1 // SPDX-License-Identifier: GPL-2.0-only
3 * This file implements the perfmon-2 subsystem which is used
4 * to program the IA-64 Performance Monitoring Unit (PMU).
6 * The initial version of perfmon.c was written by
7 * Ganesh Venkitachalam, IBM Corp.
9 * Then it was modified for perfmon-1.x by Stephane Eranian and
10 * David Mosberger, Hewlett Packard Co.
12 * Version Perfmon-2.x is a rewrite of perfmon-1.x
13 * by Stephane Eranian, Hewlett Packard Co.
15 * Copyright (C) 1999-2005 Hewlett Packard Co
16 * Stephane Eranian <eranian@hpl.hp.com>
17 * David Mosberger-Tang <davidm@hpl.hp.com>
19 * More information about perfmon available at:
20 * http://www.hpl.hp.com/research/linux/perfmon
23 #include <linux/module.h>
24 #include <linux/kernel.h>
25 #include <linux/sched.h>
26 #include <linux/sched/task.h>
27 #include <linux/sched/task_stack.h>
28 #include <linux/interrupt.h>
29 #include <linux/proc_fs.h>
30 #include <linux/seq_file.h>
31 #include <linux/init.h>
32 #include <linux/vmalloc.h>
34 #include <linux/sysctl.h>
35 #include <linux/list.h>
36 #include <linux/file.h>
37 #include <linux/poll.h>
38 #include <linux/vfs.h>
39 #include <linux/smp.h>
40 #include <linux/pagemap.h>
41 #include <linux/mount.h>
42 #include <linux/pseudo_fs.h>
43 #include <linux/bitops.h>
44 #include <linux/capability.h>
45 #include <linux/rcupdate.h>
46 #include <linux/completion.h>
47 #include <linux/tracehook.h>
48 #include <linux/slab.h>
49 #include <linux/cpu.h>
51 #include <asm/errno.h>
52 #include <asm/intrinsics.h>
54 #include <asm/perfmon.h>
55 #include <asm/processor.h>
56 #include <asm/signal.h>
57 #include <linux/uaccess.h>
58 #include <asm/delay.h>
64 * perfmon context state
66 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
67 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
68 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
69 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
71 #define PFM_INVALID_ACTIVATION (~0UL)
73 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
74 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
77 * depth of message queue
79 #define PFM_MAX_MSGS 32
80 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
83 * type of a PMU register (bitmask).
85 * bit0 : register implemented
88 * bit4 : pmc has pmc.pm
89 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
90 * bit6-7 : register type
93 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
94 #define PFM_REG_IMPL 0x1 /* register implemented */
95 #define PFM_REG_END 0x2 /* end marker */
96 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
97 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
98 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
99 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
100 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
102 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
103 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
105 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
107 /* i assumed unsigned */
108 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
109 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
111 /* XXX: these assume that register i is implemented */
112 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
113 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
114 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
115 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
117 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
118 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
119 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
120 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
122 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
123 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
125 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
126 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
127 #define PFM_CTX_TASK(h) (h)->ctx_task
129 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
131 /* XXX: does not support more than 64 PMDs */
132 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
133 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
135 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
137 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
138 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
139 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
140 #define PFM_CODE_RR 0 /* requesting code range restriction */
141 #define PFM_DATA_RR 1 /* requestion data range restriction */
143 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
144 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
145 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
147 #define RDEP(x) (1UL<<(x))
150 * context protection macros
152 * - we need to protect against CPU concurrency (spin_lock)
153 * - we need to protect against PMU overflow interrupts (local_irq_disable)
155 * - we need to protect against PMU overflow interrupts (local_irq_disable)
157 * spin_lock_irqsave()/spin_unlock_irqrestore():
158 * in SMP: local_irq_disable + spin_lock
159 * in UP : local_irq_disable
161 * spin_lock()/spin_lock():
162 * in UP : removed automatically
163 * in SMP: protect against context accesses from other CPU. interrupts
164 * are not masked. This is useful for the PMU interrupt handler
165 * because we know we will not get PMU concurrency in that code.
167 #define PROTECT_CTX(c, f) \
169 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
170 spin_lock_irqsave(&(c)->ctx_lock, f); \
171 DPRINT(("spinlocked ctx %p by [%d]\n", c, task_pid_nr(current))); \
174 #define UNPROTECT_CTX(c, f) \
176 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
177 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
180 #define PROTECT_CTX_NOPRINT(c, f) \
182 spin_lock_irqsave(&(c)->ctx_lock, f); \
186 #define UNPROTECT_CTX_NOPRINT(c, f) \
188 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
192 #define PROTECT_CTX_NOIRQ(c) \
194 spin_lock(&(c)->ctx_lock); \
197 #define UNPROTECT_CTX_NOIRQ(c) \
199 spin_unlock(&(c)->ctx_lock); \
205 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
206 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
207 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
209 #else /* !CONFIG_SMP */
210 #define SET_ACTIVATION(t) do {} while(0)
211 #define GET_ACTIVATION(t) do {} while(0)
212 #define INC_ACTIVATION(t) do {} while(0)
213 #endif /* CONFIG_SMP */
215 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
216 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
217 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
219 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
220 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
222 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
225 * cmp0 must be the value of pmc0
227 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
229 #define PFMFS_MAGIC 0xa0b4d889
234 #define PFM_DEBUGGING 1
238 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
241 #define DPRINT_ovfl(a) \
243 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
248 * 64-bit software counter structure
250 * the next_reset_type is applied to the next call to pfm_reset_regs()
253 unsigned long val
; /* virtual 64bit counter value */
254 unsigned long lval
; /* last reset value */
255 unsigned long long_reset
; /* reset value on sampling overflow */
256 unsigned long short_reset
; /* reset value on overflow */
257 unsigned long reset_pmds
[4]; /* which other pmds to reset when this counter overflows */
258 unsigned long smpl_pmds
[4]; /* which pmds are accessed when counter overflow */
259 unsigned long seed
; /* seed for random-number generator */
260 unsigned long mask
; /* mask for random-number generator */
261 unsigned int flags
; /* notify/do not notify */
262 unsigned long eventid
; /* overflow event identifier */
269 unsigned int block
:1; /* when 1, task will blocked on user notifications */
270 unsigned int system
:1; /* do system wide monitoring */
271 unsigned int using_dbreg
:1; /* using range restrictions (debug registers) */
272 unsigned int is_sampling
:1; /* true if using a custom format */
273 unsigned int excl_idle
:1; /* exclude idle task in system wide session */
274 unsigned int going_zombie
:1; /* context is zombie (MASKED+blocking) */
275 unsigned int trap_reason
:2; /* reason for going into pfm_handle_work() */
276 unsigned int no_msg
:1; /* no message sent on overflow */
277 unsigned int can_restart
:1; /* allowed to issue a PFM_RESTART */
278 unsigned int reserved
:22;
279 } pfm_context_flags_t
;
281 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
282 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
283 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
287 * perfmon context: encapsulates all the state of a monitoring session
290 typedef struct pfm_context
{
291 spinlock_t ctx_lock
; /* context protection */
293 pfm_context_flags_t ctx_flags
; /* bitmask of flags (block reason incl.) */
294 unsigned int ctx_state
; /* state: active/inactive (no bitfield) */
296 struct task_struct
*ctx_task
; /* task to which context is attached */
298 unsigned long ctx_ovfl_regs
[4]; /* which registers overflowed (notification) */
300 struct completion ctx_restart_done
; /* use for blocking notification mode */
302 unsigned long ctx_used_pmds
[4]; /* bitmask of PMD used */
303 unsigned long ctx_all_pmds
[4]; /* bitmask of all accessible PMDs */
304 unsigned long ctx_reload_pmds
[4]; /* bitmask of force reload PMD on ctxsw in */
306 unsigned long ctx_all_pmcs
[4]; /* bitmask of all accessible PMCs */
307 unsigned long ctx_reload_pmcs
[4]; /* bitmask of force reload PMC on ctxsw in */
308 unsigned long ctx_used_monitors
[4]; /* bitmask of monitor PMC being used */
310 unsigned long ctx_pmcs
[PFM_NUM_PMC_REGS
]; /* saved copies of PMC values */
312 unsigned int ctx_used_ibrs
[1]; /* bitmask of used IBR (speedup ctxsw in) */
313 unsigned int ctx_used_dbrs
[1]; /* bitmask of used DBR (speedup ctxsw in) */
314 unsigned long ctx_dbrs
[IA64_NUM_DBG_REGS
]; /* DBR values (cache) when not loaded */
315 unsigned long ctx_ibrs
[IA64_NUM_DBG_REGS
]; /* IBR values (cache) when not loaded */
317 pfm_counter_t ctx_pmds
[PFM_NUM_PMD_REGS
]; /* software state for PMDS */
319 unsigned long th_pmcs
[PFM_NUM_PMC_REGS
]; /* PMC thread save state */
320 unsigned long th_pmds
[PFM_NUM_PMD_REGS
]; /* PMD thread save state */
322 unsigned long ctx_saved_psr_up
; /* only contains psr.up value */
324 unsigned long ctx_last_activation
; /* context last activation number for last_cpu */
325 unsigned int ctx_last_cpu
; /* CPU id of current or last CPU used (SMP only) */
326 unsigned int ctx_cpu
; /* cpu to which perfmon is applied (system wide) */
328 int ctx_fd
; /* file descriptor used my this context */
329 pfm_ovfl_arg_t ctx_ovfl_arg
; /* argument to custom buffer format handler */
331 pfm_buffer_fmt_t
*ctx_buf_fmt
; /* buffer format callbacks */
332 void *ctx_smpl_hdr
; /* points to sampling buffer header kernel vaddr */
333 unsigned long ctx_smpl_size
; /* size of sampling buffer */
334 void *ctx_smpl_vaddr
; /* user level virtual address of smpl buffer */
336 wait_queue_head_t ctx_msgq_wait
;
337 pfm_msg_t ctx_msgq
[PFM_MAX_MSGS
];
340 struct fasync_struct
*ctx_async_queue
;
342 wait_queue_head_t ctx_zombieq
; /* termination cleanup wait queue */
346 * magic number used to verify that structure is really
349 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
351 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
354 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
355 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
357 #define SET_LAST_CPU(ctx, v) do {} while(0)
358 #define GET_LAST_CPU(ctx) do {} while(0)
362 #define ctx_fl_block ctx_flags.block
363 #define ctx_fl_system ctx_flags.system
364 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
365 #define ctx_fl_is_sampling ctx_flags.is_sampling
366 #define ctx_fl_excl_idle ctx_flags.excl_idle
367 #define ctx_fl_going_zombie ctx_flags.going_zombie
368 #define ctx_fl_trap_reason ctx_flags.trap_reason
369 #define ctx_fl_no_msg ctx_flags.no_msg
370 #define ctx_fl_can_restart ctx_flags.can_restart
372 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
373 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
376 * global information about all sessions
377 * mostly used to synchronize between system wide and per-process
380 spinlock_t pfs_lock
; /* lock the structure */
382 unsigned int pfs_task_sessions
; /* number of per task sessions */
383 unsigned int pfs_sys_sessions
; /* number of per system wide sessions */
384 unsigned int pfs_sys_use_dbregs
; /* incremented when a system wide session uses debug regs */
385 unsigned int pfs_ptrace_use_dbregs
; /* incremented when a process uses debug regs */
386 struct task_struct
*pfs_sys_session
[NR_CPUS
]; /* point to task owning a system-wide session */
390 * information about a PMC or PMD.
391 * dep_pmd[]: a bitmask of dependent PMD registers
392 * dep_pmc[]: a bitmask of dependent PMC registers
394 typedef int (*pfm_reg_check_t
)(struct task_struct
*task
, pfm_context_t
*ctx
, unsigned int cnum
, unsigned long *val
, struct pt_regs
*regs
);
398 unsigned long default_value
; /* power-on default value */
399 unsigned long reserved_mask
; /* bitmask of reserved bits */
400 pfm_reg_check_t read_check
;
401 pfm_reg_check_t write_check
;
402 unsigned long dep_pmd
[4];
403 unsigned long dep_pmc
[4];
406 /* assume cnum is a valid monitor */
407 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
410 * This structure is initialized at boot time and contains
411 * a description of the PMU main characteristics.
413 * If the probe function is defined, detection is based
414 * on its return value:
415 * - 0 means recognized PMU
416 * - anything else means not supported
417 * When the probe function is not defined, then the pmu_family field
418 * is used and it must match the host CPU family such that:
419 * - cpu->family & config->pmu_family != 0
422 unsigned long ovfl_val
; /* overflow value for counters */
424 pfm_reg_desc_t
*pmc_desc
; /* detailed PMC register dependencies descriptions */
425 pfm_reg_desc_t
*pmd_desc
; /* detailed PMD register dependencies descriptions */
427 unsigned int num_pmcs
; /* number of PMCS: computed at init time */
428 unsigned int num_pmds
; /* number of PMDS: computed at init time */
429 unsigned long impl_pmcs
[4]; /* bitmask of implemented PMCS */
430 unsigned long impl_pmds
[4]; /* bitmask of implemented PMDS */
432 char *pmu_name
; /* PMU family name */
433 unsigned int pmu_family
; /* cpuid family pattern used to identify pmu */
434 unsigned int flags
; /* pmu specific flags */
435 unsigned int num_ibrs
; /* number of IBRS: computed at init time */
436 unsigned int num_dbrs
; /* number of DBRS: computed at init time */
437 unsigned int num_counters
; /* PMC/PMD counting pairs : computed at init time */
438 int (*probe
)(void); /* customized probe routine */
439 unsigned int use_rr_dbregs
:1; /* set if debug registers used for range restriction */
444 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
447 * debug register related type definitions
450 unsigned long ibr_mask
:56;
451 unsigned long ibr_plm
:4;
452 unsigned long ibr_ig
:3;
453 unsigned long ibr_x
:1;
457 unsigned long dbr_mask
:56;
458 unsigned long dbr_plm
:4;
459 unsigned long dbr_ig
:2;
460 unsigned long dbr_w
:1;
461 unsigned long dbr_r
:1;
472 * perfmon command descriptions
475 int (*cmd_func
)(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
478 unsigned int cmd_narg
;
480 int (*cmd_getsize
)(void *arg
, size_t *sz
);
483 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
484 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
485 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
486 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
489 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
490 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
491 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
492 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
493 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
495 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
498 unsigned long pfm_spurious_ovfl_intr_count
; /* keep track of spurious ovfl interrupts */
499 unsigned long pfm_replay_ovfl_intr_count
; /* keep track of replayed ovfl interrupts */
500 unsigned long pfm_ovfl_intr_count
; /* keep track of ovfl interrupts */
501 unsigned long pfm_ovfl_intr_cycles
; /* cycles spent processing ovfl interrupts */
502 unsigned long pfm_ovfl_intr_cycles_min
; /* min cycles spent processing ovfl interrupts */
503 unsigned long pfm_ovfl_intr_cycles_max
; /* max cycles spent processing ovfl interrupts */
504 unsigned long pfm_smpl_handler_calls
;
505 unsigned long pfm_smpl_handler_cycles
;
506 char pad
[SMP_CACHE_BYTES
] ____cacheline_aligned
;
510 * perfmon internal variables
512 static pfm_stats_t pfm_stats
[NR_CPUS
];
513 static pfm_session_t pfm_sessions
; /* global sessions information */
515 static DEFINE_SPINLOCK(pfm_alt_install_check
);
516 static pfm_intr_handler_desc_t
*pfm_alt_intr_handler
;
518 static struct proc_dir_entry
*perfmon_dir
;
519 static pfm_uuid_t pfm_null_uuid
= {0,};
521 static spinlock_t pfm_buffer_fmt_lock
;
522 static LIST_HEAD(pfm_buffer_fmt_list
);
524 static pmu_config_t
*pmu_conf
;
526 /* sysctl() controls */
527 pfm_sysctl_t pfm_sysctl
;
528 EXPORT_SYMBOL(pfm_sysctl
);
530 static struct ctl_table pfm_ctl_table
[] = {
533 .data
= &pfm_sysctl
.debug
,
534 .maxlen
= sizeof(int),
536 .proc_handler
= proc_dointvec
,
539 .procname
= "debug_ovfl",
540 .data
= &pfm_sysctl
.debug_ovfl
,
541 .maxlen
= sizeof(int),
543 .proc_handler
= proc_dointvec
,
546 .procname
= "fastctxsw",
547 .data
= &pfm_sysctl
.fastctxsw
,
548 .maxlen
= sizeof(int),
550 .proc_handler
= proc_dointvec
,
553 .procname
= "expert_mode",
554 .data
= &pfm_sysctl
.expert_mode
,
555 .maxlen
= sizeof(int),
557 .proc_handler
= proc_dointvec
,
561 static struct ctl_table pfm_sysctl_dir
[] = {
563 .procname
= "perfmon",
565 .child
= pfm_ctl_table
,
569 static struct ctl_table pfm_sysctl_root
[] = {
571 .procname
= "kernel",
573 .child
= pfm_sysctl_dir
,
577 static struct ctl_table_header
*pfm_sysctl_header
;
579 static int pfm_context_unload(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
581 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
582 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
585 pfm_put_task(struct task_struct
*task
)
587 if (task
!= current
) put_task_struct(task
);
590 static inline unsigned long
591 pfm_protect_ctx_ctxsw(pfm_context_t
*x
)
593 spin_lock(&(x
)->ctx_lock
);
598 pfm_unprotect_ctx_ctxsw(pfm_context_t
*x
, unsigned long f
)
600 spin_unlock(&(x
)->ctx_lock
);
603 /* forward declaration */
604 static const struct dentry_operations pfmfs_dentry_operations
;
606 static int pfmfs_init_fs_context(struct fs_context
*fc
)
608 struct pseudo_fs_context
*ctx
= init_pseudo(fc
, PFMFS_MAGIC
);
611 ctx
->dops
= &pfmfs_dentry_operations
;
615 static struct file_system_type pfm_fs_type
= {
617 .init_fs_context
= pfmfs_init_fs_context
,
618 .kill_sb
= kill_anon_super
,
620 MODULE_ALIAS_FS("pfmfs");
622 DEFINE_PER_CPU(unsigned long, pfm_syst_info
);
623 DEFINE_PER_CPU(struct task_struct
*, pmu_owner
);
624 DEFINE_PER_CPU(pfm_context_t
*, pmu_ctx
);
625 DEFINE_PER_CPU(unsigned long, pmu_activation_number
);
626 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info
);
629 /* forward declaration */
630 static const struct file_operations pfm_file_ops
;
633 * forward declarations
636 static void pfm_lazy_save_regs (struct task_struct
*ta
);
639 void dump_pmu_state(const char *);
640 static int pfm_write_ibr_dbr(int mode
, pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
642 #include "perfmon_itanium.h"
643 #include "perfmon_mckinley.h"
644 #include "perfmon_montecito.h"
645 #include "perfmon_generic.h"
647 static pmu_config_t
*pmu_confs
[]={
651 &pmu_conf_gen
, /* must be last */
656 static int pfm_end_notify_user(pfm_context_t
*ctx
);
659 pfm_clear_psr_pp(void)
661 ia64_rsm(IA64_PSR_PP
);
668 ia64_ssm(IA64_PSR_PP
);
673 pfm_clear_psr_up(void)
675 ia64_rsm(IA64_PSR_UP
);
682 ia64_ssm(IA64_PSR_UP
);
686 static inline unsigned long
690 tmp
= ia64_getreg(_IA64_REG_PSR
);
696 pfm_set_psr_l(unsigned long val
)
698 ia64_setreg(_IA64_REG_PSR_L
, val
);
710 pfm_unfreeze_pmu(void)
717 pfm_restore_ibrs(unsigned long *ibrs
, unsigned int nibrs
)
721 for (i
=0; i
< nibrs
; i
++) {
722 ia64_set_ibr(i
, ibrs
[i
]);
723 ia64_dv_serialize_instruction();
729 pfm_restore_dbrs(unsigned long *dbrs
, unsigned int ndbrs
)
733 for (i
=0; i
< ndbrs
; i
++) {
734 ia64_set_dbr(i
, dbrs
[i
]);
735 ia64_dv_serialize_data();
741 * PMD[i] must be a counter. no check is made
743 static inline unsigned long
744 pfm_read_soft_counter(pfm_context_t
*ctx
, int i
)
746 return ctx
->ctx_pmds
[i
].val
+ (ia64_get_pmd(i
) & pmu_conf
->ovfl_val
);
750 * PMD[i] must be a counter. no check is made
753 pfm_write_soft_counter(pfm_context_t
*ctx
, int i
, unsigned long val
)
755 unsigned long ovfl_val
= pmu_conf
->ovfl_val
;
757 ctx
->ctx_pmds
[i
].val
= val
& ~ovfl_val
;
759 * writing to unimplemented part is ignore, so we do not need to
762 ia64_set_pmd(i
, val
& ovfl_val
);
766 pfm_get_new_msg(pfm_context_t
*ctx
)
770 next
= (ctx
->ctx_msgq_tail
+1) % PFM_MAX_MSGS
;
772 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
773 if (next
== ctx
->ctx_msgq_head
) return NULL
;
775 idx
= ctx
->ctx_msgq_tail
;
776 ctx
->ctx_msgq_tail
= next
;
778 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
, idx
));
780 return ctx
->ctx_msgq
+idx
;
784 pfm_get_next_msg(pfm_context_t
*ctx
)
788 DPRINT(("ctx=%p head=%d tail=%d\n", ctx
, ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
790 if (PFM_CTXQ_EMPTY(ctx
)) return NULL
;
795 msg
= ctx
->ctx_msgq
+ctx
->ctx_msgq_head
;
800 ctx
->ctx_msgq_head
= (ctx
->ctx_msgq_head
+1) % PFM_MAX_MSGS
;
802 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
));
808 pfm_reset_msgq(pfm_context_t
*ctx
)
810 ctx
->ctx_msgq_head
= ctx
->ctx_msgq_tail
= 0;
811 DPRINT(("ctx=%p msgq reset\n", ctx
));
814 static pfm_context_t
*
815 pfm_context_alloc(int ctx_flags
)
820 * allocate context descriptor
821 * must be able to free with interrupts disabled
823 ctx
= kzalloc(sizeof(pfm_context_t
), GFP_KERNEL
);
825 DPRINT(("alloc ctx @%p\n", ctx
));
828 * init context protection lock
830 spin_lock_init(&ctx
->ctx_lock
);
833 * context is unloaded
835 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
838 * initialization of context's flags
840 ctx
->ctx_fl_block
= (ctx_flags
& PFM_FL_NOTIFY_BLOCK
) ? 1 : 0;
841 ctx
->ctx_fl_system
= (ctx_flags
& PFM_FL_SYSTEM_WIDE
) ? 1: 0;
842 ctx
->ctx_fl_no_msg
= (ctx_flags
& PFM_FL_OVFL_NO_MSG
) ? 1: 0;
844 * will move to set properties
845 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
849 * init restart semaphore to locked
851 init_completion(&ctx
->ctx_restart_done
);
854 * activation is used in SMP only
856 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
857 SET_LAST_CPU(ctx
, -1);
860 * initialize notification message queue
862 ctx
->ctx_msgq_head
= ctx
->ctx_msgq_tail
= 0;
863 init_waitqueue_head(&ctx
->ctx_msgq_wait
);
864 init_waitqueue_head(&ctx
->ctx_zombieq
);
871 pfm_context_free(pfm_context_t
*ctx
)
874 DPRINT(("free ctx @%p\n", ctx
));
880 pfm_mask_monitoring(struct task_struct
*task
)
882 pfm_context_t
*ctx
= PFM_GET_CTX(task
);
883 unsigned long mask
, val
, ovfl_mask
;
886 DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task
)));
888 ovfl_mask
= pmu_conf
->ovfl_val
;
890 * monitoring can only be masked as a result of a valid
891 * counter overflow. In UP, it means that the PMU still
892 * has an owner. Note that the owner can be different
893 * from the current task. However the PMU state belongs
895 * In SMP, a valid overflow only happens when task is
896 * current. Therefore if we come here, we know that
897 * the PMU state belongs to the current task, therefore
898 * we can access the live registers.
