, int on)
1692 unsigned long flags;
1695 if (PFM_IS_FILE(filp) == 0) {
1696 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1700 ctx = (pfm_context_t *)filp->private_data;
1702 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1707 PROTECT_CTX(ctx, flags);
1709 ret = pfm_do_fasync(fd, filp, ctx, on);
1711 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1714 ctx->ctx_async_queue, ret));
1716 UNPROTECT_CTX(ctx, flags);
1723 * this function is exclusively called from pfm_close().
1724 * The context is not protected at that time, nor are interrupts
1725 * on the remote CPU. That's necessary to avoid deadlocks.
1728 pfm_syswide_force_stop(void *info)
1730 pfm_context_t *ctx = (pfm_context_t *)info;
1731 struct pt_regs *regs = ia64_task_regs(current);
1732 struct task_struct *owner;
1733 unsigned long flags;
1736 if (ctx->ctx_cpu != smp_processor_id()) {
1737 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1739 smp_processor_id());
1742 owner = GET_PMU_OWNER();
1743 if (owner != ctx->ctx_task) {
1744 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1746 owner->pid, ctx->ctx_task->pid);
1749 if (GET_PMU_CTX() != ctx) {
1750 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1752 GET_PMU_CTX(), ctx);
1756 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1758 * the context is already protected in pfm_close(), we simply
1759 * need to mask interrupts to avoid a PMU interrupt race on
1762 local_irq_save(flags);
1764 ret = pfm_context_unload(ctx, NULL, 0, regs);
1766 DPRINT(("context_unload returned %d\n", ret));
1770 * unmask interrupts, PMU interrupts are now spurious here
1772 local_irq_restore(flags);
1776 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1780 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1781 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1782 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1784 #endif /* CONFIG_SMP */
1787 * called for each close(). Partially free resources.
1788 * When caller is self-monitoring, the context is unloaded.
1791 pfm_flush(struct file *filp)
1794 struct task_struct *task;
1795 struct pt_regs *regs;
1796 unsigned long flags;
1797 unsigned long smpl_buf_size = 0UL;
1798 void *smpl_buf_vaddr = NULL;
1799 int state, is_system;
1801 if (PFM_IS_FILE(filp) == 0) {
1802 DPRINT(("bad magic for\n"));
1806 ctx = (pfm_context_t *)filp->private_data;
1808 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1813 * remove our file from the async queue, if we use this mode.
1814 * This can be done without the context being protected. We come
1815 * here when the context has become unreacheable by other tasks.
1817 * We may still have active monitoring at this point and we may
1818 * end up in pfm_overflow_handler(). However, fasync_helper()
1819 * operates with interrupts disabled and it cleans up the
1820 * queue. If the PMU handler is called prior to entering
1821 * fasync_helper() then it will send a signal. If it is
1822 * invoked after, it will find an empty queue and no
1823 * signal will be sent. In both case, we are safe
1825 if (filp->f_flags & FASYNC) {
1826 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1827 pfm_do_fasync (-1, filp, ctx, 0);
1830 PROTECT_CTX(ctx, flags);
1832 state = ctx->ctx_state;
1833 is_system = ctx->ctx_fl_system;
1835 task = PFM_CTX_TASK(ctx);
1836 regs = ia64_task_regs(task);
1838 DPRINT(("ctx_state=%d is_current=%d\n",
1840 task == current ? 1 : 0));
1843 * if state == UNLOADED, then task is NULL
1847 * we must stop and unload because we are losing access to the context.
1849 if (task == current) {
1852 * the task IS the owner but it migrated to another CPU: that's bad
1853 * but we must handle this cleanly. Unfortunately, the kernel does
1854 * not provide a mechanism to block migration (while the context is loaded).
1856 * We need to release the resource on the ORIGINAL cpu.
1858 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1860 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1862 * keep context protected but unmask interrupt for IPI
1864 local_irq_restore(flags);
1866 pfm_syswide_cleanup_other_cpu(ctx);
1869 * restore interrupt masking
1871 local_irq_save(flags);
1874 * context is unloaded at this point
1877 #endif /* CONFIG_SMP */
1880 DPRINT(("forcing unload\n"));
1882 * stop and unload, returning with state UNLOADED
1883 * and session unreserved.
1885 pfm_context_unload(ctx, NULL, 0, regs);
1887 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1892 * remove virtual mapping, if any, for the calling task.
1893 * cannot reset ctx field until last user is calling close().
1895 * ctx_smpl_vaddr must never be cleared because it is needed
1896 * by every task with access to the context
1898 * When called from do_exit(), the mm context is gone already, therefore
1899 * mm is NULL, i.e., the VMA is already gone and we do not have to
1902 if (ctx->ctx_smpl_vaddr && current->mm) {
1903 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1904 smpl_buf_size = ctx->ctx_smpl_size;
1907 UNPROTECT_CTX(ctx, flags);
1910 * if there was a mapping, then we systematically remove it
1911 * at this point. Cannot be done inside critical section
1912 * because some VM function reenables interrupts.
1915 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1920 * called either on explicit close() or from exit_files().
1921 * Only the LAST user of the file gets to this point, i.e., it is
1924 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1925 * (fput()),i.e, last task to access the file. Nobody else can access the
1926 * file at this point.
1928 * When called from exit_files(), the VMA has been freed because exit_mm()
1929 * is executed before exit_files().
1931 * When called from exit_files(), the current task is not yet ZOMBIE but we
1932 * flush the PMU state to the context.
1935 pfm_close(struct inode *inode, struct file *filp)
1938 struct task_struct *task;
1939 struct pt_regs *regs;
1940 DECLARE_WAITQUEUE(wait, current);
1941 unsigned long flags;
1942 unsigned long smpl_buf_size = 0UL;
1943 void *smpl_buf_addr = NULL;
1944 int free_possible = 1;
1945 int state, is_system;
1947 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1949 if (PFM_IS_FILE(filp) == 0) {
1950 DPRINT(("bad magic\n"));
1954 ctx = (pfm_context_t *)filp->private_data;
1956 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1960 PROTECT_CTX(ctx, flags);
1962 state = ctx->ctx_state;
1963 is_system = ctx->ctx_fl_system;
1965 task = PFM_CTX_TASK(ctx);
1966 regs = ia64_task_regs(task);
1968 DPRINT(("ctx_state=%d is_current=%d\n",
1970 task == current ? 1 : 0));
1973 * if task == current, then pfm_flush() unloaded the context
1975 if (state == PFM_CTX_UNLOADED) goto doit;
1978 * context is loaded/masked and task != current, we need to
1979 * either force an unload or go zombie
1983 * The task is currently blocked or will block after an overflow.
1984 * we must force it to wakeup to get out of the
1985 * MASKED state and transition to the unloaded state by itself.
1987 * This situation is only possible for per-task mode
1989 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1992 * set a "partial" zombie state to be checked
1993 * upon return from down() in pfm_handle_work().
1995 * We cannot use the ZOMBIE state, because it is checked
1996 * by pfm_load_regs() which is called upon wakeup from down().
1997 * In such case, it would free the context and then we would
1998 * return to pfm_handle_work() which would access the
1999 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2000 * but visible to pfm_handle_work().
2002 * For some window of time, we have a zombie context with
2003 * ctx_state = MASKED and not ZOMBIE
2005 ctx->ctx_fl_going_zombie = 1;
2008 * force task to wake up from MASKED state
2010 up(&ctx->ctx_restart_sem);
2012 DPRINT(("waking up ctx_state=%d\n", state));
2015 * put ourself to sleep waiting for the other
2016 * task to report completion
2018 * the context is protected by mutex, therefore there
2019 * is no risk of being notified of completion before
2020 * begin actually on the waitq.
2022 set_current_state(TASK_INTERRUPTIBLE);
2023 add_wait_queue(&ctx->ctx_zombieq, &wait);
2025 UNPROTECT_CTX(ctx, flags);
2028 * XXX: check for signals :
2029 * - ok of explicit close
2030 * - not ok when coming from exit_files()
2035 PROTECT_CTX(ctx, flags);
2038 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2039 set_current_state(TASK_RUNNING);
2042 * context is unloaded at this point
2044 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2046 else if (task != current) {
2049 * switch context to zombie state
2051 ctx->ctx_state = PFM_CTX_ZOMBIE;
2053 DPRINT(("zombie ctx for [%d]\n", task->pid));
2055 * cannot free the context on the spot. deferred until
2056 * the task notices the ZOMBIE state
2060 pfm_context_unload(ctx, NULL, 0, regs);
2065 /* reload state, may have changed during opening of critical section */
2066 state = ctx->ctx_state;
2069 * the context is still attached to a task (possibly current)
2070 * we cannot destroy it right now
2074 * we must free the sampling buffer right here because
2075 * we cannot rely on it being cleaned up later by the
2076 * monitored task. It is not possible to free vmalloc'ed
2077 * memory in pfm_load_regs(). Instead, we remove the buffer
2078 * now. should there be subsequent PMU overflow originally
2079 * meant for sampling, the will be converted to spurious
2080 * and that's fine because the monitoring tools is gone anyway.
2082 if (ctx->ctx_smpl_hdr) {
2083 smpl_buf_addr = ctx->ctx_smpl_hdr;
2084 smpl_buf_size = ctx->ctx_smpl_size;
2085 /* no more sampling */
2086 ctx->ctx_smpl_hdr = NULL;
2087 ctx->ctx_fl_is_sampling = 0;
2090 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2096 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2099 * UNLOADED that the session has already been unreserved.
2101 if (state == PFM_CTX_ZOMBIE) {
2102 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2106 * disconnect file descriptor from context must be done
2109 filp->private_data = NULL;
2112 * if we free on the spot, the context is now completely unreacheable
2113 * from the callers side. The monitored task side is also cut, so we
2116 * If we have a deferred free, only the caller side is disconnected.
2118 UNPROTECT_CTX(ctx, flags);
2121 * All memory free operations (especially for vmalloc'ed memory)
2122 * MUST be done with interrupts ENABLED.
2124 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2127 * return the memory used by the context
2129 if (free_possible) pfm_context_free(ctx);
2135 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2137 DPRINT(("pfm_no_open called\n"));
2143 static struct file_operations pfm_file_ops = {
2144 .llseek = pfm_lseek,
2149 .open = pfm_no_open, /* special open code to disallow open via /proc */
2150 .fasync = pfm_fasync,
2151 .release = pfm_close,
2156 pfmfs_delete_dentry(struct dentry *dentry)
2161 static struct dentry_operations pfmfs_dentry_operations = {
2162 .d_delete = pfmfs_delete_dentry,
2167 pfm_alloc_fd(struct file **cfile)
2170 struct file *file = NULL;
2171 struct inode * inode;
2175 fd = get_unused_fd();
2176 if (fd < 0) return -ENFILE;
2180 file = get_empty_filp();
2181 if (!file) goto out;
2184 * allocate a new inode
2186 inode = new_inode(pfmfs_mnt->mnt_sb);
2187 if (!inode) goto out;
2189 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2191 inode->i_sb = pfmfs_mnt->mnt_sb;
2192 inode->i_mode = S_IFCHR|S_IRUGO;
2194 inode->i_uid = current->fsuid;
2195 inode->i_gid = current->fsgid;
2197 sprintf(name, "[%lu]", inode->i_ino);
2199 this.len = strlen(name);
2200 this.hash = inode->i_ino;
2205 * allocate a new dcache entry
2207 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2208 if (!file->f_dentry) goto out;
2210 file->f_dentry->d_op = &pfmfs_dentry_operations;
2212 d_add(file->f_dentry, inode);
2213 file->f_vfsmnt = mntget(pfmfs_mnt);
2214 file->f_mapping = inode->i_mapping;
2216 file->f_op = &pfm_file_ops;
2217 file->f_mode = FMODE_READ;
2218 file->f_flags = O_RDONLY;
2222 * may have to delay until context is attached?
