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>
45 #include <asm/errno.h>
46 #include <asm/intrinsics.h>
48 #include <asm/perfmon.h>
49 #include <asm/processor.h>
50 #include <asm/signal.h>
51 #include <asm/system.h>
52 #include <asm/uaccess.h>
53 #include <asm/delay.h>
57 * perfmon context state
59 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
60 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
61 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
62 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
64 #define PFM_INVALID_ACTIVATION (~0UL)
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 static inline unsigned int
608 pfm_do_munmap(struct mm_struct
*mm
, unsigned long addr
, size_t len
, int acct
)
610 return do_munmap(mm
, addr
, len
);
613 static inline unsigned long
614 pfm_get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
, unsigned long pgoff
, unsigned long flags
, unsigned long exec
)
616 return get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
621 pfmfs_get_sb(struct file_system_type
*fs_type
, int flags
, const char *dev_name
, void *data
,
622 struct vfsmount
*mnt
)
624 return get_sb_pseudo(fs_type
, "pfm:", NULL
, PFMFS_MAGIC
, mnt
);
627 static struct file_system_type pfm_fs_type
= {
629 .get_sb
= pfmfs_get_sb
,
630 .kill_sb
= kill_anon_super
,
633 DEFINE_PER_CPU(unsigned long, pfm_syst_info
);
634 DEFINE_PER_CPU(struct task_struct
*, pmu_owner
);
635 DEFINE_PER_CPU(pfm_context_t
*, pmu_ctx
);
636 DEFINE_PER_CPU(unsigned long, pmu_activation_number
);
637 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info
);
640 /* forward declaration */
641 static const struct file_operations pfm_file_ops
;
644 * forward declarations
647 static void pfm_lazy_save_regs (struct task_struct
*ta
);
650 void dump_pmu_state(const char *);
651 static int pfm_write_ibr_dbr(int mode
, pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
653 #include "perfmon_itanium.h"
654 #include "perfmon_mckinley.h"
655 #include "perfmon_montecito.h"
656 #include "perfmon_generic.h"
658 static pmu_config_t
*pmu_confs
[]={
662 &pmu_conf_gen
, /* must be last */
667 static int pfm_end_notify_user(pfm_context_t
*ctx
);
670 pfm_clear_psr_pp(void)
672 ia64_rsm(IA64_PSR_PP
);
679 ia64_ssm(IA64_PSR_PP
);
684 pfm_clear_psr_up(void)
686 ia64_rsm(IA64_PSR_UP
);
693 ia64_ssm(IA64_PSR_UP
);
697 static inline unsigned long
701 tmp
= ia64_getreg(_IA64_REG_PSR
);
707 pfm_set_psr_l(unsigned long val
)
709 ia64_setreg(_IA64_REG_PSR_L
, val
);
721 pfm_unfreeze_pmu(void)
728 pfm_restore_ibrs(unsigned long *ibrs
, unsigned int nibrs
)
732 for (i
=0; i
< nibrs
; i
++) {
733 ia64_set_ibr(i
, ibrs
[i
]);
734 ia64_dv_serialize_instruction();
740 pfm_restore_dbrs(unsigned long *dbrs
, unsigned int ndbrs
)
744 for (i
=0; i
< ndbrs
; i
++) {
745 ia64_set_dbr(i
, dbrs
[i
]);
746 ia64_dv_serialize_data();
752 * PMD[i] must be a counter. no check is made
754 static inline unsigned long
755 pfm_read_soft_counter(pfm_context_t
*ctx
, int i
)
757 return ctx
->ctx_pmds
[i
].val
+ (ia64_get_pmd(i
) & pmu_conf
->ovfl_val
);
761 * PMD[i] must be a counter. no check is made
764 pfm_write_soft_counter(pfm_context_t
*ctx
, int i
, unsigned long val
)
766 unsigned long ovfl_val
= pmu_conf
->ovfl_val
;
768 ctx
->ctx_pmds
[i
].val
= val
& ~ovfl_val
;
770 * writing to unimplemented part is ignore, so we do not need to
773 ia64_set_pmd(i
, val
& ovfl_val
);
777 pfm_get_new_msg(pfm_context_t
*ctx
)
781 next
= (ctx
->ctx_msgq_tail
+1) % PFM_MAX_MSGS
;
783 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
784 if (next
== ctx
->ctx_msgq_head
) return NULL
;
786 idx
= ctx
->ctx_msgq_tail
;
787 ctx
->ctx_msgq_tail
= next
;
789 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
, idx
));
791 return ctx
->ctx_msgq
+idx
;
795 pfm_get_next_msg(pfm_context_t
*ctx
)
799 DPRINT(("ctx=%p head=%d tail=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
801 if (PFM_CTXQ_EMPTY(ctx
)) return NULL
;
806 msg
= ctx
->ctx_msgq
+ctx
->ctx_msgq_head
;
811 ctx
->ctx_msgq_head
= (ctx
->ctx_msgq_head
+1) % PFM_MAX_MSGS
;
813 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
));
819 pfm_reset_msgq(pfm_context_t
*ctx
)
821 ctx
->ctx_msgq_head
= ctx
->ctx_msgq_tail
= 0;
822 DPRINT(("ctx=%p msgq reset\n", ctx
));
826 pfm_rvmalloc(unsigned long size
)
831 size
= PAGE_ALIGN(size
);
834 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
835 memset(mem
, 0, size
);
836 addr
= (unsigned long)mem
;
838 pfm_reserve_page(addr
);
847 pfm_rvfree(void *mem
, unsigned long size
)
852 DPRINT(("freeing physical buffer @%p size=%lu\n", mem
, size
));
853 addr
= (unsigned long) mem
;
854 while ((long) size
> 0) {
855 pfm_unreserve_page(addr
);
864 static pfm_context_t
*
865 pfm_context_alloc(int ctx_flags
)
870 * allocate context descriptor
871 * must be able to free with interrupts disabled
873 ctx
= kzalloc(sizeof(pfm_context_t
), GFP_KERNEL
);
875 DPRINT(("alloc ctx @%p\n", ctx
));
878 * init context protection lock
880 spin_lock_init(&ctx
->ctx_lock
);
883 * context is unloaded
885 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
888 * initialization of context's flags
890 ctx
->ctx_fl_block
= (ctx_flags
& PFM_FL_NOTIFY_BLOCK
) ? 1 : 0;
891 ctx
->ctx_fl_system
= (ctx_flags
& PFM_FL_SYSTEM_WIDE
) ? 1: 0;
892 ctx
->ctx_fl_no_msg
= (ctx_flags
& PFM_FL_OVFL_NO_MSG
) ? 1: 0;
894 * will move to set properties
895 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
899 * init restart semaphore to locked
901 init_completion(&ctx
->ctx_restart_done
);
904 * activation is used in SMP only
906 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
907 SET_LAST_CPU(ctx
, -1);
910 * initialize notification message queue
912 ctx
->ctx_msgq_head
= ctx
->ctx_msgq_tail
= 0;
913 init_waitqueue_head(&ctx
->ctx_msgq_wait
);
914 init_waitqueue_head(&ctx
->ctx_zombieq
);
921 pfm_context_free(pfm_context_t
*ctx
)
924 DPRINT(("free ctx @%p\n", ctx
));
930 pfm_mask_monitoring(struct task_struct
*task
)
932 pfm_context_t
*ctx
= PFM_GET_CTX(task
);
933 unsigned long mask
, val
, ovfl_mask
;
936 DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task
)));
938 ovfl_mask
= pmu_conf
->ovfl_val
;
940 * monitoring can only be masked as a result of a valid
941 * counter overflow. In UP, it means that the PMU still
942 * has an owner. Note that the owner can be different
943 * from the current task. However the PMU state belongs
945 * In SMP, a valid overflow only happens when task is
946 * current. Therefore if we come here, we know that
947 * the PMU state belongs to the current task, therefore
948 * we can access the live registers.
950 * So in both cases, the live register contains the owner's
951 * state. We can ONLY touch the PMU registers and NOT the PSR.
953 * As a consequence to this call, the ctx->th_pmds[] array
954 * contains stale information which must be ignored
955 * when context is reloaded AND monitoring is active (see
958 mask
= ctx
->ctx_used_pmds
[0];
959 for (i
= 0; mask
; i
++, mask
>>=1) {
960 /* skip non used pmds */
961 if ((mask
& 0x1) == 0) continue;
962 val
= ia64_get_pmd(i
);
964 if (PMD_IS_COUNTING(i
)) {
966 * we rebuild the full 64 bit value of the counter
968 ctx
->ctx_pmds
[i
].val
+= (val
& ovfl_mask
);
970 ctx
->ctx_pmds
[i
].val
= val
;
972 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
974 ctx
->ctx_pmds
[i
].val
,
978 * mask monitoring by setting the privilege level to 0
979 * we cannot use psr.pp/psr.up for this, it is controlled by
982 * if task is current, modify actual registers, otherwise modify
983 * thread save state, i.e., what will be restored in pfm_load_regs()
985 mask
= ctx
->ctx_used_monitors
[0] >> PMU_FIRST_COUNTER
;
986 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>=1) {
987 if ((mask
& 0x1) == 0UL) continue;
988 ia64_set_pmc(i
, ctx
->th_pmcs
[i
] & ~0xfUL
);
989 ctx
->th_pmcs
[i
] &= ~0xfUL
;
990 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i
, ctx
->th_pmcs
[i
]));
993 * make all of this visible
999 * must always be done with task == current
1001 * context must be in MASKED state when calling
1004 pfm_restore_monitoring(struct task_struct
*task
)
1006 pfm_context_t
*ctx
= PFM_GET_CTX(task
);
1007 unsigned long mask
, ovfl_mask
;
1008 unsigned long psr
, val
;
1011 is_system
= ctx
->ctx_fl_system
;
1012 ovfl_mask
= pmu_conf
->ovfl_val
;
1014 if (task
!= current
) {
1015 printk(KERN_ERR
"perfmon.%d: invalid task[%d] current[%d]\n", __LINE__
, task_pid_nr(task
), task_pid_nr(current
));
1018 if (ctx
->ctx_state
!= PFM_CTX_MASKED
) {
1019 printk(KERN_ERR
"perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__
,
1020 task_pid_nr(task
), task_pid_nr(current
), ctx
->ctx_state
);
1023 psr
= pfm_get_psr();
1025 * monitoring is masked via the PMC.
1026 * As we restore their value, we do not want each counter to
1027 * restart right away. We stop monitoring using the PSR,
1028 * restore the PMC (and PMD) and then re-establish the psr
1029 * as it was. Note that there can be no pending overflow at
1030 * this point, because monitoring was MASKED.
1032 * system-wide session are pinned and self-monitoring
1034 if (is_system
&& (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP
)) {
1035 /* disable dcr pp */
1036 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) & ~IA64_DCR_PP
);
1042 * first, we restore the PMD
1044 mask
= ctx
->ctx_used_pmds
[0];
1045 for (i
= 0; mask
; i
++, mask
>>=1) {
1046 /* skip non used pmds */
1047 if ((mask
& 0x1) == 0) continue;
1049 if (PMD_IS_COUNTING(i
)) {
1051 * we split the 64bit value according to
1054 val
= ctx
->ctx_pmds
[i
].val
& ovfl_mask
;
1055 ctx
->ctx_pmds
[i
].val
&= ~ovfl_mask
;
1057 val
= ctx
->ctx_pmds
[i
].val
;
1059 ia64_set_pmd(i
, val
);
1061 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1063 ctx
->ctx_pmds
[i
].val
,
1069 mask
= ctx
->ctx_used_monitors
[0] >> PMU_FIRST_COUNTER
;
1070 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>=1) {
1071 if ((mask
& 0x1) == 0UL) continue;
1072 ctx
->th_pmcs
[i
] = ctx
->ctx_pmcs
[i
];
1073 ia64_set_pmc(i
, ctx
->th_pmcs
[i
]);
1074 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1075 task_pid_nr(task
), i
, ctx
->th_pmcs
[i
]));
1080 * must restore DBR/IBR because could be modified while masked
1081 * XXX: need to optimize
1083 if (ctx
->ctx_fl_using_dbreg
) {
1084 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
1085 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
1091 if (is_system
&& (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP
)) {
1093 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) | IA64_DCR_PP
);
1100 pfm_save_pmds(unsigned long *pmds
, unsigned long mask
)
1106 for (i
=0; mask
; i
++, mask
>>=1) {
1107 if (mask
& 0x1) pmds
[i
] = ia64_get_pmd(i
);
1112 * reload from thread state (used for ctxw only)
1115 pfm_restore_pmds(unsigned long *pmds
, unsigned long mask
)
1118 unsigned long val
, ovfl_val
= pmu_conf
->ovfl_val
;
1120 for (i
=0; mask
; i
++, mask
>>=1) {
1121 if ((mask
& 0x1) == 0) continue;
1122 val
= PMD_IS_COUNTING(i
) ? pmds
[i
] & ovfl_val
: pmds
[i
];
1123 ia64_set_pmd(i
, val
);
1129 * propagate PMD from context to thread-state
1132 pfm_copy_pmds(struct task_struct
*task
, pfm_context_t
*ctx
)
1134 unsigned long ovfl_val
= pmu_conf
->ovfl_val
;
1135 unsigned long mask
= ctx
->ctx_all_pmds
[0];
1139 DPRINT(("mask=0x%lx\n", mask
));
1141 for (i
=0; mask
; i
++, mask
>>=1) {
1143 val
= ctx
->ctx_pmds
[i
].val
;
1146 * We break up the 64 bit value into 2 pieces
1147 * the lower bits go to the machine state in the
1148 * thread (will be reloaded on ctxsw in).
1149 * The upper part stays in the soft-counter.
1151 if (PMD_IS_COUNTING(i
)) {
1152 ctx
->ctx_pmds
[i
].val
= val
& ~ovfl_val
;
1155 ctx
->th_pmds
[i
] = val
;
1157 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1160 ctx
->ctx_pmds
[i
].val
));
1165 * propagate PMC from context to thread-state
1168 pfm_copy_pmcs(struct task_struct
*task
, pfm_context_t
*ctx
)
1170 unsigned long mask
= ctx
->ctx_all_pmcs
[0];
1173 DPRINT(("mask=0x%lx\n", mask
));
1175 for (i
=0; mask
; i
++, mask
>>=1) {
1176 /* masking 0 with ovfl_val yields 0 */
1177 ctx
->th_pmcs
[i
] = ctx
->ctx_pmcs
[i
];
1178 DPRINT(("pmc[%d]=0x%lx\n", i
, ctx
->th_pmcs
[i
]));
1185 pfm_restore_pmcs(unsigned long *pmcs
, unsigned long mask
)
1189 for (i
=0; mask
; i
++, mask
>>=1) {
1190 if ((mask
& 0x1) == 0) continue;
1191 ia64_set_pmc(i
, pmcs
[i
]);
1197 pfm_uuid_cmp(pfm_uuid_t a
, pfm_uuid_t b
)
1199 return memcmp(a
, b
, sizeof(pfm_uuid_t
));
1203 pfm_buf_fmt_exit(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, void *buf
, struct pt_regs
*regs
)
1206 if (fmt
->fmt_exit
) ret
= (*fmt
->fmt_exit
)(task
, buf
, regs
);
1211 pfm_buf_fmt_getsize(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, unsigned int flags
, int cpu
, void *arg
, unsigned long *size
)
1214 if (fmt
->fmt_getsize
) ret
= (*fmt
->fmt_getsize
)(task
, flags
, cpu
, arg
, size
);
1220 pfm_buf_fmt_validate(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, unsigned int flags
,
1224 if (fmt
->fmt_validate
) ret
= (*fmt
->fmt_validate
)(task
, flags
, cpu
, arg
);
1229 pfm_buf_fmt_init(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, void *buf
, unsigned int flags
,
1233 if (fmt
->fmt_init
) ret
= (*fmt
->fmt_init
)(task
, buf
, flags
, cpu
, arg
);
1238 pfm_buf_fmt_restart(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, pfm_ovfl_ctrl_t
*ctrl
, void *buf
, struct pt_regs
*regs
)
1241 if (fmt
->fmt_restart
) ret
= (*fmt
->fmt_restart
)(task
, ctrl
, buf
, regs
);
1246 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
)
1249 if (fmt
->fmt_restart_active
) ret
= (*fmt
->fmt_restart_active
)(task
, ctrl
, buf
, regs
);
1253 static pfm_buffer_fmt_t
*
1254 __pfm_find_buffer_fmt(pfm_uuid_t uuid
)
1256 struct list_head
* pos
;
1257 pfm_buffer_fmt_t
* entry
;
1259 list_for_each(pos
, &pfm_buffer_fmt_list
) {
1260 entry
= list_entry(pos
, pfm_buffer_fmt_t
, fmt_list
);
1261 if (pfm_uuid_cmp(uuid
, entry
->fmt_uuid
) == 0)
1268 * find a buffer format based on its uuid
1270 static pfm_buffer_fmt_t
*
1271 pfm_find_buffer_fmt(pfm_uuid_t uuid
)
1273 pfm_buffer_fmt_t
* fmt
;
1274 spin_lock(&pfm_buffer_fmt_lock
);
1275 fmt
= __pfm_find_buffer_fmt(uuid
);
1276 spin_unlock(&pfm_buffer_fmt_lock
);
1281 pfm_register_buffer_fmt(pfm_buffer_fmt_t
*fmt
)
1285 /* some sanity checks */
1286 if (fmt
== NULL
|| fmt
->fmt_name
== NULL
) return -EINVAL
;
1288 /* we need at least a handler */
1289 if (fmt
->fmt_handler
== NULL
) return -EINVAL
;
1292 * XXX: need check validity of fmt_arg_size
1295 spin_lock(&pfm_buffer_fmt_lock
);
1297 if (__pfm_find_buffer_fmt(fmt
->fmt_uuid
)) {
1298 printk(KERN_ERR
"perfmon: duplicate sampling format: %s\n", fmt
->fmt_name
);
1302 list_add(&fmt
->fmt_list
, &pfm_buffer_fmt_list
);
1303 printk(KERN_INFO
"perfmon: added sampling format %s\n", fmt
->fmt_name
);
1306 spin_unlock(&pfm_buffer_fmt_lock
);
1309 EXPORT_SYMBOL(pfm_register_buffer_fmt
);
1312 pfm_unregister_buffer_fmt(pfm_uuid_t uuid
)
1314 pfm_buffer_fmt_t
*fmt
;
1317 spin_lock(&pfm_buffer_fmt_lock
);
1319 fmt
= __pfm_find_buffer_fmt(uuid
);
1321 printk(KERN_ERR
"perfmon: cannot unregister format, not found\n");
1325 list_del_init(&fmt
->fmt_list
);
1326 printk(KERN_INFO
"perfmon: removed sampling format: %s\n", fmt
->fmt_name
);
1329 spin_unlock(&pfm_buffer_fmt_lock
);
1333 EXPORT_SYMBOL(pfm_unregister_buffer_fmt
);
1335 extern void update_pal_halt_status(int);
1338 pfm_reserve_session(struct task_struct
*task
, int is_syswide
, unsigned int cpu
)
1340 unsigned long flags
;
1342 * validity checks on cpu_mask have been done upstream
1346 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1347 pfm_sessions
.pfs_sys_sessions
,
1348 pfm_sessions
.pfs_task_sessions
,
1349 pfm_sessions
.pfs_sys_use_dbregs
,
1355 * cannot mix system wide and per-task sessions
1357 if (pfm_sessions
.pfs_task_sessions
> 0UL) {
1358 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1359 pfm_sessions
.pfs_task_sessions
));
1363 if (pfm_sessions
.pfs_sys_session
[cpu
]) goto error_conflict
;
1365 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu
, smp_processor_id()));
1367 pfm_sessions
.pfs_sys_session
[cpu
] = task
;
1369 pfm_sessions
.pfs_sys_sessions
++ ;
1372 if (pfm_sessions
.pfs_sys_sessions
) goto abort
;
1373 pfm_sessions
.pfs_task_sessions
++;
1376 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1377 pfm_sessions
.pfs_sys_sessions
,
1378 pfm_sessions
.pfs_task_sessions
,
1379 pfm_sessions
.pfs_sys_use_dbregs
,
1384 * disable default_idle() to go to PAL_HALT
1386 update_pal_halt_status(0);
1393 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1394 task_pid_nr(pfm_sessions
.pfs_sys_session
[cpu
]),
1404 pfm_unreserve_session(pfm_context_t
*ctx
, int is_syswide
, unsigned int cpu
)
1406 unsigned long flags
;
1408 * validity checks on cpu_mask have been done upstream
1412 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1413 pfm_sessions
.pfs_sys_sessions
,
1414 pfm_sessions
.pfs_task_sessions
,
1415 pfm_sessions
.pfs_sys_use_dbregs
,
1421 pfm_sessions
.pfs_sys_session
[cpu
] = NULL
;
1423 * would not work with perfmon+more than one bit in cpu_mask
1425 if (ctx
&& ctx
->ctx_fl_using_dbreg
) {
1426 if (pfm_sessions
.pfs_sys_use_dbregs
== 0) {
1427 printk(KERN_ERR
"perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx
);
1429 pfm_sessions
.pfs_sys_use_dbregs
--;
1432 pfm_sessions
.pfs_sys_sessions
--;
1434 pfm_sessions
.pfs_task_sessions
--;
1436 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1437 pfm_sessions
.pfs_sys_sessions
,
1438 pfm_sessions
.pfs_task_sessions
,
1439 pfm_sessions
.pfs_sys_use_dbregs
,
1444 * if possible, enable default_idle() to go into PAL_HALT
1446 if (pfm_sessions
.pfs_task_sessions
== 0 && pfm_sessions
.pfs_sys_sessions
== 0)
1447 update_pal_halt_status(1);
1455 * removes virtual mapping of the sampling buffer.