900 * So in both cases, the live register contains the owner's
901 * state. We can ONLY touch the PMU registers and NOT the PSR.
903 * As a consequence to this call, the ctx->th_pmds[] array
904 * contains stale information which must be ignored
905 * when context is reloaded AND monitoring is active (see
908 mask
= ctx
->ctx_used_pmds
[0];
909 for (i
= 0; mask
; i
++, mask
>>=1) {
910 /* skip non used pmds */
911 if ((mask
& 0x1) == 0) continue;
912 val
= ia64_get_pmd(i
);
914 if (PMD_IS_COUNTING(i
)) {
916 * we rebuild the full 64 bit value of the counter
918 ctx
->ctx_pmds
[i
].val
+= (val
& ovfl_mask
);
920 ctx
->ctx_pmds
[i
].val
= val
;
922 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
924 ctx
->ctx_pmds
[i
].val
,
928 * mask monitoring by setting the privilege level to 0
929 * we cannot use psr.pp/psr.up for this, it is controlled by
932 * if task is current, modify actual registers, otherwise modify
933 * thread save state, i.e., what will be restored in pfm_load_regs()
935 mask
= ctx
->ctx_used_monitors
[0] >> PMU_FIRST_COUNTER
;
936 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>=1) {
937 if ((mask
& 0x1) == 0UL) continue;
938 ia64_set_pmc(i
, ctx
->th_pmcs
[i
] & ~0xfUL
);
939 ctx
->th_pmcs
[i
] &= ~0xfUL
;
940 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i
, ctx
->th_pmcs
[i
]));
943 * make all of this visible
949 * must always be done with task == current
951 * context must be in MASKED state when calling
954 pfm_restore_monitoring(struct task_struct
*task
)
956 pfm_context_t
*ctx
= PFM_GET_CTX(task
);
957 unsigned long mask
, ovfl_mask
;
958 unsigned long psr
, val
;
961 is_system
= ctx
->ctx_fl_system
;
962 ovfl_mask
= pmu_conf
->ovfl_val
;
964 if (task
!= current
) {
965 printk(KERN_ERR
"perfmon.%d: invalid task[%d] current[%d]\n", __LINE__
, task_pid_nr(task
), task_pid_nr(current
));
968 if (ctx
->ctx_state
!= PFM_CTX_MASKED
) {
969 printk(KERN_ERR
"perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__
,
970 task_pid_nr(task
), task_pid_nr(current
), ctx
->ctx_state
);
975 * monitoring is masked via the PMC.
976 * As we restore their value, we do not want each counter to
977 * restart right away. We stop monitoring using the PSR,
978 * restore the PMC (and PMD) and then re-establish the psr
979 * as it was. Note that there can be no pending overflow at
980 * this point, because monitoring was MASKED.
982 * system-wide session are pinned and self-monitoring
984 if (is_system
&& (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP
)) {
986 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) & ~IA64_DCR_PP
);
992 * first, we restore the PMD
994 mask
= ctx
->ctx_used_pmds
[0];
995 for (i
= 0; mask
; i
++, mask
>>=1) {
996 /* skip non used pmds */
997 if ((mask
& 0x1) == 0) continue;
999 if (PMD_IS_COUNTING(i
)) {
1001 * we split the 64bit value according to
1004 val
= ctx
->ctx_pmds
[i
].val
& ovfl_mask
;
1005 ctx
->ctx_pmds
[i
].val
&= ~ovfl_mask
;
1007 val
= ctx
->ctx_pmds
[i
].val
;
1009 ia64_set_pmd(i
, val
);
1011 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1013 ctx
->ctx_pmds
[i
].val
,
1019 mask
= ctx
->ctx_used_monitors
[0] >> PMU_FIRST_COUNTER
;
1020 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>=1) {
1021 if ((mask
& 0x1) == 0UL) continue;
1022 ctx
->th_pmcs
[i
] = ctx
->ctx_pmcs
[i
];
1023 ia64_set_pmc(i
, ctx
->th_pmcs
[i
]);
1024 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1025 task_pid_nr(task
), i
, ctx
->th_pmcs
[i
]));
1030 * must restore DBR/IBR because could be modified while masked
1031 * XXX: need to optimize
1033 if (ctx
->ctx_fl_using_dbreg
) {
1034 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
1035 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
1041 if (is_system
&& (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP
)) {
1043 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) | IA64_DCR_PP
);
1050 pfm_save_pmds(unsigned long *pmds
, unsigned long mask
)
1056 for (i
=0; mask
; i
++, mask
>>=1) {
1057 if (mask
& 0x1) pmds
[i
] = ia64_get_pmd(i
);
1062 * reload from thread state (used for ctxw only)
1065 pfm_restore_pmds(unsigned long *pmds
, unsigned long mask
)
1068 unsigned long val
, ovfl_val
= pmu_conf
->ovfl_val
;
1070 for (i
=0; mask
; i
++, mask
>>=1) {
1071 if ((mask
& 0x1) == 0) continue;
1072 val
= PMD_IS_COUNTING(i
) ? pmds
[i
] & ovfl_val
: pmds
[i
];
1073 ia64_set_pmd(i
, val
);
1079 * propagate PMD from context to thread-state
1082 pfm_copy_pmds(struct task_struct
*task
, pfm_context_t
*ctx
)
1084 unsigned long ovfl_val
= pmu_conf
->ovfl_val
;
1085 unsigned long mask
= ctx
->ctx_all_pmds
[0];
1089 DPRINT(("mask=0x%lx\n", mask
));
1091 for (i
=0; mask
; i
++, mask
>>=1) {
1093 val
= ctx
->ctx_pmds
[i
].val
;
1096 * We break up the 64 bit value into 2 pieces
1097 * the lower bits go to the machine state in the
1098 * thread (will be reloaded on ctxsw in).
1099 * The upper part stays in the soft-counter.
1101 if (PMD_IS_COUNTING(i
)) {
1102 ctx
->ctx_pmds
[i
].val
= val
& ~ovfl_val
;
1105 ctx
->th_pmds
[i
] = val
;
1107 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1110 ctx
->ctx_pmds
[i
].val
));
1115 * propagate PMC from context to thread-state
1118 pfm_copy_pmcs(struct task_struct
*task
, pfm_context_t
*ctx
)
1120 unsigned long mask
= ctx
->ctx_all_pmcs
[0];
1123 DPRINT(("mask=0x%lx\n", mask
));
1125 for (i
=0; mask
; i
++, mask
>>=1) {
1126 /* masking 0 with ovfl_val yields 0 */
1127 ctx
->th_pmcs
[i
] = ctx
->ctx_pmcs
[i
];
1128 DPRINT(("pmc[%d]=0x%lx\n", i
, ctx
->th_pmcs
[i
]));
1135 pfm_restore_pmcs(unsigned long *pmcs
, unsigned long mask
)
1139 for (i
=0; mask
; i
++, mask
>>=1) {
1140 if ((mask
& 0x1) == 0) continue;
1141 ia64_set_pmc(i
, pmcs
[i
]);
1147 pfm_uuid_cmp(pfm_uuid_t a
, pfm_uuid_t b
)
1149 return memcmp(a
, b
, sizeof(pfm_uuid_t
));
1153 pfm_buf_fmt_exit(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, void *buf
, struct pt_regs
*regs
)
1156 if (fmt
->fmt_exit
) ret
= (*fmt
->fmt_exit
)(task
, buf
, regs
);
1161 pfm_buf_fmt_getsize(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, unsigned int flags
, int cpu
, void *arg
, unsigned long *size
)
1164 if (fmt
->fmt_getsize
) ret
= (*fmt
->fmt_getsize
)(task
, flags
, cpu
, arg
, size
);
1170 pfm_buf_fmt_validate(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, unsigned int flags
,
1174 if (fmt
->fmt_validate
) ret
= (*fmt
->fmt_validate
)(task
, flags
, cpu
, arg
);
1179 pfm_buf_fmt_init(pfm_buffer_fmt_t
*fmt
, struct task_struct
*task
, void *buf
, unsigned int flags
,
1183 if (fmt
->fmt_init
) ret
= (*fmt
->fmt_init
)(task
, buf
, flags
, cpu
, arg
);
1188 pfm_buf_fmt_restart(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
) ret
= (*fmt
->fmt_restart
)(task
, ctrl
, buf
, regs
);
1196 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
)
1199 if (fmt
->fmt_restart_active
) ret
= (*fmt
->fmt_restart_active
)(task
, ctrl
, buf
, regs
);
1203 static pfm_buffer_fmt_t
*
1204 __pfm_find_buffer_fmt(pfm_uuid_t uuid
)
1206 struct list_head
* pos
;
1207 pfm_buffer_fmt_t
* entry
;
1209 list_for_each(pos
, &pfm_buffer_fmt_list
) {
1210 entry
= list_entry(pos
, pfm_buffer_fmt_t
, fmt_list
);
1211 if (pfm_uuid_cmp(uuid
, entry
->fmt_uuid
) == 0)
1218 * find a buffer format based on its uuid
1220 static pfm_buffer_fmt_t
*
1221 pfm_find_buffer_fmt(pfm_uuid_t uuid
)
1223 pfm_buffer_fmt_t
* fmt
;
1224 spin_lock(&pfm_buffer_fmt_lock
);
1225 fmt
= __pfm_find_buffer_fmt(uuid
);
1226 spin_unlock(&pfm_buffer_fmt_lock
);
1231 pfm_register_buffer_fmt(pfm_buffer_fmt_t
*fmt
)
1235 /* some sanity checks */
1236 if (fmt
== NULL
|| fmt
->fmt_name
== NULL
) return -EINVAL
;
1238 /* we need at least a handler */
1239 if (fmt
->fmt_handler
== NULL
) return -EINVAL
;
1242 * XXX: need check validity of fmt_arg_size
1245 spin_lock(&pfm_buffer_fmt_lock
);
1247 if (__pfm_find_buffer_fmt(fmt
->fmt_uuid
)) {
1248 printk(KERN_ERR
"perfmon: duplicate sampling format: %s\n", fmt
->fmt_name
);
1252 list_add(&fmt
->fmt_list
, &pfm_buffer_fmt_list
);
1253 printk(KERN_INFO
"perfmon: added sampling format %s\n", fmt
->fmt_name
);
1256 spin_unlock(&pfm_buffer_fmt_lock
);
1259 EXPORT_SYMBOL(pfm_register_buffer_fmt
);
1262 pfm_unregister_buffer_fmt(pfm_uuid_t uuid
)
1264 pfm_buffer_fmt_t
*fmt
;
1267 spin_lock(&pfm_buffer_fmt_lock
);
1269 fmt
= __pfm_find_buffer_fmt(uuid
);
1271 printk(KERN_ERR
"perfmon: cannot unregister format, not found\n");
1275 list_del_init(&fmt
->fmt_list
);
1276 printk(KERN_INFO
"perfmon: removed sampling format: %s\n", fmt
->fmt_name
);
1279 spin_unlock(&pfm_buffer_fmt_lock
);
1283 EXPORT_SYMBOL(pfm_unregister_buffer_fmt
);
1286 pfm_reserve_session(struct task_struct
*task
, int is_syswide
, unsigned int cpu
)
1288 unsigned long flags
;
1290 * validity checks on cpu_mask have been done upstream
1294 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1295 pfm_sessions
.pfs_sys_sessions
,
1296 pfm_sessions
.pfs_task_sessions
,
1297 pfm_sessions
.pfs_sys_use_dbregs
,
1303 * cannot mix system wide and per-task sessions
1305 if (pfm_sessions
.pfs_task_sessions
> 0UL) {
1306 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1307 pfm_sessions
.pfs_task_sessions
));
1311 if (pfm_sessions
.pfs_sys_session
[cpu
]) goto error_conflict
;
1313 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu
, smp_processor_id()));
1315 pfm_sessions
.pfs_sys_session
[cpu
] = task
;
1317 pfm_sessions
.pfs_sys_sessions
++ ;
1320 if (pfm_sessions
.pfs_sys_sessions
) goto abort
;
1321 pfm_sessions
.pfs_task_sessions
++;
1324 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1325 pfm_sessions
.pfs_sys_sessions
,
1326 pfm_sessions
.pfs_task_sessions
,
1327 pfm_sessions
.pfs_sys_use_dbregs
,
1332 * Force idle() into poll mode
1334 cpu_idle_poll_ctrl(true);
1341 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1342 task_pid_nr(pfm_sessions
.pfs_sys_session
[cpu
]),
1352 pfm_unreserve_session(pfm_context_t
*ctx
, int is_syswide
, unsigned int cpu
)
1354 unsigned long flags
;
1356 * validity checks on cpu_mask have been done upstream
1360 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1361 pfm_sessions
.pfs_sys_sessions
,
1362 pfm_sessions
.pfs_task_sessions
,
1363 pfm_sessions
.pfs_sys_use_dbregs
,
1369 pfm_sessions
.pfs_sys_session
[cpu
] = NULL
;
1371 * would not work with perfmon+more than one bit in cpu_mask
1373 if (ctx
&& ctx
->ctx_fl_using_dbreg
) {
1374 if (pfm_sessions
.pfs_sys_use_dbregs
== 0) {
1375 printk(KERN_ERR
"perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx
);
1377 pfm_sessions
.pfs_sys_use_dbregs
--;
1380 pfm_sessions
.pfs_sys_sessions
--;
1382 pfm_sessions
.pfs_task_sessions
--;
1384 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1385 pfm_sessions
.pfs_sys_sessions
,
1386 pfm_sessions
.pfs_task_sessions
,
1387 pfm_sessions
.pfs_sys_use_dbregs
,
1391 /* Undo forced polling. Last session reenables pal_halt */
1392 cpu_idle_poll_ctrl(false);
1400 * removes virtual mapping of the sampling buffer.
1401 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1402 * a PROTECT_CTX() section.
1405 pfm_remove_smpl_mapping(void *vaddr
, unsigned long size
)
1407 struct task_struct
*task
= current
;
1411 if (task
->mm
== NULL
|| size
== 0UL || vaddr
== NULL
) {
1412 printk(KERN_ERR
"perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task
), task
->mm
);
1416 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr
, size
));
1419 * does the actual unmapping
1421 r
= vm_munmap((unsigned long)vaddr
, size
);
1424 printk(KERN_ERR
"perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task
), vaddr
, size
);
1427 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr
, size
, r
));
1433 * free actual physical storage used by sampling buffer
1437 pfm_free_smpl_buffer(pfm_context_t
*ctx
)
1439 pfm_buffer_fmt_t
*fmt
;
1441 if (ctx
->ctx_smpl_hdr
== NULL
) goto invalid_free
;
1444 * we won't use the buffer format anymore
1446 fmt
= ctx
->ctx_buf_fmt
;
1448 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1451 ctx
->ctx_smpl_vaddr
));
1453 pfm_buf_fmt_exit(fmt
, current
, NULL
, NULL
);
1458 vfree(ctx
->ctx_smpl_hdr
);
1460 ctx
->ctx_smpl_hdr
= NULL
;
1461 ctx
->ctx_smpl_size
= 0UL;
1466 printk(KERN_ERR
"perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current
));
1472 pfm_exit_smpl_buffer(pfm_buffer_fmt_t
*fmt
)
1474 if (fmt
== NULL
) return;
1476 pfm_buf_fmt_exit(fmt
, current
, NULL
, NULL
);
1481 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1482 * no real gain from having the whole whorehouse mounted. So we don't need
1483 * any operations on the root directory. However, we need a non-trivial
1484 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1486 static struct vfsmount
*pfmfs_mnt __read_mostly
;
1491 int err
= register_filesystem(&pfm_fs_type
);
1493 pfmfs_mnt
= kern_mount(&pfm_fs_type
);
1494 err
= PTR_ERR(pfmfs_mnt
);
1495 if (IS_ERR(pfmfs_mnt
))
1496 unregister_filesystem(&pfm_fs_type
);
1504 pfm_read(struct file
*filp
, char __user
*buf
, size_t size
, loff_t
*ppos
)
1509 unsigned long flags
;
1510 DECLARE_WAITQUEUE(wait
, current
);
1511 if (PFM_IS_FILE(filp
) == 0) {
1512 printk(KERN_ERR
"perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current
));
1516 ctx
= filp
->private_data
;
1518 printk(KERN_ERR
"perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current
));
1523 * check even when there is no message
1525 if (size
< sizeof(pfm_msg_t
)) {
1526 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx
, sizeof(pfm_msg_t
)));
1530 PROTECT_CTX(ctx
, flags
);
1533 * put ourselves on the wait queue
1535 add_wait_queue(&ctx
->ctx_msgq_wait
, &wait
);
1543 set_current_state(TASK_INTERRUPTIBLE
);
1545 DPRINT(("head=%d tail=%d\n", ctx
->ctx_msgq_head
, ctx
->ctx_msgq_tail
));
1548 if(PFM_CTXQ_EMPTY(ctx
) == 0) break;
1550 UNPROTECT_CTX(ctx
, flags
);
1553 * check non-blocking read
1556 if(filp
->f_flags
& O_NONBLOCK
) break;
1559 * check pending signals
1561 if(signal_pending(current
)) {
1566 * no message, so wait
1570 PROTECT_CTX(ctx
, flags
);
1572 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current
), ret
));
1573 set_current_state(TASK_RUNNING
);
1574 remove_wait_queue(&ctx
->ctx_msgq_wait
, &wait
);
1576 if (ret
< 0) goto abort
;
1579 msg
= pfm_get_next_msg(ctx
);
1581 printk(KERN_ERR
"perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx
, task_pid_nr(current
));
1585 DPRINT(("fd=%d type=%d\n", msg
->pfm_gen_msg
.msg_ctx_fd
, msg
->pfm_gen_msg
.msg_type
));
1588 if(copy_to_user(buf
, msg
, sizeof(pfm_msg_t
)) == 0) ret
= sizeof(pfm_msg_t
);
1591 UNPROTECT_CTX(ctx
, flags
);
1597 pfm_write(struct file
*file
, const char __user
*ubuf
,
1598 size_t size
, loff_t
*ppos
)
1600 DPRINT(("pfm_write called\n"));
1605 pfm_poll(struct file
*filp
, poll_table
* wait
)
1608 unsigned long flags
;
1611 if (PFM_IS_FILE(filp
) == 0) {
1612 printk(KERN_ERR
"perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current
));
1616 ctx
= filp
->private_data
;
1618 printk(KERN_ERR
"perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current
));
1623 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx
->ctx_fd
));
1625 poll_wait(filp
, &ctx
->ctx_msgq_wait
, wait
);
1627 PROTECT_CTX(ctx
, flags
);
1629 if (PFM_CTXQ_EMPTY(ctx
) == 0)
1630 mask
= EPOLLIN
| EPOLLRDNORM
;
1632 UNPROTECT_CTX(ctx
, flags
);
1634 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx
->ctx_fd
, mask
));
1640 pfm_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
1642 DPRINT(("pfm_ioctl called\n"));
1647 * interrupt cannot be masked when coming here
1650 pfm_do_fasync(int fd
, struct file
*filp
, pfm_context_t
*ctx
, int on
)
1654 ret
= fasync_helper (fd
, filp
, on
, &ctx
->ctx_async_queue
);
1656 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1657 task_pid_nr(current
),
1660 ctx
->ctx_async_queue
, ret
));
1666 pfm_fasync(int fd
, struct file
*filp
, int on
)
1671 if (PFM_IS_FILE(filp
) == 0) {
1672 printk(KERN_ERR
"perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current
));
1676 ctx
= filp
->private_data
;
1678 printk(KERN_ERR
"perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current
));
1682 * we cannot mask interrupts during this call because this may
1683 * may go to sleep if memory is not readily avalaible.
1685 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1686 * done in caller. Serialization of this function is ensured by caller.
1688 ret
= pfm_do_fasync(fd
, filp
, ctx
, on
);
1691 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1694 ctx
->ctx_async_queue
, ret
));
1701 * this function is exclusively called from pfm_close().
1702 * The context is not protected at that time, nor are interrupts
1703 * on the remote CPU. That's necessary to avoid deadlocks.