2224 fd_install(fd, file);
2227 * the file structure we will use
2233 if (file) put_filp(file);
2239 pfm_free_fd(int fd, struct file *file)
2241 if (file) put_filp(file);
2246 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2250 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2253 page = pfm_kvirt_to_pa(buf);
2255 if (pfm_remap_page_range(vma, addr, page, PAGE_SIZE, PAGE_READONLY)) return -ENOMEM;
2265 * allocate a sampling buffer and remaps it into the user address space of the task
2268 pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2270 struct mm_struct *mm = task->mm;
2271 struct vm_area_struct *vma = NULL;
2277 * the fixed header + requested size and align to page boundary
2279 size = PAGE_ALIGN(rsize);
2281 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2284 * check requested size to avoid Denial-of-service attacks
2285 * XXX: may have to refine this test
2286 * Check against address space limit.
2288 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2291 if (size > task->rlim[RLIMIT_MEMLOCK].rlim_cur) return -EAGAIN;
2294 * We do the easy to undo allocations first.
2296 * pfm_rvmalloc(), clears the buffer, so there is no leak
2298 smpl_buf = pfm_rvmalloc(size);
2299 if (smpl_buf == NULL) {
2300 DPRINT(("Can't allocate sampling buffer\n"));
2304 DPRINT(("smpl_buf @%p\n", smpl_buf));
2307 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2309 DPRINT(("Cannot allocate vma\n"));
2312 memset(vma, 0, sizeof(*vma));
2315 * partially initialize the vma for the sampling buffer
2317 * The VM_DONTCOPY flag is very important as it ensures that the mapping
2318 * will never be inherited for any child process (via fork()) which is always
2322 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2323 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2326 * Now we have everything we need and we can initialize
2327 * and connect all the data structures
2330 ctx->ctx_smpl_hdr = smpl_buf;
2331 ctx->ctx_smpl_size = size; /* aligned size */
2334 * Let's do the difficult operations next.
2336 * now we atomically find some area in the address space and
2337 * remap the buffer in it.
2339 down_write(&task->mm->mmap_sem);
2341 /* find some free area in address space, must have mmap sem held */
2342 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2343 if (vma->vm_start == 0UL) {
2344 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2345 up_write(&task->mm->mmap_sem);
2348 vma->vm_end = vma->vm_start + size;
2350 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2352 /* can only be applied to current task, need to have the mm semaphore held when called */
2353 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2354 DPRINT(("Can't remap buffer\n"));
2355 up_write(&task->mm->mmap_sem);
2360 * now insert the vma in the vm list for the process, must be
2361 * done with mmap lock held
2363 insert_vm_struct(mm, vma);
2365 // mm->total_vm += size >> PAGE_SHIFT;
2366 vx_vmpages_add(mm, size >> PAGE_SHIFT);
2368 up_write(&task->mm->mmap_sem);
2371 * keep track of user level virtual address
2373 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2374 *(unsigned long *)user_vaddr = vma->vm_start;
2379 kmem_cache_free(vm_area_cachep, vma);
2381 pfm_rvfree(smpl_buf, size);
2387 * XXX: do something better here
2390 pfm_bad_permissions(struct task_struct *task)
2392 /* inspired by ptrace_attach() */
2393 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2402 return ((current->uid != task->euid)
2403 || (current->uid != task->suid)
2404 || (current->uid != task->uid)
2405 || (current->gid != task->egid)
2406 || (current->gid != task->sgid)
2407 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2411 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2417 ctx_flags = pfx->ctx_flags;
2419 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2422 * cannot block in this mode
2424 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2425 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2430 /* probably more to add here */
2436 pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2437 unsigned int cpu, pfarg_context_t *arg)
2439 pfm_buffer_fmt_t *fmt = NULL;
2440 unsigned long size = 0UL;
2442 void *fmt_arg = NULL;
2444 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2446 /* invoke and lock buffer format, if found */
2447 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2449 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2454 * buffer argument MUST be contiguous to pfarg_context_t
2456 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2458 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2460 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2462 if (ret) goto error;
2464 /* link buffer format and context */
2465 ctx->ctx_buf_fmt = fmt;
2468 * check if buffer format wants to use perfmon buffer allocation/mapping service
2470 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2471 if (ret) goto error;
2475 * buffer is always remapped into the caller's address space
2477 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2478 if (ret) goto error;
2480 /* keep track of user address of buffer */
2481 arg->ctx_smpl_vaddr = uaddr;
2483 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2490 pfm_reset_pmu_state(pfm_context_t *ctx)
2495 * install reset values for PMC.
2497 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2498 if (PMC_IS_IMPL(i) == 0) continue;
2499 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2500 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2503 * PMD registers are set to 0UL when the context in memset()
2507 * On context switched restore, we must restore ALL pmc and ALL pmd even
2508 * when they are not actively used by the task. In UP, the incoming process
2509 * may otherwise pick up left over PMC, PMD state from the previous process.
2510 * As opposed to PMD, stale PMC can cause harm to the incoming
2511 * process because they may change what is being measured.
2512 * Therefore, we must systematically reinstall the entire
2513 * PMC state. In SMP, the same thing is possible on the
2514 * same CPU but also on between 2 CPUs.
2516 * The problem with PMD is information leaking especially
2517 * to user level when psr.sp=0
2519 * There is unfortunately no easy way to avoid this problem
2520 * on either UP or SMP. This definitively slows down the
2521 * pfm_load_regs() function.
2525 * bitmask of all PMCs accessible to this context
2527 * PMC0 is treated differently.
2529 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2532 * bitmask of all PMDs that are accesible to this context
2534 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2536 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2539 * useful in case of re-enable after disable
2541 ctx->ctx_used_ibrs[0] = 0UL;
2542 ctx->ctx_used_dbrs[0] = 0UL;
2546 pfm_ctx_getsize(void *arg, size_t *sz)
2548 pfarg_context_t *req = (pfarg_context_t *)arg;
2549 pfm_buffer_fmt_t *fmt;
2553 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2555 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2557 DPRINT(("cannot find buffer format\n"));
2560 /* get just enough to copy in user parameters */
2561 *sz = fmt->fmt_arg_size;
2562 DPRINT(("arg_size=%lu\n", *sz));
2570 * cannot attach if :
2572 * - task not owned by caller
2573 * - task incompatible with context mode
2576 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2579 * no kernel task or task not owner by caller
2581 if (task->mm == NULL) {
2582 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2585 if (pfm_bad_permissions(task)) {
2586 DPRINT(("no permission to attach to [%d]\n", task->pid));
2590 * cannot block in self-monitoring mode
2592 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2593 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2597 if (task->state == TASK_ZOMBIE) {
2598 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2603 * always ok for self
2605 if (task == current) return 0;
2607 if (task->state != TASK_STOPPED) {
2608 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2612 * make sure the task is off any CPU
2614 wait_task_inactive(task);
2616 /* more to come... */
2622 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2624 struct task_struct *p = current;
2627 /* XXX: need to add more checks here */
2628 if (pid < 2) return -EPERM;
2630 if (pid != current->pid) {
2632 read_lock(&tasklist_lock);
2634 p = find_task_by_pid(pid);
2636 /* make sure task cannot go away while we operate on it */
2637 if (p) get_task_struct(p);
2639 read_unlock(&tasklist_lock);
2641 if (p == NULL) return -ESRCH;
2644 ret = pfm_task_incompatible(ctx, p);
2647 } else if (p != current) {
2656 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2658 pfarg_context_t *req = (pfarg_context_t *)arg;
2663 /* let's check the arguments first */
2664 ret = pfarg_is_sane(current, req);
2665 if (ret < 0) return ret;
2667 ctx_flags = req->ctx_flags;
2671 ctx = pfm_context_alloc();
2672 if (!ctx) goto error;
2674 req->ctx_fd = ctx->ctx_fd = pfm_alloc_fd(&filp);
2675 if (req->ctx_fd < 0) goto error_file;
2678 * attach context to file
2680 filp->private_data = ctx;
2683 * does the user want to sample?
2685 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2686 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2687 if (ret) goto buffer_error;
2691 * init context protection lock
2693 spin_lock_init(&ctx->ctx_lock);
2696 * context is unloaded
2698 ctx->ctx_state = PFM_CTX_UNLOADED;
2701 * initialization of context's flags
2703 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2704 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2705 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2706 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2708 * will move to set properties
2709 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2713 * init restart semaphore to locked
2715 sema_init(&ctx->ctx_restart_sem, 0);
2718 * activation is used in SMP only
2720 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2721 SET_LAST_CPU(ctx, -1);
2724 * initialize notification message queue
2726 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2727 init_waitqueue_head(&ctx->ctx_msgq_wait);
2728 init_waitqueue_head(&ctx->ctx_zombieq);
2730 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2735 ctx->ctx_fl_excl_idle,
2740 * initialize soft PMU state
2742 pfm_reset_pmu_state(ctx);
2747 pfm_free_fd(ctx->ctx_fd, filp);
2749 if (ctx->ctx_buf_fmt) {
2750 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2753 pfm_context_free(ctx);
2759 static inline unsigned long
2760 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2762 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2763 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2764 extern unsigned long carta_random32 (unsigned long seed);
2766 if (reg->flags & PFM_REGFL_RANDOM) {
2767 new_seed = carta_random32(old_seed);
2768 val -= (old_seed & mask); /* counter values are negative numbers! */
2769 if ((mask >> 32) != 0)
2770 /* construct a full 64-bit random value: */
2771 new_seed |= carta_random32(old_seed >> 32) << 32;
2772 reg->seed = new_seed;
2779 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2781 unsigned long mask = ovfl_regs[0];
2782 unsigned long reset_others = 0UL;
2787 * now restore reset value on sampling overflowed counters
2789 mask >>= PMU_FIRST_COUNTER;
2790 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2792 if ((mask & 0x1UL) == 0UL) continue;
2794 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2795 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2797 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2801 * Now take care of resetting the other registers
2803 for(i = 0; reset_others; i++, reset_others >>= 1) {
2805 if ((reset_others & 0x1) == 0) continue;
2807 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2809 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2810 is_long_reset ? "long" : "short", i, val));
2815 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2817 unsigned long mask = ovfl_regs[0];
2818 unsigned long reset_others = 0UL;
2822 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2824 if (ctx->ctx_state == PFM_CTX_MASKED) {
2825 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2830 * now restore reset value on sampling overflowed counters
2832 mask >>= PMU_FIRST_COUNTER;
2833 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2835 if ((mask & 0x1UL) == 0UL) continue;
2837 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2838 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2840 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2842 pfm_write_soft_counter(ctx, i, val);
2846 * Now take care of resetting the other registers
2848 for(i = 0; reset_others; i++, reset_others >>= 1) {
2850 if ((reset_others & 0x1) == 0) continue;
2852 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2854 if (PMD_IS_COUNTING(i)) {
2855 pfm_write_soft_counter(ctx, i, val);
2857 ia64_set_pmd(i, val);
2859 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2860 is_long_reset ? "long" : "short", i, val));
2866 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2868 struct thread_struct *thread = NULL;
2869 struct task_struct *task;
2870 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2871 unsigned long value, pmc_pm;
2872 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2873 unsigned int cnum, reg_flags, flags, pmc_type;
2874 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2875 int is_monitor, is_counting, state;
2877 pfm_reg_check_t wr_func;
2878 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2880 state = ctx->ctx_state;
2881 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2882 is_system = ctx->ctx_fl_system;
2883 task = ctx->ctx_task;
2884 impl_pmds = pmu_conf->impl_pmds[0];
2886 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2889 thread = &task->thread;
2891 * In system wide and when the context is loaded, access can only happen
2892 * when the caller is running on the CPU being monitored by the session.
2893 * It does not have to be the owner (ctx_task) of the context per se.
