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 const 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
= kzalloc(sizeof(pfm_context_t
), GFP_KERNEL
);
858 DPRINT(("alloc ctx @%p\n", ctx
));
864 pfm_context_free(pfm_context_t
*ctx
)
867 DPRINT(("free ctx @%p\n", ctx
));
873 pfm_mask_monitoring(struct task_struct
*task
)
875 pfm_context_t
*ctx
= PFM_GET_CTX(task
);
876 unsigned long mask
, val
, ovfl_mask
;
879 DPRINT_ovfl(("masking monitoring for [%d]\n", task
->pid
));
881 ovfl_mask
= pmu_conf
->ovfl_val
;
883 * monitoring can only be masked as a result of a valid
884 * counter overflow. In UP, it means that the PMU still
885 * has an owner. Note that the owner can be different
886 * from the current task. However the PMU state belongs
888 * In SMP, a valid overflow only happens when task is
889 * current. Therefore if we come here, we know that
890 * the PMU state belongs to the current task, therefore
891 * we can access the live registers.
893 * So in both cases, the live register contains the owner's
894 * state. We can ONLY touch the PMU registers and NOT the PSR.
896 * As a consequence to this call, the ctx->th_pmds[] array
897 * contains stale information which must be ignored
898 * when context is reloaded AND monitoring is active (see
901 mask
= ctx
->ctx_used_pmds
[0];
902 for (i
= 0; mask
; i
++, mask
>>=1) {
903 /* skip non used pmds */
904 if ((mask
& 0x1) == 0) continue;
905 val
= ia64_get_pmd(i
);
907 if (PMD_IS_COUNTING(i
)) {
909 * we rebuild the full 64 bit value of the counter
911 ctx
->ctx_pmds
[i
].val
+= (val
& ovfl_mask
);
913 ctx
->ctx_pmds
[i
].val
= val
;
915 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
917 ctx
->ctx_pmds
[i
].val
,
921 * mask monitoring by setting the privilege level to 0
922 * we cannot use psr.pp/psr.up for this, it is controlled by
925 * if task is current, modify actual registers, otherwise modify
926 * thread save state, i.e., what will be restored in pfm_load_regs()
928 mask
= ctx
->ctx_used_monitors
[0] >> PMU_FIRST_COUNTER
;
929 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>=1) {
930 if ((mask
& 0x1) == 0UL) continue;
931 ia64_set_pmc(i
, ctx
->th_pmcs
[i
] & ~0xfUL
);
932 ctx
->th_pmcs
[i
] &= ~0xfUL
;
933 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i
, ctx
->th_pmcs
[i
]));
936 * make all of this visible
942 * must always be done with task == current
944 * context must be in MASKED state when calling
947 pfm_restore_monitoring(struct task_struct
*task
)
949 pfm_context_t
*ctx
= PFM_GET_CTX(task
);
950 unsigned long mask
, ovfl_mask
;
951 unsigned long psr
, val
;
954 is_system
= ctx
->ctx_fl_system
;
955 ovfl_mask
= pmu_conf
->ovfl_val
;
957 if (task
!= current
) {
958 printk(KERN_ERR
"perfmon.%d: invalid task[%d] current[%d]\n", __LINE__
, task
->pid
, current
->pid
);
961 if (ctx
->ctx_state
!= PFM_CTX_MASKED
) {
962 printk(KERN_ERR
"perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__
,
963 task
->pid
, current
->pid
, ctx
->ctx_state
);
968 * monitoring is masked via the PMC.
969 * As we restore their value, we do not want each counter to
970 * restart right away. We stop monitoring using the PSR,
971 * restore the PMC (and PMD) and then re-establish the psr
972 * as it was. Note that there can be no pending overflow at
973 * this point, because monitoring was MASKED.
975 * system-wide session are pinned and self-monitoring
977 if (is_system
&& (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP
)) {
979 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) & ~IA64_DCR_PP
);
985 * first, we restore the PMD
987 mask
= ctx
->ctx_used_pmds
[0];
988 for (i
= 0; mask
; i
++, mask
>>=1) {
989 /* skip non used pmds */
990 if ((mask
& 0x1) == 0) continue;
992 if (PMD_IS_COUNTING(i
)) {
994 * we split the 64bit value according to
997 val
= ctx
->ctx_pmds
[i
].val
& ovfl_mask
;
998 ctx
->ctx_pmds
[i
].val
&= ~ovfl_mask
;
1000 val
= ctx
->ctx_pmds
[i
].val
;
1002 ia64_set_pmd(i
, val
);
1004 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1006 ctx
->ctx_pmds
[i
].val
,
1012 mask
= ctx
->ctx_used_monitors
[0] >> PMU_FIRST_COUNTER
;
1013 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>=1) {
1014 if ((mask
& 0x1) == 0UL) continue;
1015 ctx
->th_pmcs
[i
] = ctx
->ctx_pmcs
[i
];
1016 ia64_set_pmc(i
, ctx
->th_pmcs
[i
]);
1017 DPRINT(("[%d] pmc[%d]=0x%lx\n", task
->pid
, i
, ctx
->th_pmcs
[i
]));
1022 * must restore DBR/IBR because could be modified while masked
1023 * XXX: need to optimize
1025 if (ctx
->ctx_fl_using_dbreg
) {
1026 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
1027 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
1033 if (is_system
&& (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP
)) {
1035 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) | IA64_DCR_PP
);
1042 pfm_save_pmds(unsigned long *pmds
, unsigned long mask
)
1048 for (i
=0; mask
; i
++, mask
>>=1) {
1049 if (mask
& 0x1) pmds
[i
] = ia64_get_pmd(i
);
1054 * reload from thread state (used for ctxw only)
1057 pfm_restore_pmds(unsigned long *pmds
, unsigned long mask
)
1060 unsigned long val
, ovfl_val
= pmu_conf
->ovfl_val
;
1062 for (i
=0; mask
; i
++, mask
>>=1) {
1063 if ((mask
& 0x1) == 0) continue;
1064 val
= PMD_IS_COUNTING(i
) ? pmds
[i
] & ovfl_val
: pmds
[i
];
1065 ia64_set_pmd(i
, val
);
1071 * propagate PMD from context to thread-state
1074 pfm_copy_pmds(struct task_struct
*task
, pfm_context_t
*ctx
)
1076 unsigned long ovfl_val
= pmu_conf
->ovfl_val
;
1077 unsigned long mask
= ctx
->ctx_all_pmds
[0];
1081 DPRINT(("mask=0x%lx\n", mask
));
1083 for (i
=0; mask
; i
++, mask
>>=1) {
1085 val
= ctx
->ctx_pmds
[i
].val
;
1088 * We break up the 64 bit value into 2 pieces
1089 * the lower bits go to the machine state in the
1090 * thread (will be reloaded on ctxsw in).
1091 * The upper part stays in the soft-counter.
1093 if (PMD_IS_COUNTING(i
)) {
1094 ctx
->ctx_pmds
[i
].val
= val
& ~ovfl_val
;
1097 ctx
->th_pmds
[i
] = val
;
1099 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1102 ctx
->ctx_pmds
[i
].val
));
1107 * propagate PMC from context to thread-state
1110 pfm_copy_pmcs(struct task_struct
*task
, pfm_context_t
*ctx
)
1112 unsigned long mask
= ctx
->ctx_all_pmcs
[0];
1115 DPRINT(("mask=0x%lx\n", mask
));
1117 for (i
=0; mask
; i
++, mask
>>=1) {
1118 /* masking 0 with ovfl_val yields 0 */
1119 ctx
->th_pmcs
[i
] = ctx
->ctx_pmcs
[i
];
1120 DPRINT(("pmc[%d]=0x%lx\n", i
, ctx
->th_pmcs
[i
]));
1127 pfm_restore_pmcs(unsigned long *pmcs
, unsigned long mask
)
1131 for (i
=0; mask
; i
++, mask
>>=1) {
1132 if ((mask
& 0x1) == 0) continue;
1133 ia64_set_pmc(i
, pmcs
[i
]);
1139 pfm_uuid_cmp(pfm_uuid_t a
, pfm_uuid_t b
)
1141 return memcmp(a
, b
, sizeof(pfm_uuid_t
));
1145 pfm_buf_fmt_exit(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, void *buf
, struct pt_regs
*regs
)
1148 if (fmt
->fmt_exit
) ret
= (*fmt
->fmt_exit
)(task
, buf
, regs
);
1153 pfm_buf_fmt_getsize(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, unsigned int flags
, int cpu
, void *arg
, unsigned long *size
)
1156 if (fmt
->fmt_getsize
) ret
= (*fmt
->fmt_getsize
)(task
, flags
, cpu
, arg
, size
);
1162 pfm_buf_fmt_validate(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, unsigned int flags
,
1166 if (fmt
->fmt_validate
) ret
= (*fmt
->fmt_validate
)(task
, flags
, cpu
, arg
);
1171 pfm_buf_fmt_init(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, void *buf
, unsigned int flags
,
1175 if (fmt
->fmt_init
) ret
= (*fmt
->fmt_init
)(task
, buf
, flags
, cpu
, arg
);
1180 pfm_buf_fmt_restart(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, pfm_ovfl_ctrl_t
*ctrl
, void *buf
, struct pt_regs
*regs
)
1183 if (fmt
->fmt_restart
) ret
= (*fmt
->fmt_restart
)(task
, ctrl
, buf
, regs
);
1188 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
)
1191 if (fmt
->fmt_restart_active
) ret
= (*fmt
->fmt_restart_active
)(task
, ctrl
, buf
, regs
);
1195 static pfm_buffer_fmt_t
*
1196 __pfm_find_buffer_fmt(pfm_uuid_t uuid
)
1198 struct list_head
* pos
;
1199 pfm_buffer_fmt_t
* entry
;
1201 list_for_each(pos
, &pfm_buffer_fmt_list
) {
1202 entry
= list_entry(pos
, pfm_buffer_fmt_t
, fmt_list
);
1203 if (pfm_uuid_cmp(uuid
, entry
->fmt_uuid
) == 0)
1210 * find a buffer format based on its uuid
1212 static pfm_buffer_fmt_t
*
1213 pfm_find_buffer_fmt(pfm_uuid_t uuid
)
1215 pfm_buffer_fmt_t
* fmt
;
1216 spin_lock(&pfm_buffer_fmt_lock
);
1217 fmt
= __pfm_find_buffer_fmt(uuid
);
1218 spin_unlock(&pfm_buffer_fmt_lock
);
1223 pfm_register_buffer_fmt(pfm_buffer_fmt_t
*fmt
)
1227 /* some sanity checks */
1228 if (fmt
== NULL
|| fmt
->fmt_name
== NULL
) return -EINVAL
;
1230 /* we need at least a handler */
1231 if (fmt
->fmt_handler
== NULL
) return -EINVAL
;
1234 * XXX: need check validity of fmt_arg_size
1237 spin_lock(&pfm_buffer_fmt_lock
);
1239 if (__pfm_find_buffer_fmt(fmt
->fmt_uuid
)) {
1240 printk(KERN_ERR
"perfmon: duplicate sampling format: %s\n", fmt
->fmt_name
);
1244 list_add(&fmt
->fmt_list
, &pfm_buffer_fmt_list
);
1245 printk(KERN_INFO
"perfmon: added sampling format %s\n", fmt
->fmt_name
);
1248 spin_unlock(&pfm_buffer_fmt_lock
);
1251 EXPORT_SYMBOL(pfm_register_buffer_fmt
);
1254 pfm_unregister_buffer_fmt(pfm_uuid_t uuid
)
1256 pfm_buffer_fmt_t
*fmt
;
1259 spin_lock(&pfm_buffer_fmt_lock
);
1261 fmt
= __pfm_find_buffer_fmt(uuid
);
1263 printk(KERN_ERR
"perfmon: cannot unregister format, not found\n");
1267 list_del_init(&fmt
->fmt_list
);
1268 printk(KERN_INFO
"perfmon: removed sampling format: %s\n", fmt
->fmt_name
);
1271 spin_unlock(&pfm_buffer_fmt_lock
);
1275 EXPORT_SYMBOL(pfm_unregister_buffer_fmt
);
1277 extern void update_pal_halt_status(int);
1280 pfm_reserve_session(struct task_struct
*task
, int is_syswide
, unsigned int cpu
)
1282 unsigned long flags
;
1284 * validy checks on cpu_mask have been done upstream
1288 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1289 pfm_sessions
.pfs_sys_sessions
,
1290 pfm_sessions
.pfs_task_sessions
,
1291 pfm_sessions
.pfs_sys_use_dbregs
,
1297 * cannot mix system wide and per-task sessions
1299 if (pfm_sessions
.pfs_task_sessions
> 0UL) {
1300 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1301 pfm_sessions
.pfs_task_sessions
));
1305 if (pfm_sessions
.pfs_sys_session
[cpu
]) goto error_conflict
;
1307 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu
, smp_processor_id()));
1309 pfm_sessions
.pfs_sys_session
[cpu
] = task
;
1311 pfm_sessions
.pfs_sys_sessions
++ ;
1314 if (pfm_sessions
.pfs_sys_sessions
) goto abort
;
1315 pfm_sessions
.pfs_task_sessions
++;
1318 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1319 pfm_sessions
.pfs_sys_sessions
,
1320 pfm_sessions
.pfs_task_sessions
,
1321 pfm_sessions
.pfs_sys_use_dbregs
,
1326 * disable default_idle() to go to PAL_HALT
1328 update_pal_halt_status(0);
1335 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1336 pfm_sessions
.pfs_sys_session
[cpu
]->pid
,
1346 pfm_unreserve_session(pfm_context_t
*ctx
, int is_syswide
, unsigned int cpu
)
1348 unsigned long flags
;
1350 * validy checks on cpu_mask have been done upstream
1354 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1355 pfm_sessions
.pfs_sys_sessions
,
1356 pfm_sessions
.pfs_task_sessions
,
1357 pfm_sessions
.pfs_sys_use_dbregs
,
1363 pfm_sessions
.pfs_sys_session
[cpu
] = NULL
;
1365 * would not work with perfmon+more than one bit in cpu_mask
1367 if (ctx
&& ctx
->ctx_fl_using_dbreg
) {
1368 if (pfm_sessions
.pfs_sys_use_dbregs
== 0) {
1369 printk(KERN_ERR
"perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx
);
1371 pfm_sessions
.pfs_sys_use_dbregs
--;
1374 pfm_sessions
.pfs_sys_sessions
--;
1376 pfm_sessions
.pfs_task_sessions
--;
1378 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1379 pfm_sessions
.pfs_sys_sessions
,
1380 pfm_sessions
.pfs_task_sessions
,
1381 pfm_sessions
.pfs_sys_use_dbregs
,
1386 * if possible, enable default_idle() to go into PAL_HALT
1388 if (pfm_sessions
.pfs_task_sessions
== 0 && pfm_sessions
.pfs_sys_sessions
== 0)
1389 update_pal_halt_status(1);
1397 * removes virtual mapping of the sampling buffer.
1398 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1399 * a PROTECT_CTX() section.
1402 pfm_remove_smpl_mapping(struct task_struct
*task
, void *vaddr
, unsigned long size
)
1407 if (task
->mm
== NULL
|| size
== 0UL || vaddr
== NULL
) {
1408 printk(KERN_ERR
"perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task
->pid
, task
->mm
);
1412 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr
, size
));
1415 * does the actual unmapping
1417 down_write(&task
->mm
->mmap_sem
);
1419 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr
, size
));
1421 r
= pfm_do_munmap(task
->mm
, (unsigned long)vaddr
, size
, 0);
1423 up_write(&task
->mm
->mmap_sem
);
1425 printk(KERN_ERR
"perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task
->pid
, vaddr
, size
);
1428 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr
, size
, r
));
1434 * free actual physical storage used by sampling buffer
1438 pfm_free_smpl_buffer(pfm_context_t
*ctx
)
1440 pfm_buffer_fmt_t
*fmt
;
1442 if (ctx
->ctx_smpl_hdr
== NULL
) goto invalid_free
;
1445 * we won't use the buffer format anymore
1447 fmt
= ctx
->ctx_buf_fmt
;
1449 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1452 ctx
->ctx_smpl_vaddr
));
1454 pfm_buf_fmt_exit(fmt
, current
, NULL
, NULL
);
1459 pfm_rvfree(ctx
->ctx_smpl_hdr
, ctx
->ctx_smpl_size
);
1461 ctx
->ctx_smpl_hdr
= NULL
;
1462 ctx
->ctx_smpl_size
= 0UL;
1467 printk(KERN_ERR
"perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current
->pid
);
1473 pfm_exit_smpl_buffer(pfm_buffer_fmt_t
*fmt
)
1475 if (fmt
== NULL
) return;
1477 pfm_buf_fmt_exit(fmt
, current
, NULL
, NULL
);
1482 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1483 * no real gain from having the whole whorehouse mounted. So we don't need
1484 * any operations on the root directory. However, we need a non-trivial
1485 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1487 static struct vfsmount
*pfmfs_mnt
;
1492 int err
= register_filesystem(&pfm_fs_type
);
1494 pfmfs_mnt
= kern_mount(&pfm_fs_type
);
1495 err
= PTR_ERR(pfmfs_mnt
);
1496 if (IS_ERR(pfmfs_mnt
))
1497 unregister_filesystem(&pfm_fs_type
);
1507 unregister_filesystem(&pfm_fs_type
);
1512 pfm_read(struct file
*filp
, char __user
*buf
, size_t size
, loff_t
*ppos
)
1517 unsigned long flags
;
1518 DECLARE_WAITQUEUE(wait
, current
);
1519 if (PFM_IS_FILE(filp
) == 0) {
1520 printk(KERN_ERR
"perfmon: pfm_poll: bad magic [%d]\n", current
->pid
);
1524 ctx
= (pfm_context_t
*)filp
->private_data
;
1526 printk(KERN_ERR
"perfmon: pfm_read: NULL ctx [%d]\n", current
->pid
);
1531 * check even when there is no message
1533 if (size
< sizeof(pfm_msg_t
)) {
1534 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx
, sizeof(pfm_msg_t
)));
1538 PROTECT_CTX(ctx
, flags
);
1541 * put ourselves on the wait queue
1543 add_wait_queue(&ctx
->ctx_msgq_wait
, &wait
);
1551 set_current_state(TASK_INTERRUPTIBLE
);
1553 DPRINT(("head=%d tail=%d\n", ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
1556 if(PFM_CTXQ_EMPTY(ctx
) == 0) break;
1558 UNPROTECT_CTX(ctx
, flags
);
1561 * check non-blocking read
1564 if(filp
->f_flags
& O_NONBLOCK
) break;
1567 * check pending signals
1569 if(signal_pending(current
)) {
1574 * no message, so wait
1578 PROTECT_CTX(ctx
, flags
);
1580 DPRINT(("[%d] back to running ret=%ld\n", current
->pid
, ret
));
1581 set_current_state(TASK_RUNNING
);
1582 remove_wait_queue(&ctx
->ctx_msgq_wait
, &wait
);
1584 if (ret
< 0) goto abort
;
1587 msg
= pfm_get_next_msg(ctx
);
1589 printk(KERN_ERR
"perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx
, current
->pid
);
1593 DPRINT(("fd=%d type=%d\n", msg
->pfm_gen_msg
.msg_ctx_fd
, msg
->pfm_gen_msg
.msg_type
));
1596 if(copy_to_user(buf
, msg
, sizeof(pfm_msg_t
)) == 0) ret
= sizeof(pfm_msg_t
);
1599 UNPROTECT_CTX(ctx
, flags
);
1605 pfm_write(struct file
*file
, const char __user
*ubuf
,
1606 size_t size
, loff_t
*ppos
)
1608 DPRINT(("pfm_write called\n"));
1613 pfm_poll(struct file
*filp
, poll_table
* wait
)
1616 unsigned long flags
;
1617 unsigned int mask
= 0;
1619 if (PFM_IS_FILE(filp
) == 0) {
1620 printk(KERN_ERR
"perfmon: pfm_poll: bad magic [%d]\n", current
->pid
);
1624 ctx
= (pfm_context_t
*)filp
->private_data
;
1626 printk(KERN_ERR
"perfmon: pfm_poll: NULL ctx [%d]\n", current
->pid
);
1631 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx
->ctx_fd
));
1633 poll_wait(filp
, &ctx
->ctx_msgq_wait
, wait
);
1635 PROTECT_CTX(ctx
, flags
);
1637 if (PFM_CTXQ_EMPTY(ctx
) == 0)
1638 mask
= POLLIN
| POLLRDNORM
;
1640 UNPROTECT_CTX(ctx
, flags
);
1642 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx
->ctx_fd
, mask
));
1648 pfm_ioctl(struct inode
*inode
, struct file
*file
, unsigned int cmd
, unsigned long arg
)
1650 DPRINT(("pfm_ioctl called\n"));
1655 * interrupt cannot be masked when coming here
1658 pfm_do_fasync(int fd
, struct file
*filp
, pfm_context_t
*ctx
, int on
)
1662 ret
= fasync_helper (fd
, filp
, on
, &ctx
->ctx_async_queue
);
1664 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1668 ctx
->ctx_async_queue
, ret
));
1674 pfm_fasync(int fd
, struct file
*filp
, int on
)
1679 if (PFM_IS_FILE(filp
) == 0) {
1680 printk(KERN_ERR
"perfmon: pfm_fasync bad magic [%d]\n", current
->pid
);
1684 ctx
= (pfm_context_t
*)filp
->private_data
;
1686 printk(KERN_ERR
"perfmon: pfm_fasync NULL ctx [%d]\n", current
->pid
);
1690 * we cannot mask interrupts during this call because this may
1691 * may go to sleep if memory is not readily avalaible.
