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/smp_lock.h>
27 #include <linux/proc_fs.h>
28 #include <linux/seq_file.h>
29 #include <linux/init.h>
30 #include <linux/vmalloc.h>
32 #include <linux/sysctl.h>
33 #include <linux/list.h>
34 #include <linux/file.h>
35 #include <linux/poll.h>
36 #include <linux/vfs.h>
37 #include <linux/smp.h>
38 #include <linux/pagemap.h>
39 #include <linux/mount.h>
40 #include <linux/bitops.h>
41 #include <linux/capability.h>
42 #include <linux/rcupdate.h>
43 #include <linux/completion.h>
45 #include <asm/errno.h>
46 #include <asm/intrinsics.h>
48 #include <asm/perfmon.h>
49 #include <asm/processor.h>
50 #include <asm/signal.h>
51 #include <asm/system.h>
52 #include <asm/uaccess.h>
53 #include <asm/delay.h>
57 * perfmon context state
59 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
60 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
61 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
62 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
64 #define PFM_INVALID_ACTIVATION (~0UL)
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_lock_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, current->pid)); \
163 spin_lock_irqsave(&(c)->ctx_lock, f); \
164 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
167 #define UNPROTECT_CTX(c, f) \
169 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
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] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); 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] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); 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 u64 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
[]={
524 {1, "debug", &pfm_sysctl
.debug
, sizeof(int), 0666, NULL
, &proc_dointvec
, NULL
,},
525 {2, "debug_ovfl", &pfm_sysctl
.debug_ovfl
, sizeof(int), 0666, NULL
, &proc_dointvec
, NULL
,},
526 {3, "fastctxsw", &pfm_sysctl
.fastctxsw
, sizeof(int), 0600, NULL
, &proc_dointvec
, NULL
,},
527 {4, "expert_mode", &pfm_sysctl
.expert_mode
, sizeof(int), 0600, NULL
, &proc_dointvec
, NULL
,},
530 static ctl_table pfm_sysctl_dir
[] = {
531 {1, "perfmon", NULL
, 0, 0755, pfm_ctl_table
, },
534 static ctl_table pfm_sysctl_root
[] = {
535 {1, "kernel", NULL
, 0, 0755, pfm_sysctl_dir
, },
538 static struct ctl_table_header
*pfm_sysctl_header
;
540 static int pfm_context_unload(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
542 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
543 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
546 pfm_put_task(struct task_struct
*task
)
548 if (task
!= current
) put_task_struct(task
);
552 pfm_set_task_notify(struct task_struct
*task
)
554 struct thread_info
*info
;
556 info
= (struct thread_info
*) ((char *) task
+ IA64_TASK_SIZE
);
557 set_bit(TIF_NOTIFY_RESUME
, &info
->flags
);
561 pfm_clear_task_notify(void)
563 clear_thread_flag(TIF_NOTIFY_RESUME
);
567 pfm_reserve_page(unsigned long a
)
569 SetPageReserved(vmalloc_to_page((void *)a
));
572 pfm_unreserve_page(unsigned long a
)
574 ClearPageReserved(vmalloc_to_page((void*)a
));
577 static inline unsigned long
578 pfm_protect_ctx_ctxsw(pfm_context_t
*x
)
580 spin_lock(&(x
)->ctx_lock
);
585 pfm_unprotect_ctx_ctxsw(pfm_context_t
*x
, unsigned long f
)
587 spin_unlock(&(x
)->ctx_lock
);
590 static inline unsigned int
591 pfm_do_munmap(struct mm_struct
*mm
, unsigned long addr
, size_t len
, int acct
)
593 return do_munmap(mm
, addr
, len
);
596 static inline unsigned long
597 pfm_get_unmapped_area(struct file
*file
, unsigned long addr
, unsigned long len
, unsigned long pgoff
, unsigned long flags
, unsigned long exec
)
599 return get_unmapped_area(file
, addr
, len
, pgoff
, flags
);
604 pfmfs_get_sb(struct file_system_type
*fs_type
, int flags
, const char *dev_name
, void *data
,
605 struct vfsmount
*mnt
)
607 return get_sb_pseudo(fs_type
, "pfm:", NULL
, PFMFS_MAGIC
, mnt
);
610 static struct file_system_type pfm_fs_type
= {
612 .get_sb
= pfmfs_get_sb
,
613 .kill_sb
= kill_anon_super
,
616 DEFINE_PER_CPU(unsigned long, pfm_syst_info
);
617 DEFINE_PER_CPU(struct task_struct
*, pmu_owner
);
618 DEFINE_PER_CPU(pfm_context_t
*, pmu_ctx
);
619 DEFINE_PER_CPU(unsigned long, pmu_activation_number
);
620 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info
);
623 /* forward declaration */
624 static struct file_operations pfm_file_ops
;
627 * forward declarations
630 static void pfm_lazy_save_regs (struct task_struct
*ta
);
633 void dump_pmu_state(const char *);
634 static int pfm_write_ibr_dbr(int mode
, pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
636 #include "perfmon_itanium.h"
637 #include "perfmon_mckinley.h"
638 #include "perfmon_montecito.h"
639 #include "perfmon_generic.h"
641 static pmu_config_t
*pmu_confs
[]={
645 &pmu_conf_gen
, /* must be last */
650 static int pfm_end_notify_user(pfm_context_t
*ctx
);
653 pfm_clear_psr_pp(void)
655 ia64_rsm(IA64_PSR_PP
);
662 ia64_ssm(IA64_PSR_PP
);
667 pfm_clear_psr_up(void)
669 ia64_rsm(IA64_PSR_UP
);
676 ia64_ssm(IA64_PSR_UP
);
680 static inline unsigned long
684 tmp
= ia64_getreg(_IA64_REG_PSR
);
690 pfm_set_psr_l(unsigned long val
)
692 ia64_setreg(_IA64_REG_PSR_L
, val
);
704 pfm_unfreeze_pmu(void)
711 pfm_restore_ibrs(unsigned long *ibrs
, unsigned int nibrs
)
715 for (i
=0; i
< nibrs
; i
++) {
716 ia64_set_ibr(i
, ibrs
[i
]);
717 ia64_dv_serialize_instruction();
723 pfm_restore_dbrs(unsigned long *dbrs
, unsigned int ndbrs
)
727 for (i
=0; i
< ndbrs
; i
++) {
728 ia64_set_dbr(i
, dbrs
[i
]);
729 ia64_dv_serialize_data();
735 * PMD[i] must be a counter. no check is made
737 static inline unsigned long
738 pfm_read_soft_counter(pfm_context_t
*ctx
, int i
)
740 return ctx
->ctx_pmds
[i
].val
+ (ia64_get_pmd(i
) & pmu_conf
->ovfl_val
);
744 * PMD[i] must be a counter. no check is made
747 pfm_write_soft_counter(pfm_context_t
*ctx
, int i
, unsigned long val
)
749 unsigned long ovfl_val
= pmu_conf
->ovfl_val
;
751 ctx
->ctx_pmds
[i
].val
= val
& ~ovfl_val
;
753 * writing to unimplemented part is ignore, so we do not need to
756 ia64_set_pmd(i
, val
& ovfl_val
);
760 pfm_get_new_msg(pfm_context_t
*ctx
)
764 next
= (ctx
->ctx_msgq_tail
+1) % PFM_MAX_MSGS
;
766 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
767 if (next
== ctx
->ctx_msgq_head
) return NULL
;
769 idx
= ctx
->ctx_msgq_tail
;
770 ctx
->ctx_msgq_tail
= next
;
772 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
, idx
));
774 return ctx
->ctx_msgq
+idx
;
778 pfm_get_next_msg(pfm_context_t
*ctx
)
782 DPRINT(("ctx=%p head=%d tail=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
784 if (PFM_CTXQ_EMPTY(ctx
)) return NULL
;
789 msg
= ctx
->ctx_msgq
+ctx
->ctx_msgq_head
;
794 ctx
->ctx_msgq_head
= (ctx
->ctx_msgq_head
+1) % PFM_MAX_MSGS
;
796 DPRINT(("ctx=%p head=%d tail=%d type=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
, msg
->pfm_gen_msg
.msg_type
));
802 pfm_reset_msgq(pfm_context_t
*ctx
)
804 ctx
->ctx_msgq_head
= ctx
->ctx_msgq_tail
= 0;
805 DPRINT(("ctx=%p msgq reset\n", ctx
));
809 pfm_rvmalloc(unsigned long size
)
814 size
= PAGE_ALIGN(size
);
817 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
818 memset(mem
, 0, size
);
819 addr
= (unsigned long)mem
;
821 pfm_reserve_page(addr
);
830 pfm_rvfree(void *mem
, unsigned long size
)
835 DPRINT(("freeing physical buffer @%p size=%lu\n", mem
, size
));
836 addr
= (unsigned long) mem
;
837 while ((long) size
> 0) {
838 pfm_unreserve_page(addr
);
847 static pfm_context_t
*
848 pfm_context_alloc(void)
853 * allocate context descriptor
854 * must be able to free with interrupts disabled
856 ctx
= kmalloc(sizeof(pfm_context_t
), GFP_KERNEL
);
858 memset(ctx
, 0, sizeof(pfm_context_t
));
859 DPRINT(("alloc ctx @%p\n", ctx
));
865 pfm_context_free(pfm_context_t
*ctx
)
868 DPRINT(("free ctx @%p\n", ctx
));
874 pfm_mask_monitoring(struct task_struct
*task
)
876 pfm_context_t
*ctx
= PFM_GET_CTX(task
);
877 unsigned long mask
, val
, ovfl_mask
;
880 DPRINT_ovfl(("masking monitoring for [%d]\n", task
->pid
));
882 ovfl_mask
= pmu_conf
->ovfl_val
;
884 * monitoring can only be masked as a result of a valid
885 * counter overflow. In UP, it means that the PMU still
886 * has an owner. Note that the owner can be different
887 * from the current task. However the PMU state belongs
889 * In SMP, a valid overflow only happens when task is
890 * current. Therefore if we come here, we know that
891 * the PMU state belongs to the current task, therefore
892 * we can access the live registers.
894 * So in both cases, the live register contains the owner's
895 * state. We can ONLY touch the PMU registers and NOT the PSR.
897 * As a consequence to this call, the ctx->th_pmds[] array
898 * contains stale information which must be ignored
899 * when context is reloaded AND monitoring is active (see
902 mask
= ctx
->ctx_used_pmds
[0];
903 for (i
= 0; mask
; i
++, mask
>>=1) {
904 /* skip non used pmds */
905 if ((mask
& 0x1) == 0) continue;
906 val
= ia64_get_pmd(i
);
908 if (PMD_IS_COUNTING(i
)) {
910 * we rebuild the full 64 bit value of the counter
912 ctx
->ctx_pmds
[i
].val
+= (val
& ovfl_mask
);
914 ctx
->ctx_pmds
[i
].val
= val
;
916 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
918 ctx
->ctx_pmds
[i
].val
,
922 * mask monitoring by setting the privilege level to 0
923 * we cannot use psr.pp/psr.up for this, it is controlled by
926 * if task is current, modify actual registers, otherwise modify
927 * thread save state, i.e., what will be restored in pfm_load_regs()
929 mask
= ctx
->ctx_used_monitors
[0] >> PMU_FIRST_COUNTER
;
930 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>=1) {
931 if ((mask
& 0x1) == 0UL) continue;
932 ia64_set_pmc(i
, ctx
->th_pmcs
[i
] & ~0xfUL
);
933 ctx
->th_pmcs
[i
] &= ~0xfUL
;
934 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i
, ctx
->th_pmcs
[i
]));
937 * make all of this visible
943 * must always be done with task == current
945 * context must be in MASKED state when calling
948 pfm_restore_monitoring(struct task_struct
*task
)
950 pfm_context_t
*ctx
= PFM_GET_CTX(task
);
951 unsigned long mask
, ovfl_mask
;
952 unsigned long psr
, val
;
955 is_system
= ctx
->ctx_fl_system
;
956 ovfl_mask
= pmu_conf
->ovfl_val
;
958 if (task
!= current
) {
959 printk(KERN_ERR
"perfmon.%d: invalid task[%d] current[%d]\n", __LINE__
, task
->pid
, current
->pid
);
962 if (ctx
->ctx_state
!= PFM_CTX_MASKED
) {
963 printk(KERN_ERR
"perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__
,
964 task
->pid
, current
->pid
, ctx
->ctx_state
);
969 * monitoring is masked via the PMC.
970 * As we restore their value, we do not want each counter to
971 * restart right away. We stop monitoring using the PSR,
972 * restore the PMC (and PMD) and then re-establish the psr
973 * as it was. Note that there can be no pending overflow at
974 * this point, because monitoring was MASKED.
976 * system-wide session are pinned and self-monitoring
978 if (is_system
&& (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP
)) {
980 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) & ~IA64_DCR_PP
);
986 * first, we restore the PMD
988 mask
= ctx
->ctx_used_pmds
[0];
989 for (i
= 0; mask
; i
++, mask
>>=1) {
990 /* skip non used pmds */
991 if ((mask
& 0x1) == 0) continue;
993 if (PMD_IS_COUNTING(i
)) {
995 * we split the 64bit value according to
998 val
= ctx
->ctx_pmds
[i
].val
& ovfl_mask
;
999 ctx
->ctx_pmds
[i
].val
&= ~ovfl_mask
;
1001 val
= ctx
->ctx_pmds
[i
].val
;
1003 ia64_set_pmd(i
, val
);
1005 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1007 ctx
->ctx_pmds
[i
].val
,
1013 mask
= ctx
->ctx_used_monitors
[0] >> PMU_FIRST_COUNTER
;
1014 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>=1) {
1015 if ((mask
& 0x1) == 0UL) continue;
1016 ctx
->th_pmcs
[i
] = ctx
->ctx_pmcs
[i
];
1017 ia64_set_pmc(i
, ctx
->th_pmcs
[i
]);
1018 DPRINT(("[%d] pmc[%d]=0x%lx\n", task
->pid
, i
, ctx
->th_pmcs
[i
]));
1023 * must restore DBR/IBR because could be modified while masked
1024 * XXX: need to optimize
1026 if (ctx
->ctx_fl_using_dbreg
) {
1027 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
1028 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
1034 if (is_system
&& (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP
)) {
1036 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) | IA64_DCR_PP
);
1043 pfm_save_pmds(unsigned long *pmds
, unsigned long mask
)
1049 for (i
=0; mask
; i
++, mask
>>=1) {
1050 if (mask
& 0x1) pmds
[i
] = ia64_get_pmd(i
);
1055 * reload from thread state (used for ctxw only)
1058 pfm_restore_pmds(unsigned long *pmds
, unsigned long mask
)
1061 unsigned long val
, ovfl_val
= pmu_conf
->ovfl_val
;
1063 for (i
=0; mask
; i
++, mask
>>=1) {
1064 if ((mask
& 0x1) == 0) continue;
1065 val
= PMD_IS_COUNTING(i
) ? pmds
[i
] & ovfl_val
: pmds
[i
];
1066 ia64_set_pmd(i
, val
);
1072 * propagate PMD from context to thread-state
1075 pfm_copy_pmds(struct task_struct
*task
, pfm_context_t
*ctx
)
1077 unsigned long ovfl_val
= pmu_conf
->ovfl_val
;
1078 unsigned long mask
= ctx
->ctx_all_pmds
[0];
1082 DPRINT(("mask=0x%lx\n", mask
));
1084 for (i
=0; mask
; i
++, mask
>>=1) {
1086 val
= ctx
->ctx_pmds
[i
].val
;
1089 * We break up the 64 bit value into 2 pieces
1090 * the lower bits go to the machine state in the
1091 * thread (will be reloaded on ctxsw in).
1092 * The upper part stays in the soft-counter.
1094 if (PMD_IS_COUNTING(i
)) {
1095 ctx
->ctx_pmds
[i
].val
= val
& ~ovfl_val
;
1098 ctx
->th_pmds
[i
] = val
;
1100 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1103 ctx
->ctx_pmds
[i
].val
));
1108 * propagate PMC from context to thread-state
1111 pfm_copy_pmcs(struct task_struct
*task
, pfm_context_t
*ctx
)
1113 unsigned long mask
= ctx
->ctx_all_pmcs
[0];
1116 DPRINT(("mask=0x%lx\n", mask
));
1118 for (i
=0; mask
; i
++, mask
>>=1) {
1119 /* masking 0 with ovfl_val yields 0 */
1120 ctx
->th_pmcs
[i
] = ctx
->ctx_pmcs
[i
];
1121 DPRINT(("pmc[%d]=0x%lx\n", i
, ctx
->th_pmcs
[i
]));
1128 pfm_restore_pmcs(unsigned long *pmcs
, unsigned long mask
)
1132 for (i
=0; mask
; i
++, mask
>>=1) {
1133 if ((mask
& 0x1) == 0) continue;
1134 ia64_set_pmc(i
, pmcs
[i
]);
1140 pfm_uuid_cmp(pfm_uuid_t a
, pfm_uuid_t b
)
1142 return memcmp(a
, b
, sizeof(pfm_uuid_t
));
1146 pfm_buf_fmt_exit(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, void *buf
, struct pt_regs
*regs
)
1149 if (fmt
->fmt_exit
) ret
= (*fmt
->fmt_exit
)(task
, buf
, regs
);
1154 pfm_buf_fmt_getsize(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, unsigned int flags
, int cpu
, void *arg
, unsigned long *size
)
1157 if (fmt
->fmt_getsize
) ret
= (*fmt
->fmt_getsize
)(task
, flags
, cpu
, arg
, size
);
1163 pfm_buf_fmt_validate(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, unsigned int flags
,
1167 if (fmt
->fmt_validate
) ret
= (*fmt
->fmt_validate
)(task
, flags
, cpu
, arg
);
1172 pfm_buf_fmt_init(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, void *buf
, unsigned int flags
,
1176 if (fmt
->fmt_init
) ret
= (*fmt
->fmt_init
)(task
, buf
, flags
, cpu
, arg
);
1181 pfm_buf_fmt_restart(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, pfm_ovfl_ctrl_t
*ctrl
, void *buf
, struct pt_regs
*regs
)
1184 if (fmt
->fmt_restart
) ret
= (*fmt
->fmt_restart
)(task
, ctrl
, buf
, regs
);
1189 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
)
1192 if (fmt
->fmt_restart_active
) ret
= (*fmt
->fmt_restart_active
)(task
, ctrl
, buf
, regs
);
1196 static pfm_buffer_fmt_t
*
1197 __pfm_find_buffer_fmt(pfm_uuid_t uuid
)
1199 struct list_head
* pos
;
1200 pfm_buffer_fmt_t
* entry
;
1202 list_for_each(pos
, &pfm_buffer_fmt_list
) {
1203 entry
= list_entry(pos
, pfm_buffer_fmt_t
, fmt_list
);
1204 if (pfm_uuid_cmp(uuid
, entry
->fmt_uuid
) == 0)
1211 * find a buffer format based on its uuid
1213 static pfm_buffer_fmt_t
*
1214 pfm_find_buffer_fmt(pfm_uuid_t uuid
)
1216 pfm_buffer_fmt_t
* fmt
;
1217 spin_lock(&pfm_buffer_fmt_lock
);
1218 fmt
= __pfm_find_buffer_fmt(uuid
);
1219 spin_unlock(&pfm_buffer_fmt_lock
);
1224 pfm_register_buffer_fmt(pfm_buffer_fmt_t
*fmt
)
1228 /* some sanity checks */
1229 if (fmt
== NULL
|| fmt
->fmt_name
== NULL
) return -EINVAL
;
1231 /* we need at least a handler */
1232 if (fmt
->fmt_handler
== NULL
) return -EINVAL
;
1235 * XXX: need check validity of fmt_arg_size
1238 spin_lock(&pfm_buffer_fmt_lock
);
1240 if (__pfm_find_buffer_fmt(fmt
->fmt_uuid
)) {
1241 printk(KERN_ERR
"perfmon: duplicate sampling format: %s\n", fmt
->fmt_name
);
1245 list_add(&fmt
->fmt_list
, &pfm_buffer_fmt_list
);
1246 printk(KERN_INFO
"perfmon: added sampling format %s\n", fmt
->fmt_name
);
1249 spin_unlock(&pfm_buffer_fmt_lock
);
1252 EXPORT_SYMBOL(pfm_register_buffer_fmt
);
1255 pfm_unregister_buffer_fmt(pfm_uuid_t uuid
)
1257 pfm_buffer_fmt_t
*fmt
;
1260 spin_lock(&pfm_buffer_fmt_lock
);
1262 fmt
= __pfm_find_buffer_fmt(uuid
);
1264 printk(KERN_ERR
"perfmon: cannot unregister format, not found\n");
1268 list_del_init(&fmt
->fmt_list
);
1269 printk(KERN_INFO
"perfmon: removed sampling format: %s\n", fmt
->fmt_name
);
1272 spin_unlock(&pfm_buffer_fmt_lock
);
1276 EXPORT_SYMBOL(pfm_unregister_buffer_fmt
);
1278 extern void update_pal_halt_status(int);
1281 pfm_reserve_session(struct task_struct
*task
, int is_syswide
, unsigned int cpu
)
1283 unsigned long flags
;
1285 * validy checks on cpu_mask have been done upstream
1289 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1290 pfm_sessions
.pfs_sys_sessions
,
1291 pfm_sessions
.pfs_task_sessions
,
1292 pfm_sessions
.pfs_sys_use_dbregs
,
1298 * cannot mix system wide and per-task sessions
1300 if (pfm_sessions
.pfs_task_sessions
> 0UL) {
1301 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1302 pfm_sessions
.pfs_task_sessions
));
1306 if (pfm_sessions
.pfs_sys_session
[cpu
]) goto error_conflict
;
1308 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu
, smp_processor_id()));
1310 pfm_sessions
.pfs_sys_session
[cpu
] = task
;
1312 pfm_sessions
.pfs_sys_sessions
++ ;
1315 if (pfm_sessions
.pfs_sys_sessions
) goto abort
;
1316 pfm_sessions
.pfs_task_sessions
++;
1319 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1320 pfm_sessions
.pfs_sys_sessions
,
1321 pfm_sessions
.pfs_task_sessions
,
1322 pfm_sessions
.pfs_sys_use_dbregs
,
1327 * disable default_idle() to go to PAL_HALT
1329 update_pal_halt_status(0);
1336 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1337 pfm_sessions
.pfs_sys_session
[cpu
]->pid
,
1347 pfm_unreserve_session(pfm_context_t
*ctx
, int is_syswide
, unsigned int cpu
)
1349 unsigned long flags
;
1351 * validy checks on cpu_mask have been done upstream
1355 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1356 pfm_sessions
.pfs_sys_sessions
,
1357 pfm_sessions
.pfs_task_sessions
,
1358 pfm_sessions
.pfs_sys_use_dbregs
,
1364 pfm_sessions
.pfs_sys_session
[cpu
] = NULL
;
1366 * would not work with perfmon+more than one bit in cpu_mask
1368 if (ctx
&& ctx
->ctx_fl_using_dbreg
) {
1369 if (pfm_sessions
.pfs_sys_use_dbregs
== 0) {
1370 printk(KERN_ERR
"perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx
);
1372 pfm_sessions
.pfs_sys_use_dbregs
--;
1375 pfm_sessions
.pfs_sys_sessions
--;
1377 pfm_sessions
.pfs_task_sessions
--;
1379 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1380 pfm_sessions
.pfs_sys_sessions
,
1381 pfm_sessions
.pfs_task_sessions
,
1382 pfm_sessions
.pfs_sys_use_dbregs
,
1387 * if possible, enable default_idle() to go into PAL_HALT
1389 if (pfm_sessions
.pfs_task_sessions
== 0 && pfm_sessions
.pfs_sys_sessions
== 0)
1390 update_pal_halt_status(1);
1398 * removes virtual mapping of the sampling buffer.