1706 pfm_syswide_force_stop(void *info
)
1708 pfm_context_t
*ctx
= (pfm_context_t
*)info
;
1709 struct pt_regs
*regs
= task_pt_regs(current
);
1710 struct task_struct
*owner
;
1711 unsigned long flags
;
1714 if (ctx
->ctx_cpu
!= smp_processor_id()) {
1715 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1717 smp_processor_id());
1720 owner
= GET_PMU_OWNER();
1721 if (owner
!= ctx
->ctx_task
) {
1722 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1724 task_pid_nr(owner
), task_pid_nr(ctx
->ctx_task
));
1727 if (GET_PMU_CTX() != ctx
) {
1728 printk(KERN_ERR
"perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1730 GET_PMU_CTX(), ctx
);
1734 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx
->ctx_task
)));
1736 * the context is already protected in pfm_close(), we simply
1737 * need to mask interrupts to avoid a PMU interrupt race on
1740 local_irq_save(flags
);
1742 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
1744 DPRINT(("context_unload returned %d\n", ret
));
1748 * unmask interrupts, PMU interrupts are now spurious here
1750 local_irq_restore(flags
);
1754 pfm_syswide_cleanup_other_cpu(pfm_context_t
*ctx
)
1758 DPRINT(("calling CPU%d for cleanup\n", ctx
->ctx_cpu
));
1759 ret
= smp_call_function_single(ctx
->ctx_cpu
, pfm_syswide_force_stop
, ctx
, 1);
1760 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx
->ctx_cpu
, ret
));
1762 #endif /* CONFIG_SMP */
1765 * called for each close(). Partially free resources.
1766 * When caller is self-monitoring, the context is unloaded.
1769 pfm_flush(struct file
*filp
, fl_owner_t id
)
1772 struct task_struct
*task
;
1773 struct pt_regs
*regs
;
1774 unsigned long flags
;
1775 unsigned long smpl_buf_size
= 0UL;
1776 void *smpl_buf_vaddr
= NULL
;
1777 int state
, is_system
;
1779 if (PFM_IS_FILE(filp
) == 0) {
1780 DPRINT(("bad magic for\n"));
1784 ctx
= filp
->private_data
;
1786 printk(KERN_ERR
"perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current
));
1791 * remove our file from the async queue, if we use this mode.
1792 * This can be done without the context being protected. We come
1793 * here when the context has become unreachable by other tasks.
1795 * We may still have active monitoring at this point and we may
1796 * end up in pfm_overflow_handler(). However, fasync_helper()
1797 * operates with interrupts disabled and it cleans up the
1798 * queue. If the PMU handler is called prior to entering
1799 * fasync_helper() then it will send a signal. If it is
1800 * invoked after, it will find an empty queue and no
1801 * signal will be sent. In both case, we are safe
1803 PROTECT_CTX(ctx
, flags
);
1805 state
= ctx
->ctx_state
;
1806 is_system
= ctx
->ctx_fl_system
;
1808 task
= PFM_CTX_TASK(ctx
);
1809 regs
= task_pt_regs(task
);
1811 DPRINT(("ctx_state=%d is_current=%d\n",
1813 task
== current
? 1 : 0));
1816 * if state == UNLOADED, then task is NULL
1820 * we must stop and unload because we are losing access to the context.
1822 if (task
== current
) {
1825 * the task IS the owner but it migrated to another CPU: that's bad
1826 * but we must handle this cleanly. Unfortunately, the kernel does
1827 * not provide a mechanism to block migration (while the context is loaded).
1829 * We need to release the resource on the ORIGINAL cpu.
1831 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
1833 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
1835 * keep context protected but unmask interrupt for IPI
1837 local_irq_restore(flags
);
1839 pfm_syswide_cleanup_other_cpu(ctx
);
1842 * restore interrupt masking
1844 local_irq_save(flags
);
1847 * context is unloaded at this point
1850 #endif /* CONFIG_SMP */
1853 DPRINT(("forcing unload\n"));
1855 * stop and unload, returning with state UNLOADED
1856 * and session unreserved.
1858 pfm_context_unload(ctx
, NULL
, 0, regs
);
1860 DPRINT(("ctx_state=%d\n", ctx
->ctx_state
));
1865 * remove virtual mapping, if any, for the calling task.
1866 * cannot reset ctx field until last user is calling close().
1868 * ctx_smpl_vaddr must never be cleared because it is needed
1869 * by every task with access to the context
1871 * When called from do_exit(), the mm context is gone already, therefore
1872 * mm is NULL, i.e., the VMA is already gone and we do not have to
1875 if (ctx
->ctx_smpl_vaddr
&& current
->mm
) {
1876 smpl_buf_vaddr
= ctx
->ctx_smpl_vaddr
;
1877 smpl_buf_size
= ctx
->ctx_smpl_size
;
1880 UNPROTECT_CTX(ctx
, flags
);
1883 * if there was a mapping, then we systematically remove it
1884 * at this point. Cannot be done inside critical section
1885 * because some VM function reenables interrupts.
1888 if (smpl_buf_vaddr
) pfm_remove_smpl_mapping(smpl_buf_vaddr
, smpl_buf_size
);
1893 * called either on explicit close() or from exit_files().
1894 * Only the LAST user of the file gets to this point, i.e., it is
1897 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1898 * (fput()),i.e, last task to access the file. Nobody else can access the
1899 * file at this point.
1901 * When called from exit_files(), the VMA has been freed because exit_mm()
1902 * is executed before exit_files().
1904 * When called from exit_files(), the current task is not yet ZOMBIE but we
1905 * flush the PMU state to the context.
1908 pfm_close(struct inode
*inode
, struct file
*filp
)
1911 struct task_struct
*task
;
1912 struct pt_regs
*regs
;
1913 DECLARE_WAITQUEUE(wait
, current
);
1914 unsigned long flags
;
1915 unsigned long smpl_buf_size
= 0UL;
1916 void *smpl_buf_addr
= NULL
;
1917 int free_possible
= 1;
1918 int state
, is_system
;
1920 DPRINT(("pfm_close called private=%p\n", filp
->private_data
));
1922 if (PFM_IS_FILE(filp
) == 0) {
1923 DPRINT(("bad magic\n"));
1927 ctx
= filp
->private_data
;
1929 printk(KERN_ERR
"perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current
));
1933 PROTECT_CTX(ctx
, flags
);
1935 state
= ctx
->ctx_state
;
1936 is_system
= ctx
->ctx_fl_system
;
1938 task
= PFM_CTX_TASK(ctx
);
1939 regs
= task_pt_regs(task
);
1941 DPRINT(("ctx_state=%d is_current=%d\n",
1943 task
== current
? 1 : 0));
1946 * if task == current, then pfm_flush() unloaded the context
1948 if (state
== PFM_CTX_UNLOADED
) goto doit
;
1951 * context is loaded/masked and task != current, we need to
1952 * either force an unload or go zombie
1956 * The task is currently blocked or will block after an overflow.
1957 * we must force it to wakeup to get out of the
1958 * MASKED state and transition to the unloaded state by itself.
1960 * This situation is only possible for per-task mode
1962 if (state
== PFM_CTX_MASKED
&& CTX_OVFL_NOBLOCK(ctx
) == 0) {
1965 * set a "partial" zombie state to be checked
1966 * upon return from down() in pfm_handle_work().
1968 * We cannot use the ZOMBIE state, because it is checked
1969 * by pfm_load_regs() which is called upon wakeup from down().
1970 * In such case, it would free the context and then we would
1971 * return to pfm_handle_work() which would access the
1972 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1973 * but visible to pfm_handle_work().
1975 * For some window of time, we have a zombie context with
1976 * ctx_state = MASKED and not ZOMBIE
1978 ctx
->ctx_fl_going_zombie
= 1;
1981 * force task to wake up from MASKED state
1983 complete(&ctx
->ctx_restart_done
);
1985 DPRINT(("waking up ctx_state=%d\n", state
));
1988 * put ourself to sleep waiting for the other
1989 * task to report completion
1991 * the context is protected by mutex, therefore there
1992 * is no risk of being notified of completion before
1993 * begin actually on the waitq.
1995 set_current_state(TASK_INTERRUPTIBLE
);
1996 add_wait_queue(&ctx
->ctx_zombieq
, &wait
);
1998 UNPROTECT_CTX(ctx
, flags
);
2001 * XXX: check for signals :
2002 * - ok for explicit close
2003 * - not ok when coming from exit_files()
2008 PROTECT_CTX(ctx
, flags
);
2011 remove_wait_queue(&ctx
->ctx_zombieq
, &wait
);
2012 set_current_state(TASK_RUNNING
);
2015 * context is unloaded at this point
2017 DPRINT(("after zombie wakeup ctx_state=%d for\n", state
));
2019 else if (task
!= current
) {
2022 * switch context to zombie state
2024 ctx
->ctx_state
= PFM_CTX_ZOMBIE
;
2026 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task
)));
2028 * cannot free the context on the spot. deferred until
2029 * the task notices the ZOMBIE state
2033 pfm_context_unload(ctx
, NULL
, 0, regs
);
2038 /* reload state, may have changed during opening of critical section */
2039 state
= ctx
->ctx_state
;
2042 * the context is still attached to a task (possibly current)
2043 * we cannot destroy it right now
2047 * we must free the sampling buffer right here because
2048 * we cannot rely on it being cleaned up later by the
2049 * monitored task. It is not possible to free vmalloc'ed
2050 * memory in pfm_load_regs(). Instead, we remove the buffer
2051 * now. should there be subsequent PMU overflow originally
2052 * meant for sampling, the will be converted to spurious
2053 * and that's fine because the monitoring tools is gone anyway.
2055 if (ctx
->ctx_smpl_hdr
) {
2056 smpl_buf_addr
= ctx
->ctx_smpl_hdr
;
2057 smpl_buf_size
= ctx
->ctx_smpl_size
;
2058 /* no more sampling */
2059 ctx
->ctx_smpl_hdr
= NULL
;
2060 ctx
->ctx_fl_is_sampling
= 0;
2063 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2069 if (smpl_buf_addr
) pfm_exit_smpl_buffer(ctx
->ctx_buf_fmt
);
2072 * UNLOADED that the session has already been unreserved.
2074 if (state
== PFM_CTX_ZOMBIE
) {
2075 pfm_unreserve_session(ctx
, ctx
->ctx_fl_system
, ctx
->ctx_cpu
);
2079 * disconnect file descriptor from context must be done
2082 filp
->private_data
= NULL
;
2085 * if we free on the spot, the context is now completely unreachable
2086 * from the callers side. The monitored task side is also cut, so we
2089 * If we have a deferred free, only the caller side is disconnected.
2091 UNPROTECT_CTX(ctx
, flags
);
2094 * All memory free operations (especially for vmalloc'ed memory)
2095 * MUST be done with interrupts ENABLED.
2097 vfree(smpl_buf_addr
);
2100 * return the memory used by the context
2102 if (free_possible
) pfm_context_free(ctx
);
2107 static const struct file_operations pfm_file_ops
= {
2108 .llseek
= no_llseek
,
2112 .unlocked_ioctl
= pfm_ioctl
,
2113 .fasync
= pfm_fasync
,
2114 .release
= pfm_close
,
2118 static char *pfmfs_dname(struct dentry
*dentry
, char *buffer
, int buflen
)
2120 return dynamic_dname(dentry
, buffer
, buflen
, "pfm:[%lu]",
2121 d_inode(dentry
)->i_ino
);
2124 static const struct dentry_operations pfmfs_dentry_operations
= {
2125 .d_delete
= always_delete_dentry
,
2126 .d_dname
= pfmfs_dname
,
2130 static struct file
*
2131 pfm_alloc_file(pfm_context_t
*ctx
)
2134 struct inode
*inode
;
2136 struct qstr
this = { .name
= "" };
2139 * allocate a new inode
2141 inode
= new_inode(pfmfs_mnt
->mnt_sb
);
2143 return ERR_PTR(-ENOMEM
);
2145 DPRINT(("new inode ino=%ld @%p\n", inode
->i_ino
, inode
));
2147 inode
->i_mode
= S_IFCHR
|S_IRUGO
;
2148 inode
->i_uid
= current_fsuid();
2149 inode
->i_gid
= current_fsgid();
2152 * allocate a new dcache entry
2154 path
.dentry
= d_alloc(pfmfs_mnt
->mnt_root
, &this);
2157 return ERR_PTR(-ENOMEM
);
2159 path
.mnt
= mntget(pfmfs_mnt
);
2161 d_add(path
.dentry
, inode
);
2163 file
= alloc_file(&path
, FMODE_READ
, &pfm_file_ops
);
2169 file
->f_flags
= O_RDONLY
;
2170 file
->private_data
= ctx
;
2176 pfm_remap_buffer(struct vm_area_struct
*vma
, unsigned long buf
, unsigned long addr
, unsigned long size
)
2178 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf
, addr
, size
));
2181 unsigned long pfn
= ia64_tpa(buf
) >> PAGE_SHIFT
;
2184 if (remap_pfn_range(vma
, addr
, pfn
, PAGE_SIZE
, PAGE_READONLY
))
2195 * allocate a sampling buffer and remaps it into the user address space of the task
2198 pfm_smpl_buffer_alloc(struct task_struct
*task
, struct file
*filp
, pfm_context_t
*ctx
, unsigned long rsize
, void **user_vaddr
)
2200 struct mm_struct
*mm
= task
->mm
;
2201 struct vm_area_struct
*vma
= NULL
;
2207 * the fixed header + requested size and align to page boundary
2209 size
= PAGE_ALIGN(rsize
);
2211 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize
, size
));
2214 * check requested size to avoid Denial-of-service attacks
2215 * XXX: may have to refine this test
2216 * Check against address space limit.
2218 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2221 if (size
> task_rlimit(task
, RLIMIT_MEMLOCK
))
2225 * We do the easy to undo allocations first.
2227 smpl_buf
= vzalloc(size
);
2228 if (smpl_buf
== NULL
) {
2229 DPRINT(("Can't allocate sampling buffer\n"));
2233 DPRINT(("smpl_buf @%p\n", smpl_buf
));
2236 vma
= vm_area_alloc(mm
);
2238 DPRINT(("Cannot allocate vma\n"));
2243 * partially initialize the vma for the sampling buffer
2245 vma
->vm_file
= get_file(filp
);
2246 vma
->vm_flags
= VM_READ
|VM_MAYREAD
|VM_DONTEXPAND
|VM_DONTDUMP
;
2247 vma
->vm_page_prot
= PAGE_READONLY
; /* XXX may need to change */
2250 * Now we have everything we need and we can initialize
2251 * and connect all the data structures
2254 ctx
->ctx_smpl_hdr
= smpl_buf
;
2255 ctx
->ctx_smpl_size
= size
; /* aligned size */
2258 * Let's do the difficult operations next.
2260 * now we atomically find some area in the address space and
2261 * remap the buffer in it.
2263 mmap_write_lock(task
->mm
);
2265 /* find some free area in address space, must have mmap sem held */
2266 vma
->vm_start
= get_unmapped_area(NULL
, 0, size
, 0, MAP_PRIVATE
|MAP_ANONYMOUS
);
2267 if (IS_ERR_VALUE(vma
->vm_start
)) {
2268 DPRINT(("Cannot find unmapped area for size %ld\n", size
));
2269 mmap_write_unlock(task
->mm
);
2272 vma
->vm_end
= vma
->vm_start
+ size
;
2273 vma
->vm_pgoff
= vma
->vm_start
>> PAGE_SHIFT
;
2275 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size
, ctx
->ctx_smpl_hdr
, vma
->vm_start
));
2277 /* can only be applied to current task, need to have the mm semaphore held when called */
2278 if (pfm_remap_buffer(vma
, (unsigned long)smpl_buf
, vma
->vm_start
, size
)) {
2279 DPRINT(("Can't remap buffer\n"));
2280 mmap_write_unlock(task
->mm
);
2285 * now insert the vma in the vm list for the process, must be
2286 * done with mmap lock held
2288 insert_vm_struct(mm
, vma
);
2290 vm_stat_account(vma
->vm_mm
, vma
->vm_flags
, vma_pages(vma
));
2291 mmap_write_unlock(task
->mm
);
2294 * keep track of user level virtual address
2296 ctx
->ctx_smpl_vaddr
= (void *)vma
->vm_start
;
2297 *(unsigned long *)user_vaddr
= vma
->vm_start
;
2310 * XXX: do something better here
2313 pfm_bad_permissions(struct task_struct
*task
)
2315 const struct cred
*tcred
;
2316 kuid_t uid
= current_uid();
2317 kgid_t gid
= current_gid();
2321 tcred
= __task_cred(task
);
2323 /* inspired by ptrace_attach() */
2324 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2325 from_kuid(&init_user_ns
, uid
),
2326 from_kgid(&init_user_ns
, gid
),
2327 from_kuid(&init_user_ns
, tcred
->euid
),
2328 from_kuid(&init_user_ns
, tcred
->suid
),
2329 from_kuid(&init_user_ns
, tcred
->uid
),
2330 from_kgid(&init_user_ns
, tcred
->egid
),
2331 from_kgid(&init_user_ns
, tcred
->sgid
)));
2333 ret
= ((!uid_eq(uid
, tcred
->euid
))
2334 || (!uid_eq(uid
, tcred
->suid
))
2335 || (!uid_eq(uid
, tcred
->uid
))
2336 || (!gid_eq(gid
, tcred
->egid
))
2337 || (!gid_eq(gid
, tcred
->sgid
))
2338 || (!gid_eq(gid
, tcred
->gid
))) && !capable(CAP_SYS_PTRACE
);
2345 pfarg_is_sane(struct task_struct
*task
, pfarg_context_t
*pfx
)
2351 ctx_flags
= pfx
->ctx_flags
;
2353 if (ctx_flags
& PFM_FL_SYSTEM_WIDE
) {
2356 * cannot block in this mode
2358 if (ctx_flags
& PFM_FL_NOTIFY_BLOCK
) {
2359 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2364 /* probably more to add here */
2370 pfm_setup_buffer_fmt(struct task_struct
*task
, struct file
*filp
, pfm_context_t
*ctx
, unsigned int ctx_flags
,
2371 unsigned int cpu
, pfarg_context_t
*arg
)
2373 pfm_buffer_fmt_t
*fmt
= NULL
;
2374 unsigned long size
= 0UL;
2376 void *fmt_arg
= NULL
;
2378 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2380 /* invoke and lock buffer format, if found */
2381 fmt
= pfm_find_buffer_fmt(arg
->ctx_smpl_buf_id
);
2383 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task
)));
2388 * buffer argument MUST be contiguous to pfarg_context_t
2390 if (fmt
->fmt_arg_size
) fmt_arg
= PFM_CTXARG_BUF_ARG(arg
);
2392 ret
= pfm_buf_fmt_validate(fmt
, task
, ctx_flags
, cpu
, fmt_arg
);
2394 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task
), ctx_flags
, cpu
, fmt_arg
, ret
));
2396 if (ret
) goto error
;
2398 /* link buffer format and context */
2399 ctx
->ctx_buf_fmt
= fmt
;
2400 ctx
->ctx_fl_is_sampling
= 1; /* assume record() is defined */
2403 * check if buffer format wants to use perfmon buffer allocation/mapping service
2405 ret
= pfm_buf_fmt_getsize(fmt
, task
, ctx_flags
, cpu
, fmt_arg
, &size
);
2406 if (ret
) goto error
;
2410 * buffer is always remapped into the caller's address space
2412 ret
= pfm_smpl_buffer_alloc(current
, filp
, ctx
, size
, &uaddr
);
2413 if (ret
) goto error
;
2415 /* keep track of user address of buffer */
2416 arg
->ctx_smpl_vaddr
= uaddr
;
2418 ret
= pfm_buf_fmt_init(fmt
, task
, ctx
->ctx_smpl_hdr
, ctx_flags
, cpu
, fmt_arg
);
2425 pfm_reset_pmu_state(pfm_context_t
*ctx
)
2430 * install reset values for PMC.
2432 for (i
=1; PMC_IS_LAST(i
) == 0; i
++) {
2433 if (PMC_IS_IMPL(i
) == 0) continue;
2434 ctx
->ctx_pmcs
[i
] = PMC_DFL_VAL(i
);
2435 DPRINT(("pmc[%d]=0x%lx\n", i
, ctx
->ctx_pmcs
[i
]));
2438 * PMD registers are set to 0UL when the context in memset()
2442 * On context switched restore, we must restore ALL pmc and ALL pmd even
2443 * when they are not actively used by the task. In UP, the incoming process
2444 * may otherwise pick up left over PMC, PMD state from the previous process.
2445 * As opposed to PMD, stale PMC can cause harm to the incoming
2446 * process because they may change what is being measured.
2447 * Therefore, we must systematically reinstall the entire
2448 * PMC state. In SMP, the same thing is possible on the
2449 * same CPU but also on between 2 CPUs.
2451 * The problem with PMD is information leaking especially
2452 * to user level when psr.sp=0
2454 * There is unfortunately no easy way to avoid this problem
2455 * on either UP or SMP. This definitively slows down the
2456 * pfm_load_regs() function.
2460 * bitmask of all PMCs accessible to this context
2462 * PMC0 is treated differently.
2464 ctx
->ctx_all_pmcs
[0] = pmu_conf
->impl_pmcs
[0] & ~0x1;
2467 * bitmask of all PMDs that are accessible to this context
2469 ctx
->ctx_all_pmds
[0] = pmu_conf
->impl_pmds
[0];
2471 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx
->ctx_fd
, ctx
->ctx_all_pmcs
[0],ctx
->ctx_all_pmds
[0]));
2474 * useful in case of re-enable after disable
2476 ctx
->ctx_used_ibrs
[0] = 0UL;
2477 ctx
->ctx_used_dbrs
[0] = 0UL;
2481 pfm_ctx_getsize(void *arg
, size_t *sz
)
2483 pfarg_context_t
*req
= (pfarg_context_t
*)arg
;
2484 pfm_buffer_fmt_t
*fmt
;
2488 if (!pfm_uuid_cmp(req
->ctx_smpl_buf_id
, pfm_null_uuid
)) return 0;
2490 fmt
= pfm_find_buffer_fmt(req
->ctx_smpl_buf_id
);
2492 DPRINT(("cannot find buffer format\n"));
2495 /* get just enough to copy in user parameters */
2496 *sz
= fmt
->fmt_arg_size
;
2497 DPRINT(("arg_size=%lu\n", *sz
));
2505 * cannot attach if :
2507 * - task not owned by caller
2508 * - task incompatible with context mode
2511 pfm_task_incompatible(pfm_context_t
*ctx
, struct task_struct
*task
)
2514 * no kernel task or task not owner by caller
2516 if (task
->mm
== NULL
) {
2517 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task
)));
2520 if (pfm_bad_permissions(task
)) {
2521 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task
)));
2525 * cannot block in self-monitoring mode
2527 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && task
== current
) {
2528 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task
)));
2532 if (task
->exit_state
== EXIT_ZOMBIE
) {
2533 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task
)));
2538 * always ok for self
2540 if (task
== current
) return 0;
2542 if (!task_is_stopped_or_traced(task
)) {
2543 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task
), task
->state
));
2547 * make sure the task is off any CPU
2549 wait_task_inactive(task
, 0);
2551 /* more to come... */
2557 pfm_get_task(pfm_context_t
*ctx
, pid_t pid
, struct task_struct
**task
)
2559 struct task_struct
*p
= current
;
2562 /* XXX: need to add more checks here */
2563 if (pid
< 2) return -EPERM
;
2565 if (pid
!= task_pid_vnr(current
)) {
2566 /* make sure task cannot go away while we operate on it */
2567 p
= find_get_task_by_vpid(pid
);
2572 ret
= pfm_task_incompatible(ctx
, p
);
2575 } else if (p
!= current
) {
2584 pfm_context_create(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
2586 pfarg_context_t
*req
= (pfarg_context_t
*)arg
;
2593 /* let's check the arguments first */
2594 ret
= pfarg_is_sane(current
, req
);
2598 ctx_flags
= req
->ctx_flags
;
2602 fd
= get_unused_fd_flags(0);
2606 ctx
= pfm_context_alloc(ctx_flags
);
2610 filp
= pfm_alloc_file(ctx
);
2612 ret
= PTR_ERR(filp
);
2616 req
->ctx_fd
= ctx
->ctx_fd
= fd
;
2619 * does the user want to sample?
