2 * This file implements the perfmon-2 subsystem which is used
3 * to program the IA-64 Performance Monitoring Unit (PMU).
5 * The initial version of perfmon.c was written by
6 * Ganesh Venkitachalam, IBM Corp.
8 * Then it was modified for perfmon-1.x by Stephane Eranian and
9 * David Mosberger, Hewlett Packard Co.
11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
12 * by Stephane Eranian, Hewlett Packard Co.
14 * Copyright (C) 1999-2005 Hewlett Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 * David Mosberger-Tang <davidm@hpl.hp.com>
18 * More information about perfmon available at:
19 * http://www.hpl.hp.com/research/linux/perfmon
22 #include <linux/module.h>
23 #include <linux/kernel.h>
24 #include <linux/sched.h>
25 #include <linux/interrupt.h>
26 #include <linux/proc_fs.h>
27 #include <linux/seq_file.h>
28 #include <linux/init.h>
29 #include <linux/vmalloc.h>
31 #include <linux/sysctl.h>
32 #include <linux/list.h>
33 #include <linux/file.h>
34 #include <linux/poll.h>
35 #include <linux/vfs.h>
36 #include <linux/smp.h>
37 #include <linux/pagemap.h>
38 #include <linux/mount.h>
39 #include <linux/bitops.h>
40 #include <linux/capability.h>
41 #include <linux/rcupdate.h>
42 #include <linux/completion.h>
43 #include <linux/tracehook.h>
44 #include <linux/slab.h>
46 #include <asm/errno.h>
47 #include <asm/intrinsics.h>
49 #include <asm/perfmon.h>
50 #include <asm/processor.h>
51 #include <asm/signal.h>
52 #include <asm/uaccess.h>
53 #include <asm/delay.h>
57 * perfmon context state
59 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
60 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
61 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
62 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
64 #define PFM_INVALID_ACTIVATION (~0UL)
66 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
67 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
70 * depth of message queue
72 #define PFM_MAX_MSGS 32
73 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
76 * type of a PMU register (bitmask).
78 * bit0 : register implemented
81 * bit4 : pmc has pmc.pm
82 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
83 * bit6-7 : register type
86 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
87 #define PFM_REG_IMPL 0x1 /* register implemented */
88 #define PFM_REG_END 0x2 /* end marker */
89 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
90 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
91 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
92 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
93 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
95 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
96 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
98 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
100 /* i assumed unsigned */
101 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
102 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
104 /* XXX: these assume that register i is implemented */
105 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
106 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
107 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
108 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
110 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
111 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
112 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
113 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
115 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
116 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
118 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
119 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
120 #define PFM_CTX_TASK(h) (h)->ctx_task
122 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
124 /* XXX: does not support more than 64 PMDs */
125 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
126 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
128 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
130 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
131 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
132 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
133 #define PFM_CODE_RR 0 /* requesting code range restriction */
134 #define PFM_DATA_RR 1 /* requestion data range restriction */
136 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
137 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
138 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
140 #define RDEP(x) (1UL<<(x))
143 * context protection macros
145 * - we need to protect against CPU concurrency (spin_lock)
146 * - we need to protect against PMU overflow interrupts (local_irq_disable)
148 * - we need to protect against PMU overflow interrupts (local_irq_disable)
150 * spin_lock_irqsave()/spin_unlock_irqrestore():
151 * in SMP: local_irq_disable + spin_lock
152 * in UP : local_irq_disable
154 * spin_lock()/spin_lock():
155 * in UP : removed automatically
156 * in SMP: protect against context accesses from other CPU. interrupts
157 * are not masked. This is useful for the PMU interrupt handler
158 * because we know we will not get PMU concurrency in that code.
160 #define PROTECT_CTX(c, f) \
162 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
163 spin_lock_irqsave(&(c)->ctx_lock, f); \
164 DPRINT(("spinlocked ctx %p by [%d]\n", c, task_pid_nr(current))); \
167 #define UNPROTECT_CTX(c, f) \
169 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
170 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
173 #define PROTECT_CTX_NOPRINT(c, f) \
175 spin_lock_irqsave(&(c)->ctx_lock, f); \
179 #define UNPROTECT_CTX_NOPRINT(c, f) \
181 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
185 #define PROTECT_CTX_NOIRQ(c) \
187 spin_lock(&(c)->ctx_lock); \
190 #define UNPROTECT_CTX_NOIRQ(c) \
192 spin_unlock(&(c)->ctx_lock); \
198 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
199 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
200 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
202 #else /* !CONFIG_SMP */
203 #define SET_ACTIVATION(t) do {} while(0)
204 #define GET_ACTIVATION(t) do {} while(0)
205 #define INC_ACTIVATION(t) do {} while(0)
206 #endif /* CONFIG_SMP */
208 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
209 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
210 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
212 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
213 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
215 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
218 * cmp0 must be the value of pmc0
220 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
222 #define PFMFS_MAGIC 0xa0b4d889
227 #define PFM_DEBUGGING 1
231 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
234 #define DPRINT_ovfl(a) \
236 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
241 * 64-bit software counter structure
243 * the next_reset_type is applied to the next call to pfm_reset_regs()
246 unsigned long val; /* virtual 64bit counter value */
247 unsigned long lval; /* last reset value */
248 unsigned long long_reset; /* reset value on sampling overflow */
249 unsigned long short_reset; /* reset value on overflow */
250 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
251 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
252 unsigned long seed; /* seed for random-number generator */
253 unsigned long mask; /* mask for random-number generator */
254 unsigned int flags; /* notify/do not notify */
255 unsigned long eventid; /* overflow event identifier */
262 unsigned int block:1; /* when 1, task will blocked on user notifications */
263 unsigned int system:1; /* do system wide monitoring */
264 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
265 unsigned int is_sampling:1; /* true if using a custom format */
266 unsigned int excl_idle:1; /* exclude idle task in system wide session */
267 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
268 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
269 unsigned int no_msg:1; /* no message sent on overflow */
270 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
271 unsigned int reserved:22;
272 } pfm_context_flags_t;
274 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
275 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
276 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
280 * perfmon context: encapsulates all the state of a monitoring session
283 typedef struct pfm_context {
284 spinlock_t ctx_lock; /* context protection */
286 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
287 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
289 struct task_struct *ctx_task; /* task to which context is attached */
291 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
293 struct completion ctx_restart_done; /* use for blocking notification mode */
295 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
296 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
297 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
299 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
300 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
301 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
303 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
305 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
306 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
307 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
308 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
310 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
312 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
313 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
315 unsigned long ctx_saved_psr_up; /* only contains psr.up value */
317 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
318 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
319 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
321 int ctx_fd; /* file descriptor used my this context */
322 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
324 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
325 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
326 unsigned long ctx_smpl_size; /* size of sampling buffer */
327 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
329 wait_queue_head_t ctx_msgq_wait;
330 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
333 struct fasync_struct *ctx_async_queue;
335 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
339 * magic number used to verify that structure is really
342 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
344 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
347 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
348 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
350 #define SET_LAST_CPU(ctx, v) do {} while(0)
351 #define GET_LAST_CPU(ctx) do {} while(0)
355 #define ctx_fl_block ctx_flags.block
356 #define ctx_fl_system ctx_flags.system
357 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
358 #define ctx_fl_is_sampling ctx_flags.is_sampling
359 #define ctx_fl_excl_idle ctx_flags.excl_idle
360 #define ctx_fl_going_zombie ctx_flags.going_zombie
361 #define ctx_fl_trap_reason ctx_flags.trap_reason
362 #define ctx_fl_no_msg ctx_flags.no_msg
363 #define ctx_fl_can_restart ctx_flags.can_restart
365 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
366 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
369 * global information about all sessions
370 * mostly used to synchronize between system wide and per-process
373 spinlock_t pfs_lock; /* lock the structure */
375 unsigned int pfs_task_sessions; /* number of per task sessions */
376 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
377 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
378 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
379 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
383 * information about a PMC or PMD.
384 * dep_pmd[]: a bitmask of dependent PMD registers
385 * dep_pmc[]: a bitmask of dependent PMC registers
387 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
391 unsigned long default_value; /* power-on default value */
392 unsigned long reserved_mask; /* bitmask of reserved bits */
393 pfm_reg_check_t read_check;
394 pfm_reg_check_t write_check;
395 unsigned long dep_pmd[4];
396 unsigned long dep_pmc[4];
399 /* assume cnum is a valid monitor */
400 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
403 * This structure is initialized at boot time and contains
404 * a description of the PMU main characteristics.
406 * If the probe function is defined, detection is based
407 * on its return value:
408 * - 0 means recognized PMU
409 * - anything else means not supported
410 * When the probe function is not defined, then the pmu_family field
411 * is used and it must match the host CPU family such that:
412 * - cpu->family & config->pmu_family != 0
415 unsigned long ovfl_val; /* overflow value for counters */
417 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
418 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
420 unsigned int num_pmcs; /* number of PMCS: computed at init time */
421 unsigned int num_pmds; /* number of PMDS: computed at init time */
422 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
423 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
425 char *pmu_name; /* PMU family name */
426 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
427 unsigned int flags; /* pmu specific flags */
428 unsigned int num_ibrs; /* number of IBRS: computed at init time */
429 unsigned int num_dbrs; /* number of DBRS: computed at init time */
430 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
431 int (*probe)(void); /* customized probe routine */
432 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
437 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
440 * debug register related type definitions
443 unsigned long ibr_mask:56;
444 unsigned long ibr_plm:4;
445 unsigned long ibr_ig:3;
446 unsigned long ibr_x:1;
450 unsigned long dbr_mask:56;
451 unsigned long dbr_plm:4;
452 unsigned long dbr_ig:2;
453 unsigned long dbr_w:1;
454 unsigned long dbr_r:1;
465 * perfmon command descriptions
468 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
471 unsigned int cmd_narg;
473 int (*cmd_getsize)(void *arg, size_t *sz);
476 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
477 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
478 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
479 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
482 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
483 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
484 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
485 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
486 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
488 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
491 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
492 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
493 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
496 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
497 unsigned long pfm_smpl_handler_calls;
498 unsigned long pfm_smpl_handler_cycles;
499 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
503 * perfmon internal variables
505 static pfm_stats_t pfm_stats[NR_CPUS];
506 static pfm_session_t pfm_sessions; /* global sessions information */
508 static DEFINE_SPINLOCK(pfm_alt_install_check);
509 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
511 static struct proc_dir_entry *perfmon_dir;
512 static pfm_uuid_t pfm_null_uuid = {0,};
514 static spinlock_t pfm_buffer_fmt_lock;
515 static LIST_HEAD(pfm_buffer_fmt_list);
517 static pmu_config_t *pmu_conf;
519 /* sysctl() controls */
520 pfm_sysctl_t pfm_sysctl;
521 EXPORT_SYMBOL(pfm_sysctl);
523 static ctl_table pfm_ctl_table[]={
526 .data = &pfm_sysctl.debug,
527 .maxlen = sizeof(int),
529 .proc_handler = proc_dointvec,
532 .procname = "debug_ovfl",
533 .data = &pfm_sysctl.debug_ovfl,
534 .maxlen = sizeof(int),
536 .proc_handler = proc_dointvec,
539 .procname = "fastctxsw",
540 .data = &pfm_sysctl.fastctxsw,
541 .maxlen = sizeof(int),
543 .proc_handler = proc_dointvec,
546 .procname = "expert_mode",
547 .data = &pfm_sysctl.expert_mode,
548 .maxlen = sizeof(int),
550 .proc_handler = proc_dointvec,
554 static ctl_table pfm_sysctl_dir[] = {
556 .procname = "perfmon",
558 .child = pfm_ctl_table,
562 static ctl_table pfm_sysctl_root[] = {
564 .procname = "kernel",
566 .child = pfm_sysctl_dir,
570 static struct ctl_table_header *pfm_sysctl_header;
572 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
574 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
575 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
578 pfm_put_task(struct task_struct *task)
580 if (task != current) put_task_struct(task);
584 pfm_reserve_page(unsigned long a)
586 SetPageReserved(vmalloc_to_page((void *)a));
589 pfm_unreserve_page(unsigned long a)
591 ClearPageReserved(vmalloc_to_page((void*)a));
594 static inline unsigned long
595 pfm_protect_ctx_ctxsw(pfm_context_t *x)
597 spin_lock(&(x)->ctx_lock);
602 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
604 spin_unlock(&(x)->ctx_lock);
607 /* forward declaration */
608 static const struct dentry_operations pfmfs_dentry_operations;
610 static struct dentry *
611 pfmfs_mount(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
613 return mount_pseudo(fs_type, "pfm:", NULL, &pfmfs_dentry_operations,
617 static struct file_system_type pfm_fs_type = {
619 .mount = pfmfs_mount,
620 .kill_sb = kill_anon_super,
623 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
624 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
625 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
626 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
627 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
630 /* forward declaration */
631 static const struct file_operations pfm_file_ops;
634 * forward declarations
637 static void pfm_lazy_save_regs (struct task_struct *ta);
640 void dump_pmu_state(const char *);
641 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
643 #include "perfmon_itanium.h"
644 #include "perfmon_mckinley.h"
645 #include "perfmon_montecito.h"
646 #include "perfmon_generic.h"
648 static pmu_config_t *pmu_confs[]={
652 &pmu_conf_gen, /* must be last */
657 static int pfm_end_notify_user(pfm_context_t *ctx);
660 pfm_clear_psr_pp(void)
662 ia64_rsm(IA64_PSR_PP);
669 ia64_ssm(IA64_PSR_PP);
674 pfm_clear_psr_up(void)
676 ia64_rsm(IA64_PSR_UP);
683 ia64_ssm(IA64_PSR_UP);
687 static inline unsigned long
691 tmp = ia64_getreg(_IA64_REG_PSR);
697 pfm_set_psr_l(unsigned long val)
699 ia64_setreg(_IA64_REG_PSR_L, val);
711 pfm_unfreeze_pmu(void)
718 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
722 for (i=0; i < nibrs; i++) {
723 ia64_set_ibr(i, ibrs[i]);
724 ia64_dv_serialize_instruction();
730 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
734 for (i=0; i < ndbrs; i++) {
735 ia64_set_dbr(i, dbrs[i]);
736 ia64_dv_serialize_data();
742 * PMD[i] must be a counter. no check is made
744 static inline unsigned long
745 pfm_read_soft_counter(pfm_context_t *ctx, int i)
747 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
751 * PMD[i] must be a counter. no check is made
754 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
756 unsigned long ovfl_val = pmu_conf->ovfl_val;
758 ctx->ctx_pmds[i].val = val & ~ovfl_val;
760 * writing to unimplemented part is ignore, so we do not need to
763 ia64_set_pmd(i, val & ovfl_val);
767 pfm_get_new_msg(pfm_context_t *ctx)
771 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
773 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
774 if (next == ctx->ctx_msgq_head) return NULL;
776 idx = ctx->ctx_msgq_tail;
777 ctx->ctx_msgq_tail = next;
779 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
781 return ctx->ctx_msgq+idx;
785 pfm_get_next_msg(pfm_context_t *ctx)
789 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
791 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
796 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
801 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
803 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));
809 pfm_reset_msgq(pfm_context_t *ctx)
811 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
812 DPRINT(("ctx=%p msgq reset\n", ctx));
816 pfm_rvmalloc(unsigned long size)
821 size = PAGE_ALIGN(size);
824 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
825 addr = (unsigned long)mem;
827 pfm_reserve_page(addr);
836 pfm_rvfree(void *mem, unsigned long size)
841 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
842 addr = (unsigned long) mem;
843 while ((long) size > 0) {
844 pfm_unreserve_page(addr);
853 static pfm_context_t *
854 pfm_context_alloc(int ctx_flags)
859 * allocate context descriptor
860 * must be able to free with interrupts disabled
862 ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
864 DPRINT(("alloc ctx @%p\n", ctx));
867 * init context protection lock
869 spin_lock_init(&ctx->ctx_lock);
872 * context is unloaded
874 ctx->ctx_state = PFM_CTX_UNLOADED;
877 * initialization of context's flags
879 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
880 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
881 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
883 * will move to set properties
884 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
888 * init restart semaphore to locked
890 init_completion(&ctx->ctx_restart_done);
893 * activation is used in SMP only
895 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
896 SET_LAST_CPU(ctx, -1);
899 * initialize notification message queue
901 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
902 init_waitqueue_head(&ctx->ctx_msgq_wait);
903 init_waitqueue_head(&ctx->ctx_zombieq);
910 pfm_context_free(pfm_context_t *ctx)
913 DPRINT(("free ctx @%p\n", ctx));
919 pfm_mask_monitoring(struct task_struct *task)
921 pfm_context_t *ctx = PFM_GET_CTX(task);
922 unsigned long mask, val, ovfl_mask;
925 DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
927 ovfl_mask = pmu_conf->ovfl_val;
929 * monitoring can only be masked as a result of a valid
930 * counter overflow. In UP, it means that the PMU still
931 * has an owner. Note that the owner can be different
932 * from the current task. However the PMU state belongs
934 * In SMP, a valid overflow only happens when task is
935 * current. Therefore if we come here, we know that
936 * the PMU state belongs to the current task, therefore
937 * we can access the live registers.