1456 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1457 * a PROTECT_CTX() section.
1460 pfm_remove_smpl_mapping(struct task_struct
*task
, void *vaddr
, unsigned long size
)
1465 if (task
->mm
== NULL
|| size
== 0UL || vaddr
== NULL
) {
1466 printk(KERN_ERR
"perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task
), task
->mm
);
1470 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr
, size
));
1473 * does the actual unmapping
1475 down_write(&task
->mm
->mmap_sem
);
1477 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr
, size
));
1479 r
= pfm_do_munmap(task
->mm
, (unsigned long)vaddr
, size
, 0);
1481 up_write(&task
->mm
->mmap_sem
);
1483 printk(KERN_ERR
"perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task
), vaddr
, size
);
1486 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr
, size
, r
));
1492 * free actual physical storage used by sampling buffer
1496 pfm_free_smpl_buffer(pfm_context_t
*ctx
)
1498 pfm_buffer_fmt_t
*fmt
;
1500 if (ctx
->ctx_smpl_hdr
== NULL
) goto invalid_free
;
1503 * we won't use the buffer format anymore
1505 fmt
= ctx
->ctx_buf_fmt
;
1507 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1510 ctx
->ctx_smpl_vaddr
));
1512 pfm_buf_fmt_exit(fmt
, current
, NULL
, NULL
);
1517 pfm_rvfree(ctx
->ctx_smpl_hdr
, ctx
->ctx_smpl_size
);
1519 ctx
->ctx_smpl_hdr
= NULL
;
1520 ctx
->ctx_smpl_size
= 0UL;
1525 printk(KERN_ERR
"perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current
));
1531 pfm_exit_smpl_buffer(pfm_buffer_fmt_t
*fmt
)
1533 if (fmt
== NULL
) return;
1535 pfm_buf_fmt_exit(fmt
, current
, NULL
, NULL
);
1540 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1541 * no real gain from having the whole whorehouse mounted. So we don't need
1542 * any operations on the root directory. However, we need a non-trivial
1543 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1545 static struct vfsmount
*pfmfs_mnt
;
1550 int err
= register_filesystem(&pfm_fs_type
);
1552 pfmfs_mnt
= kern_mount(&pfm_fs_type
);
1553 err
= PTR_ERR(pfmfs_mnt
);
1554 if (IS_ERR(pfmfs_mnt
))
1555 unregister_filesystem(&pfm_fs_type
);
1563 pfm_read(struct file
*filp
, char __user
*buf
, size_t size
, loff_t
*ppos
)
1568 unsigned long flags
;
1569 DECLARE_WAITQUEUE(wait
, current
);
1570 if (PFM_IS_FILE(filp
) == 0) {
1571 printk(KERN_ERR
"perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current
));
1575 ctx
= (pfm_context_t
*)filp
->private_data
;
1577 printk(KERN_ERR
"perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current
));
1582 * check even when there is no message
1584 if (size
< sizeof(pfm_msg_t
)) {
1585 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx
, sizeof(pfm_msg_t
)));
1589 PROTECT_CTX(ctx
, flags
);
1592 * put ourselves on the wait queue
1594 add_wait_queue(&ctx
->ctx_msgq_wait
, &wait
);
1602 set_current_state(TASK_INTERRUPTIBLE
);
1604 DPRINT(("head=%d tail=%d\n", ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
1607 if(PFM_CTXQ_EMPTY(ctx
) == 0) break;
1609 UNPROTECT_CTX(ctx
, flags
);
1612 * check non-blocking read
1615 if(filp
->f_flags
& O_NONBLOCK
) break;
1618 * check pending signals
1620 if(signal_pending(current
)) {
1625 * no message, so wait
1629 PROTECT_CTX(ctx
, flags
);
1631 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current
), ret
));
1632 set_current_state(TASK_RUNNING
);
1633 remove_wait_queue(&ctx
->ctx_msgq_wait
, &wait
);
1635 if (ret
< 0) goto abort
;
1638 msg
= pfm_get_next_msg(ctx
);
1640 printk(KERN_ERR
"perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx
, task_pid_nr(current
));
1644 DPRINT(("fd=%d type=%d\n", msg
->pfm_gen_msg
.msg_ctx_fd
, msg
->pfm_gen_msg
.msg_type
));
1647 if(copy_to_user(buf
, msg
, sizeof(pfm_msg_t
)) == 0) ret
= sizeof(pfm_msg_t
);
1650 UNPROTECT_CTX(ctx
, flags
);
1656 pfm_write(struct file
*file
, const char __user
*ubuf
,
1657 size_t size
, loff_t
*ppos
)
1659 DPRINT(("pfm_write called\n"));
1664 pfm_poll(struct file
*filp
, poll_table
* wait
)
1667 unsigned long flags
;
1668 unsigned int mask
= 0;
1670 if (PFM_IS_FILE(filp
) == 0) {
1671 printk(KERN_ERR
"perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current
));
1675 ctx
= (pfm_context_t
*)filp
->private_data
;
1677 printk(KERN_ERR
"perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current
));
1682 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx
->ctx_fd
));
1684 poll_wait(filp
, &ctx
->ctx_msgq_wait
, wait
);
1686 PROTECT_CTX(ctx
, flags
);
1688 if (PFM_CTXQ_EMPTY(ctx
) == 0)
1689 mask
= POLLIN
| POLLRDNORM
;
1691 UNPROTECT_CTX(ctx
, flags
);
1693 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx
->ctx_fd
, mask
));
1699 pfm_ioctl(struct inode
*inode
, struct file
*file
, unsigned int cmd
, unsigned long arg
)
1701 DPRINT(("pfm_ioctl called\n"));
1706 * interrupt cannot be masked when coming here
1709 pfm_do_fasync(int fd
, struct file
*filp
, pfm_context_t
*ctx
, int on
)
1713 ret
= fasync_helper (fd
, filp
, on
, &ctx
->ctx_async_queue
);
1715 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1716 task_pid_nr(current
),
1719 ctx
->ctx_async_queue
, ret
));
1725 pfm_fasync(int fd
, struct file
*filp
, int on
)
1730 if (PFM_IS_FILE(filp
) == 0) {
1731 printk(KERN_ERR
"perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current
));
1735 ctx
= (pfm_context_t
*)filp
->private_data
;
1737 printk(KERN_ERR
"perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current
));
1741 * we cannot mask interrupts during this call because this may
1742 * may go to sleep if memory is not readily avalaible.
1744 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1745 * done in caller. Serialization of this function is ensured by caller.
1747 ret
= pfm_do_fasync(fd
, filp
, ctx
, on
);
1750 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1753 ctx
->ctx_async_queue
, ret
));
1760 * this function is exclusively called from pfm_close().
1761 * The context is not protected at that time, nor are interrupts
1762 * on the remote CPU. That's necessary to avoid deadlocks.
1765 pfm_syswide_force_stop(void *info
)
1767 pfm_context_t
*ctx
= (pfm_context_t
*)info
;
1768 struct pt_regs
*regs
= task_pt_regs(current
);
1769 struct task_struct
*owner
;
1770 unsigned long flags
;
1773 if (ctx
->ctx_cpu
!= smp_processor_id()) {
1774 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1776 smp_processor_id());
1779 owner
= GET_PMU_OWNER();
1780 if (owner
!= ctx
->ctx_task
) {
1781 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1783 task_pid_nr(owner
), task_pid_nr(ctx
->ctx_task
));
1786 if (GET_PMU_CTX() != ctx
) {
1787 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1789 GET_PMU_CTX(), ctx
);
1793 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx
->ctx_task
)));
1795 * the context is already protected in pfm_close(), we simply
1796 * need to mask interrupts to avoid a PMU interrupt race on
1799 local_irq_save(flags
);
1801 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
1803 DPRINT(("context_unload returned %d\n", ret
));
1807 * unmask interrupts, PMU interrupts are now spurious here
1809 local_irq_restore(flags
);
1813 pfm_syswide_cleanup_other_cpu(pfm_context_t
*ctx
)
1817 DPRINT(("calling CPU%d for cleanup\n", ctx
->ctx_cpu
));
1818 ret
= smp_call_function_single(ctx
->ctx_cpu
, pfm_syswide_force_stop
, ctx
, 1);
1819 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx
->ctx_cpu
, ret
));
1821 #endif /* CONFIG_SMP */
1824 * called for each close(). Partially free resources.
1825 * When caller is self-monitoring, the context is unloaded.
1828 pfm_flush(struct file
*filp
, fl_owner_t id
)
1831 struct task_struct
*task
;
1832 struct pt_regs
*regs
;
1833 unsigned long flags
;
1834 unsigned long smpl_buf_size
= 0UL;
1835 void *smpl_buf_vaddr
= NULL
;
1836 int state
, is_system
;
1838 if (PFM_IS_FILE(filp
) == 0) {
1839 DPRINT(("bad magic for\n"));
1843 ctx
= (pfm_context_t
*)filp
->private_data
;
1845 printk(KERN_ERR
"perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current
));
1850 * remove our file from the async queue, if we use this mode.
1851 * This can be done without the context being protected. We come
1852 * here when the context has become unreachable by other tasks.
1854 * We may still have active monitoring at this point and we may
1855 * end up in pfm_overflow_handler(). However, fasync_helper()
1856 * operates with interrupts disabled and it cleans up the
1857 * queue. If the PMU handler is called prior to entering
1858 * fasync_helper() then it will send a signal. If it is
1859 * invoked after, it will find an empty queue and no
1860 * signal will be sent. In both case, we are safe
1862 PROTECT_CTX(ctx
, flags
);
1864 state
= ctx
->ctx_state
;
1865 is_system
= ctx
->ctx_fl_system
;
1867 task
= PFM_CTX_TASK(ctx
);
1868 regs
= task_pt_regs(task
);
1870 DPRINT(("ctx_state=%d is_current=%d\n",
1872 task
== current
? 1 : 0));
1875 * if state == UNLOADED, then task is NULL
1879 * we must stop and unload because we are losing access to the context.
1881 if (task
== current
) {
1884 * the task IS the owner but it migrated to another CPU: that's bad
1885 * but we must handle this cleanly. Unfortunately, the kernel does
1886 * not provide a mechanism to block migration (while the context is loaded).
1888 * We need to release the resource on the ORIGINAL cpu.
1890 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
1892 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
1894 * keep context protected but unmask interrupt for IPI
1896 local_irq_restore(flags
);
1898 pfm_syswide_cleanup_other_cpu(ctx
);
1901 * restore interrupt masking
1903 local_irq_save(flags
);
1906 * context is unloaded at this point
1909 #endif /* CONFIG_SMP */
1912 DPRINT(("forcing unload\n"));
1914 * stop and unload, returning with state UNLOADED
1915 * and session unreserved.
1917 pfm_context_unload(ctx
, NULL
, 0, regs
);
1919 DPRINT(("ctx_state=%d\n", ctx
->ctx_state
));
1924 * remove virtual mapping, if any, for the calling task.
1925 * cannot reset ctx field until last user is calling close().
1927 * ctx_smpl_vaddr must never be cleared because it is needed
1928 * by every task with access to the context
1930 * When called from do_exit(), the mm context is gone already, therefore
1931 * mm is NULL, i.e., the VMA is already gone and we do not have to
1934 if (ctx
->ctx_smpl_vaddr
&& current
->mm
) {
1935 smpl_buf_vaddr
= ctx
->ctx_smpl_vaddr
;
1936 smpl_buf_size
= ctx
->ctx_smpl_size
;
1939 UNPROTECT_CTX(ctx
, flags
);
1942 * if there was a mapping, then we systematically remove it
1943 * at this point. Cannot be done inside critical section
1944 * because some VM function reenables interrupts.
1947 if (smpl_buf_vaddr
) pfm_remove_smpl_mapping(current
, smpl_buf_vaddr
, smpl_buf_size
);
1952 * called either on explicit close() or from exit_files().
1953 * Only the LAST user of the file gets to this point, i.e., it is
1956 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1957 * (fput()),i.e, last task to access the file. Nobody else can access the
1958 * file at this point.
1960 * When called from exit_files(), the VMA has been freed because exit_mm()
1961 * is executed before exit_files().
1963 * When called from exit_files(), the current task is not yet ZOMBIE but we
1964 * flush the PMU state to the context.
1967 pfm_close(struct inode
*inode
, struct file
*filp
)
1970 struct task_struct
*task
;
1971 struct pt_regs
*regs
;
1972 DECLARE_WAITQUEUE(wait
, current
);
1973 unsigned long flags
;
1974 unsigned long smpl_buf_size
= 0UL;
1975 void *smpl_buf_addr
= NULL
;
1976 int free_possible
= 1;
1977 int state
, is_system
;
1979 DPRINT(("pfm_close called private=%p\n", filp
->private_data
));
1981 if (PFM_IS_FILE(filp
) == 0) {
1982 DPRINT(("bad magic\n"));
1986 ctx
= (pfm_context_t
*)filp
->private_data
;
1988 printk(KERN_ERR
"perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current
));
1992 PROTECT_CTX(ctx
, flags
);
1994 state
= ctx
->ctx_state
;
1995 is_system
= ctx
->ctx_fl_system
;
1997 task
= PFM_CTX_TASK(ctx
);
1998 regs
= task_pt_regs(task
);
2000 DPRINT(("ctx_state=%d is_current=%d\n",
2002 task
== current
? 1 : 0));
2005 * if task == current, then pfm_flush() unloaded the context
2007 if (state
== PFM_CTX_UNLOADED
) goto doit
;
2010 * context is loaded/masked and task != current, we need to
2011 * either force an unload or go zombie
2015 * The task is currently blocked or will block after an overflow.
2016 * we must force it to wakeup to get out of the
2017 * MASKED state and transition to the unloaded state by itself.
2019 * This situation is only possible for per-task mode
2021 if (state
== PFM_CTX_MASKED
&& CTX_OVFL_NOBLOCK(ctx
) == 0) {
2024 * set a "partial" zombie state to be checked
2025 * upon return from down() in pfm_handle_work().
2027 * We cannot use the ZOMBIE state, because it is checked
2028 * by pfm_load_regs() which is called upon wakeup from down().
2029 * In such case, it would free the context and then we would
2030 * return to pfm_handle_work() which would access the
2031 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2032 * but visible to pfm_handle_work().
2034 * For some window of time, we have a zombie context with
2035 * ctx_state = MASKED and not ZOMBIE
2037 ctx
->ctx_fl_going_zombie
= 1;
2040 * force task to wake up from MASKED state
2042 complete(&ctx
->ctx_restart_done
);
2044 DPRINT(("waking up ctx_state=%d\n", state
));
2047 * put ourself to sleep waiting for the other
2048 * task to report completion
2050 * the context is protected by mutex, therefore there
2051 * is no risk of being notified of completion before
2052 * begin actually on the waitq.
2054 set_current_state(TASK_INTERRUPTIBLE
);
2055 add_wait_queue(&ctx
->ctx_zombieq
, &wait
);
2057 UNPROTECT_CTX(ctx
, flags
);
2060 * XXX: check for signals :
2061 * - ok for explicit close
2062 * - not ok when coming from exit_files()
2067 PROTECT_CTX(ctx
, flags
);
2070 remove_wait_queue(&ctx
->ctx_zombieq
, &wait
);
2071 set_current_state(TASK_RUNNING
);
2074 * context is unloaded at this point
2076 DPRINT(("after zombie wakeup ctx_state=%d for\n", state
));
2078 else if (task
!= current
) {
2081 * switch context to zombie state
2083 ctx
->ctx_state
= PFM_CTX_ZOMBIE
;
2085 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task
)));
2087 * cannot free the context on the spot. deferred until
2088 * the task notices the ZOMBIE state
2092 pfm_context_unload(ctx
, NULL
, 0, regs
);
2097 /* reload state, may have changed during opening of critical section */
2098 state
= ctx
->ctx_state
;
2101 * the context is still attached to a task (possibly current)
2102 * we cannot destroy it right now
2106 * we must free the sampling buffer right here because
2107 * we cannot rely on it being cleaned up later by the
2108 * monitored task. It is not possible to free vmalloc'ed
2109 * memory in pfm_load_regs(). Instead, we remove the buffer
2110 * now. should there be subsequent PMU overflow originally
2111 * meant for sampling, the will be converted to spurious
2112 * and that's fine because the monitoring tools is gone anyway.
2114 if (ctx
->ctx_smpl_hdr
) {
2115 smpl_buf_addr
= ctx
->ctx_smpl_hdr
;
2116 smpl_buf_size
= ctx
->ctx_smpl_size
;
2117 /* no more sampling */
2118 ctx
->ctx_smpl_hdr
= NULL
;
2119 ctx
->ctx_fl_is_sampling
= 0;
2122 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2128 if (smpl_buf_addr
) pfm_exit_smpl_buffer(ctx
->ctx_buf_fmt
);
2131 * UNLOADED that the session has already been unreserved.
2133 if (state
== PFM_CTX_ZOMBIE
) {
2134 pfm_unreserve_session(ctx
, ctx
->ctx_fl_system
, ctx
->ctx_cpu
);
2138 * disconnect file descriptor from context must be done
2141 filp
->private_data
= NULL
;
2144 * if we free on the spot, the context is now completely unreachable
2145 * from the callers side. The monitored task side is also cut, so we
2148 * If we have a deferred free, only the caller side is disconnected.
2150 UNPROTECT_CTX(ctx
, flags
);
2153 * All memory free operations (especially for vmalloc'ed memory)
2154 * MUST be done with interrupts ENABLED.