1399 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1400 * a PROTECT_CTX() section.
1403 pfm_remove_smpl_mapping(struct task_struct
*task
, void *vaddr
, unsigned long size
)
1408 if (task
->mm
== NULL
|| size
== 0UL || vaddr
== NULL
) {
1409 printk(KERN_ERR
"perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task
->pid
, task
->mm
);
1413 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr
, size
));
1416 * does the actual unmapping
1418 down_write(&task
->mm
->mmap_sem
);
1420 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr
, size
));
1422 r
= pfm_do_munmap(task
->mm
, (unsigned long)vaddr
, size
, 0);
1424 up_write(&task
->mm
->mmap_sem
);
1426 printk(KERN_ERR
"perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task
->pid
, vaddr
, size
);
1429 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr
, size
, r
));
1435 * free actual physical storage used by sampling buffer
1439 pfm_free_smpl_buffer(pfm_context_t
*ctx
)
1441 pfm_buffer_fmt_t
*fmt
;
1443 if (ctx
->ctx_smpl_hdr
== NULL
) goto invalid_free
;
1446 * we won't use the buffer format anymore
1448 fmt
= ctx
->ctx_buf_fmt
;
1450 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1453 ctx
->ctx_smpl_vaddr
));
1455 pfm_buf_fmt_exit(fmt
, current
, NULL
, NULL
);
1460 pfm_rvfree(ctx
->ctx_smpl_hdr
, ctx
->ctx_smpl_size
);
1462 ctx
->ctx_smpl_hdr
= NULL
;
1463 ctx
->ctx_smpl_size
= 0UL;
1468 printk(KERN_ERR
"perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current
->pid
);
1474 pfm_exit_smpl_buffer(pfm_buffer_fmt_t
*fmt
)
1476 if (fmt
== NULL
) return;
1478 pfm_buf_fmt_exit(fmt
, current
, NULL
, NULL
);
1483 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1484 * no real gain from having the whole whorehouse mounted. So we don't need
1485 * any operations on the root directory. However, we need a non-trivial
1486 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1488 static struct vfsmount
*pfmfs_mnt
;
1493 int err
= register_filesystem(&pfm_fs_type
);
1495 pfmfs_mnt
= kern_mount(&pfm_fs_type
);
1496 err
= PTR_ERR(pfmfs_mnt
);
1497 if (IS_ERR(pfmfs_mnt
))
1498 unregister_filesystem(&pfm_fs_type
);
1508 unregister_filesystem(&pfm_fs_type
);
1513 pfm_read(struct file
*filp
, char __user
*buf
, size_t size
, loff_t
*ppos
)
1518 unsigned long flags
;
1519 DECLARE_WAITQUEUE(wait
, current
);
1520 if (PFM_IS_FILE(filp
) == 0) {
1521 printk(KERN_ERR
"perfmon: pfm_poll: bad magic [%d]\n", current
->pid
);
1525 ctx
= (pfm_context_t
*)filp
->private_data
;
1527 printk(KERN_ERR
"perfmon: pfm_read: NULL ctx [%d]\n", current
->pid
);
1532 * check even when there is no message
1534 if (size
< sizeof(pfm_msg_t
)) {
1535 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx
, sizeof(pfm_msg_t
)));
1539 PROTECT_CTX(ctx
, flags
);
1542 * put ourselves on the wait queue
1544 add_wait_queue(&ctx
->ctx_msgq_wait
, &wait
);
1552 set_current_state(TASK_INTERRUPTIBLE
);
1554 DPRINT(("head=%d tail=%d\n", ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
1557 if(PFM_CTXQ_EMPTY(ctx
) == 0) break;
1559 UNPROTECT_CTX(ctx
, flags
);
1562 * check non-blocking read
1565 if(filp
->f_flags
& O_NONBLOCK
) break;
1568 * check pending signals
1570 if(signal_pending(current
)) {
1575 * no message, so wait
1579 PROTECT_CTX(ctx
, flags
);
1581 DPRINT(("[%d] back to running ret=%ld\n", current
->pid
, ret
));
1582 set_current_state(TASK_RUNNING
);
1583 remove_wait_queue(&ctx
->ctx_msgq_wait
, &wait
);
1585 if (ret
< 0) goto abort
;
1588 msg
= pfm_get_next_msg(ctx
);
1590 printk(KERN_ERR
"perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx
, current
->pid
);
1594 DPRINT(("fd=%d type=%d\n", msg
->pfm_gen_msg
.msg_ctx_fd
, msg
->pfm_gen_msg
.msg_type
));
1597 if(copy_to_user(buf
, msg
, sizeof(pfm_msg_t
)) == 0) ret
= sizeof(pfm_msg_t
);
1600 UNPROTECT_CTX(ctx
, flags
);
1606 pfm_write(struct file
*file
, const char __user
*ubuf
,
1607 size_t size
, loff_t
*ppos
)
1609 DPRINT(("pfm_write called\n"));
1614 pfm_poll(struct file
*filp
, poll_table
* wait
)
1617 unsigned long flags
;
1618 unsigned int mask
= 0;
1620 if (PFM_IS_FILE(filp
) == 0) {
1621 printk(KERN_ERR
"perfmon: pfm_poll: bad magic [%d]\n", current
->pid
);
1625 ctx
= (pfm_context_t
*)filp
->private_data
;
1627 printk(KERN_ERR
"perfmon: pfm_poll: NULL ctx [%d]\n", current
->pid
);
1632 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx
->ctx_fd
));
1634 poll_wait(filp
, &ctx
->ctx_msgq_wait
, wait
);
1636 PROTECT_CTX(ctx
, flags
);
1638 if (PFM_CTXQ_EMPTY(ctx
) == 0)
1639 mask
= POLLIN
| POLLRDNORM
;
1641 UNPROTECT_CTX(ctx
, flags
);
1643 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx
->ctx_fd
, mask
));
1649 pfm_ioctl(struct inode
*inode
, struct file
*file
, unsigned int cmd
, unsigned long arg
)
1651 DPRINT(("pfm_ioctl called\n"));
1656 * interrupt cannot be masked when coming here
1659 pfm_do_fasync(int fd
, struct file
*filp
, pfm_context_t
*ctx
, int on
)
1663 ret
= fasync_helper (fd
, filp
, on
, &ctx
->ctx_async_queue
);
1665 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1669 ctx
->ctx_async_queue
, ret
));
1675 pfm_fasync(int fd
, struct file
*filp
, int on
)
1680 if (PFM_IS_FILE(filp
) == 0) {
1681 printk(KERN_ERR
"perfmon: pfm_fasync bad magic [%d]\n", current
->pid
);
1685 ctx
= (pfm_context_t
*)filp
->private_data
;
1687 printk(KERN_ERR
"perfmon: pfm_fasync NULL ctx [%d]\n", current
->pid
);
1691 * we cannot mask interrupts during this call because this may
1692 * may go to sleep if memory is not readily avalaible.
1694 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1695 * done in caller. Serialization of this function is ensured by caller.
1697 ret
= pfm_do_fasync(fd
, filp
, ctx
, on
);
1700 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1703 ctx
->ctx_async_queue
, ret
));
1710 * this function is exclusively called from pfm_close().
1711 * The context is not protected at that time, nor are interrupts
1712 * on the remote CPU. That's necessary to avoid deadlocks.
1715 pfm_syswide_force_stop(void *info
)
1717 pfm_context_t
*ctx
= (pfm_context_t
*)info
;
1718 struct pt_regs
*regs
= task_pt_regs(current
);
1719 struct task_struct
*owner
;
1720 unsigned long flags
;
1723 if (ctx
->ctx_cpu
!= smp_processor_id()) {
1724 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1726 smp_processor_id());
1729 owner
= GET_PMU_OWNER();
1730 if (owner
!= ctx
->ctx_task
) {
1731 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1733 owner
->pid
, ctx
->ctx_task
->pid
);
1736 if (GET_PMU_CTX() != ctx
) {
1737 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1739 GET_PMU_CTX(), ctx
);
1743 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx
->ctx_task
->pid
));
1745 * the context is already protected in pfm_close(), we simply
1746 * need to mask interrupts to avoid a PMU interrupt race on
1749 local_irq_save(flags
);
1751 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
1753 DPRINT(("context_unload returned %d\n", ret
));
1757 * unmask interrupts, PMU interrupts are now spurious here
1759 local_irq_restore(flags
);
1763 pfm_syswide_cleanup_other_cpu(pfm_context_t
*ctx
)
1767 DPRINT(("calling CPU%d for cleanup\n", ctx
->ctx_cpu
));
1768 ret
= smp_call_function_single(ctx
->ctx_cpu
, pfm_syswide_force_stop
, ctx
, 0, 1);
1769 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx
->ctx_cpu
, ret
));
1771 #endif /* CONFIG_SMP */
1774 * called for each close(). Partially free resources.
1775 * When caller is self-monitoring, the context is unloaded.
1778 pfm_flush(struct file
*filp
, fl_owner_t id
)
1781 struct task_struct
*task
;
1782 struct pt_regs
*regs
;
1783 unsigned long flags
;
1784 unsigned long smpl_buf_size
= 0UL;
1785 void *smpl_buf_vaddr
= NULL
;
1786 int state
, is_system
;
1788 if (PFM_IS_FILE(filp
) == 0) {
1789 DPRINT(("bad magic for\n"));
1793 ctx
= (pfm_context_t
*)filp
->private_data
;
1795 printk(KERN_ERR
"perfmon: pfm_flush: NULL ctx [%d]\n", current
->pid
);
1800 * remove our file from the async queue, if we use this mode.
1801 * This can be done without the context being protected. We come
1802 * here when the context has become unreacheable by other tasks.
1804 * We may still have active monitoring at this point and we may
1805 * end up in pfm_overflow_handler(). However, fasync_helper()
1806 * operates with interrupts disabled and it cleans up the
1807 * queue. If the PMU handler is called prior to entering
1808 * fasync_helper() then it will send a signal. If it is
1809 * invoked after, it will find an empty queue and no
1810 * signal will be sent. In both case, we are safe
1812 if (filp
->f_flags
& FASYNC
) {
1813 DPRINT(("cleaning up async_queue=%p\n", ctx
->ctx_async_queue
));
1814 pfm_do_fasync (-1, filp
, ctx
, 0);
1817 PROTECT_CTX(ctx
, flags
);
1819 state
= ctx
->ctx_state
;
1820 is_system
= ctx
->ctx_fl_system
;
1822 task
= PFM_CTX_TASK(ctx
);
1823 regs
= task_pt_regs(task
);
1825 DPRINT(("ctx_state=%d is_current=%d\n",
1827 task
== current
? 1 : 0));
1830 * if state == UNLOADED, then task is NULL
1834 * we must stop and unload because we are losing access to the context.
1836 if (task
== current
) {
1839 * the task IS the owner but it migrated to another CPU: that's bad
1840 * but we must handle this cleanly. Unfortunately, the kernel does
1841 * not provide a mechanism to block migration (while the context is loaded).
1843 * We need to release the resource on the ORIGINAL cpu.
1845 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
1847 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
1849 * keep context protected but unmask interrupt for IPI
1851 local_irq_restore(flags
);
1853 pfm_syswide_cleanup_other_cpu(ctx
);
1856 * restore interrupt masking
1858 local_irq_save(flags
);
1861 * context is unloaded at this point
1864 #endif /* CONFIG_SMP */
1867 DPRINT(("forcing unload\n"));
1869 * stop and unload, returning with state UNLOADED
1870 * and session unreserved.
1872 pfm_context_unload(ctx
, NULL
, 0, regs
);
1874 DPRINT(("ctx_state=%d\n", ctx
->ctx_state
));
1879 * remove virtual mapping, if any, for the calling task.
1880 * cannot reset ctx field until last user is calling close().
1882 * ctx_smpl_vaddr must never be cleared because it is needed
1883 * by every task with access to the context
1885 * When called from do_exit(), the mm context is gone already, therefore
1886 * mm is NULL, i.e., the VMA is already gone and we do not have to
1889 if (ctx
->ctx_smpl_vaddr
&& current
->mm
) {
1890 smpl_buf_vaddr
= ctx
->ctx_smpl_vaddr
;
1891 smpl_buf_size
= ctx
->ctx_smpl_size
;
1894 UNPROTECT_CTX(ctx
, flags
);
1897 * if there was a mapping, then we systematically remove it
1898 * at this point. Cannot be done inside critical section
1899 * because some VM function reenables interrupts.
1902 if (smpl_buf_vaddr
) pfm_remove_smpl_mapping(current
, smpl_buf_vaddr
, smpl_buf_size
);
1907 * called either on explicit close() or from exit_files().
1908 * Only the LAST user of the file gets to this point, i.e., it is
1911 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1912 * (fput()),i.e, last task to access the file. Nobody else can access the
1913 * file at this point.
1915 * When called from exit_files(), the VMA has been freed because exit_mm()
1916 * is executed before exit_files().
1918 * When called from exit_files(), the current task is not yet ZOMBIE but we
1919 * flush the PMU state to the context.
1922 pfm_close(struct inode
*inode
, struct file
*filp
)
1925 struct task_struct
*task
;
1926 struct pt_regs
*regs
;
1927 DECLARE_WAITQUEUE(wait
, current
);
1928 unsigned long flags
;
1929 unsigned long smpl_buf_size
= 0UL;
1930 void *smpl_buf_addr
= NULL
;
1931 int free_possible
= 1;
1932 int state
, is_system
;
1934 DPRINT(("pfm_close called private=%p\n", filp
->private_data
));
1936 if (PFM_IS_FILE(filp
) == 0) {
1937 DPRINT(("bad magic\n"));
1941 ctx
= (pfm_context_t
*)filp
->private_data
;
1943 printk(KERN_ERR
"perfmon: pfm_close: NULL ctx [%d]\n", current
->pid
);
1947 PROTECT_CTX(ctx
, flags
);
1949 state
= ctx
->ctx_state
;
1950 is_system
= ctx
->ctx_fl_system
;
1952 task
= PFM_CTX_TASK(ctx
);
1953 regs
= task_pt_regs(task
);
1955 DPRINT(("ctx_state=%d is_current=%d\n",
1957 task
== current
? 1 : 0));
1960 * if task == current, then pfm_flush() unloaded the context
1962 if (state
== PFM_CTX_UNLOADED
) goto doit
;
1965 * context is loaded/masked and task != current, we need to
1966 * either force an unload or go zombie
1970 * The task is currently blocked or will block after an overflow.
1971 * we must force it to wakeup to get out of the
1972 * MASKED state and transition to the unloaded state by itself.
1974 * This situation is only possible for per-task mode
1976 if (state
== PFM_CTX_MASKED
&& CTX_OVFL_NOBLOCK(ctx
) == 0) {
1979 * set a "partial" zombie state to be checked
1980 * upon return from down() in pfm_handle_work().
1982 * We cannot use the ZOMBIE state, because it is checked
1983 * by pfm_load_regs() which is called upon wakeup from down().
1984 * In such case, it would free the context and then we would
1985 * return to pfm_handle_work() which would access the
1986 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1987 * but visible to pfm_handle_work().
1989 * For some window of time, we have a zombie context with
1990 * ctx_state = MASKED and not ZOMBIE
1992 ctx
->ctx_fl_going_zombie
= 1;
1995 * force task to wake up from MASKED state
1997 complete(&ctx
->ctx_restart_done
);
1999 DPRINT(("waking up ctx_state=%d\n", state
));
2002 * put ourself to sleep waiting for the other
2003 * task to report completion
2005 * the context is protected by mutex, therefore there
2006 * is no risk of being notified of completion before
2007 * begin actually on the waitq.
2009 set_current_state(TASK_INTERRUPTIBLE
);
2010 add_wait_queue(&ctx
->ctx_zombieq
, &wait
);
2012 UNPROTECT_CTX(ctx
, flags
);
2015 * XXX: check for signals :
2016 * - ok for explicit close
2017 * - not ok when coming from exit_files()
2022 PROTECT_CTX(ctx
, flags
);
2025 remove_wait_queue(&ctx
->ctx_zombieq
, &wait
);
2026 set_current_state(TASK_RUNNING
);
2029 * context is unloaded at this point
2031 DPRINT(("after zombie wakeup ctx_state=%d for\n", state
));
2033 else if (task
!= current
) {
2036 * switch context to zombie state
2038 ctx
->ctx_state
= PFM_CTX_ZOMBIE
;
2040 DPRINT(("zombie ctx for [%d]\n", task
->pid
));
2042 * cannot free the context on the spot. deferred until
2043 * the task notices the ZOMBIE state
2047 pfm_context_unload(ctx
, NULL
, 0, regs
);
2052 /* reload state, may have changed during opening of critical section */
2053 state
= ctx
->ctx_state
;
2056 * the context is still attached to a task (possibly current)
2057 * we cannot destroy it right now
2061 * we must free the sampling buffer right here because
2062 * we cannot rely on it being cleaned up later by the
2063 * monitored task. It is not possible to free vmalloc'ed
2064 * memory in pfm_load_regs(). Instead, we remove the buffer
2065 * now. should there be subsequent PMU overflow originally
2066 * meant for sampling, the will be converted to spurious
2067 * and that's fine because the monitoring tools is gone anyway.
2069 if (ctx
->ctx_smpl_hdr
) {
2070 smpl_buf_addr
= ctx
->ctx_smpl_hdr
;
2071 smpl_buf_size
= ctx
->ctx_smpl_size
;
2072 /* no more sampling */
2073 ctx
->ctx_smpl_hdr
= NULL
;
2074 ctx
->ctx_fl_is_sampling
= 0;
2077 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2083 if (smpl_buf_addr
) pfm_exit_smpl_buffer(ctx
->ctx_buf_fmt
);
2086 * UNLOADED that the session has already been unreserved.
2088 if (state
== PFM_CTX_ZOMBIE
) {
2089 pfm_unreserve_session(ctx
, ctx
->ctx_fl_system
, ctx
->ctx_cpu
);
2093 * disconnect file descriptor from context must be done
2096 filp
->private_data
= NULL
;
2099 * if we free on the spot, the context is now completely unreacheable
2100 * from the callers side. The monitored task side is also cut, so we
2103 * If we have a deferred free, only the caller side is disconnected.