2895 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2896 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2899 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2901 expert_mode = pfm_sysctl.expert_mode;
2903 for (i = 0; i < count; i++, req++) {
2905 cnum = req->reg_num;
2906 reg_flags = req->reg_flags;
2907 value = req->reg_value;
2908 smpl_pmds = req->reg_smpl_pmds[0];
2909 reset_pmds = req->reg_reset_pmds[0];
2913 if (cnum >= PMU_MAX_PMCS) {
2914 DPRINT(("pmc%u is invalid\n", cnum));
2918 pmc_type = pmu_conf->pmc_desc[cnum].type;
2919 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2920 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2921 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2924 * we reject all non implemented PMC as well
2925 * as attempts to modify PMC[0-3] which are used
2926 * as status registers by the PMU
2928 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2929 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2932 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2934 * If the PMC is a monitor, then if the value is not the default:
2935 * - system-wide session: PMCx.pm=1 (privileged monitor)
2936 * - per-task : PMCx.pm=0 (user monitor)
2938 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2939 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2948 * enforce generation of overflow interrupt. Necessary on all
2951 value |= 1 << PMU_PMC_OI;
2953 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2954 flags |= PFM_REGFL_OVFL_NOTIFY;
2957 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2959 /* verify validity of smpl_pmds */
2960 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2961 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2965 /* verify validity of reset_pmds */
2966 if ((reset_pmds & impl_pmds) != reset_pmds) {
2967 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2971 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2972 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2975 /* eventid on non-counting monitors are ignored */
2979 * execute write checker, if any
2981 if (likely(expert_mode == 0 && wr_func)) {
2982 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2983 if (ret) goto error;
2988 * no error on this register
2990 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2993 * Now we commit the changes to the software state
2997 * update overflow information
3001 * full flag update each time a register is programmed
3003 ctx->ctx_pmds[cnum].flags = flags;
3005 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
3006 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
3007 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
3010 * Mark all PMDS to be accessed as used.
3012 * We do not keep track of PMC because we have to
3013 * systematically restore ALL of them.
3015 * We do not update the used_monitors mask, because
3016 * if we have not programmed them, then will be in
3017 * a quiescent state, therefore we will not need to
3018 * mask/restore then when context is MASKED.
3020 CTX_USED_PMD(ctx, reset_pmds);
3021 CTX_USED_PMD(ctx, smpl_pmds);
3023 * make sure we do not try to reset on
3024 * restart because we have established new values
3026 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3029 * Needed in case the user does not initialize the equivalent
3030 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3031 * possible leak here.
3033 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3036 * keep track of the monitor PMC that we are using.
3037 * we save the value of the pmc in ctx_pmcs[] and if
3038 * the monitoring is not stopped for the context we also
3039 * place it in the saved state area so that it will be
3040 * picked up later by the context switch code.
3042 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3044 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3045 * monitoring needs to be stopped.
3047 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3050 * update context state
3052 ctx->ctx_pmcs[cnum] = value;
3056 * write thread state
3058 if (is_system == 0) thread->pmcs[cnum] = value;
3061 * write hardware register if we can
3063 if (can_access_pmu) {
3064 ia64_set_pmc(cnum, value);
3069 * per-task SMP only here
3071 * we are guaranteed that the task is not running on the other CPU,
3072 * we indicate that this PMD will need to be reloaded if the task
3073 * is rescheduled on the CPU it ran last on.
3075 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3080 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",
3085 ctx->ctx_all_pmcs[0],
3086 ctx->ctx_used_pmds[0],
3087 ctx->ctx_pmds[cnum].eventid,
3090 ctx->ctx_reload_pmcs[0],
3091 ctx->ctx_used_monitors[0],
3092 ctx->ctx_ovfl_regs[0]));
3096 * make sure the changes are visible
3098 if (can_access_pmu) ia64_srlz_d();
3102 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3107 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3109 struct thread_struct *thread = NULL;
3110 struct task_struct *task;
3111 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3112 unsigned long value, hw_value, ovfl_mask;
3114 int i, can_access_pmu = 0, state;
3115 int is_counting, is_loaded, is_system, expert_mode;
3117 pfm_reg_check_t wr_func;
3120 state = ctx->ctx_state;
3121 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3122 is_system = ctx->ctx_fl_system;
3123 ovfl_mask = pmu_conf->ovfl_val;
3124 task = ctx->ctx_task;
3126 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3129 * on both UP and SMP, we can only write to the PMC when the task is
3130 * the owner of the local PMU.
3132 if (likely(is_loaded)) {
3133 thread = &task->thread;
3135 * In system wide and when the context is loaded, access can only happen
3136 * when the caller is running on the CPU being monitored by the session.
3137 * It does not have to be the owner (ctx_task) of the context per se.
3139 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3140 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3143 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3145 expert_mode = pfm_sysctl.expert_mode;
3147 for (i = 0; i < count; i++, req++) {
3149 cnum = req->reg_num;
3150 value = req->reg_value;
3152 if (!PMD_IS_IMPL(cnum)) {
3153 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3156 is_counting = PMD_IS_COUNTING(cnum);
3157 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3160 * execute write checker, if any
3162 if (unlikely(expert_mode == 0 && wr_func)) {
3163 unsigned long v = value;
3165 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3166 if (ret) goto abort_mission;
3173 * no error on this register
3175 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3178 * now commit changes to software state
3183 * update virtualized (64bits) counter
3187 * write context state
3189 ctx->ctx_pmds[cnum].lval = value;
3192 * when context is load we use the split value
3195 hw_value = value & ovfl_mask;
3196 value = value & ~ovfl_mask;
3200 * update reset values (not just for counters)
3202 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3203 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3206 * update randomization parameters (not just for counters)
3208 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3209 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3212 * update context value
3214 ctx->ctx_pmds[cnum].val = value;
3217 * Keep track of what we use
3219 * We do not keep track of PMC because we have to
3220 * systematically restore ALL of them.
3222 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3225 * mark this PMD register used as well
3227 CTX_USED_PMD(ctx, RDEP(cnum));
3230 * make sure we do not try to reset on
3231 * restart because we have established new values
3233 if (is_counting && state == PFM_CTX_MASKED) {
3234 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3239 * write thread state
3241 if (is_system == 0) thread->pmds[cnum] = hw_value;
3244 * write hardware register if we can
3246 if (can_access_pmu) {
3247 ia64_set_pmd(cnum, hw_value);
3251 * we are guaranteed that the task is not running on the other CPU,
3252 * we indicate that this PMD will need to be reloaded if the task
3253 * is rescheduled on the CPU it ran last on.
3255 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3260 DPRINT(("pmd[%u]=0x%lx loaded=%d access_pmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3261 "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",
3267 ctx->ctx_pmds[cnum].val,
3268 ctx->ctx_pmds[cnum].short_reset,
3269 ctx->ctx_pmds[cnum].long_reset,
3270 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3271 ctx->ctx_used_pmds[0],
3272 ctx->ctx_pmds[cnum].reset_pmds[0],
3273 ctx->ctx_reload_pmds[0],
3274 ctx->ctx_all_pmds[0],
3275 ctx->ctx_ovfl_regs[0]));
3279 * make changes visible
3281 if (can_access_pmu) ia64_srlz_d();
3287 * for now, we have only one possibility for error
3289 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3294 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3295 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3296 * interrupt is delivered during the call, it will be kept pending until we leave, making
3297 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3298 * guaranteed to return consistent data to the user, it may simply be old. It is not
3299 * trivial to treat the overflow while inside the call because you may end up in
3300 * some module sampling buffer code causing deadlocks.
3303 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3305 struct thread_struct *thread = NULL;
3306 struct task_struct *task;
3307 unsigned long val = 0UL, lval, ovfl_mask, sval;
3308 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3309 unsigned int cnum, reg_flags = 0;
3310 int i, can_access_pmu = 0, state;
3311 int is_loaded, is_system, is_counting, expert_mode;
3313 pfm_reg_check_t rd_func;
3316 * access is possible when loaded only for
3317 * self-monitoring tasks or in UP mode
3320 state = ctx->ctx_state;
3321 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3322 is_system = ctx->ctx_fl_system;
3323 ovfl_mask = pmu_conf->ovfl_val;
3324 task = ctx->ctx_task;
3326 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3328 if (likely(is_loaded)) {
3329 thread = &task->thread;
3331 * In system wide and when the context is loaded, access can only happen
3332 * when the caller is running on the CPU being monitored by the session.
3333 * It does not have to be the owner (ctx_task) of the context per se.
3335 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3336 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3340 * this can be true when not self-monitoring only in UP
3342 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3344 if (can_access_pmu) ia64_srlz_d();
3346 expert_mode = pfm_sysctl.expert_mode;
3348 DPRINT(("loaded=%d access_pmu=%d ctx_state=%d\n",
3354 * on both UP and SMP, we can only read the PMD from the hardware register when
3355 * the task is the owner of the local PMU.
3358 for (i = 0; i < count; i++, req++) {
3360 cnum = req->reg_num;
3361 reg_flags = req->reg_flags;
3363 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3365 * we can only read the register that we use. That includes
3366 * the one we explicitely initialize AND the one we want included
3367 * in the sampling buffer (smpl_regs).
3369 * Having this restriction allows optimization in the ctxsw routine
3370 * without compromising security (leaks)
3372 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3374 sval = ctx->ctx_pmds[cnum].val;
3375 lval = ctx->ctx_pmds[cnum].lval;
3376 is_counting = PMD_IS_COUNTING(cnum);
3379 * If the task is not the current one, then we check if the
3380 * PMU state is still in the local live register due to lazy ctxsw.
3381 * If true, then we read directly from the registers.
3383 if (can_access_pmu){
3384 val = ia64_get_pmd(cnum);
3387 * context has been saved
3388 * if context is zombie, then task does not exist anymore.
3389 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3391 val = is_loaded ? thread->pmds[cnum] : 0UL;
3393 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3397 * XXX: need to check for overflow when loaded
3404 * execute read checker, if any
3406 if (unlikely(expert_mode == 0 && rd_func)) {
3407 unsigned long v = val;
3408 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3409 if (ret) goto error;
3414 PFM_REG_RETFLAG_SET(reg_flags, 0);
3416 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3419 * update register return value, abort all if problem during copy.
3420 * we only modify the reg_flags field. no check mode is fine because
3421 * access has been verified upfront in sys_perfmonctl().
3423 req->reg_value = val;
3424 req->reg_flags = reg_flags;
3425 req->reg_last_reset_val = lval;
3431 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3436 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3440 if (req == NULL) return -EINVAL;
3442 ctx = GET_PMU_CTX();
3444 if (ctx == NULL) return -EINVAL;
3447 * for now limit to current task, which is enough when calling
3448 * from overflow handler
3450 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3452 return pfm_write_pmcs(ctx, req, nreq, regs);
3454 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3457 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3461 if (req == NULL) return -EINVAL;
3463 ctx = GET_PMU_CTX();
3465 if (ctx == NULL) return -EINVAL;
3468 * for now limit to current task, which is enough when calling
3469 * from overflow handler
3471 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3473 return pfm_read_pmds(ctx, req, nreq, regs);
3475 EXPORT_SYMBOL(pfm_mod_read_pmds);
3478 * Only call this function when a process it trying to
3479 * write the debug registers (reading is always allowed)
3482 pfm_use_debug_registers(struct task_struct *task)
3484 pfm_context_t *ctx = task->thread.pfm_context;
3485 unsigned long flags;
3488 if (pmu_conf->use_rr_dbregs == 0) return 0;
3490 DPRINT(("called for [%d]\n", task->pid));
3495 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3498 * Even on SMP, we do not need to use an atomic here because
3499 * the only way in is via ptrace() and this is possible only when the
3500 * process is stopped. Even in the case where the ctxsw out is not totally
3501 * completed by the time we come here, there is no way the 'stopped' process
3502 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3503 * So this is always safe.
3505 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3510 * We cannot allow setting breakpoints when system wide monitoring
3511 * sessions are using the debug registers.
