2 * This file implements the perfmon-2 subsystem which is used
3 * to program the IA-64 Performance Monitoring Unit (PMU).
5 * The initial version of perfmon.c was written by
6 * Ganesh Venkitachalam, IBM Corp.
8 * Then it was modified for perfmon-1.x by Stephane Eranian and
9 * David Mosberger, Hewlett Packard Co.
11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
12 * by Stephane Eranian, Hewlett Packard Co.
14 * Copyright (C) 1999-2005 Hewlett Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 * David Mosberger-Tang <davidm@hpl.hp.com>
18 * More information about perfmon available at:
19 * http://www.hpl.hp.com/research/linux/perfmon
22 #include <linux/module.h>
23 #include <linux/kernel.h>
24 #include <linux/sched.h>
25 #include <linux/interrupt.h>
26 #include <linux/proc_fs.h>
27 #include <linux/seq_file.h>
28 #include <linux/init.h>
29 #include <linux/vmalloc.h>
31 #include <linux/sysctl.h>
32 #include <linux/list.h>
33 #include <linux/file.h>
34 #include <linux/poll.h>
35 #include <linux/vfs.h>
36 #include <linux/smp.h>
37 #include <linux/pagemap.h>
38 #include <linux/mount.h>
39 #include <linux/bitops.h>
40 #include <linux/capability.h>
41 #include <linux/rcupdate.h>
42 #include <linux/completion.h>
43 #include <linux/tracehook.h>
44 #include <linux/slab.h>
46 #include <asm/errno.h>
47 #include <asm/intrinsics.h>
49 #include <asm/perfmon.h>
50 #include <asm/processor.h>
51 #include <asm/signal.h>
52 #include <asm/uaccess.h>
53 #include <asm/delay.h>
57 * perfmon context state
59 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
60 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
61 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
62 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
64 #define PFM_INVALID_ACTIVATION (~0UL)
66 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
67 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
70 * depth of message queue
72 #define PFM_MAX_MSGS 32
73 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
76 * type of a PMU register (bitmask).
78 * bit0 : register implemented
81 * bit4 : pmc has pmc.pm
82 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
83 * bit6-7 : register type
86 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
87 #define PFM_REG_IMPL 0x1 /* register implemented */
88 #define PFM_REG_END 0x2 /* end marker */
89 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
90 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
91 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
92 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
93 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
95 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
96 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
98 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
100 /* i assumed unsigned */
101 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
102 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
104 /* XXX: these assume that register i is implemented */
105 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
106 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
107 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
108 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
110 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
111 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
112 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
113 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
115 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
116 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
118 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
119 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
120 #define PFM_CTX_TASK(h) (h)->ctx_task
122 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
124 /* XXX: does not support more than 64 PMDs */
125 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
126 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
128 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
130 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
131 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
132 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
133 #define PFM_CODE_RR 0 /* requesting code range restriction */
134 #define PFM_DATA_RR 1 /* requestion data range restriction */
136 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
137 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
138 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
140 #define RDEP(x) (1UL<<(x))
143 * context protection macros
145 * - we need to protect against CPU concurrency (spin_lock)
146 * - we need to protect against PMU overflow interrupts (local_irq_disable)
148 * - we need to protect against PMU overflow interrupts (local_irq_disable)
150 * spin_lock_irqsave()/spin_unlock_irqrestore():
151 * in SMP: local_irq_disable + spin_lock
152 * in UP : local_irq_disable
154 * spin_lock()/spin_lock():
155 * in UP : removed automatically
156 * in SMP: protect against context accesses from other CPU. interrupts
157 * are not masked. This is useful for the PMU interrupt handler
158 * because we know we will not get PMU concurrency in that code.