2105 UNPROTECT_CTX(ctx
, flags
);
2108 * All memory free operations (especially for vmalloc'ed memory)
2109 * MUST be done with interrupts ENABLED.
2111 if (smpl_buf_addr
) pfm_rvfree(smpl_buf_addr
, smpl_buf_size
);
2114 * return the memory used by the context
2116 if (free_possible
) pfm_context_free(ctx
);
2122 pfm_no_open(struct inode
*irrelevant
, struct file
*dontcare
)
2124 DPRINT(("pfm_no_open called\n"));
2130 static struct file_operations pfm_file_ops
= {
2131 .llseek
= no_llseek
,
2136 .open
= pfm_no_open
, /* special open code to disallow open via /proc */
2137 .fasync
= pfm_fasync
,
2138 .release
= pfm_close
,
2143 pfmfs_delete_dentry(struct dentry
*dentry
)
2148 static struct dentry_operations pfmfs_dentry_operations
= {
2149 .d_delete
= pfmfs_delete_dentry
,
2154 pfm_alloc_fd(struct file
**cfile
)
2157 struct file
*file
= NULL
;
2158 struct inode
* inode
;
2162 fd
= get_unused_fd();
2163 if (fd
< 0) return -ENFILE
;
2167 file
= get_empty_filp();
2168 if (!file
) goto out
;
2171 * allocate a new inode
2173 inode
= new_inode(pfmfs_mnt
->mnt_sb
);
2174 if (!inode
) goto out
;
2176 DPRINT(("new inode ino=%ld @%p\n", inode
->i_ino
, inode
));
2178 inode
->i_mode
= S_IFCHR
|S_IRUGO
;
2179 inode
->i_uid
= current
->fsuid
;
2180 inode
->i_gid
= current
->fsgid
;
2182 sprintf(name
, "[%lu]", inode
->i_ino
);
2184 this.len
= strlen(name
);
2185 this.hash
= inode
->i_ino
;
2190 * allocate a new dcache entry
2192 file
->f_dentry
= d_alloc(pfmfs_mnt
->mnt_sb
->s_root
, &this);
2193 if (!file
->f_dentry
) goto out
;
2195 file
->f_dentry
->d_op
= &pfmfs_dentry_operations
;
2197 d_add(file
->f_dentry
, inode
);
2198 file
->f_vfsmnt
= mntget(pfmfs_mnt
);
2199 file
->f_mapping
= inode
->i_mapping
;
2201 file
->f_op
= &pfm_file_ops
;
2202 file
->f_mode
= FMODE_READ
;
2203 file
->f_flags
= O_RDONLY
;
2207 * may have to delay until context is attached?
2209 fd_install(fd
, file
);
2212 * the file structure we will use
2218 if (file
) put_filp(file
);
2224 pfm_free_fd(int fd
, struct file
*file
)
2226 struct files_struct
*files
= current
->files
;
2227 struct fdtable
*fdt
;
2230 * there ie no fd_uninstall(), so we do it here
2232 spin_lock(&files
->file_lock
);
2233 fdt
= files_fdtable(files
);
2234 rcu_assign_pointer(fdt
->fd
[fd
], NULL
);
2235 spin_unlock(&files
->file_lock
);
2243 pfm_remap_buffer(struct vm_area_struct
*vma
, unsigned long buf
, unsigned long addr
, unsigned long size
)
2245 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf
, addr
, size
));
2248 unsigned long pfn
= ia64_tpa(buf
) >> PAGE_SHIFT
;
2251 if (remap_pfn_range(vma
, addr
, pfn
, PAGE_SIZE
, PAGE_READONLY
))
2262 * allocate a sampling buffer and remaps it into the user address space of the task
2265 pfm_smpl_buffer_alloc(struct task_struct
*task
, pfm_context_t
*ctx
, unsigned long rsize
, void **user_vaddr
)
2267 struct mm_struct
*mm
= task
->mm
;
2268 struct vm_area_struct
*vma
= NULL
;
2274 * the fixed header + requested size and align to page boundary
2276 size
= PAGE_ALIGN(rsize
);
2278 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize
, size
));
2281 * check requested size to avoid Denial-of-service attacks
2282 * XXX: may have to refine this test
2283 * Check against address space limit.
2285 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2288 if (size
> task
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
)
2292 * We do the easy to undo allocations first.
2294 * pfm_rvmalloc(), clears the buffer, so there is no leak
2296 smpl_buf
= pfm_rvmalloc(size
);
2297 if (smpl_buf
== NULL
) {
2298 DPRINT(("Can't allocate sampling buffer\n"));
2302 DPRINT(("smpl_buf @%p\n", smpl_buf
));
2305 vma
= kmem_cache_alloc(vm_area_cachep
, SLAB_KERNEL
);
2307 DPRINT(("Cannot allocate vma\n"));
2310 memset(vma
, 0, sizeof(*vma
));
2313 * partially initialize the vma for the sampling buffer
2316 vma
->vm_flags
= VM_READ
| VM_MAYREAD
|VM_RESERVED
;
2317 vma
->vm_page_prot
= PAGE_READONLY
; /* XXX may need to change */
2320 * Now we have everything we need and we can initialize
2321 * and connect all the data structures
2324 ctx
->ctx_smpl_hdr
= smpl_buf
;
2325 ctx
->ctx_smpl_size
= size
; /* aligned size */
2328 * Let's do the difficult operations next.
2330 * now we atomically find some area in the address space and
2331 * remap the buffer in it.
2333 down_write(&task
->mm
->mmap_sem
);
2335 /* find some free area in address space, must have mmap sem held */
2336 vma
->vm_start
= pfm_get_unmapped_area(NULL
, 0, size
, 0, MAP_PRIVATE
|MAP_ANONYMOUS
, 0);
2337 if (vma
->vm_start
== 0UL) {
2338 DPRINT(("Cannot find unmapped area for size %ld\n", size
));
2339 up_write(&task
->mm
->mmap_sem
);
2342 vma
->vm_end
= vma
->vm_start
+ size
;
2343 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2345 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size
, ctx
->ctx_smpl_hdr
, vma
->vm_start
));
2347 /* can only be applied to current task, need to have the mm semaphore held when called */
2348 if (pfm_remap_buffer(vma
, (unsigned long)smpl_buf
, vma
->vm_start
, size
)) {
2349 DPRINT(("Can't remap buffer\n"));
2350 up_write(&task
->mm
->mmap_sem
);
2355 * now insert the vma in the vm list for the process, must be
2356 * done with mmap lock held
2358 insert_vm_struct(mm
, vma
);
2360 mm
->total_vm
+= size
>> PAGE_SHIFT
;
2361 vm_stat_account(vma
->vm_mm
, vma
->vm_flags
, vma
->vm_file
,
2363 up_write(&task
->mm
->mmap_sem
);
2366 * keep track of user level virtual address
2368 ctx
->ctx_smpl_vaddr
= (void *)vma
->vm_start
;
2369 *(unsigned long *)user_vaddr
= vma
->vm_start
;
2374 kmem_cache_free(vm_area_cachep
, vma
);
2376 pfm_rvfree(smpl_buf
, size
);
2382 * XXX: do something better here
2385 pfm_bad_permissions(struct task_struct
*task
)
2387 /* inspired by ptrace_attach() */
2388 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2397 return ((current
->uid
!= task
->euid
)
2398 || (current
->uid
!= task
->suid
)
2399 || (current
->uid
!= task
->uid
)
2400 || (current
->gid
!= task
->egid
)
2401 || (current
->gid
!= task
->sgid
)
2402 || (current
->gid
!= task
->gid
)) && !capable(CAP_SYS_PTRACE
);
2406 pfarg_is_sane(struct task_struct
*task
, pfarg_context_t
*pfx
)
2412 ctx_flags
= pfx
->ctx_flags
;
2414 if (ctx_flags
& PFM_FL_SYSTEM_WIDE
) {
2417 * cannot block in this mode
2419 if (ctx_flags
& PFM_FL_NOTIFY_BLOCK
) {
2420 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2425 /* probably more to add here */
2431 pfm_setup_buffer_fmt(struct task_struct
*task
, pfm_context_t
*ctx
, unsigned int ctx_flags
,
2432 unsigned int cpu
, pfarg_context_t
*arg
)
2434 pfm_buffer_fmt_t
*fmt
= NULL
;
2435 unsigned long size
= 0UL;
2437 void *fmt_arg
= NULL
;
2439 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2441 /* invoke and lock buffer format, if found */
2442 fmt
= pfm_find_buffer_fmt(arg
->ctx_smpl_buf_id
);
2444 DPRINT(("[%d] cannot find buffer format\n", task
->pid
));
2449 * buffer argument MUST be contiguous to pfarg_context_t
2451 if (fmt
->fmt_arg_size
) fmt_arg
= PFM_CTXARG_BUF_ARG(arg
);
2453 ret
= pfm_buf_fmt_validate(fmt
, task
, ctx_flags
, cpu
, fmt_arg
);
2455 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task
->pid
, ctx_flags
, cpu
, fmt_arg
, ret
));
2457 if (ret
) goto error
;
2459 /* link buffer format and context */
2460 ctx
->ctx_buf_fmt
= fmt
;
2463 * check if buffer format wants to use perfmon buffer allocation/mapping service
2465 ret
= pfm_buf_fmt_getsize(fmt
, task
, ctx_flags
, cpu
, fmt_arg
, &size
);
2466 if (ret
) goto error
;
2470 * buffer is always remapped into the caller's address space
2472 ret
= pfm_smpl_buffer_alloc(current
, ctx
, size
, &uaddr
);
2473 if (ret
) goto error
;
2475 /* keep track of user address of buffer */
2476 arg
->ctx_smpl_vaddr
= uaddr
;
2478 ret
= pfm_buf_fmt_init(fmt
, task
, ctx
->ctx_smpl_hdr
, ctx_flags
, cpu
, fmt_arg
);
2485 pfm_reset_pmu_state(pfm_context_t
*ctx
)
2490 * install reset values for PMC.
2492 for (i
=1; PMC_IS_LAST(i
) == 0; i
++) {
2493 if (PMC_IS_IMPL(i
) == 0) continue;
2494 ctx
->ctx_pmcs
[i
] = PMC_DFL_VAL(i
);
2495 DPRINT(("pmc[%d]=0x%lx\n", i
, ctx
->ctx_pmcs
[i
]));
2498 * PMD registers are set to 0UL when the context in memset()
2502 * On context switched restore, we must restore ALL pmc and ALL pmd even
2503 * when they are not actively used by the task. In UP, the incoming process
2504 * may otherwise pick up left over PMC, PMD state from the previous process.
2505 * As opposed to PMD, stale PMC can cause harm to the incoming
2506 * process because they may change what is being measured.
2507 * Therefore, we must systematically reinstall the entire
2508 * PMC state. In SMP, the same thing is possible on the
2509 * same CPU but also on between 2 CPUs.
2511 * The problem with PMD is information leaking especially
2512 * to user level when psr.sp=0
2514 * There is unfortunately no easy way to avoid this problem
2515 * on either UP or SMP. This definitively slows down the
2516 * pfm_load_regs() function.
2520 * bitmask of all PMCs accessible to this context
2522 * PMC0 is treated differently.
2524 ctx
->ctx_all_pmcs
[0] = pmu_conf
->impl_pmcs
[0] & ~0x1;
2527 * bitmask of all PMDs that are accesible to this context
2529 ctx
->ctx_all_pmds
[0] = pmu_conf
->impl_pmds
[0];
2531 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx
->ctx_fd
, ctx
->ctx_all_pmcs
[0],ctx
->ctx_all_pmds
[0]));
2534 * useful in case of re-enable after disable
2536 ctx
->ctx_used_ibrs
[0] = 0UL;
2537 ctx
->ctx_used_dbrs
[0] = 0UL;
2541 pfm_ctx_getsize(void *arg
, size_t *sz
)
2543 pfarg_context_t
*req
= (pfarg_context_t
*)arg
;
2544 pfm_buffer_fmt_t
*fmt
;
2548 if (!pfm_uuid_cmp(req
->ctx_smpl_buf_id
, pfm_null_uuid
)) return 0;
2550 fmt
= pfm_find_buffer_fmt(req
->ctx_smpl_buf_id
);
2552 DPRINT(("cannot find buffer format\n"));
2555 /* get just enough to copy in user parameters */
2556 *sz
= fmt
->fmt_arg_size
;
2557 DPRINT(("arg_size=%lu\n", *sz
));
2565 * cannot attach if :
2567 * - task not owned by caller
2568 * - task incompatible with context mode
2571 pfm_task_incompatible(pfm_context_t
*ctx
, struct task_struct
*task
)
2574 * no kernel task or task not owner by caller
2576 if (task
->mm
== NULL
) {
2577 DPRINT(("task [%d] has not memory context (kernel thread)\n", task
->pid
));
2580 if (pfm_bad_permissions(task
)) {
2581 DPRINT(("no permission to attach to [%d]\n", task
->pid
));
2585 * cannot block in self-monitoring mode
2587 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && task
== current
) {
2588 DPRINT(("cannot load a blocking context on self for [%d]\n", task
->pid
));
2592 if (task
->exit_state
== EXIT_ZOMBIE
) {
2593 DPRINT(("cannot attach to zombie task [%d]\n", task
->pid
));
2598 * always ok for self
2600 if (task
== current
) return 0;
2602 if ((task
->state
!= TASK_STOPPED
) && (task
->state
!= TASK_TRACED
)) {
2603 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task
->pid
, task
->state
));
2607 * make sure the task is off any CPU
2609 wait_task_inactive(task
);
2611 /* more to come... */
2617 pfm_get_task(pfm_context_t
*ctx
, pid_t pid
, struct task_struct
**task
)
2619 struct task_struct
*p
= current
;
2622 /* XXX: need to add more checks here */
2623 if (pid
< 2) return -EPERM
;
2625 if (pid
!= current
->pid
) {
2627 read_lock(&tasklist_lock
);
2629 p
= find_task_by_pid(pid
);
2631 /* make sure task cannot go away while we operate on it */
2632 if (p
) get_task_struct(p
);
2634 read_unlock(&tasklist_lock
);
2636 if (p
== NULL
) return -ESRCH
;
2639 ret
= pfm_task_incompatible(ctx
, p
);
2642 } else if (p
!= current
) {
2651 pfm_context_create(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
2653 pfarg_context_t
*req
= (pfarg_context_t
*)arg
;
2658 /* let's check the arguments first */
2659 ret
= pfarg_is_sane(current
, req
);
2660 if (ret
< 0) return ret
;
2662 ctx_flags
= req
->ctx_flags
;
2666 ctx
= pfm_context_alloc();
2667 if (!ctx
) goto error
;
2669 ret
= pfm_alloc_fd(&filp
);
2670 if (ret
< 0) goto error_file
;
2672 req
->ctx_fd
= ctx
->ctx_fd
= ret
;
2675 * attach context to file
2677 filp
->private_data
= ctx
;
2680 * does the user want to sample?
2682 if (pfm_uuid_cmp(req
->ctx_smpl_buf_id
, pfm_null_uuid
)) {
2683 ret
= pfm_setup_buffer_fmt(current
, ctx
, ctx_flags
, 0, req
);
2684 if (ret
) goto buffer_error
;
2688 * init context protection lock
2690 spin_lock_init(&ctx
->ctx_lock
);
2693 * context is unloaded
2695 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
2698 * initialization of context's flags
2700 ctx
->ctx_fl_block
= (ctx_flags
& PFM_FL_NOTIFY_BLOCK
) ? 1 : 0;
2701 ctx
->ctx_fl_system
= (ctx_flags
& PFM_FL_SYSTEM_WIDE
) ? 1: 0;
2702 ctx
->ctx_fl_is_sampling
= ctx
->ctx_buf_fmt
? 1 : 0; /* assume record() is defined */
2703 ctx
->ctx_fl_no_msg
= (ctx_flags
& PFM_FL_OVFL_NO_MSG
) ? 1: 0;
2705 * will move to set properties
2706 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2710 * init restart semaphore to locked
2712 init_completion(&ctx
->ctx_restart_done
);
2715 * activation is used in SMP only
2717 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
2718 SET_LAST_CPU(ctx
, -1);
2721 * initialize notification message queue
2723 ctx
->ctx_msgq_head
= ctx
->ctx_msgq_tail
= 0;
2724 init_waitqueue_head(&ctx
->ctx_msgq_wait
);
2725 init_waitqueue_head(&ctx
->ctx_zombieq
);
2727 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2732 ctx
->ctx_fl_excl_idle
,
2737 * initialize soft PMU state
2739 pfm_reset_pmu_state(ctx
);
2744 pfm_free_fd(ctx
->ctx_fd
, filp
);
2746 if (ctx
->ctx_buf_fmt
) {
2747 pfm_buf_fmt_exit(ctx
->ctx_buf_fmt
, current
, NULL
, regs
);
2750 pfm_context_free(ctx
);
2756 static inline unsigned long
2757 pfm_new_counter_value (pfm_counter_t
*reg
, int is_long_reset
)
2759 unsigned long val
= is_long_reset
? reg
->long_reset
: reg
->short_reset
;
2760 unsigned long new_seed
, old_seed
= reg
->seed
, mask
= reg
->mask
;
2761 extern unsigned long carta_random32 (unsigned long seed
);
2763 if (reg
->flags
& PFM_REGFL_RANDOM
) {
2764 new_seed
= carta_random32(old_seed
);
2765 val
-= (old_seed
& mask
); /* counter values are negative numbers! */
2766 if ((mask
>> 32) != 0)
2767 /* construct a full 64-bit random value: */
2768 new_seed
|= carta_random32(old_seed
>> 32) << 32;
2769 reg
->seed
= new_seed
;
2776 pfm_reset_regs_masked(pfm_context_t
*ctx
, unsigned long *ovfl_regs
, int is_long_reset
)
2778 unsigned long mask
= ovfl_regs
[0];
2779 unsigned long reset_others
= 0UL;
2784 * now restore reset value on sampling overflowed counters
2786 mask
>>= PMU_FIRST_COUNTER
;
2787 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
2789 if ((mask
& 0x1UL
) == 0UL) continue;
2791 ctx
->ctx_pmds
[i
].val
= val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2792 reset_others
|= ctx
->ctx_pmds
[i
].reset_pmds
[0];
2794 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset
? "long" : "short", i
, val
));
2798 * Now take care of resetting the other registers
2800 for(i
= 0; reset_others
; i
++, reset_others
>>= 1) {
2802 if ((reset_others
& 0x1) == 0) continue;
2804 ctx
->ctx_pmds
[i
].val
= val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2806 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2807 is_long_reset
? "long" : "short", i
, val
));
2812 pfm_reset_regs(pfm_context_t
*ctx
, unsigned long *ovfl_regs
, int is_long_reset
)
2814 unsigned long mask
= ovfl_regs
[0];
2815 unsigned long reset_others
= 0UL;
2819 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs
[0], is_long_reset
));
2821 if (ctx
->ctx_state
== PFM_CTX_MASKED
) {
2822 pfm_reset_regs_masked(ctx
, ovfl_regs
, is_long_reset
);
2827 * now restore reset value on sampling overflowed counters
2829 mask
>>= PMU_FIRST_COUNTER
;
2830 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
2832 if ((mask
& 0x1UL
) == 0UL) continue;
2834 val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2835 reset_others
|= ctx
->ctx_pmds
[i
].reset_pmds
[0];
2837 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset
? "long" : "short", i
, val
));
2839 pfm_write_soft_counter(ctx
, i
, val
);
2843 * Now take care of resetting the other registers
2845 for(i
= 0; reset_others
; i
++, reset_others
>>= 1) {
2847 if ((reset_others
& 0x1) == 0) continue;
2849 val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2851 if (PMD_IS_COUNTING(i
)) {
2852 pfm_write_soft_counter(ctx
, i
, val
);
2854 ia64_set_pmd(i
, val
);
2856 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2857 is_long_reset
? "long" : "short", i
, val
));
2863 pfm_write_pmcs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
2865 struct task_struct
*task
;
2866 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
2867 unsigned long value
, pmc_pm
;
2868 unsigned long smpl_pmds
, reset_pmds
, impl_pmds
;
2869 unsigned int cnum
, reg_flags
, flags
, pmc_type
;
2870 int i
, can_access_pmu
= 0, is_loaded
, is_system
, expert_mode
;
2871 int is_monitor
, is_counting
, state
;
2873 pfm_reg_check_t wr_func
;
2874 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2876 state
= ctx
->ctx_state
;
2877 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
2878 is_system
= ctx
->ctx_fl_system
;
2879 task
= ctx
->ctx_task
;
2880 impl_pmds
= pmu_conf
->impl_pmds
[0];
2882 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
2886 * In system wide and when the context is loaded, access can only happen
2887 * when the caller is running on the CPU being monitored by the session.