3513 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3516 pfm_sessions.pfs_ptrace_use_dbregs++;
3518 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3519 pfm_sessions.pfs_ptrace_use_dbregs,
3520 pfm_sessions.pfs_sys_use_dbregs,
3529 * This function is called for every task that exits with the
3530 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3531 * able to use the debug registers for debugging purposes via
3532 * ptrace(). Therefore we know it was not using them for
3533 * perfmormance monitoring, so we only decrement the number
3534 * of "ptraced" debug register users to keep the count up to date
3537 pfm_release_debug_registers(struct task_struct *task)
3539 unsigned long flags;
3542 if (pmu_conf->use_rr_dbregs == 0) return 0;
3545 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3546 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3549 pfm_sessions.pfs_ptrace_use_dbregs--;
3558 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3560 struct task_struct *task;
3561 pfm_buffer_fmt_t *fmt;
3562 pfm_ovfl_ctrl_t rst_ctrl;
3563 int state, is_system;
3566 state = ctx->ctx_state;
3567 fmt = ctx->ctx_buf_fmt;
3568 is_system = ctx->ctx_fl_system;
3569 task = PFM_CTX_TASK(ctx);
3572 case PFM_CTX_MASKED:
3574 case PFM_CTX_LOADED:
3575 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3577 case PFM_CTX_UNLOADED:
3578 case PFM_CTX_ZOMBIE:
3579 DPRINT(("invalid state=%d\n", state));
3582 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3587 * In system wide and when the context is loaded, access can only happen
3588 * when the caller is running on the CPU being monitored by the session.
3589 * It does not have to be the owner (ctx_task) of the context per se.
3591 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3592 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3597 if (unlikely(task == NULL)) {
3598 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3602 if (task == current || is_system) {
3604 fmt = ctx->ctx_buf_fmt;
3606 DPRINT(("restarting self %d ovfl=0x%lx\n",
3608 ctx->ctx_ovfl_regs[0]));
3610 if (CTX_HAS_SMPL(ctx)) {
3612 prefetch(ctx->ctx_smpl_hdr);
3614 rst_ctrl.bits.mask_monitoring = 0;
3615 rst_ctrl.bits.reset_ovfl_pmds = 0;
3617 if (state == PFM_CTX_LOADED)
3618 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3620 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3622 rst_ctrl.bits.mask_monitoring = 0;
3623 rst_ctrl.bits.reset_ovfl_pmds = 1;
3627 if (rst_ctrl.bits.reset_ovfl_pmds)
3628 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3630 if (rst_ctrl.bits.mask_monitoring == 0) {
3631 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3633 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3635 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3637 // cannot use pfm_stop_monitoring(task, regs);
3641 * clear overflowed PMD mask to remove any stale information
3643 ctx->ctx_ovfl_regs[0] = 0UL;
3646 * back to LOADED state
3648 ctx->ctx_state = PFM_CTX_LOADED;
3651 * XXX: not really useful for self monitoring
3653 ctx->ctx_fl_can_restart = 0;
3659 * restart another task
3663 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3664 * one is seen by the task.
3666 if (state == PFM_CTX_MASKED) {
3667 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3669 * will prevent subsequent restart before this one is
3670 * seen by other task
3672 ctx->ctx_fl_can_restart = 0;
3676 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3677 * the task is blocked or on its way to block. That's the normal
3678 * restart path. If the monitoring is not masked, then the task
3679 * can be actively monitoring and we cannot directly intervene.
3680 * Therefore we use the trap mechanism to catch the task and
3681 * force it to reset the buffer/reset PMDs.
3683 * if non-blocking, then we ensure that the task will go into
3684 * pfm_handle_work() before returning to user mode.
3686 * We cannot explicitely reset another task, it MUST always
3687 * be done by the task itself. This works for system wide because
3688 * the tool that is controlling the session is logically doing
3689 * "self-monitoring".
3691 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3692 DPRINT(("unblocking [%d] \n", task->pid));
3693 up(&ctx->ctx_restart_sem);
3695 DPRINT(("[%d] armed exit trap\n", task->pid));
3697 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3699 PFM_SET_WORK_PENDING(task, 1);
3701 pfm_set_task_notify(task);
3704 * XXX: send reschedule if task runs on another CPU
3711 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3713 unsigned int m = *(unsigned int *)arg;
3715 pfm_sysctl.debug = m == 0 ? 0 : 1;
3717 pfm_debug_var = pfm_sysctl.debug;
3719 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3722 memset(pfm_stats, 0, sizeof(pfm_stats));
3723 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3729 * arg can be NULL and count can be zero for this function
3732 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3734 struct thread_struct *thread = NULL;
3735 struct task_struct *task;
3736 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3737 unsigned long flags;
3742 int i, can_access_pmu = 0;
3743 int is_system, is_loaded;
3745 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3747 state = ctx->ctx_state;
3748 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3749 is_system = ctx->ctx_fl_system;
3750 task = ctx->ctx_task;
3752 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3755 * on both UP and SMP, we can only write to the PMC when the task is
3756 * the owner of the local PMU.
3759 thread = &task->thread;
3761 * In system wide and when the context is loaded, access can only happen
3762 * when the caller is running on the CPU being monitored by the session.
3763 * It does not have to be the owner (ctx_task) of the context per se.
3765 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3766 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3769 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3773 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3774 * ensuring that no real breakpoint can be installed via this call.
3776 * IMPORTANT: regs can be NULL in this function
3779 first_time = ctx->ctx_fl_using_dbreg == 0;
3782 * don't bother if we are loaded and task is being debugged
3784 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3785 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3790 * check for debug registers in system wide mode
3792 * If though a check is done in pfm_context_load(),
3793 * we must repeat it here, in case the registers are
3794 * written after the context is loaded
3799 if (first_time && is_system) {
3800 if (pfm_sessions.pfs_ptrace_use_dbregs)
3803 pfm_sessions.pfs_sys_use_dbregs++;
3808 if (ret != 0) return ret;
3811 * mark ourself as user of the debug registers for
3814 ctx->ctx_fl_using_dbreg = 1;
3817 * clear hardware registers to make sure we don't
3818 * pick up stale state.
3820 * for a system wide session, we do not use
3821 * thread.dbr, thread.ibr because this process
3822 * never leaves the current CPU and the state
3823 * is shared by all processes running on it
3825 if (first_time && can_access_pmu) {
3826 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3827 for (i=0; i < pmu_conf->num_ibrs; i++) {
3828 ia64_set_ibr(i, 0UL);
3829 ia64_dv_serialize_instruction();
3832 for (i=0; i < pmu_conf->num_dbrs; i++) {
3833 ia64_set_dbr(i, 0UL);
3834 ia64_dv_serialize_data();
3840 * Now install the values into the registers
3842 for (i = 0; i < count; i++, req++) {
3844 rnum = req->dbreg_num;
3845 dbreg.val = req->dbreg_value;
3849 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3850 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3851 rnum, dbreg.val, mode, i, count));
3857 * make sure we do not install enabled breakpoint
3860 if (mode == PFM_CODE_RR)
3861 dbreg.ibr.ibr_x = 0;
3863 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3866 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3869 * Debug registers, just like PMC, can only be modified
3870 * by a kernel call. Moreover, perfmon() access to those
3871 * registers are centralized in this routine. The hardware
3872 * does not modify the value of these registers, therefore,
3873 * if we save them as they are written, we can avoid having
3874 * to save them on context switch out. This is made possible
3875 * by the fact that when perfmon uses debug registers, ptrace()
3876 * won't be able to modify them concurrently.
3878 if (mode == PFM_CODE_RR) {
3879 CTX_USED_IBR(ctx, rnum);
3881 if (can_access_pmu) {
3882 ia64_set_ibr(rnum, dbreg.val);
3883 ia64_dv_serialize_instruction();
3886 ctx->ctx_ibrs[rnum] = dbreg.val;
3888 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x is_loaded=%d access_pmu=%d\n",
3889 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3891 CTX_USED_DBR(ctx, rnum);
3893 if (can_access_pmu) {
3894 ia64_set_dbr(rnum, dbreg.val);
3895 ia64_dv_serialize_data();
3897 ctx->ctx_dbrs[rnum] = dbreg.val;
3899 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x is_loaded=%d access_pmu=%d\n",
3900 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3908 * in case it was our first attempt, we undo the global modifications
3912 if (ctx->ctx_fl_system) {
3913 pfm_sessions.pfs_sys_use_dbregs--;
3916 ctx->ctx_fl_using_dbreg = 0;
3919 * install error return flag
3921 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3927 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3929 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3933 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3935 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3939 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3943 if (req == NULL) return -EINVAL;
3945 ctx = GET_PMU_CTX();
3947 if (ctx == NULL) return -EINVAL;
3950 * for now limit to current task, which is enough when calling
3951 * from overflow handler
3953 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3955 return pfm_write_ibrs(ctx, req, nreq, regs);
3957 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3960 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3964 if (req == NULL) return -EINVAL;
3966 ctx = GET_PMU_CTX();
3968 if (ctx == NULL) return -EINVAL;
3971 * for now limit to current task, which is enough when calling
3972 * from overflow handler
3974 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3976 return pfm_write_dbrs(ctx, req, nreq, regs);
3978 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3982 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3984 pfarg_features_t *req = (pfarg_features_t *)arg;
3986 req->ft_version = PFM_VERSION;
3991 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3993 struct pt_regs *tregs;
3994 struct task_struct *task = PFM_CTX_TASK(ctx);
3995 int state, is_system;
3997 state = ctx->ctx_state;
3998 is_system = ctx->ctx_fl_system;
4000 if (state != PFM_CTX_LOADED && state != PFM_CTX_MASKED) return -EINVAL;
4003 * In system wide and when the context is loaded, access can only happen
4004 * when the caller is running on the CPU being monitored by the session.
4005 * It does not have to be the owner (ctx_task) of the context per se.
4007 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4008 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4011 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4012 PFM_CTX_TASK(ctx)->pid,
4016 * in system mode, we need to update the PMU directly
4017 * and the user level state of the caller, which may not
4018 * necessarily be the creator of the context.
4022 * Update local PMU first
4026 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4030 * update local cpuinfo
4032 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4035 * stop monitoring, does srlz.i
4040 * stop monitoring in the caller
4042 ia64_psr(regs)->pp = 0;
4050 if (task == current) {
4051 /* stop monitoring at kernel level */
4055 * stop monitoring at the user level
4057 ia64_psr(regs)->up = 0;
4059 tregs = ia64_task_regs(task);
4062 * stop monitoring at the user level
4064 ia64_psr(tregs)->up = 0;
4067 * monitoring disabled in kernel at next reschedule
4069 ctx->ctx_saved_psr_up = 0;
4070 DPRINT(("task=[%d]\n", task->pid));
4077 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4079 struct pt_regs *tregs;
4080 int state, is_system;
4082 state = ctx->ctx_state;
4083 is_system = ctx->ctx_fl_system;
4085 if (state != PFM_CTX_LOADED) return -EINVAL;
4088 * In system wide and when the context is loaded, access can only happen
4089 * when the caller is running on the CPU being monitored by the session.
4090 * It does not have to be the owner (ctx_task) of the context per se.
4092 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4093 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4098 * in system mode, we need to update the PMU directly
4099 * and the user level state of the caller, which may not
4100 * necessarily be the creator of the context.