2621 if (pfm_uuid_cmp(req
->ctx_smpl_buf_id
, pfm_null_uuid
)) {
2622 ret
= pfm_setup_buffer_fmt(current
, filp
, ctx
, ctx_flags
, 0, req
);
2627 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2632 ctx
->ctx_fl_excl_idle
,
2637 * initialize soft PMU state
2639 pfm_reset_pmu_state(ctx
);
2641 fd_install(fd
, filp
);
2646 path
= filp
->f_path
;
2650 if (ctx
->ctx_buf_fmt
) {
2651 pfm_buf_fmt_exit(ctx
->ctx_buf_fmt
, current
, NULL
, regs
);
2654 pfm_context_free(ctx
);
2661 static inline unsigned long
2662 pfm_new_counter_value (pfm_counter_t
*reg
, int is_long_reset
)
2664 unsigned long val
= is_long_reset
? reg
->long_reset
: reg
->short_reset
;
2665 unsigned long new_seed
, old_seed
= reg
->seed
, mask
= reg
->mask
;
2666 extern unsigned long carta_random32 (unsigned long seed
);
2668 if (reg
->flags
& PFM_REGFL_RANDOM
) {
2669 new_seed
= carta_random32(old_seed
);
2670 val
-= (old_seed
& mask
); /* counter values are negative numbers! */
2671 if ((mask
>> 32) != 0)
2672 /* construct a full 64-bit random value: */
2673 new_seed
|= carta_random32(old_seed
>> 32) << 32;
2674 reg
->seed
= new_seed
;
2681 pfm_reset_regs_masked(pfm_context_t
*ctx
, unsigned long *ovfl_regs
, int is_long_reset
)
2683 unsigned long mask
= ovfl_regs
[0];
2684 unsigned long reset_others
= 0UL;
2689 * now restore reset value on sampling overflowed counters
2691 mask
>>= PMU_FIRST_COUNTER
;
2692 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
2694 if ((mask
& 0x1UL
) == 0UL) continue;
2696 ctx
->ctx_pmds
[i
].val
= val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2697 reset_others
|= ctx
->ctx_pmds
[i
].reset_pmds
[0];
2699 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset
? "long" : "short", i
, val
));
2703 * Now take care of resetting the other registers
2705 for(i
= 0; reset_others
; i
++, reset_others
>>= 1) {
2707 if ((reset_others
& 0x1) == 0) continue;
2709 ctx
->ctx_pmds
[i
].val
= val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2711 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2712 is_long_reset
? "long" : "short", i
, val
));
2717 pfm_reset_regs(pfm_context_t
*ctx
, unsigned long *ovfl_regs
, int is_long_reset
)
2719 unsigned long mask
= ovfl_regs
[0];
2720 unsigned long reset_others
= 0UL;
2724 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs
[0], is_long_reset
));
2726 if (ctx
->ctx_state
== PFM_CTX_MASKED
) {
2727 pfm_reset_regs_masked(ctx
, ovfl_regs
, is_long_reset
);
2732 * now restore reset value on sampling overflowed counters
2734 mask
>>= PMU_FIRST_COUNTER
;
2735 for(i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
2737 if ((mask
& 0x1UL
) == 0UL) continue;
2739 val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2740 reset_others
|= ctx
->ctx_pmds
[i
].reset_pmds
[0];
2742 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset
? "long" : "short", i
, val
));
2744 pfm_write_soft_counter(ctx
, i
, val
);
2748 * Now take care of resetting the other registers
2750 for(i
= 0; reset_others
; i
++, reset_others
>>= 1) {
2752 if ((reset_others
& 0x1) == 0) continue;
2754 val
= pfm_new_counter_value(ctx
->ctx_pmds
+ i
, is_long_reset
);
2756 if (PMD_IS_COUNTING(i
)) {
2757 pfm_write_soft_counter(ctx
, i
, val
);
2759 ia64_set_pmd(i
, val
);
2761 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2762 is_long_reset
? "long" : "short", i
, val
));
2768 pfm_write_pmcs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
2770 struct task_struct
*task
;
2771 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
2772 unsigned long value
, pmc_pm
;
2773 unsigned long smpl_pmds
, reset_pmds
, impl_pmds
;
2774 unsigned int cnum
, reg_flags
, flags
, pmc_type
;
2775 int i
, can_access_pmu
= 0, is_loaded
, is_system
, expert_mode
;
2776 int is_monitor
, is_counting
, state
;
2778 pfm_reg_check_t wr_func
;
2779 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2781 state
= ctx
->ctx_state
;
2782 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
2783 is_system
= ctx
->ctx_fl_system
;
2784 task
= ctx
->ctx_task
;
2785 impl_pmds
= pmu_conf
->impl_pmds
[0];
2787 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
2791 * In system wide and when the context is loaded, access can only happen
2792 * when the caller is running on the CPU being monitored by the session.
2793 * It does not have to be the owner (ctx_task) of the context per se.
2795 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
2796 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
2799 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
2801 expert_mode
= pfm_sysctl
.expert_mode
;
2803 for (i
= 0; i
< count
; i
++, req
++) {
2805 cnum
= req
->reg_num
;
2806 reg_flags
= req
->reg_flags
;
2807 value
= req
->reg_value
;
2808 smpl_pmds
= req
->reg_smpl_pmds
[0];
2809 reset_pmds
= req
->reg_reset_pmds
[0];
2813 if (cnum
>= PMU_MAX_PMCS
) {
2814 DPRINT(("pmc%u is invalid\n", cnum
));
2818 pmc_type
= pmu_conf
->pmc_desc
[cnum
].type
;
2819 pmc_pm
= (value
>> pmu_conf
->pmc_desc
[cnum
].pm_pos
) & 0x1;
2820 is_counting
= (pmc_type
& PFM_REG_COUNTING
) == PFM_REG_COUNTING
? 1 : 0;
2821 is_monitor
= (pmc_type
& PFM_REG_MONITOR
) == PFM_REG_MONITOR
? 1 : 0;
2824 * we reject all non implemented PMC as well
2825 * as attempts to modify PMC[0-3] which are used
2826 * as status registers by the PMU
2828 if ((pmc_type
& PFM_REG_IMPL
) == 0 || (pmc_type
& PFM_REG_CONTROL
) == PFM_REG_CONTROL
) {
2829 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum
, pmc_type
));
2832 wr_func
= pmu_conf
->pmc_desc
[cnum
].write_check
;
2834 * If the PMC is a monitor, then if the value is not the default:
2835 * - system-wide session: PMCx.pm=1 (privileged monitor)
2836 * - per-task : PMCx.pm=0 (user monitor)
2838 if (is_monitor
&& value
!= PMC_DFL_VAL(cnum
) && is_system
^ pmc_pm
) {
2839 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2848 * enforce generation of overflow interrupt. Necessary on all
2851 value
|= 1 << PMU_PMC_OI
;
2853 if (reg_flags
& PFM_REGFL_OVFL_NOTIFY
) {
2854 flags
|= PFM_REGFL_OVFL_NOTIFY
;
2857 if (reg_flags
& PFM_REGFL_RANDOM
) flags
|= PFM_REGFL_RANDOM
;
2859 /* verify validity of smpl_pmds */
2860 if ((smpl_pmds
& impl_pmds
) != smpl_pmds
) {
2861 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds
, cnum
));
2865 /* verify validity of reset_pmds */
2866 if ((reset_pmds
& impl_pmds
) != reset_pmds
) {
2867 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds
, cnum
));
2871 if (reg_flags
& (PFM_REGFL_OVFL_NOTIFY
|PFM_REGFL_RANDOM
)) {
2872 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum
));
2875 /* eventid on non-counting monitors are ignored */
2879 * execute write checker, if any
2881 if (likely(expert_mode
== 0 && wr_func
)) {
2882 ret
= (*wr_func
)(task
, ctx
, cnum
, &value
, regs
);
2883 if (ret
) goto error
;
2888 * no error on this register
2890 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
2893 * Now we commit the changes to the software state
2897 * update overflow information
2901 * full flag update each time a register is programmed
2903 ctx
->ctx_pmds
[cnum
].flags
= flags
;
2905 ctx
->ctx_pmds
[cnum
].reset_pmds
[0] = reset_pmds
;
2906 ctx
->ctx_pmds
[cnum
].smpl_pmds
[0] = smpl_pmds
;
2907 ctx
->ctx_pmds
[cnum
].eventid
= req
->reg_smpl_eventid
;
2910 * Mark all PMDS to be accessed as used.
2912 * We do not keep track of PMC because we have to
2913 * systematically restore ALL of them.
2915 * We do not update the used_monitors mask, because
2916 * if we have not programmed them, then will be in
2917 * a quiescent state, therefore we will not need to
2918 * mask/restore then when context is MASKED.
2920 CTX_USED_PMD(ctx
, reset_pmds
);
2921 CTX_USED_PMD(ctx
, smpl_pmds
);
2923 * make sure we do not try to reset on
2924 * restart because we have established new values
2926 if (state
== PFM_CTX_MASKED
) ctx
->ctx_ovfl_regs
[0] &= ~1UL << cnum
;
2929 * Needed in case the user does not initialize the equivalent
2930 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2931 * possible leak here.
2933 CTX_USED_PMD(ctx
, pmu_conf
->pmc_desc
[cnum
].dep_pmd
[0]);
2936 * keep track of the monitor PMC that we are using.
2937 * we save the value of the pmc in ctx_pmcs[] and if
2938 * the monitoring is not stopped for the context we also
2939 * place it in the saved state area so that it will be
2940 * picked up later by the context switch code.
2942 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
2944 * The value in th_pmcs[] may be modified on overflow, i.e., when
2945 * monitoring needs to be stopped.
2947 if (is_monitor
) CTX_USED_MONITOR(ctx
, 1UL << cnum
);
2950 * update context state
2952 ctx
->ctx_pmcs
[cnum
] = value
;
2956 * write thread state
2958 if (is_system
== 0) ctx
->th_pmcs
[cnum
] = value
;
2961 * write hardware register if we can
2963 if (can_access_pmu
) {
2964 ia64_set_pmc(cnum
, value
);
2969 * per-task SMP only here
2971 * we are guaranteed that the task is not running on the other CPU,
2972 * we indicate that this PMD will need to be reloaded if the task
2973 * is rescheduled on the CPU it ran last on.
2975 ctx
->ctx_reload_pmcs
[0] |= 1UL << cnum
;
2980 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",
2986 ctx
->ctx_all_pmcs
[0],
2987 ctx
->ctx_used_pmds
[0],
2988 ctx
->ctx_pmds
[cnum
].eventid
,
2991 ctx
->ctx_reload_pmcs
[0],
2992 ctx
->ctx_used_monitors
[0],
2993 ctx
->ctx_ovfl_regs
[0]));
2997 * make sure the changes are visible
2999 if (can_access_pmu
) ia64_srlz_d();
3003 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3008 pfm_write_pmds(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3010 struct task_struct
*task
;
3011 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
3012 unsigned long value
, hw_value
, ovfl_mask
;
3014 int i
, can_access_pmu
= 0, state
;
3015 int is_counting
, is_loaded
, is_system
, expert_mode
;
3017 pfm_reg_check_t wr_func
;
3020 state
= ctx
->ctx_state
;
3021 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3022 is_system
= ctx
->ctx_fl_system
;
3023 ovfl_mask
= pmu_conf
->ovfl_val
;
3024 task
= ctx
->ctx_task
;
3026 if (unlikely(state
== PFM_CTX_ZOMBIE
)) return -EINVAL
;
3029 * on both UP and SMP, we can only write to the PMC when the task is
3030 * the owner of the local PMU.
3032 if (likely(is_loaded
)) {
3034 * In system wide and when the context is loaded, access can only happen
3035 * when the caller is running on the CPU being monitored by the session.
3036 * It does not have to be the owner (ctx_task) of the context per se.
3038 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3039 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3042 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3044 expert_mode
= pfm_sysctl
.expert_mode
;
3046 for (i
= 0; i
< count
; i
++, req
++) {
3048 cnum
= req
->reg_num
;
3049 value
= req
->reg_value
;
3051 if (!PMD_IS_IMPL(cnum
)) {
3052 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum
));
3055 is_counting
= PMD_IS_COUNTING(cnum
);
3056 wr_func
= pmu_conf
->pmd_desc
[cnum
].write_check
;
3059 * execute write checker, if any
3061 if (unlikely(expert_mode
== 0 && wr_func
)) {
3062 unsigned long v
= value
;
3064 ret
= (*wr_func
)(task
, ctx
, cnum
, &v
, regs
);
3065 if (ret
) goto abort_mission
;
3072 * no error on this register
3074 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
3077 * now commit changes to software state
3082 * update virtualized (64bits) counter
3086 * write context state
3088 ctx
->ctx_pmds
[cnum
].lval
= value
;
3091 * when context is load we use the split value
3094 hw_value
= value
& ovfl_mask
;
3095 value
= value
& ~ovfl_mask
;
3099 * update reset values (not just for counters)
3101 ctx
->ctx_pmds
[cnum
].long_reset
= req
->reg_long_reset
;
3102 ctx
->ctx_pmds
[cnum
].short_reset
= req
->reg_short_reset
;
3105 * update randomization parameters (not just for counters)
3107 ctx
->ctx_pmds
[cnum
].seed
= req
->reg_random_seed
;
3108 ctx
->ctx_pmds
[cnum
].mask
= req
->reg_random_mask
;
3111 * update context value
3113 ctx
->ctx_pmds
[cnum
].val
= value
;
3116 * Keep track of what we use
3118 * We do not keep track of PMC because we have to
3119 * systematically restore ALL of them.
3121 CTX_USED_PMD(ctx
, PMD_PMD_DEP(cnum
));
3124 * mark this PMD register used as well
3126 CTX_USED_PMD(ctx
, RDEP(cnum
));
3129 * make sure we do not try to reset on
3130 * restart because we have established new values
3132 if (is_counting
&& state
== PFM_CTX_MASKED
) {
3133 ctx
->ctx_ovfl_regs
[0] &= ~1UL << cnum
;
3138 * write thread state
3140 if (is_system
== 0) ctx
->th_pmds
[cnum
] = hw_value
;
3143 * write hardware register if we can
3145 if (can_access_pmu
) {
3146 ia64_set_pmd(cnum
, hw_value
);
3150 * we are guaranteed that the task is not running on the other CPU,
3151 * we indicate that this PMD will need to be reloaded if the task
3152 * is rescheduled on the CPU it ran last on.
3154 ctx
->ctx_reload_pmds
[0] |= 1UL << cnum
;
3159 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3160 "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",
3166 ctx
->ctx_pmds
[cnum
].val
,
3167 ctx
->ctx_pmds
[cnum
].short_reset
,
3168 ctx
->ctx_pmds
[cnum
].long_reset
,
3169 PMC_OVFL_NOTIFY(ctx
, cnum
) ? 'Y':'N',
3170 ctx
->ctx_pmds
[cnum
].seed
,
3171 ctx
->ctx_pmds
[cnum
].mask
,
3172 ctx
->ctx_used_pmds
[0],
3173 ctx
->ctx_pmds
[cnum
].reset_pmds
[0],
3174 ctx
->ctx_reload_pmds
[0],
3175 ctx
->ctx_all_pmds
[0],
3176 ctx
->ctx_ovfl_regs
[0]));
3180 * make changes visible
3182 if (can_access_pmu
) ia64_srlz_d();
3188 * for now, we have only one possibility for error
3190 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3195 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3196 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3197 * interrupt is delivered during the call, it will be kept pending until we leave, making
3198 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3199 * guaranteed to return consistent data to the user, it may simply be old. It is not
3200 * trivial to treat the overflow while inside the call because you may end up in
3201 * some module sampling buffer code causing deadlocks.
3204 pfm_read_pmds(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3206 struct task_struct
*task
;
3207 unsigned long val
= 0UL, lval
, ovfl_mask
, sval
;
3208 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
3209 unsigned int cnum
, reg_flags
= 0;
3210 int i
, can_access_pmu
= 0, state
;
3211 int is_loaded
, is_system
, is_counting
, expert_mode
;
3213 pfm_reg_check_t rd_func
;
3216 * access is possible when loaded only for
3217 * self-monitoring tasks or in UP mode
3220 state
= ctx
->ctx_state
;
3221 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3222 is_system
= ctx
->ctx_fl_system
;
3223 ovfl_mask
= pmu_conf
->ovfl_val
;
3224 task
= ctx
->ctx_task
;
3226 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
3228 if (likely(is_loaded
)) {
3230 * In system wide and when the context is loaded, access can only happen
3231 * when the caller is running on the CPU being monitored by the session.
3232 * It does not have to be the owner (ctx_task) of the context per se.
3234 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3235 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3239 * this can be true when not self-monitoring only in UP
3241 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3243 if (can_access_pmu
) ia64_srlz_d();
3245 expert_mode
= pfm_sysctl
.expert_mode
;
3247 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3253 * on both UP and SMP, we can only read the PMD from the hardware register when
3254 * the task is the owner of the local PMU.
3257 for (i
= 0; i
< count
; i
++, req
++) {
3259 cnum
= req
->reg_num
;
3260 reg_flags
= req
->reg_flags
;
3262 if (unlikely(!PMD_IS_IMPL(cnum
))) goto error
;
3264 * we can only read the register that we use. That includes
3265 * the one we explicitly initialize AND the one we want included
3266 * in the sampling buffer (smpl_regs).
3268 * Having this restriction allows optimization in the ctxsw routine
3269 * without compromising security (leaks)
3271 if (unlikely(!CTX_IS_USED_PMD(ctx
, cnum
))) goto error
;
3273 sval
= ctx
->ctx_pmds
[cnum
].val
;
3274 lval
= ctx
->ctx_pmds
[cnum
].lval
;
3275 is_counting
= PMD_IS_COUNTING(cnum
);
3278 * If the task is not the current one, then we check if the
3279 * PMU state is still in the local live register due to lazy ctxsw.
3280 * If true, then we read directly from the registers.
3282 if (can_access_pmu
){
3283 val
= ia64_get_pmd(cnum
);
3286 * context has been saved
3287 * if context is zombie, then task does not exist anymore.
3288 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3290 val
= is_loaded
? ctx
->th_pmds
[cnum
] : 0UL;
3292 rd_func
= pmu_conf
->pmd_desc
[cnum
].read_check
;
3296 * XXX: need to check for overflow when loaded
3303 * execute read checker, if any
3305 if (unlikely(expert_mode
== 0 && rd_func
)) {
3306 unsigned long v
= val
;
3307 ret
= (*rd_func
)(ctx
->ctx_task
, ctx
, cnum
, &v
, regs
);
3308 if (ret
) goto error
;
3313 PFM_REG_RETFLAG_SET(reg_flags
, 0);
3315 DPRINT(("pmd[%u]=0x%lx\n", cnum
, val
));
3318 * update register return value, abort all if problem during copy.
3319 * we only modify the reg_flags field. no check mode is fine because
3320 * access has been verified upfront in sys_perfmonctl().
3322 req
->reg_value
= val
;
3323 req
->reg_flags
= reg_flags
;
3324 req
->reg_last_reset_val
= lval
;
3330 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
3335 pfm_mod_write_pmcs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3339 if (req
== NULL
) return -EINVAL
;
3341 ctx
= GET_PMU_CTX();
3343 if (ctx
== NULL
) return -EINVAL
;
3346 * for now limit to current task, which is enough when calling
3347 * from overflow handler
3349 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3351 return pfm_write_pmcs(ctx
, req
, nreq
, regs
);
3353 EXPORT_SYMBOL(pfm_mod_write_pmcs
);
3356 pfm_mod_read_pmds(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3360 if (req
== NULL
) return -EINVAL
;
3362 ctx
= GET_PMU_CTX();
3364 if (ctx
== NULL
) return -EINVAL
;
3367 * for now limit to current task, which is enough when calling
3368 * from overflow handler
3370 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3372 return pfm_read_pmds(ctx
, req
, nreq
, regs
);
3374 EXPORT_SYMBOL(pfm_mod_read_pmds
);
3377 * Only call this function when a process it trying to
3378 * write the debug registers (reading is always allowed)
3381 pfm_use_debug_registers(struct task_struct
*task
)
3383 pfm_context_t
*ctx
= task
->thread
.pfm_context
;
3384 unsigned long flags
;
3387 if (pmu_conf
->use_rr_dbregs
== 0) return 0;
3389 DPRINT(("called for [%d]\n", task_pid_nr(task
)));
3394 if (task
->thread
.flags
& IA64_THREAD_DBG_VALID
) return 0;
3397 * Even on SMP, we do not need to use an atomic here because
3398 * the only way in is via ptrace() and this is possible only when the
3399 * process is stopped. Even in the case where the ctxsw out is not totally
3400 * completed by the time we come here, there is no way the 'stopped' process
3401 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3402 * So this is always safe.