939 * So in both cases, the live register contains the owner's
940 * state. We can ONLY touch the PMU registers and NOT the PSR.
942 * As a consequence to this call, the ctx->th_pmds[] array
943 * contains stale information which must be ignored
944 * when context is reloaded AND monitoring is active (see
947 mask = ctx->ctx_used_pmds[0];
948 for (i = 0; mask; i++, mask>>=1) {
949 /* skip non used pmds */
950 if ((mask & 0x1) == 0) continue;
951 val = ia64_get_pmd(i);
953 if (PMD_IS_COUNTING(i)) {
955 * we rebuild the full 64 bit value of the counter
957 ctx->ctx_pmds[i].val += (val & ovfl_mask);
959 ctx->ctx_pmds[i].val = val;
961 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
963 ctx->ctx_pmds[i].val,
967 * mask monitoring by setting the privilege level to 0
968 * we cannot use psr.pp/psr.up for this, it is controlled by
971 * if task is current, modify actual registers, otherwise modify
972 * thread save state, i.e., what will be restored in pfm_load_regs()
974 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
975 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
976 if ((mask & 0x1) == 0UL) continue;
977 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
978 ctx->th_pmcs[i] &= ~0xfUL;
979 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
982 * make all of this visible
988 * must always be done with task == current
990 * context must be in MASKED state when calling
993 pfm_restore_monitoring(struct task_struct *task)
995 pfm_context_t *ctx = PFM_GET_CTX(task);
996 unsigned long mask, ovfl_mask;
997 unsigned long psr, val;
1000 is_system = ctx->ctx_fl_system;
1001 ovfl_mask = pmu_conf->ovfl_val;
1003 if (task != current) {
1004 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1007 if (ctx->ctx_state != PFM_CTX_MASKED) {
1008 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1009 task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1012 psr = pfm_get_psr();
1014 * monitoring is masked via the PMC.
1015 * As we restore their value, we do not want each counter to
1016 * restart right away. We stop monitoring using the PSR,
1017 * restore the PMC (and PMD) and then re-establish the psr
1018 * as it was. Note that there can be no pending overflow at
1019 * this point, because monitoring was MASKED.
1021 * system-wide session are pinned and self-monitoring
1023 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1024 /* disable dcr pp */
1025 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1031 * first, we restore the PMD
1033 mask = ctx->ctx_used_pmds[0];
1034 for (i = 0; mask; i++, mask>>=1) {
1035 /* skip non used pmds */
1036 if ((mask & 0x1) == 0) continue;
1038 if (PMD_IS_COUNTING(i)) {
1040 * we split the 64bit value according to
1043 val = ctx->ctx_pmds[i].val & ovfl_mask;
1044 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1046 val = ctx->ctx_pmds[i].val;
1048 ia64_set_pmd(i, val);
1050 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1052 ctx->ctx_pmds[i].val,
1058 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1059 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1060 if ((mask & 0x1) == 0UL) continue;
1061 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1062 ia64_set_pmc(i, ctx->th_pmcs[i]);
1063 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1064 task_pid_nr(task), i, ctx->th_pmcs[i]));
1069 * must restore DBR/IBR because could be modified while masked
1070 * XXX: need to optimize
1072 if (ctx->ctx_fl_using_dbreg) {
1073 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1074 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1080 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1082 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1089 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1095 for (i=0; mask; i++, mask>>=1) {
1096 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1101 * reload from thread state (used for ctxw only)
1104 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1107 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1109 for (i=0; mask; i++, mask>>=1) {
1110 if ((mask & 0x1) == 0) continue;
1111 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1112 ia64_set_pmd(i, val);
1118 * propagate PMD from context to thread-state
1121 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1123 unsigned long ovfl_val = pmu_conf->ovfl_val;
1124 unsigned long mask = ctx->ctx_all_pmds[0];
1128 DPRINT(("mask=0x%lx\n", mask));
1130 for (i=0; mask; i++, mask>>=1) {
1132 val = ctx->ctx_pmds[i].val;
1135 * We break up the 64 bit value into 2 pieces
1136 * the lower bits go to the machine state in the
1137 * thread (will be reloaded on ctxsw in).
1138 * The upper part stays in the soft-counter.
1140 if (PMD_IS_COUNTING(i)) {
1141 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1144 ctx->th_pmds[i] = val;
1146 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1149 ctx->ctx_pmds[i].val));
1154 * propagate PMC from context to thread-state
1157 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1159 unsigned long mask = ctx->ctx_all_pmcs[0];
1162 DPRINT(("mask=0x%lx\n", mask));
1164 for (i=0; mask; i++, mask>>=1) {
1165 /* masking 0 with ovfl_val yields 0 */
1166 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1167 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1174 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1178 for (i=0; mask; i++, mask>>=1) {
1179 if ((mask & 0x1) == 0) continue;
1180 ia64_set_pmc(i, pmcs[i]);
1186 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1188 return memcmp(a, b, sizeof(pfm_uuid_t));
1192 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1195 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1200 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1203 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1209 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1213 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1218 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1222 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1227 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1230 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1235 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)
1238 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1242 static pfm_buffer_fmt_t *
1243 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1245 struct list_head * pos;
1246 pfm_buffer_fmt_t * entry;
1248 list_for_each(pos, &pfm_buffer_fmt_list) {
1249 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1250 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1257 * find a buffer format based on its uuid
1259 static pfm_buffer_fmt_t *
1260 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1262 pfm_buffer_fmt_t * fmt;
1263 spin_lock(&pfm_buffer_fmt_lock);
1264 fmt = __pfm_find_buffer_fmt(uuid);
1265 spin_unlock(&pfm_buffer_fmt_lock);
1270 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1274 /* some sanity checks */
1275 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1277 /* we need at least a handler */
1278 if (fmt->fmt_handler == NULL) return -EINVAL;
1281 * XXX: need check validity of fmt_arg_size
1284 spin_lock(&pfm_buffer_fmt_lock);
1286 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1287 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1291 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1292 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1295 spin_unlock(&pfm_buffer_fmt_lock);
1298 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1301 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1303 pfm_buffer_fmt_t *fmt;
1306 spin_lock(&pfm_buffer_fmt_lock);
1308 fmt = __pfm_find_buffer_fmt(uuid);
1310 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1314 list_del_init(&fmt->fmt_list);
1315 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1318 spin_unlock(&pfm_buffer_fmt_lock);
1322 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1324 extern void update_pal_halt_status(int);
1327 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1329 unsigned long flags;
1331 * validity checks on cpu_mask have been done upstream
1335 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1336 pfm_sessions.pfs_sys_sessions,
1337 pfm_sessions.pfs_task_sessions,
1338 pfm_sessions.pfs_sys_use_dbregs,
1344 * cannot mix system wide and per-task sessions
1346 if (pfm_sessions.pfs_task_sessions > 0UL) {
1347 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1348 pfm_sessions.pfs_task_sessions));
1352 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1354 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1356 pfm_sessions.pfs_sys_session[cpu] = task;
1358 pfm_sessions.pfs_sys_sessions++ ;
1361 if (pfm_sessions.pfs_sys_sessions) goto abort;
1362 pfm_sessions.pfs_task_sessions++;
1365 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1366 pfm_sessions.pfs_sys_sessions,
1367 pfm_sessions.pfs_task_sessions,
1368 pfm_sessions.pfs_sys_use_dbregs,
1373 * disable default_idle() to go to PAL_HALT
1375 update_pal_halt_status(0);
1382 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1383 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1393 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1395 unsigned long flags;
1397 * validity checks on cpu_mask have been done upstream
1401 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1402 pfm_sessions.pfs_sys_sessions,
1403 pfm_sessions.pfs_task_sessions,
1404 pfm_sessions.pfs_sys_use_dbregs,
1410 pfm_sessions.pfs_sys_session[cpu] = NULL;
1412 * would not work with perfmon+more than one bit in cpu_mask
1414 if (ctx && ctx->ctx_fl_using_dbreg) {
1415 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1416 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1418 pfm_sessions.pfs_sys_use_dbregs--;
1421 pfm_sessions.pfs_sys_sessions--;
1423 pfm_sessions.pfs_task_sessions--;
1425 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1426 pfm_sessions.pfs_sys_sessions,
1427 pfm_sessions.pfs_task_sessions,
1428 pfm_sessions.pfs_sys_use_dbregs,
1433 * if possible, enable default_idle() to go into PAL_HALT
1435 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1436 update_pal_halt_status(1);
1444 * removes virtual mapping of the sampling buffer.
1445 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1446 * a PROTECT_CTX() section.
1449 pfm_remove_smpl_mapping(void *vaddr, unsigned long size)
1451 struct task_struct *task = current;
1455 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1456 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1460 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1463 * does the actual unmapping
1465 r = vm_munmap((unsigned long)vaddr, size);
1468 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1471 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1477 * free actual physical storage used by sampling buffer
1481 pfm_free_smpl_buffer(pfm_context_t *ctx)
1483 pfm_buffer_fmt_t *fmt;
1485 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1488 * we won't use the buffer format anymore
1490 fmt = ctx->ctx_buf_fmt;
1492 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1495 ctx->ctx_smpl_vaddr));
1497 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1502 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1504 ctx->ctx_smpl_hdr = NULL;
1505 ctx->ctx_smpl_size = 0UL;
1510 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1516 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1518 if (fmt == NULL) return;
1520 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1525 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1526 * no real gain from having the whole whorehouse mounted. So we don't need
1527 * any operations on the root directory. However, we need a non-trivial
1528 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1530 static struct vfsmount *pfmfs_mnt __read_mostly;
1535 int err = register_filesystem(&pfm_fs_type);
1537 pfmfs_mnt = kern_mount(&pfm_fs_type);
1538 err = PTR_ERR(pfmfs_mnt);
1539 if (IS_ERR(pfmfs_mnt))
1540 unregister_filesystem(&pfm_fs_type);
1548 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1553 unsigned long flags;
1554 DECLARE_WAITQUEUE(wait, current);
1555 if (PFM_IS_FILE(filp) == 0) {
1556 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1560 ctx = filp->private_data;
1562 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1567 * check even when there is no message
1569 if (size < sizeof(pfm_msg_t)) {
1570 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1574 PROTECT_CTX(ctx, flags);
1577 * put ourselves on the wait queue
1579 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1587 set_current_state(TASK_INTERRUPTIBLE);
1589 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1592 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1594 UNPROTECT_CTX(ctx, flags);
1597 * check non-blocking read
1600 if(filp->f_flags & O_NONBLOCK) break;
1603 * check pending signals
1605 if(signal_pending(current)) {
1610 * no message, so wait
1614 PROTECT_CTX(ctx, flags);
1616 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1617 set_current_state(TASK_RUNNING);
1618 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1620 if (ret < 0) goto abort;
1623 msg = pfm_get_next_msg(ctx);
1625 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1629 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1632 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1635 UNPROTECT_CTX(ctx, flags);
1641 pfm_write(struct file *file, const char __user *ubuf,
1642 size_t size, loff_t *ppos)
1644 DPRINT(("pfm_write called\n"));
1649 pfm_poll(struct file *filp, poll_table * wait)
1652 unsigned long flags;
1653 unsigned int mask = 0;
1655 if (PFM_IS_FILE(filp) == 0) {
1656 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1660 ctx = filp->private_data;
1662 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1667 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1669 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1671 PROTECT_CTX(ctx, flags);
1673 if (PFM_CTXQ_EMPTY(ctx) == 0)
1674 mask = POLLIN | POLLRDNORM;
1676 UNPROTECT_CTX(ctx, flags);
1678 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1684 pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1686 DPRINT(("pfm_ioctl called\n"));
1691 * interrupt cannot be masked when coming here
1694 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1698 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1700 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1701 task_pid_nr(current),
1704 ctx->ctx_async_queue, ret));
1710 pfm_fasync(int fd, struct file *filp, int on)
1715 if (PFM_IS_FILE(filp) == 0) {
1716 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1720 ctx = filp->private_data;
1722 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1726 * we cannot mask interrupts during this call because this may
1727 * may go to sleep if memory is not readily avalaible.
1729 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1730 * done in caller. Serialization of this function is ensured by caller.
1732 ret = pfm_do_fasync(fd, filp, ctx, on);
1735 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1738 ctx->ctx_async_queue, ret));
1745 * this function is exclusively called from pfm_close().
1746 * The context is not protected at that time, nor are interrupts
1747 * on the remote CPU. That's necessary to avoid deadlocks.
1750 pfm_syswide_force_stop(void *info)
1752 pfm_context_t *ctx = (pfm_context_t *)info;
1753 struct pt_regs *regs = task_pt_regs(current);
1754 struct task_struct *owner;
1755 unsigned long flags;
1758 if (ctx->ctx_cpu != smp_processor_id()) {
1759 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1761 smp_processor_id());
1764 owner = GET_PMU_OWNER();
1765 if (owner != ctx->ctx_task) {
1766 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1768 task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1771 if (GET_PMU_CTX() != ctx) {
1772 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1774 GET_PMU_CTX(), ctx);
1778 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1780 * the context is already protected in pfm_close(), we simply
1781 * need to mask interrupts to avoid a PMU interrupt race on
1784 local_irq_save(flags);
1786 ret = pfm_context_unload(ctx, NULL, 0, regs);
1788 DPRINT(("context_unload returned %d\n", ret));
1792 * unmask interrupts, PMU interrupts are now spurious here
1794 local_irq_restore(flags);
1798 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1802 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1803 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1804 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1806 #endif /* CONFIG_SMP */
1809 * called for each close(). Partially free resources.