4105 * set user level psr.pp for the caller
4107 ia64_psr(regs)->pp = 1;
4110 * now update the local PMU and cpuinfo
4112 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4115 * start monitoring at kernel level
4120 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4130 if (ctx->ctx_task == current) {
4132 /* start monitoring at kernel level */
4136 * activate monitoring at user level
4138 ia64_psr(regs)->up = 1;
4141 tregs = ia64_task_regs(ctx->ctx_task);
4144 * start monitoring at the kernel level the next
4145 * time the task is scheduled
4147 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4150 * activate monitoring at user level
4152 ia64_psr(tregs)->up = 1;
4158 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4160 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4165 for (i = 0; i < count; i++, req++) {
4167 cnum = req->reg_num;
4169 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4171 req->reg_value = PMC_DFL_VAL(cnum);
4173 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4175 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4180 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4185 pfm_check_task_exist(pfm_context_t *ctx)
4187 struct task_struct *g, *t;
4190 read_lock(&tasklist_lock);
4192 do_each_thread (g, t) {
4193 if (t->thread.pfm_context == ctx) {
4197 } while_each_thread (g, t);
4199 read_unlock(&tasklist_lock);
4201 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4207 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4209 struct task_struct *task;
4210 struct thread_struct *thread;
4211 struct pfm_context_t *old;
4212 unsigned long flags;
4214 struct task_struct *owner_task = NULL;
4216 pfarg_load_t *req = (pfarg_load_t *)arg;
4217 unsigned long *pmcs_source, *pmds_source;
4220 int state, is_system, set_dbregs = 0;
4222 state = ctx->ctx_state;
4223 is_system = ctx->ctx_fl_system;
4225 * can only load from unloaded or terminated state
4227 if (state != PFM_CTX_UNLOADED) {
4228 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4234 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4236 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4237 DPRINT(("cannot use blocking mode on self\n"));
4241 ret = pfm_get_task(ctx, req->load_pid, &task);
4243 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4250 * system wide is self monitoring only
4252 if (is_system && task != current) {
4253 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4258 thread = &task->thread;
4262 * cannot load a context which is using range restrictions,
4263 * into a task that is being debugged.
4265 if (ctx->ctx_fl_using_dbreg) {
4266 if (thread->flags & IA64_THREAD_DBG_VALID) {
4268 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4274 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4275 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4278 pfm_sessions.pfs_sys_use_dbregs++;
4279 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4286 if (ret) goto error;
4290 * SMP system-wide monitoring implies self-monitoring.
4292 * The programming model expects the task to
4293 * be pinned on a CPU throughout the session.
4294 * Here we take note of the current CPU at the
4295 * time the context is loaded. No call from
4296 * another CPU will be allowed.
4298 * The pinning via shed_setaffinity()
4299 * must be done by the calling task prior
4302 * systemwide: keep track of CPU this session is supposed to run on
4304 the_cpu = ctx->ctx_cpu = smp_processor_id();
4308 * now reserve the session
4310 ret = pfm_reserve_session(current, is_system, the_cpu);
4311 if (ret) goto error;
4314 * task is necessarily stopped at this point.
4316 * If the previous context was zombie, then it got removed in
4317 * pfm_save_regs(). Therefore we should not see it here.
4318 * If we see a context, then this is an active context
4320 * XXX: needs to be atomic
4322 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4323 thread->pfm_context, ctx));
4325 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4327 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4331 pfm_reset_msgq(ctx);
4333 ctx->ctx_state = PFM_CTX_LOADED;
4336 * link context to task
4338 ctx->ctx_task = task;
4342 * we load as stopped
4344 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4345 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4347 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4349 thread->flags |= IA64_THREAD_PM_VALID;
4353 * propagate into thread-state
4355 pfm_copy_pmds(task, ctx);
4356 pfm_copy_pmcs(task, ctx);
4358 pmcs_source = thread->pmcs;
4359 pmds_source = thread->pmds;
4362 * always the case for system-wide
4364 if (task == current) {
4366 if (is_system == 0) {
4368 /* allow user level control */
4369 ia64_psr(regs)->sp = 0;
4370 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4372 SET_LAST_CPU(ctx, smp_processor_id());
4374 SET_ACTIVATION(ctx);
4377 * push the other task out, if any
4379 owner_task = GET_PMU_OWNER();
4380 if (owner_task) pfm_lazy_save_regs(owner_task);
4384 * load all PMD from ctx to PMU (as opposed to thread state)
4385 * restore all PMC from ctx to PMU
4387 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4388 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4390 ctx->ctx_reload_pmcs[0] = 0UL;
4391 ctx->ctx_reload_pmds[0] = 0UL;
4394 * guaranteed safe by earlier check against DBG_VALID
4396 if (ctx->ctx_fl_using_dbreg) {
4397 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4398 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4403 SET_PMU_OWNER(task, ctx);
4405 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4408 * when not current, task MUST be stopped, so this is safe
4410 regs = ia64_task_regs(task);
4412 /* force a full reload */
4413 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4414 SET_LAST_CPU(ctx, -1);
4416 /* initial saved psr (stopped) */
4417 ctx->ctx_saved_psr_up = 0UL;
4418 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4424 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4427 * we must undo the dbregs setting (for system-wide)
4429 if (ret && set_dbregs) {
4431 pfm_sessions.pfs_sys_use_dbregs--;
4435 * release task, there is now a link with the context
4437 if (is_system == 0 && task != current) {
4441 ret = pfm_check_task_exist(ctx);
4443 ctx->ctx_state = PFM_CTX_UNLOADED;
4444 ctx->ctx_task = NULL;
4452 * in this function, we do not need to increase the use count
4453 * for the task via get_task_struct(), because we hold the
4454 * context lock. If the task were to disappear while having
4455 * a context attached, it would go through pfm_exit_thread()
4456 * which also grabs the context lock and would therefore be blocked
4457 * until we are here.
4459 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4462 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4464 struct task_struct *task = PFM_CTX_TASK(ctx);
4465 struct pt_regs *tregs;
4466 int prev_state, is_system;
4469 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4471 prev_state = ctx->ctx_state;
4472 is_system = ctx->ctx_fl_system;
4475 * unload only when necessary
4477 if (prev_state == PFM_CTX_UNLOADED) {
4478 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4483 * clear psr and dcr bits
4485 ret = pfm_stop(ctx, NULL, 0, regs);
4486 if (ret) return ret;
4488 ctx->ctx_state = PFM_CTX_UNLOADED;
4491 * in system mode, we need to update the PMU directly
4492 * and the user level state of the caller, which may not
4493 * necessarily be the creator of the context.
4500 * local PMU is taken care of in pfm_stop()
4502 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4503 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4506 * save PMDs in context
4509 pfm_flush_pmds(current, ctx);
4512 * at this point we are done with the PMU
4513 * so we can unreserve the resource.
4515 if (prev_state != PFM_CTX_ZOMBIE)
4516 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4519 * disconnect context from task
4521 task->thread.pfm_context = NULL;
4523 * disconnect task from context
4525 ctx->ctx_task = NULL;
4528 * There is nothing more to cleanup here.
4536 tregs = task == current ? regs : ia64_task_regs(task);
4538 if (task == current) {
4540 * cancel user level control
4542 ia64_psr(regs)->sp = 1;
4544 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4547 * save PMDs to context
4550 pfm_flush_pmds(task, ctx);
4553 * at this point we are done with the PMU
4554 * so we can unreserve the resource.
4556 * when state was ZOMBIE, we have already unreserved.
4558 if (prev_state != PFM_CTX_ZOMBIE)
4559 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4562 * reset activation counter and psr
4564 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4565 SET_LAST_CPU(ctx, -1);
4568 * PMU state will not be restored
4570 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4573 * break links between context and task
4575 task->thread.pfm_context = NULL;
4576 ctx->ctx_task = NULL;
4578 PFM_SET_WORK_PENDING(task, 0);
4580 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4581 ctx->ctx_fl_can_restart = 0;
4582 ctx->ctx_fl_going_zombie = 0;
4584 DPRINT(("disconnected [%d] from context\n", task->pid));
4591 * called only from exit_thread(): task == current
4592 * we come here only if current has a context attached (loaded or masked)
4595 pfm_exit_thread(struct task_struct *task)
4598 unsigned long flags;
4599 struct pt_regs *regs = ia64_task_regs(task);
4603 ctx = PFM_GET_CTX(task);
4605 PROTECT_CTX(ctx, flags);
4607 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4609 state = ctx->ctx_state;
4611 case PFM_CTX_UNLOADED:
4613 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4614 * be in unloaded state
4616 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4618 case PFM_CTX_LOADED:
4619 case PFM_CTX_MASKED:
4620 ret = pfm_context_unload(ctx, NULL, 0, regs);
4622 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4624 DPRINT(("ctx unloaded for current state was %d\n", state));
4626 pfm_end_notify_user(ctx);
4628 case PFM_CTX_ZOMBIE:
4629 ret = pfm_context_unload(ctx, NULL, 0, regs);
4631 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4636 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4639 UNPROTECT_CTX(ctx, flags);
4641 { u64 psr = pfm_get_psr();
4642 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4643 BUG_ON(GET_PMU_OWNER());
4644 BUG_ON(ia64_psr(regs)->up);
4645 BUG_ON(ia64_psr(regs)->pp);
4649 * All memory free operations (especially for vmalloc'ed memory)
4650 * MUST be done with interrupts ENABLED.
4652 if (free_ok) pfm_context_free(ctx);
4656 * functions MUST be listed in the increasing order of their index (see permfon.h)
4658 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4659 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4660 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4661 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4662 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4664 static pfm_cmd_desc_t pfm_cmd_tab[]={
4665 /* 0 */PFM_CMD_NONE,
4666 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4667 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4668 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4669 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4670 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4671 /* 6 */PFM_CMD_NONE,
4672 /* 7 */PFM_CMD_NONE,
4673 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4674 /* 9 */PFM_CMD_NONE,
4675 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4676 /* 11 */PFM_CMD_NONE,
4677 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4678 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4679 /* 14 */PFM_CMD_NONE,
4680 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4681 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4682 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4683 /* 18 */PFM_CMD_NONE,
4684 /* 19 */PFM_CMD_NONE,
4685 /* 20 */PFM_CMD_NONE,
4686 /* 21 */PFM_CMD_NONE,
4687 /* 22 */PFM_CMD_NONE,
4688 /* 23 */PFM_CMD_NONE,
4689 /* 24 */PFM_CMD_NONE,
4690 /* 25 */PFM_CMD_NONE,
4691 /* 26 */PFM_CMD_NONE,
4692 /* 27 */PFM_CMD_NONE,
4693 /* 28 */PFM_CMD_NONE,
4694 /* 29 */PFM_CMD_NONE,
4695 /* 30 */PFM_CMD_NONE,
4696 /* 31 */PFM_CMD_NONE,
4697 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4698 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4700 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4703 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4705 struct task_struct *task;
4706 int state, old_state;
4709 state = ctx->ctx_state;
4710 task = ctx->ctx_task;
4713 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4717 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4721 task->state, PFM_CMD_STOPPED(cmd)));
4724 * self-monitoring always ok.
4726 * for system-wide the caller can either be the creator of the
4727 * context (to one to which the context is attached to) OR
4728 * a task running on the same CPU as the session.
4730 if (task == current || ctx->ctx_fl_system) return 0;
4733 * no command can operate on a zombie context
4735 if (state == PFM_CTX_ZOMBIE) {
4736 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4741 * if context is UNLOADED, MASKED we are safe to go
4743 if (state != PFM_CTX_LOADED) return 0;
4746 * context is LOADED, we must make sure the task is stopped
4747 * We could lift this restriction for UP but it would mean that
4748 * the user has no guarantee the task would not run between
4749 * two successive calls to perfmonctl(). That's probably OK.
4750 * If this user wants to ensure the task does not run, then
4751 * the task must be stopped.
4753 if (PFM_CMD_STOPPED(cmd)) {
4754 if (task->state != TASK_STOPPED) {
4755 DPRINT(("[%d] task not in stopped state\n", task->pid));
4759 * task is now stopped, wait for ctxsw out
4761 * This is an interesting point in the code.
4762 * We need to unprotect the context because
4763 * the pfm_save_regs() routines needs to grab
4764 * the same lock. There are danger in doing
4765 * this because it leaves a window open for
4766 * another task to get access to the context
4767 * and possibly change its state. The one thing
4768 * that is not possible is for the context to disappear
4769 * because we are protected by the VFS layer, i.e.,
4770 * get_fd()/put_fd().