3404 if (ctx
&& ctx
->ctx_fl_using_dbreg
== 1) return -1;
3409 * We cannot allow setting breakpoints when system wide monitoring
3410 * sessions are using the debug registers.
3412 if (pfm_sessions
.pfs_sys_use_dbregs
> 0)
3415 pfm_sessions
.pfs_ptrace_use_dbregs
++;
3417 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3418 pfm_sessions
.pfs_ptrace_use_dbregs
,
3419 pfm_sessions
.pfs_sys_use_dbregs
,
3420 task_pid_nr(task
), ret
));
3428 * This function is called for every task that exits with the
3429 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3430 * able to use the debug registers for debugging purposes via
3431 * ptrace(). Therefore we know it was not using them for
3432 * performance monitoring, so we only decrement the number
3433 * of "ptraced" debug register users to keep the count up to date
3436 pfm_release_debug_registers(struct task_struct
*task
)
3438 unsigned long flags
;
3441 if (pmu_conf
->use_rr_dbregs
== 0) return 0;
3444 if (pfm_sessions
.pfs_ptrace_use_dbregs
== 0) {
3445 printk(KERN_ERR
"perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task
));
3448 pfm_sessions
.pfs_ptrace_use_dbregs
--;
3457 pfm_restart(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3459 struct task_struct
*task
;
3460 pfm_buffer_fmt_t
*fmt
;
3461 pfm_ovfl_ctrl_t rst_ctrl
;
3462 int state
, is_system
;
3465 state
= ctx
->ctx_state
;
3466 fmt
= ctx
->ctx_buf_fmt
;
3467 is_system
= ctx
->ctx_fl_system
;
3468 task
= PFM_CTX_TASK(ctx
);
3471 case PFM_CTX_MASKED
:
3473 case PFM_CTX_LOADED
:
3474 if (CTX_HAS_SMPL(ctx
) && fmt
->fmt_restart_active
) break;
3476 case PFM_CTX_UNLOADED
:
3477 case PFM_CTX_ZOMBIE
:
3478 DPRINT(("invalid state=%d\n", state
));
3481 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state
));
3486 * In system wide and when the context is loaded, access can only happen
3487 * when the caller is running on the CPU being monitored by the session.
3488 * It does not have to be the owner (ctx_task) of the context per se.
3490 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
3491 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3496 if (unlikely(task
== NULL
)) {
3497 printk(KERN_ERR
"perfmon: [%d] pfm_restart no task\n", task_pid_nr(current
));
3501 if (task
== current
|| is_system
) {
3503 fmt
= ctx
->ctx_buf_fmt
;
3505 DPRINT(("restarting self %d ovfl=0x%lx\n",
3507 ctx
->ctx_ovfl_regs
[0]));
3509 if (CTX_HAS_SMPL(ctx
)) {
3511 prefetch(ctx
->ctx_smpl_hdr
);
3513 rst_ctrl
.bits
.mask_monitoring
= 0;
3514 rst_ctrl
.bits
.reset_ovfl_pmds
= 0;
3516 if (state
== PFM_CTX_LOADED
)
3517 ret
= pfm_buf_fmt_restart_active(fmt
, task
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
3519 ret
= pfm_buf_fmt_restart(fmt
, task
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
3521 rst_ctrl
.bits
.mask_monitoring
= 0;
3522 rst_ctrl
.bits
.reset_ovfl_pmds
= 1;
3526 if (rst_ctrl
.bits
.reset_ovfl_pmds
)
3527 pfm_reset_regs(ctx
, ctx
->ctx_ovfl_regs
, PFM_PMD_LONG_RESET
);
3529 if (rst_ctrl
.bits
.mask_monitoring
== 0) {
3530 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task
)));
3532 if (state
== PFM_CTX_MASKED
) pfm_restore_monitoring(task
);
3534 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task
)));
3536 // cannot use pfm_stop_monitoring(task, regs);
3540 * clear overflowed PMD mask to remove any stale information
3542 ctx
->ctx_ovfl_regs
[0] = 0UL;
3545 * back to LOADED state
3547 ctx
->ctx_state
= PFM_CTX_LOADED
;
3550 * XXX: not really useful for self monitoring
3552 ctx
->ctx_fl_can_restart
= 0;
3558 * restart another task
3562 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3563 * one is seen by the task.
3565 if (state
== PFM_CTX_MASKED
) {
3566 if (ctx
->ctx_fl_can_restart
== 0) return -EINVAL
;
3568 * will prevent subsequent restart before this one is
3569 * seen by other task
3571 ctx
->ctx_fl_can_restart
= 0;
3575 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3576 * the task is blocked or on its way to block. That's the normal
3577 * restart path. If the monitoring is not masked, then the task
3578 * can be actively monitoring and we cannot directly intervene.
3579 * Therefore we use the trap mechanism to catch the task and
3580 * force it to reset the buffer/reset PMDs.
3582 * if non-blocking, then we ensure that the task will go into
3583 * pfm_handle_work() before returning to user mode.
3585 * We cannot explicitly reset another task, it MUST always
3586 * be done by the task itself. This works for system wide because
3587 * the tool that is controlling the session is logically doing
3588 * "self-monitoring".
3590 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && state
== PFM_CTX_MASKED
) {
3591 DPRINT(("unblocking [%d]\n", task_pid_nr(task
)));
3592 complete(&ctx
->ctx_restart_done
);
3594 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task
)));
3596 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_RESET
;
3598 PFM_SET_WORK_PENDING(task
, 1);
3600 set_notify_resume(task
);
3603 * XXX: send reschedule if task runs on another CPU
3610 pfm_debug(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3612 unsigned int m
= *(unsigned int *)arg
;
3614 pfm_sysctl
.debug
= m
== 0 ? 0 : 1;
3616 printk(KERN_INFO
"perfmon debugging %s (timing reset)\n", pfm_sysctl
.debug
? "on" : "off");
3619 memset(pfm_stats
, 0, sizeof(pfm_stats
));
3620 for(m
=0; m
< NR_CPUS
; m
++) pfm_stats
[m
].pfm_ovfl_intr_cycles_min
= ~0UL;
3626 * arg can be NULL and count can be zero for this function
3629 pfm_write_ibr_dbr(int mode
, pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3631 struct thread_struct
*thread
= NULL
;
3632 struct task_struct
*task
;
3633 pfarg_dbreg_t
*req
= (pfarg_dbreg_t
*)arg
;
3634 unsigned long flags
;
3639 int i
, can_access_pmu
= 0;
3640 int is_system
, is_loaded
;
3642 if (pmu_conf
->use_rr_dbregs
== 0) return -EINVAL
;
3644 state
= ctx
->ctx_state
;
3645 is_loaded
= state
== PFM_CTX_LOADED
? 1 : 0;
3646 is_system
= ctx
->ctx_fl_system
;
3647 task
= ctx
->ctx_task
;
3649 if (state
== PFM_CTX_ZOMBIE
) return -EINVAL
;
3652 * on both UP and SMP, we can only write to the PMC when the task is
3653 * the owner of the local PMU.
3656 thread
= &task
->thread
;
3658 * In system wide and when the context is loaded, access can only happen
3659 * when the caller is running on the CPU being monitored by the session.
3660 * It does not have to be the owner (ctx_task) of the context per se.
3662 if (unlikely(is_system
&& ctx
->ctx_cpu
!= smp_processor_id())) {
3663 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3666 can_access_pmu
= GET_PMU_OWNER() == task
|| is_system
? 1 : 0;
3670 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3671 * ensuring that no real breakpoint can be installed via this call.
3673 * IMPORTANT: regs can be NULL in this function
3676 first_time
= ctx
->ctx_fl_using_dbreg
== 0;
3679 * don't bother if we are loaded and task is being debugged
3681 if (is_loaded
&& (thread
->flags
& IA64_THREAD_DBG_VALID
) != 0) {
3682 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task
)));
3687 * check for debug registers in system wide mode
3689 * If though a check is done in pfm_context_load(),
3690 * we must repeat it here, in case the registers are
3691 * written after the context is loaded
3696 if (first_time
&& is_system
) {
3697 if (pfm_sessions
.pfs_ptrace_use_dbregs
)
3700 pfm_sessions
.pfs_sys_use_dbregs
++;
3705 if (ret
!= 0) return ret
;
3708 * mark ourself as user of the debug registers for
3711 ctx
->ctx_fl_using_dbreg
= 1;
3714 * clear hardware registers to make sure we don't
3715 * pick up stale state.
3717 * for a system wide session, we do not use
3718 * thread.dbr, thread.ibr because this process
3719 * never leaves the current CPU and the state
3720 * is shared by all processes running on it
3722 if (first_time
&& can_access_pmu
) {
3723 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task
)));
3724 for (i
=0; i
< pmu_conf
->num_ibrs
; i
++) {
3725 ia64_set_ibr(i
, 0UL);
3726 ia64_dv_serialize_instruction();
3729 for (i
=0; i
< pmu_conf
->num_dbrs
; i
++) {
3730 ia64_set_dbr(i
, 0UL);
3731 ia64_dv_serialize_data();
3737 * Now install the values into the registers
3739 for (i
= 0; i
< count
; i
++, req
++) {
3741 rnum
= req
->dbreg_num
;
3742 dbreg
.val
= req
->dbreg_value
;
3746 if ((mode
== PFM_CODE_RR
&& rnum
>= PFM_NUM_IBRS
) || ((mode
== PFM_DATA_RR
) && rnum
>= PFM_NUM_DBRS
)) {
3747 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3748 rnum
, dbreg
.val
, mode
, i
, count
));
3754 * make sure we do not install enabled breakpoint
3757 if (mode
== PFM_CODE_RR
)
3758 dbreg
.ibr
.ibr_x
= 0;
3760 dbreg
.dbr
.dbr_r
= dbreg
.dbr
.dbr_w
= 0;
3763 PFM_REG_RETFLAG_SET(req
->dbreg_flags
, 0);
3766 * Debug registers, just like PMC, can only be modified
3767 * by a kernel call. Moreover, perfmon() access to those
3768 * registers are centralized in this routine. The hardware
3769 * does not modify the value of these registers, therefore,
3770 * if we save them as they are written, we can avoid having
3771 * to save them on context switch out. This is made possible
3772 * by the fact that when perfmon uses debug registers, ptrace()
3773 * won't be able to modify them concurrently.
3775 if (mode
== PFM_CODE_RR
) {
3776 CTX_USED_IBR(ctx
, rnum
);
3778 if (can_access_pmu
) {
3779 ia64_set_ibr(rnum
, dbreg
.val
);
3780 ia64_dv_serialize_instruction();
3783 ctx
->ctx_ibrs
[rnum
] = dbreg
.val
;
3785 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3786 rnum
, dbreg
.val
, ctx
->ctx_used_ibrs
[0], is_loaded
, can_access_pmu
));
3788 CTX_USED_DBR(ctx
, rnum
);
3790 if (can_access_pmu
) {
3791 ia64_set_dbr(rnum
, dbreg
.val
);
3792 ia64_dv_serialize_data();
3794 ctx
->ctx_dbrs
[rnum
] = dbreg
.val
;
3796 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3797 rnum
, dbreg
.val
, ctx
->ctx_used_dbrs
[0], is_loaded
, can_access_pmu
));
3805 * in case it was our first attempt, we undo the global modifications
3809 if (ctx
->ctx_fl_system
) {
3810 pfm_sessions
.pfs_sys_use_dbregs
--;
3813 ctx
->ctx_fl_using_dbreg
= 0;
3816 * install error return flag
3818 PFM_REG_RETFLAG_SET(req
->dbreg_flags
, PFM_REG_RETFL_EINVAL
);
3824 pfm_write_ibrs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3826 return pfm_write_ibr_dbr(PFM_CODE_RR
, ctx
, arg
, count
, regs
);
3830 pfm_write_dbrs(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3832 return pfm_write_ibr_dbr(PFM_DATA_RR
, ctx
, arg
, count
, regs
);
3836 pfm_mod_write_ibrs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3840 if (req
== NULL
) return -EINVAL
;
3842 ctx
= GET_PMU_CTX();
3844 if (ctx
== NULL
) return -EINVAL
;
3847 * for now limit to current task, which is enough when calling
3848 * from overflow handler
3850 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3852 return pfm_write_ibrs(ctx
, req
, nreq
, regs
);
3854 EXPORT_SYMBOL(pfm_mod_write_ibrs
);
3857 pfm_mod_write_dbrs(struct task_struct
*task
, void *req
, unsigned int nreq
, struct pt_regs
*regs
)
3861 if (req
== NULL
) return -EINVAL
;
3863 ctx
= GET_PMU_CTX();
3865 if (ctx
== NULL
) return -EINVAL
;
3868 * for now limit to current task, which is enough when calling
3869 * from overflow handler
3871 if (task
!= current
&& ctx
->ctx_fl_system
== 0) return -EBUSY
;
3873 return pfm_write_dbrs(ctx
, req
, nreq
, regs
);
3875 EXPORT_SYMBOL(pfm_mod_write_dbrs
);
3879 pfm_get_features(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3881 pfarg_features_t
*req
= (pfarg_features_t
*)arg
;
3883 req
->ft_version
= PFM_VERSION
;
3888 pfm_stop(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3890 struct pt_regs
*tregs
;
3891 struct task_struct
*task
= PFM_CTX_TASK(ctx
);
3892 int state
, is_system
;
3894 state
= ctx
->ctx_state
;
3895 is_system
= ctx
->ctx_fl_system
;
3898 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3900 if (state
== PFM_CTX_UNLOADED
) return -EINVAL
;
3903 * In system wide and when the context is loaded, access can only happen
3904 * when the caller is running on the CPU being monitored by the session.
3905 * It does not have to be the owner (ctx_task) of the context per se.
3907 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
3908 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3911 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3912 task_pid_nr(PFM_CTX_TASK(ctx
)),
3916 * in system mode, we need to update the PMU directly
3917 * and the user level state of the caller, which may not
3918 * necessarily be the creator of the context.
3922 * Update local PMU first
3926 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) & ~IA64_DCR_PP
);
3930 * update local cpuinfo
3932 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP
);
3935 * stop monitoring, does srlz.i
3940 * stop monitoring in the caller
3942 ia64_psr(regs
)->pp
= 0;
3950 if (task
== current
) {
3951 /* stop monitoring at kernel level */
3955 * stop monitoring at the user level
3957 ia64_psr(regs
)->up
= 0;
3959 tregs
= task_pt_regs(task
);
3962 * stop monitoring at the user level
3964 ia64_psr(tregs
)->up
= 0;
3967 * monitoring disabled in kernel at next reschedule
3969 ctx
->ctx_saved_psr_up
= 0;
3970 DPRINT(("task=[%d]\n", task_pid_nr(task
)));
3977 pfm_start(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
3979 struct pt_regs
*tregs
;
3980 int state
, is_system
;
3982 state
= ctx
->ctx_state
;
3983 is_system
= ctx
->ctx_fl_system
;
3985 if (state
!= PFM_CTX_LOADED
) return -EINVAL
;
3988 * In system wide and when the context is loaded, access can only happen
3989 * when the caller is running on the CPU being monitored by the session.
3990 * It does not have to be the owner (ctx_task) of the context per se.
3992 if (is_system
&& ctx
->ctx_cpu
!= smp_processor_id()) {
3993 DPRINT(("should be running on CPU%d\n", ctx
->ctx_cpu
));
3998 * in system mode, we need to update the PMU directly
3999 * and the user level state of the caller, which may not
4000 * necessarily be the creator of the context.
4005 * set user level psr.pp for the caller
4007 ia64_psr(regs
)->pp
= 1;
4010 * now update the local PMU and cpuinfo
4012 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP
);
4015 * start monitoring at kernel level
4020 ia64_setreg(_IA64_REG_CR_DCR
, ia64_getreg(_IA64_REG_CR_DCR
) | IA64_DCR_PP
);
4030 if (ctx
->ctx_task
== current
) {
4032 /* start monitoring at kernel level */
4036 * activate monitoring at user level
4038 ia64_psr(regs
)->up
= 1;
4041 tregs
= task_pt_regs(ctx
->ctx_task
);
4044 * start monitoring at the kernel level the next
4045 * time the task is scheduled
4047 ctx
->ctx_saved_psr_up
= IA64_PSR_UP
;
4050 * activate monitoring at user level
4052 ia64_psr(tregs
)->up
= 1;
4058 pfm_get_pmc_reset(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4060 pfarg_reg_t
*req
= (pfarg_reg_t
*)arg
;
4065 for (i
= 0; i
< count
; i
++, req
++) {
4067 cnum
= req
->reg_num
;
4069 if (!PMC_IS_IMPL(cnum
)) goto abort_mission
;
4071 req
->reg_value
= PMC_DFL_VAL(cnum
);
4073 PFM_REG_RETFLAG_SET(req
->reg_flags
, 0);
4075 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum
, req
->reg_value
));
4080 PFM_REG_RETFLAG_SET(req
->reg_flags
, PFM_REG_RETFL_EINVAL
);
4085 pfm_check_task_exist(pfm_context_t
*ctx
)
4087 struct task_struct
*g
, *t
;
4090 read_lock(&tasklist_lock
);
4092 do_each_thread (g
, t
) {
4093 if (t
->thread
.pfm_context
== ctx
) {
4097 } while_each_thread (g
, t
);
4099 read_unlock(&tasklist_lock
);
4101 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret
, ctx
));
4107 pfm_context_load(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4109 struct task_struct
*task
;
4110 struct thread_struct
*thread
;
4111 struct pfm_context_t
*old
;
4112 unsigned long flags
;
4114 struct task_struct
*owner_task
= NULL
;
4116 pfarg_load_t
*req
= (pfarg_load_t
*)arg
;
4117 unsigned long *pmcs_source
, *pmds_source
;
4120 int state
, is_system
, set_dbregs
= 0;
4122 state
= ctx
->ctx_state
;
4123 is_system
= ctx
->ctx_fl_system
;
4125 * can only load from unloaded or terminated state
4127 if (state
!= PFM_CTX_UNLOADED
) {
4128 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4134 DPRINT(("load_pid [%d] using_dbreg=%d\n", req
->load_pid
, ctx
->ctx_fl_using_dbreg
));
4136 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && req
->load_pid
== current
->pid
) {
4137 DPRINT(("cannot use blocking mode on self\n"));
4141 ret
= pfm_get_task(ctx
, req
->load_pid
, &task
);
4143 DPRINT(("load_pid [%d] get_task=%d\n", req
->load_pid
, ret
));
4150 * system wide is self monitoring only
4152 if (is_system
&& task
!= current
) {
4153 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4158 thread
= &task
->thread
;
4162 * cannot load a context which is using range restrictions,
4163 * into a task that is being debugged.
4165 if (ctx
->ctx_fl_using_dbreg
) {
4166 if (thread
->flags
& IA64_THREAD_DBG_VALID
) {
4168 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req
->load_pid
));
4174 if (pfm_sessions
.pfs_ptrace_use_dbregs
) {
4175 DPRINT(("cannot load [%d] dbregs in use\n",
4176 task_pid_nr(task
)));
4179 pfm_sessions
.pfs_sys_use_dbregs
++;
4180 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task
), pfm_sessions
.pfs_sys_use_dbregs
));
4187 if (ret
) goto error
;
4191 * SMP system-wide monitoring implies self-monitoring.
4193 * The programming model expects the task to
4194 * be pinned on a CPU throughout the session.
4195 * Here we take note of the current CPU at the
4196 * time the context is loaded. No call from
4197 * another CPU will be allowed.
4199 * The pinning via shed_setaffinity()
4200 * must be done by the calling task prior
4203 * systemwide: keep track of CPU this session is supposed to run on
4205 the_cpu
= ctx
->ctx_cpu
= smp_processor_id();
4209 * now reserve the session
4211 ret
= pfm_reserve_session(current
, is_system
, the_cpu
);
4212 if (ret
) goto error
;
4215 * task is necessarily stopped at this point.
4217 * If the previous context was zombie, then it got removed in
4218 * pfm_save_regs(). Therefore we should not see it here.
4219 * If we see a context, then this is an active context
4221 * XXX: needs to be atomic
4223 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4224 thread
->pfm_context
, ctx
));
4227 old
= ia64_cmpxchg(acq
, &thread
->pfm_context
, NULL
, ctx
, sizeof(pfm_context_t
*));
4229 DPRINT(("load_pid [%d] already has a context\n", req
->load_pid
));
4233 pfm_reset_msgq(ctx
);
4235 ctx
->ctx_state
= PFM_CTX_LOADED
;
4238 * link context to task
4240 ctx
->ctx_task
= task
;
4244 * we load as stopped
4246 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE
);
4247 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP
);
4249 if (ctx
->ctx_fl_excl_idle
) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE
);
4251 thread
->flags
|= IA64_THREAD_PM_VALID
;
4255 * propagate into thread-state
4257 pfm_copy_pmds(task
, ctx
);
4258 pfm_copy_pmcs(task
, ctx
);
4260 pmcs_source
= ctx
->th_pmcs
;
4261 pmds_source
= ctx
->th_pmds
;
4264 * always the case for system-wide
4266 if (task
== current
) {
4268 if (is_system
== 0) {
4270 /* allow user level control */
4271 ia64_psr(regs
)->sp
= 0;
4272 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task
)));
4274 SET_LAST_CPU(ctx
, smp_processor_id());
4276 SET_ACTIVATION(ctx
);
4279 * push the other task out, if any
4281 owner_task
= GET_PMU_OWNER();
4282 if (owner_task
) pfm_lazy_save_regs(owner_task
);
4286 * load all PMD from ctx to PMU (as opposed to thread state)
4287 * restore all PMC from ctx to PMU
4289 pfm_restore_pmds(pmds_source
, ctx
->ctx_all_pmds
[0]);
4290 pfm_restore_pmcs(pmcs_source
, ctx
->ctx_all_pmcs
[0]);
4292 ctx
->ctx_reload_pmcs
[0] = 0UL;
4293 ctx
->ctx_reload_pmds
[0] = 0UL;
4296 * guaranteed safe by earlier check against DBG_VALID
4298 if (ctx
->ctx_fl_using_dbreg
) {
4299 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
4300 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
4305 SET_PMU_OWNER(task
, ctx
);
4307 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task
)));
4310 * when not current, task MUST be stopped, so this is safe
4312 regs
= task_pt_regs(task
);
4314 /* force a full reload */
4315 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
4316 SET_LAST_CPU(ctx
, -1);
4318 /* initial saved psr (stopped) */
4319 ctx
->ctx_saved_psr_up
= 0UL;
4320 ia64_psr(regs
)->up
= ia64_psr(regs
)->pp
= 0;
4326 if (ret
) pfm_unreserve_session(ctx
, ctx
->ctx_fl_system
, the_cpu
);
4329 * we must undo the dbregs setting (for system-wide)
4331 if (ret
&& set_dbregs
) {
4333 pfm_sessions
.pfs_sys_use_dbregs
--;
4337 * release task, there is now a link with the context
4339 if (is_system
== 0 && task
!= current
) {
4343 ret
= pfm_check_task_exist(ctx
);
4345 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
4346 ctx
->ctx_task
= NULL
;
4354 * in this function, we do not need to increase the use count
4355 * for the task via get_task_struct(), because we hold the
4356 * context lock. If the task were to disappear while having
4357 * a context attached, it would go through pfm_exit_thread()
4358 * which also grabs the context lock and would therefore be blocked
4359 * until we are here.