160 #define PROTECT_CTX(c, f) \
162 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
163 spin_lock_irqsave(&(c)->ctx_lock, f); \
164 DPRINT(("spinlocked ctx %p by [%d]\n", c, task_pid_nr(current))); \
167 #define UNPROTECT_CTX(c, f) \
169 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
170 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
173 #define PROTECT_CTX_NOPRINT(c, f) \
175 spin_lock_irqsave(&(c)->ctx_lock, f); \
179 #define UNPROTECT_CTX_NOPRINT(c, f) \
181 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
185 #define PROTECT_CTX_NOIRQ(c) \
187 spin_lock(&(c)->ctx_lock); \
190 #define UNPROTECT_CTX_NOIRQ(c) \
192 spin_unlock(&(c)->ctx_lock); \
198 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
199 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
200 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
202 #else /* !CONFIG_SMP */
203 #define SET_ACTIVATION(t) do {} while(0)
204 #define GET_ACTIVATION(t) do {} while(0)
205 #define INC_ACTIVATION(t) do {} while(0)
206 #endif /* CONFIG_SMP */
208 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
209 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
210 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
212 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
213 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
215 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
218 * cmp0 must be the value of pmc0
220 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
222 #define PFMFS_MAGIC 0xa0b4d889
227 #define PFM_DEBUGGING 1
231 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
234 #define DPRINT_ovfl(a) \
236 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
241 * 64-bit software counter structure
243 * the next_reset_type is applied to the next call to pfm_reset_regs()
246 unsigned long val; /* virtual 64bit counter value */
247 unsigned long lval; /* last reset value */
248 unsigned long long_reset; /* reset value on sampling overflow */
249 unsigned long short_reset; /* reset value on overflow */
250 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
251 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
252 unsigned long seed; /* seed for random-number generator */
253 unsigned long mask; /* mask for random-number generator */
254 unsigned int flags; /* notify/do not notify */
255 unsigned long eventid; /* overflow event identifier */
262 unsigned int block:1; /* when 1, task will blocked on user notifications */
263 unsigned int system:1; /* do system wide monitoring */
264 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
265 unsigned int is_sampling:1; /* true if using a custom format */
266 unsigned int excl_idle:1; /* exclude idle task in system wide session */
267 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
268 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
269 unsigned int no_msg:1; /* no message sent on overflow */
270 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
271 unsigned int reserved:22;
272 } pfm_context_flags_t;
274 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
275 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
276 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
280 * perfmon context: encapsulates all the state of a monitoring session
283 typedef struct pfm_context {
284 spinlock_t ctx_lock; /* context protection */
286 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
287 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
289 struct task_struct *ctx_task; /* task to which context is attached */
291 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
293 struct completion ctx_restart_done; /* use for blocking notification mode */
295 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
296 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
297 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
299 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
300 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
301 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
303 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
305 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
306 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
307 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
308 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
310 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
312 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
313 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
315 unsigned long ctx_saved_psr_up; /* only contains psr.up value */
317 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
318 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
319 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
321 int ctx_fd; /* file descriptor used my this context */
322 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
324 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
325 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
326 unsigned long ctx_smpl_size; /* size of sampling buffer */
327 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
329 wait_queue_head_t ctx_msgq_wait;
330 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
333 struct fasync_struct *ctx_async_queue;
335 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
339 * magic number used to verify that structure is really
342 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
344 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
347 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
348 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
350 #define SET_LAST_CPU(ctx, v) do {} while(0)
351 #define GET_LAST_CPU(ctx) do {} while(0)
355 #define ctx_fl_block ctx_flags.block
356 #define ctx_fl_system ctx_flags.system
357 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
358 #define ctx_fl_is_sampling ctx_flags.is_sampling
359 #define ctx_fl_excl_idle ctx_flags.excl_idle
360 #define ctx_fl_going_zombie ctx_flags.going_zombie
361 #define ctx_fl_trap_reason ctx_flags.trap_reason
362 #define ctx_fl_no_msg ctx_flags.no_msg
363 #define ctx_fl_can_restart ctx_flags.can_restart
365 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
366 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
369 * global information about all sessions
370 * mostly used to synchronize between system wide and per-process
373 spinlock_t pfs_lock; /* lock the structure */
375 unsigned int pfs_task_sessions; /* number of per task sessions */
376 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
377 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
378 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
379 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
383 * information about a PMC or PMD.