4774 UNPROTECT_CTX(ctx, flags);
4776 wait_task_inactive(task);
4778 PROTECT_CTX(ctx, flags);
4781 * we must recheck to verify if state has changed
4783 if (ctx->ctx_state != old_state) {
4784 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4792 * system-call entry point (must return long)
4795 sys_perfmonctl (int fd, int cmd, void *arg, int count, long arg5, long arg6, long arg7,
4796 long arg8, long stack)
4798 struct pt_regs *regs = (struct pt_regs *)&stack;
4799 struct file *file = NULL;
4800 pfm_context_t *ctx = NULL;
4801 unsigned long flags = 0UL;
4802 void *args_k = NULL;
4803 long ret; /* will expand int return types */
4804 size_t base_sz, sz, xtra_sz = 0;
4805 int narg, completed_args = 0, call_made = 0, cmd_flags;
4806 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4807 int (*getsize)(void *arg, size_t *sz);
4808 #define PFM_MAX_ARGSIZE 4096
4811 * reject any call if perfmon was disabled at initialization
4813 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4815 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4816 DPRINT(("invalid cmd=%d\n", cmd));
4820 func = pfm_cmd_tab[cmd].cmd_func;
4821 narg = pfm_cmd_tab[cmd].cmd_narg;
4822 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4823 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4824 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4826 if (unlikely(func == NULL)) {
4827 DPRINT(("invalid cmd=%d\n", cmd));
4831 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4839 * check if number of arguments matches what the command expects
4841 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4845 sz = xtra_sz + base_sz*count;
4847 * limit abuse to min page size
4849 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4850 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4855 * allocate default-sized argument buffer
4857 if (likely(count && args_k == NULL)) {
4858 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4859 if (args_k == NULL) return -ENOMEM;
4867 * assume sz = 0 for command without parameters
4869 if (sz && copy_from_user(args_k, arg, sz)) {
4870 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4875 * check if command supports extra parameters
4877 if (completed_args == 0 && getsize) {
4879 * get extra parameters size (based on main argument)
4881 ret = (*getsize)(args_k, &xtra_sz);
4882 if (ret) goto error_args;
4886 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4888 /* retry if necessary */
4889 if (likely(xtra_sz)) goto restart_args;
4892 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4897 if (unlikely(file == NULL)) {
4898 DPRINT(("invalid fd %d\n", fd));
4901 if (unlikely(PFM_IS_FILE(file) == 0)) {
4902 DPRINT(("fd %d not related to perfmon\n", fd));
4906 ctx = (pfm_context_t *)file->private_data;
4907 if (unlikely(ctx == NULL)) {
4908 DPRINT(("no context for fd %d\n", fd));
4911 prefetch(&ctx->ctx_state);
4913 PROTECT_CTX(ctx, flags);
4916 * check task is stopped
4918 ret = pfm_check_task_state(ctx, cmd, flags);
4919 if (unlikely(ret)) goto abort_locked;
4922 ret = (*func)(ctx, args_k, count, regs);
4928 DPRINT(("context unlocked\n"));
4929 UNPROTECT_CTX(ctx, flags);
4933 /* copy argument back to user, if needed */
4934 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4937 if (args_k) kfree(args_k);
4939 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4945 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4947 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4948 pfm_ovfl_ctrl_t rst_ctrl;
4952 state = ctx->ctx_state;
4954 * Unlock sampling buffer and reset index atomically
4955 * XXX: not really needed when blocking
4957 if (CTX_HAS_SMPL(ctx)) {
4959 rst_ctrl.bits.mask_monitoring = 0;
4960 rst_ctrl.bits.reset_ovfl_pmds = 0;
4962 if (state == PFM_CTX_LOADED)
4963 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4965 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4967 rst_ctrl.bits.mask_monitoring = 0;
4968 rst_ctrl.bits.reset_ovfl_pmds = 1;
4972 if (rst_ctrl.bits.reset_ovfl_pmds) {
4973 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4975 if (rst_ctrl.bits.mask_monitoring == 0) {
4976 DPRINT(("resuming monitoring\n"));
4977 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4979 DPRINT(("stopping monitoring\n"));
4980 //pfm_stop_monitoring(current, regs);
4982 ctx->ctx_state = PFM_CTX_LOADED;
4987 * context MUST BE LOCKED when calling
4988 * can only be called for current
4991 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4993 if (ctx->ctx_fl_system) {
4994 printk(KERN_ERR "perfmon: pfm_context_force_terminate [%d] is system-wide\n", current->pid);
4998 * we stop the whole thing, we do no need to flush
4999 * we know we WERE masked
5002 ia64_psr(regs)->up = 0;
5003 ia64_psr(regs)->sp = 1;
5006 * disconnect the task from the context and vice-versa
5008 current->thread.pfm_context = NULL;
5009 current->thread.flags &= ~IA64_THREAD_PM_VALID;
5010 ctx->ctx_task = NULL;
5012 DPRINT(("context terminated\n"));
5015 * and wakeup controlling task, indicating we are now disconnected
5017 wake_up_interruptible(&ctx->ctx_zombieq);
5020 * given that context is still locked, the controlling
5021 * task will only get access when we return from
5022 * pfm_handle_work().
5026 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5029 pfm_handle_work(void)
5032 struct pt_regs *regs;
5033 unsigned long flags;
5034 unsigned long ovfl_regs;
5035 unsigned int reason;
5038 ctx = PFM_GET_CTX(current);
5040 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5044 PROTECT_CTX(ctx, flags);
5046 PFM_SET_WORK_PENDING(current, 0);
5048 pfm_clear_task_notify();
5050 regs = ia64_task_regs(current);
5053 * extract reason for being here and clear
5055 reason = ctx->ctx_fl_trap_reason;
5056 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5057 ovfl_regs = ctx->ctx_ovfl_regs[0];
5059 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5062 * must be done before we check for simple-reset mode
5064 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5067 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5068 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5070 UNPROTECT_CTX(ctx, flags);
5072 DPRINT(("before block sleeping\n"));
5075 * may go through without blocking on SMP systems
5076 * if restart has been received already by the time we call down()
5078 ret = down_interruptible(&ctx->ctx_restart_sem);
5080 DPRINT(("after block sleeping ret=%d\n", ret));
5082 PROTECT_CTX(ctx, flags);
5085 * we need to read the ovfl_regs only after wake-up
5086 * because we may have had pfm_write_pmds() in between
5087 * and that can changed PMD values and therefore
5088 * ovfl_regs is reset for these new PMD values.
5090 ovfl_regs = ctx->ctx_ovfl_regs[0];
5092 if (ctx->ctx_fl_going_zombie) {
5094 DPRINT(("context is zombie, bailing out\n"));
5095 pfm_context_force_terminate(ctx, regs);
5099 * in case of interruption of down() we don't restart anything
5101 if (ret < 0) goto nothing_to_do;
5104 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5105 ctx->ctx_ovfl_regs[0] = 0UL;
5109 UNPROTECT_CTX(ctx, flags);
5113 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5115 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5116 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5120 DPRINT(("waking up somebody\n"));
5122 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5125 * safe, we are not in intr handler, nor in ctxsw when
5128 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5134 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5136 pfm_msg_t *msg = NULL;
5138 if (ctx->ctx_fl_no_msg == 0) {
5139 msg = pfm_get_new_msg(ctx);
5141 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5145 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5146 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5147 msg->pfm_ovfl_msg.msg_active_set = 0;
5148 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5149 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5150 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5151 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5152 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5155 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5161 return pfm_notify_user(ctx, msg);
5165 pfm_end_notify_user(pfm_context_t *ctx)
5169 msg = pfm_get_new_msg(ctx);
5171 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5175 memset(msg, 0, sizeof(*msg));
5177 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5178 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5179 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5181 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5186 return pfm_notify_user(ctx, msg);
5190 * main overflow processing routine.
5191 * it can be called from the interrupt path or explicitely during the context switch code
5194 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5196 pfm_ovfl_arg_t *ovfl_arg;
5198 unsigned long old_val, ovfl_val, new_val;
5199 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5200 unsigned long tstamp;
5201 pfm_ovfl_ctrl_t ovfl_ctrl;
5202 unsigned int i, has_smpl;
5203 int must_notify = 0;
5205 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5208 * sanity test. Should never happen
5210 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5212 tstamp = ia64_get_itc();
5213 mask = pmc0 >> PMU_FIRST_COUNTER;
5214 ovfl_val = pmu_conf->ovfl_val;
5215 has_smpl = CTX_HAS_SMPL(ctx);
5217 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5218 "used_pmds=0x%lx\n",
5220 task ? task->pid: -1,
5221 (regs ? regs->cr_iip : 0),
5222 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5223 ctx->ctx_used_pmds[0]));
5227 * first we update the virtual counters
5228 * assume there was a prior ia64_srlz_d() issued
5230 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5232 /* skip pmd which did not overflow */
5233 if ((mask & 0x1) == 0) continue;
5236 * Note that the pmd is not necessarily 0 at this point as qualified events
5237 * may have happened before the PMU was frozen. The residual count is not
5238 * taken into consideration here but will be with any read of the pmd via
5241 old_val = new_val = ctx->ctx_pmds[i].val;
5242 new_val += 1 + ovfl_val;
5243 ctx->ctx_pmds[i].val = new_val;
5246 * check for overflow condition
5248 if (likely(old_val > new_val)) {
5249 ovfl_pmds |= 1UL << i;
5250 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5253 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5257 ia64_get_pmd(i) & ovfl_val,
5263 * there was no 64-bit overflow, nothing else to do
5265 if (ovfl_pmds == 0UL) return;
5268 * reset all control bits
5274 * if a sampling format module exists, then we "cache" the overflow by
5275 * calling the module's handler() routine.
5278 unsigned long start_cycles, end_cycles;
5279 unsigned long pmd_mask;
5281 int this_cpu = smp_processor_id();
5283 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5284 ovfl_arg = &ctx->ctx_ovfl_arg;
5286 prefetch(ctx->ctx_smpl_hdr);
5288 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5292 if ((pmd_mask & 0x1) == 0) continue;
5294 ovfl_arg->ovfl_pmd = (unsigned char )i;
5295 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5296 ovfl_arg->active_set = 0;
5297 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5298 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5300 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5301 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5302 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5305 * copy values of pmds of interest. Sampling format may copy them
5306 * into sampling buffer.
5309 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5310 if ((smpl_pmds & 0x1) == 0) continue;
5311 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5312 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5316 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5318 start_cycles = ia64_get_itc();
5321 * call custom buffer format record (handler) routine
5323 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5325 end_cycles = ia64_get_itc();
5328 * For those controls, we take the union because they have
5329 * an all or nothing behavior.
5331 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5332 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5333 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5335 * build the bitmask of pmds to reset now
5337 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5339 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5342 * when the module cannot handle the rest of the overflows, we abort right here
5344 if (ret && pmd_mask) {
5345 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5346 pmd_mask<<PMU_FIRST_COUNTER));
5349 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5351 ovfl_pmds &= ~reset_pmds;
5354 * when no sampling module is used, then the default
5355 * is to notify on overflow if requested by user
5357 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5358 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5359 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5360 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5362 * if needed, we reset all overflowed pmds
5364 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5367 DPRINT(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n",
5371 * reset the requested PMD registers using the short reset values
5374 unsigned long bm = reset_pmds;
5375 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5378 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5380 * keep track of what to reset when unblocking
5382 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5385 * check for blocking context
5387 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5389 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5392 * set the perfmon specific checking pending work for the task
5394 PFM_SET_WORK_PENDING(task, 1);
5397 * when coming from ctxsw, current still points to the
5398 * previous task, therefore we must work with task and not current.
5400 pfm_set_task_notify(task);
5403 * defer until state is changed (shorten spin window). the context is locked
5404 * anyway, so the signal receiver would come spin for nothing.
5409 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5410 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5411 PFM_GET_WORK_PENDING(task),
5412 ctx->ctx_fl_trap_reason,
5415 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5417 * in case monitoring must be stopped, we toggle the psr bits
5419 if (ovfl_ctrl.bits.mask_monitoring) {
5420 pfm_mask_monitoring(task);
5421 ctx->ctx_state = PFM_CTX_MASKED;
5422 ctx->ctx_fl_can_restart = 1;
5426 * send notification now
5428 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5433 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5435 task ? task->pid : -1,
5441 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5442 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5443 * come here as zombie only if the task is the current task. In which case, we
5444 * can access the PMU hardware directly.