1810 * When caller is self-monitoring, the context is unloaded.
1813 pfm_flush(struct file *filp, fl_owner_t id)
1816 struct task_struct *task;
1817 struct pt_regs *regs;
1818 unsigned long flags;
1819 unsigned long smpl_buf_size = 0UL;
1820 void *smpl_buf_vaddr = NULL;
1821 int state, is_system;
1823 if (PFM_IS_FILE(filp) == 0) {
1824 DPRINT(("bad magic for\n"));
1828 ctx = filp->private_data;
1830 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1835 * remove our file from the async queue, if we use this mode.
1836 * This can be done without the context being protected. We come
1837 * here when the context has become unreachable by other tasks.
1839 * We may still have active monitoring at this point and we may
1840 * end up in pfm_overflow_handler(). However, fasync_helper()
1841 * operates with interrupts disabled and it cleans up the
1842 * queue. If the PMU handler is called prior to entering
1843 * fasync_helper() then it will send a signal. If it is
1844 * invoked after, it will find an empty queue and no
1845 * signal will be sent. In both case, we are safe
1847 PROTECT_CTX(ctx, flags);
1849 state = ctx->ctx_state;
1850 is_system = ctx->ctx_fl_system;
1852 task = PFM_CTX_TASK(ctx);
1853 regs = task_pt_regs(task);
1855 DPRINT(("ctx_state=%d is_current=%d\n",
1857 task == current ? 1 : 0));
1860 * if state == UNLOADED, then task is NULL
1864 * we must stop and unload because we are losing access to the context.
1866 if (task == current) {
1869 * the task IS the owner but it migrated to another CPU: that's bad
1870 * but we must handle this cleanly. Unfortunately, the kernel does
1871 * not provide a mechanism to block migration (while the context is loaded).
1873 * We need to release the resource on the ORIGINAL cpu.
1875 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1877 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1879 * keep context protected but unmask interrupt for IPI
1881 local_irq_restore(flags);
1883 pfm_syswide_cleanup_other_cpu(ctx);
1886 * restore interrupt masking
1888 local_irq_save(flags);
1891 * context is unloaded at this point
1894 #endif /* CONFIG_SMP */
1897 DPRINT(("forcing unload\n"));
1899 * stop and unload, returning with state UNLOADED
1900 * and session unreserved.
1902 pfm_context_unload(ctx, NULL, 0, regs);
1904 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1909 * remove virtual mapping, if any, for the calling task.
1910 * cannot reset ctx field until last user is calling close().
1912 * ctx_smpl_vaddr must never be cleared because it is needed
1913 * by every task with access to the context
1915 * When called from do_exit(), the mm context is gone already, therefore
1916 * mm is NULL, i.e., the VMA is already gone and we do not have to
1919 if (ctx->ctx_smpl_vaddr && current->mm) {
1920 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1921 smpl_buf_size = ctx->ctx_smpl_size;
1924 UNPROTECT_CTX(ctx, flags);
1927 * if there was a mapping, then we systematically remove it
1928 * at this point. Cannot be done inside critical section
1929 * because some VM function reenables interrupts.
1932 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(smpl_buf_vaddr, smpl_buf_size);
1937 * called either on explicit close() or from exit_files().
1938 * Only the LAST user of the file gets to this point, i.e., it is
1941 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1942 * (fput()),i.e, last task to access the file. Nobody else can access the
1943 * file at this point.
1945 * When called from exit_files(), the VMA has been freed because exit_mm()
1946 * is executed before exit_files().
1948 * When called from exit_files(), the current task is not yet ZOMBIE but we
1949 * flush the PMU state to the context.
1952 pfm_close(struct inode *inode, struct file *filp)
1955 struct task_struct *task;
1956 struct pt_regs *regs;
1957 DECLARE_WAITQUEUE(wait, current);
1958 unsigned long flags;
1959 unsigned long smpl_buf_size = 0UL;
1960 void *smpl_buf_addr = NULL;
1961 int free_possible = 1;
1962 int state, is_system;
1964 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1966 if (PFM_IS_FILE(filp) == 0) {
1967 DPRINT(("bad magic\n"));
1971 ctx = filp->private_data;
1973 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1977 PROTECT_CTX(ctx, flags);
1979 state = ctx->ctx_state;
1980 is_system = ctx->ctx_fl_system;
1982 task = PFM_CTX_TASK(ctx);
1983 regs = task_pt_regs(task);
1985 DPRINT(("ctx_state=%d is_current=%d\n",
1987 task == current ? 1 : 0));
1990 * if task == current, then pfm_flush() unloaded the context
1992 if (state == PFM_CTX_UNLOADED) goto doit;
1995 * context is loaded/masked and task != current, we need to
1996 * either force an unload or go zombie
2000 * The task is currently blocked or will block after an overflow.
2001 * we must force it to wakeup to get out of the
2002 * MASKED state and transition to the unloaded state by itself.
2004 * This situation is only possible for per-task mode
2006 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2009 * set a "partial" zombie state to be checked
2010 * upon return from down() in pfm_handle_work().
2012 * We cannot use the ZOMBIE state, because it is checked
2013 * by pfm_load_regs() which is called upon wakeup from down().
2014 * In such case, it would free the context and then we would
2015 * return to pfm_handle_work() which would access the
2016 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2017 * but visible to pfm_handle_work().
2019 * For some window of time, we have a zombie context with
2020 * ctx_state = MASKED and not ZOMBIE
2022 ctx->ctx_fl_going_zombie = 1;
2025 * force task to wake up from MASKED state
2027 complete(&ctx->ctx_restart_done);
2029 DPRINT(("waking up ctx_state=%d\n", state));
2032 * put ourself to sleep waiting for the other
2033 * task to report completion
2035 * the context is protected by mutex, therefore there
2036 * is no risk of being notified of completion before
2037 * begin actually on the waitq.
2039 set_current_state(TASK_INTERRUPTIBLE);
2040 add_wait_queue(&ctx->ctx_zombieq, &wait);
2042 UNPROTECT_CTX(ctx, flags);
2045 * XXX: check for signals :
2046 * - ok for explicit close
2047 * - not ok when coming from exit_files()
2052 PROTECT_CTX(ctx, flags);
2055 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2056 set_current_state(TASK_RUNNING);
2059 * context is unloaded at this point
2061 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2063 else if (task != current) {
2066 * switch context to zombie state
2068 ctx->ctx_state = PFM_CTX_ZOMBIE;
2070 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2072 * cannot free the context on the spot. deferred until
2073 * the task notices the ZOMBIE state
2077 pfm_context_unload(ctx, NULL, 0, regs);
2082 /* reload state, may have changed during opening of critical section */
2083 state = ctx->ctx_state;
2086 * the context is still attached to a task (possibly current)
2087 * we cannot destroy it right now
2091 * we must free the sampling buffer right here because
2092 * we cannot rely on it being cleaned up later by the
2093 * monitored task. It is not possible to free vmalloc'ed
2094 * memory in pfm_load_regs(). Instead, we remove the buffer
2095 * now. should there be subsequent PMU overflow originally
2096 * meant for sampling, the will be converted to spurious
2097 * and that's fine because the monitoring tools is gone anyway.
2099 if (ctx->ctx_smpl_hdr) {
2100 smpl_buf_addr = ctx->ctx_smpl_hdr;
2101 smpl_buf_size = ctx->ctx_smpl_size;
2102 /* no more sampling */
2103 ctx->ctx_smpl_hdr = NULL;
2104 ctx->ctx_fl_is_sampling = 0;
2107 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2113 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2116 * UNLOADED that the session has already been unreserved.
2118 if (state == PFM_CTX_ZOMBIE) {
2119 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2123 * disconnect file descriptor from context must be done
2126 filp->private_data = NULL;
2129 * if we free on the spot, the context is now completely unreachable
2130 * from the callers side. The monitored task side is also cut, so we
2133 * If we have a deferred free, only the caller side is disconnected.
2135 UNPROTECT_CTX(ctx, flags);
2138 * All memory free operations (especially for vmalloc'ed memory)
2139 * MUST be done with interrupts ENABLED.
2141 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2144 * return the memory used by the context
2146 if (free_possible) pfm_context_free(ctx);
2152 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2154 DPRINT(("pfm_no_open called\n"));
2160 static const struct file_operations pfm_file_ops = {
2161 .llseek = no_llseek,
2165 .unlocked_ioctl = pfm_ioctl,
2166 .open = pfm_no_open, /* special open code to disallow open via /proc */
2167 .fasync = pfm_fasync,
2168 .release = pfm_close,
2173 pfmfs_delete_dentry(const struct dentry *dentry)
2178 static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2180 return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2181 dentry->d_inode->i_ino);
2184 static const struct dentry_operations pfmfs_dentry_operations = {
2185 .d_delete = pfmfs_delete_dentry,
2186 .d_dname = pfmfs_dname,
2190 static struct file *
2191 pfm_alloc_file(pfm_context_t *ctx)
2194 struct inode *inode;
2196 struct qstr this = { .name = "" };
2199 * allocate a new inode
2201 inode = new_inode(pfmfs_mnt->mnt_sb);
2203 return ERR_PTR(-ENOMEM);
2205 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2207 inode->i_mode = S_IFCHR|S_IRUGO;
2208 inode->i_uid = current_fsuid();
2209 inode->i_gid = current_fsgid();
2212 * allocate a new dcache entry
2214 path.dentry = d_alloc(pfmfs_mnt->mnt_root, &this);
2217 return ERR_PTR(-ENOMEM);
2219 path.mnt = mntget(pfmfs_mnt);
2221 d_add(path.dentry, inode);
2223 file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2226 return ERR_PTR(-ENFILE);
2229 file->f_flags = O_RDONLY;
2230 file->private_data = ctx;
2236 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2238 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2241 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2244 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2255 * allocate a sampling buffer and remaps it into the user address space of the task
2258 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2260 struct mm_struct *mm = task->mm;
2261 struct vm_area_struct *vma = NULL;
2267 * the fixed header + requested size and align to page boundary
2269 size = PAGE_ALIGN(rsize);
2271 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2274 * check requested size to avoid Denial-of-service attacks
2275 * XXX: may have to refine this test
2276 * Check against address space limit.
2278 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2281 if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2285 * We do the easy to undo allocations first.
2287 * pfm_rvmalloc(), clears the buffer, so there is no leak
2289 smpl_buf = pfm_rvmalloc(size);
2290 if (smpl_buf == NULL) {
2291 DPRINT(("Can't allocate sampling buffer\n"));
2295 DPRINT(("smpl_buf @%p\n", smpl_buf));
2298 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2300 DPRINT(("Cannot allocate vma\n"));
2303 INIT_LIST_HEAD(&vma->anon_vma_chain);
2306 * partially initialize the vma for the sampling buffer
2309 vma->vm_file = filp;
2310 vma->vm_flags = VM_READ | VM_MAYREAD | VM_DONTEXPAND | VM_DONTDUMP;
2311 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2314 * Now we have everything we need and we can initialize
2315 * and connect all the data structures
2318 ctx->ctx_smpl_hdr = smpl_buf;
2319 ctx->ctx_smpl_size = size; /* aligned size */
2322 * Let's do the difficult operations next.
2324 * now we atomically find some area in the address space and
2325 * remap the buffer in it.
2327 down_write(&task->mm->mmap_sem);
2329 /* find some free area in address space, must have mmap sem held */
2330 vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2331 if (IS_ERR_VALUE(vma->vm_start)) {
2332 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2333 up_write(&task->mm->mmap_sem);
2336 vma->vm_end = vma->vm_start + size;
2337 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2339 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2341 /* can only be applied to current task, need to have the mm semaphore held when called */
2342 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2343 DPRINT(("Can't remap buffer\n"));
2344 up_write(&task->mm->mmap_sem);
2351 * now insert the vma in the vm list for the process, must be
2352 * done with mmap lock held
2354 insert_vm_struct(mm, vma);
2356 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2358 up_write(&task->mm->mmap_sem);
2361 * keep track of user level virtual address
2363 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2364 *(unsigned long *)user_vaddr = vma->vm_start;
2369 kmem_cache_free(vm_area_cachep, vma);
2371 pfm_rvfree(smpl_buf, size);
2377 * XXX: do something better here
2380 pfm_bad_permissions(struct task_struct *task)
2382 const struct cred *tcred;
2383 uid_t uid = current_uid();
2384 gid_t gid = current_gid();
2388 tcred = __task_cred(task);
2390 /* inspired by ptrace_attach() */
2391 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2400 ret = ((uid != tcred->euid)
2401 || (uid != tcred->suid)
2402 || (uid != tcred->uid)
2403 || (gid != tcred->egid)
2404 || (gid != tcred->sgid)
2405 || (gid != tcred->gid)) && !capable(CAP_SYS_PTRACE);
2412 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2418 ctx_flags = pfx->ctx_flags;
2420 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2423 * cannot block in this mode
2425 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2426 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2431 /* probably more to add here */
2437 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2438 unsigned int cpu, pfarg_context_t *arg)
2440 pfm_buffer_fmt_t *fmt = NULL;
2441 unsigned long size = 0UL;
2443 void *fmt_arg = NULL;
2445 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2447 /* invoke and lock buffer format, if found */
2448 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2450 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2455 * buffer argument MUST be contiguous to pfarg_context_t
2457 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2459 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2461 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2463 if (ret) goto error;
2465 /* link buffer format and context */
2466 ctx->ctx_buf_fmt = fmt;
2467 ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2470 * check if buffer format wants to use perfmon buffer allocation/mapping service
2472 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2473 if (ret) goto error;
2477 * buffer is always remapped into the caller's address space
2479 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2480 if (ret) goto error;
2482 /* keep track of user address of buffer */
2483 arg->ctx_smpl_vaddr = uaddr;
2485 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2492 pfm_reset_pmu_state(pfm_context_t *ctx)
2497 * install reset values for PMC.
2499 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2500 if (PMC_IS_IMPL(i) == 0) continue;
2501 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2502 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2505 * PMD registers are set to 0UL when the context in memset()
2509 * On context switched restore, we must restore ALL pmc and ALL pmd even
2510 * when they are not actively used by the task. In UP, the incoming process
2511 * may otherwise pick up left over PMC, PMD state from the previous process.
2512 * As opposed to PMD, stale PMC can cause harm to the incoming
2513 * process because they may change what is being measured.
2514 * Therefore, we must systematically reinstall the entire
2515 * PMC state. In SMP, the same thing is possible on the
2516 * same CPU but also on between 2 CPUs.
2518 * The problem with PMD is information leaking especially
2519 * to user level when psr.sp=0
2521 * There is unfortunately no easy way to avoid this problem
2522 * on either UP or SMP. This definitively slows down the
2523 * pfm_load_regs() function.
2527 * bitmask of all PMCs accessible to this context
2529 * PMC0 is treated differently.