2888 * It does not have to be the owner (ctx_task) of the context per se.
2890 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
2891 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
2894 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
2896 expert_mode
= pfm_sysctl
.expert_mode
;
2898 for (i
= 0; i
< count
; i
++, req
++) {
2900 cnum
= req
->reg_num
;
2901 reg_flags
= req
->reg_flags
;
2902 value
= req
->reg_value
;
2903 smpl_pmds
= req
->reg_smpl_pmds
[0];
2904 reset_pmds
= req
->reg_reset_pmds
[0];
2908 if (cnum
>= PMU_MAX_PMCS
) {
2909 DPRINT(("pmc%u is invalid\n", cnum
));
2913 pmc_type
= pmu_conf
->pmc_desc
[cnum
].type
;
2914 pmc_pm
= (value
>> pmu_conf
->pmc_desc
[cnum
].pm_pos
) & 0x1;
2915 is_counting
= (pmc_type
& PFM_REG_COUNTING
) == PFM_REG_COUNTING
? 1 : 0;
2916 is_monitor
= (pmc_type
& PFM_REG_MONITOR
) == PFM_REG_MONITOR
? 1 : 0;
2919 * we reject all non implemented PMC as well
2920 * as attempts to modify PMC[0-3] which are used
2921 * as status registers by the PMU
2923 if ((pmc_type
& PFM_REG_IMPL
) == 0 || (pmc_type
& PFM_REG_CONTROL
) == PFM_REG_CONTROL
) {
2924 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum
, pmc_type
));
2927 wr_func
= pmu_conf
->pmc_desc
[cnum
].write_check
;
2929 * If the PMC is a monitor, then if the value is not the default:
2930 * - system-wide session: PMCx.pm=1 (privileged monitor)
2931 * - per-task : PMCx.pm=0 (user monitor)
2933 if (is_monitor
&& value
!= PMC_DFL_VAL(cnum
) && is_system
^ pmc_pm
) {
2934 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2943 * enforce generation of overflow interrupt. Necessary on all
2946 value
|= 1 << PMU_PMC_OI
;
2948 if (reg_flags
& PFM_REGFL_OVFL_NOTIFY
) {
2949 flags
|= PFM_REGFL_OVFL_NOTIFY
;
2952 if (reg_flags
& PFM_REGFL_RANDOM
) flags
|= PFM_REGFL_RANDOM
;
2954 /* verify validity of smpl_pmds */
2955 if ((smpl_pmds
& impl_pmds
) != smpl_pmds
) {
2956 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds
, cnum
));
2960 /* verify validity of reset_pmds */
2961 if ((reset_pmds
& impl_pmds
) != reset_pmds
) {
2962 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds
, cnum
));
2966 if (reg_flags
& (PFM_REGFL_OVFL_NOTIFY
|PFM_REGFL_RANDOM
)) {
2967 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum
));
2970 /* eventid on non-counting monitors are ignored */
2974 * execute write checker, if any
2976 if (likely(expert_mode
== 0 && wr_func
)) {
2977 ret
= (*wr_func
)(task
, ctx
, cnum
, &value
, regs
);
2978 if (ret
) goto error
;
2983 * no error on this register
2985 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
2988 * Now we commit the changes to the software state
2992 * update overflow information
2996 * full flag update each time a register is programmed
2998 ctx
->ctx_pmds
[cnum
].flags
= flags
;
3000 ctx
->ctx_pmds
[cnum
].reset_pmds
[0] = reset_pmds
;
3001 ctx
->ctx_pmds
[cnum
].smpl_pmds
[0] = smpl_pmds
;
3002 ctx
->ctx_pmds
[cnum
].eventid
= req
->reg_smpl_eventid
;
3005 * Mark all PMDS to be accessed as used.
3007 * We do not keep track of PMC because we have to
3008 * systematically restore ALL of them.
3010 * We do not update the used_monitors mask, because
3011 * if we have not programmed them, then will be in
3012 * a quiescent state, therefore we will not need to
3013 * mask/restore then when context is MASKED.
3015 CTX_USED_PMD(ctx
, reset_pmds
);
3016 CTX_USED_PMD(ctx
, smpl_pmds
);
3018 * make sure we do not try to reset on
3019 * restart because we have established new values
3021 if (state
== PFM_CTX_MASKED
) ctx
->ctx_ovfl_regs
[0] &= ~1UL << cnum
;
3024 * Needed in case the user does not initialize the equivalent
3025 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3026 * possible leak here.
3028 CTX_USED_PMD(ctx
, pmu_conf
->pmc_desc
[cnum
].dep_pmd
[0]);
3031 * keep track of the monitor PMC that we are using.
3032 * we save the value of the pmc in ctx_pmcs[] and if
3033 * the monitoring is not stopped for the context we also
3034 * place it in the saved state area so that it will be
3035 * picked up later by the context switch code.
3037 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3039 * The value in th_pmcs[] may be modified on overflow, i.e., when
3040 * monitoring needs to be stopped.
3042 if (is_monitor
) CTX_USED_MONITOR(ctx
, 1UL << cnum
);
3045 * update context state
3047 ctx
->ctx_pmcs
[cnum
] = value
;
3051 * write thread state
3053 if (is_system
== 0) ctx
->th_pmcs
[cnum
] = value
;
3056 * write hardware register if we can
3058 if (can_access_pmu
) {
3059 ia64_set_pmc(cnum
, value
);
3064 * per-task SMP only here
3066 * we are guaranteed that the task is not running on the other CPU,
3067 * we indicate that this PMD will need to be reloaded if the task
3068 * is rescheduled on the CPU it ran last on.
3070 ctx
->ctx_reload_pmcs
[0] |= 1UL << cnum
;
3075 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",
3081 ctx
->ctx_all_pmcs
[0],
3082 ctx
->ctx_used_pmds
[0],
3083 ctx
->ctx_pmds
[cnum
].eventid
,
3086 ctx
->ctx_reload_pmcs
[0],
3087 ctx
->ctx_used_monitors
[0],
3088 ctx
->ctx_ovfl_regs
[0]));
3092 * make sure the changes are visible
3094 if (can_access_pmu
) ia64_srlz_d();
3098 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3103 pfm_write_pmds(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3105 struct task_struct
*task
;
3106 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
3107 unsigned long value
, hw_value
, ovfl_mask
;
3109 int i
, can_access_pmu
= 0, state
;
3110 int is_counting
, is_loaded
, is_system
, expert_mode
;
3112 pfm_reg_check_t wr_func
;
3115 state
= ctx
->ctx_state
;
3116 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3117 is_system
= ctx
->ctx_fl_system
;
3118 ovfl_mask
= pmu_conf
->ovfl_val
;
3119 task
= ctx
->ctx_task
;
3121 if (unlikely(state
== PFM_CTX_ZOMBIE
)) return -EINVAL
;
3124 * on both UP and SMP, we can only write to the PMC when the task is
3125 * the owner of the local PMU.
3127 if (likely(is_loaded
)) {
3129 * In system wide and when the context is loaded, access can only happen
3130 * when the caller is running on the CPU being monitored by the session.
3131 * It does not have to be the owner (ctx_task) of the context per se.
3133 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3134 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3137 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3139 expert_mode
= pfm_sysctl
.expert_mode
;
3141 for (i
= 0; i
< count
; i
++, req
++) {
3143 cnum
= req
->reg_num
;
3144 value
= req
->reg_value
;
3146 if (!PMD_IS_IMPL(cnum
)) {
3147 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum
));
3150 is_counting
= PMD_IS_COUNTING(cnum
);
3151 wr_func
= pmu_conf
->pmd_desc
[cnum
].write_check
;
3154 * execute write checker, if any
3156 if (unlikely(expert_mode
== 0 && wr_func
)) {
3157 unsigned long v
= value
;
3159 ret
= (*wr_func
)(task
, ctx
, cnum
, &v
, regs
);
3160 if (ret
) goto abort_mission
;
3167 * no error on this register
3169 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
3172 * now commit changes to software state
3177 * update virtualized (64bits) counter
3181 * write context state
3183 ctx
->ctx_pmds
[cnum
].lval
= value
;
3186 * when context is load we use the split value
3189 hw_value
= value
& ovfl_mask
;
3190 value
= value
& ~ovfl_mask
;
3194 * update reset values (not just for counters)
3196 ctx
->ctx_pmds
[cnum
].long_reset
= req
->reg_long_reset
;
3197 ctx
->ctx_pmds
[cnum
].short_reset
= req
->reg_short_reset
;
3200 * update randomization parameters (not just for counters)
3202 ctx
->ctx_pmds
[cnum
].seed
= req
->reg_random_seed
;
3203 ctx
->ctx_pmds
[cnum
].mask
= req
->reg_random_mask
;
3206 * update context value
3208 ctx
->ctx_pmds
[cnum
].val
= value
;
3211 * Keep track of what we use
3213 * We do not keep track of PMC because we have to
3214 * systematically restore ALL of them.
3216 CTX_USED_PMD(ctx
, PMD_PMD_DEP(cnum
));
3219 * mark this PMD register used as well
3221 CTX_USED_PMD(ctx
, RDEP(cnum
));
3224 * make sure we do not try to reset on
3225 * restart because we have established new values
3227 if (is_counting
&& state
== PFM_CTX_MASKED
) {
3228 ctx
->ctx_ovfl_regs
[0] &= ~1UL << cnum
;
3233 * write thread state
3235 if (is_system
== 0) ctx
->th_pmds
[cnum
] = hw_value
;
3238 * write hardware register if we can
3240 if (can_access_pmu
) {
3241 ia64_set_pmd(cnum
, hw_value
);
3245 * we are guaranteed that the task is not running on the other CPU,
3246 * we indicate that this PMD will need to be reloaded if the task
3247 * is rescheduled on the CPU it ran last on.
3249 ctx
->ctx_reload_pmds
[0] |= 1UL << cnum
;
3254 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3255 "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",
3261 ctx
->ctx_pmds
[cnum
].val
,
3262 ctx
->ctx_pmds
[cnum
].short_reset
,
3263 ctx
->ctx_pmds
[cnum
].long_reset
,
3264 PMC_OVFL_NOTIFY(ctx
, cnum
) ? 'Y':'N',
3265 ctx
->ctx_pmds
[cnum
].seed
,
3266 ctx
->ctx_pmds
[cnum
].mask
,
3267 ctx
->ctx_used_pmds
[0],
3268 ctx
->ctx_pmds
[cnum
].reset_pmds
[0],
3269 ctx
->ctx_reload_pmds
[0],
3270 ctx
->ctx_all_pmds
[0],
3271 ctx
->ctx_ovfl_regs
[0]));
3275 * make changes visible
3277 if (can_access_pmu
) ia64_srlz_d();
3283 * for now, we have only one possibility for error
3285 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3290 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3291 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3292 * interrupt is delivered during the call, it will be kept pending until we leave, making
3293 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3294 * guaranteed to return consistent data to the user, it may simply be old. It is not
3295 * trivial to treat the overflow while inside the call because you may end up in
3296 * some module sampling buffer code causing deadlocks.
3299 pfm_read_pmds(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3301 struct task_struct
*task
;
3302 unsigned long val
= 0UL, lval
, ovfl_mask
, sval
;
3303 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
3304 unsigned int cnum
, reg_flags
= 0;
3305 int i
, can_access_pmu
= 0, state
;
3306 int is_loaded
, is_system
, is_counting
, expert_mode
;
3308 pfm_reg_check_t rd_func
;
3311 * access is possible when loaded only for
3312 * self-monitoring tasks or in UP mode
3315 state
= ctx
->ctx_state
;
3316 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3317 is_system
= ctx
->ctx_fl_system
;
3318 ovfl_mask
= pmu_conf
->ovfl_val
;
3319 task
= ctx
->ctx_task
;
3321 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
3323 if (likely(is_loaded
)) {
3325 * In system wide and when the context is loaded, access can only happen
3326 * when the caller is running on the CPU being monitored by the session.
3327 * It does not have to be the owner (ctx_task) of the context per se.
3329 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3330 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3334 * this can be true when not self-monitoring only in UP
3336 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3338 if (can_access_pmu
) ia64_srlz_d();
3340 expert_mode
= pfm_sysctl
.expert_mode
;
3342 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3348 * on both UP and SMP, we can only read the PMD from the hardware register when
3349 * the task is the owner of the local PMU.
3352 for (i
= 0; i
< count
; i
++, req
++) {
3354 cnum
= req
->reg_num
;
3355 reg_flags
= req
->reg_flags
;
3357 if (unlikely(!PMD_IS_IMPL(cnum
))) goto error
;
3359 * we can only read the register that we use. That includes
3360 * the one we explicitely initialize AND the one we want included
3361 * in the sampling buffer (smpl_regs).
3363 * Having this restriction allows optimization in the ctxsw routine
3364 * without compromising security (leaks)
3366 if (unlikely(!CTX_IS_USED_PMD(ctx
, cnum
))) goto error
;
3368 sval
= ctx
->ctx_pmds
[cnum
].val
;
3369 lval
= ctx
->ctx_pmds
[cnum
].lval
;
3370 is_counting
= PMD_IS_COUNTING(cnum
);
3373 * If the task is not the current one, then we check if the
3374 * PMU state is still in the local live register due to lazy ctxsw.
3375 * If true, then we read directly from the registers.
3377 if (can_access_pmu
){
3378 val
= ia64_get_pmd(cnum
);
3381 * context has been saved
3382 * if context is zombie, then task does not exist anymore.
3383 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3385 val
= is_loaded
? ctx
->th_pmds
[cnum
] : 0UL;
3387 rd_func
= pmu_conf
->pmd_desc
[cnum
].read_check
;
3391 * XXX: need to check for overflow when loaded
3398 * execute read checker, if any
3400 if (unlikely(expert_mode
== 0 && rd_func
)) {
3401 unsigned long v
= val
;
3402 ret
= (*rd_func
)(ctx
->ctx_task
, ctx
, cnum
, &v
, regs
);
3403 if (ret
) goto error
;
3408 PFM_REG_RETFLAG_SET(reg_flags
, 0);
3410 DPRINT(("pmd[%u]=0x%lx\n", cnum
, val
));
3413 * update register return value, abort all if problem during copy.
3414 * we only modify the reg_flags field. no check mode is fine because
3415 * access has been verified upfront in sys_perfmonctl().
3417 req
->reg_value
= val
;
3418 req
->reg_flags
= reg_flags
;
3419 req
->reg_last_reset_val
= lval
;
3425 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3430 pfm_mod_write_pmcs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3434 if (req
== NULL
) return -EINVAL
;
3436 ctx
= GET_PMU_CTX();
3438 if (ctx
== NULL
) return -EINVAL
;
3441 * for now limit to current task, which is enough when calling
3442 * from overflow handler
3444 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3446 return pfm_write_pmcs(ctx
, req
, nreq
, regs
);
3448 EXPORT_SYMBOL(pfm_mod_write_pmcs
);
3451 pfm_mod_read_pmds(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3455 if (req
== NULL
) return -EINVAL
;
3457 ctx
= GET_PMU_CTX();
3459 if (ctx
== NULL
) return -EINVAL
;
3462 * for now limit to current task, which is enough when calling
3463 * from overflow handler
3465 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3467 return pfm_read_pmds(ctx
, req
, nreq
, regs
);
3469 EXPORT_SYMBOL(pfm_mod_read_pmds
);
3472 * Only call this function when a process it trying to
3473 * write the debug registers (reading is always allowed)
3476 pfm_use_debug_registers(struct task_struct
*task
)
3478 pfm_context_t
*ctx
= task
->thread
.pfm_context
;
3479 unsigned long flags
;
3482 if (pmu_conf
->use_rr_dbregs
== 0) return 0;
3484 DPRINT(("called for [%d]\n", task
->pid
));
3489 if (task
->thread
.flags
& IA64_THREAD_DBG_VALID
) return 0;
3492 * Even on SMP, we do not need to use an atomic here because
3493 * the only way in is via ptrace() and this is possible only when the
3494 * process is stopped. Even in the case where the ctxsw out is not totally
3495 * completed by the time we come here, there is no way the 'stopped' process
3496 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3497 * So this is always safe.
3499 if (ctx
&& ctx
->ctx_fl_using_dbreg
== 1) return -1;
3504 * We cannot allow setting breakpoints when system wide monitoring
3505 * sessions are using the debug registers.
3507 if (pfm_sessions
.pfs_sys_use_dbregs
> 0)
3510 pfm_sessions
.pfs_ptrace_use_dbregs
++;
3512 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3513 pfm_sessions
.pfs_ptrace_use_dbregs
,
3514 pfm_sessions
.pfs_sys_use_dbregs
,
3523 * This function is called for every task that exits with the
3524 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3525 * able to use the debug registers for debugging purposes via
3526 * ptrace(). Therefore we know it was not using them for
3527 * perfmormance monitoring, so we only decrement the number
3528 * of "ptraced" debug register users to keep the count up to date
3531 pfm_release_debug_registers(struct task_struct
*task
)
3533 unsigned long flags
;
3536 if (pmu_conf
->use_rr_dbregs
== 0) return 0;
3539 if (pfm_sessions
.pfs_ptrace_use_dbregs
== 0) {
3540 printk(KERN_ERR
"perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task
->pid
);
3543 pfm_sessions
.pfs_ptrace_use_dbregs
--;
3552 pfm_restart(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3554 struct task_struct
*task
;
3555 pfm_buffer_fmt_t
*fmt
;
3556 pfm_ovfl_ctrl_t rst_ctrl
;
3557 int state
, is_system
;
3560 state
= ctx
->ctx_state
;
3561 fmt
= ctx
->ctx_buf_fmt
;
3562 is_system
= ctx
->ctx_fl_system
;
3563 task
= PFM_CTX_TASK(ctx
);
3566 case PFM_CTX_MASKED
:
3568 case PFM_CTX_LOADED
:
3569 if (CTX_HAS_SMPL(ctx
) && fmt
->fmt_restart_active
) break;
3571 case PFM_CTX_UNLOADED
:
3572 case PFM_CTX_ZOMBIE
:
3573 DPRINT(("invalid state=%d\n", state
));
3576 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state
));
3581 * In system wide and when the context is loaded, access can only happen
3582 * when the caller is running on the CPU being monitored by the session.
3583 * It does not have to be the owner (ctx_task) of the context per se.
3585 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
3586 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3591 if (unlikely(task
== NULL
)) {
3592 printk(KERN_ERR
"perfmon: [%d] pfm_restart no task\n", current
->pid
);
3596 if (task
== current
|| is_system
) {
3598 fmt
= ctx
->ctx_buf_fmt
;
3600 DPRINT(("restarting self %d ovfl=0x%lx\n",
3602 ctx
->ctx_ovfl_regs
[0]));
3604 if (CTX_HAS_SMPL(ctx
)) {
3606 prefetch(ctx
->ctx_smpl_hdr
);
3608 rst_ctrl
.bits
.mask_monitoring
= 0;
3609 rst_ctrl
.bits
.reset_ovfl_pmds
= 0;
3611 if (state
== PFM_CTX_LOADED
)
3612 ret
= pfm_buf_fmt_restart_active(fmt
, task
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
3614 ret
= pfm_buf_fmt_restart(fmt
, task
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
3616 rst_ctrl
.bits
.mask_monitoring
= 0;
3617 rst_ctrl
.bits
.reset_ovfl_pmds
= 1;
3621 if (rst_ctrl
.bits
.reset_ovfl_pmds
)
3622 pfm_reset_regs(ctx
, ctx
->ctx_ovfl_regs
, PFM_PMD_LONG_RESET
);
3624 if (rst_ctrl
.bits
.mask_monitoring
== 0) {
3625 DPRINT(("resuming monitoring for [%d]\n", task
->pid
));
3627 if (state
== PFM_CTX_MASKED
) pfm_restore_monitoring(task
);
3629 DPRINT(("keeping monitoring stopped for [%d]\n", task
->pid
));
3631 // cannot use pfm_stop_monitoring(task, regs);
3635 * clear overflowed PMD mask to remove any stale information
3637 ctx
->ctx_ovfl_regs
[0] = 0UL;
3640 * back to LOADED state
3642 ctx
->ctx_state
= PFM_CTX_LOADED
;
3645 * XXX: not really useful for self monitoring
3647 ctx
->ctx_fl_can_restart
= 0;
3653 * restart another task
3657 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3658 * one is seen by the task.
3660 if (state
== PFM_CTX_MASKED
) {
3661 if (ctx
->ctx_fl_can_restart
== 0) return -EINVAL
;
3663 * will prevent subsequent restart before this one is
3664 * seen by other task
3666 ctx
->ctx_fl_can_restart
= 0;
3670 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3671 * the task is blocked or on its way to block. That's the normal
3672 * restart path. If the monitoring is not masked, then the task
3673 * can be actively monitoring and we cannot directly intervene.