1693 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1694 * done in caller. Serialization of this function is ensured by caller.
1696 ret
= pfm_do_fasync(fd
, filp
, ctx
, on
);
1699 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1702 ctx
->ctx_async_queue
, ret
));
1709 * this function is exclusively called from pfm_close().
1710 * The context is not protected at that time, nor are interrupts
1711 * on the remote CPU. That's necessary to avoid deadlocks.
1714 pfm_syswide_force_stop(void *info
)
1716 pfm_context_t
*ctx
= (pfm_context_t
*)info
;
1717 struct pt_regs
*regs
= task_pt_regs(current
);
1718 struct task_struct
*owner
;
1719 unsigned long flags
;
1722 if (ctx
->ctx_cpu
!= smp_processor_id()) {
1723 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1725 smp_processor_id());
1728 owner
= GET_PMU_OWNER();
1729 if (owner
!= ctx
->ctx_task
) {
1730 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1732 owner
->pid
, ctx
->ctx_task
->pid
);
1735 if (GET_PMU_CTX() != ctx
) {
1736 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1738 GET_PMU_CTX(), ctx
);
1742 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx
->ctx_task
->pid
));
1744 * the context is already protected in pfm_close(), we simply
1745 * need to mask interrupts to avoid a PMU interrupt race on
1748 local_irq_save(flags
);
1750 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
1752 DPRINT(("context_unload returned %d\n", ret
));
1756 * unmask interrupts, PMU interrupts are now spurious here
1758 local_irq_restore(flags
);
1762 pfm_syswide_cleanup_other_cpu(pfm_context_t
*ctx
)
1766 DPRINT(("calling CPU%d for cleanup\n", ctx
->ctx_cpu
));
1767 ret
= smp_call_function_single(ctx
->ctx_cpu
, pfm_syswide_force_stop
, ctx
, 0, 1);
1768 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx
->ctx_cpu
, ret
));
1770 #endif /* CONFIG_SMP */
1773 * called for each close(). Partially free resources.
1774 * When caller is self-monitoring, the context is unloaded.
1777 pfm_flush(struct file
*filp
, fl_owner_t id
)
1780 struct task_struct
*task
;
1781 struct pt_regs
*regs
;
1782 unsigned long flags
;
1783 unsigned long smpl_buf_size
= 0UL;
1784 void *smpl_buf_vaddr
= NULL
;
1785 int state
, is_system
;
1787 if (PFM_IS_FILE(filp
) == 0) {
1788 DPRINT(("bad magic for\n"));
1792 ctx
= (pfm_context_t
*)filp
->private_data
;
1794 printk(KERN_ERR
"perfmon: pfm_flush: NULL ctx [%d]\n", current
->pid
);
1799 * remove our file from the async queue, if we use this mode.
1800 * This can be done without the context being protected. We come
1801 * here when the context has become unreacheable by other tasks.
1803 * We may still have active monitoring at this point and we may
1804 * end up in pfm_overflow_handler(). However, fasync_helper()
1805 * operates with interrupts disabled and it cleans up the
1806 * queue. If the PMU handler is called prior to entering
1807 * fasync_helper() then it will send a signal. If it is
1808 * invoked after, it will find an empty queue and no
1809 * signal will be sent. In both case, we are safe
1811 if (filp
->f_flags
& FASYNC
) {
1812 DPRINT(("cleaning up async_queue=%p\n", ctx
->ctx_async_queue
));
1813 pfm_do_fasync (-1, filp
, ctx
, 0);
1816 PROTECT_CTX(ctx
, flags
);
1818 state
= ctx
->ctx_state
;
1819 is_system
= ctx
->ctx_fl_system
;
1821 task
= PFM_CTX_TASK(ctx
);
1822 regs
= task_pt_regs(task
);
1824 DPRINT(("ctx_state=%d is_current=%d\n",
1826 task
== current
? 1 : 0));
1829 * if state == UNLOADED, then task is NULL
1833 * we must stop and unload because we are losing access to the context.
1835 if (task
== current
) {
1838 * the task IS the owner but it migrated to another CPU: that's bad
1839 * but we must handle this cleanly. Unfortunately, the kernel does
1840 * not provide a mechanism to block migration (while the context is loaded).
1842 * We need to release the resource on the ORIGINAL cpu.
1844 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
1846 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
1848 * keep context protected but unmask interrupt for IPI
1850 local_irq_restore(flags
);
1852 pfm_syswide_cleanup_other_cpu(ctx
);
1855 * restore interrupt masking
1857 local_irq_save(flags
);
1860 * context is unloaded at this point
1863 #endif /* CONFIG_SMP */
1866 DPRINT(("forcing unload\n"));
1868 * stop and unload, returning with state UNLOADED
1869 * and session unreserved.
1871 pfm_context_unload(ctx
, NULL
, 0, regs
);
1873 DPRINT(("ctx_state=%d\n", ctx
->ctx_state
));
1878 * remove virtual mapping, if any, for the calling task.
1879 * cannot reset ctx field until last user is calling close().
1881 * ctx_smpl_vaddr must never be cleared because it is needed
1882 * by every task with access to the context
1884 * When called from do_exit(), the mm context is gone already, therefore
1885 * mm is NULL, i.e., the VMA is already gone and we do not have to
1888 if (ctx
->ctx_smpl_vaddr
&& current
->mm
) {
1889 smpl_buf_vaddr
= ctx
->ctx_smpl_vaddr
;
1890 smpl_buf_size
= ctx
->ctx_smpl_size
;
1893 UNPROTECT_CTX(ctx
, flags
);
1896 * if there was a mapping, then we systematically remove it
1897 * at this point. Cannot be done inside critical section
1898 * because some VM function reenables interrupts.
1901 if (smpl_buf_vaddr
) pfm_remove_smpl_mapping(current
, smpl_buf_vaddr
, smpl_buf_size
);
1906 * called either on explicit close() or from exit_files().
1907 * Only the LAST user of the file gets to this point, i.e., it is
1910 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1911 * (fput()),i.e, last task to access the file. Nobody else can access the
1912 * file at this point.
1914 * When called from exit_files(), the VMA has been freed because exit_mm()
1915 * is executed before exit_files().
1917 * When called from exit_files(), the current task is not yet ZOMBIE but we
1918 * flush the PMU state to the context.
1921 pfm_close(struct inode
*inode
, struct file
*filp
)
1924 struct task_struct
*task
;
1925 struct pt_regs
*regs
;
1926 DECLARE_WAITQUEUE(wait
, current
);
1927 unsigned long flags
;
1928 unsigned long smpl_buf_size
= 0UL;
1929 void *smpl_buf_addr
= NULL
;
1930 int free_possible
= 1;
1931 int state
, is_system
;
1933 DPRINT(("pfm_close called private=%p\n", filp
->private_data
));
1935 if (PFM_IS_FILE(filp
) == 0) {
1936 DPRINT(("bad magic\n"));
1940 ctx
= (pfm_context_t
*)filp
->private_data
;
1942 printk(KERN_ERR
"perfmon: pfm_close: NULL ctx [%d]\n", current
->pid
);
1946 PROTECT_CTX(ctx
, flags
);
1948 state
= ctx
->ctx_state
;
1949 is_system
= ctx
->ctx_fl_system
;
1951 task
= PFM_CTX_TASK(ctx
);
1952 regs
= task_pt_regs(task
);
1954 DPRINT(("ctx_state=%d is_current=%d\n",
1956 task
== current
? 1 : 0));
1959 * if task == current, then pfm_flush() unloaded the context
1961 if (state
== PFM_CTX_UNLOADED
) goto doit
;
1964 * context is loaded/masked and task != current, we need to
1965 * either force an unload or go zombie
1969 * The task is currently blocked or will block after an overflow.
1970 * we must force it to wakeup to get out of the
1971 * MASKED state and transition to the unloaded state by itself.
1973 * This situation is only possible for per-task mode
1975 if (state
== PFM_CTX_MASKED
&& CTX_OVFL_NOBLOCK(ctx
) == 0) {
1978 * set a "partial" zombie state to be checked
1979 * upon return from down() in pfm_handle_work().
1981 * We cannot use the ZOMBIE state, because it is checked
1982 * by pfm_load_regs() which is called upon wakeup from down().
1983 * In such case, it would free the context and then we would
1984 * return to pfm_handle_work() which would access the
1985 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1986 * but visible to pfm_handle_work().
1988 * For some window of time, we have a zombie context with
1989 * ctx_state = MASKED and not ZOMBIE
1991 ctx
->ctx_fl_going_zombie
= 1;
1994 * force task to wake up from MASKED state
1996 complete(&ctx
->ctx_restart_done
);
1998 DPRINT(("waking up ctx_state=%d\n", state
));
2001 * put ourself to sleep waiting for the other
2002 * task to report completion
2004 * the context is protected by mutex, therefore there
2005 * is no risk of being notified of completion before
2006 * begin actually on the waitq.
2008 set_current_state(TASK_INTERRUPTIBLE
);
2009 add_wait_queue(&ctx
->ctx_zombieq
, &wait
);
2011 UNPROTECT_CTX(ctx
, flags
);
2014 * XXX: check for signals :
2015 * - ok for explicit close
2016 * - not ok when coming from exit_files()
2021 PROTECT_CTX(ctx
, flags
);
2024 remove_wait_queue(&ctx
->ctx_zombieq
, &wait
);
2025 set_current_state(TASK_RUNNING
);
2028 * context is unloaded at this point
2030 DPRINT(("after zombie wakeup ctx_state=%d for\n", state
));
2032 else if (task
!= current
) {
2035 * switch context to zombie state
2037 ctx
->ctx_state
= PFM_CTX_ZOMBIE
;
2039 DPRINT(("zombie ctx for [%d]\n", task
->pid
));
2041 * cannot free the context on the spot. deferred until
2042 * the task notices the ZOMBIE state
2046 pfm_context_unload(ctx
, NULL
, 0, regs
);
2051 /* reload state, may have changed during opening of critical section */
2052 state
= ctx
->ctx_state
;
2055 * the context is still attached to a task (possibly current)
2056 * we cannot destroy it right now
2060 * we must free the sampling buffer right here because
2061 * we cannot rely on it being cleaned up later by the
2062 * monitored task. It is not possible to free vmalloc'ed
2063 * memory in pfm_load_regs(). Instead, we remove the buffer
2064 * now. should there be subsequent PMU overflow originally
2065 * meant for sampling, the will be converted to spurious
2066 * and that's fine because the monitoring tools is gone anyway.
2068 if (ctx
->ctx_smpl_hdr
) {
2069 smpl_buf_addr
= ctx
->ctx_smpl_hdr
;
2070 smpl_buf_size
= ctx
->ctx_smpl_size
;
2071 /* no more sampling */
2072 ctx
->ctx_smpl_hdr
= NULL
;
2073 ctx
->ctx_fl_is_sampling
= 0;
2076 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2082 if (smpl_buf_addr
) pfm_exit_smpl_buffer(ctx
->ctx_buf_fmt
);
2085 * UNLOADED that the session has already been unreserved.
2087 if (state
== PFM_CTX_ZOMBIE
) {
2088 pfm_unreserve_session(ctx
, ctx
->ctx_fl_system
, ctx
->ctx_cpu
);
2092 * disconnect file descriptor from context must be done
2095 filp
->private_data
= NULL
;
2098 * if we free on the spot, the context is now completely unreacheable
2099 * from the callers side. The monitored task side is also cut, so we
2102 * If we have a deferred free, only the caller side is disconnected.
2104 UNPROTECT_CTX(ctx
, flags
);
2107 * All memory free operations (especially for vmalloc'ed memory)
2108 * MUST be done with interrupts ENABLED.
2110 if (smpl_buf_addr
) pfm_rvfree(smpl_buf_addr
, smpl_buf_size
);
2113 * return the memory used by the context
2115 if (free_possible
) pfm_context_free(ctx
);
2121 pfm_no_open(struct inode
*irrelevant
, struct file
*dontcare
)
2123 DPRINT(("pfm_no_open called\n"));
2129 static const struct file_operations pfm_file_ops
= {
2130 .llseek
= no_llseek
,
2135 .open
= pfm_no_open
, /* special open code to disallow open via /proc */
2136 .fasync
= pfm_fasync
,
2137 .release
= pfm_close
,
2142 pfmfs_delete_dentry(struct dentry
*dentry
)
2147 static struct dentry_operations pfmfs_dentry_operations
= {
2148 .d_delete
= pfmfs_delete_dentry
,
2153 pfm_alloc_fd(struct file
**cfile
)
2156 struct file
*file
= NULL
;
2157 struct inode
* inode
;
2161 fd
= get_unused_fd();
2162 if (fd
< 0) return -ENFILE
;
2166 file
= get_empty_filp();
2167 if (!file
) goto out
;
2170 * allocate a new inode
2172 inode
= new_inode(pfmfs_mnt
->mnt_sb
);
2173 if (!inode
) goto out
;
2175 DPRINT(("new inode ino=%ld @%p\n", inode
->i_ino
, inode
));
2177 inode
->i_mode
= S_IFCHR
|S_IRUGO
;
2178 inode
->i_uid
= current
->fsuid
;
2179 inode
->i_gid
= current
->fsgid
;
2181 sprintf(name
, "[%lu]", inode
->i_ino
);
2183 this.len
= strlen(name
);
2184 this.hash
= inode
->i_ino
;
2189 * allocate a new dcache entry
2191 file
->f_path
.dentry
= d_alloc(pfmfs_mnt
->mnt_sb
->s_root
, &this);
2192 if (!file
->f_path
.dentry
) goto out
;
2194 file
->f_path
.dentry
->d_op
= &pfmfs_dentry_operations
;
2196 d_add(file
->f_path
.dentry
, inode
);
2197 file
->f_path
.mnt
= mntget(pfmfs_mnt
);
2198 file
->f_mapping
= inode
->i_mapping
;
2200 file
->f_op
= &pfm_file_ops
;
2201 file
->f_mode
= FMODE_READ
;
2202 file
->f_flags
= O_RDONLY
;
2206 * may have to delay until context is attached?
2208 fd_install(fd
, file
);
2211 * the file structure we will use
2217 if (file
) put_filp(file
);
2223 pfm_free_fd(int fd
, struct file
*file
)
2225 struct files_struct
*files
= current
->files
;
2226 struct fdtable
*fdt
;
2229 * there ie no fd_uninstall(), so we do it here
2231 spin_lock(&files
->file_lock
);
2232 fdt
= files_fdtable(files
);
2233 rcu_assign_pointer(fdt
->fd
[fd
], NULL
);
2234 spin_unlock(&files
->file_lock
);
2242 pfm_remap_buffer(struct vm_area_struct
*vma
, unsigned long buf
, unsigned long addr
, unsigned long size
)
2244 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf
, addr
, size
));
2247 unsigned long pfn
= ia64_tpa(buf
) >> PAGE_SHIFT
;
2250 if (remap_pfn_range(vma
, addr
, pfn
, PAGE_SIZE
, PAGE_READONLY
))
2261 * allocate a sampling buffer and remaps it into the user address space of the task
2264 pfm_smpl_buffer_alloc(struct task_struct
*task
, pfm_context_t
*ctx
, unsigned long rsize
, void **user_vaddr
)
2266 struct mm_struct
*mm
= task
->mm
;
2267 struct vm_area_struct
*vma
= NULL
;
2273 * the fixed header + requested size and align to page boundary
2275 size
= PAGE_ALIGN(rsize
);
2277 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize
, size
));
2280 * check requested size to avoid Denial-of-service attacks
2281 * XXX: may have to refine this test
2282 * Check against address space limit.
2284 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2287 if (size
> task
->signal
->rlim
[RLIMIT_MEMLOCK
].rlim_cur
)
2291 * We do the easy to undo allocations first.
2293 * pfm_rvmalloc(), clears the buffer, so there is no leak
2295 smpl_buf
= pfm_rvmalloc(size
);
2296 if (smpl_buf
== NULL
) {
2297 DPRINT(("Can't allocate sampling buffer\n"));
2301 DPRINT(("smpl_buf @%p\n", smpl_buf
));
2304 vma
= kmem_cache_zalloc(vm_area_cachep
, GFP_KERNEL
);
2306 DPRINT(("Cannot allocate vma\n"));
2311 * partially initialize the vma for the sampling buffer
2314 vma
->vm_flags
= VM_READ
| VM_MAYREAD
|VM_RESERVED
;
2315 vma
->vm_page_prot
= PAGE_READONLY
; /* XXX may need to change */
2318 * Now we have everything we need and we can initialize
2319 * and connect all the data structures
2322 ctx
->ctx_smpl_hdr
= smpl_buf
;
2323 ctx
->ctx_smpl_size
= size
; /* aligned size */
2326 * Let's do the difficult operations next.
2328 * now we atomically find some area in the address space and
2329 * remap the buffer in it.