2531 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2534 * bitmask of all PMDs that are accessible to this context
2536 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2538 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2541 * useful in case of re-enable after disable
2543 ctx->ctx_used_ibrs[0] = 0UL;
2544 ctx->ctx_used_dbrs[0] = 0UL;
2548 pfm_ctx_getsize(void *arg, size_t *sz)
2550 pfarg_context_t *req = (pfarg_context_t *)arg;
2551 pfm_buffer_fmt_t *fmt;
2555 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2557 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2559 DPRINT(("cannot find buffer format\n"));
2562 /* get just enough to copy in user parameters */
2563 *sz = fmt->fmt_arg_size;
2564 DPRINT(("arg_size=%lu\n", *sz));
2572 * cannot attach if :
2574 * - task not owned by caller
2575 * - task incompatible with context mode
2578 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2581 * no kernel task or task not owner by caller
2583 if (task->mm == NULL) {
2584 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2587 if (pfm_bad_permissions(task)) {
2588 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task)));
2592 * cannot block in self-monitoring mode
2594 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2595 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2599 if (task->exit_state == EXIT_ZOMBIE) {
2600 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task)));
2605 * always ok for self
2607 if (task == current) return 0;
2609 if (!task_is_stopped_or_traced(task)) {
2610 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2614 * make sure the task is off any CPU
2616 wait_task_inactive(task, 0);
2618 /* more to come... */
2624 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2626 struct task_struct *p = current;
2629 /* XXX: need to add more checks here */
2630 if (pid < 2) return -EPERM;
2632 if (pid != task_pid_vnr(current)) {
2634 read_lock(&tasklist_lock);
2636 p = find_task_by_vpid(pid);
2638 /* make sure task cannot go away while we operate on it */
2639 if (p) get_task_struct(p);
2641 read_unlock(&tasklist_lock);
2643 if (p == NULL) return -ESRCH;
2646 ret = pfm_task_incompatible(ctx, p);
2649 } else if (p != current) {
2658 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2660 pfarg_context_t *req = (pfarg_context_t *)arg;
2667 /* let's check the arguments first */
2668 ret = pfarg_is_sane(current, req);
2672 ctx_flags = req->ctx_flags;
2676 fd = get_unused_fd();
2680 ctx = pfm_context_alloc(ctx_flags);
2684 filp = pfm_alloc_file(ctx);
2686 ret = PTR_ERR(filp);
2690 req->ctx_fd = ctx->ctx_fd = fd;
2693 * does the user want to sample?
2695 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2696 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2701 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2706 ctx->ctx_fl_excl_idle,
2711 * initialize soft PMU state
2713 pfm_reset_pmu_state(ctx);
2715 fd_install(fd, filp);
2720 path = filp->f_path;
2724 if (ctx->ctx_buf_fmt) {
2725 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2728 pfm_context_free(ctx);
2735 static inline unsigned long
2736 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2738 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2739 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2740 extern unsigned long carta_random32 (unsigned long seed);
2742 if (reg->flags & PFM_REGFL_RANDOM) {
2743 new_seed = carta_random32(old_seed);
2744 val -= (old_seed & mask); /* counter values are negative numbers! */
2745 if ((mask >> 32) != 0)
2746 /* construct a full 64-bit random value: */
2747 new_seed |= carta_random32(old_seed >> 32) << 32;
2748 reg->seed = new_seed;
2755 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2757 unsigned long mask = ovfl_regs[0];
2758 unsigned long reset_others = 0UL;
2763 * now restore reset value on sampling overflowed counters
2765 mask >>= PMU_FIRST_COUNTER;
2766 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2768 if ((mask & 0x1UL) == 0UL) continue;
2770 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2771 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2773 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2777 * Now take care of resetting the other registers
2779 for(i = 0; reset_others; i++, reset_others >>= 1) {
2781 if ((reset_others & 0x1) == 0) continue;
2783 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2785 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2786 is_long_reset ? "long" : "short", i, val));
2791 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2793 unsigned long mask = ovfl_regs[0];
2794 unsigned long reset_others = 0UL;
2798 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2800 if (ctx->ctx_state == PFM_CTX_MASKED) {
2801 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2806 * now restore reset value on sampling overflowed counters
2808 mask >>= PMU_FIRST_COUNTER;
2809 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2811 if ((mask & 0x1UL) == 0UL) continue;
2813 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2814 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2816 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2818 pfm_write_soft_counter(ctx, i, val);
2822 * Now take care of resetting the other registers
2824 for(i = 0; reset_others; i++, reset_others >>= 1) {
2826 if ((reset_others & 0x1) == 0) continue;
2828 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2830 if (PMD_IS_COUNTING(i)) {
2831 pfm_write_soft_counter(ctx, i, val);
2833 ia64_set_pmd(i, val);
2835 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2836 is_long_reset ? "long" : "short", i, val));
2842 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2844 struct task_struct *task;
2845 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2846 unsigned long value, pmc_pm;
2847 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2848 unsigned int cnum, reg_flags, flags, pmc_type;
2849 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2850 int is_monitor, is_counting, state;
2852 pfm_reg_check_t wr_func;
2853 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2855 state = ctx->ctx_state;
2856 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2857 is_system = ctx->ctx_fl_system;
2858 task = ctx->ctx_task;
2859 impl_pmds = pmu_conf->impl_pmds[0];
2861 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2865 * In system wide and when the context is loaded, access can only happen
2866 * when the caller is running on the CPU being monitored by the session.
2867 * It does not have to be the owner (ctx_task) of the context per se.
2869 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2870 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2873 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2875 expert_mode = pfm_sysctl.expert_mode;
2877 for (i = 0; i < count; i++, req++) {
2879 cnum = req->reg_num;
2880 reg_flags = req->reg_flags;
2881 value = req->reg_value;
2882 smpl_pmds = req->reg_smpl_pmds[0];
2883 reset_pmds = req->reg_reset_pmds[0];
2887 if (cnum >= PMU_MAX_PMCS) {
2888 DPRINT(("pmc%u is invalid\n", cnum));
2892 pmc_type = pmu_conf->pmc_desc[cnum].type;
2893 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2894 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2895 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2898 * we reject all non implemented PMC as well
2899 * as attempts to modify PMC[0-3] which are used
2900 * as status registers by the PMU
2902 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2903 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2906 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2908 * If the PMC is a monitor, then if the value is not the default:
2909 * - system-wide session: PMCx.pm=1 (privileged monitor)
2910 * - per-task : PMCx.pm=0 (user monitor)
2912 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2913 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2922 * enforce generation of overflow interrupt. Necessary on all
2925 value |= 1 << PMU_PMC_OI;
2927 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2928 flags |= PFM_REGFL_OVFL_NOTIFY;
2931 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2933 /* verify validity of smpl_pmds */
2934 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2935 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2939 /* verify validity of reset_pmds */
2940 if ((reset_pmds & impl_pmds) != reset_pmds) {
2941 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2945 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2946 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2949 /* eventid on non-counting monitors are ignored */
2953 * execute write checker, if any
2955 if (likely(expert_mode == 0 && wr_func)) {
2956 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2957 if (ret) goto error;
2962 * no error on this register
2964 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2967 * Now we commit the changes to the software state
2971 * update overflow information
2975 * full flag update each time a register is programmed
2977 ctx->ctx_pmds[cnum].flags = flags;
2979 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2980 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2981 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2984 * Mark all PMDS to be accessed as used.
2986 * We do not keep track of PMC because we have to
2987 * systematically restore ALL of them.
2989 * We do not update the used_monitors mask, because
2990 * if we have not programmed them, then will be in
2991 * a quiescent state, therefore we will not need to
2992 * mask/restore then when context is MASKED.
2994 CTX_USED_PMD(ctx, reset_pmds);
2995 CTX_USED_PMD(ctx, smpl_pmds);
2997 * make sure we do not try to reset on
2998 * restart because we have established new values
3000 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3003 * Needed in case the user does not initialize the equivalent
3004 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3005 * possible leak here.
3007 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3010 * keep track of the monitor PMC that we are using.
3011 * we save the value of the pmc in ctx_pmcs[] and if
3012 * the monitoring is not stopped for the context we also
3013 * place it in the saved state area so that it will be
3014 * picked up later by the context switch code.
3016 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3018 * The value in th_pmcs[] may be modified on overflow, i.e., when
3019 * monitoring needs to be stopped.
3021 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3024 * update context state
3026 ctx->ctx_pmcs[cnum] = value;
3030 * write thread state
3032 if (is_system == 0) ctx->th_pmcs[cnum] = value;
3035 * write hardware register if we can
3037 if (can_access_pmu) {
3038 ia64_set_pmc(cnum, value);
3043 * per-task SMP only here
3045 * we are guaranteed that the task is not running on the other CPU,
3046 * we indicate that this PMD will need to be reloaded if the task
3047 * is rescheduled on the CPU it ran last on.
3049 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3054 DPRINT(("pmc[%u]=0x%lx ld=%d apmu=%d flags=0x%x all_pmcs=0x%lx used_pmds=0x%lx eventid=%ld smpl_pmds=0x%lx reset_pmds=0x%lx reloads_pmcs=0x%lx used_monitors=0x%lx ovfl_regs=0x%lx\n",
3060 ctx->ctx_all_pmcs[0],
3061 ctx->ctx_used_pmds[0],
3062 ctx->ctx_pmds[cnum].eventid,
3065 ctx->ctx_reload_pmcs[0],
3066 ctx->ctx_used_monitors[0],
3067 ctx->ctx_ovfl_regs[0]));
3071 * make sure the changes are visible
3073 if (can_access_pmu) ia64_srlz_d();
3077 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3082 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3084 struct task_struct *task;
3085 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3086 unsigned long value, hw_value, ovfl_mask;
3088 int i, can_access_pmu = 0, state;
3089 int is_counting, is_loaded, is_system, expert_mode;
3091 pfm_reg_check_t wr_func;
3094 state = ctx->ctx_state;
3095 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3096 is_system = ctx->ctx_fl_system;
3097 ovfl_mask = pmu_conf->ovfl_val;
3098 task = ctx->ctx_task;
3100 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3103 * on both UP and SMP, we can only write to the PMC when the task is
3104 * the owner of the local PMU.
3106 if (likely(is_loaded)) {
3108 * In system wide and when the context is loaded, access can only happen
3109 * when the caller is running on the CPU being monitored by the session.
3110 * It does not have to be the owner (ctx_task) of the context per se.
3112 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3113 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3116 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3118 expert_mode = pfm_sysctl.expert_mode;
3120 for (i = 0; i < count; i++, req++) {
3122 cnum = req->reg_num;
3123 value = req->reg_value;
3125 if (!PMD_IS_IMPL(cnum)) {
3126 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3129 is_counting = PMD_IS_COUNTING(cnum);
3130 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3133 * execute write checker, if any
3135 if (unlikely(expert_mode == 0 && wr_func)) {
3136 unsigned long v = value;
3138 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3139 if (ret) goto abort_mission;
3146 * no error on this register
3148 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3151 * now commit changes to software state
3156 * update virtualized (64bits) counter
3160 * write context state
3162 ctx->ctx_pmds[cnum].lval = value;
3165 * when context is load we use the split value
3168 hw_value = value & ovfl_mask;
3169 value = value & ~ovfl_mask;
3173 * update reset values (not just for counters)
3175 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3176 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3179 * update randomization parameters (not just for counters)
3181 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3182 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3185 * update context value
3187 ctx->ctx_pmds[cnum].val = value;
3190 * Keep track of what we use
3192 * We do not keep track of PMC because we have to
3193 * systematically restore ALL of them.
3195 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3198 * mark this PMD register used as well
3200 CTX_USED_PMD(ctx, RDEP(cnum));
3203 * make sure we do not try to reset on
3204 * restart because we have established new values
3206 if (is_counting && state == PFM_CTX_MASKED) {
3207 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3212 * write thread state
3214 if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3217 * write hardware register if we can
3219 if (can_access_pmu) {
3220 ia64_set_pmd(cnum, hw_value);
3224 * we are guaranteed that the task is not running on the other CPU,
3225 * we indicate that this PMD will need to be reloaded if the task
3226 * is rescheduled on the CPU it ran last on.
3228 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3233 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3234 "long_reset=0x%lx notify=%c seed=0x%lx mask=0x%lx used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3240 ctx->ctx_pmds[cnum].val,
3241 ctx->ctx_pmds[cnum].short_reset,
3242 ctx->ctx_pmds[cnum].long_reset,
3243 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3244 ctx->ctx_pmds[cnum].seed,
3245 ctx->ctx_pmds[cnum].mask,
3246 ctx->ctx_used_pmds[0],
3247 ctx->ctx_pmds[cnum].reset_pmds[0],
3248 ctx->ctx_reload_pmds[0],
3249 ctx->ctx_all_pmds[0],
3250 ctx->ctx_ovfl_regs[0]));
3254 * make changes visible
3256 if (can_access_pmu) ia64_srlz_d();
3262 * for now, we have only one possibility for error
3264 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3269 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3270 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3271 * interrupt is delivered during the call, it will be kept pending until we leave, making
3272 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3273 * guaranteed to return consistent data to the user, it may simply be old. It is not
3274 * trivial to treat the overflow while inside the call because you may end up in
3275 * some module sampling buffer code causing deadlocks.
3278 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3280 struct task_struct *task;
3281 unsigned long val = 0UL, lval, ovfl_mask, sval;
3282 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3283 unsigned int cnum, reg_flags = 0;
3284 int i, can_access_pmu = 0, state;
3285 int is_loaded, is_system, is_counting, expert_mode;
3287 pfm_reg_check_t rd_func;
3290 * access is possible when loaded only for
3291 * self-monitoring tasks or in UP mode
3294 state = ctx->ctx_state;
3295 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3296 is_system = ctx->ctx_fl_system;
3297 ovfl_mask = pmu_conf->ovfl_val;
3298 task = ctx->ctx_task;
3300 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3302 if (likely(is_loaded)) {
3304 * In system wide and when the context is loaded, access can only happen
3305 * when the caller is running on the CPU being monitored by the session.
3306 * It does not have to be the owner (ctx_task) of the context per se.
3308 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3309 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3313 * this can be true when not self-monitoring only in UP
3315 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3317 if (can_access_pmu) ia64_srlz_d();
3319 expert_mode = pfm_sysctl.expert_mode;
3321 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3327 * on both UP and SMP, we can only read the PMD from the hardware register when
3328 * the task is the owner of the local PMU.
3331 for (i = 0; i < count; i++, req++) {
3333 cnum = req->reg_num;
3334 reg_flags = req->reg_flags;
3336 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3338 * we can only read the register that we use. That includes
3339 * the one we explicitly initialize AND the one we want included
3340 * in the sampling buffer (smpl_regs).
3342 * Having this restriction allows optimization in the ctxsw routine
3343 * without compromising security (leaks)
3345 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3347 sval = ctx->ctx_pmds[cnum].val;
3348 lval = ctx->ctx_pmds[cnum].lval;
3349 is_counting = PMD_IS_COUNTING(cnum);
3352 * If the task is not the current one, then we check if the
3353 * PMU state is still in the local live register due to lazy ctxsw.
3354 * If true, then we read directly from the registers.
3356 if (can_access_pmu){
3357 val = ia64_get_pmd(cnum);
3360 * context has been saved
3361 * if context is zombie, then task does not exist anymore.
3362 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3364 val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3366 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3370 * XXX: need to check for overflow when loaded
3377 * execute read checker, if any
3379 if (unlikely(expert_mode == 0 && rd_func)) {
3380 unsigned long v = val;
3381 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3382 if (ret) goto error;
3387 PFM_REG_RETFLAG_SET(reg_flags, 0);
3389 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3392 * update register return value, abort all if problem during copy.
3393 * we only modify the reg_flags field. no check mode is fine because
3394 * access has been verified upfront in sys_perfmonctl().