2156 if (smpl_buf_addr
) pfm_rvfree(smpl_buf_addr
, smpl_buf_size
);
2159 * return the memory used by the context
2161 if (free_possible
) pfm_context_free(ctx
);
2167 pfm_no_open(struct inode
*irrelevant
, struct file
*dontcare
)
2169 DPRINT(("pfm_no_open called\n"));
2175 static const struct file_operations pfm_file_ops
= {
2176 .llseek
= no_llseek
,
2181 .open
= pfm_no_open
, /* special open code to disallow open via /proc */
2182 .fasync
= pfm_fasync
,
2183 .release
= pfm_close
,
2188 pfmfs_delete_dentry(struct dentry
*dentry
)
2193 static const struct dentry_operations pfmfs_dentry_operations
= {
2194 .d_delete
= pfmfs_delete_dentry
,
2198 static struct file
*
2199 pfm_alloc_file(pfm_context_t
*ctx
)
2202 struct inode
*inode
;
2208 * allocate a new inode
2210 inode
= new_inode(pfmfs_mnt
->mnt_sb
);
2212 return ERR_PTR(-ENOMEM
);
2214 DPRINT(("new inode ino=%ld @%p\n", inode
->i_ino
, inode
));
2216 inode
->i_mode
= S_IFCHR
|S_IRUGO
;
2217 inode
->i_uid
= current_fsuid();
2218 inode
->i_gid
= current_fsgid();
2220 sprintf(name
, "[%lu]", inode
->i_ino
);
2222 this.len
= strlen(name
);
2223 this.hash
= inode
->i_ino
;
2226 * allocate a new dcache entry
2228 path
.dentry
= d_alloc(pfmfs_mnt
->mnt_sb
->s_root
, &this);
2231 return ERR_PTR(-ENOMEM
);
2233 path
.mnt
= mntget(pfmfs_mnt
);
2235 path
.dentry
->d_op
= &pfmfs_dentry_operations
;
2236 d_add(path
.dentry
, inode
);
2238 file
= alloc_file(&path
, FMODE_READ
, &pfm_file_ops
);
2241 return ERR_PTR(-ENFILE
);
2244 file
->f_flags
= O_RDONLY
;
2245 file
->private_data
= ctx
;
2251 pfm_remap_buffer(struct vm_area_struct
*vma
, unsigned long buf
, unsigned long addr
, unsigned long size
)
2253 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf
, addr
, size
));
2256 unsigned long pfn
= ia64_tpa(buf
) >> PAGE_SHIFT
;
2259 if (remap_pfn_range(vma
, addr
, pfn
, PAGE_SIZE
, PAGE_READONLY
))
2270 * allocate a sampling buffer and remaps it into the user address space of the task
2273 pfm_smpl_buffer_alloc(struct task_struct
*task
, struct file
*filp
, pfm_context_t
*ctx
, unsigned long rsize
, void **user_vaddr
)
2275 struct mm_struct
*mm
= task
->mm
;
2276 struct vm_area_struct
*vma
= NULL
;
2282 * the fixed header + requested size and align to page boundary
2284 size
= PAGE_ALIGN(rsize
);
2286 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize
, size
));
2289 * check requested size to avoid Denial-of-service attacks
2290 * XXX: may have to refine this test
2291 * Check against address space limit.
2293 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2296 if (size
> task_rlimit(task
, RLIMIT_MEMLOCK
))
2300 * We do the easy to undo allocations first.
2302 * pfm_rvmalloc(), clears the buffer, so there is no leak
2304 smpl_buf
= pfm_rvmalloc(size
);
2305 if (smpl_buf
== NULL
) {
2306 DPRINT(("Can't allocate sampling buffer\n"));
2310 DPRINT(("smpl_buf @%p\n", smpl_buf
));
2313 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2315 DPRINT(("Cannot allocate vma\n"));
2318 INIT_LIST_HEAD(&vma
->anon_vma_chain
);
2321 * partially initialize the vma for the sampling buffer
2324 vma
->vm_file
= filp
;
2325 vma
->vm_flags
= VM_READ
| VM_MAYREAD
|VM_RESERVED
;
2326 vma
->vm_page_prot
= PAGE_READONLY
; /* XXX may need to change */
2329 * Now we have everything we need and we can initialize
2330 * and connect all the data structures
2333 ctx
->ctx_smpl_hdr
= smpl_buf
;
2334 ctx
->ctx_smpl_size
= size
; /* aligned size */
2337 * Let's do the difficult operations next.
2339 * now we atomically find some area in the address space and
2340 * remap the buffer in it.
2342 down_write(&task
->mm
->mmap_sem
);
2344 /* find some free area in address space, must have mmap sem held */
2345 vma
->vm_start
= pfm_get_unmapped_area(NULL
, 0, size
, 0, MAP_PRIVATE
|MAP_ANONYMOUS
, 0);
2346 if (vma
->vm_start
== 0UL) {
2347 DPRINT(("Cannot find unmapped area for size %ld\n", size
));
2348 up_write(&task
->mm
->mmap_sem
);
2351 vma
->vm_end
= vma
->vm_start
+ size
;
2352 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2354 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size
, ctx
->ctx_smpl_hdr
, vma
->vm_start
));
2356 /* can only be applied to current task, need to have the mm semaphore held when called */
2357 if (pfm_remap_buffer(vma
, (unsigned long)smpl_buf
, vma
->vm_start
, size
)) {
2358 DPRINT(("Can't remap buffer\n"));
2359 up_write(&task
->mm
->mmap_sem
);
2366 * now insert the vma in the vm list for the process, must be
2367 * done with mmap lock held
2369 insert_vm_struct(mm
, vma
);
2371 mm
->total_vm
+= size
>> PAGE_SHIFT
;
2372 vm_stat_account(vma
->vm_mm
, vma
->vm_flags
, vma
->vm_file
,
2374 up_write(&task
->mm
->mmap_sem
);
2377 * keep track of user level virtual address
2379 ctx
->ctx_smpl_vaddr
= (void *)vma
->vm_start
;
2380 *(unsigned long *)user_vaddr
= vma
->vm_start
;
2385 kmem_cache_free(vm_area_cachep
, vma
);
2387 pfm_rvfree(smpl_buf
, size
);
2393 * XXX: do something better here
2396 pfm_bad_permissions(struct task_struct
*task
)
2398 const struct cred
*tcred
;
2399 uid_t uid
= current_uid();
2400 gid_t gid
= current_gid();
2404 tcred
= __task_cred(task
);
2406 /* inspired by ptrace_attach() */
2407 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2416 ret
= ((uid
!= tcred
->euid
)
2417 || (uid
!= tcred
->suid
)
2418 || (uid
!= tcred
->uid
)
2419 || (gid
!= tcred
->egid
)
2420 || (gid
!= tcred
->sgid
)
2421 || (gid
!= tcred
->gid
)) && !capable(CAP_SYS_PTRACE
);
2428 pfarg_is_sane(struct task_struct
*task
, pfarg_context_t
*pfx
)
2434 ctx_flags
= pfx
->ctx_flags
;
2436 if (ctx_flags
& PFM_FL_SYSTEM_WIDE
) {
2439 * cannot block in this mode
2441 if (ctx_flags
& PFM_FL_NOTIFY_BLOCK
) {
2442 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2447 /* probably more to add here */
2453 pfm_setup_buffer_fmt(struct task_struct
*task
, struct file
*filp
, pfm_context_t
*ctx
, unsigned int ctx_flags
,
2454 unsigned int cpu
, pfarg_context_t
*arg
)
2456 pfm_buffer_fmt_t
*fmt
= NULL
;
2457 unsigned long size
= 0UL;
2459 void *fmt_arg
= NULL
;
2461 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2463 /* invoke and lock buffer format, if found */
2464 fmt
= pfm_find_buffer_fmt(arg
->ctx_smpl_buf_id
);
2466 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task
)));
2471 * buffer argument MUST be contiguous to pfarg_context_t
2473 if (fmt
->fmt_arg_size
) fmt_arg
= PFM_CTXARG_BUF_ARG(arg
);
2475 ret
= pfm_buf_fmt_validate(fmt
, task
, ctx_flags
, cpu
, fmt_arg
);
2477 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task
), ctx_flags
, cpu
, fmt_arg
, ret
));
2479 if (ret
) goto error
;
2481 /* link buffer format and context */
2482 ctx
->ctx_buf_fmt
= fmt
;
2483 ctx
->ctx_fl_is_sampling
= 1; /* assume record() is defined */
2486 * check if buffer format wants to use perfmon buffer allocation/mapping service
2488 ret
= pfm_buf_fmt_getsize(fmt
, task
, ctx_flags
, cpu
, fmt_arg
, &size
);
2489 if (ret
) goto error
;
2493 * buffer is always remapped into the caller's address space
2495 ret
= pfm_smpl_buffer_alloc(current
, filp
, ctx
, size
, &uaddr
);
2496 if (ret
) goto error
;
2498 /* keep track of user address of buffer */
2499 arg
->ctx_smpl_vaddr
= uaddr
;
2501 ret
= pfm_buf_fmt_init(fmt
, task
, ctx
->ctx_smpl_hdr
, ctx_flags
, cpu
, fmt_arg
);
2508 pfm_reset_pmu_state(pfm_context_t
*ctx
)
2513 * install reset values for PMC.
2515 for (i
=1; PMC_IS_LAST(i
) == 0; i
++) {
2516 if (PMC_IS_IMPL(i
) == 0) continue;
2517 ctx
->ctx_pmcs
[i
] = PMC_DFL_VAL(i
);
2518 DPRINT(("pmc[%d]=0x%lx\n", i
, ctx
->ctx_pmcs
[i
]));
2521 * PMD registers are set to 0UL when the context in memset()
2525 * On context switched restore, we must restore ALL pmc and ALL pmd even
2526 * when they are not actively used by the task. In UP, the incoming process
2527 * may otherwise pick up left over PMC, PMD state from the previous process.
2528 * As opposed to PMD, stale PMC can cause harm to the incoming
2529 * process because they may change what is being measured.
2530 * Therefore, we must systematically reinstall the entire
2531 * PMC state. In SMP, the same thing is possible on the
2532 * same CPU but also on between 2 CPUs.
2534 * The problem with PMD is information leaking especially
2535 * to user level when psr.sp=0
2537 * There is unfortunately no easy way to avoid this problem
2538 * on either UP or SMP. This definitively slows down the
2539 * pfm_load_regs() function.
2543 * bitmask of all PMCs accessible to this context
2545 * PMC0 is treated differently.
2547 ctx
->ctx_all_pmcs
[0] = pmu_conf
->impl_pmcs
[0] & ~0x1;
2550 * bitmask of all PMDs that are accessible to this context
2552 ctx
->ctx_all_pmds
[0] = pmu_conf
->impl_pmds
[0];
2554 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx
->ctx_fd
, ctx
->ctx_all_pmcs
[0],ctx
->ctx_all_pmds
[0]));
2557 * useful in case of re-enable after disable
2559 ctx
->ctx_used_ibrs
[0] = 0UL;
2560 ctx
->ctx_used_dbrs
[0] = 0UL;
2564 pfm_ctx_getsize(void *arg
, size_t *sz
)
2566 pfarg_context_t
*req
= (pfarg_context_t
*)arg
;
2567 pfm_buffer_fmt_t
*fmt
;
2571 if (!pfm_uuid_cmp(req
->ctx_smpl_buf_id
, pfm_null_uuid
)) return 0;
2573 fmt
= pfm_find_buffer_fmt(req
->ctx_smpl_buf_id
);
2575 DPRINT(("cannot find buffer format\n"));
2578 /* get just enough to copy in user parameters */
2579 *sz
= fmt
->fmt_arg_size
;
2580 DPRINT(("arg_size=%lu\n", *sz
));
2588 * cannot attach if :
2590 * - task not owned by caller
2591 * - task incompatible with context mode
2594 pfm_task_incompatible(pfm_context_t
*ctx
, struct task_struct
*task
)
2597 * no kernel task or task not owner by caller
2599 if (task
->mm
== NULL
) {
2600 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task
)));
2603 if (pfm_bad_permissions(task
)) {
2604 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task
)));
2608 * cannot block in self-monitoring mode
2610 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && task
== current
) {
2611 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task
)));
2615 if (task
->exit_state
== EXIT_ZOMBIE
) {
2616 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task
)));
2621 * always ok for self
2623 if (task
== current
) return 0;
2625 if (!task_is_stopped_or_traced(task
)) {
2626 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task
), task
->state
));
2630 * make sure the task is off any CPU
2632 wait_task_inactive(task
, 0);
2634 /* more to come... */
2640 pfm_get_task(pfm_context_t
*ctx
, pid_t pid
, struct task_struct
**task
)
2642 struct task_struct
*p
= current
;
2645 /* XXX: need to add more checks here */
2646 if (pid
< 2) return -EPERM
;
2648 if (pid
!= task_pid_vnr(current
)) {
2650 read_lock(&tasklist_lock
);
2652 p
= find_task_by_vpid(pid
);
2654 /* make sure task cannot go away while we operate on it */
2655 if (p
) get_task_struct(p
);
2657 read_unlock(&tasklist_lock
);
2659 if (p
== NULL
) return -ESRCH
;
2662 ret
= pfm_task_incompatible(ctx
, p
);
2665 } else if (p
!= current
) {
2674 pfm_context_create(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
2676 pfarg_context_t
*req
= (pfarg_context_t
*)arg
;
2683 /* let's check the arguments first */
2684 ret
= pfarg_is_sane(current
, req
);
2688 ctx_flags
= req
->ctx_flags
;
2692 fd
= get_unused_fd();
2696 ctx
= pfm_context_alloc(ctx_flags
);
2700 filp
= pfm_alloc_file(ctx
);
2702 ret
= PTR_ERR(filp
);
2706 req
->ctx_fd
= ctx
->ctx_fd
= fd
;
2709 * does the user want to sample?
2711 if (pfm_uuid_cmp(req
->ctx_smpl_buf_id
, pfm_null_uuid
)) {
2712 ret
= pfm_setup_buffer_fmt(current
, filp
, ctx
, ctx_flags
, 0, req
);
2717 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2722 ctx
->ctx_fl_excl_idle
,
2727 * initialize soft PMU state
2729 pfm_reset_pmu_state(ctx
);
2731 fd_install(fd
, filp
);
2736 path
= filp
->f_path
;
2740 if (ctx
->ctx_buf_fmt
) {
2741 pfm_buf_fmt_exit(ctx
->ctx_buf_fmt
, current
, NULL
, regs
);
2744 pfm_context_free(ctx
);
2751 static inline unsigned long
2752 pfm_new_counter_value (pfm_counter_t
*reg
, int is_long_reset
)
2754 unsigned long val
= is_long_reset
? reg
->long_reset
: reg
->short_reset
;
2755 unsigned long new_seed
, old_seed
= reg
->seed
, mask
= reg
->mask
;
2756 extern unsigned long carta_random32 (unsigned long seed
);
2758 if (reg
->flags
& PFM_REGFL_RANDOM
) {
2759 new_seed
= carta_random32(old_seed
);
2760 val
-= (old_seed
& mask
); /* counter values are negative numbers! */
2761 if ((mask
>> 32) != 0)
2762 /* construct a full 64-bit random value: */
2763 new_seed
|= carta_random32(old_seed
>> 32) << 32;
2764 reg
->seed
= new_seed
;
2771 pfm_reset_regs_masked(pfm_context_t
*ctx
, unsigned long *ovfl_regs
, int is_long_reset
)
2773 unsigned long mask
= ovfl_regs
[0];
2774 unsigned long reset_others
= 0UL;
2779 * now restore reset value on sampling overflowed counters
2781 mask
>>= PMU_FIRST_COUNTER
;
2782 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
2784 if ((mask
& 0x1UL
) == 0UL) continue;
2786 ctx
->ctx_pmds
[i
].val
= val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2787 reset_others
|= ctx
->ctx_pmds
[i
].reset_pmds
[0];
2789 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset
? "long" : "short", i
, val
));
2793 * Now take care of resetting the other registers
2795 for(i
= 0; reset_others
; i
++, reset_others
>>= 1) {
2797 if ((reset_others
& 0x1) == 0) continue;
2799 ctx
->ctx_pmds
[i
].val
= val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2801 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2802 is_long_reset
? "long" : "short", i
, val
));
2807 pfm_reset_regs(pfm_context_t
*ctx
, unsigned long *ovfl_regs
, int is_long_reset
)
2809 unsigned long mask
= ovfl_regs
[0];
2810 unsigned long reset_others
= 0UL;
2814 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs
[0], is_long_reset
));
2816 if (ctx
->ctx_state
== PFM_CTX_MASKED
) {
2817 pfm_reset_regs_masked(ctx
, ovfl_regs
, is_long_reset
);
2822 * now restore reset value on sampling overflowed counters
2824 mask
>>= PMU_FIRST_COUNTER
;
2825 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
2827 if ((mask
& 0x1UL
) == 0UL) continue;
2829 val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2830 reset_others
|= ctx
->ctx_pmds
[i
].reset_pmds
[0];
2832 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset
? "long" : "short", i
, val
));
2834 pfm_write_soft_counter(ctx
, i
, val
);
2838 * Now take care of resetting the other registers
2840 for(i
= 0; reset_others
; i
++, reset_others
>>= 1) {
2842 if ((reset_others
& 0x1) == 0) continue;
2844 val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2846 if (PMD_IS_COUNTING(i
)) {
2847 pfm_write_soft_counter(ctx
, i
, val
);
2849 ia64_set_pmd(i
, val
);
2851 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2852 is_long_reset
? "long" : "short", i
, val
));
2858 pfm_write_pmcs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
2860 struct task_struct
*task
;
2861 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
2862 unsigned long value
, pmc_pm
;
2863 unsigned long smpl_pmds
, reset_pmds
, impl_pmds
;
2864 unsigned int cnum
, reg_flags
, flags
, pmc_type
;
2865 int i
, can_access_pmu
= 0, is_loaded
, is_system
, expert_mode
;
2866 int is_monitor
, is_counting
, state
;
2868 pfm_reg_check_t wr_func
;
2869 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2871 state
= ctx
->ctx_state
;
2872 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
2873 is_system
= ctx
->ctx_fl_system
;
2874 task
= ctx
->ctx_task
;
2875 impl_pmds
= pmu_conf
->impl_pmds
[0];
2877 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
2881 * In system wide and when the context is loaded, access can only happen
2882 * when the caller is running on the CPU being monitored by the session.
2883 * It does not have to be the owner (ctx_task) of the context per se.
2885 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
2886 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
2889 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
2891 expert_mode
= pfm_sysctl
.expert_mode
;
2893 for (i
= 0; i
< count
; i
++, req
++) {
2895 cnum
= req
->reg_num
;
2896 reg_flags
= req
->reg_flags
;
2897 value
= req
->reg_value
;
2898 smpl_pmds
= req
->reg_smpl_pmds
[0];
2899 reset_pmds
= req
->reg_reset_pmds
[0];
2903 if (cnum
>= PMU_MAX_PMCS
) {
2904 DPRINT(("pmc%u is invalid\n", cnum
));
2908 pmc_type
= pmu_conf
->pmc_desc
[cnum
].type
;
2909 pmc_pm
= (value
>> pmu_conf
->pmc_desc
[cnum
].pm_pos
) & 0x1;
2910 is_counting
= (pmc_type
& PFM_REG_COUNTING
) == PFM_REG_COUNTING
? 1 : 0;
2911 is_monitor
= (pmc_type
& PFM_REG_MONITOR
) == PFM_REG_MONITOR
? 1 : 0;
2914 * we reject all non implemented PMC as well
2915 * as attempts to modify PMC[0-3] which are used
2916 * as status registers by the PMU
2918 if ((pmc_type
& PFM_REG_IMPL
) == 0 || (pmc_type
& PFM_REG_CONTROL
) == PFM_REG_CONTROL
) {
2919 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum
, pmc_type
));
2922 wr_func
= pmu_conf
->pmc_desc
[cnum
].write_check
;
2924 * If the PMC is a monitor, then if the value is not the default:
2925 * - system-wide session: PMCx.pm=1 (privileged monitor)
2926 * - per-task : PMCx.pm=0 (user monitor)
2928 if (is_monitor
&& value
!= PMC_DFL_VAL(cnum
) && is_system
^ pmc_pm
) {
2929 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2938 * enforce generation of overflow interrupt. Necessary on all
2941 value
|= 1 << PMU_PMC_OI
;
2943 if (reg_flags
& PFM_REGFL_OVFL_NOTIFY
) {
2944 flags
|= PFM_REGFL_OVFL_NOTIFY
;
2947 if (reg_flags
& PFM_REGFL_RANDOM
) flags
|= PFM_REGFL_RANDOM
;
2949 /* verify validity of smpl_pmds */
2950 if ((smpl_pmds
& impl_pmds
) != smpl_pmds
) {
2951 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds
, cnum
));
2955 /* verify validity of reset_pmds */
2956 if ((reset_pmds
& impl_pmds
) != reset_pmds
) {
2957 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds
, cnum
));
2961 if (reg_flags
& (PFM_REGFL_OVFL_NOTIFY
|PFM_REGFL_RANDOM
)) {
2962 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum
));
2965 /* eventid on non-counting monitors are ignored */
2969 * execute write checker, if any
2971 if (likely(expert_mode
== 0 && wr_func
)) {
2972 ret
= (*wr_func
)(task
, ctx
, cnum
, &value
, regs
);
2973 if (ret
) goto error
;
2978 * no error on this register
2980 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
2983 * Now we commit the changes to the software state
2987 * update overflow information
2991 * full flag update each time a register is programmed
2993 ctx
->ctx_pmds
[cnum
].flags
= flags
;
2995 ctx
->ctx_pmds
[cnum
].reset_pmds
[0] = reset_pmds
;
2996 ctx
->ctx_pmds
[cnum
].smpl_pmds
[0] = smpl_pmds
;
2997 ctx
->ctx_pmds
[cnum
].eventid
= req
->reg_smpl_eventid
;
3000 * Mark all PMDS to be accessed as used.