2331 down_write(&task
->mm
->mmap_sem
);
2333 /* find some free area in address space, must have mmap sem held */
2334 vma
->vm_start
= pfm_get_unmapped_area(NULL
, 0, size
, 0, MAP_PRIVATE
|MAP_ANONYMOUS
, 0);
2335 if (vma
->vm_start
== 0UL) {
2336 DPRINT(("Cannot find unmapped area for size %ld\n", size
));
2337 up_write(&task
->mm
->mmap_sem
);
2340 vma
->vm_end
= vma
->vm_start
+ size
;
2341 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2343 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size
, ctx
->ctx_smpl_hdr
, vma
->vm_start
));
2345 /* can only be applied to current task, need to have the mm semaphore held when called */
2346 if (pfm_remap_buffer(vma
, (unsigned long)smpl_buf
, vma
->vm_start
, size
)) {
2347 DPRINT(("Can't remap buffer\n"));
2348 up_write(&task
->mm
->mmap_sem
);
2353 * now insert the vma in the vm list for the process, must be
2354 * done with mmap lock held
2356 insert_vm_struct(mm
, vma
);
2358 mm
->total_vm
+= size
>> PAGE_SHIFT
;
2359 vm_stat_account(vma
->vm_mm
, vma
->vm_flags
, vma
->vm_file
,
2361 up_write(&task
->mm
->mmap_sem
);
2364 * keep track of user level virtual address
2366 ctx
->ctx_smpl_vaddr
= (void *)vma
->vm_start
;
2367 *(unsigned long *)user_vaddr
= vma
->vm_start
;
2372 kmem_cache_free(vm_area_cachep
, vma
);
2374 pfm_rvfree(smpl_buf
, size
);
2380 * XXX: do something better here
2383 pfm_bad_permissions(struct task_struct
*task
)
2385 /* inspired by ptrace_attach() */
2386 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2395 return ((current
->uid
!= task
->euid
)
2396 || (current
->uid
!= task
->suid
)
2397 || (current
->uid
!= task
->uid
)
2398 || (current
->gid
!= task
->egid
)
2399 || (current
->gid
!= task
->sgid
)
2400 || (current
->gid
!= task
->gid
)) && !capable(CAP_SYS_PTRACE
);
2404 pfarg_is_sane(struct task_struct
*task
, pfarg_context_t
*pfx
)
2410 ctx_flags
= pfx
->ctx_flags
;
2412 if (ctx_flags
& PFM_FL_SYSTEM_WIDE
) {
2415 * cannot block in this mode
2417 if (ctx_flags
& PFM_FL_NOTIFY_BLOCK
) {
2418 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2423 /* probably more to add here */
2429 pfm_setup_buffer_fmt(struct task_struct
*task
, pfm_context_t
*ctx
, unsigned int ctx_flags
,
2430 unsigned int cpu
, pfarg_context_t
*arg
)
2432 pfm_buffer_fmt_t
*fmt
= NULL
;
2433 unsigned long size
= 0UL;
2435 void *fmt_arg
= NULL
;
2437 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2439 /* invoke and lock buffer format, if found */
2440 fmt
= pfm_find_buffer_fmt(arg
->ctx_smpl_buf_id
);
2442 DPRINT(("[%d] cannot find buffer format\n", task
->pid
));
2447 * buffer argument MUST be contiguous to pfarg_context_t
2449 if (fmt
->fmt_arg_size
) fmt_arg
= PFM_CTXARG_BUF_ARG(arg
);
2451 ret
= pfm_buf_fmt_validate(fmt
, task
, ctx_flags
, cpu
, fmt_arg
);
2453 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task
->pid
, ctx_flags
, cpu
, fmt_arg
, ret
));
2455 if (ret
) goto error
;
2457 /* link buffer format and context */
2458 ctx
->ctx_buf_fmt
= fmt
;
2461 * check if buffer format wants to use perfmon buffer allocation/mapping service
2463 ret
= pfm_buf_fmt_getsize(fmt
, task
, ctx_flags
, cpu
, fmt_arg
, &size
);
2464 if (ret
) goto error
;
2468 * buffer is always remapped into the caller's address space
2470 ret
= pfm_smpl_buffer_alloc(current
, ctx
, size
, &uaddr
);
2471 if (ret
) goto error
;
2473 /* keep track of user address of buffer */
2474 arg
->ctx_smpl_vaddr
= uaddr
;
2476 ret
= pfm_buf_fmt_init(fmt
, task
, ctx
->ctx_smpl_hdr
, ctx_flags
, cpu
, fmt_arg
);
2483 pfm_reset_pmu_state(pfm_context_t
*ctx
)
2488 * install reset values for PMC.
2490 for (i
=1; PMC_IS_LAST(i
) == 0; i
++) {
2491 if (PMC_IS_IMPL(i
) == 0) continue;
2492 ctx
->ctx_pmcs
[i
] = PMC_DFL_VAL(i
);
2493 DPRINT(("pmc[%d]=0x%lx\n", i
, ctx
->ctx_pmcs
[i
]));
2496 * PMD registers are set to 0UL when the context in memset()
2500 * On context switched restore, we must restore ALL pmc and ALL pmd even
2501 * when they are not actively used by the task. In UP, the incoming process
2502 * may otherwise pick up left over PMC, PMD state from the previous process.
2503 * As opposed to PMD, stale PMC can cause harm to the incoming
2504 * process because they may change what is being measured.
2505 * Therefore, we must systematically reinstall the entire
2506 * PMC state. In SMP, the same thing is possible on the
2507 * same CPU but also on between 2 CPUs.
2509 * The problem with PMD is information leaking especially
2510 * to user level when psr.sp=0
2512 * There is unfortunately no easy way to avoid this problem
2513 * on either UP or SMP. This definitively slows down the
2514 * pfm_load_regs() function.
2518 * bitmask of all PMCs accessible to this context
2520 * PMC0 is treated differently.
2522 ctx
->ctx_all_pmcs
[0] = pmu_conf
->impl_pmcs
[0] & ~0x1;
2525 * bitmask of all PMDs that are accesible to this context
2527 ctx
->ctx_all_pmds
[0] = pmu_conf
->impl_pmds
[0];
2529 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx
->ctx_fd
, ctx
->ctx_all_pmcs
[0],ctx
->ctx_all_pmds
[0]));
2532 * useful in case of re-enable after disable
2534 ctx
->ctx_used_ibrs
[0] = 0UL;
2535 ctx
->ctx_used_dbrs
[0] = 0UL;
2539 pfm_ctx_getsize(void *arg
, size_t *sz
)
2541 pfarg_context_t
*req
= (pfarg_context_t
*)arg
;
2542 pfm_buffer_fmt_t
*fmt
;
2546 if (!pfm_uuid_cmp(req
->ctx_smpl_buf_id
, pfm_null_uuid
)) return 0;
2548 fmt
= pfm_find_buffer_fmt(req
->ctx_smpl_buf_id
);
2550 DPRINT(("cannot find buffer format\n"));
2553 /* get just enough to copy in user parameters */
2554 *sz
= fmt
->fmt_arg_size
;
2555 DPRINT(("arg_size=%lu\n", *sz
));
2563 * cannot attach if :
2565 * - task not owned by caller
2566 * - task incompatible with context mode
2569 pfm_task_incompatible(pfm_context_t
*ctx
, struct task_struct
*task
)
2572 * no kernel task or task not owner by caller
2574 if (task
->mm
== NULL
) {
2575 DPRINT(("task [%d] has not memory context (kernel thread)\n", task
->pid
));
2578 if (pfm_bad_permissions(task
)) {
2579 DPRINT(("no permission to attach to [%d]\n", task
->pid
));
2583 * cannot block in self-monitoring mode
2585 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && task
== current
) {
2586 DPRINT(("cannot load a blocking context on self for [%d]\n", task
->pid
));
2590 if (task
->exit_state
== EXIT_ZOMBIE
) {
2591 DPRINT(("cannot attach to zombie task [%d]\n", task
->pid
));
2596 * always ok for self
2598 if (task
== current
) return 0;
2600 if ((task
->state
!= TASK_STOPPED
) && (task
->state
!= TASK_TRACED
)) {
2601 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task
->pid
, task
->state
));
2605 * make sure the task is off any CPU
2607 wait_task_inactive(task
);
2609 /* more to come... */
2615 pfm_get_task(pfm_context_t
*ctx
, pid_t pid
, struct task_struct
**task
)
2617 struct task_struct
*p
= current
;
2620 /* XXX: need to add more checks here */
2621 if (pid
< 2) return -EPERM
;
2623 if (pid
!= current
->pid
) {
2625 read_lock(&tasklist_lock
);
2627 p
= find_task_by_pid(pid
);
2629 /* make sure task cannot go away while we operate on it */
2630 if (p
) get_task_struct(p
);
2632 read_unlock(&tasklist_lock
);
2634 if (p
== NULL
) return -ESRCH
;
2637 ret
= pfm_task_incompatible(ctx
, p
);
2640 } else if (p
!= current
) {
2649 pfm_context_create(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
2651 pfarg_context_t
*req
= (pfarg_context_t
*)arg
;
2656 /* let's check the arguments first */
2657 ret
= pfarg_is_sane(current
, req
);
2658 if (ret
< 0) return ret
;
2660 ctx_flags
= req
->ctx_flags
;
2664 ctx
= pfm_context_alloc();
2665 if (!ctx
) goto error
;
2667 ret
= pfm_alloc_fd(&filp
);
2668 if (ret
< 0) goto error_file
;
2670 req
->ctx_fd
= ctx
->ctx_fd
= ret
;
2673 * attach context to file
2675 filp
->private_data
= ctx
;
2678 * does the user want to sample?
2680 if (pfm_uuid_cmp(req
->ctx_smpl_buf_id
, pfm_null_uuid
)) {
2681 ret
= pfm_setup_buffer_fmt(current
, ctx
, ctx_flags
, 0, req
);
2682 if (ret
) goto buffer_error
;
2686 * init context protection lock
2688 spin_lock_init(&ctx
->ctx_lock
);
2691 * context is unloaded
2693 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
2696 * initialization of context's flags
2698 ctx
->ctx_fl_block
= (ctx_flags
& PFM_FL_NOTIFY_BLOCK
) ? 1 : 0;
2699 ctx
->ctx_fl_system
= (ctx_flags
& PFM_FL_SYSTEM_WIDE
) ? 1: 0;
2700 ctx
->ctx_fl_is_sampling
= ctx
->ctx_buf_fmt
? 1 : 0; /* assume record() is defined */
2701 ctx
->ctx_fl_no_msg
= (ctx_flags
& PFM_FL_OVFL_NO_MSG
) ? 1: 0;
2703 * will move to set properties
2704 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2708 * init restart semaphore to locked
2710 init_completion(&ctx
->ctx_restart_done
);
2713 * activation is used in SMP only
2715 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
2716 SET_LAST_CPU(ctx
, -1);
2719 * initialize notification message queue
2721 ctx
->ctx_msgq_head
= ctx
->ctx_msgq_tail
= 0;
2722 init_waitqueue_head(&ctx
->ctx_msgq_wait
);
2723 init_waitqueue_head(&ctx
->ctx_zombieq
);
2725 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2730 ctx
->ctx_fl_excl_idle
,
2735 * initialize soft PMU state
2737 pfm_reset_pmu_state(ctx
);
2742 pfm_free_fd(ctx
->ctx_fd
, filp
);
2744 if (ctx
->ctx_buf_fmt
) {
2745 pfm_buf_fmt_exit(ctx
->ctx_buf_fmt
, current
, NULL
, regs
);
2748 pfm_context_free(ctx
);
2754 static inline unsigned long
2755 pfm_new_counter_value (pfm_counter_t
*reg
, int is_long_reset
)
2757 unsigned long val
= is_long_reset
? reg
->long_reset
: reg
->short_reset
;
2758 unsigned long new_seed
, old_seed
= reg
->seed
, mask
= reg
->mask
;
2759 extern unsigned long carta_random32 (unsigned long seed
);
2761 if (reg
->flags
& PFM_REGFL_RANDOM
) {
2762 new_seed
= carta_random32(old_seed
);
2763 val
-= (old_seed
& mask
); /* counter values are negative numbers! */
2764 if ((mask
>> 32) != 0)
2765 /* construct a full 64-bit random value: */
2766 new_seed
|= carta_random32(old_seed
>> 32) << 32;
2767 reg
->seed
= new_seed
;
2774 pfm_reset_regs_masked(pfm_context_t
*ctx
, unsigned long *ovfl_regs
, int is_long_reset
)
2776 unsigned long mask
= ovfl_regs
[0];
2777 unsigned long reset_others
= 0UL;
2782 * now restore reset value on sampling overflowed counters
2784 mask
>>= PMU_FIRST_COUNTER
;
2785 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
2787 if ((mask
& 0x1UL
) == 0UL) continue;
2789 ctx
->ctx_pmds
[i
].val
= val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2790 reset_others
|= ctx
->ctx_pmds
[i
].reset_pmds
[0];
2792 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset
? "long" : "short", i
, val
));
2796 * Now take care of resetting the other registers
2798 for(i
= 0; reset_others
; i
++, reset_others
>>= 1) {
2800 if ((reset_others
& 0x1) == 0) continue;
2802 ctx
->ctx_pmds
[i
].val
= val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2804 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2805 is_long_reset
? "long" : "short", i
, val
));
2810 pfm_reset_regs(pfm_context_t
*ctx
, unsigned long *ovfl_regs
, int is_long_reset
)
2812 unsigned long mask
= ovfl_regs
[0];
2813 unsigned long reset_others
= 0UL;
2817 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs
[0], is_long_reset
));
2819 if (ctx
->ctx_state
== PFM_CTX_MASKED
) {
2820 pfm_reset_regs_masked(ctx
, ovfl_regs
, is_long_reset
);
2825 * now restore reset value on sampling overflowed counters
2827 mask
>>= PMU_FIRST_COUNTER
;
2828 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
2830 if ((mask
& 0x1UL
) == 0UL) continue;
2832 val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2833 reset_others
|= ctx
->ctx_pmds
[i
].reset_pmds
[0];
2835 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset
? "long" : "short", i
, val
));
2837 pfm_write_soft_counter(ctx
, i
, val
);
2841 * Now take care of resetting the other registers
2843 for(i
= 0; reset_others
; i
++, reset_others
>>= 1) {
2845 if ((reset_others
& 0x1) == 0) continue;
2847 val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2849 if (PMD_IS_COUNTING(i
)) {
2850 pfm_write_soft_counter(ctx
, i
, val
);
2852 ia64_set_pmd(i
, val
);
2854 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2855 is_long_reset
? "long" : "short", i
, val
));
2861 pfm_write_pmcs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
2863 struct task_struct
*task
;
2864 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
2865 unsigned long value
, pmc_pm
;
2866 unsigned long smpl_pmds
, reset_pmds
, impl_pmds
;
2867 unsigned int cnum
, reg_flags
, flags
, pmc_type
;
2868 int i
, can_access_pmu
= 0, is_loaded
, is_system
, expert_mode
;
2869 int is_monitor
, is_counting
, state
;
2871 pfm_reg_check_t wr_func
;
2872 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2874 state
= ctx
->ctx_state
;
2875 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
2876 is_system
= ctx
->ctx_fl_system
;
2877 task
= ctx
->ctx_task
;
2878 impl_pmds
= pmu_conf
->impl_pmds
[0];
2880 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
2884 * In system wide and when the context is loaded, access can only happen
2885 * when the caller is running on the CPU being monitored by the session.
2886 * It does not have to be the owner (ctx_task) of the context per se.
2888 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
2889 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
2892 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
2894 expert_mode
= pfm_sysctl
.expert_mode
;
2896 for (i
= 0; i
< count
; i
++, req
++) {
2898 cnum
= req
->reg_num
;
2899 reg_flags
= req
->reg_flags
;
2900 value
= req
->reg_value
;
2901 smpl_pmds
= req
->reg_smpl_pmds
[0];
2902 reset_pmds
= req
->reg_reset_pmds
[0];
2906 if (cnum
>= PMU_MAX_PMCS
) {
2907 DPRINT(("pmc%u is invalid\n", cnum
));
2911 pmc_type
= pmu_conf
->pmc_desc
[cnum
].type
;
2912 pmc_pm
= (value
>> pmu_conf
->pmc_desc
[cnum
].pm_pos
) & 0x1;
2913 is_counting
= (pmc_type
& PFM_REG_COUNTING
) == PFM_REG_COUNTING
? 1 : 0;
2914 is_monitor
= (pmc_type
& PFM_REG_MONITOR
) == PFM_REG_MONITOR
? 1 : 0;
2917 * we reject all non implemented PMC as well
2918 * as attempts to modify PMC[0-3] which are used
2919 * as status registers by the PMU
2921 if ((pmc_type
& PFM_REG_IMPL
) == 0 || (pmc_type
& PFM_REG_CONTROL
) == PFM_REG_CONTROL
) {
2922 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum
, pmc_type
));
2925 wr_func
= pmu_conf
->pmc_desc
[cnum
].write_check
;
2927 * If the PMC is a monitor, then if the value is not the default:
2928 * - system-wide session: PMCx.pm=1 (privileged monitor)
2929 * - per-task : PMCx.pm=0 (user monitor)
2931 if (is_monitor
&& value
!= PMC_DFL_VAL(cnum
) && is_system
^ pmc_pm
) {
2932 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2941 * enforce generation of overflow interrupt. Necessary on all
2944 value
|= 1 << PMU_PMC_OI
;
2946 if (reg_flags
& PFM_REGFL_OVFL_NOTIFY
) {
2947 flags
|= PFM_REGFL_OVFL_NOTIFY
;
2950 if (reg_flags
& PFM_REGFL_RANDOM
) flags
|= PFM_REGFL_RANDOM
;
2952 /* verify validity of smpl_pmds */
2953 if ((smpl_pmds
& impl_pmds
) != smpl_pmds
) {
2954 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds
, cnum
));
2958 /* verify validity of reset_pmds */
2959 if ((reset_pmds
& impl_pmds
) != reset_pmds
) {
2960 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds
, cnum
));
2964 if (reg_flags
& (PFM_REGFL_OVFL_NOTIFY
|PFM_REGFL_RANDOM
)) {
2965 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum
));
2968 /* eventid on non-counting monitors are ignored */
2972 * execute write checker, if any
2974 if (likely(expert_mode
== 0 && wr_func
)) {
2975 ret
= (*wr_func
)(task
, ctx
, cnum
, &value
, regs
);
2976 if (ret
) goto error
;
2981 * no error on this register
2983 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
2986 * Now we commit the changes to the software state
2990 * update overflow information
2994 * full flag update each time a register is programmed
2996 ctx
->ctx_pmds
[cnum
].flags
= flags
;
2998 ctx
->ctx_pmds
[cnum
].reset_pmds
[0] = reset_pmds
;
2999 ctx
->ctx_pmds
[cnum
].smpl_pmds
[0] = smpl_pmds
;
3000 ctx
->ctx_pmds
[cnum
].eventid
= req
->reg_smpl_eventid
;
3003 * Mark all PMDS to be accessed as used.
3005 * We do not keep track of PMC because we have to
3006 * systematically restore ALL of them.
3008 * We do not update the used_monitors mask, because
3009 * if we have not programmed them, then will be in
3010 * a quiescent state, therefore we will not need to
3011 * mask/restore then when context is MASKED.
3013 CTX_USED_PMD(ctx
, reset_pmds
);
3014 CTX_USED_PMD(ctx
, smpl_pmds
);
3016 * make sure we do not try to reset on
3017 * restart because we have established new values
3019 if (state
== PFM_CTX_MASKED
) ctx
->ctx_ovfl_regs
[0] &= ~1UL << cnum
;
3022 * Needed in case the user does not initialize the equivalent
3023 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3024 * possible leak here.
3026 CTX_USED_PMD(ctx
, pmu_conf
->pmc_desc
[cnum
].dep_pmd
[0]);
3029 * keep track of the monitor PMC that we are using.
3030 * we save the value of the pmc in ctx_pmcs[] and if
3031 * the monitoring is not stopped for the context we also
3032 * place it in the saved state area so that it will be
3033 * picked up later by the context switch code.
3035 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3037 * The value in th_pmcs[] may be modified on overflow, i.e., when
3038 * monitoring needs to be stopped.
3040 if (is_monitor
) CTX_USED_MONITOR(ctx
, 1UL << cnum
);
3043 * update context state
3045 ctx
->ctx_pmcs
[cnum
] = value
;
3049 * write thread state
3051 if (is_system
== 0) ctx
->th_pmcs
[cnum
] = value
;
3054 * write hardware register if we can
3056 if (can_access_pmu
) {
3057 ia64_set_pmc(cnum
, value
);
3062 * per-task SMP only here
3064 * we are guaranteed that the task is not running on the other CPU,
3065 * we indicate that this PMD will need to be reloaded if the task
3066 * is rescheduled on the CPU it ran last on.