3396 req->reg_value = val;
3397 req->reg_flags = reg_flags;
3398 req->reg_last_reset_val = lval;
3404 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3409 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3413 if (req == NULL) return -EINVAL;
3415 ctx = GET_PMU_CTX();
3417 if (ctx == NULL) return -EINVAL;
3420 * for now limit to current task, which is enough when calling
3421 * from overflow handler
3423 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3425 return pfm_write_pmcs(ctx, req, nreq, regs);
3427 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3430 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3434 if (req == NULL) return -EINVAL;
3436 ctx = GET_PMU_CTX();
3438 if (ctx == NULL) return -EINVAL;
3441 * for now limit to current task, which is enough when calling
3442 * from overflow handler
3444 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3446 return pfm_read_pmds(ctx, req, nreq, regs);
3448 EXPORT_SYMBOL(pfm_mod_read_pmds);
3451 * Only call this function when a process it trying to
3452 * write the debug registers (reading is always allowed)
3455 pfm_use_debug_registers(struct task_struct *task)
3457 pfm_context_t *ctx = task->thread.pfm_context;
3458 unsigned long flags;
3461 if (pmu_conf->use_rr_dbregs == 0) return 0;
3463 DPRINT(("called for [%d]\n", task_pid_nr(task)));
3468 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3471 * Even on SMP, we do not need to use an atomic here because
3472 * the only way in is via ptrace() and this is possible only when the
3473 * process is stopped. Even in the case where the ctxsw out is not totally
3474 * completed by the time we come here, there is no way the 'stopped' process
3475 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3476 * So this is always safe.
3478 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3483 * We cannot allow setting breakpoints when system wide monitoring
3484 * sessions are using the debug registers.
3486 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3489 pfm_sessions.pfs_ptrace_use_dbregs++;
3491 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3492 pfm_sessions.pfs_ptrace_use_dbregs,
3493 pfm_sessions.pfs_sys_use_dbregs,
3494 task_pid_nr(task), ret));
3502 * This function is called for every task that exits with the
3503 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3504 * able to use the debug registers for debugging purposes via
3505 * ptrace(). Therefore we know it was not using them for
3506 * performance monitoring, so we only decrement the number
3507 * of "ptraced" debug register users to keep the count up to date
3510 pfm_release_debug_registers(struct task_struct *task)
3512 unsigned long flags;
3515 if (pmu_conf->use_rr_dbregs == 0) return 0;
3518 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3519 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3522 pfm_sessions.pfs_ptrace_use_dbregs--;
3531 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3533 struct task_struct *task;
3534 pfm_buffer_fmt_t *fmt;
3535 pfm_ovfl_ctrl_t rst_ctrl;
3536 int state, is_system;
3539 state = ctx->ctx_state;
3540 fmt = ctx->ctx_buf_fmt;
3541 is_system = ctx->ctx_fl_system;
3542 task = PFM_CTX_TASK(ctx);
3545 case PFM_CTX_MASKED:
3547 case PFM_CTX_LOADED:
3548 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3550 case PFM_CTX_UNLOADED:
3551 case PFM_CTX_ZOMBIE:
3552 DPRINT(("invalid state=%d\n", state));
3555 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3560 * In system wide and when the context is loaded, access can only happen
3561 * when the caller is running on the CPU being monitored by the session.
3562 * It does not have to be the owner (ctx_task) of the context per se.
3564 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3565 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3570 if (unlikely(task == NULL)) {
3571 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3575 if (task == current || is_system) {
3577 fmt = ctx->ctx_buf_fmt;
3579 DPRINT(("restarting self %d ovfl=0x%lx\n",
3581 ctx->ctx_ovfl_regs[0]));
3583 if (CTX_HAS_SMPL(ctx)) {
3585 prefetch(ctx->ctx_smpl_hdr);
3587 rst_ctrl.bits.mask_monitoring = 0;
3588 rst_ctrl.bits.reset_ovfl_pmds = 0;
3590 if (state == PFM_CTX_LOADED)
3591 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3593 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3595 rst_ctrl.bits.mask_monitoring = 0;
3596 rst_ctrl.bits.reset_ovfl_pmds = 1;
3600 if (rst_ctrl.bits.reset_ovfl_pmds)
3601 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3603 if (rst_ctrl.bits.mask_monitoring == 0) {
3604 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3606 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3608 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3610 // cannot use pfm_stop_monitoring(task, regs);
3614 * clear overflowed PMD mask to remove any stale information
3616 ctx->ctx_ovfl_regs[0] = 0UL;
3619 * back to LOADED state
3621 ctx->ctx_state = PFM_CTX_LOADED;
3624 * XXX: not really useful for self monitoring
3626 ctx->ctx_fl_can_restart = 0;
3632 * restart another task
3636 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3637 * one is seen by the task.
3639 if (state == PFM_CTX_MASKED) {
3640 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3642 * will prevent subsequent restart before this one is
3643 * seen by other task
3645 ctx->ctx_fl_can_restart = 0;
3649 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3650 * the task is blocked or on its way to block. That's the normal
3651 * restart path. If the monitoring is not masked, then the task
3652 * can be actively monitoring and we cannot directly intervene.
3653 * Therefore we use the trap mechanism to catch the task and
3654 * force it to reset the buffer/reset PMDs.
3656 * if non-blocking, then we ensure that the task will go into
3657 * pfm_handle_work() before returning to user mode.
3659 * We cannot explicitly reset another task, it MUST always
3660 * be done by the task itself. This works for system wide because
3661 * the tool that is controlling the session is logically doing
3662 * "self-monitoring".
3664 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3665 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3666 complete(&ctx->ctx_restart_done);
3668 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3670 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3672 PFM_SET_WORK_PENDING(task, 1);
3674 set_notify_resume(task);
3677 * XXX: send reschedule if task runs on another CPU
3684 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3686 unsigned int m = *(unsigned int *)arg;
3688 pfm_sysctl.debug = m == 0 ? 0 : 1;
3690 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3693 memset(pfm_stats, 0, sizeof(pfm_stats));
3694 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3700 * arg can be NULL and count can be zero for this function
3703 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3705 struct thread_struct *thread = NULL;
3706 struct task_struct *task;
3707 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3708 unsigned long flags;
3713 int i, can_access_pmu = 0;
3714 int is_system, is_loaded;
3716 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3718 state = ctx->ctx_state;
3719 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3720 is_system = ctx->ctx_fl_system;
3721 task = ctx->ctx_task;
3723 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3726 * on both UP and SMP, we can only write to the PMC when the task is
3727 * the owner of the local PMU.
3730 thread = &task->thread;
3732 * In system wide and when the context is loaded, access can only happen
3733 * when the caller is running on the CPU being monitored by the session.
3734 * It does not have to be the owner (ctx_task) of the context per se.
3736 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3737 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3740 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3744 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3745 * ensuring that no real breakpoint can be installed via this call.
3747 * IMPORTANT: regs can be NULL in this function
3750 first_time = ctx->ctx_fl_using_dbreg == 0;
3753 * don't bother if we are loaded and task is being debugged
3755 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3756 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3761 * check for debug registers in system wide mode
3763 * If though a check is done in pfm_context_load(),
3764 * we must repeat it here, in case the registers are
3765 * written after the context is loaded
3770 if (first_time && is_system) {
3771 if (pfm_sessions.pfs_ptrace_use_dbregs)
3774 pfm_sessions.pfs_sys_use_dbregs++;
3779 if (ret != 0) return ret;
3782 * mark ourself as user of the debug registers for
3785 ctx->ctx_fl_using_dbreg = 1;
3788 * clear hardware registers to make sure we don't
3789 * pick up stale state.
3791 * for a system wide session, we do not use
3792 * thread.dbr, thread.ibr because this process
3793 * never leaves the current CPU and the state
3794 * is shared by all processes running on it
3796 if (first_time && can_access_pmu) {
3797 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3798 for (i=0; i < pmu_conf->num_ibrs; i++) {
3799 ia64_set_ibr(i, 0UL);
3800 ia64_dv_serialize_instruction();
3803 for (i=0; i < pmu_conf->num_dbrs; i++) {
3804 ia64_set_dbr(i, 0UL);
3805 ia64_dv_serialize_data();
3811 * Now install the values into the registers
3813 for (i = 0; i < count; i++, req++) {
3815 rnum = req->dbreg_num;
3816 dbreg.val = req->dbreg_value;
3820 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3821 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3822 rnum, dbreg.val, mode, i, count));
3828 * make sure we do not install enabled breakpoint
3831 if (mode == PFM_CODE_RR)
3832 dbreg.ibr.ibr_x = 0;
3834 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3837 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3840 * Debug registers, just like PMC, can only be modified
3841 * by a kernel call. Moreover, perfmon() access to those
3842 * registers are centralized in this routine. The hardware
3843 * does not modify the value of these registers, therefore,
3844 * if we save them as they are written, we can avoid having
3845 * to save them on context switch out. This is made possible
3846 * by the fact that when perfmon uses debug registers, ptrace()
3847 * won't be able to modify them concurrently.
3849 if (mode == PFM_CODE_RR) {
3850 CTX_USED_IBR(ctx, rnum);
3852 if (can_access_pmu) {
3853 ia64_set_ibr(rnum, dbreg.val);
3854 ia64_dv_serialize_instruction();
3857 ctx->ctx_ibrs[rnum] = dbreg.val;
3859 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3860 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3862 CTX_USED_DBR(ctx, rnum);
3864 if (can_access_pmu) {
3865 ia64_set_dbr(rnum, dbreg.val);
3866 ia64_dv_serialize_data();
3868 ctx->ctx_dbrs[rnum] = dbreg.val;
3870 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3871 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3879 * in case it was our first attempt, we undo the global modifications
3883 if (ctx->ctx_fl_system) {
3884 pfm_sessions.pfs_sys_use_dbregs--;
3887 ctx->ctx_fl_using_dbreg = 0;
3890 * install error return flag
3892 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3898 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3900 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3904 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3906 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3910 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3914 if (req == NULL) return -EINVAL;
3916 ctx = GET_PMU_CTX();
3918 if (ctx == NULL) return -EINVAL;
3921 * for now limit to current task, which is enough when calling
3922 * from overflow handler
3924 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3926 return pfm_write_ibrs(ctx, req, nreq, regs);
3928 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3931 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3935 if (req == NULL) return -EINVAL;
3937 ctx = GET_PMU_CTX();
3939 if (ctx == NULL) return -EINVAL;
3942 * for now limit to current task, which is enough when calling
3943 * from overflow handler
3945 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3947 return pfm_write_dbrs(ctx, req, nreq, regs);
3949 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3953 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3955 pfarg_features_t *req = (pfarg_features_t *)arg;
3957 req->ft_version = PFM_VERSION;
3962 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3964 struct pt_regs *tregs;
3965 struct task_struct *task = PFM_CTX_TASK(ctx);
3966 int state, is_system;
3968 state = ctx->ctx_state;
3969 is_system = ctx->ctx_fl_system;
3972 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3974 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3977 * In system wide and when the context is loaded, access can only happen
3978 * when the caller is running on the CPU being monitored by the session.
3979 * It does not have to be the owner (ctx_task) of the context per se.
3981 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3982 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3985 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3986 task_pid_nr(PFM_CTX_TASK(ctx)),
3990 * in system mode, we need to update the PMU directly
3991 * and the user level state of the caller, which may not
3992 * necessarily be the creator of the context.
3996 * Update local PMU first
4000 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4004 * update local cpuinfo
4006 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4009 * stop monitoring, does srlz.i
4014 * stop monitoring in the caller
4016 ia64_psr(regs)->pp = 0;
4024 if (task == current) {
4025 /* stop monitoring at kernel level */
4029 * stop monitoring at the user level
4031 ia64_psr(regs)->up = 0;
4033 tregs = task_pt_regs(task);
4036 * stop monitoring at the user level
4038 ia64_psr(tregs)->up = 0;
4041 * monitoring disabled in kernel at next reschedule
4043 ctx->ctx_saved_psr_up = 0;
4044 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4051 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4053 struct pt_regs *tregs;
4054 int state, is_system;
4056 state = ctx->ctx_state;
4057 is_system = ctx->ctx_fl_system;
4059 if (state != PFM_CTX_LOADED) return -EINVAL;
4062 * In system wide and when the context is loaded, access can only happen
4063 * when the caller is running on the CPU being monitored by the session.
4064 * It does not have to be the owner (ctx_task) of the context per se.
4066 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4067 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4072 * in system mode, we need to update the PMU directly
4073 * and the user level state of the caller, which may not
4074 * necessarily be the creator of the context.
4079 * set user level psr.pp for the caller
4081 ia64_psr(regs)->pp = 1;
4084 * now update the local PMU and cpuinfo
4086 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4089 * start monitoring at kernel level
4094 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4104 if (ctx->ctx_task == current) {
4106 /* start monitoring at kernel level */
4110 * activate monitoring at user level
4112 ia64_psr(regs)->up = 1;
4115 tregs = task_pt_regs(ctx->ctx_task);
4118 * start monitoring at the kernel level the next
4119 * time the task is scheduled
4121 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4124 * activate monitoring at user level
4126 ia64_psr(tregs)->up = 1;
4132 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4134 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4139 for (i = 0; i < count; i++, req++) {
4141 cnum = req->reg_num;
4143 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4145 req->reg_value = PMC_DFL_VAL(cnum);
4147 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4149 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4154 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4159 pfm_check_task_exist(pfm_context_t *ctx)
4161 struct task_struct *g, *t;
4164 read_lock(&tasklist_lock);
4166 do_each_thread (g, t) {
4167 if (t->thread.pfm_context == ctx) {
4171 } while_each_thread (g, t);
4173 read_unlock(&tasklist_lock);
4175 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4181 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4183 struct task_struct *task;
4184 struct thread_struct *thread;
4185 struct pfm_context_t *old;
4186 unsigned long flags;
4188 struct task_struct *owner_task = NULL;
4190 pfarg_load_t *req = (pfarg_load_t *)arg;
4191 unsigned long *pmcs_source, *pmds_source;
4194 int state, is_system, set_dbregs = 0;
4196 state = ctx->ctx_state;
4197 is_system = ctx->ctx_fl_system;
4199 * can only load from unloaded or terminated state
4201 if (state != PFM_CTX_UNLOADED) {
4202 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4208 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4210 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4211 DPRINT(("cannot use blocking mode on self\n"));
4215 ret = pfm_get_task(ctx, req->load_pid, &task);
4217 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4224 * system wide is self monitoring only
4226 if (is_system && task != current) {
4227 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4232 thread = &task->thread;
4236 * cannot load a context which is using range restrictions,
4237 * into a task that is being debugged.
4239 if (ctx->ctx_fl_using_dbreg) {
4240 if (thread->flags & IA64_THREAD_DBG_VALID) {
4242 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4248 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4249 DPRINT(("cannot load [%d] dbregs in use\n",
4250 task_pid_nr(task)));
4253 pfm_sessions.pfs_sys_use_dbregs++;
4254 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4261 if (ret) goto error;
4265 * SMP system-wide monitoring implies self-monitoring.
4267 * The programming model expects the task to
4268 * be pinned on a CPU throughout the session.
4269 * Here we take note of the current CPU at the
4270 * time the context is loaded. No call from
4271 * another CPU will be allowed.