3002 * We do not keep track of PMC because we have to
3003 * systematically restore ALL of them.
3005 * We do not update the used_monitors mask, because
3006 * if we have not programmed them, then will be in
3007 * a quiescent state, therefore we will not need to
3008 * mask/restore then when context is MASKED.
3010 CTX_USED_PMD(ctx
, reset_pmds
);
3011 CTX_USED_PMD(ctx
, smpl_pmds
);
3013 * make sure we do not try to reset on
3014 * restart because we have established new values
3016 if (state
== PFM_CTX_MASKED
) ctx
->ctx_ovfl_regs
[0] &= ~1UL << cnum
;
3019 * Needed in case the user does not initialize the equivalent
3020 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3021 * possible leak here.
3023 CTX_USED_PMD(ctx
, pmu_conf
->pmc_desc
[cnum
].dep_pmd
[0]);
3026 * keep track of the monitor PMC that we are using.
3027 * we save the value of the pmc in ctx_pmcs[] and if
3028 * the monitoring is not stopped for the context we also
3029 * place it in the saved state area so that it will be
3030 * picked up later by the context switch code.
3032 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3034 * The value in th_pmcs[] may be modified on overflow, i.e., when
3035 * monitoring needs to be stopped.
3037 if (is_monitor
) CTX_USED_MONITOR(ctx
, 1UL << cnum
);
3040 * update context state
3042 ctx
->ctx_pmcs
[cnum
] = value
;
3046 * write thread state
3048 if (is_system
== 0) ctx
->th_pmcs
[cnum
] = value
;
3051 * write hardware register if we can
3053 if (can_access_pmu
) {
3054 ia64_set_pmc(cnum
, value
);
3059 * per-task SMP only here
3061 * we are guaranteed that the task is not running on the other CPU,
3062 * we indicate that this PMD will need to be reloaded if the task
3063 * is rescheduled on the CPU it ran last on.
3065 ctx
->ctx_reload_pmcs
[0] |= 1UL << cnum
;
3070 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",
3076 ctx
->ctx_all_pmcs
[0],
3077 ctx
->ctx_used_pmds
[0],
3078 ctx
->ctx_pmds
[cnum
].eventid
,
3081 ctx
->ctx_reload_pmcs
[0],
3082 ctx
->ctx_used_monitors
[0],
3083 ctx
->ctx_ovfl_regs
[0]));
3087 * make sure the changes are visible
3089 if (can_access_pmu
) ia64_srlz_d();
3093 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3098 pfm_write_pmds(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3100 struct task_struct
*task
;
3101 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
3102 unsigned long value
, hw_value
, ovfl_mask
;
3104 int i
, can_access_pmu
= 0, state
;
3105 int is_counting
, is_loaded
, is_system
, expert_mode
;
3107 pfm_reg_check_t wr_func
;
3110 state
= ctx
->ctx_state
;
3111 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3112 is_system
= ctx
->ctx_fl_system
;
3113 ovfl_mask
= pmu_conf
->ovfl_val
;
3114 task
= ctx
->ctx_task
;
3116 if (unlikely(state
== PFM_CTX_ZOMBIE
)) return -EINVAL
;
3119 * on both UP and SMP, we can only write to the PMC when the task is
3120 * the owner of the local PMU.
3122 if (likely(is_loaded
)) {
3124 * In system wide and when the context is loaded, access can only happen
3125 * when the caller is running on the CPU being monitored by the session.
3126 * It does not have to be the owner (ctx_task) of the context per se.
3128 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3129 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3132 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3134 expert_mode
= pfm_sysctl
.expert_mode
;
3136 for (i
= 0; i
< count
; i
++, req
++) {
3138 cnum
= req
->reg_num
;
3139 value
= req
->reg_value
;
3141 if (!PMD_IS_IMPL(cnum
)) {
3142 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum
));
3145 is_counting
= PMD_IS_COUNTING(cnum
);
3146 wr_func
= pmu_conf
->pmd_desc
[cnum
].write_check
;
3149 * execute write checker, if any
3151 if (unlikely(expert_mode
== 0 && wr_func
)) {
3152 unsigned long v
= value
;
3154 ret
= (*wr_func
)(task
, ctx
, cnum
, &v
, regs
);
3155 if (ret
) goto abort_mission
;
3162 * no error on this register
3164 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
3167 * now commit changes to software state
3172 * update virtualized (64bits) counter
3176 * write context state
3178 ctx
->ctx_pmds
[cnum
].lval
= value
;
3181 * when context is load we use the split value
3184 hw_value
= value
& ovfl_mask
;
3185 value
= value
& ~ovfl_mask
;
3189 * update reset values (not just for counters)
3191 ctx
->ctx_pmds
[cnum
].long_reset
= req
->reg_long_reset
;
3192 ctx
->ctx_pmds
[cnum
].short_reset
= req
->reg_short_reset
;
3195 * update randomization parameters (not just for counters)
3197 ctx
->ctx_pmds
[cnum
].seed
= req
->reg_random_seed
;
3198 ctx
->ctx_pmds
[cnum
].mask
= req
->reg_random_mask
;
3201 * update context value
3203 ctx
->ctx_pmds
[cnum
].val
= value
;
3206 * Keep track of what we use
3208 * We do not keep track of PMC because we have to
3209 * systematically restore ALL of them.
3211 CTX_USED_PMD(ctx
, PMD_PMD_DEP(cnum
));
3214 * mark this PMD register used as well
3216 CTX_USED_PMD(ctx
, RDEP(cnum
));
3219 * make sure we do not try to reset on
3220 * restart because we have established new values
3222 if (is_counting
&& state
== PFM_CTX_MASKED
) {
3223 ctx
->ctx_ovfl_regs
[0] &= ~1UL << cnum
;
3228 * write thread state
3230 if (is_system
== 0) ctx
->th_pmds
[cnum
] = hw_value
;
3233 * write hardware register if we can
3235 if (can_access_pmu
) {
3236 ia64_set_pmd(cnum
, hw_value
);
3240 * we are guaranteed that the task is not running on the other CPU,
3241 * we indicate that this PMD will need to be reloaded if the task
3242 * is rescheduled on the CPU it ran last on.
3244 ctx
->ctx_reload_pmds
[0] |= 1UL << cnum
;
3249 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3250 "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",
3256 ctx
->ctx_pmds
[cnum
].val
,
3257 ctx
->ctx_pmds
[cnum
].short_reset
,
3258 ctx
->ctx_pmds
[cnum
].long_reset
,
3259 PMC_OVFL_NOTIFY(ctx
, cnum
) ? 'Y':'N',
3260 ctx
->ctx_pmds
[cnum
].seed
,
3261 ctx
->ctx_pmds
[cnum
].mask
,
3262 ctx
->ctx_used_pmds
[0],
3263 ctx
->ctx_pmds
[cnum
].reset_pmds
[0],
3264 ctx
->ctx_reload_pmds
[0],
3265 ctx
->ctx_all_pmds
[0],
3266 ctx
->ctx_ovfl_regs
[0]));
3270 * make changes visible
3272 if (can_access_pmu
) ia64_srlz_d();
3278 * for now, we have only one possibility for error
3280 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3285 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3286 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3287 * interrupt is delivered during the call, it will be kept pending until we leave, making
3288 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3289 * guaranteed to return consistent data to the user, it may simply be old. It is not
3290 * trivial to treat the overflow while inside the call because you may end up in
3291 * some module sampling buffer code causing deadlocks.
3294 pfm_read_pmds(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3296 struct task_struct
*task
;
3297 unsigned long val
= 0UL, lval
, ovfl_mask
, sval
;
3298 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
3299 unsigned int cnum
, reg_flags
= 0;
3300 int i
, can_access_pmu
= 0, state
;
3301 int is_loaded
, is_system
, is_counting
, expert_mode
;
3303 pfm_reg_check_t rd_func
;
3306 * access is possible when loaded only for
3307 * self-monitoring tasks or in UP mode
3310 state
= ctx
->ctx_state
;
3311 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3312 is_system
= ctx
->ctx_fl_system
;
3313 ovfl_mask
= pmu_conf
->ovfl_val
;
3314 task
= ctx
->ctx_task
;
3316 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
3318 if (likely(is_loaded
)) {
3320 * In system wide and when the context is loaded, access can only happen
3321 * when the caller is running on the CPU being monitored by the session.
3322 * It does not have to be the owner (ctx_task) of the context per se.
3324 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3325 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3329 * this can be true when not self-monitoring only in UP
3331 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3333 if (can_access_pmu
) ia64_srlz_d();
3335 expert_mode
= pfm_sysctl
.expert_mode
;
3337 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3343 * on both UP and SMP, we can only read the PMD from the hardware register when
3344 * the task is the owner of the local PMU.
3347 for (i
= 0; i
< count
; i
++, req
++) {
3349 cnum
= req
->reg_num
;
3350 reg_flags
= req
->reg_flags
;
3352 if (unlikely(!PMD_IS_IMPL(cnum
))) goto error
;
3354 * we can only read the register that we use. That includes
3355 * the one we explicitly initialize AND the one we want included
3356 * in the sampling buffer (smpl_regs).
3358 * Having this restriction allows optimization in the ctxsw routine
3359 * without compromising security (leaks)
3361 if (unlikely(!CTX_IS_USED_PMD(ctx
, cnum
))) goto error
;
3363 sval
= ctx
->ctx_pmds
[cnum
].val
;
3364 lval
= ctx
->ctx_pmds
[cnum
].lval
;
3365 is_counting
= PMD_IS_COUNTING(cnum
);
3368 * If the task is not the current one, then we check if the
3369 * PMU state is still in the local live register due to lazy ctxsw.
3370 * If true, then we read directly from the registers.
3372 if (can_access_pmu
){
3373 val
= ia64_get_pmd(cnum
);
3376 * context has been saved
3377 * if context is zombie, then task does not exist anymore.
3378 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3380 val
= is_loaded
? ctx
->th_pmds
[cnum
] : 0UL;
3382 rd_func
= pmu_conf
->pmd_desc
[cnum
].read_check
;
3386 * XXX: need to check for overflow when loaded
3393 * execute read checker, if any
3395 if (unlikely(expert_mode
== 0 && rd_func
)) {
3396 unsigned long v
= val
;
3397 ret
= (*rd_func
)(ctx
->ctx_task
, ctx
, cnum
, &v
, regs
);
3398 if (ret
) goto error
;
3403 PFM_REG_RETFLAG_SET(reg_flags
, 0);
3405 DPRINT(("pmd[%u]=0x%lx\n", cnum
, val
));
3408 * update register return value, abort all if problem during copy.
3409 * we only modify the reg_flags field. no check mode is fine because
3410 * access has been verified upfront in sys_perfmonctl().
3412 req
->reg_value
= val
;
3413 req
->reg_flags
= reg_flags
;
3414 req
->reg_last_reset_val
= lval
;
3420 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3425 pfm_mod_write_pmcs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3429 if (req
== NULL
) return -EINVAL
;
3431 ctx
= GET_PMU_CTX();
3433 if (ctx
== NULL
) return -EINVAL
;
3436 * for now limit to current task, which is enough when calling
3437 * from overflow handler
3439 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3441 return pfm_write_pmcs(ctx
, req
, nreq
, regs
);
3443 EXPORT_SYMBOL(pfm_mod_write_pmcs
);
3446 pfm_mod_read_pmds(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3450 if (req
== NULL
) return -EINVAL
;
3452 ctx
= GET_PMU_CTX();
3454 if (ctx
== NULL
) return -EINVAL
;
3457 * for now limit to current task, which is enough when calling
3458 * from overflow handler
3460 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3462 return pfm_read_pmds(ctx
, req
, nreq
, regs
);
3464 EXPORT_SYMBOL(pfm_mod_read_pmds
);
3467 * Only call this function when a process it trying to
3468 * write the debug registers (reading is always allowed)
3471 pfm_use_debug_registers(struct task_struct
*task
)
3473 pfm_context_t
*ctx
= task
->thread
.pfm_context
;
3474 unsigned long flags
;
3477 if (pmu_conf
->use_rr_dbregs
== 0) return 0;
3479 DPRINT(("called for [%d]\n", task_pid_nr(task
)));
3484 if (task
->thread
.flags
& IA64_THREAD_DBG_VALID
) return 0;
3487 * Even on SMP, we do not need to use an atomic here because
3488 * the only way in is via ptrace() and this is possible only when the
3489 * process is stopped. Even in the case where the ctxsw out is not totally
3490 * completed by the time we come here, there is no way the 'stopped' process
3491 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3492 * So this is always safe.
3494 if (ctx
&& ctx
->ctx_fl_using_dbreg
== 1) return -1;
3499 * We cannot allow setting breakpoints when system wide monitoring
3500 * sessions are using the debug registers.
3502 if (pfm_sessions
.pfs_sys_use_dbregs
> 0)
3505 pfm_sessions
.pfs_ptrace_use_dbregs
++;
3507 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3508 pfm_sessions
.pfs_ptrace_use_dbregs
,
3509 pfm_sessions
.pfs_sys_use_dbregs
,
3510 task_pid_nr(task
), ret
));
3518 * This function is called for every task that exits with the
3519 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3520 * able to use the debug registers for debugging purposes via
3521 * ptrace(). Therefore we know it was not using them for
3522 * performance monitoring, so we only decrement the number
3523 * of "ptraced" debug register users to keep the count up to date
3526 pfm_release_debug_registers(struct task_struct
*task
)
3528 unsigned long flags
;
3531 if (pmu_conf
->use_rr_dbregs
== 0) return 0;
3534 if (pfm_sessions
.pfs_ptrace_use_dbregs
== 0) {
3535 printk(KERN_ERR
"perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task
));
3538 pfm_sessions
.pfs_ptrace_use_dbregs
--;
3547 pfm_restart(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3549 struct task_struct
*task
;
3550 pfm_buffer_fmt_t
*fmt
;
3551 pfm_ovfl_ctrl_t rst_ctrl
;
3552 int state
, is_system
;
3555 state
= ctx
->ctx_state
;
3556 fmt
= ctx
->ctx_buf_fmt
;
3557 is_system
= ctx
->ctx_fl_system
;
3558 task
= PFM_CTX_TASK(ctx
);
3561 case PFM_CTX_MASKED
:
3563 case PFM_CTX_LOADED
:
3564 if (CTX_HAS_SMPL(ctx
) && fmt
->fmt_restart_active
) break;
3566 case PFM_CTX_UNLOADED
:
3567 case PFM_CTX_ZOMBIE
:
3568 DPRINT(("invalid state=%d\n", state
));
3571 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state
));
3576 * In system wide and when the context is loaded, access can only happen
3577 * when the caller is running on the CPU being monitored by the session.
3578 * It does not have to be the owner (ctx_task) of the context per se.
3580 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
3581 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3586 if (unlikely(task
== NULL
)) {
3587 printk(KERN_ERR
"perfmon: [%d] pfm_restart no task\n", task_pid_nr(current
));
3591 if (task
== current
|| is_system
) {
3593 fmt
= ctx
->ctx_buf_fmt
;
3595 DPRINT(("restarting self %d ovfl=0x%lx\n",
3597 ctx
->ctx_ovfl_regs
[0]));
3599 if (CTX_HAS_SMPL(ctx
)) {
3601 prefetch(ctx
->ctx_smpl_hdr
);
3603 rst_ctrl
.bits
.mask_monitoring
= 0;
3604 rst_ctrl
.bits
.reset_ovfl_pmds
= 0;
3606 if (state
== PFM_CTX_LOADED
)
3607 ret
= pfm_buf_fmt_restart_active(fmt
, task
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
3609 ret
= pfm_buf_fmt_restart(fmt
, task
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
3611 rst_ctrl
.bits
.mask_monitoring
= 0;
3612 rst_ctrl
.bits
.reset_ovfl_pmds
= 1;
3616 if (rst_ctrl
.bits
.reset_ovfl_pmds
)
3617 pfm_reset_regs(ctx
, ctx
->ctx_ovfl_regs
, PFM_PMD_LONG_RESET
);
3619 if (rst_ctrl
.bits
.mask_monitoring
== 0) {
3620 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task
)));
3622 if (state
== PFM_CTX_MASKED
) pfm_restore_monitoring(task
);
3624 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task
)));
3626 // cannot use pfm_stop_monitoring(task, regs);
3630 * clear overflowed PMD mask to remove any stale information
3632 ctx
->ctx_ovfl_regs
[0] = 0UL;
3635 * back to LOADED state
3637 ctx
->ctx_state
= PFM_CTX_LOADED
;
3640 * XXX: not really useful for self monitoring
3642 ctx
->ctx_fl_can_restart
= 0;
3648 * restart another task
3652 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3653 * one is seen by the task.
3655 if (state
== PFM_CTX_MASKED
) {
3656 if (ctx
->ctx_fl_can_restart
== 0) return -EINVAL
;
3658 * will prevent subsequent restart before this one is
3659 * seen by other task
3661 ctx
->ctx_fl_can_restart
= 0;
3665 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3666 * the task is blocked or on its way to block. That's the normal
3667 * restart path. If the monitoring is not masked, then the task
3668 * can be actively monitoring and we cannot directly intervene.
3669 * Therefore we use the trap mechanism to catch the task and
3670 * force it to reset the buffer/reset PMDs.
3672 * if non-blocking, then we ensure that the task will go into
3673 * pfm_handle_work() before returning to user mode.
3675 * We cannot explicitly reset another task, it MUST always
3676 * be done by the task itself. This works for system wide because
3677 * the tool that is controlling the session is logically doing
3678 * "self-monitoring".
3680 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && state
== PFM_CTX_MASKED
) {
3681 DPRINT(("unblocking [%d]\n", task_pid_nr(task
)));
3682 complete(&ctx
->ctx_restart_done
);
3684 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task
)));
3686 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_RESET
;
3688 PFM_SET_WORK_PENDING(task
, 1);
3690 set_notify_resume(task
);
3693 * XXX: send reschedule if task runs on another CPU
3700 pfm_debug(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3702 unsigned int m
= *(unsigned int *)arg
;
3704 pfm_sysctl
.debug
= m
== 0 ? 0 : 1;
3706 printk(KERN_INFO
"perfmon debugging %s (timing reset)\n", pfm_sysctl
.debug
? "on" : "off");
3709 memset(pfm_stats
, 0, sizeof(pfm_stats
));
3710 for(m
=0; m
< NR_CPUS
; m
++) pfm_stats
[m
].pfm_ovfl_intr_cycles_min
= ~0UL;
3716 * arg can be NULL and count can be zero for this function
3719 pfm_write_ibr_dbr(int mode
, pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3721 struct thread_struct
*thread
= NULL
;
3722 struct task_struct
*task
;
3723 pfarg_dbreg_t
*req
= (pfarg_dbreg_t
*)arg
;
3724 unsigned long flags
;
3729 int i
, can_access_pmu
= 0;
3730 int is_system
, is_loaded
;
3732 if (pmu_conf
->use_rr_dbregs
== 0) return -EINVAL
;
3734 state
= ctx
->ctx_state
;
3735 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3736 is_system
= ctx
->ctx_fl_system
;
3737 task
= ctx
->ctx_task
;
3739 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
3742 * on both UP and SMP, we can only write to the PMC when the task is
3743 * the owner of the local PMU.
3746 thread
= &task
->thread
;
3748 * In system wide and when the context is loaded, access can only happen
3749 * when the caller is running on the CPU being monitored by the session.