3068 ctx
->ctx_reload_pmcs
[0] |= 1UL << cnum
;
3073 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",
3079 ctx
->ctx_all_pmcs
[0],
3080 ctx
->ctx_used_pmds
[0],
3081 ctx
->ctx_pmds
[cnum
].eventid
,
3084 ctx
->ctx_reload_pmcs
[0],
3085 ctx
->ctx_used_monitors
[0],
3086 ctx
->ctx_ovfl_regs
[0]));
3090 * make sure the changes are visible
3092 if (can_access_pmu
) ia64_srlz_d();
3096 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3101 pfm_write_pmds(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3103 struct task_struct
*task
;
3104 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
3105 unsigned long value
, hw_value
, ovfl_mask
;
3107 int i
, can_access_pmu
= 0, state
;
3108 int is_counting
, is_loaded
, is_system
, expert_mode
;
3110 pfm_reg_check_t wr_func
;
3113 state
= ctx
->ctx_state
;
3114 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3115 is_system
= ctx
->ctx_fl_system
;
3116 ovfl_mask
= pmu_conf
->ovfl_val
;
3117 task
= ctx
->ctx_task
;
3119 if (unlikely(state
== PFM_CTX_ZOMBIE
)) return -EINVAL
;
3122 * on both UP and SMP, we can only write to the PMC when the task is
3123 * the owner of the local PMU.
3125 if (likely(is_loaded
)) {
3127 * In system wide and when the context is loaded, access can only happen
3128 * when the caller is running on the CPU being monitored by the session.
3129 * It does not have to be the owner (ctx_task) of the context per se.
3131 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3132 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3135 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3137 expert_mode
= pfm_sysctl
.expert_mode
;
3139 for (i
= 0; i
< count
; i
++, req
++) {
3141 cnum
= req
->reg_num
;
3142 value
= req
->reg_value
;
3144 if (!PMD_IS_IMPL(cnum
)) {
3145 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum
));
3148 is_counting
= PMD_IS_COUNTING(cnum
);
3149 wr_func
= pmu_conf
->pmd_desc
[cnum
].write_check
;
3152 * execute write checker, if any
3154 if (unlikely(expert_mode
== 0 && wr_func
)) {
3155 unsigned long v
= value
;
3157 ret
= (*wr_func
)(task
, ctx
, cnum
, &v
, regs
);
3158 if (ret
) goto abort_mission
;
3165 * no error on this register
3167 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
3170 * now commit changes to software state
3175 * update virtualized (64bits) counter
3179 * write context state
3181 ctx
->ctx_pmds
[cnum
].lval
= value
;
3184 * when context is load we use the split value
3187 hw_value
= value
& ovfl_mask
;
3188 value
= value
& ~ovfl_mask
;
3192 * update reset values (not just for counters)
3194 ctx
->ctx_pmds
[cnum
].long_reset
= req
->reg_long_reset
;
3195 ctx
->ctx_pmds
[cnum
].short_reset
= req
->reg_short_reset
;
3198 * update randomization parameters (not just for counters)
3200 ctx
->ctx_pmds
[cnum
].seed
= req
->reg_random_seed
;
3201 ctx
->ctx_pmds
[cnum
].mask
= req
->reg_random_mask
;
3204 * update context value
3206 ctx
->ctx_pmds
[cnum
].val
= value
;
3209 * Keep track of what we use
3211 * We do not keep track of PMC because we have to
3212 * systematically restore ALL of them.
3214 CTX_USED_PMD(ctx
, PMD_PMD_DEP(cnum
));
3217 * mark this PMD register used as well
3219 CTX_USED_PMD(ctx
, RDEP(cnum
));
3222 * make sure we do not try to reset on
3223 * restart because we have established new values
3225 if (is_counting
&& state
== PFM_CTX_MASKED
) {
3226 ctx
->ctx_ovfl_regs
[0] &= ~1UL << cnum
;
3231 * write thread state
3233 if (is_system
== 0) ctx
->th_pmds
[cnum
] = hw_value
;
3236 * write hardware register if we can
3238 if (can_access_pmu
) {
3239 ia64_set_pmd(cnum
, hw_value
);
3243 * we are guaranteed that the task is not running on the other CPU,
3244 * we indicate that this PMD will need to be reloaded if the task
3245 * is rescheduled on the CPU it ran last on.
3247 ctx
->ctx_reload_pmds
[0] |= 1UL << cnum
;
3252 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3253 "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",
3259 ctx
->ctx_pmds
[cnum
].val
,
3260 ctx
->ctx_pmds
[cnum
].short_reset
,
3261 ctx
->ctx_pmds
[cnum
].long_reset
,
3262 PMC_OVFL_NOTIFY(ctx
, cnum
) ? 'Y':'N',
3263 ctx
->ctx_pmds
[cnum
].seed
,
3264 ctx
->ctx_pmds
[cnum
].mask
,
3265 ctx
->ctx_used_pmds
[0],
3266 ctx
->ctx_pmds
[cnum
].reset_pmds
[0],
3267 ctx
->ctx_reload_pmds
[0],
3268 ctx
->ctx_all_pmds
[0],
3269 ctx
->ctx_ovfl_regs
[0]));
3273 * make changes visible
3275 if (can_access_pmu
) ia64_srlz_d();
3281 * for now, we have only one possibility for error
3283 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3288 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3289 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3290 * interrupt is delivered during the call, it will be kept pending until we leave, making
3291 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3292 * guaranteed to return consistent data to the user, it may simply be old. It is not
3293 * trivial to treat the overflow while inside the call because you may end up in
3294 * some module sampling buffer code causing deadlocks.
3297 pfm_read_pmds(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3299 struct task_struct
*task
;
3300 unsigned long val
= 0UL, lval
, ovfl_mask
, sval
;
3301 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
3302 unsigned int cnum
, reg_flags
= 0;
3303 int i
, can_access_pmu
= 0, state
;
3304 int is_loaded
, is_system
, is_counting
, expert_mode
;
3306 pfm_reg_check_t rd_func
;
3309 * access is possible when loaded only for
3310 * self-monitoring tasks or in UP mode
3313 state
= ctx
->ctx_state
;
3314 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3315 is_system
= ctx
->ctx_fl_system
;
3316 ovfl_mask
= pmu_conf
->ovfl_val
;
3317 task
= ctx
->ctx_task
;
3319 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
3321 if (likely(is_loaded
)) {
3323 * In system wide and when the context is loaded, access can only happen
3324 * when the caller is running on the CPU being monitored by the session.
3325 * It does not have to be the owner (ctx_task) of the context per se.
3327 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3328 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3332 * this can be true when not self-monitoring only in UP
3334 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3336 if (can_access_pmu
) ia64_srlz_d();
3338 expert_mode
= pfm_sysctl
.expert_mode
;
3340 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3346 * on both UP and SMP, we can only read the PMD from the hardware register when
3347 * the task is the owner of the local PMU.
3350 for (i
= 0; i
< count
; i
++, req
++) {
3352 cnum
= req
->reg_num
;
3353 reg_flags
= req
->reg_flags
;
3355 if (unlikely(!PMD_IS_IMPL(cnum
))) goto error
;
3357 * we can only read the register that we use. That includes
3358 * the one we explicitely initialize AND the one we want included
3359 * in the sampling buffer (smpl_regs).
3361 * Having this restriction allows optimization in the ctxsw routine
3362 * without compromising security (leaks)
3364 if (unlikely(!CTX_IS_USED_PMD(ctx
, cnum
))) goto error
;
3366 sval
= ctx
->ctx_pmds
[cnum
].val
;
3367 lval
= ctx
->ctx_pmds
[cnum
].lval
;
3368 is_counting
= PMD_IS_COUNTING(cnum
);
3371 * If the task is not the current one, then we check if the
3372 * PMU state is still in the local live register due to lazy ctxsw.
3373 * If true, then we read directly from the registers.
3375 if (can_access_pmu
){
3376 val
= ia64_get_pmd(cnum
);
3379 * context has been saved
3380 * if context is zombie, then task does not exist anymore.
3381 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3383 val
= is_loaded
? ctx
->th_pmds
[cnum
] : 0UL;
3385 rd_func
= pmu_conf
->pmd_desc
[cnum
].read_check
;
3389 * XXX: need to check for overflow when loaded
3396 * execute read checker, if any
3398 if (unlikely(expert_mode
== 0 && rd_func
)) {
3399 unsigned long v
= val
;
3400 ret
= (*rd_func
)(ctx
->ctx_task
, ctx
, cnum
, &v
, regs
);
3401 if (ret
) goto error
;
3406 PFM_REG_RETFLAG_SET(reg_flags
, 0);
3408 DPRINT(("pmd[%u]=0x%lx\n", cnum
, val
));
3411 * update register return value, abort all if problem during copy.
3412 * we only modify the reg_flags field. no check mode is fine because
3413 * access has been verified upfront in sys_perfmonctl().
3415 req
->reg_value
= val
;
3416 req
->reg_flags
= reg_flags
;
3417 req
->reg_last_reset_val
= lval
;
3423 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3428 pfm_mod_write_pmcs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3432 if (req
== NULL
) return -EINVAL
;
3434 ctx
= GET_PMU_CTX();
3436 if (ctx
== NULL
) return -EINVAL
;
3439 * for now limit to current task, which is enough when calling
3440 * from overflow handler
3442 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3444 return pfm_write_pmcs(ctx
, req
, nreq
, regs
);
3446 EXPORT_SYMBOL(pfm_mod_write_pmcs
);
3449 pfm_mod_read_pmds(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3453 if (req
== NULL
) return -EINVAL
;
3455 ctx
= GET_PMU_CTX();
3457 if (ctx
== NULL
) return -EINVAL
;
3460 * for now limit to current task, which is enough when calling
3461 * from overflow handler
3463 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3465 return pfm_read_pmds(ctx
, req
, nreq
, regs
);
3467 EXPORT_SYMBOL(pfm_mod_read_pmds
);
3470 * Only call this function when a process it trying to
3471 * write the debug registers (reading is always allowed)
3474 pfm_use_debug_registers(struct task_struct
*task
)
3476 pfm_context_t
*ctx
= task
->thread
.pfm_context
;
3477 unsigned long flags
;
3480 if (pmu_conf
->use_rr_dbregs
== 0) return 0;
3482 DPRINT(("called for [%d]\n", task
->pid
));
3487 if (task
->thread
.flags
& IA64_THREAD_DBG_VALID
) return 0;
3490 * Even on SMP, we do not need to use an atomic here because
3491 * the only way in is via ptrace() and this is possible only when the
3492 * process is stopped. Even in the case where the ctxsw out is not totally
3493 * completed by the time we come here, there is no way the 'stopped' process
3494 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3495 * So this is always safe.
3497 if (ctx
&& ctx
->ctx_fl_using_dbreg
== 1) return -1;
3502 * We cannot allow setting breakpoints when system wide monitoring
3503 * sessions are using the debug registers.
3505 if (pfm_sessions
.pfs_sys_use_dbregs
> 0)
3508 pfm_sessions
.pfs_ptrace_use_dbregs
++;
3510 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3511 pfm_sessions
.pfs_ptrace_use_dbregs
,
3512 pfm_sessions
.pfs_sys_use_dbregs
,
3521 * This function is called for every task that exits with the
3522 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3523 * able to use the debug registers for debugging purposes via
3524 * ptrace(). Therefore we know it was not using them for
3525 * perfmormance monitoring, so we only decrement the number
3526 * of "ptraced" debug register users to keep the count up to date
3529 pfm_release_debug_registers(struct task_struct
*task
)
3531 unsigned long flags
;
3534 if (pmu_conf
->use_rr_dbregs
== 0) return 0;
3537 if (pfm_sessions
.pfs_ptrace_use_dbregs
== 0) {
3538 printk(KERN_ERR
"perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task
->pid
);
3541 pfm_sessions
.pfs_ptrace_use_dbregs
--;
3550 pfm_restart(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3552 struct task_struct
*task
;
3553 pfm_buffer_fmt_t
*fmt
;
3554 pfm_ovfl_ctrl_t rst_ctrl
;
3555 int state
, is_system
;
3558 state
= ctx
->ctx_state
;
3559 fmt
= ctx
->ctx_buf_fmt
;
3560 is_system
= ctx
->ctx_fl_system
;
3561 task
= PFM_CTX_TASK(ctx
);
3564 case PFM_CTX_MASKED
:
3566 case PFM_CTX_LOADED
:
3567 if (CTX_HAS_SMPL(ctx
) && fmt
->fmt_restart_active
) break;
3569 case PFM_CTX_UNLOADED
:
3570 case PFM_CTX_ZOMBIE
:
3571 DPRINT(("invalid state=%d\n", state
));
3574 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state
));
3579 * In system wide and when the context is loaded, access can only happen
3580 * when the caller is running on the CPU being monitored by the session.
3581 * It does not have to be the owner (ctx_task) of the context per se.
3583 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
3584 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3589 if (unlikely(task
== NULL
)) {
3590 printk(KERN_ERR
"perfmon: [%d] pfm_restart no task\n", current
->pid
);
3594 if (task
== current
|| is_system
) {
3596 fmt
= ctx
->ctx_buf_fmt
;
3598 DPRINT(("restarting self %d ovfl=0x%lx\n",
3600 ctx
->ctx_ovfl_regs
[0]));
3602 if (CTX_HAS_SMPL(ctx
)) {
3604 prefetch(ctx
->ctx_smpl_hdr
);
3606 rst_ctrl
.bits
.mask_monitoring
= 0;
3607 rst_ctrl
.bits
.reset_ovfl_pmds
= 0;
3609 if (state
== PFM_CTX_LOADED
)
3610 ret
= pfm_buf_fmt_restart_active(fmt
, task
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
3612 ret
= pfm_buf_fmt_restart(fmt
, task
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
3614 rst_ctrl
.bits
.mask_monitoring
= 0;
3615 rst_ctrl
.bits
.reset_ovfl_pmds
= 1;
3619 if (rst_ctrl
.bits
.reset_ovfl_pmds
)
3620 pfm_reset_regs(ctx
, ctx
->ctx_ovfl_regs
, PFM_PMD_LONG_RESET
);
3622 if (rst_ctrl
.bits
.mask_monitoring
== 0) {
3623 DPRINT(("resuming monitoring for [%d]\n", task
->pid
));
3625 if (state
== PFM_CTX_MASKED
) pfm_restore_monitoring(task
);
3627 DPRINT(("keeping monitoring stopped for [%d]\n", task
->pid
));
3629 // cannot use pfm_stop_monitoring(task, regs);
3633 * clear overflowed PMD mask to remove any stale information
3635 ctx
->ctx_ovfl_regs
[0] = 0UL;
3638 * back to LOADED state
3640 ctx
->ctx_state
= PFM_CTX_LOADED
;
3643 * XXX: not really useful for self monitoring
3645 ctx
->ctx_fl_can_restart
= 0;
3651 * restart another task
3655 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3656 * one is seen by the task.
3658 if (state
== PFM_CTX_MASKED
) {
3659 if (ctx
->ctx_fl_can_restart
== 0) return -EINVAL
;
3661 * will prevent subsequent restart before this one is
3662 * seen by other task
3664 ctx
->ctx_fl_can_restart
= 0;
3668 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3669 * the task is blocked or on its way to block. That's the normal
3670 * restart path. If the monitoring is not masked, then the task
3671 * can be actively monitoring and we cannot directly intervene.
3672 * Therefore we use the trap mechanism to catch the task and
3673 * force it to reset the buffer/reset PMDs.
3675 * if non-blocking, then we ensure that the task will go into
3676 * pfm_handle_work() before returning to user mode.
3678 * We cannot explicitely reset another task, it MUST always
3679 * be done by the task itself. This works for system wide because
3680 * the tool that is controlling the session is logically doing
3681 * "self-monitoring".
3683 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && state
== PFM_CTX_MASKED
) {
3684 DPRINT(("unblocking [%d] \n", task
->pid
));
3685 complete(&ctx
->ctx_restart_done
);
3687 DPRINT(("[%d] armed exit trap\n", task
->pid
));
3689 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_RESET
;
3691 PFM_SET_WORK_PENDING(task
, 1);
3693 pfm_set_task_notify(task
);
3696 * XXX: send reschedule if task runs on another CPU
3703 pfm_debug(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3705 unsigned int m
= *(unsigned int *)arg
;
3707 pfm_sysctl
.debug
= m
== 0 ? 0 : 1;
3709 printk(KERN_INFO
"perfmon debugging %s (timing reset)\n", pfm_sysctl
.debug
? "on" : "off");
3712 memset(pfm_stats
, 0, sizeof(pfm_stats
));
3713 for(m
=0; m
< NR_CPUS
; m
++) pfm_stats
[m
].pfm_ovfl_intr_cycles_min
= ~0UL;
3719 * arg can be NULL and count can be zero for this function
3722 pfm_write_ibr_dbr(int mode
, pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3724 struct thread_struct
*thread
= NULL
;
3725 struct task_struct
*task
;
3726 pfarg_dbreg_t
*req
= (pfarg_dbreg_t
*)arg
;
3727 unsigned long flags
;
3732 int i
, can_access_pmu
= 0;
3733 int is_system
, is_loaded
;
3735 if (pmu_conf
->use_rr_dbregs
== 0) return -EINVAL
;
3737 state
= ctx
->ctx_state
;
3738 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3739 is_system
= ctx
->ctx_fl_system
;
3740 task
= ctx
->ctx_task
;
3742 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
3745 * on both UP and SMP, we can only write to the PMC when the task is
3746 * the owner of the local PMU.
3749 thread
= &task
->thread
;
3751 * In system wide and when the context is loaded, access can only happen
3752 * when the caller is running on the CPU being monitored by the session.
3753 * It does not have to be the owner (ctx_task) of the context per se.
3755 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3756 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3759 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3763 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3764 * ensuring that no real breakpoint can be installed via this call.
3766 * IMPORTANT: regs can be NULL in this function
3769 first_time
= ctx
->ctx_fl_using_dbreg
== 0;
3772 * don't bother if we are loaded and task is being debugged
3774 if (is_loaded
&& (thread
->flags
& IA64_THREAD_DBG_VALID
) != 0) {
3775 DPRINT(("debug registers already in use for [%d]\n", task
->pid
));
3780 * check for debug registers in system wide mode
3782 * If though a check is done in pfm_context_load(),
3783 * we must repeat it here, in case the registers are
3784 * written after the context is loaded
3789 if (first_time
&& is_system
) {
3790 if (pfm_sessions
.pfs_ptrace_use_dbregs
)
3793 pfm_sessions
.pfs_sys_use_dbregs
++;
3798 if (ret
!= 0) return ret
;
3801 * mark ourself as user of the debug registers for
3804 ctx
->ctx_fl_using_dbreg
= 1;
3807 * clear hardware registers to make sure we don't
3808 * pick up stale state.
3810 * for a system wide session, we do not use
3811 * thread.dbr, thread.ibr because this process
3812 * never leaves the current CPU and the state
3813 * is shared by all processes running on it
3815 if (first_time
&& can_access_pmu
) {
3816 DPRINT(("[%d] clearing ibrs, dbrs\n", task
->pid
));
3817 for (i
=0; i
< pmu_conf
->num_ibrs
; i
++) {
3818 ia64_set_ibr(i
, 0UL);
3819 ia64_dv_serialize_instruction();
3822 for (i
=0; i
< pmu_conf
->num_dbrs
; i
++) {
3823 ia64_set_dbr(i
, 0UL);
3824 ia64_dv_serialize_data();
3830 * Now install the values into the registers
3832 for (i
= 0; i
< count
; i
++, req
++) {
3834 rnum
= req
->dbreg_num
;
3835 dbreg
.val
= req
->dbreg_value
;
3839 if ((mode
== PFM_CODE_RR
&& rnum
>= PFM_NUM_IBRS
) || ((mode
== PFM_DATA_RR
) && rnum
>= PFM_NUM_DBRS
)) {
3840 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3841 rnum
, dbreg
.val
, mode
, i
, count
));
3847 * make sure we do not install enabled breakpoint
3850 if (mode
== PFM_CODE_RR
)
3851 dbreg
.ibr
.ibr_x
= 0;
3853 dbreg
.dbr
.dbr_r
= dbreg
.dbr
.dbr_w
= 0;
3856 PFM_REG_RETFLAG_SET(req
->dbreg_flags
, 0);
3859 * Debug registers, just like PMC, can only be modified
3860 * by a kernel call. Moreover, perfmon() access to those
3861 * registers are centralized in this routine. The hardware
3862 * does not modify the value of these registers, therefore,
3863 * if we save them as they are written, we can avoid having
3864 * to save them on context switch out. This is made possible
3865 * by the fact that when perfmon uses debug registers, ptrace()
3866 * won't be able to modify them concurrently.