3674 * Therefore we use the trap mechanism to catch the task and
3675 * force it to reset the buffer/reset PMDs.
3677 * if non-blocking, then we ensure that the task will go into
3678 * pfm_handle_work() before returning to user mode.
3680 * We cannot explicitely reset another task, it MUST always
3681 * be done by the task itself. This works for system wide because
3682 * the tool that is controlling the session is logically doing
3683 * "self-monitoring".
3685 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && state
== PFM_CTX_MASKED
) {
3686 DPRINT(("unblocking [%d] \n", task
->pid
));
3687 complete(&ctx
->ctx_restart_done
);
3689 DPRINT(("[%d] armed exit trap\n", task
->pid
));
3691 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_RESET
;
3693 PFM_SET_WORK_PENDING(task
, 1);
3695 pfm_set_task_notify(task
);
3698 * XXX: send reschedule if task runs on another CPU
3705 pfm_debug(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3707 unsigned int m
= *(unsigned int *)arg
;
3709 pfm_sysctl
.debug
= m
== 0 ? 0 : 1;
3711 printk(KERN_INFO
"perfmon debugging %s (timing reset)\n", pfm_sysctl
.debug
? "on" : "off");
3714 memset(pfm_stats
, 0, sizeof(pfm_stats
));
3715 for(m
=0; m
< NR_CPUS
; m
++) pfm_stats
[m
].pfm_ovfl_intr_cycles_min
= ~0UL;
3721 * arg can be NULL and count can be zero for this function
3724 pfm_write_ibr_dbr(int mode
, pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3726 struct thread_struct
*thread
= NULL
;
3727 struct task_struct
*task
;
3728 pfarg_dbreg_t
*req
= (pfarg_dbreg_t
*)arg
;
3729 unsigned long flags
;
3734 int i
, can_access_pmu
= 0;
3735 int is_system
, is_loaded
;
3737 if (pmu_conf
->use_rr_dbregs
== 0) return -EINVAL
;
3739 state
= ctx
->ctx_state
;
3740 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3741 is_system
= ctx
->ctx_fl_system
;
3742 task
= ctx
->ctx_task
;
3744 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
3747 * on both UP and SMP, we can only write to the PMC when the task is
3748 * the owner of the local PMU.
3751 thread
= &task
->thread
;
3753 * In system wide and when the context is loaded, access can only happen
3754 * when the caller is running on the CPU being monitored by the session.
3755 * It does not have to be the owner (ctx_task) of the context per se.
3757 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3758 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3761 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3765 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3766 * ensuring that no real breakpoint can be installed via this call.
3768 * IMPORTANT: regs can be NULL in this function
3771 first_time
= ctx
->ctx_fl_using_dbreg
== 0;
3774 * don't bother if we are loaded and task is being debugged
3776 if (is_loaded
&& (thread
->flags
& IA64_THREAD_DBG_VALID
) != 0) {
3777 DPRINT(("debug registers already in use for [%d]\n", task
->pid
));
3782 * check for debug registers in system wide mode
3784 * If though a check is done in pfm_context_load(),
3785 * we must repeat it here, in case the registers are
3786 * written after the context is loaded
3791 if (first_time
&& is_system
) {
3792 if (pfm_sessions
.pfs_ptrace_use_dbregs
)
3795 pfm_sessions
.pfs_sys_use_dbregs
++;
3800 if (ret
!= 0) return ret
;
3803 * mark ourself as user of the debug registers for
3806 ctx
->ctx_fl_using_dbreg
= 1;
3809 * clear hardware registers to make sure we don't
3810 * pick up stale state.
3812 * for a system wide session, we do not use
3813 * thread.dbr, thread.ibr because this process
3814 * never leaves the current CPU and the state
3815 * is shared by all processes running on it
3817 if (first_time
&& can_access_pmu
) {
3818 DPRINT(("[%d] clearing ibrs, dbrs\n", task
->pid
));
3819 for (i
=0; i
< pmu_conf
->num_ibrs
; i
++) {
3820 ia64_set_ibr(i
, 0UL);
3821 ia64_dv_serialize_instruction();
3824 for (i
=0; i
< pmu_conf
->num_dbrs
; i
++) {
3825 ia64_set_dbr(i
, 0UL);
3826 ia64_dv_serialize_data();
3832 * Now install the values into the registers
3834 for (i
= 0; i
< count
; i
++, req
++) {
3836 rnum
= req
->dbreg_num
;
3837 dbreg
.val
= req
->dbreg_value
;
3841 if ((mode
== PFM_CODE_RR
&& rnum
>= PFM_NUM_IBRS
) || ((mode
== PFM_DATA_RR
) && rnum
>= PFM_NUM_DBRS
)) {
3842 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3843 rnum
, dbreg
.val
, mode
, i
, count
));
3849 * make sure we do not install enabled breakpoint
3852 if (mode
== PFM_CODE_RR
)
3853 dbreg
.ibr
.ibr_x
= 0;
3855 dbreg
.dbr
.dbr_r
= dbreg
.dbr
.dbr_w
= 0;
3858 PFM_REG_RETFLAG_SET(req
->dbreg_flags
, 0);
3861 * Debug registers, just like PMC, can only be modified
3862 * by a kernel call. Moreover, perfmon() access to those
3863 * registers are centralized in this routine. The hardware
3864 * does not modify the value of these registers, therefore,
3865 * if we save them as they are written, we can avoid having
3866 * to save them on context switch out. This is made possible
3867 * by the fact that when perfmon uses debug registers, ptrace()
3868 * won't be able to modify them concurrently.
3870 if (mode
== PFM_CODE_RR
) {
3871 CTX_USED_IBR(ctx
, rnum
);
3873 if (can_access_pmu
) {
3874 ia64_set_ibr(rnum
, dbreg
.val
);
3875 ia64_dv_serialize_instruction();
3878 ctx
->ctx_ibrs
[rnum
] = dbreg
.val
;
3880 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3881 rnum
, dbreg
.val
, ctx
->ctx_used_ibrs
[0], is_loaded
, can_access_pmu
));
3883 CTX_USED_DBR(ctx
, rnum
);
3885 if (can_access_pmu
) {
3886 ia64_set_dbr(rnum
, dbreg
.val
);
3887 ia64_dv_serialize_data();
3889 ctx
->ctx_dbrs
[rnum
] = dbreg
.val
;
3891 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3892 rnum
, dbreg
.val
, ctx
->ctx_used_dbrs
[0], is_loaded
, can_access_pmu
));
3900 * in case it was our first attempt, we undo the global modifications
3904 if (ctx
->ctx_fl_system
) {
3905 pfm_sessions
.pfs_sys_use_dbregs
--;
3908 ctx
->ctx_fl_using_dbreg
= 0;
3911 * install error return flag
3913 PFM_REG_RETFLAG_SET(req
->dbreg_flags
, PFM_REG_RETFL_EINVAL
);
3919 pfm_write_ibrs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3921 return pfm_write_ibr_dbr(PFM_CODE_RR
, ctx
, arg
, count
, regs
);
3925 pfm_write_dbrs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3927 return pfm_write_ibr_dbr(PFM_DATA_RR
, ctx
, arg
, count
, regs
);
3931 pfm_mod_write_ibrs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3935 if (req
== NULL
) return -EINVAL
;
3937 ctx
= GET_PMU_CTX();
3939 if (ctx
== NULL
) return -EINVAL
;
3942 * for now limit to current task, which is enough when calling
3943 * from overflow handler
3945 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3947 return pfm_write_ibrs(ctx
, req
, nreq
, regs
);
3949 EXPORT_SYMBOL(pfm_mod_write_ibrs
);
3952 pfm_mod_write_dbrs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3956 if (req
== NULL
) return -EINVAL
;
3958 ctx
= GET_PMU_CTX();
3960 if (ctx
== NULL
) return -EINVAL
;
3963 * for now limit to current task, which is enough when calling
3964 * from overflow handler
3966 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3968 return pfm_write_dbrs(ctx
, req
, nreq
, regs
);
3970 EXPORT_SYMBOL(pfm_mod_write_dbrs
);
3974 pfm_get_features(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3976 pfarg_features_t
*req
= (pfarg_features_t
*)arg
;
3978 req
->ft_version
= PFM_VERSION
;
3983 pfm_stop(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3985 struct pt_regs
*tregs
;
3986 struct task_struct
*task
= PFM_CTX_TASK(ctx
);
3987 int state
, is_system
;
3989 state
= ctx
->ctx_state
;
3990 is_system
= ctx
->ctx_fl_system
;
3993 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3995 if (state
== PFM_CTX_UNLOADED
) return -EINVAL
;
3998 * In system wide and when the context is loaded, access can only happen
3999 * when the caller is running on the CPU being monitored by the session.
4000 * It does not have to be the owner (ctx_task) of the context per se.
4002 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
4003 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
4006 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4007 PFM_CTX_TASK(ctx
)->pid
,
4011 * in system mode, we need to update the PMU directly
4012 * and the user level state of the caller, which may not
4013 * necessarily be the creator of the context.
4017 * Update local PMU first
4021 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) & ~IA64_DCR_PP
);
4025 * update local cpuinfo
4027 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP
);
4030 * stop monitoring, does srlz.i
4035 * stop monitoring in the caller
4037 ia64_psr(regs
)->pp
= 0;
4045 if (task
== current
) {
4046 /* stop monitoring at kernel level */
4050 * stop monitoring at the user level
4052 ia64_psr(regs
)->up
= 0;
4054 tregs
= task_pt_regs(task
);
4057 * stop monitoring at the user level
4059 ia64_psr(tregs
)->up
= 0;
4062 * monitoring disabled in kernel at next reschedule
4064 ctx
->ctx_saved_psr_up
= 0;
4065 DPRINT(("task=[%d]\n", task
->pid
));
4072 pfm_start(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4074 struct pt_regs
*tregs
;
4075 int state
, is_system
;
4077 state
= ctx
->ctx_state
;
4078 is_system
= ctx
->ctx_fl_system
;
4080 if (state
!= PFM_CTX_LOADED
) return -EINVAL
;
4083 * In system wide and when the context is loaded, access can only happen
4084 * when the caller is running on the CPU being monitored by the session.
4085 * It does not have to be the owner (ctx_task) of the context per se.
4087 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
4088 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
4093 * in system mode, we need to update the PMU directly
4094 * and the user level state of the caller, which may not
4095 * necessarily be the creator of the context.
4100 * set user level psr.pp for the caller
4102 ia64_psr(regs
)->pp
= 1;
4105 * now update the local PMU and cpuinfo
4107 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP
);
4110 * start monitoring at kernel level
4115 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) | IA64_DCR_PP
);
4125 if (ctx
->ctx_task
== current
) {
4127 /* start monitoring at kernel level */
4131 * activate monitoring at user level
4133 ia64_psr(regs
)->up
= 1;
4136 tregs
= task_pt_regs(ctx
->ctx_task
);
4139 * start monitoring at the kernel level the next
4140 * time the task is scheduled
4142 ctx
->ctx_saved_psr_up
= IA64_PSR_UP
;
4145 * activate monitoring at user level
4147 ia64_psr(tregs
)->up
= 1;
4153 pfm_get_pmc_reset(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4155 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
4160 for (i
= 0; i
< count
; i
++, req
++) {
4162 cnum
= req
->reg_num
;
4164 if (!PMC_IS_IMPL(cnum
)) goto abort_mission
;
4166 req
->reg_value
= PMC_DFL_VAL(cnum
);
4168 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
4170 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum
, req
->reg_value
));
4175 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
4180 pfm_check_task_exist(pfm_context_t
*ctx
)
4182 struct task_struct
*g
, *t
;
4185 read_lock(&tasklist_lock
);
4187 do_each_thread (g
, t
) {
4188 if (t
->thread
.pfm_context
== ctx
) {
4192 } while_each_thread (g
, t
);
4194 read_unlock(&tasklist_lock
);
4196 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret
, ctx
));
4202 pfm_context_load(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4204 struct task_struct
*task
;
4205 struct thread_struct
*thread
;
4206 struct pfm_context_t
*old
;
4207 unsigned long flags
;
4209 struct task_struct
*owner_task
= NULL
;
4211 pfarg_load_t
*req
= (pfarg_load_t
*)arg
;
4212 unsigned long *pmcs_source
, *pmds_source
;
4215 int state
, is_system
, set_dbregs
= 0;
4217 state
= ctx
->ctx_state
;
4218 is_system
= ctx
->ctx_fl_system
;
4220 * can only load from unloaded or terminated state
4222 if (state
!= PFM_CTX_UNLOADED
) {
4223 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4229 DPRINT(("load_pid [%d] using_dbreg=%d\n", req
->load_pid
, ctx
->ctx_fl_using_dbreg
));
4231 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && req
->load_pid
== current
->pid
) {
4232 DPRINT(("cannot use blocking mode on self\n"));
4236 ret
= pfm_get_task(ctx
, req
->load_pid
, &task
);
4238 DPRINT(("load_pid [%d] get_task=%d\n", req
->load_pid
, ret
));
4245 * system wide is self monitoring only
4247 if (is_system
&& task
!= current
) {
4248 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4253 thread
= &task
->thread
;
4257 * cannot load a context which is using range restrictions,
4258 * into a task that is being debugged.
4260 if (ctx
->ctx_fl_using_dbreg
) {
4261 if (thread
->flags
& IA64_THREAD_DBG_VALID
) {
4263 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req
->load_pid
));
4269 if (pfm_sessions
.pfs_ptrace_use_dbregs
) {
4270 DPRINT(("cannot load [%d] dbregs in use\n", task
->pid
));
4273 pfm_sessions
.pfs_sys_use_dbregs
++;
4274 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task
->pid
, pfm_sessions
.pfs_sys_use_dbregs
));
4281 if (ret
) goto error
;
4285 * SMP system-wide monitoring implies self-monitoring.
4287 * The programming model expects the task to
4288 * be pinned on a CPU throughout the session.
4289 * Here we take note of the current CPU at the
4290 * time the context is loaded. No call from
4291 * another CPU will be allowed.
4293 * The pinning via shed_setaffinity()
4294 * must be done by the calling task prior
4297 * systemwide: keep track of CPU this session is supposed to run on
4299 the_cpu
= ctx
->ctx_cpu
= smp_processor_id();
4303 * now reserve the session
4305 ret
= pfm_reserve_session(current
, is_system
, the_cpu
);
4306 if (ret
) goto error
;
4309 * task is necessarily stopped at this point.
4311 * If the previous context was zombie, then it got removed in
4312 * pfm_save_regs(). Therefore we should not see it here.
4313 * If we see a context, then this is an active context
4315 * XXX: needs to be atomic
4317 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4318 thread
->pfm_context
, ctx
));
4321 old
= ia64_cmpxchg(acq
, &thread
->pfm_context
, NULL
, ctx
, sizeof(pfm_context_t
*));
4323 DPRINT(("load_pid [%d] already has a context\n", req
->load_pid
));
4327 pfm_reset_msgq(ctx
);
4329 ctx
->ctx_state
= PFM_CTX_LOADED
;
4332 * link context to task
4334 ctx
->ctx_task
= task
;
4338 * we load as stopped
4340 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE
);
4341 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP
);
4343 if (ctx
->ctx_fl_excl_idle
) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE
);
4345 thread
->flags
|= IA64_THREAD_PM_VALID
;
4349 * propagate into thread-state
4351 pfm_copy_pmds(task
, ctx
);
4352 pfm_copy_pmcs(task
, ctx
);
4354 pmcs_source
= ctx
->th_pmcs
;
4355 pmds_source
= ctx
->th_pmds
;
4358 * always the case for system-wide
4360 if (task
== current
) {
4362 if (is_system
== 0) {
4364 /* allow user level control */
4365 ia64_psr(regs
)->sp
= 0;
4366 DPRINT(("clearing psr.sp for [%d]\n", task
->pid
));
4368 SET_LAST_CPU(ctx
, smp_processor_id());
4370 SET_ACTIVATION(ctx
);
4373 * push the other task out, if any
4375 owner_task
= GET_PMU_OWNER();
4376 if (owner_task
) pfm_lazy_save_regs(owner_task
);
4380 * load all PMD from ctx to PMU (as opposed to thread state)
4381 * restore all PMC from ctx to PMU
4383 pfm_restore_pmds(pmds_source
, ctx
->ctx_all_pmds
[0]);
4384 pfm_restore_pmcs(pmcs_source
, ctx
->ctx_all_pmcs
[0]);
4386 ctx
->ctx_reload_pmcs
[0] = 0UL;
4387 ctx
->ctx_reload_pmds
[0] = 0UL;
4390 * guaranteed safe by earlier check against DBG_VALID
4392 if (ctx
->ctx_fl_using_dbreg
) {
4393 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
4394 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
4399 SET_PMU_OWNER(task
, ctx
);
4401 DPRINT(("context loaded on PMU for [%d]\n", task
->pid
));
4404 * when not current, task MUST be stopped, so this is safe
4406 regs
= task_pt_regs(task
);
4408 /* force a full reload */
4409 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
4410 SET_LAST_CPU(ctx
, -1);
4412 /* initial saved psr (stopped) */
4413 ctx
->ctx_saved_psr_up
= 0UL;
4414 ia64_psr(regs
)->up
= ia64_psr(regs
)->pp
= 0;
4420 if (ret
) pfm_unreserve_session(ctx
, ctx
->ctx_fl_system
, the_cpu
);
4423 * we must undo the dbregs setting (for system-wide)
4425 if (ret
&& set_dbregs
) {
4427 pfm_sessions
.pfs_sys_use_dbregs
--;
4431 * release task, there is now a link with the context
4433 if (is_system
== 0 && task
!= current
) {
4437 ret
= pfm_check_task_exist(ctx
);
4439 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
4440 ctx
->ctx_task
= NULL
;
4448 * in this function, we do not need to increase the use count
4449 * for the task via get_task_struct(), because we hold the
4450 * context lock. If the task were to disappear while having
4451 * a context attached, it would go through pfm_exit_thread()
4452 * which also grabs the context lock and would therefore be blocked
4453 * until we are here.
4455 static void pfm_flush_pmds(struct task_struct
*, pfm_context_t
*ctx
);
4458 pfm_context_unload(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4460 struct task_struct
*task
= PFM_CTX_TASK(ctx
);
4461 struct pt_regs
*tregs
;
4462 int prev_state
, is_system
;
4465 DPRINT(("ctx_state=%d task [%d]\n", ctx
->ctx_state
, task
? task
->pid
: -1));
4467 prev_state
= ctx
->ctx_state
;
4468 is_system
= ctx
->ctx_fl_system
;
4471 * unload only when necessary
4473 if (prev_state
== PFM_CTX_UNLOADED
) {
4474 DPRINT(("ctx_state=%d, nothing to do\n", prev_state
));
4479 * clear psr and dcr bits
4481 ret
= pfm_stop(ctx
, NULL
, 0, regs
);
4482 if (ret
) return ret
;
4484 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
4487 * in system mode, we need to update the PMU directly
4488 * and the user level state of the caller, which may not
4489 * necessarily be the creator of the context.
4496 * local PMU is taken care of in pfm_stop()
4498 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE
);
4499 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE
);
4502 * save PMDs in context
4505 pfm_flush_pmds(current
, ctx
);
4508 * at this point we are done with the PMU
4509 * so we can unreserve the resource.
4511 if (prev_state
!= PFM_CTX_ZOMBIE
)
4512 pfm_unreserve_session(ctx
, 1 , ctx
->ctx_cpu
);
4515 * disconnect context from task
4517 task
->thread
.pfm_context
= NULL
;
4519 * disconnect task from context
4521 ctx
->ctx_task
= NULL
;
4524 * There is nothing more to cleanup here.
4532 tregs
= task
== current
? regs
: task_pt_regs(task
);
4534 if (task
== current
) {
4536 * cancel user level control
4538 ia64_psr(regs
)->sp
= 1;
4540 DPRINT(("setting psr.sp for [%d]\n", task
->pid
));
4543 * save PMDs to context
4546 pfm_flush_pmds(task
, ctx
);
4549 * at this point we are done with the PMU
4550 * so we can unreserve the resource.
4552 * when state was ZOMBIE, we have already unreserved.