3750 * It does not have to be the owner (ctx_task) of the context per se.
3752 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3753 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3756 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3760 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3761 * ensuring that no real breakpoint can be installed via this call.
3763 * IMPORTANT: regs can be NULL in this function
3766 first_time
= ctx
->ctx_fl_using_dbreg
== 0;
3769 * don't bother if we are loaded and task is being debugged
3771 if (is_loaded
&& (thread
->flags
& IA64_THREAD_DBG_VALID
) != 0) {
3772 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task
)));
3777 * check for debug registers in system wide mode
3779 * If though a check is done in pfm_context_load(),
3780 * we must repeat it here, in case the registers are
3781 * written after the context is loaded
3786 if (first_time
&& is_system
) {
3787 if (pfm_sessions
.pfs_ptrace_use_dbregs
)
3790 pfm_sessions
.pfs_sys_use_dbregs
++;
3795 if (ret
!= 0) return ret
;
3798 * mark ourself as user of the debug registers for
3801 ctx
->ctx_fl_using_dbreg
= 1;
3804 * clear hardware registers to make sure we don't
3805 * pick up stale state.
3807 * for a system wide session, we do not use
3808 * thread.dbr, thread.ibr because this process
3809 * never leaves the current CPU and the state
3810 * is shared by all processes running on it
3812 if (first_time
&& can_access_pmu
) {
3813 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task
)));
3814 for (i
=0; i
< pmu_conf
->num_ibrs
; i
++) {
3815 ia64_set_ibr(i
, 0UL);
3816 ia64_dv_serialize_instruction();
3819 for (i
=0; i
< pmu_conf
->num_dbrs
; i
++) {
3820 ia64_set_dbr(i
, 0UL);
3821 ia64_dv_serialize_data();
3827 * Now install the values into the registers
3829 for (i
= 0; i
< count
; i
++, req
++) {
3831 rnum
= req
->dbreg_num
;
3832 dbreg
.val
= req
->dbreg_value
;
3836 if ((mode
== PFM_CODE_RR
&& rnum
>= PFM_NUM_IBRS
) || ((mode
== PFM_DATA_RR
) && rnum
>= PFM_NUM_DBRS
)) {
3837 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3838 rnum
, dbreg
.val
, mode
, i
, count
));
3844 * make sure we do not install enabled breakpoint
3847 if (mode
== PFM_CODE_RR
)
3848 dbreg
.ibr
.ibr_x
= 0;
3850 dbreg
.dbr
.dbr_r
= dbreg
.dbr
.dbr_w
= 0;
3853 PFM_REG_RETFLAG_SET(req
->dbreg_flags
, 0);
3856 * Debug registers, just like PMC, can only be modified
3857 * by a kernel call. Moreover, perfmon() access to those
3858 * registers are centralized in this routine. The hardware
3859 * does not modify the value of these registers, therefore,
3860 * if we save them as they are written, we can avoid having
3861 * to save them on context switch out. This is made possible
3862 * by the fact that when perfmon uses debug registers, ptrace()
3863 * won't be able to modify them concurrently.
3865 if (mode
== PFM_CODE_RR
) {
3866 CTX_USED_IBR(ctx
, rnum
);
3868 if (can_access_pmu
) {
3869 ia64_set_ibr(rnum
, dbreg
.val
);
3870 ia64_dv_serialize_instruction();
3873 ctx
->ctx_ibrs
[rnum
] = dbreg
.val
;
3875 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3876 rnum
, dbreg
.val
, ctx
->ctx_used_ibrs
[0], is_loaded
, can_access_pmu
));
3878 CTX_USED_DBR(ctx
, rnum
);
3880 if (can_access_pmu
) {
3881 ia64_set_dbr(rnum
, dbreg
.val
);
3882 ia64_dv_serialize_data();
3884 ctx
->ctx_dbrs
[rnum
] = dbreg
.val
;
3886 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3887 rnum
, dbreg
.val
, ctx
->ctx_used_dbrs
[0], is_loaded
, can_access_pmu
));
3895 * in case it was our first attempt, we undo the global modifications
3899 if (ctx
->ctx_fl_system
) {
3900 pfm_sessions
.pfs_sys_use_dbregs
--;
3903 ctx
->ctx_fl_using_dbreg
= 0;
3906 * install error return flag
3908 PFM_REG_RETFLAG_SET(req
->dbreg_flags
, PFM_REG_RETFL_EINVAL
);
3914 pfm_write_ibrs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3916 return pfm_write_ibr_dbr(PFM_CODE_RR
, ctx
, arg
, count
, regs
);
3920 pfm_write_dbrs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3922 return pfm_write_ibr_dbr(PFM_DATA_RR
, ctx
, arg
, count
, regs
);
3926 pfm_mod_write_ibrs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3930 if (req
== NULL
) return -EINVAL
;
3932 ctx
= GET_PMU_CTX();
3934 if (ctx
== NULL
) return -EINVAL
;
3937 * for now limit to current task, which is enough when calling
3938 * from overflow handler
3940 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3942 return pfm_write_ibrs(ctx
, req
, nreq
, regs
);
3944 EXPORT_SYMBOL(pfm_mod_write_ibrs
);
3947 pfm_mod_write_dbrs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3951 if (req
== NULL
) return -EINVAL
;
3953 ctx
= GET_PMU_CTX();
3955 if (ctx
== NULL
) return -EINVAL
;
3958 * for now limit to current task, which is enough when calling
3959 * from overflow handler
3961 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3963 return pfm_write_dbrs(ctx
, req
, nreq
, regs
);
3965 EXPORT_SYMBOL(pfm_mod_write_dbrs
);
3969 pfm_get_features(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3971 pfarg_features_t
*req
= (pfarg_features_t
*)arg
;
3973 req
->ft_version
= PFM_VERSION
;
3978 pfm_stop(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3980 struct pt_regs
*tregs
;
3981 struct task_struct
*task
= PFM_CTX_TASK(ctx
);
3982 int state
, is_system
;
3984 state
= ctx
->ctx_state
;
3985 is_system
= ctx
->ctx_fl_system
;
3988 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3990 if (state
== PFM_CTX_UNLOADED
) return -EINVAL
;
3993 * In system wide and when the context is loaded, access can only happen
3994 * when the caller is running on the CPU being monitored by the session.
3995 * It does not have to be the owner (ctx_task) of the context per se.
3997 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
3998 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
4001 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4002 task_pid_nr(PFM_CTX_TASK(ctx
)),
4006 * in system mode, we need to update the PMU directly
4007 * and the user level state of the caller, which may not
4008 * necessarily be the creator of the context.
4012 * Update local PMU first
4016 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) & ~IA64_DCR_PP
);
4020 * update local cpuinfo
4022 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP
);
4025 * stop monitoring, does srlz.i
4030 * stop monitoring in the caller
4032 ia64_psr(regs
)->pp
= 0;
4040 if (task
== current
) {
4041 /* stop monitoring at kernel level */
4045 * stop monitoring at the user level
4047 ia64_psr(regs
)->up
= 0;
4049 tregs
= task_pt_regs(task
);
4052 * stop monitoring at the user level
4054 ia64_psr(tregs
)->up
= 0;
4057 * monitoring disabled in kernel at next reschedule
4059 ctx
->ctx_saved_psr_up
= 0;
4060 DPRINT(("task=[%d]\n", task_pid_nr(task
)));
4067 pfm_start(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4069 struct pt_regs
*tregs
;
4070 int state
, is_system
;
4072 state
= ctx
->ctx_state
;
4073 is_system
= ctx
->ctx_fl_system
;
4075 if (state
!= PFM_CTX_LOADED
) return -EINVAL
;
4078 * In system wide and when the context is loaded, access can only happen
4079 * when the caller is running on the CPU being monitored by the session.
4080 * It does not have to be the owner (ctx_task) of the context per se.
4082 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
4083 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
4088 * in system mode, we need to update the PMU directly
4089 * and the user level state of the caller, which may not
4090 * necessarily be the creator of the context.
4095 * set user level psr.pp for the caller
4097 ia64_psr(regs
)->pp
= 1;
4100 * now update the local PMU and cpuinfo
4102 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP
);
4105 * start monitoring at kernel level
4110 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) | IA64_DCR_PP
);
4120 if (ctx
->ctx_task
== current
) {
4122 /* start monitoring at kernel level */
4126 * activate monitoring at user level
4128 ia64_psr(regs
)->up
= 1;
4131 tregs
= task_pt_regs(ctx
->ctx_task
);
4134 * start monitoring at the kernel level the next
4135 * time the task is scheduled
4137 ctx
->ctx_saved_psr_up
= IA64_PSR_UP
;
4140 * activate monitoring at user level
4142 ia64_psr(tregs
)->up
= 1;
4148 pfm_get_pmc_reset(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4150 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
4155 for (i
= 0; i
< count
; i
++, req
++) {
4157 cnum
= req
->reg_num
;
4159 if (!PMC_IS_IMPL(cnum
)) goto abort_mission
;
4161 req
->reg_value
= PMC_DFL_VAL(cnum
);
4163 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
4165 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum
, req
->reg_value
));
4170 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
4175 pfm_check_task_exist(pfm_context_t
*ctx
)
4177 struct task_struct
*g
, *t
;
4180 read_lock(&tasklist_lock
);
4182 do_each_thread (g
, t
) {
4183 if (t
->thread
.pfm_context
== ctx
) {
4187 } while_each_thread (g
, t
);
4189 read_unlock(&tasklist_lock
);
4191 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret
, ctx
));
4197 pfm_context_load(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4199 struct task_struct
*task
;
4200 struct thread_struct
*thread
;
4201 struct pfm_context_t
*old
;
4202 unsigned long flags
;
4204 struct task_struct
*owner_task
= NULL
;
4206 pfarg_load_t
*req
= (pfarg_load_t
*)arg
;
4207 unsigned long *pmcs_source
, *pmds_source
;
4210 int state
, is_system
, set_dbregs
= 0;
4212 state
= ctx
->ctx_state
;
4213 is_system
= ctx
->ctx_fl_system
;
4215 * can only load from unloaded or terminated state
4217 if (state
!= PFM_CTX_UNLOADED
) {
4218 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4224 DPRINT(("load_pid [%d] using_dbreg=%d\n", req
->load_pid
, ctx
->ctx_fl_using_dbreg
));
4226 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && req
->load_pid
== current
->pid
) {
4227 DPRINT(("cannot use blocking mode on self\n"));
4231 ret
= pfm_get_task(ctx
, req
->load_pid
, &task
);
4233 DPRINT(("load_pid [%d] get_task=%d\n", req
->load_pid
, ret
));
4240 * system wide is self monitoring only
4242 if (is_system
&& task
!= current
) {
4243 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4248 thread
= &task
->thread
;
4252 * cannot load a context which is using range restrictions,
4253 * into a task that is being debugged.
4255 if (ctx
->ctx_fl_using_dbreg
) {
4256 if (thread
->flags
& IA64_THREAD_DBG_VALID
) {
4258 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req
->load_pid
));
4264 if (pfm_sessions
.pfs_ptrace_use_dbregs
) {
4265 DPRINT(("cannot load [%d] dbregs in use\n",
4266 task_pid_nr(task
)));
4269 pfm_sessions
.pfs_sys_use_dbregs
++;
4270 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task
), pfm_sessions
.pfs_sys_use_dbregs
));
4277 if (ret
) goto error
;
4281 * SMP system-wide monitoring implies self-monitoring.
4283 * The programming model expects the task to
4284 * be pinned on a CPU throughout the session.
4285 * Here we take note of the current CPU at the
4286 * time the context is loaded. No call from
4287 * another CPU will be allowed.
4289 * The pinning via shed_setaffinity()
4290 * must be done by the calling task prior
4293 * systemwide: keep track of CPU this session is supposed to run on
4295 the_cpu
= ctx
->ctx_cpu
= smp_processor_id();
4299 * now reserve the session
4301 ret
= pfm_reserve_session(current
, is_system
, the_cpu
);
4302 if (ret
) goto error
;
4305 * task is necessarily stopped at this point.
4307 * If the previous context was zombie, then it got removed in
4308 * pfm_save_regs(). Therefore we should not see it here.
4309 * If we see a context, then this is an active context
4311 * XXX: needs to be atomic
4313 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4314 thread
->pfm_context
, ctx
));
4317 old
= ia64_cmpxchg(acq
, &thread
->pfm_context
, NULL
, ctx
, sizeof(pfm_context_t
*));
4319 DPRINT(("load_pid [%d] already has a context\n", req
->load_pid
));
4323 pfm_reset_msgq(ctx
);
4325 ctx
->ctx_state
= PFM_CTX_LOADED
;
4328 * link context to task
4330 ctx
->ctx_task
= task
;
4334 * we load as stopped
4336 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE
);
4337 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP
);
4339 if (ctx
->ctx_fl_excl_idle
) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE
);
4341 thread
->flags
|= IA64_THREAD_PM_VALID
;
4345 * propagate into thread-state
4347 pfm_copy_pmds(task
, ctx
);
4348 pfm_copy_pmcs(task
, ctx
);
4350 pmcs_source
= ctx
->th_pmcs
;
4351 pmds_source
= ctx
->th_pmds
;
4354 * always the case for system-wide
4356 if (task
== current
) {
4358 if (is_system
== 0) {
4360 /* allow user level control */
4361 ia64_psr(regs
)->sp
= 0;
4362 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task
)));
4364 SET_LAST_CPU(ctx
, smp_processor_id());
4366 SET_ACTIVATION(ctx
);
4369 * push the other task out, if any
4371 owner_task
= GET_PMU_OWNER();
4372 if (owner_task
) pfm_lazy_save_regs(owner_task
);
4376 * load all PMD from ctx to PMU (as opposed to thread state)
4377 * restore all PMC from ctx to PMU
4379 pfm_restore_pmds(pmds_source
, ctx
->ctx_all_pmds
[0]);
4380 pfm_restore_pmcs(pmcs_source
, ctx
->ctx_all_pmcs
[0]);
4382 ctx
->ctx_reload_pmcs
[0] = 0UL;
4383 ctx
->ctx_reload_pmds
[0] = 0UL;
4386 * guaranteed safe by earlier check against DBG_VALID
4388 if (ctx
->ctx_fl_using_dbreg
) {
4389 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
4390 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
4395 SET_PMU_OWNER(task
, ctx
);
4397 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task
)));
4400 * when not current, task MUST be stopped, so this is safe
4402 regs
= task_pt_regs(task
);
4404 /* force a full reload */
4405 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
4406 SET_LAST_CPU(ctx
, -1);
4408 /* initial saved psr (stopped) */
4409 ctx
->ctx_saved_psr_up
= 0UL;
4410 ia64_psr(regs
)->up
= ia64_psr(regs
)->pp
= 0;
4416 if (ret
) pfm_unreserve_session(ctx
, ctx
->ctx_fl_system
, the_cpu
);
4419 * we must undo the dbregs setting (for system-wide)
4421 if (ret
&& set_dbregs
) {
4423 pfm_sessions
.pfs_sys_use_dbregs
--;
4427 * release task, there is now a link with the context
4429 if (is_system
== 0 && task
!= current
) {
4433 ret
= pfm_check_task_exist(ctx
);
4435 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
4436 ctx
->ctx_task
= NULL
;
4444 * in this function, we do not need to increase the use count
4445 * for the task via get_task_struct(), because we hold the
4446 * context lock. If the task were to disappear while having
4447 * a context attached, it would go through pfm_exit_thread()
4448 * which also grabs the context lock and would therefore be blocked
4449 * until we are here.
4451 static void pfm_flush_pmds(struct task_struct
*, pfm_context_t
*ctx
);
4454 pfm_context_unload(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4456 struct task_struct
*task
= PFM_CTX_TASK(ctx
);
4457 struct pt_regs
*tregs
;
4458 int prev_state
, is_system
;
4461 DPRINT(("ctx_state=%d task [%d]\n", ctx
->ctx_state
, task
? task_pid_nr(task
) : -1));
4463 prev_state
= ctx
->ctx_state
;
4464 is_system
= ctx
->ctx_fl_system
;
4467 * unload only when necessary
4469 if (prev_state
== PFM_CTX_UNLOADED
) {
4470 DPRINT(("ctx_state=%d, nothing to do\n", prev_state
));
4475 * clear psr and dcr bits
4477 ret
= pfm_stop(ctx
, NULL
, 0, regs
);
4478 if (ret
) return ret
;
4480 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
4483 * in system mode, we need to update the PMU directly
4484 * and the user level state of the caller, which may not
4485 * necessarily be the creator of the context.
4492 * local PMU is taken care of in pfm_stop()
4494 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE
);
4495 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE
);
4498 * save PMDs in context
4501 pfm_flush_pmds(current
, ctx
);
4504 * at this point we are done with the PMU
4505 * so we can unreserve the resource.
4507 if (prev_state
!= PFM_CTX_ZOMBIE
)
4508 pfm_unreserve_session(ctx
, 1 , ctx
->ctx_cpu
);
4511 * disconnect context from task
4513 task
->thread
.pfm_context
= NULL
;
4515 * disconnect task from context
4517 ctx
->ctx_task
= NULL
;
4520 * There is nothing more to cleanup here.
4528 tregs
= task
== current
? regs
: task_pt_regs(task
);
4530 if (task
== current
) {
4532 * cancel user level control
4534 ia64_psr(regs
)->sp
= 1;
4536 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task
)));
4539 * save PMDs to context
4542 pfm_flush_pmds(task
, ctx
);
4545 * at this point we are done with the PMU
4546 * so we can unreserve the resource.
4548 * when state was ZOMBIE, we have already unreserved.
4550 if (prev_state
!= PFM_CTX_ZOMBIE
)
4551 pfm_unreserve_session(ctx
, 0 , ctx
->ctx_cpu
);
4554 * reset activation counter and psr
4556 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
4557 SET_LAST_CPU(ctx
, -1);
4560 * PMU state will not be restored
4562 task
->thread
.flags
&= ~IA64_THREAD_PM_VALID
;
4565 * break links between context and task
4567 task
->thread
.pfm_context
= NULL
;
4568 ctx
->ctx_task
= NULL
;
4570 PFM_SET_WORK_PENDING(task
, 0);
4572 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_NONE
;
4573 ctx
->ctx_fl_can_restart
= 0;
4574 ctx
->ctx_fl_going_zombie
= 0;
4576 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task
)));
4583 * called only from exit_thread(): task == current
4584 * we come here only if current has a context attached (loaded or masked)
4587 pfm_exit_thread(struct task_struct
*task
)
4590 unsigned long flags
;
4591 struct pt_regs
*regs
= task_pt_regs(task
);
4595 ctx
= PFM_GET_CTX(task
);
4597 PROTECT_CTX(ctx
, flags
);
4599 DPRINT(("state=%d task [%d]\n", ctx
->ctx_state
, task_pid_nr(task
)));
4601 state
= ctx
->ctx_state
;
4603 case PFM_CTX_UNLOADED
:
4605 * only comes to this function if pfm_context is not NULL, i.e., cannot
4606 * be in unloaded state
4608 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task
));
4610 case PFM_CTX_LOADED
:
4611 case PFM_CTX_MASKED
:
4612 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
4614 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task
), state
, ret
);
4616 DPRINT(("ctx unloaded for current state was %d\n", state
));
4618 pfm_end_notify_user(ctx
);
4620 case PFM_CTX_ZOMBIE
:
4621 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
4623 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task
), state
, ret
);
4628 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task
), state
);
4631 UNPROTECT_CTX(ctx
, flags
);
4633 { u64 psr
= pfm_get_psr();
4634 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
4635 BUG_ON(GET_PMU_OWNER());
4636 BUG_ON(ia64_psr(regs
)->up
);
4637 BUG_ON(ia64_psr(regs
)->pp
);
4641 * All memory free operations (especially for vmalloc'ed memory)
4642 * MUST be done with interrupts ENABLED.