3868 if (mode
== PFM_CODE_RR
) {
3869 CTX_USED_IBR(ctx
, rnum
);
3871 if (can_access_pmu
) {
3872 ia64_set_ibr(rnum
, dbreg
.val
);
3873 ia64_dv_serialize_instruction();
3876 ctx
->ctx_ibrs
[rnum
] = dbreg
.val
;
3878 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3879 rnum
, dbreg
.val
, ctx
->ctx_used_ibrs
[0], is_loaded
, can_access_pmu
));
3881 CTX_USED_DBR(ctx
, rnum
);
3883 if (can_access_pmu
) {
3884 ia64_set_dbr(rnum
, dbreg
.val
);
3885 ia64_dv_serialize_data();
3887 ctx
->ctx_dbrs
[rnum
] = dbreg
.val
;
3889 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3890 rnum
, dbreg
.val
, ctx
->ctx_used_dbrs
[0], is_loaded
, can_access_pmu
));
3898 * in case it was our first attempt, we undo the global modifications
3902 if (ctx
->ctx_fl_system
) {
3903 pfm_sessions
.pfs_sys_use_dbregs
--;
3906 ctx
->ctx_fl_using_dbreg
= 0;
3909 * install error return flag
3911 PFM_REG_RETFLAG_SET(req
->dbreg_flags
, PFM_REG_RETFL_EINVAL
);
3917 pfm_write_ibrs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3919 return pfm_write_ibr_dbr(PFM_CODE_RR
, ctx
, arg
, count
, regs
);
3923 pfm_write_dbrs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3925 return pfm_write_ibr_dbr(PFM_DATA_RR
, ctx
, arg
, count
, regs
);
3929 pfm_mod_write_ibrs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3933 if (req
== NULL
) return -EINVAL
;
3935 ctx
= GET_PMU_CTX();
3937 if (ctx
== NULL
) return -EINVAL
;
3940 * for now limit to current task, which is enough when calling
3941 * from overflow handler
3943 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3945 return pfm_write_ibrs(ctx
, req
, nreq
, regs
);
3947 EXPORT_SYMBOL(pfm_mod_write_ibrs
);
3950 pfm_mod_write_dbrs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3954 if (req
== NULL
) return -EINVAL
;
3956 ctx
= GET_PMU_CTX();
3958 if (ctx
== NULL
) return -EINVAL
;
3961 * for now limit to current task, which is enough when calling
3962 * from overflow handler
3964 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3966 return pfm_write_dbrs(ctx
, req
, nreq
, regs
);
3968 EXPORT_SYMBOL(pfm_mod_write_dbrs
);
3972 pfm_get_features(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3974 pfarg_features_t
*req
= (pfarg_features_t
*)arg
;
3976 req
->ft_version
= PFM_VERSION
;
3981 pfm_stop(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3983 struct pt_regs
*tregs
;
3984 struct task_struct
*task
= PFM_CTX_TASK(ctx
);
3985 int state
, is_system
;
3987 state
= ctx
->ctx_state
;
3988 is_system
= ctx
->ctx_fl_system
;
3991 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3993 if (state
== PFM_CTX_UNLOADED
) return -EINVAL
;
3996 * In system wide and when the context is loaded, access can only happen
3997 * when the caller is running on the CPU being monitored by the session.
3998 * It does not have to be the owner (ctx_task) of the context per se.
4000 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
4001 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
4004 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4005 PFM_CTX_TASK(ctx
)->pid
,
4009 * in system mode, we need to update the PMU directly
4010 * and the user level state of the caller, which may not
4011 * necessarily be the creator of the context.
4015 * Update local PMU first
4019 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) & ~IA64_DCR_PP
);
4023 * update local cpuinfo
4025 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP
);
4028 * stop monitoring, does srlz.i
4033 * stop monitoring in the caller
4035 ia64_psr(regs
)->pp
= 0;
4043 if (task
== current
) {
4044 /* stop monitoring at kernel level */
4048 * stop monitoring at the user level
4050 ia64_psr(regs
)->up
= 0;
4052 tregs
= task_pt_regs(task
);
4055 * stop monitoring at the user level
4057 ia64_psr(tregs
)->up
= 0;
4060 * monitoring disabled in kernel at next reschedule
4062 ctx
->ctx_saved_psr_up
= 0;
4063 DPRINT(("task=[%d]\n", task
->pid
));
4070 pfm_start(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4072 struct pt_regs
*tregs
;
4073 int state
, is_system
;
4075 state
= ctx
->ctx_state
;
4076 is_system
= ctx
->ctx_fl_system
;
4078 if (state
!= PFM_CTX_LOADED
) return -EINVAL
;
4081 * In system wide and when the context is loaded, access can only happen
4082 * when the caller is running on the CPU being monitored by the session.
4083 * It does not have to be the owner (ctx_task) of the context per se.
4085 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
4086 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
4091 * in system mode, we need to update the PMU directly
4092 * and the user level state of the caller, which may not
4093 * necessarily be the creator of the context.
4098 * set user level psr.pp for the caller
4100 ia64_psr(regs
)->pp
= 1;
4103 * now update the local PMU and cpuinfo
4105 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP
);
4108 * start monitoring at kernel level
4113 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) | IA64_DCR_PP
);
4123 if (ctx
->ctx_task
== current
) {
4125 /* start monitoring at kernel level */
4129 * activate monitoring at user level
4131 ia64_psr(regs
)->up
= 1;
4134 tregs
= task_pt_regs(ctx
->ctx_task
);
4137 * start monitoring at the kernel level the next
4138 * time the task is scheduled
4140 ctx
->ctx_saved_psr_up
= IA64_PSR_UP
;
4143 * activate monitoring at user level
4145 ia64_psr(tregs
)->up
= 1;
4151 pfm_get_pmc_reset(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4153 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
4158 for (i
= 0; i
< count
; i
++, req
++) {
4160 cnum
= req
->reg_num
;
4162 if (!PMC_IS_IMPL(cnum
)) goto abort_mission
;
4164 req
->reg_value
= PMC_DFL_VAL(cnum
);
4166 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
4168 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum
, req
->reg_value
));
4173 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
4178 pfm_check_task_exist(pfm_context_t
*ctx
)
4180 struct task_struct
*g
, *t
;
4183 read_lock(&tasklist_lock
);
4185 do_each_thread (g
, t
) {
4186 if (t
->thread
.pfm_context
== ctx
) {
4190 } while_each_thread (g
, t
);
4192 read_unlock(&tasklist_lock
);
4194 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret
, ctx
));
4200 pfm_context_load(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4202 struct task_struct
*task
;
4203 struct thread_struct
*thread
;
4204 struct pfm_context_t
*old
;
4205 unsigned long flags
;
4207 struct task_struct
*owner_task
= NULL
;
4209 pfarg_load_t
*req
= (pfarg_load_t
*)arg
;
4210 unsigned long *pmcs_source
, *pmds_source
;
4213 int state
, is_system
, set_dbregs
= 0;
4215 state
= ctx
->ctx_state
;
4216 is_system
= ctx
->ctx_fl_system
;
4218 * can only load from unloaded or terminated state
4220 if (state
!= PFM_CTX_UNLOADED
) {
4221 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4227 DPRINT(("load_pid [%d] using_dbreg=%d\n", req
->load_pid
, ctx
->ctx_fl_using_dbreg
));
4229 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && req
->load_pid
== current
->pid
) {
4230 DPRINT(("cannot use blocking mode on self\n"));
4234 ret
= pfm_get_task(ctx
, req
->load_pid
, &task
);
4236 DPRINT(("load_pid [%d] get_task=%d\n", req
->load_pid
, ret
));
4243 * system wide is self monitoring only
4245 if (is_system
&& task
!= current
) {
4246 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4251 thread
= &task
->thread
;
4255 * cannot load a context which is using range restrictions,
4256 * into a task that is being debugged.
4258 if (ctx
->ctx_fl_using_dbreg
) {
4259 if (thread
->flags
& IA64_THREAD_DBG_VALID
) {
4261 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req
->load_pid
));
4267 if (pfm_sessions
.pfs_ptrace_use_dbregs
) {
4268 DPRINT(("cannot load [%d] dbregs in use\n", task
->pid
));
4271 pfm_sessions
.pfs_sys_use_dbregs
++;
4272 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task
->pid
, pfm_sessions
.pfs_sys_use_dbregs
));
4279 if (ret
) goto error
;
4283 * SMP system-wide monitoring implies self-monitoring.
4285 * The programming model expects the task to
4286 * be pinned on a CPU throughout the session.
4287 * Here we take note of the current CPU at the
4288 * time the context is loaded. No call from
4289 * another CPU will be allowed.
4291 * The pinning via shed_setaffinity()
4292 * must be done by the calling task prior
4295 * systemwide: keep track of CPU this session is supposed to run on
4297 the_cpu
= ctx
->ctx_cpu
= smp_processor_id();
4301 * now reserve the session
4303 ret
= pfm_reserve_session(current
, is_system
, the_cpu
);
4304 if (ret
) goto error
;
4307 * task is necessarily stopped at this point.
4309 * If the previous context was zombie, then it got removed in
4310 * pfm_save_regs(). Therefore we should not see it here.
4311 * If we see a context, then this is an active context
4313 * XXX: needs to be atomic
4315 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4316 thread
->pfm_context
, ctx
));
4319 old
= ia64_cmpxchg(acq
, &thread
->pfm_context
, NULL
, ctx
, sizeof(pfm_context_t
*));
4321 DPRINT(("load_pid [%d] already has a context\n", req
->load_pid
));
4325 pfm_reset_msgq(ctx
);
4327 ctx
->ctx_state
= PFM_CTX_LOADED
;
4330 * link context to task
4332 ctx
->ctx_task
= task
;
4336 * we load as stopped
4338 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE
);
4339 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP
);
4341 if (ctx
->ctx_fl_excl_idle
) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE
);
4343 thread
->flags
|= IA64_THREAD_PM_VALID
;
4347 * propagate into thread-state
4349 pfm_copy_pmds(task
, ctx
);
4350 pfm_copy_pmcs(task
, ctx
);
4352 pmcs_source
= ctx
->th_pmcs
;
4353 pmds_source
= ctx
->th_pmds
;
4356 * always the case for system-wide
4358 if (task
== current
) {
4360 if (is_system
== 0) {
4362 /* allow user level control */
4363 ia64_psr(regs
)->sp
= 0;
4364 DPRINT(("clearing psr.sp for [%d]\n", task
->pid
));
4366 SET_LAST_CPU(ctx
, smp_processor_id());
4368 SET_ACTIVATION(ctx
);
4371 * push the other task out, if any
4373 owner_task
= GET_PMU_OWNER();
4374 if (owner_task
) pfm_lazy_save_regs(owner_task
);
4378 * load all PMD from ctx to PMU (as opposed to thread state)
4379 * restore all PMC from ctx to PMU
4381 pfm_restore_pmds(pmds_source
, ctx
->ctx_all_pmds
[0]);
4382 pfm_restore_pmcs(pmcs_source
, ctx
->ctx_all_pmcs
[0]);
4384 ctx
->ctx_reload_pmcs
[0] = 0UL;
4385 ctx
->ctx_reload_pmds
[0] = 0UL;
4388 * guaranteed safe by earlier check against DBG_VALID
4390 if (ctx
->ctx_fl_using_dbreg
) {
4391 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
4392 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
4397 SET_PMU_OWNER(task
, ctx
);
4399 DPRINT(("context loaded on PMU for [%d]\n", task
->pid
));
4402 * when not current, task MUST be stopped, so this is safe
4404 regs
= task_pt_regs(task
);
4406 /* force a full reload */
4407 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
4408 SET_LAST_CPU(ctx
, -1);
4410 /* initial saved psr (stopped) */
4411 ctx
->ctx_saved_psr_up
= 0UL;
4412 ia64_psr(regs
)->up
= ia64_psr(regs
)->pp
= 0;
4418 if (ret
) pfm_unreserve_session(ctx
, ctx
->ctx_fl_system
, the_cpu
);
4421 * we must undo the dbregs setting (for system-wide)
4423 if (ret
&& set_dbregs
) {
4425 pfm_sessions
.pfs_sys_use_dbregs
--;
4429 * release task, there is now a link with the context
4431 if (is_system
== 0 && task
!= current
) {
4435 ret
= pfm_check_task_exist(ctx
);
4437 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
4438 ctx
->ctx_task
= NULL
;
4446 * in this function, we do not need to increase the use count
4447 * for the task via get_task_struct(), because we hold the
4448 * context lock. If the task were to disappear while having
4449 * a context attached, it would go through pfm_exit_thread()
4450 * which also grabs the context lock and would therefore be blocked
4451 * until we are here.
4453 static void pfm_flush_pmds(struct task_struct
*, pfm_context_t
*ctx
);
4456 pfm_context_unload(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4458 struct task_struct
*task
= PFM_CTX_TASK(ctx
);
4459 struct pt_regs
*tregs
;
4460 int prev_state
, is_system
;
4463 DPRINT(("ctx_state=%d task [%d]\n", ctx
->ctx_state
, task
? task
->pid
: -1));
4465 prev_state
= ctx
->ctx_state
;
4466 is_system
= ctx
->ctx_fl_system
;
4469 * unload only when necessary
4471 if (prev_state
== PFM_CTX_UNLOADED
) {
4472 DPRINT(("ctx_state=%d, nothing to do\n", prev_state
));
4477 * clear psr and dcr bits
4479 ret
= pfm_stop(ctx
, NULL
, 0, regs
);
4480 if (ret
) return ret
;
4482 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
4485 * in system mode, we need to update the PMU directly
4486 * and the user level state of the caller, which may not
4487 * necessarily be the creator of the context.
4494 * local PMU is taken care of in pfm_stop()
4496 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE
);
4497 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE
);
4500 * save PMDs in context
4503 pfm_flush_pmds(current
, ctx
);
4506 * at this point we are done with the PMU
4507 * so we can unreserve the resource.
4509 if (prev_state
!= PFM_CTX_ZOMBIE
)
4510 pfm_unreserve_session(ctx
, 1 , ctx
->ctx_cpu
);
4513 * disconnect context from task
4515 task
->thread
.pfm_context
= NULL
;
4517 * disconnect task from context
4519 ctx
->ctx_task
= NULL
;
4522 * There is nothing more to cleanup here.
4530 tregs
= task
== current
? regs
: task_pt_regs(task
);
4532 if (task
== current
) {
4534 * cancel user level control
4536 ia64_psr(regs
)->sp
= 1;
4538 DPRINT(("setting psr.sp for [%d]\n", task
->pid
));
4541 * save PMDs to context
4544 pfm_flush_pmds(task
, ctx
);
4547 * at this point we are done with the PMU
4548 * so we can unreserve the resource.
4550 * when state was ZOMBIE, we have already unreserved.
4552 if (prev_state
!= PFM_CTX_ZOMBIE
)
4553 pfm_unreserve_session(ctx
, 0 , ctx
->ctx_cpu
);
4556 * reset activation counter and psr
4558 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
4559 SET_LAST_CPU(ctx
, -1);
4562 * PMU state will not be restored
4564 task
->thread
.flags
&= ~IA64_THREAD_PM_VALID
;
4567 * break links between context and task
4569 task
->thread
.pfm_context
= NULL
;
4570 ctx
->ctx_task
= NULL
;
4572 PFM_SET_WORK_PENDING(task
, 0);
4574 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_NONE
;
4575 ctx
->ctx_fl_can_restart
= 0;
4576 ctx
->ctx_fl_going_zombie
= 0;
4578 DPRINT(("disconnected [%d] from context\n", task
->pid
));
4585 * called only from exit_thread(): task == current
4586 * we come here only if current has a context attached (loaded or masked)
4589 pfm_exit_thread(struct task_struct
*task
)
4592 unsigned long flags
;
4593 struct pt_regs
*regs
= task_pt_regs(task
);
4597 ctx
= PFM_GET_CTX(task
);
4599 PROTECT_CTX(ctx
, flags
);
4601 DPRINT(("state=%d task [%d]\n", ctx
->ctx_state
, task
->pid
));
4603 state
= ctx
->ctx_state
;
4605 case PFM_CTX_UNLOADED
:
4607 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4608 * be in unloaded state
4610 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] ctx unloaded\n", task
->pid
);
4612 case PFM_CTX_LOADED
:
4613 case PFM_CTX_MASKED
:
4614 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
4616 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task
->pid
, state
, ret
);
4618 DPRINT(("ctx unloaded for current state was %d\n", state
));
4620 pfm_end_notify_user(ctx
);
4622 case PFM_CTX_ZOMBIE
:
4623 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
4625 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task
->pid
, state
, ret
);
4630 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task
->pid
, state
);
4633 UNPROTECT_CTX(ctx
, flags
);
4635 { u64 psr
= pfm_get_psr();
4636 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
4637 BUG_ON(GET_PMU_OWNER());
4638 BUG_ON(ia64_psr(regs
)->up
);
4639 BUG_ON(ia64_psr(regs
)->pp
);
4643 * All memory free operations (especially for vmalloc'ed memory)
4644 * MUST be done with interrupts ENABLED.
4646 if (free_ok
) pfm_context_free(ctx
);
4650 * functions MUST be listed in the increasing order of their index (see permfon.h)
4652 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4653 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4654 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4655 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4656 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4658 static pfm_cmd_desc_t pfm_cmd_tab
[]={
4659 /* 0 */PFM_CMD_NONE
,
4660 /* 1 */PFM_CMD(pfm_write_pmcs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4661 /* 2 */PFM_CMD(pfm_write_pmds
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4662 /* 3 */PFM_CMD(pfm_read_pmds
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4663 /* 4 */PFM_CMD_S(pfm_stop
, PFM_CMD_PCLRWS
),
4664 /* 5 */PFM_CMD_S(pfm_start
, PFM_CMD_PCLRWS
),
4665 /* 6 */PFM_CMD_NONE
,
4666 /* 7 */PFM_CMD_NONE
,
4667 /* 8 */PFM_CMD(pfm_context_create
, PFM_CMD_ARG_RW
, 1, pfarg_context_t
, pfm_ctx_getsize
),
4668 /* 9 */PFM_CMD_NONE
,
4669 /* 10 */PFM_CMD_S(pfm_restart
, PFM_CMD_PCLRW
),
4670 /* 11 */PFM_CMD_NONE
,
4671 /* 12 */PFM_CMD(pfm_get_features
, PFM_CMD_ARG_RW
, 1, pfarg_features_t
, NULL
),
4672 /* 13 */PFM_CMD(pfm_debug
, 0, 1, unsigned int, NULL
),
4673 /* 14 */PFM_CMD_NONE
,
4674 /* 15 */PFM_CMD(pfm_get_pmc_reset
, PFM_CMD_ARG_RW
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4675 /* 16 */PFM_CMD(pfm_context_load
, PFM_CMD_PCLRWS
, 1, pfarg_load_t
, NULL
),
4676 /* 17 */PFM_CMD_S(pfm_context_unload
, PFM_CMD_PCLRWS
),
4677 /* 18 */PFM_CMD_NONE
,
4678 /* 19 */PFM_CMD_NONE
,
4679 /* 20 */PFM_CMD_NONE
,
4680 /* 21 */PFM_CMD_NONE
,
4681 /* 22 */PFM_CMD_NONE
,
4682 /* 23 */PFM_CMD_NONE
,
4683 /* 24 */PFM_CMD_NONE
,
4684 /* 25 */PFM_CMD_NONE
,
4685 /* 26 */PFM_CMD_NONE
,
4686 /* 27 */PFM_CMD_NONE
,
4687 /* 28 */PFM_CMD_NONE
,
4688 /* 29 */PFM_CMD_NONE
,
4689 /* 30 */PFM_CMD_NONE
,
4690 /* 31 */PFM_CMD_NONE
,
4691 /* 32 */PFM_CMD(pfm_write_ibrs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_dbreg_t
, NULL
),
4692 /* 33 */PFM_CMD(pfm_write_dbrs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_dbreg_t
, NULL
)
4694 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4697 pfm_check_task_state(pfm_context_t
*ctx
, int cmd
, unsigned long flags
)
4699 struct task_struct
*task
;
4700 int state
, old_state
;
4703 state
= ctx
->ctx_state
;
4704 task
= ctx
->ctx_task
;
4707 DPRINT(("context %d no task, state=%d\n", ctx
->ctx_fd
, state
));
4711 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4715 task
->state
, PFM_CMD_STOPPED(cmd
)));
4718 * self-monitoring always ok.