5446 * Note that zombies do have PM_VALID set. So here we do the minimal.
5448 * In case the context was zombified it could not be reclaimed at the time
5449 * the monitoring program exited. At this point, the PMU reservation has been
5450 * returned, the sampiing buffer has been freed. We must convert this call
5451 * into a spurious interrupt. However, we must also avoid infinite overflows
5452 * by stopping monitoring for this task. We can only come here for a per-task
5453 * context. All we need to do is to stop monitoring using the psr bits which
5454 * are always task private. By re-enabling secure montioring, we ensure that
5455 * the monitored task will not be able to re-activate monitoring.
5456 * The task will eventually be context switched out, at which point the context
5457 * will be reclaimed (that includes releasing ownership of the PMU).
5459 * So there might be a window of time where the number of per-task session is zero
5460 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5461 * context. This is safe because if a per-task session comes in, it will push this one
5462 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5463 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5464 * also push our zombie context out.
5466 * Overall pretty hairy stuff....
5468 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5470 ia64_psr(regs)->up = 0;
5471 ia64_psr(regs)->sp = 1;
5476 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5478 struct task_struct *task;
5480 unsigned long flags;
5482 int this_cpu = smp_processor_id();
5485 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5488 * srlz.d done before arriving here
5490 pmc0 = ia64_get_pmc(0);
5492 task = GET_PMU_OWNER();
5493 ctx = GET_PMU_CTX();
5496 * if we have some pending bits set
5497 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5499 if (PMC0_HAS_OVFL(pmc0) && task) {
5501 * we assume that pmc0.fr is always set here
5505 if (!ctx) goto report_spurious1;
5507 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5508 goto report_spurious2;
5510 PROTECT_CTX_NOPRINT(ctx, flags);
5512 pfm_overflow_handler(task, ctx, pmc0, regs);
5514 UNPROTECT_CTX_NOPRINT(ctx, flags);
5517 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5521 * keep it unfrozen at all times
5528 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5529 this_cpu, task->pid);
5533 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5541 pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5543 unsigned long start_cycles, total_cycles;
5544 unsigned long min, max;
5548 this_cpu = get_cpu();
5549 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5550 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5552 start_cycles = ia64_get_itc();
5554 ret = pfm_do_interrupt_handler(irq, arg, regs);
5556 total_cycles = ia64_get_itc();
5559 * don't measure spurious interrupts
5561 if (likely(ret == 0)) {
5562 total_cycles -= start_cycles;
5564 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5565 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5567 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5569 put_cpu_no_resched();
5574 * /proc/perfmon interface, for debug only
5577 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5580 pfm_proc_start(struct seq_file *m, loff_t *pos)
5583 return PFM_PROC_SHOW_HEADER;
5586 while (*pos <= NR_CPUS) {
5587 if (cpu_online(*pos - 1)) {
5588 return (void *)*pos;
5596 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5599 return pfm_proc_start(m, pos);
5603 pfm_proc_stop(struct seq_file *m, void *v)
5608 pfm_proc_show_header(struct seq_file *m)
5610 struct list_head * pos;
5611 pfm_buffer_fmt_t * entry;
5612 unsigned long flags;
5615 "perfmon version : %u.%u\n"
5618 "expert mode : %s\n"
5619 "ovfl_mask : 0x%lx\n"
5620 "PMU flags : 0x%x\n",
5621 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5623 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5624 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5631 "proc_sessions : %u\n"
5632 "sys_sessions : %u\n"
5633 "sys_use_dbregs : %u\n"
5634 "ptrace_use_dbregs : %u\n",
5635 pfm_sessions.pfs_task_sessions,
5636 pfm_sessions.pfs_sys_sessions,
5637 pfm_sessions.pfs_sys_use_dbregs,
5638 pfm_sessions.pfs_ptrace_use_dbregs);
5642 spin_lock(&pfm_buffer_fmt_lock);
5644 list_for_each(pos, &pfm_buffer_fmt_list) {
5645 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5646 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5657 entry->fmt_uuid[10],
5658 entry->fmt_uuid[11],
5659 entry->fmt_uuid[12],
5660 entry->fmt_uuid[13],
5661 entry->fmt_uuid[14],
5662 entry->fmt_uuid[15],
5665 spin_unlock(&pfm_buffer_fmt_lock);
5670 pfm_proc_show(struct seq_file *m, void *v)
5676 if (v == PFM_PROC_SHOW_HEADER) {
5677 pfm_proc_show_header(m);
5681 /* show info for CPU (v - 1) */
5685 "CPU%-2d overflow intrs : %lu\n"
5686 "CPU%-2d overflow cycles : %lu\n"
5687 "CPU%-2d overflow min : %lu\n"
5688 "CPU%-2d overflow max : %lu\n"
5689 "CPU%-2d smpl handler calls : %lu\n"
5690 "CPU%-2d smpl handler cycles : %lu\n"
5691 "CPU%-2d spurious intrs : %lu\n"
5692 "CPU%-2d replay intrs : %lu\n"
5693 "CPU%-2d syst_wide : %d\n"
5694 "CPU%-2d dcr_pp : %d\n"
5695 "CPU%-2d exclude idle : %d\n"
5696 "CPU%-2d owner : %d\n"
5697 "CPU%-2d context : %p\n"
5698 "CPU%-2d activations : %lu\n",
5699 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5700 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5701 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5702 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5703 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5704 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5705 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5706 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5707 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5708 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5709 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5710 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5711 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5712 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5714 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5716 psr = pfm_get_psr();
5721 "CPU%-2d psr : 0x%lx\n"
5722 "CPU%-2d pmc0 : 0x%lx\n",
5724 cpu, ia64_get_pmc(0));
5726 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5727 if (PMC_IS_COUNTING(i) == 0) continue;
5729 "CPU%-2d pmc%u : 0x%lx\n"
5730 "CPU%-2d pmd%u : 0x%lx\n",
5731 cpu, i, ia64_get_pmc(i),
5732 cpu, i, ia64_get_pmd(i));
5738 struct seq_operations pfm_seq_ops = {
5739 .start = pfm_proc_start,
5740 .next = pfm_proc_next,
5741 .stop = pfm_proc_stop,
5742 .show = pfm_proc_show
5746 pfm_proc_open(struct inode *inode, struct file *file)
5748 return seq_open(file, &pfm_seq_ops);
5753 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5754 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5755 * is active or inactive based on mode. We must rely on the value in
5756 * local_cpu_data->pfm_syst_info
5759 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5761 struct pt_regs *regs;
5763 unsigned long dcr_pp;
5765 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5768 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5769 * on every CPU, so we can rely on the pid to identify the idle task.
5771 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5772 regs = ia64_task_regs(task);
5773 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5777 * if monitoring has started
5780 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5782 * context switching in?
5785 /* mask monitoring for the idle task */
5786 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5792 * context switching out
5793 * restore monitoring for next task
5795 * Due to inlining this odd if-then-else construction generates
5798 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5807 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5809 struct task_struct *task = ctx->ctx_task;
5811 ia64_psr(regs)->up = 0;
5812 ia64_psr(regs)->sp = 1;
5814 if (GET_PMU_OWNER() == task) {
5815 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5816 SET_PMU_OWNER(NULL, NULL);
5820 * disconnect the task from the context and vice-versa
5822 PFM_SET_WORK_PENDING(task, 0);
5824 task->thread.pfm_context = NULL;
5825 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5827 DPRINT(("force cleanup for [%d]\n", task->pid));
5832 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5835 pfm_save_regs(struct task_struct *task)
5838 struct thread_struct *t;
5839 unsigned long flags;
5843 ctx = PFM_GET_CTX(task);
5844 if (ctx == NULL) return;
5848 * we always come here with interrupts ALREADY disabled by
5849 * the scheduler. So we simply need to protect against concurrent
5850 * access, not CPU concurrency.
5852 flags = pfm_protect_ctx_ctxsw(ctx);
5854 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5855 struct pt_regs *regs = ia64_task_regs(task);
5859 pfm_force_cleanup(ctx, regs);
5861 BUG_ON(ctx->ctx_smpl_hdr);
5863 pfm_unprotect_ctx_ctxsw(ctx, flags);
5865 pfm_context_free(ctx);
5872 if (ctx->ctx_last_activation != GET_ACTIVATION()) {
5873 pfm_unprotect_ctx_ctxsw(ctx, flags);
5878 * save current PSR: needed because we modify it
5881 psr = pfm_get_psr();
5883 BUG_ON(psr & (IA64_PSR_I));
5887 * This is the last instruction which may generate an overflow
5889 * We do not need to set psr.sp because, it is irrelevant in kernel.
5890 * It will be restored from ipsr when going back to user level
5895 * keep a copy of psr.up (for reload)
5897 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5900 * release ownership of this PMU.
5901 * PM interrupts are masked, so nothing
5904 SET_PMU_OWNER(NULL, NULL);
5907 * we systematically save the PMD as we have no
5908 * guarantee we will be schedule at that same
5911 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5914 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5915 * we will need it on the restore path to check
5916 * for pending overflow.
5918 t->pmcs[0] = ia64_get_pmc(0);
5921 * unfreeze PMU if had pending overflows
5923 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5926 * finally, allow context access.
5927 * interrupts will still be masked after this call.
5929 pfm_unprotect_ctx_ctxsw(ctx, flags);
5932 #else /* !CONFIG_SMP */
5934 pfm_save_regs(struct task_struct *task)
5939 ctx = PFM_GET_CTX(task);
5940 if (ctx == NULL) return;
5943 * save current PSR: needed because we modify it
5945 psr = pfm_get_psr();
5947 BUG_ON(psr & (IA64_PSR_I));
5951 * This is the last instruction which may generate an overflow
5953 * We do not need to set psr.sp because, it is irrelevant in kernel.
5954 * It will be restored from ipsr when going back to user level
5959 * keep a copy of psr.up (for reload)
5961 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5965 pfm_lazy_save_regs (struct task_struct *task)
5968 struct thread_struct *t;
5969 unsigned long flags;
5971 { u64 psr = pfm_get_psr();
5972 BUG_ON(psr & IA64_PSR_UP);
5975 ctx = PFM_GET_CTX(task);
5979 * we need to mask PMU overflow here to
5980 * make sure that we maintain pmc0 until
5981 * we save it. overflow interrupts are
5982 * treated as spurious if there is no
5985 * XXX: I don't think this is necessary
5987 PROTECT_CTX(ctx,flags);
5990 * release ownership of this PMU.
5991 * must be done before we save the registers.
5993 * after this call any PMU interrupt is treated
5996 SET_PMU_OWNER(NULL, NULL);
5999 * save all the pmds we use
6001 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
6004 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6005 * it is needed to check for pended overflow
6006 * on the restore path
6008 t->pmcs[0] = ia64_get_pmc(0);
6011 * unfreeze PMU if had pending overflows
6013 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6016 * now get can unmask PMU interrupts, they will
6017 * be treated as purely spurious and we will not
6018 * lose any information
6020 UNPROTECT_CTX(ctx,flags);
6022 #endif /* CONFIG_SMP */
6026 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6029 pfm_load_regs (struct task_struct *task)
6032 struct thread_struct *t;
6033 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6034 unsigned long flags;
6036 int need_irq_resend;
6038 ctx = PFM_GET_CTX(task);
6039 if (unlikely(ctx == NULL)) return;
6041 BUG_ON(GET_PMU_OWNER());
6045 * possible on unload
6047 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6050 * we always come here with interrupts ALREADY disabled by
6051 * the scheduler. So we simply need to protect against concurrent
6052 * access, not CPU concurrency.