4644 if (free_ok
) pfm_context_free(ctx
);
4648 * functions MUST be listed in the increasing order of their index (see permfon.h)
4650 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4651 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4652 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4653 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4654 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4656 static pfm_cmd_desc_t pfm_cmd_tab
[]={
4657 /* 0 */PFM_CMD_NONE
,
4658 /* 1 */PFM_CMD(pfm_write_pmcs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4659 /* 2 */PFM_CMD(pfm_write_pmds
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4660 /* 3 */PFM_CMD(pfm_read_pmds
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4661 /* 4 */PFM_CMD_S(pfm_stop
, PFM_CMD_PCLRWS
),
4662 /* 5 */PFM_CMD_S(pfm_start
, PFM_CMD_PCLRWS
),
4663 /* 6 */PFM_CMD_NONE
,
4664 /* 7 */PFM_CMD_NONE
,
4665 /* 8 */PFM_CMD(pfm_context_create
, PFM_CMD_ARG_RW
, 1, pfarg_context_t
, pfm_ctx_getsize
),
4666 /* 9 */PFM_CMD_NONE
,
4667 /* 10 */PFM_CMD_S(pfm_restart
, PFM_CMD_PCLRW
),
4668 /* 11 */PFM_CMD_NONE
,
4669 /* 12 */PFM_CMD(pfm_get_features
, PFM_CMD_ARG_RW
, 1, pfarg_features_t
, NULL
),
4670 /* 13 */PFM_CMD(pfm_debug
, 0, 1, unsigned int, NULL
),
4671 /* 14 */PFM_CMD_NONE
,
4672 /* 15 */PFM_CMD(pfm_get_pmc_reset
, PFM_CMD_ARG_RW
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4673 /* 16 */PFM_CMD(pfm_context_load
, PFM_CMD_PCLRWS
, 1, pfarg_load_t
, NULL
),
4674 /* 17 */PFM_CMD_S(pfm_context_unload
, PFM_CMD_PCLRWS
),
4675 /* 18 */PFM_CMD_NONE
,
4676 /* 19 */PFM_CMD_NONE
,
4677 /* 20 */PFM_CMD_NONE
,
4678 /* 21 */PFM_CMD_NONE
,
4679 /* 22 */PFM_CMD_NONE
,
4680 /* 23 */PFM_CMD_NONE
,
4681 /* 24 */PFM_CMD_NONE
,
4682 /* 25 */PFM_CMD_NONE
,
4683 /* 26 */PFM_CMD_NONE
,
4684 /* 27 */PFM_CMD_NONE
,
4685 /* 28 */PFM_CMD_NONE
,
4686 /* 29 */PFM_CMD_NONE
,
4687 /* 30 */PFM_CMD_NONE
,
4688 /* 31 */PFM_CMD_NONE
,
4689 /* 32 */PFM_CMD(pfm_write_ibrs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_dbreg_t
, NULL
),
4690 /* 33 */PFM_CMD(pfm_write_dbrs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_dbreg_t
, NULL
)
4692 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4695 pfm_check_task_state(pfm_context_t
*ctx
, int cmd
, unsigned long flags
)
4697 struct task_struct
*task
;
4698 int state
, old_state
;
4701 state
= ctx
->ctx_state
;
4702 task
= ctx
->ctx_task
;
4705 DPRINT(("context %d no task, state=%d\n", ctx
->ctx_fd
, state
));
4709 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4713 task
->state
, PFM_CMD_STOPPED(cmd
)));
4716 * self-monitoring always ok.
4718 * for system-wide the caller can either be the creator of the
4719 * context (to one to which the context is attached to) OR
4720 * a task running on the same CPU as the session.
4722 if (task
== current
|| ctx
->ctx_fl_system
) return 0;
4725 * we are monitoring another thread
4728 case PFM_CTX_UNLOADED
:
4730 * if context is UNLOADED we are safe to go
4733 case PFM_CTX_ZOMBIE
:
4735 * no command can operate on a zombie context
4737 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd
));
4739 case PFM_CTX_MASKED
:
4741 * PMU state has been saved to software even though
4742 * the thread may still be running.
4744 if (cmd
!= PFM_UNLOAD_CONTEXT
) return 0;
4748 * context is LOADED or MASKED. Some commands may need to have
4751 * We could lift this restriction for UP but it would mean that
4752 * the user has no guarantee the task would not run between
4753 * two successive calls to perfmonctl(). That's probably OK.
4754 * If this user wants to ensure the task does not run, then
4755 * the task must be stopped.
4757 if (PFM_CMD_STOPPED(cmd
)) {
4758 if (!task_is_stopped_or_traced(task
)) {
4759 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task
)));
4763 * task is now stopped, wait for ctxsw out
4765 * This is an interesting point in the code.
4766 * We need to unprotect the context because
4767 * the pfm_save_regs() routines needs to grab
4768 * the same lock. There are danger in doing
4769 * this because it leaves a window open for
4770 * another task to get access to the context
4771 * and possibly change its state. The one thing
4772 * that is not possible is for the context to disappear
4773 * because we are protected by the VFS layer, i.e.,
4774 * get_fd()/put_fd().
4778 UNPROTECT_CTX(ctx
, flags
);
4780 wait_task_inactive(task
, 0);
4782 PROTECT_CTX(ctx
, flags
);
4785 * we must recheck to verify if state has changed
4787 if (ctx
->ctx_state
!= old_state
) {
4788 DPRINT(("old_state=%d new_state=%d\n", old_state
, ctx
->ctx_state
));
4796 * system-call entry point (must return long)
4799 sys_perfmonctl (int fd
, int cmd
, void __user
*arg
, int count
)
4801 struct file
*file
= NULL
;
4802 pfm_context_t
*ctx
= NULL
;
4803 unsigned long flags
= 0UL;
4804 void *args_k
= NULL
;
4805 long ret
; /* will expand int return types */
4806 size_t base_sz
, sz
, xtra_sz
= 0;
4807 int narg
, completed_args
= 0, call_made
= 0, cmd_flags
;
4808 int (*func
)(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
4809 int (*getsize
)(void *arg
, size_t *sz
);
4810 #define PFM_MAX_ARGSIZE 4096
4813 * reject any call if perfmon was disabled at initialization
4815 if (unlikely(pmu_conf
== NULL
)) return -ENOSYS
;
4817 if (unlikely(cmd
< 0 || cmd
>= PFM_CMD_COUNT
)) {
4818 DPRINT(("invalid cmd=%d\n", cmd
));
4822 func
= pfm_cmd_tab
[cmd
].cmd_func
;
4823 narg
= pfm_cmd_tab
[cmd
].cmd_narg
;
4824 base_sz
= pfm_cmd_tab
[cmd
].cmd_argsize
;
4825 getsize
= pfm_cmd_tab
[cmd
].cmd_getsize
;
4826 cmd_flags
= pfm_cmd_tab
[cmd
].cmd_flags
;
4828 if (unlikely(func
== NULL
)) {
4829 DPRINT(("invalid cmd=%d\n", cmd
));
4833 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4841 * check if number of arguments matches what the command expects
4843 if (unlikely((narg
== PFM_CMD_ARG_MANY
&& count
<= 0) || (narg
> 0 && narg
!= count
)))
4847 sz
= xtra_sz
+ base_sz
*count
;
4849 * limit abuse to min page size
4851 if (unlikely(sz
> PFM_MAX_ARGSIZE
)) {
4852 printk(KERN_ERR
"perfmon: [%d] argument too big %lu\n", task_pid_nr(current
), sz
);
4857 * allocate default-sized argument buffer
4859 if (likely(count
&& args_k
== NULL
)) {
4860 args_k
= kmalloc(PFM_MAX_ARGSIZE
, GFP_KERNEL
);
4861 if (args_k
== NULL
) return -ENOMEM
;
4869 * assume sz = 0 for command without parameters
4871 if (sz
&& copy_from_user(args_k
, arg
, sz
)) {
4872 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz
, arg
));
4877 * check if command supports extra parameters
4879 if (completed_args
== 0 && getsize
) {
4881 * get extra parameters size (based on main argument)
4883 ret
= (*getsize
)(args_k
, &xtra_sz
);
4884 if (ret
) goto error_args
;
4888 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz
, xtra_sz
));
4890 /* retry if necessary */
4891 if (likely(xtra_sz
)) goto restart_args
;
4894 if (unlikely((cmd_flags
& PFM_CMD_FD
) == 0)) goto skip_fd
;
4899 if (unlikely(file
== NULL
)) {
4900 DPRINT(("invalid fd %d\n", fd
));
4903 if (unlikely(PFM_IS_FILE(file
) == 0)) {
4904 DPRINT(("fd %d not related to perfmon\n", fd
));
4908 ctx
= (pfm_context_t
*)file
->private_data
;
4909 if (unlikely(ctx
== NULL
)) {
4910 DPRINT(("no context for fd %d\n", fd
));
4913 prefetch(&ctx
->ctx_state
);
4915 PROTECT_CTX(ctx
, flags
);
4918 * check task is stopped
4920 ret
= pfm_check_task_state(ctx
, cmd
, flags
);
4921 if (unlikely(ret
)) goto abort_locked
;
4924 ret
= (*func
)(ctx
, args_k
, count
, task_pt_regs(current
));
4930 DPRINT(("context unlocked\n"));
4931 UNPROTECT_CTX(ctx
, flags
);
4934 /* copy argument back to user, if needed */
4935 if (call_made
&& PFM_CMD_RW_ARG(cmd
) && copy_to_user(arg
, args_k
, base_sz
*count
)) ret
= -EFAULT
;
4943 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd
), ret
));
4949 pfm_resume_after_ovfl(pfm_context_t
*ctx
, unsigned long ovfl_regs
, struct pt_regs
*regs
)
4951 pfm_buffer_fmt_t
*fmt
= ctx
->ctx_buf_fmt
;
4952 pfm_ovfl_ctrl_t rst_ctrl
;
4956 state
= ctx
->ctx_state
;
4958 * Unlock sampling buffer and reset index atomically
4959 * XXX: not really needed when blocking
4961 if (CTX_HAS_SMPL(ctx
)) {
4963 rst_ctrl
.bits
.mask_monitoring
= 0;
4964 rst_ctrl
.bits
.reset_ovfl_pmds
= 0;
4966 if (state
== PFM_CTX_LOADED
)
4967 ret
= pfm_buf_fmt_restart_active(fmt
, current
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
4969 ret
= pfm_buf_fmt_restart(fmt
, current
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
4971 rst_ctrl
.bits
.mask_monitoring
= 0;
4972 rst_ctrl
.bits
.reset_ovfl_pmds
= 1;
4976 if (rst_ctrl
.bits
.reset_ovfl_pmds
) {
4977 pfm_reset_regs(ctx
, &ovfl_regs
, PFM_PMD_LONG_RESET
);
4979 if (rst_ctrl
.bits
.mask_monitoring
== 0) {
4980 DPRINT(("resuming monitoring\n"));
4981 if (ctx
->ctx_state
== PFM_CTX_MASKED
) pfm_restore_monitoring(current
);
4983 DPRINT(("stopping monitoring\n"));
4984 //pfm_stop_monitoring(current, regs);
4986 ctx
->ctx_state
= PFM_CTX_LOADED
;
4991 * context MUST BE LOCKED when calling
4992 * can only be called for current
4995 pfm_context_force_terminate(pfm_context_t
*ctx
, struct pt_regs
*regs
)
4999 DPRINT(("entering for [%d]\n", task_pid_nr(current
)));
5001 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
5003 printk(KERN_ERR
"pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current
), ret
);
5007 * and wakeup controlling task, indicating we are now disconnected
5009 wake_up_interruptible(&ctx
->ctx_zombieq
);
5012 * given that context is still locked, the controlling
5013 * task will only get access when we return from
5014 * pfm_handle_work().
5018 static int pfm_ovfl_notify_user(pfm_context_t
*ctx
, unsigned long ovfl_pmds
);
5021 * pfm_handle_work() can be called with interrupts enabled
5022 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5023 * call may sleep, therefore we must re-enable interrupts
5024 * to avoid deadlocks. It is safe to do so because this function
5025 * is called ONLY when returning to user level (pUStk=1), in which case
5026 * there is no risk of kernel stack overflow due to deep
5027 * interrupt nesting.
5030 pfm_handle_work(void)
5033 struct pt_regs
*regs
;
5034 unsigned long flags
, dummy_flags
;
5035 unsigned long ovfl_regs
;
5036 unsigned int reason
;
5039 ctx
= PFM_GET_CTX(current
);
5041 printk(KERN_ERR
"perfmon: [%d] has no PFM context\n",
5042 task_pid_nr(current
));
5046 PROTECT_CTX(ctx
, flags
);
5048 PFM_SET_WORK_PENDING(current
, 0);
5050 regs
= task_pt_regs(current
);
5053 * extract reason for being here and clear
5055 reason
= ctx
->ctx_fl_trap_reason
;
5056 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_NONE
;
5057 ovfl_regs
= ctx
->ctx_ovfl_regs
[0];
5059 DPRINT(("reason=%d state=%d\n", reason
, ctx
->ctx_state
));
5062 * must be done before we check for simple-reset mode
5064 if (ctx
->ctx_fl_going_zombie
|| ctx
->ctx_state
== PFM_CTX_ZOMBIE
)
5067 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5068 if (reason
== PFM_TRAP_REASON_RESET
)
5072 * restore interrupt mask to what it was on entry.
5073 * Could be enabled/diasbled.
5075 UNPROTECT_CTX(ctx
, flags
);
5078 * force interrupt enable because of down_interruptible()
5082 DPRINT(("before block sleeping\n"));
5085 * may go through without blocking on SMP systems
5086 * if restart has been received already by the time we call down()
5088 ret
= wait_for_completion_interruptible(&ctx
->ctx_restart_done
);
5090 DPRINT(("after block sleeping ret=%d\n", ret
));
5093 * lock context and mask interrupts again
5094 * We save flags into a dummy because we may have
5095 * altered interrupts mask compared to entry in this
5098 PROTECT_CTX(ctx
, dummy_flags
);
5101 * we need to read the ovfl_regs only after wake-up
5102 * because we may have had pfm_write_pmds() in between
5103 * and that can changed PMD values and therefore
5104 * ovfl_regs is reset for these new PMD values.
5106 ovfl_regs
= ctx
->ctx_ovfl_regs
[0];
5108 if (ctx
->ctx_fl_going_zombie
) {
5110 DPRINT(("context is zombie, bailing out\n"));
5111 pfm_context_force_terminate(ctx
, regs
);
5115 * in case of interruption of down() we don't restart anything
5121 pfm_resume_after_ovfl(ctx
, ovfl_regs
, regs
);
5122 ctx
->ctx_ovfl_regs
[0] = 0UL;
5126 * restore flags as they were upon entry
5128 UNPROTECT_CTX(ctx
, flags
);
5132 pfm_notify_user(pfm_context_t
*ctx
, pfm_msg_t
*msg
)
5134 if (ctx
->ctx_state
== PFM_CTX_ZOMBIE
) {
5135 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5139 DPRINT(("waking up somebody\n"));
5141 if (msg
) wake_up_interruptible(&ctx
->ctx_msgq_wait
);
5144 * safe, we are not in intr handler, nor in ctxsw when
5147 kill_fasync (&ctx
->ctx_async_queue
, SIGIO
, POLL_IN
);
5153 pfm_ovfl_notify_user(pfm_context_t
*ctx
, unsigned long ovfl_pmds
)
5155 pfm_msg_t
*msg
= NULL
;
5157 if (ctx
->ctx_fl_no_msg
== 0) {
5158 msg
= pfm_get_new_msg(ctx
);
5160 printk(KERN_ERR
"perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5164 msg
->pfm_ovfl_msg
.msg_type
= PFM_MSG_OVFL
;
5165 msg
->pfm_ovfl_msg
.msg_ctx_fd
= ctx
->ctx_fd
;
5166 msg
->pfm_ovfl_msg
.msg_active_set
= 0;
5167 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[0] = ovfl_pmds
;
5168 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[1] = 0UL;
5169 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[2] = 0UL;
5170 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[3] = 0UL;
5171 msg
->pfm_ovfl_msg
.msg_tstamp
= 0UL;
5174 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5180 return pfm_notify_user(ctx
, msg
);
5184 pfm_end_notify_user(pfm_context_t
*ctx
)
5188 msg
= pfm_get_new_msg(ctx
);
5190 printk(KERN_ERR
"perfmon: pfm_end_notify_user no more notification msgs\n");
5194 memset(msg
, 0, sizeof(*msg
));
5196 msg
->pfm_end_msg
.msg_type
= PFM_MSG_END
;
5197 msg
->pfm_end_msg
.msg_ctx_fd
= ctx
->ctx_fd
;
5198 msg
->pfm_ovfl_msg
.msg_tstamp
= 0UL;
5200 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5205 return pfm_notify_user(ctx
, msg
);
5209 * main overflow processing routine.
5210 * it can be called from the interrupt path or explicitly during the context switch code
5212 static void pfm_overflow_handler(struct task_struct
*task
, pfm_context_t
*ctx
,
5213 unsigned long pmc0
, struct pt_regs
*regs
)
5215 pfm_ovfl_arg_t
*ovfl_arg
;
5217 unsigned long old_val
, ovfl_val
, new_val
;
5218 unsigned long ovfl_notify
= 0UL, ovfl_pmds
= 0UL, smpl_pmds
= 0UL, reset_pmds
;
5219 unsigned long tstamp
;
5220 pfm_ovfl_ctrl_t ovfl_ctrl
;
5221 unsigned int i
, has_smpl
;
5222 int must_notify
= 0;
5224 if (unlikely(ctx
->ctx_state
== PFM_CTX_ZOMBIE
)) goto stop_monitoring
;
5227 * sanity test. Should never happen
5229 if (unlikely((pmc0
& 0x1) == 0)) goto sanity_check
;
5231 tstamp
= ia64_get_itc();
5232 mask
= pmc0
>> PMU_FIRST_COUNTER
;
5233 ovfl_val
= pmu_conf
->ovfl_val
;
5234 has_smpl
= CTX_HAS_SMPL(ctx
);
5236 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5237 "used_pmds=0x%lx\n",
5239 task
? task_pid_nr(task
): -1,
5240 (regs
? regs
->cr_iip
: 0),
5241 CTX_OVFL_NOBLOCK(ctx
) ? "nonblocking" : "blocking",
5242 ctx
->ctx_used_pmds
[0]));
5246 * first we update the virtual counters
5247 * assume there was a prior ia64_srlz_d() issued
5249 for (i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
5251 /* skip pmd which did not overflow */
5252 if ((mask
& 0x1) == 0) continue;
5255 * Note that the pmd is not necessarily 0 at this point as qualified events
5256 * may have happened before the PMU was frozen. The residual count is not
5257 * taken into consideration here but will be with any read of the pmd via
5260 old_val
= new_val
= ctx
->ctx_pmds
[i
].val
;
5261 new_val
+= 1 + ovfl_val
;
5262 ctx
->ctx_pmds
[i
].val
= new_val
;
5265 * check for overflow condition
5267 if (likely(old_val
> new_val
)) {
5268 ovfl_pmds
|= 1UL << i
;
5269 if (PMC_OVFL_NOTIFY(ctx
, i
)) ovfl_notify
|= 1UL << i
;
5272 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5276 ia64_get_pmd(i
) & ovfl_val
,
5282 * there was no 64-bit overflow, nothing else to do
5284 if (ovfl_pmds
== 0UL) return;
5287 * reset all control bits
5293 * if a sampling format module exists, then we "cache" the overflow by
5294 * calling the module's handler() routine.
5297 unsigned long start_cycles
, end_cycles
;
5298 unsigned long pmd_mask
;
5300 int this_cpu
= smp_processor_id();
5302 pmd_mask
= ovfl_pmds
>> PMU_FIRST_COUNTER
;
5303 ovfl_arg
= &ctx
->ctx_ovfl_arg
;
5305 prefetch(ctx
->ctx_smpl_hdr
);
5307 for(i
=PMU_FIRST_COUNTER
; pmd_mask
&& ret
== 0; i
++, pmd_mask
>>=1) {
5311 if ((pmd_mask
& 0x1) == 0) continue;
5313 ovfl_arg
->ovfl_pmd
= (unsigned char )i
;
5314 ovfl_arg
->ovfl_notify
= ovfl_notify
& mask
? 1 : 0;
5315 ovfl_arg
->active_set
= 0;
5316 ovfl_arg
->ovfl_ctrl
.val
= 0; /* module must fill in all fields */
5317 ovfl_arg
->smpl_pmds
[0] = smpl_pmds
= ctx
->ctx_pmds
[i
].smpl_pmds
[0];
5319 ovfl_arg
->pmd_value
= ctx
->ctx_pmds
[i
].val
;
5320 ovfl_arg
->pmd_last_reset
= ctx
->ctx_pmds
[i
].lval
;
5321 ovfl_arg
->pmd_eventid
= ctx
->ctx_pmds
[i
].eventid
;
5324 * copy values of pmds of interest. Sampling format may copy them
5325 * into sampling buffer.