4720 * for system-wide the caller can either be the creator of the
4721 * context (to one to which the context is attached to) OR
4722 * a task running on the same CPU as the session.
4724 if (task
== current
|| ctx
->ctx_fl_system
) return 0;
4727 * we are monitoring another thread
4730 case PFM_CTX_UNLOADED
:
4732 * if context is UNLOADED we are safe to go
4735 case PFM_CTX_ZOMBIE
:
4737 * no command can operate on a zombie context
4739 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd
));
4741 case PFM_CTX_MASKED
:
4743 * PMU state has been saved to software even though
4744 * the thread may still be running.
4746 if (cmd
!= PFM_UNLOAD_CONTEXT
) return 0;
4750 * context is LOADED or MASKED. Some commands may need to have
4753 * We could lift this restriction for UP but it would mean that
4754 * the user has no guarantee the task would not run between
4755 * two successive calls to perfmonctl(). That's probably OK.
4756 * If this user wants to ensure the task does not run, then
4757 * the task must be stopped.
4759 if (PFM_CMD_STOPPED(cmd
)) {
4760 if ((task
->state
!= TASK_STOPPED
) && (task
->state
!= TASK_TRACED
)) {
4761 DPRINT(("[%d] task not in stopped state\n", task
->pid
));
4765 * task is now stopped, wait for ctxsw out
4767 * This is an interesting point in the code.
4768 * We need to unprotect the context because
4769 * the pfm_save_regs() routines needs to grab
4770 * the same lock. There are danger in doing
4771 * this because it leaves a window open for
4772 * another task to get access to the context
4773 * and possibly change its state. The one thing
4774 * that is not possible is for the context to disappear
4775 * because we are protected by the VFS layer, i.e.,
4776 * get_fd()/put_fd().
4780 UNPROTECT_CTX(ctx
, flags
);
4782 wait_task_inactive(task
);
4784 PROTECT_CTX(ctx
, flags
);
4787 * we must recheck to verify if state has changed
4789 if (ctx
->ctx_state
!= old_state
) {
4790 DPRINT(("old_state=%d new_state=%d\n", old_state
, ctx
->ctx_state
));
4798 * system-call entry point (must return long)
4801 sys_perfmonctl (int fd
, int cmd
, void __user
*arg
, int count
)
4803 struct file
*file
= NULL
;
4804 pfm_context_t
*ctx
= NULL
;
4805 unsigned long flags
= 0UL;
4806 void *args_k
= NULL
;
4807 long ret
; /* will expand int return types */
4808 size_t base_sz
, sz
, xtra_sz
= 0;
4809 int narg
, completed_args
= 0, call_made
= 0, cmd_flags
;
4810 int (*func
)(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
4811 int (*getsize
)(void *arg
, size_t *sz
);
4812 #define PFM_MAX_ARGSIZE 4096
4815 * reject any call if perfmon was disabled at initialization
4817 if (unlikely(pmu_conf
== NULL
)) return -ENOSYS
;
4819 if (unlikely(cmd
< 0 || cmd
>= PFM_CMD_COUNT
)) {
4820 DPRINT(("invalid cmd=%d\n", cmd
));
4824 func
= pfm_cmd_tab
[cmd
].cmd_func
;
4825 narg
= pfm_cmd_tab
[cmd
].cmd_narg
;
4826 base_sz
= pfm_cmd_tab
[cmd
].cmd_argsize
;
4827 getsize
= pfm_cmd_tab
[cmd
].cmd_getsize
;
4828 cmd_flags
= pfm_cmd_tab
[cmd
].cmd_flags
;
4830 if (unlikely(func
== NULL
)) {
4831 DPRINT(("invalid cmd=%d\n", cmd
));
4835 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4843 * check if number of arguments matches what the command expects
4845 if (unlikely((narg
== PFM_CMD_ARG_MANY
&& count
<= 0) || (narg
> 0 && narg
!= count
)))
4849 sz
= xtra_sz
+ base_sz
*count
;
4851 * limit abuse to min page size
4853 if (unlikely(sz
> PFM_MAX_ARGSIZE
)) {
4854 printk(KERN_ERR
"perfmon: [%d] argument too big %lu\n", current
->pid
, sz
);
4859 * allocate default-sized argument buffer
4861 if (likely(count
&& args_k
== NULL
)) {
4862 args_k
= kmalloc(PFM_MAX_ARGSIZE
, GFP_KERNEL
);
4863 if (args_k
== NULL
) return -ENOMEM
;
4871 * assume sz = 0 for command without parameters
4873 if (sz
&& copy_from_user(args_k
, arg
, sz
)) {
4874 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz
, arg
));
4879 * check if command supports extra parameters
4881 if (completed_args
== 0 && getsize
) {
4883 * get extra parameters size (based on main argument)
4885 ret
= (*getsize
)(args_k
, &xtra_sz
);
4886 if (ret
) goto error_args
;
4890 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz
, xtra_sz
));
4892 /* retry if necessary */
4893 if (likely(xtra_sz
)) goto restart_args
;
4896 if (unlikely((cmd_flags
& PFM_CMD_FD
) == 0)) goto skip_fd
;
4901 if (unlikely(file
== NULL
)) {
4902 DPRINT(("invalid fd %d\n", fd
));
4905 if (unlikely(PFM_IS_FILE(file
) == 0)) {
4906 DPRINT(("fd %d not related to perfmon\n", fd
));
4910 ctx
= (pfm_context_t
*)file
->private_data
;
4911 if (unlikely(ctx
== NULL
)) {
4912 DPRINT(("no context for fd %d\n", fd
));
4915 prefetch(&ctx
->ctx_state
);
4917 PROTECT_CTX(ctx
, flags
);
4920 * check task is stopped
4922 ret
= pfm_check_task_state(ctx
, cmd
, flags
);
4923 if (unlikely(ret
)) goto abort_locked
;
4926 ret
= (*func
)(ctx
, args_k
, count
, task_pt_regs(current
));
4932 DPRINT(("context unlocked\n"));
4933 UNPROTECT_CTX(ctx
, flags
);
4936 /* copy argument back to user, if needed */
4937 if (call_made
&& PFM_CMD_RW_ARG(cmd
) && copy_to_user(arg
, args_k
, base_sz
*count
)) ret
= -EFAULT
;
4945 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd
), ret
));
4951 pfm_resume_after_ovfl(pfm_context_t
*ctx
, unsigned long ovfl_regs
, struct pt_regs
*regs
)
4953 pfm_buffer_fmt_t
*fmt
= ctx
->ctx_buf_fmt
;
4954 pfm_ovfl_ctrl_t rst_ctrl
;
4958 state
= ctx
->ctx_state
;
4960 * Unlock sampling buffer and reset index atomically
4961 * XXX: not really needed when blocking
4963 if (CTX_HAS_SMPL(ctx
)) {
4965 rst_ctrl
.bits
.mask_monitoring
= 0;
4966 rst_ctrl
.bits
.reset_ovfl_pmds
= 0;
4968 if (state
== PFM_CTX_LOADED
)
4969 ret
= pfm_buf_fmt_restart_active(fmt
, current
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
4971 ret
= pfm_buf_fmt_restart(fmt
, current
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
4973 rst_ctrl
.bits
.mask_monitoring
= 0;
4974 rst_ctrl
.bits
.reset_ovfl_pmds
= 1;
4978 if (rst_ctrl
.bits
.reset_ovfl_pmds
) {
4979 pfm_reset_regs(ctx
, &ovfl_regs
, PFM_PMD_LONG_RESET
);
4981 if (rst_ctrl
.bits
.mask_monitoring
== 0) {
4982 DPRINT(("resuming monitoring\n"));
4983 if (ctx
->ctx_state
== PFM_CTX_MASKED
) pfm_restore_monitoring(current
);
4985 DPRINT(("stopping monitoring\n"));
4986 //pfm_stop_monitoring(current, regs);
4988 ctx
->ctx_state
= PFM_CTX_LOADED
;
4993 * context MUST BE LOCKED when calling
4994 * can only be called for current
4997 pfm_context_force_terminate(pfm_context_t
*ctx
, struct pt_regs
*regs
)
5001 DPRINT(("entering for [%d]\n", current
->pid
));
5003 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
5005 printk(KERN_ERR
"pfm_context_force_terminate: [%d] unloaded failed with %d\n", current
->pid
, ret
);
5009 * and wakeup controlling task, indicating we are now disconnected
5011 wake_up_interruptible(&ctx
->ctx_zombieq
);
5014 * given that context is still locked, the controlling
5015 * task will only get access when we return from
5016 * pfm_handle_work().
5020 static int pfm_ovfl_notify_user(pfm_context_t
*ctx
, unsigned long ovfl_pmds
);
5022 * pfm_handle_work() can be called with interrupts enabled
5023 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5024 * call may sleep, therefore we must re-enable interrupts
5025 * to avoid deadlocks. It is safe to do so because this function
5026 * is called ONLY when returning to user level (PUStk=1), in which case
5027 * there is no risk of kernel stack overflow due to deep
5028 * interrupt nesting.
5031 pfm_handle_work(void)
5034 struct pt_regs
*regs
;
5035 unsigned long flags
, dummy_flags
;
5036 unsigned long ovfl_regs
;
5037 unsigned int reason
;
5040 ctx
= PFM_GET_CTX(current
);
5042 printk(KERN_ERR
"perfmon: [%d] has no PFM context\n", current
->pid
);
5046 PROTECT_CTX(ctx
, flags
);
5048 PFM_SET_WORK_PENDING(current
, 0);
5050 pfm_clear_task_notify();
5052 regs
= task_pt_regs(current
);
5055 * extract reason for being here and clear
5057 reason
= ctx
->ctx_fl_trap_reason
;
5058 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_NONE
;
5059 ovfl_regs
= ctx
->ctx_ovfl_regs
[0];
5061 DPRINT(("reason=%d state=%d\n", reason
, ctx
->ctx_state
));
5064 * must be done before we check for simple-reset mode
5066 if (ctx
->ctx_fl_going_zombie
|| ctx
->ctx_state
== PFM_CTX_ZOMBIE
) goto do_zombie
;
5069 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5070 if (reason
== PFM_TRAP_REASON_RESET
) goto skip_blocking
;
5073 * restore interrupt mask to what it was on entry.
5074 * Could be enabled/diasbled.
5076 UNPROTECT_CTX(ctx
, flags
);
5079 * force interrupt enable because of down_interruptible()
5083 DPRINT(("before block sleeping\n"));
5086 * may go through without blocking on SMP systems
5087 * if restart has been received already by the time we call down()
5089 ret
= wait_for_completion_interruptible(&ctx
->ctx_restart_done
);
5091 DPRINT(("after block sleeping ret=%d\n", ret
));
5094 * lock context and mask interrupts again
5095 * We save flags into a dummy because we may have
5096 * altered interrupts mask compared to entry in this
5099 PROTECT_CTX(ctx
, dummy_flags
);
5102 * we need to read the ovfl_regs only after wake-up
5103 * because we may have had pfm_write_pmds() in between
5104 * and that can changed PMD values and therefore
5105 * ovfl_regs is reset for these new PMD values.
5107 ovfl_regs
= ctx
->ctx_ovfl_regs
[0];
5109 if (ctx
->ctx_fl_going_zombie
) {
5111 DPRINT(("context is zombie, bailing out\n"));
5112 pfm_context_force_terminate(ctx
, regs
);
5116 * in case of interruption of down() we don't restart anything
5118 if (ret
< 0) goto nothing_to_do
;
5121 pfm_resume_after_ovfl(ctx
, ovfl_regs
, regs
);
5122 ctx
->ctx_ovfl_regs
[0] = 0UL;
5126 * restore flags as they were upon entry
5128 UNPROTECT_CTX(ctx
, flags
);
5132 pfm_notify_user(pfm_context_t
*ctx
, pfm_msg_t
*msg
)
5134 if (ctx
->ctx_state
== PFM_CTX_ZOMBIE
) {
5135 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5139 DPRINT(("waking up somebody\n"));
5141 if (msg
) wake_up_interruptible(&ctx
->ctx_msgq_wait
);
5144 * safe, we are not in intr handler, nor in ctxsw when
5147 kill_fasync (&ctx
->ctx_async_queue
, SIGIO
, POLL_IN
);
5153 pfm_ovfl_notify_user(pfm_context_t
*ctx
, unsigned long ovfl_pmds
)
5155 pfm_msg_t
*msg
= NULL
;
5157 if (ctx
->ctx_fl_no_msg
== 0) {
5158 msg
= pfm_get_new_msg(ctx
);
5160 printk(KERN_ERR
"perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5164 msg
->pfm_ovfl_msg
.msg_type
= PFM_MSG_OVFL
;
5165 msg
->pfm_ovfl_msg
.msg_ctx_fd
= ctx
->ctx_fd
;
5166 msg
->pfm_ovfl_msg
.msg_active_set
= 0;
5167 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[0] = ovfl_pmds
;
5168 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[1] = 0UL;
5169 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[2] = 0UL;
5170 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[3] = 0UL;
5171 msg
->pfm_ovfl_msg
.msg_tstamp
= 0UL;
5174 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5180 return pfm_notify_user(ctx
, msg
);
5184 pfm_end_notify_user(pfm_context_t
*ctx
)
5188 msg
= pfm_get_new_msg(ctx
);
5190 printk(KERN_ERR
"perfmon: pfm_end_notify_user no more notification msgs\n");
5194 memset(msg
, 0, sizeof(*msg
));
5196 msg
->pfm_end_msg
.msg_type
= PFM_MSG_END
;
5197 msg
->pfm_end_msg
.msg_ctx_fd
= ctx
->ctx_fd
;
5198 msg
->pfm_ovfl_msg
.msg_tstamp
= 0UL;
5200 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5205 return pfm_notify_user(ctx
, msg
);
5209 * main overflow processing routine.
5210 * it can be called from the interrupt path or explicitely during the context switch code
5213 pfm_overflow_handler(struct task_struct
*task
, pfm_context_t
*ctx
, u64 pmc0
, struct pt_regs
*regs
)
5215 pfm_ovfl_arg_t
*ovfl_arg
;
5217 unsigned long old_val
, ovfl_val
, new_val
;
5218 unsigned long ovfl_notify
= 0UL, ovfl_pmds
= 0UL, smpl_pmds
= 0UL, reset_pmds
;
5219 unsigned long tstamp
;
5220 pfm_ovfl_ctrl_t ovfl_ctrl
;
5221 unsigned int i
, has_smpl
;
5222 int must_notify
= 0;
5224 if (unlikely(ctx
->ctx_state
== PFM_CTX_ZOMBIE
)) goto stop_monitoring
;
5227 * sanity test. Should never happen
5229 if (unlikely((pmc0
& 0x1) == 0)) goto sanity_check
;
5231 tstamp
= ia64_get_itc();
5232 mask
= pmc0
>> PMU_FIRST_COUNTER
;
5233 ovfl_val
= pmu_conf
->ovfl_val
;
5234 has_smpl
= CTX_HAS_SMPL(ctx
);
5236 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5237 "used_pmds=0x%lx\n",
5239 task
? task
->pid
: -1,
5240 (regs
? regs
->cr_iip
: 0),
5241 CTX_OVFL_NOBLOCK(ctx
) ? "nonblocking" : "blocking",
5242 ctx
->ctx_used_pmds
[0]));
5246 * first we update the virtual counters
5247 * assume there was a prior ia64_srlz_d() issued
5249 for (i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
5251 /* skip pmd which did not overflow */
5252 if ((mask
& 0x1) == 0) continue;
5255 * Note that the pmd is not necessarily 0 at this point as qualified events
5256 * may have happened before the PMU was frozen. The residual count is not
5257 * taken into consideration here but will be with any read of the pmd via
5260 old_val
= new_val
= ctx
->ctx_pmds
[i
].val
;
5261 new_val
+= 1 + ovfl_val
;
5262 ctx
->ctx_pmds
[i
].val
= new_val
;
5265 * check for overflow condition
5267 if (likely(old_val
> new_val
)) {
5268 ovfl_pmds
|= 1UL << i
;
5269 if (PMC_OVFL_NOTIFY(ctx
, i
)) ovfl_notify
|= 1UL << i
;
5272 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5276 ia64_get_pmd(i
) & ovfl_val
,
5282 * there was no 64-bit overflow, nothing else to do
5284 if (ovfl_pmds
== 0UL) return;
5287 * reset all control bits
5293 * if a sampling format module exists, then we "cache" the overflow by
5294 * calling the module's handler() routine.
5297 unsigned long start_cycles
, end_cycles
;
5298 unsigned long pmd_mask
;
5300 int this_cpu
= smp_processor_id();
5302 pmd_mask
= ovfl_pmds
>> PMU_FIRST_COUNTER
;
5303 ovfl_arg
= &ctx
->ctx_ovfl_arg
;
5305 prefetch(ctx
->ctx_smpl_hdr
);
5307 for(i
=PMU_FIRST_COUNTER
; pmd_mask
&& ret
== 0; i
++, pmd_mask
>>=1) {
5311 if ((pmd_mask
& 0x1) == 0) continue;
5313 ovfl_arg
->ovfl_pmd
= (unsigned char )i
;
5314 ovfl_arg
->ovfl_notify
= ovfl_notify
& mask
? 1 : 0;
5315 ovfl_arg
->active_set
= 0;
5316 ovfl_arg
->ovfl_ctrl
.val
= 0; /* module must fill in all fields */
5317 ovfl_arg
->smpl_pmds
[0] = smpl_pmds
= ctx
->ctx_pmds
[i
].smpl_pmds
[0];
5319 ovfl_arg
->pmd_value
= ctx
->ctx_pmds
[i
].val
;
5320 ovfl_arg
->pmd_last_reset
= ctx
->ctx_pmds
[i
].lval
;
5321 ovfl_arg
->pmd_eventid
= ctx
->ctx_pmds
[i
].eventid
;
5324 * copy values of pmds of interest. Sampling format may copy them
5325 * into sampling buffer.