6054 flags = pfm_protect_ctx_ctxsw(ctx);
6055 psr = pfm_get_psr();
6057 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6059 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6060 BUG_ON(psr & IA64_PSR_I);
6062 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6063 struct pt_regs *regs = ia64_task_regs(task);
6065 BUG_ON(ctx->ctx_smpl_hdr);
6067 pfm_force_cleanup(ctx, regs);
6069 pfm_unprotect_ctx_ctxsw(ctx, flags);
6072 * this one (kmalloc'ed) is fine with interrupts disabled
6074 pfm_context_free(ctx);
6080 * we restore ALL the debug registers to avoid picking up
6083 if (ctx->ctx_fl_using_dbreg) {
6084 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6085 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6088 * retrieve saved psr.up
6090 psr_up = ctx->ctx_saved_psr_up;
6093 * if we were the last user of the PMU on that CPU,
6094 * then nothing to do except restore psr
6096 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6099 * retrieve partial reload masks (due to user modifications)
6101 pmc_mask = ctx->ctx_reload_pmcs[0];
6102 pmd_mask = ctx->ctx_reload_pmds[0];
6106 * To avoid leaking information to the user level when psr.sp=0,
6107 * we must reload ALL implemented pmds (even the ones we don't use).
6108 * In the kernel we only allow PFM_READ_PMDS on registers which
6109 * we initialized or requested (sampling) so there is no risk there.
6111 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6114 * ALL accessible PMCs are systematically reloaded, unused registers
6115 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6116 * up stale configuration.
6118 * PMC0 is never in the mask. It is always restored separately.
6120 pmc_mask = ctx->ctx_all_pmcs[0];
6123 * when context is MASKED, we will restore PMC with plm=0
6124 * and PMD with stale information, but that's ok, nothing
6127 * XXX: optimize here
6129 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6130 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6133 * check for pending overflow at the time the state
6136 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6138 * reload pmc0 with the overflow information
6139 * On McKinley PMU, this will trigger a PMU interrupt
6141 ia64_set_pmc(0, t->pmcs[0]);
6146 * will replay the PMU interrupt
6148 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6150 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6154 * we just did a reload, so we reset the partial reload fields
6156 ctx->ctx_reload_pmcs[0] = 0UL;
6157 ctx->ctx_reload_pmds[0] = 0UL;
6159 SET_LAST_CPU(ctx, smp_processor_id());
6162 * dump activation value for this PMU
6166 * record current activation for this context
6168 SET_ACTIVATION(ctx);
6171 * establish new ownership.
6173 SET_PMU_OWNER(task, ctx);
6176 * restore the psr.up bit. measurement
6178 * no PMU interrupt can happen at this point
6179 * because we still have interrupts disabled.
6181 if (likely(psr_up)) pfm_set_psr_up();
6184 * allow concurrent access to context
6186 pfm_unprotect_ctx_ctxsw(ctx, flags);
6188 #else /* !CONFIG_SMP */
6190 * reload PMU state for UP kernels
6191 * in 2.5 we come here with interrupts disabled
6194 pfm_load_regs (struct task_struct *task)
6196 struct thread_struct *t;
6198 struct task_struct *owner;
6199 unsigned long pmd_mask, pmc_mask;
6201 int need_irq_resend;
6203 owner = GET_PMU_OWNER();
6204 ctx = PFM_GET_CTX(task);
6206 psr = pfm_get_psr();
6208 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6209 BUG_ON(psr & IA64_PSR_I);
6212 * we restore ALL the debug registers to avoid picking up
6215 * This must be done even when the task is still the owner
6216 * as the registers may have been modified via ptrace()
6217 * (not perfmon) by the previous task.
6219 if (ctx->ctx_fl_using_dbreg) {
6220 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6221 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6225 * retrieved saved psr.up
6227 psr_up = ctx->ctx_saved_psr_up;
6228 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6231 * short path, our state is still there, just
6232 * need to restore psr and we go
6234 * we do not touch either PMC nor PMD. the psr is not touched
6235 * by the overflow_handler. So we are safe w.r.t. to interrupt
6236 * concurrency even without interrupt masking.
6238 if (likely(owner == task)) {
6239 if (likely(psr_up)) pfm_set_psr_up();
6244 * someone else is still using the PMU, first push it out and
6245 * then we'll be able to install our stuff !
6247 * Upon return, there will be no owner for the current PMU
6249 if (owner) pfm_lazy_save_regs(owner);
6252 * To avoid leaking information to the user level when psr.sp=0,
6253 * we must reload ALL implemented pmds (even the ones we don't use).
6254 * In the kernel we only allow PFM_READ_PMDS on registers which
6255 * we initialized or requested (sampling) so there is no risk there.
6257 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6260 * ALL accessible PMCs are systematically reloaded, unused registers
6261 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6262 * up stale configuration.
6264 * PMC0 is never in the mask. It is always restored separately
6266 pmc_mask = ctx->ctx_all_pmcs[0];
6268 pfm_restore_pmds(t->pmds, pmd_mask);
6269 pfm_restore_pmcs(t->pmcs, pmc_mask);
6272 * check for pending overflow at the time the state
6275 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6277 * reload pmc0 with the overflow information
6278 * On McKinley PMU, this will trigger a PMU interrupt
6280 ia64_set_pmc(0, t->pmcs[0]);
6286 * will replay the PMU interrupt
6288 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6290 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6294 * establish new ownership.
6296 SET_PMU_OWNER(task, ctx);
6299 * restore the psr.up bit. measurement
6301 * no PMU interrupt can happen at this point
6302 * because we still have interrupts disabled.
6304 if (likely(psr_up)) pfm_set_psr_up();
6306 #endif /* CONFIG_SMP */
6309 * this function assumes monitoring is stopped
6312 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6315 unsigned long mask2, val, pmd_val, ovfl_val;
6316 int i, can_access_pmu = 0;
6320 * is the caller the task being monitored (or which initiated the
6321 * session for system wide measurements)
6323 is_self = ctx->ctx_task == task ? 1 : 0;
6326 if (task == current) {
6329 * in UP, the state can still be in the registers
6331 if (task == current || GET_PMU_OWNER() == task) {
6335 * Mark the PMU as not owned
6336 * This will cause the interrupt handler to do nothing in case an overflow
6337 * interrupt was in-flight
6338 * This also guarantees that pmc0 will contain the final state
6339 * It virtually gives us full control on overflow processing from that point
6342 SET_PMU_OWNER(NULL, NULL);
6345 * read current overflow status:
6347 * we are guaranteed to read the final stable state
6350 pmc0 = ia64_get_pmc(0); /* slow */
6353 * reset freeze bit, overflow status information destroyed
6357 pmc0 = task->thread.pmcs[0];
6359 * clear whatever overflow status bits there were
6361 task->thread.pmcs[0] = 0;
6363 ovfl_val = pmu_conf->ovfl_val;
6365 * we save all the used pmds
6366 * we take care of overflows for counting PMDs
6368 * XXX: sampling situation is not taken into account here
6370 mask2 = ctx->ctx_used_pmds[0];
6371 for (i = 0; mask2; i++, mask2>>=1) {
6373 /* skip non used pmds */
6374 if ((mask2 & 0x1) == 0) continue;
6377 * can access PMU always true in system wide mode
6379 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6381 if (PMD_IS_COUNTING(i)) {
6382 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6385 ctx->ctx_pmds[i].val,
6389 * we rebuild the full 64 bit value of the counter
6391 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6394 * now everything is in ctx_pmds[] and we need
6395 * to clear the saved context from save_regs() such that
6396 * pfm_read_pmds() gets the correct value
6401 * take care of overflow inline
6403 if (pmc0 & (1UL << i)) {
6404 val += 1 + ovfl_val;
6405 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6409 DPRINT(("[%d] is_self=%d ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, is_self, i, val, pmd_val));
6411 if (is_self) task->thread.pmds[i] = pmd_val;
6413 ctx->ctx_pmds[i].val = val;
6417 static struct irqaction perfmon_irqaction = {
6418 .handler = pfm_interrupt_handler,
6419 .flags = SA_INTERRUPT,
6424 * perfmon initialization routine, called from the initcall() table
6426 static int init_pfm_fs(void);
6434 family = local_cpu_data->family;
6439 if ((*p)->probe() == 0) goto found;
6440 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6451 static struct file_operations pfm_proc_fops = {
6452 .open = pfm_proc_open,
6454 .llseek = seq_lseek,
6455 .release = seq_release,
6461 unsigned int n, n_counters, i;
6463 printk("perfmon: version %u.%u IRQ %u\n",
6466 IA64_PERFMON_VECTOR);
6468 if (pfm_probe_pmu()) {
6469 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6470 local_cpu_data->family);
6475 * compute the number of implemented PMD/PMC from the
6476 * description tables
6479 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6480 if (PMC_IS_IMPL(i) == 0) continue;
6481 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6484 pmu_conf->num_pmcs = n;
6486 n = 0; n_counters = 0;
6487 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6488 if (PMD_IS_IMPL(i) == 0) continue;
6489 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6491 if (PMD_IS_COUNTING(i)) n_counters++;
6493 pmu_conf->num_pmds = n;
6494 pmu_conf->num_counters = n_counters;
6497 * sanity checks on the number of debug registers
6499 if (pmu_conf->use_rr_dbregs) {
6500 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6501 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6505 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6506 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6512 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6516 pmu_conf->num_counters,
6517 ffz(pmu_conf->ovfl_val));
6520 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6521 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6527 * create /proc/perfmon (mostly for debugging purposes)
6529 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6530 if (perfmon_dir == NULL) {
6531 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6536 * install customized file operations for /proc/perfmon entry
6538 perfmon_dir->proc_fops = &pfm_proc_fops;
6541 * create /proc/sys/kernel/perfmon (for debugging purposes)
6543 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6546 * initialize all our spinlocks
6548 spin_lock_init(&pfm_sessions.pfs_lock);
6549 spin_lock_init(&pfm_buffer_fmt_lock);
6553 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6558 __initcall(pfm_init);
6561 * this function is called before pfm_init()
6564 pfm_init_percpu (void)
6567 * make sure no measurement is active
6568 * (may inherit programmed PMCs from EFI).
6574 * we run with the PMU not frozen at all times
6578 if (smp_processor_id() == 0)
6579 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6581 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6586 * used for debug purposes only
6589 dump_pmu_state(const char *from)
6591 struct task_struct *task;
6592 struct thread_struct *t;
6593 struct pt_regs *regs;
6595 unsigned long psr, dcr, info, flags;
6598 local_irq_save(flags);
6600 this_cpu = smp_processor_id();
6601 regs = ia64_task_regs(current);
6602 info = PFM_CPUINFO_GET();
6603 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6605 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6606 local_irq_restore(flags);
6610 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6617 task = GET_PMU_OWNER();
6618 ctx = GET_PMU_CTX();
6620 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6622 psr = pfm_get_psr();
6624 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",
6627 psr & IA64_PSR_PP ? 1 : 0,
6628 psr & IA64_PSR_UP ? 1 : 0,
6629 dcr & IA64_DCR_PP ? 1 : 0,
6632 ia64_psr(regs)->pp);
6634 ia64_psr(regs)->up = 0;
6635 ia64_psr(regs)->pp = 0;
6637 t = ¤t->thread;
6639 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6640 if (PMC_IS_IMPL(i) == 0) continue;
6641 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6644 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6645 if (PMD_IS_IMPL(i) == 0) continue;
6646 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6650 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6653 ctx->ctx_smpl_vaddr,
6657 ctx->ctx_saved_psr_up);
6659 local_irq_restore(flags);
6663 * called from process.c:copy_thread(). task is new child.
6666 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6668 struct thread_struct *thread;
6670 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6672 thread = &task->thread;
6675 * cut links inherited from parent (current)
6677 thread->pfm_context = NULL;
6679 PFM_SET_WORK_PENDING(task, 0);
6682 * the psr bits are already set properly in copy_threads()
6685 #else /* !CONFIG_PERFMON */
6687 sys_perfmonctl (int fd, int cmd, void *arg, int count, long arg5, long arg6, long arg7,
6688 long arg8, long stack)
6692 #endif /* CONFIG_PERFMON */