5328 for(j
=0, k
=0; smpl_pmds
; j
++, smpl_pmds
>>=1) {
5329 if ((smpl_pmds
& 0x1) == 0) continue;
5330 ovfl_arg
->smpl_pmds_values
[k
++] = PMD_IS_COUNTING(j
) ? pfm_read_soft_counter(ctx
, j
) : ia64_get_pmd(j
);
5331 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k
-1, j
, ovfl_arg
->smpl_pmds_values
[k
-1]));
5335 pfm_stats
[this_cpu
].pfm_smpl_handler_calls
++;
5337 start_cycles
= ia64_get_itc();
5340 * call custom buffer format record (handler) routine
5342 ret
= (*ctx
->ctx_buf_fmt
->fmt_handler
)(task
, ctx
->ctx_smpl_hdr
, ovfl_arg
, regs
, tstamp
);
5344 end_cycles
= ia64_get_itc();
5347 * For those controls, we take the union because they have
5348 * an all or nothing behavior.
5350 ovfl_ctrl
.bits
.notify_user
|= ovfl_arg
->ovfl_ctrl
.bits
.notify_user
;
5351 ovfl_ctrl
.bits
.block_task
|= ovfl_arg
->ovfl_ctrl
.bits
.block_task
;
5352 ovfl_ctrl
.bits
.mask_monitoring
|= ovfl_arg
->ovfl_ctrl
.bits
.mask_monitoring
;
5354 * build the bitmask of pmds to reset now
5356 if (ovfl_arg
->ovfl_ctrl
.bits
.reset_ovfl_pmds
) reset_pmds
|= mask
;
5358 pfm_stats
[this_cpu
].pfm_smpl_handler_cycles
+= end_cycles
- start_cycles
;
5361 * when the module cannot handle the rest of the overflows, we abort right here
5363 if (ret
&& pmd_mask
) {
5364 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5365 pmd_mask
<<PMU_FIRST_COUNTER
));
5368 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5370 ovfl_pmds
&= ~reset_pmds
;
5373 * when no sampling module is used, then the default
5374 * is to notify on overflow if requested by user
5376 ovfl_ctrl
.bits
.notify_user
= ovfl_notify
? 1 : 0;
5377 ovfl_ctrl
.bits
.block_task
= ovfl_notify
? 1 : 0;
5378 ovfl_ctrl
.bits
.mask_monitoring
= ovfl_notify
? 1 : 0; /* XXX: change for saturation */
5379 ovfl_ctrl
.bits
.reset_ovfl_pmds
= ovfl_notify
? 0 : 1;
5381 * if needed, we reset all overflowed pmds
5383 if (ovfl_notify
== 0) reset_pmds
= ovfl_pmds
;
5386 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds
, reset_pmds
));
5389 * reset the requested PMD registers using the short reset values
5392 unsigned long bm
= reset_pmds
;
5393 pfm_reset_regs(ctx
, &bm
, PFM_PMD_SHORT_RESET
);
5396 if (ovfl_notify
&& ovfl_ctrl
.bits
.notify_user
) {
5398 * keep track of what to reset when unblocking
5400 ctx
->ctx_ovfl_regs
[0] = ovfl_pmds
;
5403 * check for blocking context
5405 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && ovfl_ctrl
.bits
.block_task
) {
5407 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_BLOCK
;
5410 * set the perfmon specific checking pending work for the task
5412 PFM_SET_WORK_PENDING(task
, 1);
5415 * when coming from ctxsw, current still points to the
5416 * previous task, therefore we must work with task and not current.
5418 set_notify_resume(task
);
5421 * defer until state is changed (shorten spin window). the context is locked
5422 * anyway, so the signal receiver would come spin for nothing.
5427 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5428 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5429 PFM_GET_WORK_PENDING(task
),
5430 ctx
->ctx_fl_trap_reason
,
5433 ovfl_ctrl
.bits
.mask_monitoring
? 1 : 0));
5435 * in case monitoring must be stopped, we toggle the psr bits
5437 if (ovfl_ctrl
.bits
.mask_monitoring
) {
5438 pfm_mask_monitoring(task
);
5439 ctx
->ctx_state
= PFM_CTX_MASKED
;
5440 ctx
->ctx_fl_can_restart
= 1;
5444 * send notification now
5446 if (must_notify
) pfm_ovfl_notify_user(ctx
, ovfl_notify
);
5451 printk(KERN_ERR
"perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5453 task
? task_pid_nr(task
) : -1,
5459 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5460 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5461 * come here as zombie only if the task is the current task. In which case, we
5462 * can access the PMU hardware directly.
5464 * Note that zombies do have PM_VALID set. So here we do the minimal.
5466 * In case the context was zombified it could not be reclaimed at the time
5467 * the monitoring program exited. At this point, the PMU reservation has been
5468 * returned, the sampiing buffer has been freed. We must convert this call
5469 * into a spurious interrupt. However, we must also avoid infinite overflows
5470 * by stopping monitoring for this task. We can only come here for a per-task
5471 * context. All we need to do is to stop monitoring using the psr bits which
5472 * are always task private. By re-enabling secure montioring, we ensure that
5473 * the monitored task will not be able to re-activate monitoring.
5474 * The task will eventually be context switched out, at which point the context
5475 * will be reclaimed (that includes releasing ownership of the PMU).
5477 * So there might be a window of time where the number of per-task session is zero
5478 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5479 * context. This is safe because if a per-task session comes in, it will push this one
5480 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5481 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5482 * also push our zombie context out.
5484 * Overall pretty hairy stuff....
5486 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task
? task_pid_nr(task
): -1));
5488 ia64_psr(regs
)->up
= 0;
5489 ia64_psr(regs
)->sp
= 1;
5494 pfm_do_interrupt_handler(void *arg
, struct pt_regs
*regs
)
5496 struct task_struct
*task
;
5498 unsigned long flags
;
5500 int this_cpu
= smp_processor_id();
5503 pfm_stats
[this_cpu
].pfm_ovfl_intr_count
++;
5506 * srlz.d done before arriving here
5508 pmc0
= ia64_get_pmc(0);
5510 task
= GET_PMU_OWNER();
5511 ctx
= GET_PMU_CTX();
5514 * if we have some pending bits set
5515 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5517 if (PMC0_HAS_OVFL(pmc0
) && task
) {
5519 * we assume that pmc0.fr is always set here
5523 if (!ctx
) goto report_spurious1
;
5525 if (ctx
->ctx_fl_system
== 0 && (task
->thread
.flags
& IA64_THREAD_PM_VALID
) == 0)
5526 goto report_spurious2
;
5528 PROTECT_CTX_NOPRINT(ctx
, flags
);
5530 pfm_overflow_handler(task
, ctx
, pmc0
, regs
);
5532 UNPROTECT_CTX_NOPRINT(ctx
, flags
);
5535 pfm_stats
[this_cpu
].pfm_spurious_ovfl_intr_count
++;
5539 * keep it unfrozen at all times
5546 printk(KERN_INFO
"perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5547 this_cpu
, task_pid_nr(task
));
5551 printk(KERN_INFO
"perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5559 pfm_interrupt_handler(int irq
, void *arg
)
5561 unsigned long start_cycles
, total_cycles
;
5562 unsigned long min
, max
;
5565 struct pt_regs
*regs
= get_irq_regs();
5567 this_cpu
= get_cpu();
5568 if (likely(!pfm_alt_intr_handler
)) {
5569 min
= pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_min
;
5570 max
= pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_max
;
5572 start_cycles
= ia64_get_itc();
5574 ret
= pfm_do_interrupt_handler(arg
, regs
);
5576 total_cycles
= ia64_get_itc();
5579 * don't measure spurious interrupts
5581 if (likely(ret
== 0)) {
5582 total_cycles
-= start_cycles
;
5584 if (total_cycles
< min
) pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_min
= total_cycles
;
5585 if (total_cycles
> max
) pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_max
= total_cycles
;
5587 pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles
+= total_cycles
;
5591 (*pfm_alt_intr_handler
->handler
)(irq
, arg
, regs
);
5599 * /proc/perfmon interface, for debug only
5602 #define PFM_PROC_SHOW_HEADER ((void *)(long)nr_cpu_ids+1)
5605 pfm_proc_start(struct seq_file
*m
, loff_t
*pos
)
5608 return PFM_PROC_SHOW_HEADER
;
5611 while (*pos
<= nr_cpu_ids
) {
5612 if (cpu_online(*pos
- 1)) {
5613 return (void *)*pos
;
5621 pfm_proc_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
5624 return pfm_proc_start(m
, pos
);
5628 pfm_proc_stop(struct seq_file
*m
, void *v
)
5633 pfm_proc_show_header(struct seq_file
*m
)
5635 struct list_head
* pos
;
5636 pfm_buffer_fmt_t
* entry
;
5637 unsigned long flags
;
5640 "perfmon version : %u.%u\n"
5643 "expert mode : %s\n"
5644 "ovfl_mask : 0x%lx\n"
5645 "PMU flags : 0x%x\n",
5646 PFM_VERSION_MAJ
, PFM_VERSION_MIN
,
5648 pfm_sysctl
.fastctxsw
> 0 ? "Yes": "No",
5649 pfm_sysctl
.expert_mode
> 0 ? "Yes": "No",
5656 "proc_sessions : %u\n"
5657 "sys_sessions : %u\n"
5658 "sys_use_dbregs : %u\n"
5659 "ptrace_use_dbregs : %u\n",
5660 pfm_sessions
.pfs_task_sessions
,
5661 pfm_sessions
.pfs_sys_sessions
,
5662 pfm_sessions
.pfs_sys_use_dbregs
,
5663 pfm_sessions
.pfs_ptrace_use_dbregs
);
5667 spin_lock(&pfm_buffer_fmt_lock
);
5669 list_for_each(pos
, &pfm_buffer_fmt_list
) {
5670 entry
= list_entry(pos
, pfm_buffer_fmt_t
, fmt_list
);
5671 seq_printf(m
, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5682 entry
->fmt_uuid
[10],
5683 entry
->fmt_uuid
[11],
5684 entry
->fmt_uuid
[12],
5685 entry
->fmt_uuid
[13],
5686 entry
->fmt_uuid
[14],
5687 entry
->fmt_uuid
[15],
5690 spin_unlock(&pfm_buffer_fmt_lock
);
5695 pfm_proc_show(struct seq_file
*m
, void *v
)
5701 if (v
== PFM_PROC_SHOW_HEADER
) {
5702 pfm_proc_show_header(m
);
5706 /* show info for CPU (v - 1) */
5710 "CPU%-2d overflow intrs : %lu\n"
5711 "CPU%-2d overflow cycles : %lu\n"
5712 "CPU%-2d overflow min : %lu\n"
5713 "CPU%-2d overflow max : %lu\n"
5714 "CPU%-2d smpl handler calls : %lu\n"
5715 "CPU%-2d smpl handler cycles : %lu\n"
5716 "CPU%-2d spurious intrs : %lu\n"
5717 "CPU%-2d replay intrs : %lu\n"
5718 "CPU%-2d syst_wide : %d\n"
5719 "CPU%-2d dcr_pp : %d\n"
5720 "CPU%-2d exclude idle : %d\n"
5721 "CPU%-2d owner : %d\n"
5722 "CPU%-2d context : %p\n"
5723 "CPU%-2d activations : %lu\n",
5724 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_count
,
5725 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles
,
5726 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles_min
,
5727 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles_max
,
5728 cpu
, pfm_stats
[cpu
].pfm_smpl_handler_calls
,
5729 cpu
, pfm_stats
[cpu
].pfm_smpl_handler_cycles
,
5730 cpu
, pfm_stats
[cpu
].pfm_spurious_ovfl_intr_count
,
5731 cpu
, pfm_stats
[cpu
].pfm_replay_ovfl_intr_count
,
5732 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_SYST_WIDE
? 1 : 0,
5733 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_DCR_PP
? 1 : 0,
5734 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_EXCL_IDLE
? 1 : 0,
5735 cpu
, pfm_get_cpu_data(pmu_owner
, cpu
) ? pfm_get_cpu_data(pmu_owner
, cpu
)->pid
: -1,
5736 cpu
, pfm_get_cpu_data(pmu_ctx
, cpu
),
5737 cpu
, pfm_get_cpu_data(pmu_activation_number
, cpu
));
5739 if (num_online_cpus() == 1 && pfm_sysctl
.debug
> 0) {
5741 psr
= pfm_get_psr();
5746 "CPU%-2d psr : 0x%lx\n"
5747 "CPU%-2d pmc0 : 0x%lx\n",
5749 cpu
, ia64_get_pmc(0));
5751 for (i
=0; PMC_IS_LAST(i
) == 0; i
++) {
5752 if (PMC_IS_COUNTING(i
) == 0) continue;
5754 "CPU%-2d pmc%u : 0x%lx\n"
5755 "CPU%-2d pmd%u : 0x%lx\n",
5756 cpu
, i
, ia64_get_pmc(i
),
5757 cpu
, i
, ia64_get_pmd(i
));
5763 const struct seq_operations pfm_seq_ops
= {
5764 .start
= pfm_proc_start
,
5765 .next
= pfm_proc_next
,
5766 .stop
= pfm_proc_stop
,
5767 .show
= pfm_proc_show
5771 pfm_proc_open(struct inode
*inode
, struct file
*file
)
5773 return seq_open(file
, &pfm_seq_ops
);
5778 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5779 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5780 * is active or inactive based on mode. We must rely on the value in
5781 * local_cpu_data->pfm_syst_info
5784 pfm_syst_wide_update_task(struct task_struct
*task
, unsigned long info
, int is_ctxswin
)
5786 struct pt_regs
*regs
;
5788 unsigned long dcr_pp
;
5790 dcr_pp
= info
& PFM_CPUINFO_DCR_PP
? 1 : 0;
5793 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5794 * on every CPU, so we can rely on the pid to identify the idle task.
5796 if ((info
& PFM_CPUINFO_EXCL_IDLE
) == 0 || task
->pid
) {
5797 regs
= task_pt_regs(task
);
5798 ia64_psr(regs
)->pp
= is_ctxswin
? dcr_pp
: 0;
5802 * if monitoring has started
5805 dcr
= ia64_getreg(_IA64_REG_CR_DCR
);
5807 * context switching in?
5810 /* mask monitoring for the idle task */
5811 ia64_setreg(_IA64_REG_CR_DCR
, dcr
& ~IA64_DCR_PP
);
5817 * context switching out
5818 * restore monitoring for next task
5820 * Due to inlining this odd if-then-else construction generates
5823 ia64_setreg(_IA64_REG_CR_DCR
, dcr
|IA64_DCR_PP
);
5832 pfm_force_cleanup(pfm_context_t
*ctx
, struct pt_regs
*regs
)
5834 struct task_struct
*task
= ctx
->ctx_task
;
5836 ia64_psr(regs
)->up
= 0;
5837 ia64_psr(regs
)->sp
= 1;
5839 if (GET_PMU_OWNER() == task
) {
5840 DPRINT(("cleared ownership for [%d]\n",
5841 task_pid_nr(ctx
->ctx_task
)));
5842 SET_PMU_OWNER(NULL
, NULL
);
5846 * disconnect the task from the context and vice-versa
5848 PFM_SET_WORK_PENDING(task
, 0);
5850 task
->thread
.pfm_context
= NULL
;
5851 task
->thread
.flags
&= ~IA64_THREAD_PM_VALID
;
5853 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task
)));
5858 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5861 pfm_save_regs(struct task_struct
*task
)
5864 unsigned long flags
;
5868 ctx
= PFM_GET_CTX(task
);
5869 if (ctx
== NULL
) return;
5872 * we always come here with interrupts ALREADY disabled by
5873 * the scheduler. So we simply need to protect against concurrent
5874 * access, not CPU concurrency.
5876 flags
= pfm_protect_ctx_ctxsw(ctx
);
5878 if (ctx
->ctx_state
== PFM_CTX_ZOMBIE
) {
5879 struct pt_regs
*regs
= task_pt_regs(task
);
5883 pfm_force_cleanup(ctx
, regs
);
5885 BUG_ON(ctx
->ctx_smpl_hdr
);
5887 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
5889 pfm_context_free(ctx
);
5894 * save current PSR: needed because we modify it
5897 psr
= pfm_get_psr();
5899 BUG_ON(psr
& (IA64_PSR_I
));
5903 * This is the last instruction which may generate an overflow
5905 * We do not need to set psr.sp because, it is irrelevant in kernel.
5906 * It will be restored from ipsr when going back to user level
5911 * keep a copy of psr.up (for reload)
5913 ctx
->ctx_saved_psr_up
= psr
& IA64_PSR_UP
;
5916 * release ownership of this PMU.
5917 * PM interrupts are masked, so nothing
5920 SET_PMU_OWNER(NULL
, NULL
);
5923 * we systematically save the PMD as we have no
5924 * guarantee we will be schedule at that same
5927 pfm_save_pmds(ctx
->th_pmds
, ctx
->ctx_used_pmds
[0]);
5930 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5931 * we will need it on the restore path to check
5932 * for pending overflow.
5934 ctx
->th_pmcs
[0] = ia64_get_pmc(0);
5937 * unfreeze PMU if had pending overflows
5939 if (ctx
->th_pmcs
[0] & ~0x1UL
) pfm_unfreeze_pmu();
5942 * finally, allow context access.
5943 * interrupts will still be masked after this call.
5945 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
5948 #else /* !CONFIG_SMP */
5950 pfm_save_regs(struct task_struct
*task
)
5955 ctx
= PFM_GET_CTX(task
);
5956 if (ctx
== NULL
) return;
5959 * save current PSR: needed because we modify it
5961 psr
= pfm_get_psr();
5963 BUG_ON(psr
& (IA64_PSR_I
));
5967 * This is the last instruction which may generate an overflow
5969 * We do not need to set psr.sp because, it is irrelevant in kernel.
5970 * It will be restored from ipsr when going back to user level
5975 * keep a copy of psr.up (for reload)
5977 ctx
->ctx_saved_psr_up
= psr
& IA64_PSR_UP
;
5981 pfm_lazy_save_regs (struct task_struct
*task
)
5984 unsigned long flags
;
5986 { u64 psr
= pfm_get_psr();
5987 BUG_ON(psr
& IA64_PSR_UP
);
5990 ctx
= PFM_GET_CTX(task
);
5993 * we need to mask PMU overflow here to
5994 * make sure that we maintain pmc0 until
5995 * we save it. overflow interrupts are
5996 * treated as spurious if there is no
5999 * XXX: I don't think this is necessary
6001 PROTECT_CTX(ctx
,flags
);
6004 * release ownership of this PMU.
6005 * must be done before we save the registers.
6007 * after this call any PMU interrupt is treated
6010 SET_PMU_OWNER(NULL
, NULL
);
6013 * save all the pmds we use
6015 pfm_save_pmds(ctx
->th_pmds
, ctx
->ctx_used_pmds
[0]);
6018 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6019 * it is needed to check for pended overflow
6020 * on the restore path
6022 ctx
->th_pmcs
[0] = ia64_get_pmc(0);
6025 * unfreeze PMU if had pending overflows
6027 if (ctx
->th_pmcs
[0] & ~0x1UL
) pfm_unfreeze_pmu();
6030 * now get can unmask PMU interrupts, they will
6031 * be treated as purely spurious and we will not
6032 * lose any information
6034 UNPROTECT_CTX(ctx
,flags
);
6036 #endif /* CONFIG_SMP */
6040 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6043 pfm_load_regs (struct task_struct
*task
)
6046 unsigned long pmc_mask
= 0UL, pmd_mask
= 0UL;
6047 unsigned long flags
;
6049 int need_irq_resend
;
6051 ctx
= PFM_GET_CTX(task
);
6052 if (unlikely(ctx
== NULL
)) return;
6054 BUG_ON(GET_PMU_OWNER());
6057 * possible on unload
6059 if (unlikely((task
->thread
.flags
& IA64_THREAD_PM_VALID
) == 0)) return;
6062 * we always come here with interrupts ALREADY disabled by
6063 * the scheduler. So we simply need to protect against concurrent
6064 * access, not CPU concurrency.