5328 for(j
=0, k
=0; smpl_pmds
; j
++, smpl_pmds
>>=1) {
5329 if ((smpl_pmds
& 0x1) == 0) continue;
5330 ovfl_arg
->smpl_pmds_values
[k
++] = PMD_IS_COUNTING(j
) ? pfm_read_soft_counter(ctx
, j
) : ia64_get_pmd(j
);
5331 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k
-1, j
, ovfl_arg
->smpl_pmds_values
[k
-1]));
5335 pfm_stats
[this_cpu
].pfm_smpl_handler_calls
++;
5337 start_cycles
= ia64_get_itc();
5340 * call custom buffer format record (handler) routine
5342 ret
= (*ctx
->ctx_buf_fmt
->fmt_handler
)(task
, ctx
->ctx_smpl_hdr
, ovfl_arg
, regs
, tstamp
);
5344 end_cycles
= ia64_get_itc();
5347 * For those controls, we take the union because they have
5348 * an all or nothing behavior.
5350 ovfl_ctrl
.bits
.notify_user
|= ovfl_arg
->ovfl_ctrl
.bits
.notify_user
;
5351 ovfl_ctrl
.bits
.block_task
|= ovfl_arg
->ovfl_ctrl
.bits
.block_task
;
5352 ovfl_ctrl
.bits
.mask_monitoring
|= ovfl_arg
->ovfl_ctrl
.bits
.mask_monitoring
;
5354 * build the bitmask of pmds to reset now
5356 if (ovfl_arg
->ovfl_ctrl
.bits
.reset_ovfl_pmds
) reset_pmds
|= mask
;
5358 pfm_stats
[this_cpu
].pfm_smpl_handler_cycles
+= end_cycles
- start_cycles
;
5361 * when the module cannot handle the rest of the overflows, we abort right here
5363 if (ret
&& pmd_mask
) {
5364 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5365 pmd_mask
<<PMU_FIRST_COUNTER
));
5368 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5370 ovfl_pmds
&= ~reset_pmds
;
5373 * when no sampling module is used, then the default
5374 * is to notify on overflow if requested by user
5376 ovfl_ctrl
.bits
.notify_user
= ovfl_notify
? 1 : 0;
5377 ovfl_ctrl
.bits
.block_task
= ovfl_notify
? 1 : 0;
5378 ovfl_ctrl
.bits
.mask_monitoring
= ovfl_notify
? 1 : 0; /* XXX: change for saturation */
5379 ovfl_ctrl
.bits
.reset_ovfl_pmds
= ovfl_notify
? 0 : 1;
5381 * if needed, we reset all overflowed pmds
5383 if (ovfl_notify
== 0) reset_pmds
= ovfl_pmds
;
5386 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds
, reset_pmds
));
5389 * reset the requested PMD registers using the short reset values
5392 unsigned long bm
= reset_pmds
;
5393 pfm_reset_regs(ctx
, &bm
, PFM_PMD_SHORT_RESET
);
5396 if (ovfl_notify
&& ovfl_ctrl
.bits
.notify_user
) {
5398 * keep track of what to reset when unblocking
5400 ctx
->ctx_ovfl_regs
[0] = ovfl_pmds
;
5403 * check for blocking context
5405 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && ovfl_ctrl
.bits
.block_task
) {
5407 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_BLOCK
;
5410 * set the perfmon specific checking pending work for the task
5412 PFM_SET_WORK_PENDING(task
, 1);
5415 * when coming from ctxsw, current still points to the
5416 * previous task, therefore we must work with task and not current.
5418 pfm_set_task_notify(task
);
5421 * defer until state is changed (shorten spin window). the context is locked
5422 * anyway, so the signal receiver would come spin for nothing.
5427 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5428 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid
: -1,
5429 PFM_GET_WORK_PENDING(task
),
5430 ctx
->ctx_fl_trap_reason
,
5433 ovfl_ctrl
.bits
.mask_monitoring
? 1 : 0));
5435 * in case monitoring must be stopped, we toggle the psr bits
5437 if (ovfl_ctrl
.bits
.mask_monitoring
) {
5438 pfm_mask_monitoring(task
);
5439 ctx
->ctx_state
= PFM_CTX_MASKED
;
5440 ctx
->ctx_fl_can_restart
= 1;
5444 * send notification now
5446 if (must_notify
) pfm_ovfl_notify_user(ctx
, ovfl_notify
);
5451 printk(KERN_ERR
"perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5453 task
? task
->pid
: -1,
5459 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5460 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5461 * come here as zombie only if the task is the current task. In which case, we
5462 * can access the PMU hardware directly.
5464 * Note that zombies do have PM_VALID set. So here we do the minimal.
5466 * In case the context was zombified it could not be reclaimed at the time
5467 * the monitoring program exited. At this point, the PMU reservation has been
5468 * returned, the sampiing buffer has been freed. We must convert this call
5469 * into a spurious interrupt. However, we must also avoid infinite overflows
5470 * by stopping monitoring for this task. We can only come here for a per-task
5471 * context. All we need to do is to stop monitoring using the psr bits which
5472 * are always task private. By re-enabling secure montioring, we ensure that
5473 * the monitored task will not be able to re-activate monitoring.
5474 * The task will eventually be context switched out, at which point the context
5475 * will be reclaimed (that includes releasing ownership of the PMU).
5477 * So there might be a window of time where the number of per-task session is zero
5478 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5479 * context. This is safe because if a per-task session comes in, it will push this one
5480 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5481 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5482 * also push our zombie context out.
5484 * Overall pretty hairy stuff....
5486 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task
? task
->pid
: -1));
5488 ia64_psr(regs
)->up
= 0;
5489 ia64_psr(regs
)->sp
= 1;
5494 pfm_do_interrupt_handler(int irq
, void *arg
, struct pt_regs
*regs
)
5496 struct task_struct
*task
;
5498 unsigned long flags
;
5500 int this_cpu
= smp_processor_id();
5503 pfm_stats
[this_cpu
].pfm_ovfl_intr_count
++;
5506 * srlz.d done before arriving here
5508 pmc0
= ia64_get_pmc(0);
5510 task
= GET_PMU_OWNER();
5511 ctx
= GET_PMU_CTX();
5514 * if we have some pending bits set
5515 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5517 if (PMC0_HAS_OVFL(pmc0
) && task
) {
5519 * we assume that pmc0.fr is always set here
5523 if (!ctx
) goto report_spurious1
;
5525 if (ctx
->ctx_fl_system
== 0 && (task
->thread
.flags
& IA64_THREAD_PM_VALID
) == 0)
5526 goto report_spurious2
;
5528 PROTECT_CTX_NOPRINT(ctx
, flags
);
5530 pfm_overflow_handler(task
, ctx
, pmc0
, regs
);
5532 UNPROTECT_CTX_NOPRINT(ctx
, flags
);
5535 pfm_stats
[this_cpu
].pfm_spurious_ovfl_intr_count
++;
5539 * keep it unfrozen at all times
5546 printk(KERN_INFO
"perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5547 this_cpu
, task
->pid
);
5551 printk(KERN_INFO
"perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5559 pfm_interrupt_handler(int irq
, void *arg
)
5561 unsigned long start_cycles
, total_cycles
;
5562 unsigned long min
, max
;
5565 struct pt_regs
*regs
= get_irq_regs();
5567 this_cpu
= get_cpu();
5568 if (likely(!pfm_alt_intr_handler
)) {
5569 min
= pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_min
;
5570 max
= pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_max
;
5572 start_cycles
= ia64_get_itc();
5574 ret
= pfm_do_interrupt_handler(irq
, arg
, regs
);
5576 total_cycles
= ia64_get_itc();
5579 * don't measure spurious interrupts
5581 if (likely(ret
== 0)) {
5582 total_cycles
-= start_cycles
;
5584 if (total_cycles
< min
) pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_min
= total_cycles
;
5585 if (total_cycles
> max
) pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_max
= total_cycles
;
5587 pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles
+= total_cycles
;
5591 (*pfm_alt_intr_handler
->handler
)(irq
, arg
, regs
);
5594 put_cpu_no_resched();
5599 * /proc/perfmon interface, for debug only
5602 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5605 pfm_proc_start(struct seq_file
*m
, loff_t
*pos
)
5608 return PFM_PROC_SHOW_HEADER
;
5611 while (*pos
<= NR_CPUS
) {
5612 if (cpu_online(*pos
- 1)) {
5613 return (void *)*pos
;
5621 pfm_proc_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
5624 return pfm_proc_start(m
, pos
);
5628 pfm_proc_stop(struct seq_file
*m
, void *v
)
5633 pfm_proc_show_header(struct seq_file
*m
)
5635 struct list_head
* pos
;
5636 pfm_buffer_fmt_t
* entry
;
5637 unsigned long flags
;
5640 "perfmon version : %u.%u\n"
5643 "expert mode : %s\n"
5644 "ovfl_mask : 0x%lx\n"
5645 "PMU flags : 0x%x\n",
5646 PFM_VERSION_MAJ
, PFM_VERSION_MIN
,
5648 pfm_sysctl
.fastctxsw
> 0 ? "Yes": "No",
5649 pfm_sysctl
.expert_mode
> 0 ? "Yes": "No",
5656 "proc_sessions : %u\n"
5657 "sys_sessions : %u\n"
5658 "sys_use_dbregs : %u\n"
5659 "ptrace_use_dbregs : %u\n",
5660 pfm_sessions
.pfs_task_sessions
,
5661 pfm_sessions
.pfs_sys_sessions
,
5662 pfm_sessions
.pfs_sys_use_dbregs
,
5663 pfm_sessions
.pfs_ptrace_use_dbregs
);
5667 spin_lock(&pfm_buffer_fmt_lock
);
5669 list_for_each(pos
, &pfm_buffer_fmt_list
) {
5670 entry
= list_entry(pos
, pfm_buffer_fmt_t
, fmt_list
);
5671 seq_printf(m
, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5682 entry
->fmt_uuid
[10],
5683 entry
->fmt_uuid
[11],
5684 entry
->fmt_uuid
[12],
5685 entry
->fmt_uuid
[13],
5686 entry
->fmt_uuid
[14],
5687 entry
->fmt_uuid
[15],
5690 spin_unlock(&pfm_buffer_fmt_lock
);
5695 pfm_proc_show(struct seq_file
*m
, void *v
)
5701 if (v
== PFM_PROC_SHOW_HEADER
) {
5702 pfm_proc_show_header(m
);
5706 /* show info for CPU (v - 1) */
5710 "CPU%-2d overflow intrs : %lu\n"
5711 "CPU%-2d overflow cycles : %lu\n"
5712 "CPU%-2d overflow min : %lu\n"
5713 "CPU%-2d overflow max : %lu\n"
5714 "CPU%-2d smpl handler calls : %lu\n"
5715 "CPU%-2d smpl handler cycles : %lu\n"
5716 "CPU%-2d spurious intrs : %lu\n"
5717 "CPU%-2d replay intrs : %lu\n"
5718 "CPU%-2d syst_wide : %d\n"
5719 "CPU%-2d dcr_pp : %d\n"
5720 "CPU%-2d exclude idle : %d\n"
5721 "CPU%-2d owner : %d\n"
5722 "CPU%-2d context : %p\n"
5723 "CPU%-2d activations : %lu\n",
5724 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_count
,
5725 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles
,
5726 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles_min
,
5727 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles_max
,
5728 cpu
, pfm_stats
[cpu
].pfm_smpl_handler_calls
,
5729 cpu
, pfm_stats
[cpu
].pfm_smpl_handler_cycles
,
5730 cpu
, pfm_stats
[cpu
].pfm_spurious_ovfl_intr_count
,
5731 cpu
, pfm_stats
[cpu
].pfm_replay_ovfl_intr_count
,
5732 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_SYST_WIDE
? 1 : 0,
5733 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_DCR_PP
? 1 : 0,
5734 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_EXCL_IDLE
? 1 : 0,
5735 cpu
, pfm_get_cpu_data(pmu_owner
, cpu
) ? pfm_get_cpu_data(pmu_owner
, cpu
)->pid
: -1,
5736 cpu
, pfm_get_cpu_data(pmu_ctx
, cpu
),
5737 cpu
, pfm_get_cpu_data(pmu_activation_number
, cpu
));
5739 if (num_online_cpus() == 1 && pfm_sysctl
.debug
> 0) {
5741 psr
= pfm_get_psr();
5746 "CPU%-2d psr : 0x%lx\n"
5747 "CPU%-2d pmc0 : 0x%lx\n",
5749 cpu
, ia64_get_pmc(0));
5751 for (i
=0; PMC_IS_LAST(i
) == 0; i
++) {
5752 if (PMC_IS_COUNTING(i
) == 0) continue;
5754 "CPU%-2d pmc%u : 0x%lx\n"
5755 "CPU%-2d pmd%u : 0x%lx\n",
5756 cpu
, i
, ia64_get_pmc(i
),
5757 cpu
, i
, ia64_get_pmd(i
));
5763 struct seq_operations pfm_seq_ops
= {
5764 .start
= pfm_proc_start
,
5765 .next
= pfm_proc_next
,
5766 .stop
= pfm_proc_stop
,
5767 .show
= pfm_proc_show
5771 pfm_proc_open(struct inode
*inode
, struct file
*file
)
5773 return seq_open(file
, &pfm_seq_ops
);
5778 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5779 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5780 * is active or inactive based on mode. We must rely on the value in
5781 * local_cpu_data->pfm_syst_info
5784 pfm_syst_wide_update_task(struct task_struct
*task
, unsigned long info
, int is_ctxswin
)
5786 struct pt_regs
*regs
;
5788 unsigned long dcr_pp
;
5790 dcr_pp
= info
& PFM_CPUINFO_DCR_PP
? 1 : 0;
5793 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5794 * on every CPU, so we can rely on the pid to identify the idle task.
5796 if ((info
& PFM_CPUINFO_EXCL_IDLE
) == 0 || task
->pid
) {
5797 regs
= task_pt_regs(task
);
5798 ia64_psr(regs
)->pp
= is_ctxswin
? dcr_pp
: 0;
5802 * if monitoring has started
5805 dcr
= ia64_getreg(_IA64_REG_CR_DCR
);
5807 * context switching in?
5810 /* mask monitoring for the idle task */
5811 ia64_setreg(_IA64_REG_CR_DCR
, dcr
& ~IA64_DCR_PP
);
5817 * context switching out
5818 * restore monitoring for next task
5820 * Due to inlining this odd if-then-else construction generates
5823 ia64_setreg(_IA64_REG_CR_DCR
, dcr
|IA64_DCR_PP
);
5832 pfm_force_cleanup(pfm_context_t
*ctx
, struct pt_regs
*regs
)
5834 struct task_struct
*task
= ctx
->ctx_task
;
5836 ia64_psr(regs
)->up
= 0;
5837 ia64_psr(regs
)->sp
= 1;
5839 if (GET_PMU_OWNER() == task
) {
5840 DPRINT(("cleared ownership for [%d]\n", ctx
->ctx_task
->pid
));
5841 SET_PMU_OWNER(NULL
, NULL
);
5845 * disconnect the task from the context and vice-versa
5847 PFM_SET_WORK_PENDING(task
, 0);
5849 task
->thread
.pfm_context
= NULL
;
5850 task
->thread
.flags
&= ~IA64_THREAD_PM_VALID
;
5852 DPRINT(("force cleanup for [%d]\n", task
->pid
));
5857 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5860 pfm_save_regs(struct task_struct
*task
)
5863 unsigned long flags
;
5867 ctx
= PFM_GET_CTX(task
);
5868 if (ctx
== NULL
) return;
5871 * we always come here with interrupts ALREADY disabled by
5872 * the scheduler. So we simply need to protect against concurrent
5873 * access, not CPU concurrency.
5875 flags
= pfm_protect_ctx_ctxsw(ctx
);
5877 if (ctx
->ctx_state
== PFM_CTX_ZOMBIE
) {
5878 struct pt_regs
*regs
= task_pt_regs(task
);
5882 pfm_force_cleanup(ctx
, regs
);
5884 BUG_ON(ctx
->ctx_smpl_hdr
);
5886 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
5888 pfm_context_free(ctx
);
5893 * save current PSR: needed because we modify it
5896 psr
= pfm_get_psr();
5898 BUG_ON(psr
& (IA64_PSR_I
));
5902 * This is the last instruction which may generate an overflow
5904 * We do not need to set psr.sp because, it is irrelevant in kernel.
5905 * It will be restored from ipsr when going back to user level
5910 * keep a copy of psr.up (for reload)
5912 ctx
->ctx_saved_psr_up
= psr
& IA64_PSR_UP
;
5915 * release ownership of this PMU.
5916 * PM interrupts are masked, so nothing
5919 SET_PMU_OWNER(NULL
, NULL
);
5922 * we systematically save the PMD as we have no
5923 * guarantee we will be schedule at that same
5926 pfm_save_pmds(ctx
->th_pmds
, ctx
->ctx_used_pmds
[0]);
5929 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5930 * we will need it on the restore path to check
5931 * for pending overflow.
5933 ctx
->th_pmcs
[0] = ia64_get_pmc(0);
5936 * unfreeze PMU if had pending overflows
5938 if (ctx
->th_pmcs
[0] & ~0x1UL
) pfm_unfreeze_pmu();
5941 * finally, allow context access.
5942 * interrupts will still be masked after this call.
5944 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
5947 #else /* !CONFIG_SMP */
5949 pfm_save_regs(struct task_struct
*task
)
5954 ctx
= PFM_GET_CTX(task
);
5955 if (ctx
== NULL
) return;
5958 * save current PSR: needed because we modify it
5960 psr
= pfm_get_psr();
5962 BUG_ON(psr
& (IA64_PSR_I
));
5966 * This is the last instruction which may generate an overflow
5968 * We do not need to set psr.sp because, it is irrelevant in kernel.
5969 * It will be restored from ipsr when going back to user level
5974 * keep a copy of psr.up (for reload)
5976 ctx
->ctx_saved_psr_up
= psr
& IA64_PSR_UP
;
5980 pfm_lazy_save_regs (struct task_struct
*task
)
5983 unsigned long flags
;
5985 { u64 psr
= pfm_get_psr();
5986 BUG_ON(psr
& IA64_PSR_UP
);
5989 ctx
= PFM_GET_CTX(task
);
5992 * we need to mask PMU overflow here to
5993 * make sure that we maintain pmc0 until
5994 * we save it. overflow interrupts are
5995 * treated as spurious if there is no
5998 * XXX: I don't think this is necessary
6000 PROTECT_CTX(ctx
,flags
);
6003 * release ownership of this PMU.
6004 * must be done before we save the registers.
6006 * after this call any PMU interrupt is treated
6009 SET_PMU_OWNER(NULL
, NULL
);
6012 * save all the pmds we use
6014 pfm_save_pmds(ctx
->th_pmds
, ctx
->ctx_used_pmds
[0]);
6017 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6018 * it is needed to check for pended overflow
6019 * on the restore path
6021 ctx
->th_pmcs
[0] = ia64_get_pmc(0);
6024 * unfreeze PMU if had pending overflows
6026 if (ctx
->th_pmcs
[0] & ~0x1UL
) pfm_unfreeze_pmu();
6029 * now get can unmask PMU interrupts, they will
6030 * be treated as purely spurious and we will not
6031 * lose any information
6033 UNPROTECT_CTX(ctx
,flags
);
6035 #endif /* CONFIG_SMP */
6039 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6042 pfm_load_regs (struct task_struct
*task
)
6045 unsigned long pmc_mask
= 0UL, pmd_mask
= 0UL;
6046 unsigned long flags
;
6048 int need_irq_resend
;
6050 ctx
= PFM_GET_CTX(task
);
6051 if (unlikely(ctx
== NULL
)) return;
6053 BUG_ON(GET_PMU_OWNER());
6056 * possible on unload
6058 if (unlikely((task
->thread
.flags
& IA64_THREAD_PM_VALID
) == 0)) return;
6061 * we always come here with interrupts ALREADY disabled by
6062 * the scheduler. So we simply need to protect against concurrent
6063 * access, not CPU concurrency.