4554 if (prev_state
!= PFM_CTX_ZOMBIE
)
4555 pfm_unreserve_session(ctx
, 0 , ctx
->ctx_cpu
);
4558 * reset activation counter and psr
4560 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
4561 SET_LAST_CPU(ctx
, -1);
4564 * PMU state will not be restored
4566 task
->thread
.flags
&= ~IA64_THREAD_PM_VALID
;
4569 * break links between context and task
4571 task
->thread
.pfm_context
= NULL
;
4572 ctx
->ctx_task
= NULL
;
4574 PFM_SET_WORK_PENDING(task
, 0);
4576 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_NONE
;
4577 ctx
->ctx_fl_can_restart
= 0;
4578 ctx
->ctx_fl_going_zombie
= 0;
4580 DPRINT(("disconnected [%d] from context\n", task
->pid
));
4587 * called only from exit_thread(): task == current
4588 * we come here only if current has a context attached (loaded or masked)
4591 pfm_exit_thread(struct task_struct
*task
)
4594 unsigned long flags
;
4595 struct pt_regs
*regs
= task_pt_regs(task
);
4599 ctx
= PFM_GET_CTX(task
);
4601 PROTECT_CTX(ctx
, flags
);
4603 DPRINT(("state=%d task [%d]\n", ctx
->ctx_state
, task
->pid
));
4605 state
= ctx
->ctx_state
;
4607 case PFM_CTX_UNLOADED
:
4609 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4610 * be in unloaded state
4612 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] ctx unloaded\n", task
->pid
);
4614 case PFM_CTX_LOADED
:
4615 case PFM_CTX_MASKED
:
4616 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
4618 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task
->pid
, state
, ret
);
4620 DPRINT(("ctx unloaded for current state was %d\n", state
));
4622 pfm_end_notify_user(ctx
);
4624 case PFM_CTX_ZOMBIE
:
4625 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
4627 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task
->pid
, state
, ret
);
4632 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task
->pid
, state
);
4635 UNPROTECT_CTX(ctx
, flags
);
4637 { u64 psr
= pfm_get_psr();
4638 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
4639 BUG_ON(GET_PMU_OWNER());
4640 BUG_ON(ia64_psr(regs
)->up
);
4641 BUG_ON(ia64_psr(regs
)->pp
);
4645 * All memory free operations (especially for vmalloc'ed memory)
4646 * MUST be done with interrupts ENABLED.
4648 if (free_ok
) pfm_context_free(ctx
);
4652 * functions MUST be listed in the increasing order of their index (see permfon.h)
4654 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4655 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4656 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4657 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4658 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4660 static pfm_cmd_desc_t pfm_cmd_tab
[]={
4661 /* 0 */PFM_CMD_NONE
,
4662 /* 1 */PFM_CMD(pfm_write_pmcs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4663 /* 2 */PFM_CMD(pfm_write_pmds
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4664 /* 3 */PFM_CMD(pfm_read_pmds
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4665 /* 4 */PFM_CMD_S(pfm_stop
, PFM_CMD_PCLRWS
),
4666 /* 5 */PFM_CMD_S(pfm_start
, PFM_CMD_PCLRWS
),
4667 /* 6 */PFM_CMD_NONE
,
4668 /* 7 */PFM_CMD_NONE
,
4669 /* 8 */PFM_CMD(pfm_context_create
, PFM_CMD_ARG_RW
, 1, pfarg_context_t
, pfm_ctx_getsize
),
4670 /* 9 */PFM_CMD_NONE
,
4671 /* 10 */PFM_CMD_S(pfm_restart
, PFM_CMD_PCLRW
),
4672 /* 11 */PFM_CMD_NONE
,
4673 /* 12 */PFM_CMD(pfm_get_features
, PFM_CMD_ARG_RW
, 1, pfarg_features_t
, NULL
),
4674 /* 13 */PFM_CMD(pfm_debug
, 0, 1, unsigned int, NULL
),
4675 /* 14 */PFM_CMD_NONE
,
4676 /* 15 */PFM_CMD(pfm_get_pmc_reset
, PFM_CMD_ARG_RW
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4677 /* 16 */PFM_CMD(pfm_context_load
, PFM_CMD_PCLRWS
, 1, pfarg_load_t
, NULL
),
4678 /* 17 */PFM_CMD_S(pfm_context_unload
, PFM_CMD_PCLRWS
),
4679 /* 18 */PFM_CMD_NONE
,
4680 /* 19 */PFM_CMD_NONE
,
4681 /* 20 */PFM_CMD_NONE
,
4682 /* 21 */PFM_CMD_NONE
,
4683 /* 22 */PFM_CMD_NONE
,
4684 /* 23 */PFM_CMD_NONE
,
4685 /* 24 */PFM_CMD_NONE
,
4686 /* 25 */PFM_CMD_NONE
,
4687 /* 26 */PFM_CMD_NONE
,
4688 /* 27 */PFM_CMD_NONE
,
4689 /* 28 */PFM_CMD_NONE
,
4690 /* 29 */PFM_CMD_NONE
,
4691 /* 30 */PFM_CMD_NONE
,
4692 /* 31 */PFM_CMD_NONE
,
4693 /* 32 */PFM_CMD(pfm_write_ibrs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_dbreg_t
, NULL
),
4694 /* 33 */PFM_CMD(pfm_write_dbrs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_dbreg_t
, NULL
)
4696 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4699 pfm_check_task_state(pfm_context_t
*ctx
, int cmd
, unsigned long flags
)
4701 struct task_struct
*task
;
4702 int state
, old_state
;
4705 state
= ctx
->ctx_state
;
4706 task
= ctx
->ctx_task
;
4709 DPRINT(("context %d no task, state=%d\n", ctx
->ctx_fd
, state
));
4713 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4717 task
->state
, PFM_CMD_STOPPED(cmd
)));
4720 * self-monitoring always ok.
4722 * for system-wide the caller can either be the creator of the
4723 * context (to one to which the context is attached to) OR
4724 * a task running on the same CPU as the session.
4726 if (task
== current
|| ctx
->ctx_fl_system
) return 0;
4729 * we are monitoring another thread
4732 case PFM_CTX_UNLOADED
:
4734 * if context is UNLOADED we are safe to go
4737 case PFM_CTX_ZOMBIE
:
4739 * no command can operate on a zombie context
4741 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd
));
4743 case PFM_CTX_MASKED
:
4745 * PMU state has been saved to software even though
4746 * the thread may still be running.
4748 if (cmd
!= PFM_UNLOAD_CONTEXT
) return 0;
4752 * context is LOADED or MASKED. Some commands may need to have
4755 * We could lift this restriction for UP but it would mean that
4756 * the user has no guarantee the task would not run between
4757 * two successive calls to perfmonctl(). That's probably OK.
4758 * If this user wants to ensure the task does not run, then
4759 * the task must be stopped.
4761 if (PFM_CMD_STOPPED(cmd
)) {
4762 if ((task
->state
!= TASK_STOPPED
) && (task
->state
!= TASK_TRACED
)) {
4763 DPRINT(("[%d] task not in stopped state\n", task
->pid
));
4767 * task is now stopped, wait for ctxsw out
4769 * This is an interesting point in the code.
4770 * We need to unprotect the context because
4771 * the pfm_save_regs() routines needs to grab
4772 * the same lock. There are danger in doing
4773 * this because it leaves a window open for
4774 * another task to get access to the context
4775 * and possibly change its state. The one thing
4776 * that is not possible is for the context to disappear
4777 * because we are protected by the VFS layer, i.e.,
4778 * get_fd()/put_fd().
4782 UNPROTECT_CTX(ctx
, flags
);
4784 wait_task_inactive(task
);
4786 PROTECT_CTX(ctx
, flags
);
4789 * we must recheck to verify if state has changed
4791 if (ctx
->ctx_state
!= old_state
) {
4792 DPRINT(("old_state=%d new_state=%d\n", old_state
, ctx
->ctx_state
));
4800 * system-call entry point (must return long)
4803 sys_perfmonctl (int fd
, int cmd
, void __user
*arg
, int count
)
4805 struct file
*file
= NULL
;
4806 pfm_context_t
*ctx
= NULL
;
4807 unsigned long flags
= 0UL;
4808 void *args_k
= NULL
;
4809 long ret
; /* will expand int return types */
4810 size_t base_sz
, sz
, xtra_sz
= 0;
4811 int narg
, completed_args
= 0, call_made
= 0, cmd_flags
;
4812 int (*func
)(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
4813 int (*getsize
)(void *arg
, size_t *sz
);
4814 #define PFM_MAX_ARGSIZE 4096
4817 * reject any call if perfmon was disabled at initialization
4819 if (unlikely(pmu_conf
== NULL
)) return -ENOSYS
;
4821 if (unlikely(cmd
< 0 || cmd
>= PFM_CMD_COUNT
)) {
4822 DPRINT(("invalid cmd=%d\n", cmd
));
4826 func
= pfm_cmd_tab
[cmd
].cmd_func
;
4827 narg
= pfm_cmd_tab
[cmd
].cmd_narg
;
4828 base_sz
= pfm_cmd_tab
[cmd
].cmd_argsize
;
4829 getsize
= pfm_cmd_tab
[cmd
].cmd_getsize
;
4830 cmd_flags
= pfm_cmd_tab
[cmd
].cmd_flags
;
4832 if (unlikely(func
== NULL
)) {
4833 DPRINT(("invalid cmd=%d\n", cmd
));
4837 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4845 * check if number of arguments matches what the command expects
4847 if (unlikely((narg
== PFM_CMD_ARG_MANY
&& count
<= 0) || (narg
> 0 && narg
!= count
)))
4851 sz
= xtra_sz
+ base_sz
*count
;
4853 * limit abuse to min page size
4855 if (unlikely(sz
> PFM_MAX_ARGSIZE
)) {
4856 printk(KERN_ERR
"perfmon: [%d] argument too big %lu\n", current
->pid
, sz
);
4861 * allocate default-sized argument buffer
4863 if (likely(count
&& args_k
== NULL
)) {
4864 args_k
= kmalloc(PFM_MAX_ARGSIZE
, GFP_KERNEL
);
4865 if (args_k
== NULL
) return -ENOMEM
;
4873 * assume sz = 0 for command without parameters
4875 if (sz
&& copy_from_user(args_k
, arg
, sz
)) {
4876 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz
, arg
));
4881 * check if command supports extra parameters
4883 if (completed_args
== 0 && getsize
) {
4885 * get extra parameters size (based on main argument)
4887 ret
= (*getsize
)(args_k
, &xtra_sz
);
4888 if (ret
) goto error_args
;
4892 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz
, xtra_sz
));
4894 /* retry if necessary */
4895 if (likely(xtra_sz
)) goto restart_args
;
4898 if (unlikely((cmd_flags
& PFM_CMD_FD
) == 0)) goto skip_fd
;
4903 if (unlikely(file
== NULL
)) {
4904 DPRINT(("invalid fd %d\n", fd
));
4907 if (unlikely(PFM_IS_FILE(file
) == 0)) {
4908 DPRINT(("fd %d not related to perfmon\n", fd
));
4912 ctx
= (pfm_context_t
*)file
->private_data
;
4913 if (unlikely(ctx
== NULL
)) {
4914 DPRINT(("no context for fd %d\n", fd
));
4917 prefetch(&ctx
->ctx_state
);
4919 PROTECT_CTX(ctx
, flags
);
4922 * check task is stopped
4924 ret
= pfm_check_task_state(ctx
, cmd
, flags
);
4925 if (unlikely(ret
)) goto abort_locked
;
4928 ret
= (*func
)(ctx
, args_k
, count
, task_pt_regs(current
));
4934 DPRINT(("context unlocked\n"));
4935 UNPROTECT_CTX(ctx
, flags
);
4938 /* copy argument back to user, if needed */
4939 if (call_made
&& PFM_CMD_RW_ARG(cmd
) && copy_to_user(arg
, args_k
, base_sz
*count
)) ret
= -EFAULT
;
4947 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd
), ret
));
4953 pfm_resume_after_ovfl(pfm_context_t
*ctx
, unsigned long ovfl_regs
, struct pt_regs
*regs
)
4955 pfm_buffer_fmt_t
*fmt
= ctx
->ctx_buf_fmt
;
4956 pfm_ovfl_ctrl_t rst_ctrl
;
4960 state
= ctx
->ctx_state
;
4962 * Unlock sampling buffer and reset index atomically
4963 * XXX: not really needed when blocking
4965 if (CTX_HAS_SMPL(ctx
)) {
4967 rst_ctrl
.bits
.mask_monitoring
= 0;
4968 rst_ctrl
.bits
.reset_ovfl_pmds
= 0;
4970 if (state
== PFM_CTX_LOADED
)
4971 ret
= pfm_buf_fmt_restart_active(fmt
, current
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
4973 ret
= pfm_buf_fmt_restart(fmt
, current
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
4975 rst_ctrl
.bits
.mask_monitoring
= 0;
4976 rst_ctrl
.bits
.reset_ovfl_pmds
= 1;
4980 if (rst_ctrl
.bits
.reset_ovfl_pmds
) {
4981 pfm_reset_regs(ctx
, &ovfl_regs
, PFM_PMD_LONG_RESET
);
4983 if (rst_ctrl
.bits
.mask_monitoring
== 0) {
4984 DPRINT(("resuming monitoring\n"));
4985 if (ctx
->ctx_state
== PFM_CTX_MASKED
) pfm_restore_monitoring(current
);
4987 DPRINT(("stopping monitoring\n"));
4988 //pfm_stop_monitoring(current, regs);
4990 ctx
->ctx_state
= PFM_CTX_LOADED
;
4995 * context MUST BE LOCKED when calling
4996 * can only be called for current
4999 pfm_context_force_terminate(pfm_context_t
*ctx
, struct pt_regs
*regs
)
5003 DPRINT(("entering for [%d]\n", current
->pid
));
5005 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
5007 printk(KERN_ERR
"pfm_context_force_terminate: [%d] unloaded failed with %d\n", current
->pid
, ret
);
5011 * and wakeup controlling task, indicating we are now disconnected
5013 wake_up_interruptible(&ctx
->ctx_zombieq
);
5016 * given that context is still locked, the controlling
5017 * task will only get access when we return from
5018 * pfm_handle_work().
5022 static int pfm_ovfl_notify_user(pfm_context_t
*ctx
, unsigned long ovfl_pmds
);
5024 * pfm_handle_work() can be called with interrupts enabled
5025 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5026 * call may sleep, therefore we must re-enable interrupts
5027 * to avoid deadlocks. It is safe to do so because this function
5028 * is called ONLY when returning to user level (PUStk=1), in which case
5029 * there is no risk of kernel stack overflow due to deep
5030 * interrupt nesting.
5033 pfm_handle_work(void)
5036 struct pt_regs
*regs
;
5037 unsigned long flags
, dummy_flags
;
5038 unsigned long ovfl_regs
;
5039 unsigned int reason
;
5042 ctx
= PFM_GET_CTX(current
);
5044 printk(KERN_ERR
"perfmon: [%d] has no PFM context\n", current
->pid
);
5048 PROTECT_CTX(ctx
, flags
);
5050 PFM_SET_WORK_PENDING(current
, 0);
5052 pfm_clear_task_notify();
5054 regs
= task_pt_regs(current
);
5057 * extract reason for being here and clear
5059 reason
= ctx
->ctx_fl_trap_reason
;
5060 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_NONE
;
5061 ovfl_regs
= ctx
->ctx_ovfl_regs
[0];
5063 DPRINT(("reason=%d state=%d\n", reason
, ctx
->ctx_state
));
5066 * must be done before we check for simple-reset mode
5068 if (ctx
->ctx_fl_going_zombie
|| ctx
->ctx_state
== PFM_CTX_ZOMBIE
) goto do_zombie
;
5071 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5072 if (reason
== PFM_TRAP_REASON_RESET
) goto skip_blocking
;
5075 * restore interrupt mask to what it was on entry.
5076 * Could be enabled/diasbled.
5078 UNPROTECT_CTX(ctx
, flags
);
5081 * force interrupt enable because of down_interruptible()
5085 DPRINT(("before block sleeping\n"));
5088 * may go through without blocking on SMP systems
5089 * if restart has been received already by the time we call down()
5091 ret
= wait_for_completion_interruptible(&ctx
->ctx_restart_done
);
5093 DPRINT(("after block sleeping ret=%d\n", ret
));
5096 * lock context and mask interrupts again
5097 * We save flags into a dummy because we may have
5098 * altered interrupts mask compared to entry in this
5101 PROTECT_CTX(ctx
, dummy_flags
);
5104 * we need to read the ovfl_regs only after wake-up
5105 * because we may have had pfm_write_pmds() in between
5106 * and that can changed PMD values and therefore
5107 * ovfl_regs is reset for these new PMD values.
5109 ovfl_regs
= ctx
->ctx_ovfl_regs
[0];
5111 if (ctx
->ctx_fl_going_zombie
) {
5113 DPRINT(("context is zombie, bailing out\n"));
5114 pfm_context_force_terminate(ctx
, regs
);
5118 * in case of interruption of down() we don't restart anything
5120 if (ret
< 0) goto nothing_to_do
;
5123 pfm_resume_after_ovfl(ctx
, ovfl_regs
, regs
);
5124 ctx
->ctx_ovfl_regs
[0] = 0UL;
5128 * restore flags as they were upon entry
5130 UNPROTECT_CTX(ctx
, flags
);
5134 pfm_notify_user(pfm_context_t
*ctx
, pfm_msg_t
*msg
)
5136 if (ctx
->ctx_state
== PFM_CTX_ZOMBIE
) {
5137 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5141 DPRINT(("waking up somebody\n"));
5143 if (msg
) wake_up_interruptible(&ctx
->ctx_msgq_wait
);
5146 * safe, we are not in intr handler, nor in ctxsw when
5149 kill_fasync (&ctx
->ctx_async_queue
, SIGIO
, POLL_IN
);
5155 pfm_ovfl_notify_user(pfm_context_t
*ctx
, unsigned long ovfl_pmds
)
5157 pfm_msg_t
*msg
= NULL
;
5159 if (ctx
->ctx_fl_no_msg
== 0) {
5160 msg
= pfm_get_new_msg(ctx
);
5162 printk(KERN_ERR
"perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5166 msg
->pfm_ovfl_msg
.msg_type
= PFM_MSG_OVFL
;
5167 msg
->pfm_ovfl_msg
.msg_ctx_fd
= ctx
->ctx_fd
;
5168 msg
->pfm_ovfl_msg
.msg_active_set
= 0;
5169 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[0] = ovfl_pmds
;
5170 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[1] = 0UL;
5171 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[2] = 0UL;
5172 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[3] = 0UL;
5173 msg
->pfm_ovfl_msg
.msg_tstamp
= 0UL;
5176 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5182 return pfm_notify_user(ctx
, msg
);
5186 pfm_end_notify_user(pfm_context_t
*ctx
)
5190 msg
= pfm_get_new_msg(ctx
);
5192 printk(KERN_ERR
"perfmon: pfm_end_notify_user no more notification msgs\n");
5196 memset(msg
, 0, sizeof(*msg
));
5198 msg
->pfm_end_msg
.msg_type
= PFM_MSG_END
;
5199 msg
->pfm_end_msg
.msg_ctx_fd
= ctx
->ctx_fd
;
5200 msg
->pfm_ovfl_msg
.msg_tstamp
= 0UL;
5202 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5207 return pfm_notify_user(ctx
, msg
);
5211 * main overflow processing routine.
5212 * it can be called from the interrupt path or explicitely during the context switch code
5215 pfm_overflow_handler(struct task_struct
*task
, pfm_context_t
*ctx
, u64 pmc0
, struct pt_regs
*regs
)
5217 pfm_ovfl_arg_t
*ovfl_arg
;
5219 unsigned long old_val
, ovfl_val
, new_val
;
5220 unsigned long ovfl_notify
= 0UL, ovfl_pmds
= 0UL, smpl_pmds
= 0UL, reset_pmds
;
5221 unsigned long tstamp
;
5222 pfm_ovfl_ctrl_t ovfl_ctrl
;
5223 unsigned int i
, has_smpl
;
5224 int must_notify
= 0;
5226 if (unlikely(ctx
->ctx_state
== PFM_CTX_ZOMBIE
)) goto stop_monitoring
;
5229 * sanity test. Should never happen
5231 if (unlikely((pmc0
& 0x1) == 0)) goto sanity_check
;
5233 tstamp
= ia64_get_itc();
5234 mask
= pmc0
>> PMU_FIRST_COUNTER
;
5235 ovfl_val
= pmu_conf
->ovfl_val
;
5236 has_smpl
= CTX_HAS_SMPL(ctx
);
5238 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5239 "used_pmds=0x%lx\n",
5241 task
? task
->pid
: -1,
5242 (regs
? regs
->cr_iip
: 0),
5243 CTX_OVFL_NOBLOCK(ctx
) ? "nonblocking" : "blocking",
5244 ctx
->ctx_used_pmds
[0]));
5248 * first we update the virtual counters
5249 * assume there was a prior ia64_srlz_d() issued
5251 for (i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
5253 /* skip pmd which did not overflow */
5254 if ((mask
& 0x1) == 0) continue;
5257 * Note that the pmd is not necessarily 0 at this point as qualified events
5258 * may have happened before the PMU was frozen. The residual count is not
5259 * taken into consideration here but will be with any read of the pmd via
5262 old_val
= new_val
= ctx
->ctx_pmds
[i
].val
;
5263 new_val
+= 1 + ovfl_val
;
5264 ctx
->ctx_pmds
[i
].val
= new_val
;
5267 * check for overflow condition
5269 if (likely(old_val
> new_val
)) {
5270 ovfl_pmds
|= 1UL << i
;
5271 if (PMC_OVFL_NOTIFY(ctx
, i
)) ovfl_notify
|= 1UL << i
;
5274 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5278 ia64_get_pmd(i
) & ovfl_val
,
5284 * there was no 64-bit overflow, nothing else to do
5286 if (ovfl_pmds
== 0UL) return;
5289 * reset all control bits
5295 * if a sampling format module exists, then we "cache" the overflow by
5296 * calling the module's handler() routine.