6066 flags
= pfm_protect_ctx_ctxsw(ctx
);
6067 psr
= pfm_get_psr();
6069 need_irq_resend
= pmu_conf
->flags
& PFM_PMU_IRQ_RESEND
;
6071 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
6072 BUG_ON(psr
& IA64_PSR_I
);
6074 if (unlikely(ctx
->ctx_state
== PFM_CTX_ZOMBIE
)) {
6075 struct pt_regs
*regs
= task_pt_regs(task
);
6077 BUG_ON(ctx
->ctx_smpl_hdr
);
6079 pfm_force_cleanup(ctx
, regs
);
6081 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
6084 * this one (kmalloc'ed) is fine with interrupts disabled
6086 pfm_context_free(ctx
);
6092 * we restore ALL the debug registers to avoid picking up
6095 if (ctx
->ctx_fl_using_dbreg
) {
6096 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
6097 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
6100 * retrieve saved psr.up
6102 psr_up
= ctx
->ctx_saved_psr_up
;
6105 * if we were the last user of the PMU on that CPU,
6106 * then nothing to do except restore psr
6108 if (GET_LAST_CPU(ctx
) == smp_processor_id() && ctx
->ctx_last_activation
== GET_ACTIVATION()) {
6111 * retrieve partial reload masks (due to user modifications)
6113 pmc_mask
= ctx
->ctx_reload_pmcs
[0];
6114 pmd_mask
= ctx
->ctx_reload_pmds
[0];
6118 * To avoid leaking information to the user level when psr.sp=0,
6119 * we must reload ALL implemented pmds (even the ones we don't use).
6120 * In the kernel we only allow PFM_READ_PMDS on registers which
6121 * we initialized or requested (sampling) so there is no risk there.
6123 pmd_mask
= pfm_sysctl
.fastctxsw
? ctx
->ctx_used_pmds
[0] : ctx
->ctx_all_pmds
[0];
6126 * ALL accessible PMCs are systematically reloaded, unused registers
6127 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6128 * up stale configuration.
6130 * PMC0 is never in the mask. It is always restored separately.
6132 pmc_mask
= ctx
->ctx_all_pmcs
[0];
6135 * when context is MASKED, we will restore PMC with plm=0
6136 * and PMD with stale information, but that's ok, nothing
6139 * XXX: optimize here
6141 if (pmd_mask
) pfm_restore_pmds(ctx
->th_pmds
, pmd_mask
);
6142 if (pmc_mask
) pfm_restore_pmcs(ctx
->th_pmcs
, pmc_mask
);
6145 * check for pending overflow at the time the state
6148 if (unlikely(PMC0_HAS_OVFL(ctx
->th_pmcs
[0]))) {
6150 * reload pmc0 with the overflow information
6151 * On McKinley PMU, this will trigger a PMU interrupt
6153 ia64_set_pmc(0, ctx
->th_pmcs
[0]);
6155 ctx
->th_pmcs
[0] = 0UL;
6158 * will replay the PMU interrupt
6160 if (need_irq_resend
) ia64_resend_irq(IA64_PERFMON_VECTOR
);
6162 pfm_stats
[smp_processor_id()].pfm_replay_ovfl_intr_count
++;
6166 * we just did a reload, so we reset the partial reload fields
6168 ctx
->ctx_reload_pmcs
[0] = 0UL;
6169 ctx
->ctx_reload_pmds
[0] = 0UL;
6171 SET_LAST_CPU(ctx
, smp_processor_id());
6174 * dump activation value for this PMU
6178 * record current activation for this context
6180 SET_ACTIVATION(ctx
);
6183 * establish new ownership.
6185 SET_PMU_OWNER(task
, ctx
);
6188 * restore the psr.up bit. measurement
6190 * no PMU interrupt can happen at this point
6191 * because we still have interrupts disabled.
6193 if (likely(psr_up
)) pfm_set_psr_up();
6196 * allow concurrent access to context
6198 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
6200 #else /* !CONFIG_SMP */
6202 * reload PMU state for UP kernels
6203 * in 2.5 we come here with interrupts disabled
6206 pfm_load_regs (struct task_struct
*task
)
6209 struct task_struct
*owner
;
6210 unsigned long pmd_mask
, pmc_mask
;
6212 int need_irq_resend
;
6214 owner
= GET_PMU_OWNER();
6215 ctx
= PFM_GET_CTX(task
);
6216 psr
= pfm_get_psr();
6218 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
6219 BUG_ON(psr
& IA64_PSR_I
);
6222 * we restore ALL the debug registers to avoid picking up
6225 * This must be done even when the task is still the owner
6226 * as the registers may have been modified via ptrace()
6227 * (not perfmon) by the previous task.
6229 if (ctx
->ctx_fl_using_dbreg
) {
6230 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
6231 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
6235 * retrieved saved psr.up
6237 psr_up
= ctx
->ctx_saved_psr_up
;
6238 need_irq_resend
= pmu_conf
->flags
& PFM_PMU_IRQ_RESEND
;
6241 * short path, our state is still there, just
6242 * need to restore psr and we go
6244 * we do not touch either PMC nor PMD. the psr is not touched
6245 * by the overflow_handler. So we are safe w.r.t. to interrupt
6246 * concurrency even without interrupt masking.
6248 if (likely(owner
== task
)) {
6249 if (likely(psr_up
)) pfm_set_psr_up();
6254 * someone else is still using the PMU, first push it out and
6255 * then we'll be able to install our stuff !
6257 * Upon return, there will be no owner for the current PMU
6259 if (owner
) pfm_lazy_save_regs(owner
);
6262 * To avoid leaking information to the user level when psr.sp=0,
6263 * we must reload ALL implemented pmds (even the ones we don't use).
6264 * In the kernel we only allow PFM_READ_PMDS on registers which
6265 * we initialized or requested (sampling) so there is no risk there.
6267 pmd_mask
= pfm_sysctl
.fastctxsw
? ctx
->ctx_used_pmds
[0] : ctx
->ctx_all_pmds
[0];
6270 * ALL accessible PMCs are systematically reloaded, unused registers
6271 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6272 * up stale configuration.
6274 * PMC0 is never in the mask. It is always restored separately
6276 pmc_mask
= ctx
->ctx_all_pmcs
[0];
6278 pfm_restore_pmds(ctx
->th_pmds
, pmd_mask
);
6279 pfm_restore_pmcs(ctx
->th_pmcs
, pmc_mask
);
6282 * check for pending overflow at the time the state
6285 if (unlikely(PMC0_HAS_OVFL(ctx
->th_pmcs
[0]))) {
6287 * reload pmc0 with the overflow information
6288 * On McKinley PMU, this will trigger a PMU interrupt
6290 ia64_set_pmc(0, ctx
->th_pmcs
[0]);
6293 ctx
->th_pmcs
[0] = 0UL;
6296 * will replay the PMU interrupt
6298 if (need_irq_resend
) ia64_resend_irq(IA64_PERFMON_VECTOR
);
6300 pfm_stats
[smp_processor_id()].pfm_replay_ovfl_intr_count
++;
6304 * establish new ownership.
6306 SET_PMU_OWNER(task
, ctx
);
6309 * restore the psr.up bit. measurement
6311 * no PMU interrupt can happen at this point
6312 * because we still have interrupts disabled.
6314 if (likely(psr_up
)) pfm_set_psr_up();
6316 #endif /* CONFIG_SMP */
6319 * this function assumes monitoring is stopped
6322 pfm_flush_pmds(struct task_struct
*task
, pfm_context_t
*ctx
)
6325 unsigned long mask2
, val
, pmd_val
, ovfl_val
;
6326 int i
, can_access_pmu
= 0;
6330 * is the caller the task being monitored (or which initiated the
6331 * session for system wide measurements)
6333 is_self
= ctx
->ctx_task
== task
? 1 : 0;
6336 * can access PMU is task is the owner of the PMU state on the current CPU
6337 * or if we are running on the CPU bound to the context in system-wide mode
6338 * (that is not necessarily the task the context is attached to in this mode).
6339 * In system-wide we always have can_access_pmu true because a task running on an
6340 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6342 can_access_pmu
= (GET_PMU_OWNER() == task
) || (ctx
->ctx_fl_system
&& ctx
->ctx_cpu
== smp_processor_id());
6343 if (can_access_pmu
) {
6345 * Mark the PMU as not owned
6346 * This will cause the interrupt handler to do nothing in case an overflow
6347 * interrupt was in-flight
6348 * This also guarantees that pmc0 will contain the final state
6349 * It virtually gives us full control on overflow processing from that point
6352 SET_PMU_OWNER(NULL
, NULL
);
6353 DPRINT(("releasing ownership\n"));
6356 * read current overflow status:
6358 * we are guaranteed to read the final stable state
6361 pmc0
= ia64_get_pmc(0); /* slow */
6364 * reset freeze bit, overflow status information destroyed
6368 pmc0
= ctx
->th_pmcs
[0];
6370 * clear whatever overflow status bits there were
6372 ctx
->th_pmcs
[0] = 0;
6374 ovfl_val
= pmu_conf
->ovfl_val
;
6376 * we save all the used pmds
6377 * we take care of overflows for counting PMDs
6379 * XXX: sampling situation is not taken into account here
6381 mask2
= ctx
->ctx_used_pmds
[0];
6383 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self
, ovfl_val
, mask2
));
6385 for (i
= 0; mask2
; i
++, mask2
>>=1) {
6387 /* skip non used pmds */
6388 if ((mask2
& 0x1) == 0) continue;
6391 * can access PMU always true in system wide mode
6393 val
= pmd_val
= can_access_pmu
? ia64_get_pmd(i
) : ctx
->th_pmds
[i
];
6395 if (PMD_IS_COUNTING(i
)) {
6396 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6399 ctx
->ctx_pmds
[i
].val
,
6403 * we rebuild the full 64 bit value of the counter
6405 val
= ctx
->ctx_pmds
[i
].val
+ (val
& ovfl_val
);
6408 * now everything is in ctx_pmds[] and we need
6409 * to clear the saved context from save_regs() such that
6410 * pfm_read_pmds() gets the correct value
6415 * take care of overflow inline
6417 if (pmc0
& (1UL << i
)) {
6418 val
+= 1 + ovfl_val
;
6419 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task
), i
));
6423 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task
), i
, val
, pmd_val
));
6425 if (is_self
) ctx
->th_pmds
[i
] = pmd_val
;
6427 ctx
->ctx_pmds
[i
].val
= val
;
6431 static struct irqaction perfmon_irqaction
= {
6432 .handler
= pfm_interrupt_handler
,
6433 .flags
= IRQF_DISABLED
,
6438 pfm_alt_save_pmu_state(void *data
)
6440 struct pt_regs
*regs
;
6442 regs
= task_pt_regs(current
);
6444 DPRINT(("called\n"));
6447 * should not be necessary but
6448 * let's take not risk
6452 ia64_psr(regs
)->pp
= 0;
6455 * This call is required
6456 * May cause a spurious interrupt on some processors
6464 pfm_alt_restore_pmu_state(void *data
)
6466 struct pt_regs
*regs
;
6468 regs
= task_pt_regs(current
);
6470 DPRINT(("called\n"));
6473 * put PMU back in state expected
6478 ia64_psr(regs
)->pp
= 0;
6481 * perfmon runs with PMU unfrozen at all times
6489 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t
*hdl
)
6494 /* some sanity checks */
6495 if (hdl
== NULL
|| hdl
->handler
== NULL
) return -EINVAL
;
6497 /* do the easy test first */
6498 if (pfm_alt_intr_handler
) return -EBUSY
;
6500 /* one at a time in the install or remove, just fail the others */
6501 if (!spin_trylock(&pfm_alt_install_check
)) {
6505 /* reserve our session */
6506 for_each_online_cpu(reserve_cpu
) {
6507 ret
= pfm_reserve_session(NULL
, 1, reserve_cpu
);
6508 if (ret
) goto cleanup_reserve
;
6511 /* save the current system wide pmu states */
6512 ret
= on_each_cpu(pfm_alt_save_pmu_state
, NULL
, 1);
6514 DPRINT(("on_each_cpu() failed: %d\n", ret
));
6515 goto cleanup_reserve
;
6518 /* officially change to the alternate interrupt handler */
6519 pfm_alt_intr_handler
= hdl
;
6521 spin_unlock(&pfm_alt_install_check
);
6526 for_each_online_cpu(i
) {
6527 /* don't unreserve more than we reserved */
6528 if (i
>= reserve_cpu
) break;
6530 pfm_unreserve_session(NULL
, 1, i
);
6533 spin_unlock(&pfm_alt_install_check
);
6537 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt
);
6540 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t
*hdl
)
6545 if (hdl
== NULL
) return -EINVAL
;
6547 /* cannot remove someone else's handler! */
6548 if (pfm_alt_intr_handler
!= hdl
) return -EINVAL
;
6550 /* one at a time in the install or remove, just fail the others */
6551 if (!spin_trylock(&pfm_alt_install_check
)) {
6555 pfm_alt_intr_handler
= NULL
;
6557 ret
= on_each_cpu(pfm_alt_restore_pmu_state
, NULL
, 1);
6559 DPRINT(("on_each_cpu() failed: %d\n", ret
));
6562 for_each_online_cpu(i
) {
6563 pfm_unreserve_session(NULL
, 1, i
);
6566 spin_unlock(&pfm_alt_install_check
);
6570 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt
);
6573 * perfmon initialization routine, called from the initcall() table
6575 static int init_pfm_fs(void);
6583 family
= local_cpu_data
->family
;
6588 if ((*p
)->probe() == 0) goto found
;
6589 } else if ((*p
)->pmu_family
== family
|| (*p
)->pmu_family
== 0xff) {
6600 static const struct file_operations pfm_proc_fops
= {
6601 .open
= pfm_proc_open
,
6603 .llseek
= seq_lseek
,
6604 .release
= seq_release
,
6610 unsigned int n
, n_counters
, i
;
6612 printk("perfmon: version %u.%u IRQ %u\n",
6615 IA64_PERFMON_VECTOR
);
6617 if (pfm_probe_pmu()) {
6618 printk(KERN_INFO
"perfmon: disabled, there is no support for processor family %d\n",
6619 local_cpu_data
->family
);
6624 * compute the number of implemented PMD/PMC from the
6625 * description tables
6628 for (i
=0; PMC_IS_LAST(i
) == 0; i
++) {
6629 if (PMC_IS_IMPL(i
) == 0) continue;
6630 pmu_conf
->impl_pmcs
[i
>>6] |= 1UL << (i
&63);
6633 pmu_conf
->num_pmcs
= n
;
6635 n
= 0; n_counters
= 0;
6636 for (i
=0; PMD_IS_LAST(i
) == 0; i
++) {
6637 if (PMD_IS_IMPL(i
) == 0) continue;
6638 pmu_conf
->impl_pmds
[i
>>6] |= 1UL << (i
&63);
6640 if (PMD_IS_COUNTING(i
)) n_counters
++;
6642 pmu_conf
->num_pmds
= n
;
6643 pmu_conf
->num_counters
= n_counters
;
6646 * sanity checks on the number of debug registers
6648 if (pmu_conf
->use_rr_dbregs
) {
6649 if (pmu_conf
->num_ibrs
> IA64_NUM_DBG_REGS
) {
6650 printk(KERN_INFO
"perfmon: unsupported number of code debug registers (%u)\n", pmu_conf
->num_ibrs
);
6654 if (pmu_conf
->num_dbrs
> IA64_NUM_DBG_REGS
) {
6655 printk(KERN_INFO
"perfmon: unsupported number of data debug registers (%u)\n", pmu_conf
->num_ibrs
);
6661 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6665 pmu_conf
->num_counters
,
6666 ffz(pmu_conf
->ovfl_val
));
6669 if (pmu_conf
->num_pmds
>= PFM_NUM_PMD_REGS
|| pmu_conf
->num_pmcs
>= PFM_NUM_PMC_REGS
) {
6670 printk(KERN_ERR
"perfmon: not enough pmc/pmd, perfmon disabled\n");
6676 * create /proc/perfmon (mostly for debugging purposes)
6678 perfmon_dir
= proc_create("perfmon", S_IRUGO
, NULL
, &pfm_proc_fops
);
6679 if (perfmon_dir
== NULL
) {
6680 printk(KERN_ERR
"perfmon: cannot create /proc entry, perfmon disabled\n");
6686 * create /proc/sys/kernel/perfmon (for debugging purposes)
6688 pfm_sysctl_header
= register_sysctl_table(pfm_sysctl_root
);
6691 * initialize all our spinlocks
6693 spin_lock_init(&pfm_sessions
.pfs_lock
);
6694 spin_lock_init(&pfm_buffer_fmt_lock
);
6698 for(i
=0; i
< NR_CPUS
; i
++) pfm_stats
[i
].pfm_ovfl_intr_cycles_min
= ~0UL;
6703 __initcall(pfm_init
);
6706 * this function is called before pfm_init()
6709 pfm_init_percpu (void)
6711 static int first_time
=1;
6713 * make sure no measurement is active
6714 * (may inherit programmed PMCs from EFI).
6720 * we run with the PMU not frozen at all times
6725 register_percpu_irq(IA64_PERFMON_VECTOR
, &perfmon_irqaction
);
6729 ia64_setreg(_IA64_REG_CR_PMV
, IA64_PERFMON_VECTOR
);
6734 * used for debug purposes only
6737 dump_pmu_state(const char *from
)
6739 struct task_struct
*task
;
6740 struct pt_regs
*regs
;
6742 unsigned long psr
, dcr
, info
, flags
;
6745 local_irq_save(flags
);
6747 this_cpu
= smp_processor_id();
6748 regs
= task_pt_regs(current
);
6749 info
= PFM_CPUINFO_GET();
6750 dcr
= ia64_getreg(_IA64_REG_CR_DCR
);
6752 if (info
== 0 && ia64_psr(regs
)->pp
== 0 && (dcr
& IA64_DCR_PP
) == 0) {
6753 local_irq_restore(flags
);
6757 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6760 task_pid_nr(current
),
6764 task
= GET_PMU_OWNER();
6765 ctx
= GET_PMU_CTX();
6767 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu
, task
? task_pid_nr(task
) : -1, ctx
);
6769 psr
= pfm_get_psr();
6771 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",
6774 psr
& IA64_PSR_PP
? 1 : 0,
6775 psr
& IA64_PSR_UP
? 1 : 0,
6776 dcr
& IA64_DCR_PP
? 1 : 0,
6779 ia64_psr(regs
)->pp
);
6781 ia64_psr(regs
)->up
= 0;
6782 ia64_psr(regs
)->pp
= 0;
6784 for (i
=1; PMC_IS_LAST(i
) == 0; i
++) {
6785 if (PMC_IS_IMPL(i
) == 0) continue;
6786 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
]);
6789 for (i
=1; PMD_IS_LAST(i
) == 0; i
++) {
6790 if (PMD_IS_IMPL(i
) == 0) continue;
6791 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
]);
6795 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6798 ctx
->ctx_smpl_vaddr
,
6802 ctx
->ctx_saved_psr_up
);
6804 local_irq_restore(flags
);
6808 * called from process.c:copy_thread(). task is new child.
6811 pfm_inherit(struct task_struct
*task
, struct pt_regs
*regs
)
6813 struct thread_struct
*thread
;
6815 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task
)));
6817 thread
= &task
->thread
;
6820 * cut links inherited from parent (current)
6822 thread
->pfm_context
= NULL
;
6824 PFM_SET_WORK_PENDING(task
, 0);
6827 * the psr bits are already set properly in copy_threads()
6830 #else /* !CONFIG_PERFMON */
6832 sys_perfmonctl (int fd
, int cmd
, void *arg
, int count
)
6836 #endif /* CONFIG_PERFMON */