6065 flags
= pfm_protect_ctx_ctxsw(ctx
);
6066 psr
= pfm_get_psr();
6068 need_irq_resend
= pmu_conf
->flags
& PFM_PMU_IRQ_RESEND
;
6070 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
6071 BUG_ON(psr
& IA64_PSR_I
);
6073 if (unlikely(ctx
->ctx_state
== PFM_CTX_ZOMBIE
)) {
6074 struct pt_regs
*regs
= task_pt_regs(task
);
6076 BUG_ON(ctx
->ctx_smpl_hdr
);
6078 pfm_force_cleanup(ctx
, regs
);
6080 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
6083 * this one (kmalloc'ed) is fine with interrupts disabled
6085 pfm_context_free(ctx
);
6091 * we restore ALL the debug registers to avoid picking up
6094 if (ctx
->ctx_fl_using_dbreg
) {
6095 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
6096 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
6099 * retrieve saved psr.up
6101 psr_up
= ctx
->ctx_saved_psr_up
;
6104 * if we were the last user of the PMU on that CPU,
6105 * then nothing to do except restore psr
6107 if (GET_LAST_CPU(ctx
) == smp_processor_id() && ctx
->ctx_last_activation
== GET_ACTIVATION()) {
6110 * retrieve partial reload masks (due to user modifications)
6112 pmc_mask
= ctx
->ctx_reload_pmcs
[0];
6113 pmd_mask
= ctx
->ctx_reload_pmds
[0];
6117 * To avoid leaking information to the user level when psr.sp=0,
6118 * we must reload ALL implemented pmds (even the ones we don't use).
6119 * In the kernel we only allow PFM_READ_PMDS on registers which
6120 * we initialized or requested (sampling) so there is no risk there.
6122 pmd_mask
= pfm_sysctl
.fastctxsw
? ctx
->ctx_used_pmds
[0] : ctx
->ctx_all_pmds
[0];
6125 * ALL accessible PMCs are systematically reloaded, unused registers
6126 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6127 * up stale configuration.
6129 * PMC0 is never in the mask. It is always restored separately.
6131 pmc_mask
= ctx
->ctx_all_pmcs
[0];
6134 * when context is MASKED, we will restore PMC with plm=0
6135 * and PMD with stale information, but that's ok, nothing
6138 * XXX: optimize here
6140 if (pmd_mask
) pfm_restore_pmds(ctx
->th_pmds
, pmd_mask
);
6141 if (pmc_mask
) pfm_restore_pmcs(ctx
->th_pmcs
, pmc_mask
);
6144 * check for pending overflow at the time the state
6147 if (unlikely(PMC0_HAS_OVFL(ctx
->th_pmcs
[0]))) {
6149 * reload pmc0 with the overflow information
6150 * On McKinley PMU, this will trigger a PMU interrupt
6152 ia64_set_pmc(0, ctx
->th_pmcs
[0]);
6154 ctx
->th_pmcs
[0] = 0UL;
6157 * will replay the PMU interrupt
6159 if (need_irq_resend
) ia64_resend_irq(IA64_PERFMON_VECTOR
);
6161 pfm_stats
[smp_processor_id()].pfm_replay_ovfl_intr_count
++;
6165 * we just did a reload, so we reset the partial reload fields
6167 ctx
->ctx_reload_pmcs
[0] = 0UL;
6168 ctx
->ctx_reload_pmds
[0] = 0UL;
6170 SET_LAST_CPU(ctx
, smp_processor_id());
6173 * dump activation value for this PMU
6177 * record current activation for this context
6179 SET_ACTIVATION(ctx
);
6182 * establish new ownership.
6184 SET_PMU_OWNER(task
, ctx
);
6187 * restore the psr.up bit. measurement
6189 * no PMU interrupt can happen at this point
6190 * because we still have interrupts disabled.
6192 if (likely(psr_up
)) pfm_set_psr_up();
6195 * allow concurrent access to context
6197 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
6199 #else /* !CONFIG_SMP */
6201 * reload PMU state for UP kernels
6202 * in 2.5 we come here with interrupts disabled
6205 pfm_load_regs (struct task_struct
*task
)
6208 struct task_struct
*owner
;
6209 unsigned long pmd_mask
, pmc_mask
;
6211 int need_irq_resend
;
6213 owner
= GET_PMU_OWNER();
6214 ctx
= PFM_GET_CTX(task
);
6215 psr
= pfm_get_psr();
6217 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
6218 BUG_ON(psr
& IA64_PSR_I
);
6221 * we restore ALL the debug registers to avoid picking up
6224 * This must be done even when the task is still the owner
6225 * as the registers may have been modified via ptrace()
6226 * (not perfmon) by the previous task.
6228 if (ctx
->ctx_fl_using_dbreg
) {
6229 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
6230 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
6234 * retrieved saved psr.up
6236 psr_up
= ctx
->ctx_saved_psr_up
;
6237 need_irq_resend
= pmu_conf
->flags
& PFM_PMU_IRQ_RESEND
;
6240 * short path, our state is still there, just
6241 * need to restore psr and we go
6243 * we do not touch either PMC nor PMD. the psr is not touched
6244 * by the overflow_handler. So we are safe w.r.t. to interrupt
6245 * concurrency even without interrupt masking.
6247 if (likely(owner
== task
)) {
6248 if (likely(psr_up
)) pfm_set_psr_up();
6253 * someone else is still using the PMU, first push it out and
6254 * then we'll be able to install our stuff !
6256 * Upon return, there will be no owner for the current PMU
6258 if (owner
) pfm_lazy_save_regs(owner
);
6261 * To avoid leaking information to the user level when psr.sp=0,
6262 * we must reload ALL implemented pmds (even the ones we don't use).
6263 * In the kernel we only allow PFM_READ_PMDS on registers which
6264 * we initialized or requested (sampling) so there is no risk there.
6266 pmd_mask
= pfm_sysctl
.fastctxsw
? ctx
->ctx_used_pmds
[0] : ctx
->ctx_all_pmds
[0];
6269 * ALL accessible PMCs are systematically reloaded, unused registers
6270 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6271 * up stale configuration.
6273 * PMC0 is never in the mask. It is always restored separately
6275 pmc_mask
= ctx
->ctx_all_pmcs
[0];
6277 pfm_restore_pmds(ctx
->th_pmds
, pmd_mask
);
6278 pfm_restore_pmcs(ctx
->th_pmcs
, pmc_mask
);
6281 * check for pending overflow at the time the state
6284 if (unlikely(PMC0_HAS_OVFL(ctx
->th_pmcs
[0]))) {
6286 * reload pmc0 with the overflow information
6287 * On McKinley PMU, this will trigger a PMU interrupt
6289 ia64_set_pmc(0, ctx
->th_pmcs
[0]);
6292 ctx
->th_pmcs
[0] = 0UL;
6295 * will replay the PMU interrupt
6297 if (need_irq_resend
) ia64_resend_irq(IA64_PERFMON_VECTOR
);
6299 pfm_stats
[smp_processor_id()].pfm_replay_ovfl_intr_count
++;
6303 * establish new ownership.
6305 SET_PMU_OWNER(task
, ctx
);
6308 * restore the psr.up bit. measurement
6310 * no PMU interrupt can happen at this point
6311 * because we still have interrupts disabled.
6313 if (likely(psr_up
)) pfm_set_psr_up();
6315 #endif /* CONFIG_SMP */
6318 * this function assumes monitoring is stopped
6321 pfm_flush_pmds(struct task_struct
*task
, pfm_context_t
*ctx
)
6324 unsigned long mask2
, val
, pmd_val
, ovfl_val
;
6325 int i
, can_access_pmu
= 0;
6329 * is the caller the task being monitored (or which initiated the
6330 * session for system wide measurements)
6332 is_self
= ctx
->ctx_task
== task
? 1 : 0;
6335 * can access PMU is task is the owner of the PMU state on the current CPU
6336 * or if we are running on the CPU bound to the context in system-wide mode
6337 * (that is not necessarily the task the context is attached to in this mode).
6338 * In system-wide we always have can_access_pmu true because a task running on an
6339 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6341 can_access_pmu
= (GET_PMU_OWNER() == task
) || (ctx
->ctx_fl_system
&& ctx
->ctx_cpu
== smp_processor_id());
6342 if (can_access_pmu
) {
6344 * Mark the PMU as not owned
6345 * This will cause the interrupt handler to do nothing in case an overflow
6346 * interrupt was in-flight
6347 * This also guarantees that pmc0 will contain the final state
6348 * It virtually gives us full control on overflow processing from that point
6351 SET_PMU_OWNER(NULL
, NULL
);
6352 DPRINT(("releasing ownership\n"));
6355 * read current overflow status:
6357 * we are guaranteed to read the final stable state
6360 pmc0
= ia64_get_pmc(0); /* slow */
6363 * reset freeze bit, overflow status information destroyed
6367 pmc0
= ctx
->th_pmcs
[0];
6369 * clear whatever overflow status bits there were
6371 ctx
->th_pmcs
[0] = 0;
6373 ovfl_val
= pmu_conf
->ovfl_val
;
6375 * we save all the used pmds
6376 * we take care of overflows for counting PMDs
6378 * XXX: sampling situation is not taken into account here
6380 mask2
= ctx
->ctx_used_pmds
[0];
6382 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self
, ovfl_val
, mask2
));
6384 for (i
= 0; mask2
; i
++, mask2
>>=1) {
6386 /* skip non used pmds */
6387 if ((mask2
& 0x1) == 0) continue;
6390 * can access PMU always true in system wide mode
6392 val
= pmd_val
= can_access_pmu
? ia64_get_pmd(i
) : ctx
->th_pmds
[i
];
6394 if (PMD_IS_COUNTING(i
)) {
6395 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6398 ctx
->ctx_pmds
[i
].val
,
6402 * we rebuild the full 64 bit value of the counter
6404 val
= ctx
->ctx_pmds
[i
].val
+ (val
& ovfl_val
);
6407 * now everything is in ctx_pmds[] and we need
6408 * to clear the saved context from save_regs() such that
6409 * pfm_read_pmds() gets the correct value
6414 * take care of overflow inline
6416 if (pmc0
& (1UL << i
)) {
6417 val
+= 1 + ovfl_val
;
6418 DPRINT(("[%d] pmd[%d] overflowed\n", task
->pid
, i
));
6422 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task
->pid
, i
, val
, pmd_val
));
6424 if (is_self
) ctx
->th_pmds
[i
] = pmd_val
;
6426 ctx
->ctx_pmds
[i
].val
= val
;
6430 static struct irqaction perfmon_irqaction
= {
6431 .handler
= pfm_interrupt_handler
,
6432 .flags
= IRQF_DISABLED
,
6437 pfm_alt_save_pmu_state(void *data
)
6439 struct pt_regs
*regs
;
6441 regs
= task_pt_regs(current
);
6443 DPRINT(("called\n"));
6446 * should not be necessary but
6447 * let's take not risk
6451 ia64_psr(regs
)->pp
= 0;
6454 * This call is required
6455 * May cause a spurious interrupt on some processors
6463 pfm_alt_restore_pmu_state(void *data
)
6465 struct pt_regs
*regs
;
6467 regs
= task_pt_regs(current
);
6469 DPRINT(("called\n"));
6472 * put PMU back in state expected
6477 ia64_psr(regs
)->pp
= 0;
6480 * perfmon runs with PMU unfrozen at all times
6488 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t
*hdl
)
6493 /* some sanity checks */
6494 if (hdl
== NULL
|| hdl
->handler
== NULL
) return -EINVAL
;
6496 /* do the easy test first */
6497 if (pfm_alt_intr_handler
) return -EBUSY
;
6499 /* one at a time in the install or remove, just fail the others */
6500 if (!spin_trylock(&pfm_alt_install_check
)) {
6504 /* reserve our session */
6505 for_each_online_cpu(reserve_cpu
) {
6506 ret
= pfm_reserve_session(NULL
, 1, reserve_cpu
);
6507 if (ret
) goto cleanup_reserve
;
6510 /* save the current system wide pmu states */
6511 ret
= on_each_cpu(pfm_alt_save_pmu_state
, NULL
, 0, 1);
6513 DPRINT(("on_each_cpu() failed: %d\n", ret
));
6514 goto cleanup_reserve
;
6517 /* officially change to the alternate interrupt handler */
6518 pfm_alt_intr_handler
= hdl
;
6520 spin_unlock(&pfm_alt_install_check
);
6525 for_each_online_cpu(i
) {
6526 /* don't unreserve more than we reserved */
6527 if (i
>= reserve_cpu
) break;
6529 pfm_unreserve_session(NULL
, 1, i
);
6532 spin_unlock(&pfm_alt_install_check
);
6536 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt
);
6539 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t
*hdl
)
6544 if (hdl
== NULL
) return -EINVAL
;
6546 /* cannot remove someone else's handler! */
6547 if (pfm_alt_intr_handler
!= hdl
) return -EINVAL
;
6549 /* one at a time in the install or remove, just fail the others */
6550 if (!spin_trylock(&pfm_alt_install_check
)) {
6554 pfm_alt_intr_handler
= NULL
;
6556 ret
= on_each_cpu(pfm_alt_restore_pmu_state
, NULL
, 0, 1);
6558 DPRINT(("on_each_cpu() failed: %d\n", ret
));
6561 for_each_online_cpu(i
) {
6562 pfm_unreserve_session(NULL
, 1, i
);
6565 spin_unlock(&pfm_alt_install_check
);
6569 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt
);
6572 * perfmon initialization routine, called from the initcall() table
6574 static int init_pfm_fs(void);
6582 family
= local_cpu_data
->family
;
6587 if ((*p
)->probe() == 0) goto found
;
6588 } else if ((*p
)->pmu_family
== family
|| (*p
)->pmu_family
== 0xff) {
6599 static const struct file_operations pfm_proc_fops
= {
6600 .open
= pfm_proc_open
,
6602 .llseek
= seq_lseek
,
6603 .release
= seq_release
,
6609 unsigned int n
, n_counters
, i
;
6611 printk("perfmon: version %u.%u IRQ %u\n",
6614 IA64_PERFMON_VECTOR
);
6616 if (pfm_probe_pmu()) {
6617 printk(KERN_INFO
"perfmon: disabled, there is no support for processor family %d\n",
6618 local_cpu_data
->family
);
6623 * compute the number of implemented PMD/PMC from the
6624 * description tables
6627 for (i
=0; PMC_IS_LAST(i
) == 0; i
++) {
6628 if (PMC_IS_IMPL(i
) == 0) continue;
6629 pmu_conf
->impl_pmcs
[i
>>6] |= 1UL << (i
&63);
6632 pmu_conf
->num_pmcs
= n
;
6634 n
= 0; n_counters
= 0;
6635 for (i
=0; PMD_IS_LAST(i
) == 0; i
++) {
6636 if (PMD_IS_IMPL(i
) == 0) continue;
6637 pmu_conf
->impl_pmds
[i
>>6] |= 1UL << (i
&63);
6639 if (PMD_IS_COUNTING(i
)) n_counters
++;
6641 pmu_conf
->num_pmds
= n
;
6642 pmu_conf
->num_counters
= n_counters
;
6645 * sanity checks on the number of debug registers
6647 if (pmu_conf
->use_rr_dbregs
) {
6648 if (pmu_conf
->num_ibrs
> IA64_NUM_DBG_REGS
) {
6649 printk(KERN_INFO
"perfmon: unsupported number of code debug registers (%u)\n", pmu_conf
->num_ibrs
);
6653 if (pmu_conf
->num_dbrs
> IA64_NUM_DBG_REGS
) {
6654 printk(KERN_INFO
"perfmon: unsupported number of data debug registers (%u)\n", pmu_conf
->num_ibrs
);
6660 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6664 pmu_conf
->num_counters
,
6665 ffz(pmu_conf
->ovfl_val
));
6668 if (pmu_conf
->num_pmds
>= PFM_NUM_PMD_REGS
|| pmu_conf
->num_pmcs
>= PFM_NUM_PMC_REGS
) {
6669 printk(KERN_ERR
"perfmon: not enough pmc/pmd, perfmon disabled\n");
6675 * create /proc/perfmon (mostly for debugging purposes)
6677 perfmon_dir
= create_proc_entry("perfmon", S_IRUGO
, NULL
);
6678 if (perfmon_dir
== NULL
) {
6679 printk(KERN_ERR
"perfmon: cannot create /proc entry, perfmon disabled\n");
6684 * install customized file operations for /proc/perfmon entry
6686 perfmon_dir
->proc_fops
= &pfm_proc_fops
;
6689 * create /proc/sys/kernel/perfmon (for debugging purposes)
6691 pfm_sysctl_header
= register_sysctl_table(pfm_sysctl_root
, 0);
6694 * initialize all our spinlocks
6696 spin_lock_init(&pfm_sessions
.pfs_lock
);
6697 spin_lock_init(&pfm_buffer_fmt_lock
);
6701 for(i
=0; i
< NR_CPUS
; i
++) pfm_stats
[i
].pfm_ovfl_intr_cycles_min
= ~0UL;
6706 __initcall(pfm_init
);
6709 * this function is called before pfm_init()
6712 pfm_init_percpu (void)
6714 static int first_time
=1;
6716 * make sure no measurement is active
6717 * (may inherit programmed PMCs from EFI).
6723 * we run with the PMU not frozen at all times
6728 register_percpu_irq(IA64_PERFMON_VECTOR
, &perfmon_irqaction
);
6732 ia64_setreg(_IA64_REG_CR_PMV
, IA64_PERFMON_VECTOR
);
6737 * used for debug purposes only
6740 dump_pmu_state(const char *from
)
6742 struct task_struct
*task
;
6743 struct pt_regs
*regs
;
6745 unsigned long psr
, dcr
, info
, flags
;
6748 local_irq_save(flags
);
6750 this_cpu
= smp_processor_id();
6751 regs
= task_pt_regs(current
);
6752 info
= PFM_CPUINFO_GET();
6753 dcr
= ia64_getreg(_IA64_REG_CR_DCR
);
6755 if (info
== 0 && ia64_psr(regs
)->pp
== 0 && (dcr
& IA64_DCR_PP
) == 0) {
6756 local_irq_restore(flags
);
6760 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6767 task
= GET_PMU_OWNER();
6768 ctx
= GET_PMU_CTX();
6770 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu
, task
? task
->pid
: -1, ctx
);
6772 psr
= pfm_get_psr();
6774 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",
6777 psr
& IA64_PSR_PP
? 1 : 0,
6778 psr
& IA64_PSR_UP
? 1 : 0,
6779 dcr
& IA64_DCR_PP
? 1 : 0,
6782 ia64_psr(regs
)->pp
);
6784 ia64_psr(regs
)->up
= 0;
6785 ia64_psr(regs
)->pp
= 0;
6787 for (i
=1; PMC_IS_LAST(i
) == 0; i
++) {
6788 if (PMC_IS_IMPL(i
) == 0) continue;
6789 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
]);
6792 for (i
=1; PMD_IS_LAST(i
) == 0; i
++) {
6793 if (PMD_IS_IMPL(i
) == 0) continue;
6794 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
]);
6798 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6801 ctx
->ctx_smpl_vaddr
,
6805 ctx
->ctx_saved_psr_up
);
6807 local_irq_restore(flags
);
6811 * called from process.c:copy_thread(). task is new child.
6814 pfm_inherit(struct task_struct
*task
, struct pt_regs
*regs
)
6816 struct thread_struct
*thread
;
6818 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task
->pid
));
6820 thread
= &task
->thread
;
6823 * cut links inherited from parent (current)
6825 thread
->pfm_context
= NULL
;
6827 PFM_SET_WORK_PENDING(task
, 0);
6830 * the psr bits are already set properly in copy_threads()
6833 #else /* !CONFIG_PERFMON */
6835 sys_perfmonctl (int fd
, int cmd
, void *arg
, int count
)
6839 #endif /* CONFIG_PERFMON */