5299 unsigned long start_cycles
, end_cycles
;
5300 unsigned long pmd_mask
;
5302 int this_cpu
= smp_processor_id();
5304 pmd_mask
= ovfl_pmds
>> PMU_FIRST_COUNTER
;
5305 ovfl_arg
= &ctx
->ctx_ovfl_arg
;
5307 prefetch(ctx
->ctx_smpl_hdr
);
5309 for(i
=PMU_FIRST_COUNTER
; pmd_mask
&& ret
== 0; i
++, pmd_mask
>>=1) {
5313 if ((pmd_mask
& 0x1) == 0) continue;
5315 ovfl_arg
->ovfl_pmd
= (unsigned char )i
;
5316 ovfl_arg
->ovfl_notify
= ovfl_notify
& mask
? 1 : 0;
5317 ovfl_arg
->active_set
= 0;
5318 ovfl_arg
->ovfl_ctrl
.val
= 0; /* module must fill in all fields */
5319 ovfl_arg
->smpl_pmds
[0] = smpl_pmds
= ctx
->ctx_pmds
[i
].smpl_pmds
[0];
5321 ovfl_arg
->pmd_value
= ctx
->ctx_pmds
[i
].val
;
5322 ovfl_arg
->pmd_last_reset
= ctx
->ctx_pmds
[i
].lval
;
5323 ovfl_arg
->pmd_eventid
= ctx
->ctx_pmds
[i
].eventid
;
5326 * copy values of pmds of interest. Sampling format may copy them
5327 * into sampling buffer.
5330 for(j
=0, k
=0; smpl_pmds
; j
++, smpl_pmds
>>=1) {
5331 if ((smpl_pmds
& 0x1) == 0) continue;
5332 ovfl_arg
->smpl_pmds_values
[k
++] = PMD_IS_COUNTING(j
) ? pfm_read_soft_counter(ctx
, j
) : ia64_get_pmd(j
);
5333 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k
-1, j
, ovfl_arg
->smpl_pmds_values
[k
-1]));
5337 pfm_stats
[this_cpu
].pfm_smpl_handler_calls
++;
5339 start_cycles
= ia64_get_itc();
5342 * call custom buffer format record (handler) routine
5344 ret
= (*ctx
->ctx_buf_fmt
->fmt_handler
)(task
, ctx
->ctx_smpl_hdr
, ovfl_arg
, regs
, tstamp
);
5346 end_cycles
= ia64_get_itc();
5349 * For those controls, we take the union because they have
5350 * an all or nothing behavior.
5352 ovfl_ctrl
.bits
.notify_user
|= ovfl_arg
->ovfl_ctrl
.bits
.notify_user
;
5353 ovfl_ctrl
.bits
.block_task
|= ovfl_arg
->ovfl_ctrl
.bits
.block_task
;
5354 ovfl_ctrl
.bits
.mask_monitoring
|= ovfl_arg
->ovfl_ctrl
.bits
.mask_monitoring
;
5356 * build the bitmask of pmds to reset now
5358 if (ovfl_arg
->ovfl_ctrl
.bits
.reset_ovfl_pmds
) reset_pmds
|= mask
;
5360 pfm_stats
[this_cpu
].pfm_smpl_handler_cycles
+= end_cycles
- start_cycles
;
5363 * when the module cannot handle the rest of the overflows, we abort right here
5365 if (ret
&& pmd_mask
) {
5366 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5367 pmd_mask
<<PMU_FIRST_COUNTER
));
5370 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5372 ovfl_pmds
&= ~reset_pmds
;
5375 * when no sampling module is used, then the default
5376 * is to notify on overflow if requested by user
5378 ovfl_ctrl
.bits
.notify_user
= ovfl_notify
? 1 : 0;
5379 ovfl_ctrl
.bits
.block_task
= ovfl_notify
? 1 : 0;
5380 ovfl_ctrl
.bits
.mask_monitoring
= ovfl_notify
? 1 : 0; /* XXX: change for saturation */
5381 ovfl_ctrl
.bits
.reset_ovfl_pmds
= ovfl_notify
? 0 : 1;
5383 * if needed, we reset all overflowed pmds
5385 if (ovfl_notify
== 0) reset_pmds
= ovfl_pmds
;
5388 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds
, reset_pmds
));
5391 * reset the requested PMD registers using the short reset values
5394 unsigned long bm
= reset_pmds
;
5395 pfm_reset_regs(ctx
, &bm
, PFM_PMD_SHORT_RESET
);
5398 if (ovfl_notify
&& ovfl_ctrl
.bits
.notify_user
) {
5400 * keep track of what to reset when unblocking
5402 ctx
->ctx_ovfl_regs
[0] = ovfl_pmds
;
5405 * check for blocking context
5407 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && ovfl_ctrl
.bits
.block_task
) {
5409 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_BLOCK
;
5412 * set the perfmon specific checking pending work for the task
5414 PFM_SET_WORK_PENDING(task
, 1);
5417 * when coming from ctxsw, current still points to the
5418 * previous task, therefore we must work with task and not current.
5420 pfm_set_task_notify(task
);
5423 * defer until state is changed (shorten spin window). the context is locked
5424 * anyway, so the signal receiver would come spin for nothing.
5429 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5430 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid
: -1,
5431 PFM_GET_WORK_PENDING(task
),
5432 ctx
->ctx_fl_trap_reason
,
5435 ovfl_ctrl
.bits
.mask_monitoring
? 1 : 0));
5437 * in case monitoring must be stopped, we toggle the psr bits
5439 if (ovfl_ctrl
.bits
.mask_monitoring
) {
5440 pfm_mask_monitoring(task
);
5441 ctx
->ctx_state
= PFM_CTX_MASKED
;
5442 ctx
->ctx_fl_can_restart
= 1;
5446 * send notification now
5448 if (must_notify
) pfm_ovfl_notify_user(ctx
, ovfl_notify
);
5453 printk(KERN_ERR
"perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5455 task
? task
->pid
: -1,
5461 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5462 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5463 * come here as zombie only if the task is the current task. In which case, we
5464 * can access the PMU hardware directly.
5466 * Note that zombies do have PM_VALID set. So here we do the minimal.
5468 * In case the context was zombified it could not be reclaimed at the time
5469 * the monitoring program exited. At this point, the PMU reservation has been
5470 * returned, the sampiing buffer has been freed. We must convert this call
5471 * into a spurious interrupt. However, we must also avoid infinite overflows
5472 * by stopping monitoring for this task. We can only come here for a per-task
5473 * context. All we need to do is to stop monitoring using the psr bits which
5474 * are always task private. By re-enabling secure montioring, we ensure that
5475 * the monitored task will not be able to re-activate monitoring.
5476 * The task will eventually be context switched out, at which point the context
5477 * will be reclaimed (that includes releasing ownership of the PMU).
5479 * So there might be a window of time where the number of per-task session is zero
5480 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5481 * context. This is safe because if a per-task session comes in, it will push this one
5482 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5483 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5484 * also push our zombie context out.
5486 * Overall pretty hairy stuff....
5488 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task
? task
->pid
: -1));
5490 ia64_psr(regs
)->up
= 0;
5491 ia64_psr(regs
)->sp
= 1;
5496 pfm_do_interrupt_handler(int irq
, void *arg
, struct pt_regs
*regs
)
5498 struct task_struct
*task
;
5500 unsigned long flags
;
5502 int this_cpu
= smp_processor_id();
5505 pfm_stats
[this_cpu
].pfm_ovfl_intr_count
++;
5508 * srlz.d done before arriving here
5510 pmc0
= ia64_get_pmc(0);
5512 task
= GET_PMU_OWNER();
5513 ctx
= GET_PMU_CTX();
5516 * if we have some pending bits set
5517 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5519 if (PMC0_HAS_OVFL(pmc0
) && task
) {
5521 * we assume that pmc0.fr is always set here
5525 if (!ctx
) goto report_spurious1
;
5527 if (ctx
->ctx_fl_system
== 0 && (task
->thread
.flags
& IA64_THREAD_PM_VALID
) == 0)
5528 goto report_spurious2
;
5530 PROTECT_CTX_NOPRINT(ctx
, flags
);
5532 pfm_overflow_handler(task
, ctx
, pmc0
, regs
);
5534 UNPROTECT_CTX_NOPRINT(ctx
, flags
);
5537 pfm_stats
[this_cpu
].pfm_spurious_ovfl_intr_count
++;
5541 * keep it unfrozen at all times
5548 printk(KERN_INFO
"perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5549 this_cpu
, task
->pid
);
5553 printk(KERN_INFO
"perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5561 pfm_interrupt_handler(int irq
, void *arg
)
5563 unsigned long start_cycles
, total_cycles
;
5564 unsigned long min
, max
;
5567 struct pt_regs
*regs
= get_irq_regs();
5569 this_cpu
= get_cpu();
5570 if (likely(!pfm_alt_intr_handler
)) {
5571 min
= pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_min
;
5572 max
= pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_max
;
5574 start_cycles
= ia64_get_itc();
5576 ret
= pfm_do_interrupt_handler(irq
, arg
, regs
);
5578 total_cycles
= ia64_get_itc();
5581 * don't measure spurious interrupts
5583 if (likely(ret
== 0)) {
5584 total_cycles
-= start_cycles
;
5586 if (total_cycles
< min
) pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_min
= total_cycles
;
5587 if (total_cycles
> max
) pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_max
= total_cycles
;
5589 pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles
+= total_cycles
;
5593 (*pfm_alt_intr_handler
->handler
)(irq
, arg
, regs
);
5596 put_cpu_no_resched();
5601 * /proc/perfmon interface, for debug only
5604 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5607 pfm_proc_start(struct seq_file
*m
, loff_t
*pos
)
5610 return PFM_PROC_SHOW_HEADER
;
5613 while (*pos
<= NR_CPUS
) {
5614 if (cpu_online(*pos
- 1)) {
5615 return (void *)*pos
;
5623 pfm_proc_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
5626 return pfm_proc_start(m
, pos
);
5630 pfm_proc_stop(struct seq_file
*m
, void *v
)
5635 pfm_proc_show_header(struct seq_file
*m
)
5637 struct list_head
* pos
;
5638 pfm_buffer_fmt_t
* entry
;
5639 unsigned long flags
;
5642 "perfmon version : %u.%u\n"
5645 "expert mode : %s\n"
5646 "ovfl_mask : 0x%lx\n"
5647 "PMU flags : 0x%x\n",
5648 PFM_VERSION_MAJ
, PFM_VERSION_MIN
,
5650 pfm_sysctl
.fastctxsw
> 0 ? "Yes": "No",
5651 pfm_sysctl
.expert_mode
> 0 ? "Yes": "No",
5658 "proc_sessions : %u\n"
5659 "sys_sessions : %u\n"
5660 "sys_use_dbregs : %u\n"
5661 "ptrace_use_dbregs : %u\n",
5662 pfm_sessions
.pfs_task_sessions
,
5663 pfm_sessions
.pfs_sys_sessions
,
5664 pfm_sessions
.pfs_sys_use_dbregs
,
5665 pfm_sessions
.pfs_ptrace_use_dbregs
);
5669 spin_lock(&pfm_buffer_fmt_lock
);
5671 list_for_each(pos
, &pfm_buffer_fmt_list
) {
5672 entry
= list_entry(pos
, pfm_buffer_fmt_t
, fmt_list
);
5673 seq_printf(m
, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5684 entry
->fmt_uuid
[10],
5685 entry
->fmt_uuid
[11],
5686 entry
->fmt_uuid
[12],
5687 entry
->fmt_uuid
[13],
5688 entry
->fmt_uuid
[14],
5689 entry
->fmt_uuid
[15],
5692 spin_unlock(&pfm_buffer_fmt_lock
);
5697 pfm_proc_show(struct seq_file
*m
, void *v
)
5703 if (v
== PFM_PROC_SHOW_HEADER
) {
5704 pfm_proc_show_header(m
);
5708 /* show info for CPU (v - 1) */
5712 "CPU%-2d overflow intrs : %lu\n"
5713 "CPU%-2d overflow cycles : %lu\n"
5714 "CPU%-2d overflow min : %lu\n"
5715 "CPU%-2d overflow max : %lu\n"
5716 "CPU%-2d smpl handler calls : %lu\n"
5717 "CPU%-2d smpl handler cycles : %lu\n"
5718 "CPU%-2d spurious intrs : %lu\n"
5719 "CPU%-2d replay intrs : %lu\n"
5720 "CPU%-2d syst_wide : %d\n"
5721 "CPU%-2d dcr_pp : %d\n"
5722 "CPU%-2d exclude idle : %d\n"
5723 "CPU%-2d owner : %d\n"
5724 "CPU%-2d context : %p\n"
5725 "CPU%-2d activations : %lu\n",
5726 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_count
,
5727 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles
,
5728 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles_min
,
5729 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles_max
,
5730 cpu
, pfm_stats
[cpu
].pfm_smpl_handler_calls
,
5731 cpu
, pfm_stats
[cpu
].pfm_smpl_handler_cycles
,
5732 cpu
, pfm_stats
[cpu
].pfm_spurious_ovfl_intr_count
,
5733 cpu
, pfm_stats
[cpu
].pfm_replay_ovfl_intr_count
,
5734 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_SYST_WIDE
? 1 : 0,
5735 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_DCR_PP
? 1 : 0,
5736 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_EXCL_IDLE
? 1 : 0,
5737 cpu
, pfm_get_cpu_data(pmu_owner
, cpu
) ? pfm_get_cpu_data(pmu_owner
, cpu
)->pid
: -1,
5738 cpu
, pfm_get_cpu_data(pmu_ctx
, cpu
),
5739 cpu
, pfm_get_cpu_data(pmu_activation_number
, cpu
));
5741 if (num_online_cpus() == 1 && pfm_sysctl
.debug
> 0) {
5743 psr
= pfm_get_psr();
5748 "CPU%-2d psr : 0x%lx\n"
5749 "CPU%-2d pmc0 : 0x%lx\n",
5751 cpu
, ia64_get_pmc(0));
5753 for (i
=0; PMC_IS_LAST(i
) == 0; i
++) {
5754 if (PMC_IS_COUNTING(i
) == 0) continue;
5756 "CPU%-2d pmc%u : 0x%lx\n"
5757 "CPU%-2d pmd%u : 0x%lx\n",
5758 cpu
, i
, ia64_get_pmc(i
),
5759 cpu
, i
, ia64_get_pmd(i
));
5765 struct seq_operations pfm_seq_ops
= {
5766 .start
= pfm_proc_start
,
5767 .next
= pfm_proc_next
,
5768 .stop
= pfm_proc_stop
,
5769 .show
= pfm_proc_show
5773 pfm_proc_open(struct inode
*inode
, struct file
*file
)
5775 return seq_open(file
, &pfm_seq_ops
);
5780 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5781 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5782 * is active or inactive based on mode. We must rely on the value in
5783 * local_cpu_data->pfm_syst_info
5786 pfm_syst_wide_update_task(struct task_struct
*task
, unsigned long info
, int is_ctxswin
)
5788 struct pt_regs
*regs
;
5790 unsigned long dcr_pp
;
5792 dcr_pp
= info
& PFM_CPUINFO_DCR_PP
? 1 : 0;
5795 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5796 * on every CPU, so we can rely on the pid to identify the idle task.
5798 if ((info
& PFM_CPUINFO_EXCL_IDLE
) == 0 || task
->pid
) {
5799 regs
= task_pt_regs(task
);
5800 ia64_psr(regs
)->pp
= is_ctxswin
? dcr_pp
: 0;
5804 * if monitoring has started
5807 dcr
= ia64_getreg(_IA64_REG_CR_DCR
);
5809 * context switching in?
5812 /* mask monitoring for the idle task */
5813 ia64_setreg(_IA64_REG_CR_DCR
, dcr
& ~IA64_DCR_PP
);
5819 * context switching out
5820 * restore monitoring for next task
5822 * Due to inlining this odd if-then-else construction generates
5825 ia64_setreg(_IA64_REG_CR_DCR
, dcr
|IA64_DCR_PP
);
5834 pfm_force_cleanup(pfm_context_t
*ctx
, struct pt_regs
*regs
)
5836 struct task_struct
*task
= ctx
->ctx_task
;
5838 ia64_psr(regs
)->up
= 0;
5839 ia64_psr(regs
)->sp
= 1;
5841 if (GET_PMU_OWNER() == task
) {
5842 DPRINT(("cleared ownership for [%d]\n", ctx
->ctx_task
->pid
));
5843 SET_PMU_OWNER(NULL
, NULL
);
5847 * disconnect the task from the context and vice-versa
5849 PFM_SET_WORK_PENDING(task
, 0);
5851 task
->thread
.pfm_context
= NULL
;
5852 task
->thread
.flags
&= ~IA64_THREAD_PM_VALID
;
5854 DPRINT(("force cleanup for [%d]\n", task
->pid
));
5859 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5862 pfm_save_regs(struct task_struct
*task
)
5865 unsigned long flags
;
5869 ctx
= PFM_GET_CTX(task
);
5870 if (ctx
== NULL
) return;
5873 * we always come here with interrupts ALREADY disabled by
5874 * the scheduler. So we simply need to protect against concurrent
5875 * access, not CPU concurrency.
5877 flags
= pfm_protect_ctx_ctxsw(ctx
);
5879 if (ctx
->ctx_state
== PFM_CTX_ZOMBIE
) {
5880 struct pt_regs
*regs
= task_pt_regs(task
);
5884 pfm_force_cleanup(ctx
, regs
);
5886 BUG_ON(ctx
->ctx_smpl_hdr
);
5888 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
5890 pfm_context_free(ctx
);
5895 * save current PSR: needed because we modify it
5898 psr
= pfm_get_psr();
5900 BUG_ON(psr
& (IA64_PSR_I
));
5904 * This is the last instruction which may generate an overflow
5906 * We do not need to set psr.sp because, it is irrelevant in kernel.
5907 * It will be restored from ipsr when going back to user level
5912 * keep a copy of psr.up (for reload)
5914 ctx
->ctx_saved_psr_up
= psr
& IA64_PSR_UP
;
5917 * release ownership of this PMU.
5918 * PM interrupts are masked, so nothing
5921 SET_PMU_OWNER(NULL
, NULL
);
5924 * we systematically save the PMD as we have no
5925 * guarantee we will be schedule at that same
5928 pfm_save_pmds(ctx
->th_pmds
, ctx
->ctx_used_pmds
[0]);
5931 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5932 * we will need it on the restore path to check
5933 * for pending overflow.
5935 ctx
->th_pmcs
[0] = ia64_get_pmc(0);
5938 * unfreeze PMU if had pending overflows
5940 if (ctx
->th_pmcs
[0] & ~0x1UL
) pfm_unfreeze_pmu();
5943 * finally, allow context access.
5944 * interrupts will still be masked after this call.
5946 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
5949 #else /* !CONFIG_SMP */
5951 pfm_save_regs(struct task_struct
*task
)
5956 ctx
= PFM_GET_CTX(task
);
5957 if (ctx
== NULL
) return;
5960 * save current PSR: needed because we modify it
5962 psr
= pfm_get_psr();
5964 BUG_ON(psr
& (IA64_PSR_I
));
5968 * This is the last instruction which may generate an overflow
5970 * We do not need to set psr.sp because, it is irrelevant in kernel.
5971 * It will be restored from ipsr when going back to user level
5976 * keep a copy of psr.up (for reload)
5978 ctx
->ctx_saved_psr_up
= psr
& IA64_PSR_UP
;
5982 pfm_lazy_save_regs (struct task_struct
*task
)
5985 unsigned long flags
;
5987 { u64 psr
= pfm_get_psr();
5988 BUG_ON(psr
& IA64_PSR_UP
);
5991 ctx
= PFM_GET_CTX(task
);
5994 * we need to mask PMU overflow here to
5995 * make sure that we maintain pmc0 until
5996 * we save it. overflow interrupts are
5997 * treated as spurious if there is no
6000 * XXX: I don't think this is necessary
6002 PROTECT_CTX(ctx
,flags
);
6005 * release ownership of this PMU.
6006 * must be done before we save the registers.