4361 static void pfm_flush_pmds(struct task_struct
*, pfm_context_t
*ctx
);
4364 pfm_context_unload(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
)
4366 struct task_struct
*task
= PFM_CTX_TASK(ctx
);
4367 struct pt_regs
*tregs
;
4368 int prev_state
, is_system
;
4371 DPRINT(("ctx_state=%d task [%d]\n", ctx
->ctx_state
, task
? task_pid_nr(task
) : -1));
4373 prev_state
= ctx
->ctx_state
;
4374 is_system
= ctx
->ctx_fl_system
;
4377 * unload only when necessary
4379 if (prev_state
== PFM_CTX_UNLOADED
) {
4380 DPRINT(("ctx_state=%d, nothing to do\n", prev_state
));
4385 * clear psr and dcr bits
4387 ret
= pfm_stop(ctx
, NULL
, 0, regs
);
4388 if (ret
) return ret
;
4390 ctx
->ctx_state
= PFM_CTX_UNLOADED
;
4393 * in system mode, we need to update the PMU directly
4394 * and the user level state of the caller, which may not
4395 * necessarily be the creator of the context.
4402 * local PMU is taken care of in pfm_stop()
4404 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE
);
4405 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE
);
4408 * save PMDs in context
4411 pfm_flush_pmds(current
, ctx
);
4414 * at this point we are done with the PMU
4415 * so we can unreserve the resource.
4417 if (prev_state
!= PFM_CTX_ZOMBIE
)
4418 pfm_unreserve_session(ctx
, 1 , ctx
->ctx_cpu
);
4421 * disconnect context from task
4423 task
->thread
.pfm_context
= NULL
;
4425 * disconnect task from context
4427 ctx
->ctx_task
= NULL
;
4430 * There is nothing more to cleanup here.
4438 tregs
= task
== current
? regs
: task_pt_regs(task
);
4440 if (task
== current
) {
4442 * cancel user level control
4444 ia64_psr(regs
)->sp
= 1;
4446 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task
)));
4449 * save PMDs to context
4452 pfm_flush_pmds(task
, ctx
);
4455 * at this point we are done with the PMU
4456 * so we can unreserve the resource.
4458 * when state was ZOMBIE, we have already unreserved.
4460 if (prev_state
!= PFM_CTX_ZOMBIE
)
4461 pfm_unreserve_session(ctx
, 0 , ctx
->ctx_cpu
);
4464 * reset activation counter and psr
4466 ctx
->ctx_last_activation
= PFM_INVALID_ACTIVATION
;
4467 SET_LAST_CPU(ctx
, -1);
4470 * PMU state will not be restored
4472 task
->thread
.flags
&= ~IA64_THREAD_PM_VALID
;
4475 * break links between context and task
4477 task
->thread
.pfm_context
= NULL
;
4478 ctx
->ctx_task
= NULL
;
4480 PFM_SET_WORK_PENDING(task
, 0);
4482 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_NONE
;
4483 ctx
->ctx_fl_can_restart
= 0;
4484 ctx
->ctx_fl_going_zombie
= 0;
4486 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task
)));
4493 * called only from exit_thread()
4494 * we come here only if the task has a context attached (loaded or masked)
4497 pfm_exit_thread(struct task_struct
*task
)
4500 unsigned long flags
;
4501 struct pt_regs
*regs
= task_pt_regs(task
);
4505 ctx
= PFM_GET_CTX(task
);
4507 PROTECT_CTX(ctx
, flags
);
4509 DPRINT(("state=%d task [%d]\n", ctx
->ctx_state
, task_pid_nr(task
)));
4511 state
= ctx
->ctx_state
;
4513 case PFM_CTX_UNLOADED
:
4515 * only comes to this function if pfm_context is not NULL, i.e., cannot
4516 * be in unloaded state
4518 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task
));
4520 case PFM_CTX_LOADED
:
4521 case PFM_CTX_MASKED
:
4522 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
4524 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task
), state
, ret
);
4526 DPRINT(("ctx unloaded for current state was %d\n", state
));
4528 pfm_end_notify_user(ctx
);
4530 case PFM_CTX_ZOMBIE
:
4531 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
4533 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task
), state
, ret
);
4538 printk(KERN_ERR
"perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task
), state
);
4541 UNPROTECT_CTX(ctx
, flags
);
4543 { u64 psr
= pfm_get_psr();
4544 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
4545 BUG_ON(GET_PMU_OWNER());
4546 BUG_ON(ia64_psr(regs
)->up
);
4547 BUG_ON(ia64_psr(regs
)->pp
);
4551 * All memory free operations (especially for vmalloc'ed memory)
4552 * MUST be done with interrupts ENABLED.
4554 if (free_ok
) pfm_context_free(ctx
);
4558 * functions MUST be listed in the increasing order of their index (see permfon.h)
4560 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4561 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4562 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4563 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4564 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4566 static pfm_cmd_desc_t pfm_cmd_tab
[]={
4567 /* 0 */PFM_CMD_NONE
,
4568 /* 1 */PFM_CMD(pfm_write_pmcs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4569 /* 2 */PFM_CMD(pfm_write_pmds
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4570 /* 3 */PFM_CMD(pfm_read_pmds
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4571 /* 4 */PFM_CMD_S(pfm_stop
, PFM_CMD_PCLRWS
),
4572 /* 5 */PFM_CMD_S(pfm_start
, PFM_CMD_PCLRWS
),
4573 /* 6 */PFM_CMD_NONE
,
4574 /* 7 */PFM_CMD_NONE
,
4575 /* 8 */PFM_CMD(pfm_context_create
, PFM_CMD_ARG_RW
, 1, pfarg_context_t
, pfm_ctx_getsize
),
4576 /* 9 */PFM_CMD_NONE
,
4577 /* 10 */PFM_CMD_S(pfm_restart
, PFM_CMD_PCLRW
),
4578 /* 11 */PFM_CMD_NONE
,
4579 /* 12 */PFM_CMD(pfm_get_features
, PFM_CMD_ARG_RW
, 1, pfarg_features_t
, NULL
),
4580 /* 13 */PFM_CMD(pfm_debug
, 0, 1, unsigned int, NULL
),
4581 /* 14 */PFM_CMD_NONE
,
4582 /* 15 */PFM_CMD(pfm_get_pmc_reset
, PFM_CMD_ARG_RW
, PFM_CMD_ARG_MANY
, pfarg_reg_t
, NULL
),
4583 /* 16 */PFM_CMD(pfm_context_load
, PFM_CMD_PCLRWS
, 1, pfarg_load_t
, NULL
),
4584 /* 17 */PFM_CMD_S(pfm_context_unload
, PFM_CMD_PCLRWS
),
4585 /* 18 */PFM_CMD_NONE
,
4586 /* 19 */PFM_CMD_NONE
,
4587 /* 20 */PFM_CMD_NONE
,
4588 /* 21 */PFM_CMD_NONE
,
4589 /* 22 */PFM_CMD_NONE
,
4590 /* 23 */PFM_CMD_NONE
,
4591 /* 24 */PFM_CMD_NONE
,
4592 /* 25 */PFM_CMD_NONE
,
4593 /* 26 */PFM_CMD_NONE
,
4594 /* 27 */PFM_CMD_NONE
,
4595 /* 28 */PFM_CMD_NONE
,
4596 /* 29 */PFM_CMD_NONE
,
4597 /* 30 */PFM_CMD_NONE
,
4598 /* 31 */PFM_CMD_NONE
,
4599 /* 32 */PFM_CMD(pfm_write_ibrs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_dbreg_t
, NULL
),
4600 /* 33 */PFM_CMD(pfm_write_dbrs
, PFM_CMD_PCLRWS
, PFM_CMD_ARG_MANY
, pfarg_dbreg_t
, NULL
)
4602 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4605 pfm_check_task_state(pfm_context_t
*ctx
, int cmd
, unsigned long flags
)
4607 struct task_struct
*task
;
4608 int state
, old_state
;
4611 state
= ctx
->ctx_state
;
4612 task
= ctx
->ctx_task
;
4615 DPRINT(("context %d no task, state=%d\n", ctx
->ctx_fd
, state
));
4619 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4623 task
->state
, PFM_CMD_STOPPED(cmd
)));
4626 * self-monitoring always ok.
4628 * for system-wide the caller can either be the creator of the
4629 * context (to one to which the context is attached to) OR
4630 * a task running on the same CPU as the session.
4632 if (task
== current
|| ctx
->ctx_fl_system
) return 0;
4635 * we are monitoring another thread
4638 case PFM_CTX_UNLOADED
:
4640 * if context is UNLOADED we are safe to go
4643 case PFM_CTX_ZOMBIE
:
4645 * no command can operate on a zombie context
4647 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd
));
4649 case PFM_CTX_MASKED
:
4651 * PMU state has been saved to software even though
4652 * the thread may still be running.
4654 if (cmd
!= PFM_UNLOAD_CONTEXT
) return 0;
4658 * context is LOADED or MASKED. Some commands may need to have
4661 * We could lift this restriction for UP but it would mean that
4662 * the user has no guarantee the task would not run between
4663 * two successive calls to perfmonctl(). That's probably OK.
4664 * If this user wants to ensure the task does not run, then
4665 * the task must be stopped.
4667 if (PFM_CMD_STOPPED(cmd
)) {
4668 if (!task_is_stopped_or_traced(task
)) {
4669 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task
)));
4673 * task is now stopped, wait for ctxsw out
4675 * This is an interesting point in the code.
4676 * We need to unprotect the context because
4677 * the pfm_save_regs() routines needs to grab
4678 * the same lock. There are danger in doing
4679 * this because it leaves a window open for
4680 * another task to get access to the context
4681 * and possibly change its state. The one thing
4682 * that is not possible is for the context to disappear
4683 * because we are protected by the VFS layer, i.e.,
4684 * get_fd()/put_fd().
4688 UNPROTECT_CTX(ctx
, flags
);
4690 wait_task_inactive(task
, 0);
4692 PROTECT_CTX(ctx
, flags
);
4695 * we must recheck to verify if state has changed
4697 if (ctx
->ctx_state
!= old_state
) {
4698 DPRINT(("old_state=%d new_state=%d\n", old_state
, ctx
->ctx_state
));
4706 * system-call entry point (must return long)
4709 sys_perfmonctl (int fd
, int cmd
, void __user
*arg
, int count
)
4711 struct fd f
= {NULL
, 0};
4712 pfm_context_t
*ctx
= NULL
;
4713 unsigned long flags
= 0UL;
4714 void *args_k
= NULL
;
4715 long ret
; /* will expand int return types */
4716 size_t base_sz
, sz
, xtra_sz
= 0;
4717 int narg
, completed_args
= 0, call_made
= 0, cmd_flags
;
4718 int (*func
)(pfm_context_t
*ctx
, void *arg
, int count
, struct pt_regs
*regs
);
4719 int (*getsize
)(void *arg
, size_t *sz
);
4720 #define PFM_MAX_ARGSIZE 4096
4723 * reject any call if perfmon was disabled at initialization
4725 if (unlikely(pmu_conf
== NULL
)) return -ENOSYS
;
4727 if (unlikely(cmd
< 0 || cmd
>= PFM_CMD_COUNT
)) {
4728 DPRINT(("invalid cmd=%d\n", cmd
));
4732 func
= pfm_cmd_tab
[cmd
].cmd_func
;
4733 narg
= pfm_cmd_tab
[cmd
].cmd_narg
;
4734 base_sz
= pfm_cmd_tab
[cmd
].cmd_argsize
;
4735 getsize
= pfm_cmd_tab
[cmd
].cmd_getsize
;
4736 cmd_flags
= pfm_cmd_tab
[cmd
].cmd_flags
;
4738 if (unlikely(func
== NULL
)) {
4739 DPRINT(("invalid cmd=%d\n", cmd
));
4743 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4751 * check if number of arguments matches what the command expects
4753 if (unlikely((narg
== PFM_CMD_ARG_MANY
&& count
<= 0) || (narg
> 0 && narg
!= count
)))
4757 sz
= xtra_sz
+ base_sz
*count
;
4759 * limit abuse to min page size
4761 if (unlikely(sz
> PFM_MAX_ARGSIZE
)) {
4762 printk(KERN_ERR
"perfmon: [%d] argument too big %lu\n", task_pid_nr(current
), sz
);
4767 * allocate default-sized argument buffer
4769 if (likely(count
&& args_k
== NULL
)) {
4770 args_k
= kmalloc(PFM_MAX_ARGSIZE
, GFP_KERNEL
);
4771 if (args_k
== NULL
) return -ENOMEM
;
4779 * assume sz = 0 for command without parameters
4781 if (sz
&& copy_from_user(args_k
, arg
, sz
)) {
4782 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz
, arg
));
4787 * check if command supports extra parameters
4789 if (completed_args
== 0 && getsize
) {
4791 * get extra parameters size (based on main argument)
4793 ret
= (*getsize
)(args_k
, &xtra_sz
);
4794 if (ret
) goto error_args
;
4798 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz
, xtra_sz
));
4800 /* retry if necessary */
4801 if (likely(xtra_sz
)) goto restart_args
;
4804 if (unlikely((cmd_flags
& PFM_CMD_FD
) == 0)) goto skip_fd
;
4809 if (unlikely(f
.file
== NULL
)) {
4810 DPRINT(("invalid fd %d\n", fd
));
4813 if (unlikely(PFM_IS_FILE(f
.file
) == 0)) {
4814 DPRINT(("fd %d not related to perfmon\n", fd
));
4818 ctx
= f
.file
->private_data
;
4819 if (unlikely(ctx
== NULL
)) {
4820 DPRINT(("no context for fd %d\n", fd
));
4823 prefetch(&ctx
->ctx_state
);
4825 PROTECT_CTX(ctx
, flags
);
4828 * check task is stopped
4830 ret
= pfm_check_task_state(ctx
, cmd
, flags
);
4831 if (unlikely(ret
)) goto abort_locked
;
4834 ret
= (*func
)(ctx
, args_k
, count
, task_pt_regs(current
));
4840 DPRINT(("context unlocked\n"));
4841 UNPROTECT_CTX(ctx
, flags
);
4844 /* copy argument back to user, if needed */
4845 if (call_made
&& PFM_CMD_RW_ARG(cmd
) && copy_to_user(arg
, args_k
, base_sz
*count
)) ret
= -EFAULT
;
4853 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd
), ret
));
4859 pfm_resume_after_ovfl(pfm_context_t
*ctx
, unsigned long ovfl_regs
, struct pt_regs
*regs
)
4861 pfm_buffer_fmt_t
*fmt
= ctx
->ctx_buf_fmt
;
4862 pfm_ovfl_ctrl_t rst_ctrl
;
4866 state
= ctx
->ctx_state
;
4868 * Unlock sampling buffer and reset index atomically
4869 * XXX: not really needed when blocking
4871 if (CTX_HAS_SMPL(ctx
)) {
4873 rst_ctrl
.bits
.mask_monitoring
= 0;
4874 rst_ctrl
.bits
.reset_ovfl_pmds
= 0;
4876 if (state
== PFM_CTX_LOADED
)
4877 ret
= pfm_buf_fmt_restart_active(fmt
, current
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
4879 ret
= pfm_buf_fmt_restart(fmt
, current
, &rst_ctrl
, ctx
->ctx_smpl_hdr
, regs
);
4881 rst_ctrl
.bits
.mask_monitoring
= 0;
4882 rst_ctrl
.bits
.reset_ovfl_pmds
= 1;
4886 if (rst_ctrl
.bits
.reset_ovfl_pmds
) {
4887 pfm_reset_regs(ctx
, &ovfl_regs
, PFM_PMD_LONG_RESET
);
4889 if (rst_ctrl
.bits
.mask_monitoring
== 0) {
4890 DPRINT(("resuming monitoring\n"));
4891 if (ctx
->ctx_state
== PFM_CTX_MASKED
) pfm_restore_monitoring(current
);
4893 DPRINT(("stopping monitoring\n"));
4894 //pfm_stop_monitoring(current, regs);
4896 ctx
->ctx_state
= PFM_CTX_LOADED
;
4901 * context MUST BE LOCKED when calling
4902 * can only be called for current
4905 pfm_context_force_terminate(pfm_context_t
*ctx
, struct pt_regs
*regs
)
4909 DPRINT(("entering for [%d]\n", task_pid_nr(current
)));
4911 ret
= pfm_context_unload(ctx
, NULL
, 0, regs
);
4913 printk(KERN_ERR
"pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current
), ret
);
4917 * and wakeup controlling task, indicating we are now disconnected
4919 wake_up_interruptible(&ctx
->ctx_zombieq
);
4922 * given that context is still locked, the controlling
4923 * task will only get access when we return from
4924 * pfm_handle_work().
4928 static int pfm_ovfl_notify_user(pfm_context_t
*ctx
, unsigned long ovfl_pmds
);
4931 * pfm_handle_work() can be called with interrupts enabled
4932 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4933 * call may sleep, therefore we must re-enable interrupts
4934 * to avoid deadlocks. It is safe to do so because this function
4935 * is called ONLY when returning to user level (pUStk=1), in which case
4936 * there is no risk of kernel stack overflow due to deep
4937 * interrupt nesting.
4940 pfm_handle_work(void)
4943 struct pt_regs
*regs
;
4944 unsigned long flags
, dummy_flags
;
4945 unsigned long ovfl_regs
;
4946 unsigned int reason
;
4949 ctx
= PFM_GET_CTX(current
);
4951 printk(KERN_ERR
"perfmon: [%d] has no PFM context\n",
4952 task_pid_nr(current
));
4956 PROTECT_CTX(ctx
, flags
);
4958 PFM_SET_WORK_PENDING(current
, 0);
4960 regs
= task_pt_regs(current
);
4963 * extract reason for being here and clear
4965 reason
= ctx
->ctx_fl_trap_reason
;
4966 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_NONE
;
4967 ovfl_regs
= ctx
->ctx_ovfl_regs
[0];
4969 DPRINT(("reason=%d state=%d\n", reason
, ctx
->ctx_state
));
4972 * must be done before we check for simple-reset mode
4974 if (ctx
->ctx_fl_going_zombie
|| ctx
->ctx_state
== PFM_CTX_ZOMBIE
)
4977 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
4978 if (reason
== PFM_TRAP_REASON_RESET
)
4982 * restore interrupt mask to what it was on entry.
4983 * Could be enabled/diasbled.
4985 UNPROTECT_CTX(ctx
, flags
);
4988 * force interrupt enable because of down_interruptible()
4992 DPRINT(("before block sleeping\n"));
4995 * may go through without blocking on SMP systems
4996 * if restart has been received already by the time we call down()
4998 ret
= wait_for_completion_interruptible(&ctx
->ctx_restart_done
);
5000 DPRINT(("after block sleeping ret=%d\n", ret
));
5003 * lock context and mask interrupts again
5004 * We save flags into a dummy because we may have
5005 * altered interrupts mask compared to entry in this
5008 PROTECT_CTX(ctx
, dummy_flags
);
5011 * we need to read the ovfl_regs only after wake-up
5012 * because we may have had pfm_write_pmds() in between
5013 * and that can changed PMD values and therefore
5014 * ovfl_regs is reset for these new PMD values.
5016 ovfl_regs
= ctx
->ctx_ovfl_regs
[0];
5018 if (ctx
->ctx_fl_going_zombie
) {
5020 DPRINT(("context is zombie, bailing out\n"));
5021 pfm_context_force_terminate(ctx
, regs
);
5025 * in case of interruption of down() we don't restart anything
5031 pfm_resume_after_ovfl(ctx
, ovfl_regs
, regs
);
5032 ctx
->ctx_ovfl_regs
[0] = 0UL;
5036 * restore flags as they were upon entry
5038 UNPROTECT_CTX(ctx
, flags
);
5042 pfm_notify_user(pfm_context_t
*ctx
, pfm_msg_t
*msg
)
5044 if (ctx
->ctx_state
== PFM_CTX_ZOMBIE
) {
5045 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5049 DPRINT(("waking up somebody\n"));
5051 if (msg
) wake_up_interruptible(&ctx
->ctx_msgq_wait
);
5054 * safe, we are not in intr handler, nor in ctxsw when
5057 kill_fasync (&ctx
->ctx_async_queue
, SIGIO
, POLL_IN
);
5063 pfm_ovfl_notify_user(pfm_context_t
*ctx
, unsigned long ovfl_pmds
)
5065 pfm_msg_t
*msg
= NULL
;
5067 if (ctx
->ctx_fl_no_msg
== 0) {
5068 msg
= pfm_get_new_msg(ctx
);
5070 printk(KERN_ERR
"perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5074 msg
->pfm_ovfl_msg
.msg_type
= PFM_MSG_OVFL
;
5075 msg
->pfm_ovfl_msg
.msg_ctx_fd
= ctx
->ctx_fd
;
5076 msg
->pfm_ovfl_msg
.msg_active_set
= 0;
5077 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[0] = ovfl_pmds
;
5078 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[1] = 0UL;
5079 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[2] = 0UL;
5080 msg
->pfm_ovfl_msg
.msg_ovfl_pmds
[3] = 0UL;
5081 msg
->pfm_ovfl_msg
.msg_tstamp
= 0UL;
5084 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5090 return pfm_notify_user(ctx
, msg
);
5094 pfm_end_notify_user(pfm_context_t
*ctx
)
5098 msg
= pfm_get_new_msg(ctx
);
5100 printk(KERN_ERR
"perfmon: pfm_end_notify_user no more notification msgs\n");
5104 memset(msg
, 0, sizeof(*msg
));
5106 msg
->pfm_end_msg
.msg_type
= PFM_MSG_END
;
5107 msg
->pfm_end_msg
.msg_ctx_fd
= ctx
->ctx_fd
;
5108 msg
->pfm_ovfl_msg
.msg_tstamp
= 0UL;
5110 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5115 return pfm_notify_user(ctx
, msg
);
5119 * main overflow processing routine.