4273 * The pinning via shed_setaffinity()
4274 * must be done by the calling task prior
4277 * systemwide: keep track of CPU this session is supposed to run on
4279 the_cpu = ctx->ctx_cpu = smp_processor_id();
4283 * now reserve the session
4285 ret = pfm_reserve_session(current, is_system, the_cpu);
4286 if (ret) goto error;
4289 * task is necessarily stopped at this point.
4291 * If the previous context was zombie, then it got removed in
4292 * pfm_save_regs(). Therefore we should not see it here.
4293 * If we see a context, then this is an active context
4295 * XXX: needs to be atomic
4297 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4298 thread->pfm_context, ctx));
4301 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4303 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4307 pfm_reset_msgq(ctx);
4309 ctx->ctx_state = PFM_CTX_LOADED;
4312 * link context to task
4314 ctx->ctx_task = task;
4318 * we load as stopped
4320 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4321 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4323 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4325 thread->flags |= IA64_THREAD_PM_VALID;
4329 * propagate into thread-state
4331 pfm_copy_pmds(task, ctx);
4332 pfm_copy_pmcs(task, ctx);
4334 pmcs_source = ctx->th_pmcs;
4335 pmds_source = ctx->th_pmds;
4338 * always the case for system-wide
4340 if (task == current) {
4342 if (is_system == 0) {
4344 /* allow user level control */
4345 ia64_psr(regs)->sp = 0;
4346 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4348 SET_LAST_CPU(ctx, smp_processor_id());
4350 SET_ACTIVATION(ctx);
4353 * push the other task out, if any
4355 owner_task = GET_PMU_OWNER();
4356 if (owner_task) pfm_lazy_save_regs(owner_task);
4360 * load all PMD from ctx to PMU (as opposed to thread state)
4361 * restore all PMC from ctx to PMU
4363 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4364 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4366 ctx->ctx_reload_pmcs[0] = 0UL;
4367 ctx->ctx_reload_pmds[0] = 0UL;
4370 * guaranteed safe by earlier check against DBG_VALID
4372 if (ctx->ctx_fl_using_dbreg) {
4373 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4374 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4379 SET_PMU_OWNER(task, ctx);
4381 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4384 * when not current, task MUST be stopped, so this is safe
4386 regs = task_pt_regs(task);
4388 /* force a full reload */
4389 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4390 SET_LAST_CPU(ctx, -1);
4392 /* initial saved psr (stopped) */
4393 ctx->ctx_saved_psr_up = 0UL;
4394 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4400 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4403 * we must undo the dbregs setting (for system-wide)
4405 if (ret && set_dbregs) {
4407 pfm_sessions.pfs_sys_use_dbregs--;
4411 * release task, there is now a link with the context
4413 if (is_system == 0 && task != current) {
4417 ret = pfm_check_task_exist(ctx);
4419 ctx->ctx_state = PFM_CTX_UNLOADED;
4420 ctx->ctx_task = NULL;
4428 * in this function, we do not need to increase the use count
4429 * for the task via get_task_struct(), because we hold the
4430 * context lock. If the task were to disappear while having
4431 * a context attached, it would go through pfm_exit_thread()
4432 * which also grabs the context lock and would therefore be blocked
4433 * until we are here.
4435 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4438 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4440 struct task_struct *task = PFM_CTX_TASK(ctx);
4441 struct pt_regs *tregs;
4442 int prev_state, is_system;
4445 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4447 prev_state = ctx->ctx_state;
4448 is_system = ctx->ctx_fl_system;
4451 * unload only when necessary
4453 if (prev_state == PFM_CTX_UNLOADED) {
4454 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4459 * clear psr and dcr bits
4461 ret = pfm_stop(ctx, NULL, 0, regs);
4462 if (ret) return ret;
4464 ctx->ctx_state = PFM_CTX_UNLOADED;
4467 * in system mode, we need to update the PMU directly
4468 * and the user level state of the caller, which may not
4469 * necessarily be the creator of the context.
4476 * local PMU is taken care of in pfm_stop()
4478 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4479 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4482 * save PMDs in context
4485 pfm_flush_pmds(current, ctx);
4488 * at this point we are done with the PMU
4489 * so we can unreserve the resource.
4491 if (prev_state != PFM_CTX_ZOMBIE)
4492 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4495 * disconnect context from task
4497 task->thread.pfm_context = NULL;
4499 * disconnect task from context
4501 ctx->ctx_task = NULL;
4504 * There is nothing more to cleanup here.
4512 tregs = task == current ? regs : task_pt_regs(task);
4514 if (task == current) {
4516 * cancel user level control
4518 ia64_psr(regs)->sp = 1;
4520 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4523 * save PMDs to context
4526 pfm_flush_pmds(task, ctx);
4529 * at this point we are done with the PMU
4530 * so we can unreserve the resource.
4532 * when state was ZOMBIE, we have already unreserved.
4534 if (prev_state != PFM_CTX_ZOMBIE)
4535 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4538 * reset activation counter and psr
4540 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4541 SET_LAST_CPU(ctx, -1);
4544 * PMU state will not be restored
4546 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4549 * break links between context and task
4551 task->thread.pfm_context = NULL;
4552 ctx->ctx_task = NULL;
4554 PFM_SET_WORK_PENDING(task, 0);
4556 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4557 ctx->ctx_fl_can_restart = 0;
4558 ctx->ctx_fl_going_zombie = 0;
4560 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4567 * called only from exit_thread(): task == current
4568 * we come here only if current has a context attached (loaded or masked)
4571 pfm_exit_thread(struct task_struct *task)
4574 unsigned long flags;
4575 struct pt_regs *regs = task_pt_regs(task);
4579 ctx = PFM_GET_CTX(task);
4581 PROTECT_CTX(ctx, flags);
4583 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4585 state = ctx->ctx_state;
4587 case PFM_CTX_UNLOADED:
4589 * only comes to this function if pfm_context is not NULL, i.e., cannot
4590 * be in unloaded state
4592 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4594 case PFM_CTX_LOADED:
4595 case PFM_CTX_MASKED:
4596 ret = pfm_context_unload(ctx, NULL, 0, regs);
4598 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4600 DPRINT(("ctx unloaded for current state was %d\n", state));
4602 pfm_end_notify_user(ctx);
4604 case PFM_CTX_ZOMBIE:
4605 ret = pfm_context_unload(ctx, NULL, 0, regs);
4607 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4612 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4615 UNPROTECT_CTX(ctx, flags);
4617 { u64 psr = pfm_get_psr();
4618 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4619 BUG_ON(GET_PMU_OWNER());
4620 BUG_ON(ia64_psr(regs)->up);
4621 BUG_ON(ia64_psr(regs)->pp);
4625 * All memory free operations (especially for vmalloc'ed memory)
4626 * MUST be done with interrupts ENABLED.
4628 if (free_ok) pfm_context_free(ctx);
4632 * functions MUST be listed in the increasing order of their index (see permfon.h)
4634 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4635 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4636 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4637 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4638 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4640 static pfm_cmd_desc_t pfm_cmd_tab[]={
4641 /* 0 */PFM_CMD_NONE,
4642 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4643 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4644 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4645 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4646 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4647 /* 6 */PFM_CMD_NONE,
4648 /* 7 */PFM_CMD_NONE,
4649 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4650 /* 9 */PFM_CMD_NONE,
4651 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4652 /* 11 */PFM_CMD_NONE,
4653 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4654 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4655 /* 14 */PFM_CMD_NONE,
4656 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4657 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4658 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4659 /* 18 */PFM_CMD_NONE,
4660 /* 19 */PFM_CMD_NONE,
4661 /* 20 */PFM_CMD_NONE,
4662 /* 21 */PFM_CMD_NONE,
4663 /* 22 */PFM_CMD_NONE,
4664 /* 23 */PFM_CMD_NONE,
4665 /* 24 */PFM_CMD_NONE,
4666 /* 25 */PFM_CMD_NONE,
4667 /* 26 */PFM_CMD_NONE,
4668 /* 27 */PFM_CMD_NONE,
4669 /* 28 */PFM_CMD_NONE,
4670 /* 29 */PFM_CMD_NONE,
4671 /* 30 */PFM_CMD_NONE,
4672 /* 31 */PFM_CMD_NONE,
4673 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4674 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4676 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4679 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4681 struct task_struct *task;
4682 int state, old_state;
4685 state = ctx->ctx_state;
4686 task = ctx->ctx_task;
4689 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4693 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4697 task->state, PFM_CMD_STOPPED(cmd)));
4700 * self-monitoring always ok.
4702 * for system-wide the caller can either be the creator of the
4703 * context (to one to which the context is attached to) OR
4704 * a task running on the same CPU as the session.
4706 if (task == current || ctx->ctx_fl_system) return 0;
4709 * we are monitoring another thread
4712 case PFM_CTX_UNLOADED:
4714 * if context is UNLOADED we are safe to go
4717 case PFM_CTX_ZOMBIE:
4719 * no command can operate on a zombie context
4721 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4723 case PFM_CTX_MASKED:
4725 * PMU state has been saved to software even though
4726 * the thread may still be running.
4728 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4732 * context is LOADED or MASKED. Some commands may need to have
4735 * We could lift this restriction for UP but it would mean that
4736 * the user has no guarantee the task would not run between
4737 * two successive calls to perfmonctl(). That's probably OK.
4738 * If this user wants to ensure the task does not run, then
4739 * the task must be stopped.
4741 if (PFM_CMD_STOPPED(cmd)) {
4742 if (!task_is_stopped_or_traced(task)) {
4743 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4747 * task is now stopped, wait for ctxsw out
4749 * This is an interesting point in the code.
4750 * We need to unprotect the context because
4751 * the pfm_save_regs() routines needs to grab
4752 * the same lock. There are danger in doing
4753 * this because it leaves a window open for
4754 * another task to get access to the context
4755 * and possibly change its state. The one thing
4756 * that is not possible is for the context to disappear
4757 * because we are protected by the VFS layer, i.e.,
4758 * get_fd()/put_fd().
4762 UNPROTECT_CTX(ctx, flags);
4764 wait_task_inactive(task, 0);
4766 PROTECT_CTX(ctx, flags);
4769 * we must recheck to verify if state has changed
4771 if (ctx->ctx_state != old_state) {
4772 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4780 * system-call entry point (must return long)
4783 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4785 struct file *file = NULL;
4786 pfm_context_t *ctx = NULL;
4787 unsigned long flags = 0UL;
4788 void *args_k = NULL;
4789 long ret; /* will expand int return types */
4790 size_t base_sz, sz, xtra_sz = 0;
4791 int narg, completed_args = 0, call_made = 0, cmd_flags;
4792 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4793 int (*getsize)(void *arg, size_t *sz);
4794 #define PFM_MAX_ARGSIZE 4096
4797 * reject any call if perfmon was disabled at initialization
4799 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4801 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4802 DPRINT(("invalid cmd=%d\n", cmd));
4806 func = pfm_cmd_tab[cmd].cmd_func;
4807 narg = pfm_cmd_tab[cmd].cmd_narg;
4808 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4809 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4810 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4812 if (unlikely(func == NULL)) {
4813 DPRINT(("invalid cmd=%d\n", cmd));
4817 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4825 * check if number of arguments matches what the command expects
4827 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4831 sz = xtra_sz + base_sz*count;
4833 * limit abuse to min page size
4835 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4836 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4841 * allocate default-sized argument buffer
4843 if (likely(count && args_k == NULL)) {
4844 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4845 if (args_k == NULL) return -ENOMEM;
4853 * assume sz = 0 for command without parameters
4855 if (sz && copy_from_user(args_k, arg, sz)) {
4856 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4861 * check if command supports extra parameters
4863 if (completed_args == 0 && getsize) {
4865 * get extra parameters size (based on main argument)
4867 ret = (*getsize)(args_k, &xtra_sz);
4868 if (ret) goto error_args;
4872 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4874 /* retry if necessary */
4875 if (likely(xtra_sz)) goto restart_args;
4878 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4883 if (unlikely(file == NULL)) {
4884 DPRINT(("invalid fd %d\n", fd));
4887 if (unlikely(PFM_IS_FILE(file) == 0)) {
4888 DPRINT(("fd %d not related to perfmon\n", fd));
4892 ctx = file->private_data;
4893 if (unlikely(ctx == NULL)) {
4894 DPRINT(("no context for fd %d\n", fd));
4897 prefetch(&ctx->ctx_state);
4899 PROTECT_CTX(ctx, flags);
4902 * check task is stopped
4904 ret = pfm_check_task_state(ctx, cmd, flags);
4905 if (unlikely(ret)) goto abort_locked;
4908 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4914 DPRINT(("context unlocked\n"));
4915 UNPROTECT_CTX(ctx, flags);
4918 /* copy argument back to user, if needed */
4919 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4927 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4933 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4935 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4936 pfm_ovfl_ctrl_t rst_ctrl;
4940 state = ctx->ctx_state;
4942 * Unlock sampling buffer and reset index atomically
4943 * XXX: not really needed when blocking
4945 if (CTX_HAS_SMPL(ctx)) {
4947 rst_ctrl.bits.mask_monitoring = 0;
4948 rst_ctrl.bits.reset_ovfl_pmds = 0;
4950 if (state == PFM_CTX_LOADED)
4951 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4953 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4955 rst_ctrl.bits.mask_monitoring = 0;
4956 rst_ctrl.bits.reset_ovfl_pmds = 1;
4960 if (rst_ctrl.bits.reset_ovfl_pmds) {
4961 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4963 if (rst_ctrl.bits.mask_monitoring == 0) {
4964 DPRINT(("resuming monitoring\n"));
4965 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4967 DPRINT(("stopping monitoring\n"));
4968 //pfm_stop_monitoring(current, regs);
4970 ctx->ctx_state = PFM_CTX_LOADED;
4975 * context MUST BE LOCKED when calling
4976 * can only be called for current
4979 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4983 DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4985 ret = pfm_context_unload(ctx, NULL, 0, regs);
4987 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4991 * and wakeup controlling task, indicating we are now disconnected
4993 wake_up_interruptible(&ctx->ctx_zombieq);
4996 * given that context is still locked, the controlling
4997 * task will only get access when we return from
4998 * pfm_handle_work().
5002 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5005 * pfm_handle_work() can be called with interrupts enabled
5006 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5007 * call may sleep, therefore we must re-enable interrupts
5008 * to avoid deadlocks. It is safe to do so because this function
5009 * is called ONLY when returning to user level (pUStk=1), in which case
5010 * there is no risk of kernel stack overflow due to deep
5011 * interrupt nesting.
5014 pfm_handle_work(void)
5017 struct pt_regs *regs;
5018 unsigned long flags, dummy_flags;
5019 unsigned long ovfl_regs;
5020 unsigned int reason;
5023 ctx = PFM_GET_CTX(current);
5025 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5026 task_pid_nr(current));
5030 PROTECT_CTX(ctx, flags);
5032 PFM_SET_WORK_PENDING(current, 0);
5034 regs = task_pt_regs(current);
5037 * extract reason for being here and clear
5039 reason = ctx->ctx_fl_trap_reason;
5040 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5041 ovfl_regs = ctx->ctx_ovfl_regs[0];
5043 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5046 * must be done before we check for simple-reset mode
5048 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5051 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5052 if (reason == PFM_TRAP_REASON_RESET)
5056 * restore interrupt mask to what it was on entry.
5057 * Could be enabled/diasbled.