6008 * after this call any PMU interrupt is treated
6011 SET_PMU_OWNER(NULL
, NULL
);
6014 * save all the pmds we use
6016 pfm_save_pmds(ctx
->th_pmds
, ctx
->ctx_used_pmds
[0]);
6019 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6020 * it is needed to check for pended overflow
6021 * on the restore path
6023 ctx
->th_pmcs
[0] = ia64_get_pmc(0);
6026 * unfreeze PMU if had pending overflows
6028 if (ctx
->th_pmcs
[0] & ~0x1UL
) pfm_unfreeze_pmu();
6031 * now get can unmask PMU interrupts, they will
6032 * be treated as purely spurious and we will not
6033 * lose any information
6035 UNPROTECT_CTX(ctx
,flags
);
6037 #endif /* CONFIG_SMP */
6041 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6044 pfm_load_regs (struct task_struct
*task
)
6047 unsigned long pmc_mask
= 0UL, pmd_mask
= 0UL;
6048 unsigned long flags
;
6050 int need_irq_resend
;
6052 ctx
= PFM_GET_CTX(task
);
6053 if (unlikely(ctx
== NULL
)) return;
6055 BUG_ON(GET_PMU_OWNER());
6058 * possible on unload
6060 if (unlikely((task
->thread
.flags
& IA64_THREAD_PM_VALID
) == 0)) return;
6063 * we always come here with interrupts ALREADY disabled by
6064 * the scheduler. So we simply need to protect against concurrent
6065 * access, not CPU concurrency.
6067 flags
= pfm_protect_ctx_ctxsw(ctx
);
6068 psr
= pfm_get_psr();
6070 need_irq_resend
= pmu_conf
->flags
& PFM_PMU_IRQ_RESEND
;
6072 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
6073 BUG_ON(psr
& IA64_PSR_I
);
6075 if (unlikely(ctx
->ctx_state
== PFM_CTX_ZOMBIE
)) {
6076 struct pt_regs
*regs
= task_pt_regs(task
);
6078 BUG_ON(ctx
->ctx_smpl_hdr
);
6080 pfm_force_cleanup(ctx
, regs
);
6082 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
6085 * this one (kmalloc'ed) is fine with interrupts disabled
6087 pfm_context_free(ctx
);
6093 * we restore ALL the debug registers to avoid picking up
6096 if (ctx
->ctx_fl_using_dbreg
) {
6097 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
6098 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
6101 * retrieve saved psr.up
6103 psr_up
= ctx
->ctx_saved_psr_up
;
6106 * if we were the last user of the PMU on that CPU,
6107 * then nothing to do except restore psr
6109 if (GET_LAST_CPU(ctx
) == smp_processor_id() && ctx
->ctx_last_activation
== GET_ACTIVATION()) {
6112 * retrieve partial reload masks (due to user modifications)
6114 pmc_mask
= ctx
->ctx_reload_pmcs
[0];
6115 pmd_mask
= ctx
->ctx_reload_pmds
[0];
6119 * To avoid leaking information to the user level when psr.sp=0,
6120 * we must reload ALL implemented pmds (even the ones we don't use).
6121 * In the kernel we only allow PFM_READ_PMDS on registers which
6122 * we initialized or requested (sampling) so there is no risk there.
6124 pmd_mask
= pfm_sysctl
.fastctxsw
? ctx
->ctx_used_pmds
[0] : ctx
->ctx_all_pmds
[0];
6127 * ALL accessible PMCs are systematically reloaded, unused registers
6128 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6129 * up stale configuration.
6131 * PMC0 is never in the mask. It is always restored separately.
6133 pmc_mask
= ctx
->ctx_all_pmcs
[0];
6136 * when context is MASKED, we will restore PMC with plm=0
6137 * and PMD with stale information, but that's ok, nothing
6140 * XXX: optimize here
6142 if (pmd_mask
) pfm_restore_pmds(ctx
->th_pmds
, pmd_mask
);
6143 if (pmc_mask
) pfm_restore_pmcs(ctx
->th_pmcs
, pmc_mask
);
6146 * check for pending overflow at the time the state
6149 if (unlikely(PMC0_HAS_OVFL(ctx
->th_pmcs
[0]))) {
6151 * reload pmc0 with the overflow information
6152 * On McKinley PMU, this will trigger a PMU interrupt
6154 ia64_set_pmc(0, ctx
->th_pmcs
[0]);
6156 ctx
->th_pmcs
[0] = 0UL;
6159 * will replay the PMU interrupt
6161 if (need_irq_resend
) ia64_resend_irq(IA64_PERFMON_VECTOR
);
6163 pfm_stats
[smp_processor_id()].pfm_replay_ovfl_intr_count
++;
6167 * we just did a reload, so we reset the partial reload fields
6169 ctx
->ctx_reload_pmcs
[0] = 0UL;
6170 ctx
->ctx_reload_pmds
[0] = 0UL;
6172 SET_LAST_CPU(ctx
, smp_processor_id());
6175 * dump activation value for this PMU
6179 * record current activation for this context
6181 SET_ACTIVATION(ctx
);
6184 * establish new ownership.
6186 SET_PMU_OWNER(task
, ctx
);
6189 * restore the psr.up bit. measurement
6191 * no PMU interrupt can happen at this point
6192 * because we still have interrupts disabled.
6194 if (likely(psr_up
)) pfm_set_psr_up();
6197 * allow concurrent access to context
6199 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
6201 #else /* !CONFIG_SMP */
6203 * reload PMU state for UP kernels
6204 * in 2.5 we come here with interrupts disabled
6207 pfm_load_regs (struct task_struct
*task
)
6210 struct task_struct
*owner
;
6211 unsigned long pmd_mask
, pmc_mask
;
6213 int need_irq_resend
;
6215 owner
= GET_PMU_OWNER();
6216 ctx
= PFM_GET_CTX(task
);
6217 psr
= pfm_get_psr();
6219 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
6220 BUG_ON(psr
& IA64_PSR_I
);
6223 * we restore ALL the debug registers to avoid picking up
6226 * This must be done even when the task is still the owner
6227 * as the registers may have been modified via ptrace()
6228 * (not perfmon) by the previous task.
6230 if (ctx
->ctx_fl_using_dbreg
) {
6231 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
6232 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
6236 * retrieved saved psr.up
6238 psr_up
= ctx
->ctx_saved_psr_up
;
6239 need_irq_resend
= pmu_conf
->flags
& PFM_PMU_IRQ_RESEND
;
6242 * short path, our state is still there, just
6243 * need to restore psr and we go
6245 * we do not touch either PMC nor PMD. the psr is not touched
6246 * by the overflow_handler. So we are safe w.r.t. to interrupt
6247 * concurrency even without interrupt masking.
6249 if (likely(owner
== task
)) {
6250 if (likely(psr_up
)) pfm_set_psr_up();
6255 * someone else is still using the PMU, first push it out and
6256 * then we'll be able to install our stuff !
6258 * Upon return, there will be no owner for the current PMU
6260 if (owner
) pfm_lazy_save_regs(owner
);
6263 * To avoid leaking information to the user level when psr.sp=0,
6264 * we must reload ALL implemented pmds (even the ones we don't use).
6265 * In the kernel we only allow PFM_READ_PMDS on registers which
6266 * we initialized or requested (sampling) so there is no risk there.
6268 pmd_mask
= pfm_sysctl
.fastctxsw
? ctx
->ctx_used_pmds
[0] : ctx
->ctx_all_pmds
[0];
6271 * ALL accessible PMCs are systematically reloaded, unused registers
6272 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6273 * up stale configuration.
6275 * PMC0 is never in the mask. It is always restored separately
6277 pmc_mask
= ctx
->ctx_all_pmcs
[0];
6279 pfm_restore_pmds(ctx
->th_pmds
, pmd_mask
);
6280 pfm_restore_pmcs(ctx
->th_pmcs
, pmc_mask
);
6283 * check for pending overflow at the time the state
6286 if (unlikely(PMC0_HAS_OVFL(ctx
->th_pmcs
[0]))) {
6288 * reload pmc0 with the overflow information
6289 * On McKinley PMU, this will trigger a PMU interrupt
6291 ia64_set_pmc(0, ctx
->th_pmcs
[0]);
6294 ctx
->th_pmcs
[0] = 0UL;
6297 * will replay the PMU interrupt
6299 if (need_irq_resend
) ia64_resend_irq(IA64_PERFMON_VECTOR
);
6301 pfm_stats
[smp_processor_id()].pfm_replay_ovfl_intr_count
++;
6305 * establish new ownership.
6307 SET_PMU_OWNER(task
, ctx
);
6310 * restore the psr.up bit. measurement
6312 * no PMU interrupt can happen at this point
6313 * because we still have interrupts disabled.
6315 if (likely(psr_up
)) pfm_set_psr_up();
6317 #endif /* CONFIG_SMP */
6320 * this function assumes monitoring is stopped
6323 pfm_flush_pmds(struct task_struct
*task
, pfm_context_t
*ctx
)
6326 unsigned long mask2
, val
, pmd_val
, ovfl_val
;
6327 int i
, can_access_pmu
= 0;
6331 * is the caller the task being monitored (or which initiated the
6332 * session for system wide measurements)
6334 is_self
= ctx
->ctx_task
== task
? 1 : 0;
6337 * can access PMU is task is the owner of the PMU state on the current CPU
6338 * or if we are running on the CPU bound to the context in system-wide mode
6339 * (that is not necessarily the task the context is attached to in this mode).
6340 * In system-wide we always have can_access_pmu true because a task running on an
6341 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6343 can_access_pmu
= (GET_PMU_OWNER() == task
) || (ctx
->ctx_fl_system
&& ctx
->ctx_cpu
== smp_processor_id());
6344 if (can_access_pmu
) {
6346 * Mark the PMU as not owned
6347 * This will cause the interrupt handler to do nothing in case an overflow
6348 * interrupt was in-flight
6349 * This also guarantees that pmc0 will contain the final state
6350 * It virtually gives us full control on overflow processing from that point
6353 SET_PMU_OWNER(NULL
, NULL
);
6354 DPRINT(("releasing ownership\n"));
6357 * read current overflow status:
6359 * we are guaranteed to read the final stable state
6362 pmc0
= ia64_get_pmc(0); /* slow */
6365 * reset freeze bit, overflow status information destroyed
6369 pmc0
= ctx
->th_pmcs
[0];
6371 * clear whatever overflow status bits there were
6373 ctx
->th_pmcs
[0] = 0;
6375 ovfl_val
= pmu_conf
->ovfl_val
;
6377 * we save all the used pmds
6378 * we take care of overflows for counting PMDs
6380 * XXX: sampling situation is not taken into account here
6382 mask2
= ctx
->ctx_used_pmds
[0];
6384 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self
, ovfl_val
, mask2
));
6386 for (i
= 0; mask2
; i
++, mask2
>>=1) {
6388 /* skip non used pmds */
6389 if ((mask2
& 0x1) == 0) continue;
6392 * can access PMU always true in system wide mode
6394 val
= pmd_val
= can_access_pmu
? ia64_get_pmd(i
) : ctx
->th_pmds
[i
];
6396 if (PMD_IS_COUNTING(i
)) {
6397 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6400 ctx
->ctx_pmds
[i
].val
,
6404 * we rebuild the full 64 bit value of the counter
6406 val
= ctx
->ctx_pmds
[i
].val
+ (val
& ovfl_val
);
6409 * now everything is in ctx_pmds[] and we need
6410 * to clear the saved context from save_regs() such that
6411 * pfm_read_pmds() gets the correct value
6416 * take care of overflow inline
6418 if (pmc0
& (1UL << i
)) {
6419 val
+= 1 + ovfl_val
;
6420 DPRINT(("[%d] pmd[%d] overflowed\n", task
->pid
, i
));
6424 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task
->pid
, i
, val
, pmd_val
));
6426 if (is_self
) ctx
->th_pmds
[i
] = pmd_val
;
6428 ctx
->ctx_pmds
[i
].val
= val
;
6432 static struct irqaction perfmon_irqaction
= {
6433 .handler
= pfm_interrupt_handler
,
6434 .flags
= IRQF_DISABLED
,
6439 pfm_alt_save_pmu_state(void *data
)
6441 struct pt_regs
*regs
;
6443 regs
= task_pt_regs(current
);
6445 DPRINT(("called\n"));
6448 * should not be necessary but
6449 * let's take not risk
6453 ia64_psr(regs
)->pp
= 0;
6456 * This call is required
6457 * May cause a spurious interrupt on some processors
6465 pfm_alt_restore_pmu_state(void *data
)
6467 struct pt_regs
*regs
;
6469 regs
= task_pt_regs(current
);
6471 DPRINT(("called\n"));
6474 * put PMU back in state expected
6479 ia64_psr(regs
)->pp
= 0;
6482 * perfmon runs with PMU unfrozen at all times
6490 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t
*hdl
)
6495 /* some sanity checks */
6496 if (hdl
== NULL
|| hdl
->handler
== NULL
) return -EINVAL
;
6498 /* do the easy test first */
6499 if (pfm_alt_intr_handler
) return -EBUSY
;
6501 /* one at a time in the install or remove, just fail the others */
6502 if (!spin_trylock(&pfm_alt_install_check
)) {
6506 /* reserve our session */
6507 for_each_online_cpu(reserve_cpu
) {
6508 ret
= pfm_reserve_session(NULL
, 1, reserve_cpu
);
6509 if (ret
) goto cleanup_reserve
;
6512 /* save the current system wide pmu states */
6513 ret
= on_each_cpu(pfm_alt_save_pmu_state
, NULL
, 0, 1);
6515 DPRINT(("on_each_cpu() failed: %d\n", ret
));
6516 goto cleanup_reserve
;
6519 /* officially change to the alternate interrupt handler */
6520 pfm_alt_intr_handler
= hdl
;
6522 spin_unlock(&pfm_alt_install_check
);
6527 for_each_online_cpu(i
) {
6528 /* don't unreserve more than we reserved */
6529 if (i
>= reserve_cpu
) break;
6531 pfm_unreserve_session(NULL
, 1, i
);
6534 spin_unlock(&pfm_alt_install_check
);
6538 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt
);
6541 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t
*hdl
)
6546 if (hdl
== NULL
) return -EINVAL
;
6548 /* cannot remove someone else's handler! */
6549 if (pfm_alt_intr_handler
!= hdl
) return -EINVAL
;
6551 /* one at a time in the install or remove, just fail the others */
6552 if (!spin_trylock(&pfm_alt_install_check
)) {
6556 pfm_alt_intr_handler
= NULL
;
6558 ret
= on_each_cpu(pfm_alt_restore_pmu_state
, NULL
, 0, 1);
6560 DPRINT(("on_each_cpu() failed: %d\n", ret
));
6563 for_each_online_cpu(i
) {
6564 pfm_unreserve_session(NULL
, 1, i
);
6567 spin_unlock(&pfm_alt_install_check
);
6571 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt
);
6574 * perfmon initialization routine, called from the initcall() table
6576 static int init_pfm_fs(void);
6584 family
= local_cpu_data
->family
;
6589 if ((*p
)->probe() == 0) goto found
;
6590 } else if ((*p
)->pmu_family
== family
|| (*p
)->pmu_family
== 0xff) {
6601 static struct file_operations pfm_proc_fops
= {
6602 .open
= pfm_proc_open
,
6604 .llseek
= seq_lseek
,
6605 .release
= seq_release
,
6611 unsigned int n
, n_counters
, i
;
6613 printk("perfmon: version %u.%u IRQ %u\n",
6616 IA64_PERFMON_VECTOR
);
6618 if (pfm_probe_pmu()) {
6619 printk(KERN_INFO
"perfmon: disabled, there is no support for processor family %d\n",
6620 local_cpu_data
->family
);
6625 * compute the number of implemented PMD/PMC from the
6626 * description tables
6629 for (i
=0; PMC_IS_LAST(i
) == 0; i
++) {
6630 if (PMC_IS_IMPL(i
) == 0) continue;
6631 pmu_conf
->impl_pmcs
[i
>>6] |= 1UL << (i
&63);
6634 pmu_conf
->num_pmcs
= n
;
6636 n
= 0; n_counters
= 0;
6637 for (i
=0; PMD_IS_LAST(i
) == 0; i
++) {
6638 if (PMD_IS_IMPL(i
) == 0) continue;
6639 pmu_conf
->impl_pmds
[i
>>6] |= 1UL << (i
&63);
6641 if (PMD_IS_COUNTING(i
)) n_counters
++;
6643 pmu_conf
->num_pmds
= n
;
6644 pmu_conf
->num_counters
= n_counters
;
6647 * sanity checks on the number of debug registers
6649 if (pmu_conf
->use_rr_dbregs
) {
6650 if (pmu_conf
->num_ibrs
> IA64_NUM_DBG_REGS
) {
6651 printk(KERN_INFO
"perfmon: unsupported number of code debug registers (%u)\n", pmu_conf
->num_ibrs
);
6655 if (pmu_conf
->num_dbrs
> IA64_NUM_DBG_REGS
) {
6656 printk(KERN_INFO
"perfmon: unsupported number of data debug registers (%u)\n", pmu_conf
->num_ibrs
);
6662 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6666 pmu_conf
->num_counters
,
6667 ffz(pmu_conf
->ovfl_val
));
6670 if (pmu_conf
->num_pmds
>= PFM_NUM_PMD_REGS
|| pmu_conf
->num_pmcs
>= PFM_NUM_PMC_REGS
) {
6671 printk(KERN_ERR
"perfmon: not enough pmc/pmd, perfmon disabled\n");
6677 * create /proc/perfmon (mostly for debugging purposes)
6679 perfmon_dir
= create_proc_entry("perfmon", S_IRUGO
, NULL
);
6680 if (perfmon_dir
== NULL
) {
6681 printk(KERN_ERR
"perfmon: cannot create /proc entry, perfmon disabled\n");
6686 * install customized file operations for /proc/perfmon entry
6688 perfmon_dir
->proc_fops
= &pfm_proc_fops
;
6691 * create /proc/sys/kernel/perfmon (for debugging purposes)
6693 pfm_sysctl_header
= register_sysctl_table(pfm_sysctl_root
, 0);
6696 * initialize all our spinlocks
6698 spin_lock_init(&pfm_sessions
.pfs_lock
);
6699 spin_lock_init(&pfm_buffer_fmt_lock
);
6703 for(i
=0; i
< NR_CPUS
; i
++) pfm_stats
[i
].pfm_ovfl_intr_cycles_min
= ~0UL;
6708 __initcall(pfm_init
);
6711 * this function is called before pfm_init()
6714 pfm_init_percpu (void)
6716 static int first_time
=1;
6718 * make sure no measurement is active
6719 * (may inherit programmed PMCs from EFI).
6725 * we run with the PMU not frozen at all times
6730 register_percpu_irq(IA64_PERFMON_VECTOR
, &perfmon_irqaction
);
6734 ia64_setreg(_IA64_REG_CR_PMV
, IA64_PERFMON_VECTOR
);
6739 * used for debug purposes only
6742 dump_pmu_state(const char *from
)
6744 struct task_struct
*task
;
6745 struct pt_regs
*regs
;
6747 unsigned long psr
, dcr
, info
, flags
;
6750 local_irq_save(flags
);
6752 this_cpu
= smp_processor_id();
6753 regs
= task_pt_regs(current
);
6754 info
= PFM_CPUINFO_GET();
6755 dcr
= ia64_getreg(_IA64_REG_CR_DCR
);
6757 if (info
== 0 && ia64_psr(regs
)->pp
== 0 && (dcr
& IA64_DCR_PP
) == 0) {
6758 local_irq_restore(flags
);
6762 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6769 task
= GET_PMU_OWNER();
6770 ctx
= GET_PMU_CTX();
6772 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu
, task
? task
->pid
: -1, ctx
);
6774 psr
= pfm_get_psr();
6776 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",
6779 psr
& IA64_PSR_PP
? 1 : 0,
6780 psr
& IA64_PSR_UP
? 1 : 0,
6781 dcr
& IA64_DCR_PP
? 1 : 0,
6784 ia64_psr(regs
)->pp
);
6786 ia64_psr(regs
)->up
= 0;
6787 ia64_psr(regs
)->pp
= 0;
6789 for (i
=1; PMC_IS_LAST(i
) == 0; i
++) {
6790 if (PMC_IS_IMPL(i
) == 0) continue;
6791 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
]);
6794 for (i
=1; PMD_IS_LAST(i
) == 0; i
++) {
6795 if (PMD_IS_IMPL(i
) == 0) continue;
6796 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
]);
6800 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6803 ctx
->ctx_smpl_vaddr
,
6807 ctx
->ctx_saved_psr_up
);
6809 local_irq_restore(flags
);
6813 * called from process.c:copy_thread(). task is new child.
6816 pfm_inherit(struct task_struct
*task
, struct pt_regs
*regs
)
6818 struct thread_struct
*thread
;
6820 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task
->pid
));
6822 thread
= &task
->thread
;
6825 * cut links inherited from parent (current)
6827 thread
->pfm_context
= NULL
;
6829 PFM_SET_WORK_PENDING(task
, 0);
6832 * the psr bits are already set properly in copy_threads()
6835 #else /* !CONFIG_PERFMON */
6837 sys_perfmonctl (int fd
, int cmd
, void *arg
, int count
)
6841 #endif /* CONFIG_PERFMON */