5120 * it can be called from the interrupt path or explicitly during the context switch code
5122 static void pfm_overflow_handler(struct task_struct
*task
, pfm_context_t
*ctx
,
5123 unsigned long pmc0
, struct pt_regs
*regs
)
5125 pfm_ovfl_arg_t
*ovfl_arg
;
5127 unsigned long old_val
, ovfl_val
, new_val
;
5128 unsigned long ovfl_notify
= 0UL, ovfl_pmds
= 0UL, smpl_pmds
= 0UL, reset_pmds
;
5129 unsigned long tstamp
;
5130 pfm_ovfl_ctrl_t ovfl_ctrl
;
5131 unsigned int i
, has_smpl
;
5132 int must_notify
= 0;
5134 if (unlikely(ctx
->ctx_state
== PFM_CTX_ZOMBIE
)) goto stop_monitoring
;
5137 * sanity test. Should never happen
5139 if (unlikely((pmc0
& 0x1) == 0)) goto sanity_check
;
5141 tstamp
= ia64_get_itc();
5142 mask
= pmc0
>> PMU_FIRST_COUNTER
;
5143 ovfl_val
= pmu_conf
->ovfl_val
;
5144 has_smpl
= CTX_HAS_SMPL(ctx
);
5146 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5147 "used_pmds=0x%lx\n",
5149 task
? task_pid_nr(task
): -1,
5150 (regs
? regs
->cr_iip
: 0),
5151 CTX_OVFL_NOBLOCK(ctx
) ? "nonblocking" : "blocking",
5152 ctx
->ctx_used_pmds
[0]));
5156 * first we update the virtual counters
5157 * assume there was a prior ia64_srlz_d() issued
5159 for (i
= PMU_FIRST_COUNTER
; mask
; i
++, mask
>>= 1) {
5161 /* skip pmd which did not overflow */
5162 if ((mask
& 0x1) == 0) continue;
5165 * Note that the pmd is not necessarily 0 at this point as qualified events
5166 * may have happened before the PMU was frozen. The residual count is not
5167 * taken into consideration here but will be with any read of the pmd via
5170 old_val
= new_val
= ctx
->ctx_pmds
[i
].val
;
5171 new_val
+= 1 + ovfl_val
;
5172 ctx
->ctx_pmds
[i
].val
= new_val
;
5175 * check for overflow condition
5177 if (likely(old_val
> new_val
)) {
5178 ovfl_pmds
|= 1UL << i
;
5179 if (PMC_OVFL_NOTIFY(ctx
, i
)) ovfl_notify
|= 1UL << i
;
5182 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5186 ia64_get_pmd(i
) & ovfl_val
,
5192 * there was no 64-bit overflow, nothing else to do
5194 if (ovfl_pmds
== 0UL) return;
5197 * reset all control bits
5203 * if a sampling format module exists, then we "cache" the overflow by
5204 * calling the module's handler() routine.
5207 unsigned long start_cycles
, end_cycles
;
5208 unsigned long pmd_mask
;
5210 int this_cpu
= smp_processor_id();
5212 pmd_mask
= ovfl_pmds
>> PMU_FIRST_COUNTER
;
5213 ovfl_arg
= &ctx
->ctx_ovfl_arg
;
5215 prefetch(ctx
->ctx_smpl_hdr
);
5217 for(i
=PMU_FIRST_COUNTER
; pmd_mask
&& ret
== 0; i
++, pmd_mask
>>=1) {
5221 if ((pmd_mask
& 0x1) == 0) continue;
5223 ovfl_arg
->ovfl_pmd
= (unsigned char )i
;
5224 ovfl_arg
->ovfl_notify
= ovfl_notify
& mask
? 1 : 0;
5225 ovfl_arg
->active_set
= 0;
5226 ovfl_arg
->ovfl_ctrl
.val
= 0; /* module must fill in all fields */
5227 ovfl_arg
->smpl_pmds
[0] = smpl_pmds
= ctx
->ctx_pmds
[i
].smpl_pmds
[0];
5229 ovfl_arg
->pmd_value
= ctx
->ctx_pmds
[i
].val
;
5230 ovfl_arg
->pmd_last_reset
= ctx
->ctx_pmds
[i
].lval
;
5231 ovfl_arg
->pmd_eventid
= ctx
->ctx_pmds
[i
].eventid
;
5234 * copy values of pmds of interest. Sampling format may copy them
5235 * into sampling buffer.
5238 for(j
=0, k
=0; smpl_pmds
; j
++, smpl_pmds
>>=1) {
5239 if ((smpl_pmds
& 0x1) == 0) continue;
5240 ovfl_arg
->smpl_pmds_values
[k
++] = PMD_IS_COUNTING(j
) ? pfm_read_soft_counter(ctx
, j
) : ia64_get_pmd(j
);
5241 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k
-1, j
, ovfl_arg
->smpl_pmds_values
[k
-1]));
5245 pfm_stats
[this_cpu
].pfm_smpl_handler_calls
++;
5247 start_cycles
= ia64_get_itc();
5250 * call custom buffer format record (handler) routine
5252 ret
= (*ctx
->ctx_buf_fmt
->fmt_handler
)(task
, ctx
->ctx_smpl_hdr
, ovfl_arg
, regs
, tstamp
);
5254 end_cycles
= ia64_get_itc();
5257 * For those controls, we take the union because they have
5258 * an all or nothing behavior.
5260 ovfl_ctrl
.bits
.notify_user
|= ovfl_arg
->ovfl_ctrl
.bits
.notify_user
;
5261 ovfl_ctrl
.bits
.block_task
|= ovfl_arg
->ovfl_ctrl
.bits
.block_task
;
5262 ovfl_ctrl
.bits
.mask_monitoring
|= ovfl_arg
->ovfl_ctrl
.bits
.mask_monitoring
;
5264 * build the bitmask of pmds to reset now
5266 if (ovfl_arg
->ovfl_ctrl
.bits
.reset_ovfl_pmds
) reset_pmds
|= mask
;
5268 pfm_stats
[this_cpu
].pfm_smpl_handler_cycles
+= end_cycles
- start_cycles
;
5271 * when the module cannot handle the rest of the overflows, we abort right here
5273 if (ret
&& pmd_mask
) {
5274 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5275 pmd_mask
<<PMU_FIRST_COUNTER
));
5278 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5280 ovfl_pmds
&= ~reset_pmds
;
5283 * when no sampling module is used, then the default
5284 * is to notify on overflow if requested by user
5286 ovfl_ctrl
.bits
.notify_user
= ovfl_notify
? 1 : 0;
5287 ovfl_ctrl
.bits
.block_task
= ovfl_notify
? 1 : 0;
5288 ovfl_ctrl
.bits
.mask_monitoring
= ovfl_notify
? 1 : 0; /* XXX: change for saturation */
5289 ovfl_ctrl
.bits
.reset_ovfl_pmds
= ovfl_notify
? 0 : 1;
5291 * if needed, we reset all overflowed pmds
5293 if (ovfl_notify
== 0) reset_pmds
= ovfl_pmds
;
5296 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds
, reset_pmds
));
5299 * reset the requested PMD registers using the short reset values
5302 unsigned long bm
= reset_pmds
;
5303 pfm_reset_regs(ctx
, &bm
, PFM_PMD_SHORT_RESET
);
5306 if (ovfl_notify
&& ovfl_ctrl
.bits
.notify_user
) {
5308 * keep track of what to reset when unblocking
5310 ctx
->ctx_ovfl_regs
[0] = ovfl_pmds
;
5313 * check for blocking context
5315 if (CTX_OVFL_NOBLOCK(ctx
) == 0 && ovfl_ctrl
.bits
.block_task
) {
5317 ctx
->ctx_fl_trap_reason
= PFM_TRAP_REASON_BLOCK
;
5320 * set the perfmon specific checking pending work for the task
5322 PFM_SET_WORK_PENDING(task
, 1);
5325 * when coming from ctxsw, current still points to the
5326 * previous task, therefore we must work with task and not current.
5328 set_notify_resume(task
);
5331 * defer until state is changed (shorten spin window). the context is locked
5332 * anyway, so the signal receiver would come spin for nothing.
5337 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5338 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5339 PFM_GET_WORK_PENDING(task
),
5340 ctx
->ctx_fl_trap_reason
,
5343 ovfl_ctrl
.bits
.mask_monitoring
? 1 : 0));
5345 * in case monitoring must be stopped, we toggle the psr bits
5347 if (ovfl_ctrl
.bits
.mask_monitoring
) {
5348 pfm_mask_monitoring(task
);
5349 ctx
->ctx_state
= PFM_CTX_MASKED
;
5350 ctx
->ctx_fl_can_restart
= 1;
5354 * send notification now
5356 if (must_notify
) pfm_ovfl_notify_user(ctx
, ovfl_notify
);
5361 printk(KERN_ERR
"perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5363 task
? task_pid_nr(task
) : -1,
5369 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5370 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5371 * come here as zombie only if the task is the current task. In which case, we
5372 * can access the PMU hardware directly.
5374 * Note that zombies do have PM_VALID set. So here we do the minimal.
5376 * In case the context was zombified it could not be reclaimed at the time
5377 * the monitoring program exited. At this point, the PMU reservation has been
5378 * returned, the sampiing buffer has been freed. We must convert this call
5379 * into a spurious interrupt. However, we must also avoid infinite overflows
5380 * by stopping monitoring for this task. We can only come here for a per-task
5381 * context. All we need to do is to stop monitoring using the psr bits which
5382 * are always task private. By re-enabling secure montioring, we ensure that
5383 * the monitored task will not be able to re-activate monitoring.
5384 * The task will eventually be context switched out, at which point the context
5385 * will be reclaimed (that includes releasing ownership of the PMU).
5387 * So there might be a window of time where the number of per-task session is zero
5388 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5389 * context. This is safe because if a per-task session comes in, it will push this one
5390 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5391 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5392 * also push our zombie context out.
5394 * Overall pretty hairy stuff....
5396 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task
? task_pid_nr(task
): -1));
5398 ia64_psr(regs
)->up
= 0;
5399 ia64_psr(regs
)->sp
= 1;
5404 pfm_do_interrupt_handler(void *arg
, struct pt_regs
*regs
)
5406 struct task_struct
*task
;
5408 unsigned long flags
;
5410 int this_cpu
= smp_processor_id();
5413 pfm_stats
[this_cpu
].pfm_ovfl_intr_count
++;
5416 * srlz.d done before arriving here
5418 pmc0
= ia64_get_pmc(0);
5420 task
= GET_PMU_OWNER();
5421 ctx
= GET_PMU_CTX();
5424 * if we have some pending bits set
5425 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5427 if (PMC0_HAS_OVFL(pmc0
) && task
) {
5429 * we assume that pmc0.fr is always set here
5433 if (!ctx
) goto report_spurious1
;
5435 if (ctx
->ctx_fl_system
== 0 && (task
->thread
.flags
& IA64_THREAD_PM_VALID
) == 0)
5436 goto report_spurious2
;
5438 PROTECT_CTX_NOPRINT(ctx
, flags
);
5440 pfm_overflow_handler(task
, ctx
, pmc0
, regs
);
5442 UNPROTECT_CTX_NOPRINT(ctx
, flags
);
5445 pfm_stats
[this_cpu
].pfm_spurious_ovfl_intr_count
++;
5449 * keep it unfrozen at all times
5456 printk(KERN_INFO
"perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5457 this_cpu
, task_pid_nr(task
));
5461 printk(KERN_INFO
"perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5469 pfm_interrupt_handler(int irq
, void *arg
)
5471 unsigned long start_cycles
, total_cycles
;
5472 unsigned long min
, max
;
5475 struct pt_regs
*regs
= get_irq_regs();
5477 this_cpu
= get_cpu();
5478 if (likely(!pfm_alt_intr_handler
)) {
5479 min
= pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_min
;
5480 max
= pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_max
;
5482 start_cycles
= ia64_get_itc();
5484 ret
= pfm_do_interrupt_handler(arg
, regs
);
5486 total_cycles
= ia64_get_itc();
5489 * don't measure spurious interrupts
5491 if (likely(ret
== 0)) {
5492 total_cycles
-= start_cycles
;
5494 if (total_cycles
< min
) pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_min
= total_cycles
;
5495 if (total_cycles
> max
) pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles_max
= total_cycles
;
5497 pfm_stats
[this_cpu
].pfm_ovfl_intr_cycles
+= total_cycles
;
5501 (*pfm_alt_intr_handler
->handler
)(irq
, arg
, regs
);
5509 * /proc/perfmon interface, for debug only
5512 #define PFM_PROC_SHOW_HEADER ((void *)(long)nr_cpu_ids+1)
5515 pfm_proc_start(struct seq_file
*m
, loff_t
*pos
)
5518 return PFM_PROC_SHOW_HEADER
;
5521 while (*pos
<= nr_cpu_ids
) {
5522 if (cpu_online(*pos
- 1)) {
5523 return (void *)*pos
;
5531 pfm_proc_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
5534 return pfm_proc_start(m
, pos
);
5538 pfm_proc_stop(struct seq_file
*m
, void *v
)
5543 pfm_proc_show_header(struct seq_file
*m
)
5545 struct list_head
* pos
;
5546 pfm_buffer_fmt_t
* entry
;
5547 unsigned long flags
;
5550 "perfmon version : %u.%u\n"
5553 "expert mode : %s\n"
5554 "ovfl_mask : 0x%lx\n"
5555 "PMU flags : 0x%x\n",
5556 PFM_VERSION_MAJ
, PFM_VERSION_MIN
,
5558 pfm_sysctl
.fastctxsw
> 0 ? "Yes": "No",
5559 pfm_sysctl
.expert_mode
> 0 ? "Yes": "No",
5566 "proc_sessions : %u\n"
5567 "sys_sessions : %u\n"
5568 "sys_use_dbregs : %u\n"
5569 "ptrace_use_dbregs : %u\n",
5570 pfm_sessions
.pfs_task_sessions
,
5571 pfm_sessions
.pfs_sys_sessions
,
5572 pfm_sessions
.pfs_sys_use_dbregs
,
5573 pfm_sessions
.pfs_ptrace_use_dbregs
);
5577 spin_lock(&pfm_buffer_fmt_lock
);
5579 list_for_each(pos
, &pfm_buffer_fmt_list
) {
5580 entry
= list_entry(pos
, pfm_buffer_fmt_t
, fmt_list
);
5581 seq_printf(m
, "format : %16phD %s\n",
5582 entry
->fmt_uuid
, entry
->fmt_name
);
5584 spin_unlock(&pfm_buffer_fmt_lock
);
5589 pfm_proc_show(struct seq_file
*m
, void *v
)
5595 if (v
== PFM_PROC_SHOW_HEADER
) {
5596 pfm_proc_show_header(m
);
5600 /* show info for CPU (v - 1) */
5604 "CPU%-2d overflow intrs : %lu\n"
5605 "CPU%-2d overflow cycles : %lu\n"
5606 "CPU%-2d overflow min : %lu\n"
5607 "CPU%-2d overflow max : %lu\n"
5608 "CPU%-2d smpl handler calls : %lu\n"
5609 "CPU%-2d smpl handler cycles : %lu\n"
5610 "CPU%-2d spurious intrs : %lu\n"
5611 "CPU%-2d replay intrs : %lu\n"
5612 "CPU%-2d syst_wide : %d\n"
5613 "CPU%-2d dcr_pp : %d\n"
5614 "CPU%-2d exclude idle : %d\n"
5615 "CPU%-2d owner : %d\n"
5616 "CPU%-2d context : %p\n"
5617 "CPU%-2d activations : %lu\n",
5618 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_count
,
5619 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles
,
5620 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles_min
,
5621 cpu
, pfm_stats
[cpu
].pfm_ovfl_intr_cycles_max
,
5622 cpu
, pfm_stats
[cpu
].pfm_smpl_handler_calls
,
5623 cpu
, pfm_stats
[cpu
].pfm_smpl_handler_cycles
,
5624 cpu
, pfm_stats
[cpu
].pfm_spurious_ovfl_intr_count
,
5625 cpu
, pfm_stats
[cpu
].pfm_replay_ovfl_intr_count
,
5626 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_SYST_WIDE
? 1 : 0,
5627 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_DCR_PP
? 1 : 0,
5628 cpu
, pfm_get_cpu_data(pfm_syst_info
, cpu
) & PFM_CPUINFO_EXCL_IDLE
? 1 : 0,
5629 cpu
, pfm_get_cpu_data(pmu_owner
, cpu
) ? pfm_get_cpu_data(pmu_owner
, cpu
)->pid
: -1,
5630 cpu
, pfm_get_cpu_data(pmu_ctx
, cpu
),
5631 cpu
, pfm_get_cpu_data(pmu_activation_number
, cpu
));
5633 if (num_online_cpus() == 1 && pfm_sysctl
.debug
> 0) {
5635 psr
= pfm_get_psr();
5640 "CPU%-2d psr : 0x%lx\n"
5641 "CPU%-2d pmc0 : 0x%lx\n",
5643 cpu
, ia64_get_pmc(0));
5645 for (i
=0; PMC_IS_LAST(i
) == 0; i
++) {
5646 if (PMC_IS_COUNTING(i
) == 0) continue;
5648 "CPU%-2d pmc%u : 0x%lx\n"
5649 "CPU%-2d pmd%u : 0x%lx\n",
5650 cpu
, i
, ia64_get_pmc(i
),
5651 cpu
, i
, ia64_get_pmd(i
));
5657 const struct seq_operations pfm_seq_ops
= {
5658 .start
= pfm_proc_start
,
5659 .next
= pfm_proc_next
,
5660 .stop
= pfm_proc_stop
,
5661 .show
= pfm_proc_show
5665 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5666 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5667 * is active or inactive based on mode. We must rely on the value in
5668 * local_cpu_data->pfm_syst_info
5671 pfm_syst_wide_update_task(struct task_struct
*task
, unsigned long info
, int is_ctxswin
)
5673 struct pt_regs
*regs
;
5675 unsigned long dcr_pp
;
5677 dcr_pp
= info
& PFM_CPUINFO_DCR_PP
? 1 : 0;
5680 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5681 * on every CPU, so we can rely on the pid to identify the idle task.
5683 if ((info
& PFM_CPUINFO_EXCL_IDLE
) == 0 || task
->pid
) {
5684 regs
= task_pt_regs(task
);
5685 ia64_psr(regs
)->pp
= is_ctxswin
? dcr_pp
: 0;
5689 * if monitoring has started
5692 dcr
= ia64_getreg(_IA64_REG_CR_DCR
);
5694 * context switching in?
5697 /* mask monitoring for the idle task */
5698 ia64_setreg(_IA64_REG_CR_DCR
, dcr
& ~IA64_DCR_PP
);
5704 * context switching out
5705 * restore monitoring for next task
5707 * Due to inlining this odd if-then-else construction generates
5710 ia64_setreg(_IA64_REG_CR_DCR
, dcr
|IA64_DCR_PP
);
5719 pfm_force_cleanup(pfm_context_t
*ctx
, struct pt_regs
*regs
)
5721 struct task_struct
*task
= ctx
->ctx_task
;
5723 ia64_psr(regs
)->up
= 0;
5724 ia64_psr(regs
)->sp
= 1;
5726 if (GET_PMU_OWNER() == task
) {
5727 DPRINT(("cleared ownership for [%d]\n",
5728 task_pid_nr(ctx
->ctx_task
)));
5729 SET_PMU_OWNER(NULL
, NULL
);
5733 * disconnect the task from the context and vice-versa
5735 PFM_SET_WORK_PENDING(task
, 0);
5737 task
->thread
.pfm_context
= NULL
;
5738 task
->thread
.flags
&= ~IA64_THREAD_PM_VALID
;
5740 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task
)));
5745 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5748 pfm_save_regs(struct task_struct
*task
)
5751 unsigned long flags
;
5755 ctx
= PFM_GET_CTX(task
);
5756 if (ctx
== NULL
) return;
5759 * we always come here with interrupts ALREADY disabled by
5760 * the scheduler. So we simply need to protect against concurrent
5761 * access, not CPU concurrency.
5763 flags
= pfm_protect_ctx_ctxsw(ctx
);
5765 if (ctx
->ctx_state
== PFM_CTX_ZOMBIE
) {
5766 struct pt_regs
*regs
= task_pt_regs(task
);
5770 pfm_force_cleanup(ctx
, regs
);
5772 BUG_ON(ctx
->ctx_smpl_hdr
);
5774 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
5776 pfm_context_free(ctx
);
5781 * save current PSR: needed because we modify it
5784 psr
= pfm_get_psr();
5786 BUG_ON(psr
& (IA64_PSR_I
));
5790 * This is the last instruction which may generate an overflow
5792 * We do not need to set psr.sp because, it is irrelevant in kernel.
5793 * It will be restored from ipsr when going back to user level
5798 * keep a copy of psr.up (for reload)
5800 ctx
->ctx_saved_psr_up
= psr
& IA64_PSR_UP
;
5803 * release ownership of this PMU.
5804 * PM interrupts are masked, so nothing
5807 SET_PMU_OWNER(NULL
, NULL
);
5810 * we systematically save the PMD as we have no
5811 * guarantee we will be schedule at that same
5814 pfm_save_pmds(ctx
->th_pmds
, ctx
->ctx_used_pmds
[0]);
5817 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5818 * we will need it on the restore path to check
5819 * for pending overflow.
5821 ctx
->th_pmcs
[0] = ia64_get_pmc(0);
5824 * unfreeze PMU if had pending overflows
5826 if (ctx
->th_pmcs
[0] & ~0x1UL
) pfm_unfreeze_pmu();
5829 * finally, allow context access.