5059 UNPROTECT_CTX(ctx, flags);
5062 * force interrupt enable because of down_interruptible()
5066 DPRINT(("before block sleeping\n"));
5069 * may go through without blocking on SMP systems
5070 * if restart has been received already by the time we call down()
5072 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5074 DPRINT(("after block sleeping ret=%d\n", ret));
5077 * lock context and mask interrupts again
5078 * We save flags into a dummy because we may have
5079 * altered interrupts mask compared to entry in this
5082 PROTECT_CTX(ctx, dummy_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
5105 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5106 ctx->ctx_ovfl_regs[0] = 0UL;
5110 * restore flags as they were upon entry
5112 UNPROTECT_CTX(ctx, flags);
5116 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5118 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5119 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5123 DPRINT(("waking up somebody\n"));
5125 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5128 * safe, we are not in intr handler, nor in ctxsw when
5131 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5137 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5139 pfm_msg_t *msg = NULL;
5141 if (ctx->ctx_fl_no_msg == 0) {
5142 msg = pfm_get_new_msg(ctx);
5144 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5148 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5149 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5150 msg->pfm_ovfl_msg.msg_active_set = 0;
5151 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5152 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5153 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5154 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5155 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5158 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5164 return pfm_notify_user(ctx, msg);
5168 pfm_end_notify_user(pfm_context_t *ctx)
5172 msg = pfm_get_new_msg(ctx);
5174 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5178 memset(msg, 0, sizeof(*msg));
5180 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5181 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5182 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5184 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5189 return pfm_notify_user(ctx, msg);
5193 * main overflow processing routine.
5194 * it can be called from the interrupt path or explicitly during the context switch code
5196 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5197 unsigned long pmc0, struct pt_regs *regs)
5199 pfm_ovfl_arg_t *ovfl_arg;
5201 unsigned long old_val, ovfl_val, new_val;
5202 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5203 unsigned long tstamp;
5204 pfm_ovfl_ctrl_t ovfl_ctrl;
5205 unsigned int i, has_smpl;
5206 int must_notify = 0;
5208 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5211 * sanity test. Should never happen
5213 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5215 tstamp = ia64_get_itc();
5216 mask = pmc0 >> PMU_FIRST_COUNTER;
5217 ovfl_val = pmu_conf->ovfl_val;
5218 has_smpl = CTX_HAS_SMPL(ctx);
5220 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5221 "used_pmds=0x%lx\n",
5223 task ? task_pid_nr(task): -1,
5224 (regs ? regs->cr_iip : 0),
5225 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5226 ctx->ctx_used_pmds[0]));
5230 * first we update the virtual counters
5231 * assume there was a prior ia64_srlz_d() issued
5233 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5235 /* skip pmd which did not overflow */
5236 if ((mask & 0x1) == 0) continue;
5239 * Note that the pmd is not necessarily 0 at this point as qualified events
5240 * may have happened before the PMU was frozen. The residual count is not
5241 * taken into consideration here but will be with any read of the pmd via
5244 old_val = new_val = ctx->ctx_pmds[i].val;
5245 new_val += 1 + ovfl_val;
5246 ctx->ctx_pmds[i].val = new_val;
5249 * check for overflow condition
5251 if (likely(old_val > new_val)) {
5252 ovfl_pmds |= 1UL << i;
5253 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5256 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5260 ia64_get_pmd(i) & ovfl_val,
5266 * there was no 64-bit overflow, nothing else to do
5268 if (ovfl_pmds == 0UL) return;
5271 * reset all control bits
5277 * if a sampling format module exists, then we "cache" the overflow by
5278 * calling the module's handler() routine.
5281 unsigned long start_cycles, end_cycles;
5282 unsigned long pmd_mask;
5284 int this_cpu = smp_processor_id();
5286 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5287 ovfl_arg = &ctx->ctx_ovfl_arg;
5289 prefetch(ctx->ctx_smpl_hdr);
5291 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5295 if ((pmd_mask & 0x1) == 0) continue;
5297 ovfl_arg->ovfl_pmd = (unsigned char )i;
5298 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5299 ovfl_arg->active_set = 0;
5300 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5301 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5303 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5304 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5305 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5308 * copy values of pmds of interest. Sampling format may copy them
5309 * into sampling buffer.
5312 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5313 if ((smpl_pmds & 0x1) == 0) continue;
5314 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5315 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5319 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5321 start_cycles = ia64_get_itc();
5324 * call custom buffer format record (handler) routine
5326 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5328 end_cycles = ia64_get_itc();
5331 * For those controls, we take the union because they have
5332 * an all or nothing behavior.
5334 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5335 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5336 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5338 * build the bitmask of pmds to reset now
5340 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5342 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5345 * when the module cannot handle the rest of the overflows, we abort right here
5347 if (ret && pmd_mask) {
5348 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5349 pmd_mask<<PMU_FIRST_COUNTER));
5352 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5354 ovfl_pmds &= ~reset_pmds;
5357 * when no sampling module is used, then the default
5358 * is to notify on overflow if requested by user
5360 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5361 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5362 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5363 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5365 * if needed, we reset all overflowed pmds
5367 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5370 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5373 * reset the requested PMD registers using the short reset values
5376 unsigned long bm = reset_pmds;
5377 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5380 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5382 * keep track of what to reset when unblocking
5384 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5387 * check for blocking context
5389 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5391 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5394 * set the perfmon specific checking pending work for the task
5396 PFM_SET_WORK_PENDING(task, 1);
5399 * when coming from ctxsw, current still points to the
5400 * previous task, therefore we must work with task and not current.
5402 set_notify_resume(task);
5405 * defer until state is changed (shorten spin window). the context is locked
5406 * anyway, so the signal receiver would come spin for nothing.
5411 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5412 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5413 PFM_GET_WORK_PENDING(task),
5414 ctx->ctx_fl_trap_reason,
5417 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5419 * in case monitoring must be stopped, we toggle the psr bits
5421 if (ovfl_ctrl.bits.mask_monitoring) {
5422 pfm_mask_monitoring(task);
5423 ctx->ctx_state = PFM_CTX_MASKED;
5424 ctx->ctx_fl_can_restart = 1;
5428 * send notification now
5430 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5435 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5437 task ? task_pid_nr(task) : -1,
5443 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5444 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5445 * come here as zombie only if the task is the current task. In which case, we
5446 * can access the PMU hardware directly.
5448 * Note that zombies do have PM_VALID set. So here we do the minimal.
5450 * In case the context was zombified it could not be reclaimed at the time
5451 * the monitoring program exited. At this point, the PMU reservation has been
5452 * returned, the sampiing buffer has been freed. We must convert this call
5453 * into a spurious interrupt. However, we must also avoid infinite overflows
5454 * by stopping monitoring for this task. We can only come here for a per-task
5455 * context. All we need to do is to stop monitoring using the psr bits which
5456 * are always task private. By re-enabling secure montioring, we ensure that
5457 * the monitored task will not be able to re-activate monitoring.
5458 * The task will eventually be context switched out, at which point the context
5459 * will be reclaimed (that includes releasing ownership of the PMU).
5461 * So there might be a window of time where the number of per-task session is zero
5462 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5463 * context. This is safe because if a per-task session comes in, it will push this one
5464 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5465 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5466 * also push our zombie context out.
5468 * Overall pretty hairy stuff....
5470 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5472 ia64_psr(regs)->up = 0;
5473 ia64_psr(regs)->sp = 1;
5478 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5480 struct task_struct *task;
5482 unsigned long flags;
5484 int this_cpu = smp_processor_id();
5487 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5490 * srlz.d done before arriving here
5492 pmc0 = ia64_get_pmc(0);
5494 task = GET_PMU_OWNER();
5495 ctx = GET_PMU_CTX();
5498 * if we have some pending bits set
5499 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5501 if (PMC0_HAS_OVFL(pmc0) && task) {
5503 * we assume that pmc0.fr is always set here
5507 if (!ctx) goto report_spurious1;
5509 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5510 goto report_spurious2;
5512 PROTECT_CTX_NOPRINT(ctx, flags);
5514 pfm_overflow_handler(task, ctx, pmc0, regs);
5516 UNPROTECT_CTX_NOPRINT(ctx, flags);
5519 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5523 * keep it unfrozen at all times
5530 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5531 this_cpu, task_pid_nr(task));
5535 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5543 pfm_interrupt_handler(int irq, void *arg)
5545 unsigned long start_cycles, total_cycles;
5546 unsigned long min, max;
5549 struct pt_regs *regs = get_irq_regs();
5551 this_cpu = get_cpu();
5552 if (likely(!pfm_alt_intr_handler)) {
5553 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5554 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5556 start_cycles = ia64_get_itc();
5558 ret = pfm_do_interrupt_handler(arg, regs);
5560 total_cycles = ia64_get_itc();
5563 * don't measure spurious interrupts
5565 if (likely(ret == 0)) {
5566 total_cycles -= start_cycles;
5568 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5569 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5571 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5575 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5583 * /proc/perfmon interface, for debug only
5586 #define PFM_PROC_SHOW_HEADER ((void *)(long)nr_cpu_ids+1)
5589 pfm_proc_start(struct seq_file *m, loff_t *pos)
5592 return PFM_PROC_SHOW_HEADER;
5595 while (*pos <= nr_cpu_ids) {
5596 if (cpu_online(*pos - 1)) {
5597 return (void *)*pos;
5605 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5608 return pfm_proc_start(m, pos);
5612 pfm_proc_stop(struct seq_file *m, void *v)
5617 pfm_proc_show_header(struct seq_file *m)
5619 struct list_head * pos;
5620 pfm_buffer_fmt_t * entry;
5621 unsigned long flags;
5624 "perfmon version : %u.%u\n"
5627 "expert mode : %s\n"
5628 "ovfl_mask : 0x%lx\n"
5629 "PMU flags : 0x%x\n",
5630 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5632 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5633 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5640 "proc_sessions : %u\n"
5641 "sys_sessions : %u\n"
5642 "sys_use_dbregs : %u\n"
5643 "ptrace_use_dbregs : %u\n",
5644 pfm_sessions.pfs_task_sessions,
5645 pfm_sessions.pfs_sys_sessions,
5646 pfm_sessions.pfs_sys_use_dbregs,
5647 pfm_sessions.pfs_ptrace_use_dbregs);
5651 spin_lock(&pfm_buffer_fmt_lock);
5653 list_for_each(pos, &pfm_buffer_fmt_list) {
5654 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5655 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5666 entry->fmt_uuid[10],
5667 entry->fmt_uuid[11],
5668 entry->fmt_uuid[12],
5669 entry->fmt_uuid[13],
5670 entry->fmt_uuid[14],
5671 entry->fmt_uuid[15],
5674 spin_unlock(&pfm_buffer_fmt_lock);
5679 pfm_proc_show(struct seq_file *m, void *v)
5685 if (v == PFM_PROC_SHOW_HEADER) {
5686 pfm_proc_show_header(m);
5690 /* show info for CPU (v - 1) */
5694 "CPU%-2d overflow intrs : %lu\n"
5695 "CPU%-2d overflow cycles : %lu\n"
5696 "CPU%-2d overflow min : %lu\n"
5697 "CPU%-2d overflow max : %lu\n"
5698 "CPU%-2d smpl handler calls : %lu\n"
5699 "CPU%-2d smpl handler cycles : %lu\n"
5700 "CPU%-2d spurious intrs : %lu\n"
5701 "CPU%-2d replay intrs : %lu\n"
5702 "CPU%-2d syst_wide : %d\n"
5703 "CPU%-2d dcr_pp : %d\n"
5704 "CPU%-2d exclude idle : %d\n"
5705 "CPU%-2d owner : %d\n"
5706 "CPU%-2d context : %p\n"
5707 "CPU%-2d activations : %lu\n",
5708 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5709 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5710 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5711 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5712 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5713 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5714 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5715 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5716 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5717 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5718 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5719 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5720 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5721 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5723 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5725 psr = pfm_get_psr();
5730 "CPU%-2d psr : 0x%lx\n"
5731 "CPU%-2d pmc0 : 0x%lx\n",
5733 cpu, ia64_get_pmc(0));
5735 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5736 if (PMC_IS_COUNTING(i) == 0) continue;
5738 "CPU%-2d pmc%u : 0x%lx\n"
5739 "CPU%-2d pmd%u : 0x%lx\n",
5740 cpu, i, ia64_get_pmc(i),
5741 cpu, i, ia64_get_pmd(i));
5747 const struct seq_operations pfm_seq_ops = {
5748 .start = pfm_proc_start,
5749 .next = pfm_proc_next,
5750 .stop = pfm_proc_stop,
5751 .show = pfm_proc_show
5755 pfm_proc_open(struct inode *inode, struct file *file)
5757 return seq_open(file, &pfm_seq_ops);
5762 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5763 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5764 * is active or inactive based on mode. We must rely on the value in
5765 * local_cpu_data->pfm_syst_info
5768 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5770 struct pt_regs *regs;
5772 unsigned long dcr_pp;
5774 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5777 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5778 * on every CPU, so we can rely on the pid to identify the idle task.
5780 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5781 regs = task_pt_regs(task);
5782 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5786 * if monitoring has started
5789 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5791 * context switching in?
5794 /* mask monitoring for the idle task */
5795 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5801 * context switching out
5802 * restore monitoring for next task
5804 * Due to inlining this odd if-then-else construction generates
5807 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5816 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5818 struct task_struct *task = ctx->ctx_task;
5820 ia64_psr(regs)->up = 0;
5821 ia64_psr(regs)->sp = 1;
5823 if (GET_PMU_OWNER() == task) {
5824 DPRINT(("cleared ownership for [%d]\n",
5825 task_pid_nr(ctx->ctx_task)));
5826 SET_PMU_OWNER(NULL, NULL);
5830 * disconnect the task from the context and vice-versa
5832 PFM_SET_WORK_PENDING(task, 0);
5834 task->thread.pfm_context = NULL;
5835 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5837 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task)));
5842 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5845 pfm_save_regs(struct task_struct *task)
5848 unsigned long flags;
5852 ctx = PFM_GET_CTX(task);
5853 if (ctx == NULL) return;
5856 * we always come here with interrupts ALREADY disabled by
5857 * the scheduler. So we simply need to protect against concurrent
5858 * access, not CPU concurrency.
5860 flags = pfm_protect_ctx_ctxsw(ctx);
5862 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5863 struct pt_regs *regs = task_pt_regs(task);
5867 pfm_force_cleanup(ctx, regs);
5869 BUG_ON(ctx->ctx_smpl_hdr);
5871 pfm_unprotect_ctx_ctxsw(ctx, flags);
5873 pfm_context_free(ctx);
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(ctx->th_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 ctx->th_pmcs[0] = ia64_get_pmc(0);
5921 * unfreeze PMU if had pending overflows
5923 if (ctx->th_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 unsigned long flags;
5970 { u64 psr = pfm_get_psr();
5971 BUG_ON(psr & IA64_PSR_UP);
5974 ctx = PFM_GET_CTX(task);
5977 * we need to mask PMU overflow here to
5978 * make sure that we maintain pmc0 until
5979 * we save it. overflow interrupts are
5980 * treated as spurious if there is no
5983 * XXX: I don't think this is necessary
5985 PROTECT_CTX(ctx,flags);
5988 * release ownership of this PMU.
5989 * must be done before we save the registers.
5991 * after this call any PMU interrupt is treated
5994 SET_PMU_OWNER(NULL, NULL);
5997 * save all the pmds we use
5999 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6002 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6003 * it is needed to check for pended overflow
6004 * on the restore path
6006 ctx->th_pmcs[0] = ia64_get_pmc(0);
6009 * unfreeze PMU if had pending overflows
6011 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6014 * now get can unmask PMU interrupts, they will
6015 * be treated as purely spurious and we will not
6016 * lose any information
6018 UNPROTECT_CTX(ctx,flags);
6020 #endif /* CONFIG_SMP */
6024 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6027 pfm_load_regs (struct task_struct *task)
6030 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6031 unsigned long flags;
6033 int need_irq_resend;
6035 ctx = PFM_GET_CTX(task);
6036 if (unlikely(ctx == NULL)) return;
6038 BUG_ON(GET_PMU_OWNER());
6041 * possible on unload
6043 if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6046 * we always come here with interrupts ALREADY disabled by
6047 * the scheduler. So we simply need to protect against concurrent
6048 * access, not CPU concurrency.
6050 flags = pfm_protect_ctx_ctxsw(ctx);
6051 psr = pfm_get_psr();
6053 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6055 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6056 BUG_ON(psr & IA64_PSR_I);
6058 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6059 struct pt_regs *regs = task_pt_regs(task);
6061 BUG_ON(ctx->ctx_smpl_hdr);
6063 pfm_force_cleanup(ctx, regs);
6065 pfm_unprotect_ctx_ctxsw(ctx, flags);
6068 * this one (kmalloc'ed) is fine with interrupts disabled
6070 pfm_context_free(ctx);
6076 * we restore ALL the debug registers to avoid picking up
6079 if (ctx->ctx_fl_using_dbreg) {
6080 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6081 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6084 * retrieve saved psr.up
6086 psr_up = ctx->ctx_saved_psr_up;
6089 * if we were the last user of the PMU on that CPU,
6090 * then nothing to do except restore psr
6092 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6095 * retrieve partial reload masks (due to user modifications)
6097 pmc_mask = ctx->ctx_reload_pmcs[0];
6098 pmd_mask = ctx->ctx_reload_pmds[0];
6102 * To avoid leaking information to the user level when psr.sp=0,
6103 * we must reload ALL implemented pmds (even the ones we don't use).
6104 * In the kernel we only allow PFM_READ_PMDS on registers which
6105 * we initialized or requested (sampling) so there is no risk there.
6107 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6110 * ALL accessible PMCs are systematically reloaded, unused registers
6111 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6112 * up stale configuration.
6114 * PMC0 is never in the mask. It is always restored separately.
6116 pmc_mask = ctx->ctx_all_pmcs[0];
6119 * when context is MASKED, we will restore PMC with plm=0
6120 * and PMD with stale information, but that's ok, nothing
6123 * XXX: optimize here
6125 if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6126 if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6129 * check for pending overflow at the time the state
6132 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6134 * reload pmc0 with the overflow information
6135 * On McKinley PMU, this will trigger a PMU interrupt
6137 ia64_set_pmc(0, ctx->th_pmcs[0]);
6139 ctx->th_pmcs[0] = 0UL;
6142 * will replay the PMU interrupt
6144 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6146 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6150 * we just did a reload, so we reset the partial reload fields
6152 ctx->ctx_reload_pmcs[0] = 0UL;
6153 ctx->ctx_reload_pmds[0] = 0UL;
6155 SET_LAST_CPU(ctx, smp_processor_id());
6158 * dump activation value for this PMU
6162 * record current activation for this context
6164 SET_ACTIVATION(ctx);
6167 * establish new ownership.
6169 SET_PMU_OWNER(task, ctx);
6172 * restore the psr.up bit. measurement
6174 * no PMU interrupt can happen at this point
6175 * because we still have interrupts disabled.
6177 if (likely(psr_up)) pfm_set_psr_up();
6180 * allow concurrent access to context
6182 pfm_unprotect_ctx_ctxsw(ctx, flags);
6184 #else /* !CONFIG_SMP */
6186 * reload PMU state for UP kernels
6187 * in 2.5 we come here with interrupts disabled
6190 pfm_load_regs (struct task_struct *task)
6193 struct task_struct *owner;
6194 unsigned long pmd_mask, pmc_mask;
6196 int need_irq_resend;
6198 owner = GET_PMU_OWNER();
6199 ctx = PFM_GET_CTX(task);
6200 psr = pfm_get_psr();
6202 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6203 BUG_ON(psr & IA64_PSR_I);
6206 * we restore ALL the debug registers to avoid picking up
6209 * This must be done even when the task is still the owner
6210 * as the registers may have been modified via ptrace()
6211 * (not perfmon) by the previous task.
6213 if (ctx->ctx_fl_using_dbreg) {
6214 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6215 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6219 * retrieved saved psr.up
6221 psr_up = ctx->ctx_saved_psr_up;
6222 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6225 * short path, our state is still there, just
6226 * need to restore psr and we go
6228 * we do not touch either PMC nor PMD. the psr is not touched
6229 * by the overflow_handler. So we are safe w.r.t. to interrupt
6230 * concurrency even without interrupt masking.
6232 if (likely(owner == task)) {
6233 if (likely(psr_up)) pfm_set_psr_up();
6238 * someone else is still using the PMU, first push it out and
6239 * then we'll be able to install our stuff !
6241 * Upon return, there will be no owner for the current PMU
6243 if (owner) pfm_lazy_save_regs(owner);
6246 * To avoid leaking information to the user level when psr.sp=0,
6247 * we must reload ALL implemented pmds (even the ones we don't use).
6248 * In the kernel we only allow PFM_READ_PMDS on registers which
6249 * we initialized or requested (sampling) so there is no risk there.
6251 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6254 * ALL accessible PMCs are systematically reloaded, unused registers
6255 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6256 * up stale configuration.
6258 * PMC0 is never in the mask. It is always restored separately
6260 pmc_mask = ctx->ctx_all_pmcs[0];
6262 pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6263 pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6266 * check for pending overflow at the time the state
6269 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6271 * reload pmc0 with the overflow information
6272 * On McKinley PMU, this will trigger a PMU interrupt
6274 ia64_set_pmc(0, ctx->th_pmcs[0]);
6277 ctx->th_pmcs[0] = 0UL;
6280 * will replay the PMU interrupt
6282 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6284 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6288 * establish new ownership.
6290 SET_PMU_OWNER(task, ctx);
6293 * restore the psr.up bit. measurement
6295 * no PMU interrupt can happen at this point
6296 * because we still have interrupts disabled.
6298 if (likely(psr_up)) pfm_set_psr_up();
6300 #endif /* CONFIG_SMP */
6303 * this function assumes monitoring is stopped
6306 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6309 unsigned long mask2, val, pmd_val, ovfl_val;
6310 int i, can_access_pmu = 0;
6314 * is the caller the task being monitored (or which initiated the
6315 * session for system wide measurements)
6317 is_self = ctx->ctx_task == task ? 1 : 0;
6320 * can access PMU is task is the owner of the PMU state on the current CPU
6321 * or if we are running on the CPU bound to the context in system-wide mode
6322 * (that is not necessarily the task the context is attached to in this mode).
6323 * In system-wide we always have can_access_pmu true because a task running on an
6324 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6326 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6327 if (can_access_pmu) {
6329 * Mark the PMU as not owned
6330 * This will cause the interrupt handler to do nothing in case an overflow
6331 * interrupt was in-flight
6332 * This also guarantees that pmc0 will contain the final state
6333 * It virtually gives us full control on overflow processing from that point
6336 SET_PMU_OWNER(NULL, NULL);
6337 DPRINT(("releasing ownership\n"));
6340 * read current overflow status:
6342 * we are guaranteed to read the final stable state
6345 pmc0 = ia64_get_pmc(0); /* slow */
6348 * reset freeze bit, overflow status information destroyed
6352 pmc0 = ctx->th_pmcs[0];
6354 * clear whatever overflow status bits there were
6356 ctx->th_pmcs[0] = 0;
6358 ovfl_val = pmu_conf->ovfl_val;
6360 * we save all the used pmds
6361 * we take care of overflows for counting PMDs
6363 * XXX: sampling situation is not taken into account here
6365 mask2 = ctx->ctx_used_pmds[0];
6367 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6369 for (i = 0; mask2; i++, mask2>>=1) {
6371 /* skip non used pmds */
6372 if ((mask2 & 0x1) == 0) continue;
6375 * can access PMU always true in system wide mode
6377 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6379 if (PMD_IS_COUNTING(i)) {
6380 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6383 ctx->ctx_pmds[i].val,
6387 * we rebuild the full 64 bit value of the counter
6389 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6392 * now everything is in ctx_pmds[] and we need
6393 * to clear the saved context from save_regs() such that
6394 * pfm_read_pmds() gets the correct value
6399 * take care of overflow inline
6401 if (pmc0 & (1UL << i)) {
6402 val += 1 + ovfl_val;
6403 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6407 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6409 if (is_self) ctx->th_pmds[i] = pmd_val;
6411 ctx->ctx_pmds[i].val = val;
6415 static struct irqaction perfmon_irqaction = {
6416 .handler = pfm_interrupt_handler,
6417 .flags = IRQF_DISABLED,
6422 pfm_alt_save_pmu_state(void *data)
6424 struct pt_regs *regs;
6426 regs = task_pt_regs(current);
6428 DPRINT(("called\n"));
6431 * should not be necessary but
6432 * let's take not risk
6436 ia64_psr(regs)->pp = 0;
6439 * This call is required
6440 * May cause a spurious interrupt on some processors
6448 pfm_alt_restore_pmu_state(void *data)
6450 struct pt_regs *regs;
6452 regs = task_pt_regs(current);
6454 DPRINT(("called\n"));
6457 * put PMU back in state expected
6462 ia64_psr(regs)->pp = 0;
6465 * perfmon runs with PMU unfrozen at all times
6473 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6478 /* some sanity checks */
6479 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6481 /* do the easy test first */
6482 if (pfm_alt_intr_handler) return -EBUSY;
6484 /* one at a time in the install or remove, just fail the others */
6485 if (!spin_trylock(&pfm_alt_install_check)) {
6489 /* reserve our session */
6490 for_each_online_cpu(reserve_cpu) {
6491 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6492 if (ret) goto cleanup_reserve;
6495 /* save the current system wide pmu states */
6496 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6498 DPRINT(("on_each_cpu() failed: %d\n", ret));
6499 goto cleanup_reserve;
6502 /* officially change to the alternate interrupt handler */
6503 pfm_alt_intr_handler = hdl;
6505 spin_unlock(&pfm_alt_install_check);
6510 for_each_online_cpu(i) {
6511 /* don't unreserve more than we reserved */
6512 if (i >= reserve_cpu) break;
6514 pfm_unreserve_session(NULL, 1, i);
6517 spin_unlock(&pfm_alt_install_check);
6521 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6524 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6529 if (hdl == NULL) return -EINVAL;
6531 /* cannot remove someone else's handler! */
6532 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6534 /* one at a time in the install or remove, just fail the others */
6535 if (!spin_trylock(&pfm_alt_install_check)) {
6539 pfm_alt_intr_handler = NULL;
6541 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6543 DPRINT(("on_each_cpu() failed: %d\n", ret));
6546 for_each_online_cpu(i) {
6547 pfm_unreserve_session(NULL, 1, i);
6550 spin_unlock(&pfm_alt_install_check);
6554 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6557 * perfmon initialization routine, called from the initcall() table
6559 static int init_pfm_fs(void);
6567 family = local_cpu_data->family;
6572 if ((*p)->probe() == 0) goto found;
6573 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6584 static const struct file_operations pfm_proc_fops = {
6585 .open = pfm_proc_open,
6587 .llseek = seq_lseek,
6588 .release = seq_release,
6594 unsigned int n, n_counters, i;
6596 printk("perfmon: version %u.%u IRQ %u\n",
6599 IA64_PERFMON_VECTOR);
6601 if (pfm_probe_pmu()) {
6602 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6603 local_cpu_data->family);
6608 * compute the number of implemented PMD/PMC from the
6609 * description tables
6612 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6613 if (PMC_IS_IMPL(i) == 0) continue;
6614 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6617 pmu_conf->num_pmcs = n;
6619 n = 0; n_counters = 0;
6620 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6621 if (PMD_IS_IMPL(i) == 0) continue;
6622 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6624 if (PMD_IS_COUNTING(i)) n_counters++;
6626 pmu_conf->num_pmds = n;
6627 pmu_conf->num_counters = n_counters;
6630 * sanity checks on the number of debug registers
6632 if (pmu_conf->use_rr_dbregs) {
6633 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6634 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6638 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6639 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6645 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6649 pmu_conf->num_counters,
6650 ffz(pmu_conf->ovfl_val));
6653 if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6654 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6660 * create /proc/perfmon (mostly for debugging purposes)
6662 perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6663 if (perfmon_dir == NULL) {
6664 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6670 * create /proc/sys/kernel/perfmon (for debugging purposes)
6672 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6675 * initialize all our spinlocks
6677 spin_lock_init(&pfm_sessions.pfs_lock);
6678 spin_lock_init(&pfm_buffer_fmt_lock);
6682 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6687 __initcall(pfm_init);
6690 * this function is called before pfm_init()
6693 pfm_init_percpu (void)
6695 static int first_time=1;
6697 * make sure no measurement is active
6698 * (may inherit programmed PMCs from EFI).
6704 * we run with the PMU not frozen at all times
6709 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6713 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6718 * used for debug purposes only
6721 dump_pmu_state(const char *from)
6723 struct task_struct *task;
6724 struct pt_regs *regs;
6726 unsigned long psr, dcr, info, flags;
6729 local_irq_save(flags);
6731 this_cpu = smp_processor_id();
6732 regs = task_pt_regs(current);
6733 info = PFM_CPUINFO_GET();
6734 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6736 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6737 local_irq_restore(flags);
6741 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6744 task_pid_nr(current),
6748 task = GET_PMU_OWNER();
6749 ctx = GET_PMU_CTX();
6751 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6753 psr = pfm_get_psr();
6755 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",
6758 psr & IA64_PSR_PP ? 1 : 0,
6759 psr & IA64_PSR_UP ? 1 : 0,
6760 dcr & IA64_DCR_PP ? 1 : 0,
6763 ia64_psr(regs)->pp);
6765 ia64_psr(regs)->up = 0;
6766 ia64_psr(regs)->pp = 0;
6768 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6769 if (PMC_IS_IMPL(i) == 0) continue;
6770 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, ctx->th_pmcs[i]);
6773 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6774 if (PMD_IS_IMPL(i) == 0) continue;
6775 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, ctx->th_pmds[i]);
6779 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6782 ctx->ctx_smpl_vaddr,
6786 ctx->ctx_saved_psr_up);
6788 local_irq_restore(flags);
6792 * called from process.c:copy_thread(). task is new child.
6795 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6797 struct thread_struct *thread;
6799 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6801 thread = &task->thread;
6804 * cut links inherited from parent (current)
6806 thread->pfm_context = NULL;
6808 PFM_SET_WORK_PENDING(task, 0);
6811 * the psr bits are already set properly in copy_threads()
6814 #else /* !CONFIG_PERFMON */
6816 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6820 #endif /* CONFIG_PERFMON */