5830 * interrupts will still be masked after this call.
5832 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
5835 #else /* !CONFIG_SMP */
5837 pfm_save_regs(struct task_struct
*task
)
5842 ctx
= PFM_GET_CTX(task
);
5843 if (ctx
== NULL
) return;
5846 * save current PSR: needed because we modify it
5848 psr
= pfm_get_psr();
5850 BUG_ON(psr
& (IA64_PSR_I
));
5854 * This is the last instruction which may generate an overflow
5856 * We do not need to set psr.sp because, it is irrelevant in kernel.
5857 * It will be restored from ipsr when going back to user level
5862 * keep a copy of psr.up (for reload)
5864 ctx
->ctx_saved_psr_up
= psr
& IA64_PSR_UP
;
5868 pfm_lazy_save_regs (struct task_struct
*task
)
5871 unsigned long flags
;
5873 { u64 psr
= pfm_get_psr();
5874 BUG_ON(psr
& IA64_PSR_UP
);
5877 ctx
= PFM_GET_CTX(task
);
5880 * we need to mask PMU overflow here to
5881 * make sure that we maintain pmc0 until
5882 * we save it. overflow interrupts are
5883 * treated as spurious if there is no
5886 * XXX: I don't think this is necessary
5888 PROTECT_CTX(ctx
,flags
);
5891 * release ownership of this PMU.
5892 * must be done before we save the registers.
5894 * after this call any PMU interrupt is treated
5897 SET_PMU_OWNER(NULL
, NULL
);
5900 * save all the pmds we use
5902 pfm_save_pmds(ctx
->th_pmds
, ctx
->ctx_used_pmds
[0]);
5905 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5906 * it is needed to check for pended overflow
5907 * on the restore path
5909 ctx
->th_pmcs
[0] = ia64_get_pmc(0);
5912 * unfreeze PMU if had pending overflows
5914 if (ctx
->th_pmcs
[0] & ~0x1UL
) pfm_unfreeze_pmu();
5917 * now get can unmask PMU interrupts, they will
5918 * be treated as purely spurious and we will not
5919 * lose any information
5921 UNPROTECT_CTX(ctx
,flags
);
5923 #endif /* CONFIG_SMP */
5927 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5930 pfm_load_regs (struct task_struct
*task
)
5933 unsigned long pmc_mask
= 0UL, pmd_mask
= 0UL;
5934 unsigned long flags
;
5936 int need_irq_resend
;
5938 ctx
= PFM_GET_CTX(task
);
5939 if (unlikely(ctx
== NULL
)) return;
5941 BUG_ON(GET_PMU_OWNER());
5944 * possible on unload
5946 if (unlikely((task
->thread
.flags
& IA64_THREAD_PM_VALID
) == 0)) return;
5949 * we always come here with interrupts ALREADY disabled by
5950 * the scheduler. So we simply need to protect against concurrent
5951 * access, not CPU concurrency.
5953 flags
= pfm_protect_ctx_ctxsw(ctx
);
5954 psr
= pfm_get_psr();
5956 need_irq_resend
= pmu_conf
->flags
& PFM_PMU_IRQ_RESEND
;
5958 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
5959 BUG_ON(psr
& IA64_PSR_I
);
5961 if (unlikely(ctx
->ctx_state
== PFM_CTX_ZOMBIE
)) {
5962 struct pt_regs
*regs
= task_pt_regs(task
);
5964 BUG_ON(ctx
->ctx_smpl_hdr
);
5966 pfm_force_cleanup(ctx
, regs
);
5968 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
5971 * this one (kmalloc'ed) is fine with interrupts disabled
5973 pfm_context_free(ctx
);
5979 * we restore ALL the debug registers to avoid picking up
5982 if (ctx
->ctx_fl_using_dbreg
) {
5983 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
5984 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
5987 * retrieve saved psr.up
5989 psr_up
= ctx
->ctx_saved_psr_up
;
5992 * if we were the last user of the PMU on that CPU,
5993 * then nothing to do except restore psr
5995 if (GET_LAST_CPU(ctx
) == smp_processor_id() && ctx
->ctx_last_activation
== GET_ACTIVATION()) {
5998 * retrieve partial reload masks (due to user modifications)
6000 pmc_mask
= ctx
->ctx_reload_pmcs
[0];
6001 pmd_mask
= ctx
->ctx_reload_pmds
[0];
6005 * To avoid leaking information to the user level when psr.sp=0,
6006 * we must reload ALL implemented pmds (even the ones we don't use).
6007 * In the kernel we only allow PFM_READ_PMDS on registers which
6008 * we initialized or requested (sampling) so there is no risk there.
6010 pmd_mask
= pfm_sysctl
.fastctxsw
? ctx
->ctx_used_pmds
[0] : ctx
->ctx_all_pmds
[0];
6013 * ALL accessible PMCs are systematically reloaded, unused registers
6014 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6015 * up stale configuration.
6017 * PMC0 is never in the mask. It is always restored separately.
6019 pmc_mask
= ctx
->ctx_all_pmcs
[0];
6022 * when context is MASKED, we will restore PMC with plm=0
6023 * and PMD with stale information, but that's ok, nothing
6026 * XXX: optimize here
6028 if (pmd_mask
) pfm_restore_pmds(ctx
->th_pmds
, pmd_mask
);
6029 if (pmc_mask
) pfm_restore_pmcs(ctx
->th_pmcs
, pmc_mask
);
6032 * check for pending overflow at the time the state
6035 if (unlikely(PMC0_HAS_OVFL(ctx
->th_pmcs
[0]))) {
6037 * reload pmc0 with the overflow information
6038 * On McKinley PMU, this will trigger a PMU interrupt
6040 ia64_set_pmc(0, ctx
->th_pmcs
[0]);
6042 ctx
->th_pmcs
[0] = 0UL;
6045 * will replay the PMU interrupt
6047 if (need_irq_resend
) ia64_resend_irq(IA64_PERFMON_VECTOR
);
6049 pfm_stats
[smp_processor_id()].pfm_replay_ovfl_intr_count
++;
6053 * we just did a reload, so we reset the partial reload fields
6055 ctx
->ctx_reload_pmcs
[0] = 0UL;
6056 ctx
->ctx_reload_pmds
[0] = 0UL;
6058 SET_LAST_CPU(ctx
, smp_processor_id());
6061 * dump activation value for this PMU
6065 * record current activation for this context
6067 SET_ACTIVATION(ctx
);
6070 * establish new ownership.
6072 SET_PMU_OWNER(task
, ctx
);
6075 * restore the psr.up bit. measurement
6077 * no PMU interrupt can happen at this point
6078 * because we still have interrupts disabled.
6080 if (likely(psr_up
)) pfm_set_psr_up();
6083 * allow concurrent access to context
6085 pfm_unprotect_ctx_ctxsw(ctx
, flags
);
6087 #else /* !CONFIG_SMP */
6089 * reload PMU state for UP kernels
6090 * in 2.5 we come here with interrupts disabled
6093 pfm_load_regs (struct task_struct
*task
)
6096 struct task_struct
*owner
;
6097 unsigned long pmd_mask
, pmc_mask
;
6099 int need_irq_resend
;
6101 owner
= GET_PMU_OWNER();
6102 ctx
= PFM_GET_CTX(task
);
6103 psr
= pfm_get_psr();
6105 BUG_ON(psr
& (IA64_PSR_UP
|IA64_PSR_PP
));
6106 BUG_ON(psr
& IA64_PSR_I
);
6109 * we restore ALL the debug registers to avoid picking up
6112 * This must be done even when the task is still the owner
6113 * as the registers may have been modified via ptrace()
6114 * (not perfmon) by the previous task.
6116 if (ctx
->ctx_fl_using_dbreg
) {
6117 pfm_restore_ibrs(ctx
->ctx_ibrs
, pmu_conf
->num_ibrs
);
6118 pfm_restore_dbrs(ctx
->ctx_dbrs
, pmu_conf
->num_dbrs
);
6122 * retrieved saved psr.up
6124 psr_up
= ctx
->ctx_saved_psr_up
;
6125 need_irq_resend
= pmu_conf
->flags
& PFM_PMU_IRQ_RESEND
;
6128 * short path, our state is still there, just
6129 * need to restore psr and we go
6131 * we do not touch either PMC nor PMD. the psr is not touched
6132 * by the overflow_handler. So we are safe w.r.t. to interrupt
6133 * concurrency even without interrupt masking.
6135 if (likely(owner
== task
)) {
6136 if (likely(psr_up
)) pfm_set_psr_up();
6141 * someone else is still using the PMU, first push it out and
6142 * then we'll be able to install our stuff !
6144 * Upon return, there will be no owner for the current PMU
6146 if (owner
) pfm_lazy_save_regs(owner
);
6149 * To avoid leaking information to the user level when psr.sp=0,
6150 * we must reload ALL implemented pmds (even the ones we don't use).
6151 * In the kernel we only allow PFM_READ_PMDS on registers which
6152 * we initialized or requested (sampling) so there is no risk there.
6154 pmd_mask
= pfm_sysctl
.fastctxsw
? ctx
->ctx_used_pmds
[0] : ctx
->ctx_all_pmds
[0];
6157 * ALL accessible PMCs are systematically reloaded, unused registers
6158 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6159 * up stale configuration.
6161 * PMC0 is never in the mask. It is always restored separately
6163 pmc_mask
= ctx
->ctx_all_pmcs
[0];
6165 pfm_restore_pmds(ctx
->th_pmds
, pmd_mask
);
6166 pfm_restore_pmcs(ctx
->th_pmcs
, pmc_mask
);
6169 * check for pending overflow at the time the state
6172 if (unlikely(PMC0_HAS_OVFL(ctx
->th_pmcs
[0]))) {
6174 * reload pmc0 with the overflow information
6175 * On McKinley PMU, this will trigger a PMU interrupt
6177 ia64_set_pmc(0, ctx
->th_pmcs
[0]);
6180 ctx
->th_pmcs
[0] = 0UL;
6183 * will replay the PMU interrupt
6185 if (need_irq_resend
) ia64_resend_irq(IA64_PERFMON_VECTOR
);
6187 pfm_stats
[smp_processor_id()].pfm_replay_ovfl_intr_count
++;
6191 * establish new ownership.
6193 SET_PMU_OWNER(task
, ctx
);
6196 * restore the psr.up bit. measurement
6198 * no PMU interrupt can happen at this point
6199 * because we still have interrupts disabled.
6201 if (likely(psr_up
)) pfm_set_psr_up();
6203 #endif /* CONFIG_SMP */
6206 * this function assumes monitoring is stopped
6209 pfm_flush_pmds(struct task_struct
*task
, pfm_context_t
*ctx
)
6212 unsigned long mask2
, val
, pmd_val
, ovfl_val
;
6213 int i
, can_access_pmu
= 0;
6217 * is the caller the task being monitored (or which initiated the
6218 * session for system wide measurements)
6220 is_self
= ctx
->ctx_task
== task
? 1 : 0;
6223 * can access PMU is task is the owner of the PMU state on the current CPU
6224 * or if we are running on the CPU bound to the context in system-wide mode
6225 * (that is not necessarily the task the context is attached to in this mode).
6226 * In system-wide we always have can_access_pmu true because a task running on an
6227 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6229 can_access_pmu
= (GET_PMU_OWNER() == task
) || (ctx
->ctx_fl_system
&& ctx
->ctx_cpu
== smp_processor_id());
6230 if (can_access_pmu
) {
6232 * Mark the PMU as not owned
6233 * This will cause the interrupt handler to do nothing in case an overflow
6234 * interrupt was in-flight
6235 * This also guarantees that pmc0 will contain the final state
6236 * It virtually gives us full control on overflow processing from that point
6239 SET_PMU_OWNER(NULL
, NULL
);
6240 DPRINT(("releasing ownership\n"));
6243 * read current overflow status:
6245 * we are guaranteed to read the final stable state
6248 pmc0
= ia64_get_pmc(0); /* slow */
6251 * reset freeze bit, overflow status information destroyed
6255 pmc0
= ctx
->th_pmcs
[0];
6257 * clear whatever overflow status bits there were
6259 ctx
->th_pmcs
[0] = 0;
6261 ovfl_val
= pmu_conf
->ovfl_val
;
6263 * we save all the used pmds
6264 * we take care of overflows for counting PMDs
6266 * XXX: sampling situation is not taken into account here
6268 mask2
= ctx
->ctx_used_pmds
[0];
6270 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self
, ovfl_val
, mask2
));
6272 for (i
= 0; mask2
; i
++, mask2
>>=1) {
6274 /* skip non used pmds */
6275 if ((mask2
& 0x1) == 0) continue;
6278 * can access PMU always true in system wide mode
6280 val
= pmd_val
= can_access_pmu
? ia64_get_pmd(i
) : ctx
->th_pmds
[i
];
6282 if (PMD_IS_COUNTING(i
)) {
6283 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6286 ctx
->ctx_pmds
[i
].val
,
6290 * we rebuild the full 64 bit value of the counter
6292 val
= ctx
->ctx_pmds
[i
].val
+ (val
& ovfl_val
);
6295 * now everything is in ctx_pmds[] and we need
6296 * to clear the saved context from save_regs() such that
6297 * pfm_read_pmds() gets the correct value
6302 * take care of overflow inline
6304 if (pmc0
& (1UL << i
)) {
6305 val
+= 1 + ovfl_val
;
6306 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task
), i
));
6310 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task
), i
, val
, pmd_val
));
6312 if (is_self
) ctx
->th_pmds
[i
] = pmd_val
;
6314 ctx
->ctx_pmds
[i
].val
= val
;
6319 pfm_alt_save_pmu_state(void *data
)
6321 struct pt_regs
*regs
;
6323 regs
= task_pt_regs(current
);
6325 DPRINT(("called\n"));
6328 * should not be necessary but
6329 * let's take not risk
6333 ia64_psr(regs
)->pp
= 0;
6336 * This call is required
6337 * May cause a spurious interrupt on some processors
6345 pfm_alt_restore_pmu_state(void *data
)
6347 struct pt_regs
*regs
;
6349 regs
= task_pt_regs(current
);
6351 DPRINT(("called\n"));
6354 * put PMU back in state expected
6359 ia64_psr(regs
)->pp
= 0;
6362 * perfmon runs with PMU unfrozen at all times
6370 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t
*hdl
)
6375 /* some sanity checks */
6376 if (hdl
== NULL
|| hdl
->handler
== NULL
) return -EINVAL
;
6378 /* do the easy test first */
6379 if (pfm_alt_intr_handler
) return -EBUSY
;
6381 /* one at a time in the install or remove, just fail the others */
6382 if (!spin_trylock(&pfm_alt_install_check
)) {
6386 /* reserve our session */
6387 for_each_online_cpu(reserve_cpu
) {
6388 ret
= pfm_reserve_session(NULL
, 1, reserve_cpu
);
6389 if (ret
) goto cleanup_reserve
;
6392 /* save the current system wide pmu states */
6393 on_each_cpu(pfm_alt_save_pmu_state
, NULL
, 1);
6395 /* officially change to the alternate interrupt handler */
6396 pfm_alt_intr_handler
= hdl
;
6398 spin_unlock(&pfm_alt_install_check
);
6403 for_each_online_cpu(i
) {
6404 /* don't unreserve more than we reserved */
6405 if (i
>= reserve_cpu
) break;
6407 pfm_unreserve_session(NULL
, 1, i
);
6410 spin_unlock(&pfm_alt_install_check
);
6414 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt
);
6417 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t
*hdl
)
6421 if (hdl
== NULL
) return -EINVAL
;
6423 /* cannot remove someone else's handler! */
6424 if (pfm_alt_intr_handler
!= hdl
) return -EINVAL
;
6426 /* one at a time in the install or remove, just fail the others */
6427 if (!spin_trylock(&pfm_alt_install_check
)) {
6431 pfm_alt_intr_handler
= NULL
;
6433 on_each_cpu(pfm_alt_restore_pmu_state
, NULL
, 1);
6435 for_each_online_cpu(i
) {
6436 pfm_unreserve_session(NULL
, 1, i
);
6439 spin_unlock(&pfm_alt_install_check
);
6443 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt
);
6446 * perfmon initialization routine, called from the initcall() table
6448 static int init_pfm_fs(void);
6456 family
= local_cpu_data
->family
;
6461 if ((*p
)->probe() == 0) goto found
;
6462 } else if ((*p
)->pmu_family
== family
|| (*p
)->pmu_family
== 0xff) {
6476 unsigned int n
, n_counters
, i
;
6478 printk("perfmon: version %u.%u IRQ %u\n",
6481 IA64_PERFMON_VECTOR
);
6483 if (pfm_probe_pmu()) {
6484 printk(KERN_INFO
"perfmon: disabled, there is no support for processor family %d\n",
6485 local_cpu_data
->family
);
6490 * compute the number of implemented PMD/PMC from the
6491 * description tables
6494 for (i
=0; PMC_IS_LAST(i
) == 0; i
++) {
6495 if (PMC_IS_IMPL(i
) == 0) continue;
6496 pmu_conf
->impl_pmcs
[i
>>6] |= 1UL << (i
&63);
6499 pmu_conf
->num_pmcs
= n
;
6501 n
= 0; n_counters
= 0;
6502 for (i
=0; PMD_IS_LAST(i
) == 0; i
++) {
6503 if (PMD_IS_IMPL(i
) == 0) continue;
6504 pmu_conf
->impl_pmds
[i
>>6] |= 1UL << (i
&63);
6506 if (PMD_IS_COUNTING(i
)) n_counters
++;
6508 pmu_conf
->num_pmds
= n
;
6509 pmu_conf
->num_counters
= n_counters
;
6512 * sanity checks on the number of debug registers
6514 if (pmu_conf
->use_rr_dbregs
) {
6515 if (pmu_conf
->num_ibrs
> IA64_NUM_DBG_REGS
) {
6516 printk(KERN_INFO
"perfmon: unsupported number of code debug registers (%u)\n", pmu_conf
->num_ibrs
);
6520 if (pmu_conf
->num_dbrs
> IA64_NUM_DBG_REGS
) {
6521 printk(KERN_INFO
"perfmon: unsupported number of data debug registers (%u)\n", pmu_conf
->num_ibrs
);
6527 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6531 pmu_conf
->num_counters
,
6532 ffz(pmu_conf
->ovfl_val
));
6535 if (pmu_conf
->num_pmds
>= PFM_NUM_PMD_REGS
|| pmu_conf
->num_pmcs
>= PFM_NUM_PMC_REGS
) {
6536 printk(KERN_ERR
"perfmon: not enough pmc/pmd, perfmon disabled\n");
6542 * create /proc/perfmon (mostly for debugging purposes)
6544 perfmon_dir
= proc_create_seq("perfmon", S_IRUGO
, NULL
, &pfm_seq_ops
);
6545 if (perfmon_dir
== NULL
) {
6546 printk(KERN_ERR
"perfmon: cannot create /proc entry, perfmon disabled\n");
6552 * create /proc/sys/kernel/perfmon (for debugging purposes)
6554 pfm_sysctl_header
= register_sysctl_table(pfm_sysctl_root
);
6557 * initialize all our spinlocks
6559 spin_lock_init(&pfm_sessions
.pfs_lock
);
6560 spin_lock_init(&pfm_buffer_fmt_lock
);
6564 for(i
=0; i
< NR_CPUS
; i
++) pfm_stats
[i
].pfm_ovfl_intr_cycles_min
= ~0UL;
6569 __initcall(pfm_init
);
6572 * this function is called before pfm_init()
6575 pfm_init_percpu (void)
6577 static int first_time
=1;
6579 * make sure no measurement is active
6580 * (may inherit programmed PMCs from EFI).
6586 * we run with the PMU not frozen at all times
6591 register_percpu_irq(IA64_PERFMON_VECTOR
, pfm_interrupt_handler
,
6596 ia64_setreg(_IA64_REG_CR_PMV
, IA64_PERFMON_VECTOR
);
6601 * used for debug purposes only
6604 dump_pmu_state(const char *from
)
6606 struct task_struct
*task
;
6607 struct pt_regs
*regs
;
6609 unsigned long psr
, dcr
, info
, flags
;
6612 local_irq_save(flags
);
6614 this_cpu
= smp_processor_id();
6615 regs
= task_pt_regs(current
);
6616 info
= PFM_CPUINFO_GET();
6617 dcr
= ia64_getreg(_IA64_REG_CR_DCR
);
6619 if (info
== 0 && ia64_psr(regs
)->pp
== 0 && (dcr
& IA64_DCR_PP
) == 0) {
6620 local_irq_restore(flags
);
6624 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6627 task_pid_nr(current
),
6631 task
= GET_PMU_OWNER();
6632 ctx
= GET_PMU_CTX();
6634 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu
, task
? task_pid_nr(task
) : -1, ctx
);
6636 psr
= pfm_get_psr();
6638 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",
6641 psr
& IA64_PSR_PP
? 1 : 0,
6642 psr
& IA64_PSR_UP
? 1 : 0,
6643 dcr
& IA64_DCR_PP
? 1 : 0,
6646 ia64_psr(regs
)->pp
);
6648 ia64_psr(regs
)->up
= 0;
6649 ia64_psr(regs
)->pp
= 0;
6651 for (i
=1; PMC_IS_LAST(i
) == 0; i
++) {
6652 if (PMC_IS_IMPL(i
) == 0) continue;
6653 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
]);
6656 for (i
=1; PMD_IS_LAST(i
) == 0; i
++) {
6657 if (PMD_IS_IMPL(i
) == 0) continue;
6658 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
]);
6662 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6665 ctx
->ctx_smpl_vaddr
,
6669 ctx
->ctx_saved_psr_up
);
6671 local_irq_restore(flags
);
6675 * called from process.c:copy_thread(). task is new child.
6678 pfm_inherit(struct task_struct
*task
, struct pt_regs
*regs
)
6680 struct thread_struct
*thread
;
6682 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task
)));
6684 thread
= &task
->thread
;
6687 * cut links inherited from parent (current)
6689 thread
->pfm_context
= NULL
;
6691 PFM_SET_WORK_PENDING(task
, 0);
6694 * the psr bits are already set properly in copy_threads()
6697 #else /* !CONFIG_PERFMON */
6699 sys_perfmonctl (int fd
, int cmd
, void *arg
, int count
)
6703 #endif /* CONFIG_PERFMON */