384 * dep_pmd[]: a bitmask of dependent PMD registers
385 * dep_pmc[]: a bitmask of dependent PMC registers
387 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
391 unsigned long default_value; /* power-on default value */
392 unsigned long reserved_mask; /* bitmask of reserved bits */
393 pfm_reg_check_t read_check;
394 pfm_reg_check_t write_check;
395 unsigned long dep_pmd[4];
396 unsigned long dep_pmc[4];
399 /* assume cnum is a valid monitor */
400 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
403 * This structure is initialized at boot time and contains
404 * a description of the PMU main characteristics.
406 * If the probe function is defined, detection is based
407 * on its return value:
408 * - 0 means recognized PMU
409 * - anything else means not supported
410 * When the probe function is not defined, then the pmu_family field
411 * is used and it must match the host CPU family such that:
412 * - cpu->family & config->pmu_family != 0
415 unsigned long ovfl_val; /* overflow value for counters */
417 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
418 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
420 unsigned int num_pmcs; /* number of PMCS: computed at init time */
421 unsigned int num_pmds; /* number of PMDS: computed at init time */
422 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
423 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
425 char *pmu_name; /* PMU family name */
426 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
427 unsigned int flags; /* pmu specific flags */
428 unsigned int num_ibrs; /* number of IBRS: computed at init time */
429 unsigned int num_dbrs; /* number of DBRS: computed at init time */
430 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
431 int (*probe)(void); /* customized probe routine */
432 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
437 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
440 * debug register related type definitions
443 unsigned long ibr_mask:56;
444 unsigned long ibr_plm:4;
445 unsigned long ibr_ig:3;
446 unsigned long ibr_x:1;
450 unsigned long dbr_mask:56;
451 unsigned long dbr_plm:4;
452 unsigned long dbr_ig:2;
453 unsigned long dbr_w:1;
454 unsigned long dbr_r:1;
465 * perfmon command descriptions
468 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
471 unsigned int cmd_narg;
473 int (*cmd_getsize)(void *arg, size_t *sz);
476 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
477 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
478 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
479 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
482 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
483 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
484 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
485 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
486 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
488 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
491 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
492 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
493 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
496 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
497 unsigned long pfm_smpl_handler_calls;
498 unsigned long pfm_smpl_handler_cycles;
499 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
503 * perfmon internal variables
505 static pfm_stats_t pfm_stats[NR_CPUS];
506 static pfm_session_t pfm_sessions; /* global sessions information */
508 static DEFINE_SPINLOCK(pfm_alt_install_check);
509 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
511 static struct proc_dir_entry *perfmon_dir;
512 static pfm_uuid_t pfm_null_uuid = {0,};
514 static spinlock_t pfm_buffer_fmt_lock;
515 static LIST_HEAD(pfm_buffer_fmt_list);
517 static pmu_config_t *pmu_conf;
519 /* sysctl() controls */
520 pfm_sysctl_t pfm_sysctl;
521 EXPORT_SYMBOL(pfm_sysctl);
523 static ctl_table pfm_ctl_table[]={
526 .data = &pfm_sysctl.debug,
527 .maxlen = sizeof(int),
529 .proc_handler = proc_dointvec,
532 .procname = "debug_ovfl",
533 .data = &pfm_sysctl.debug_ovfl,
534 .maxlen = sizeof(int),
536 .proc_handler = proc_dointvec,
539 .procname = "fastctxsw",
540 .data = &pfm_sysctl.fastctxsw,
541 .maxlen = sizeof(int),
543 .proc_handler = proc_dointvec,
546 .procname = "expert_mode",
547 .data = &pfm_sysctl.expert_mode,
548 .maxlen = sizeof(int),
550 .proc_handler = proc_dointvec,
554 static ctl_table pfm_sysctl_dir[] = {
556 .procname = "perfmon",
558 .child = pfm_ctl_table,
562 static ctl_table pfm_sysctl_root[] = {
564 .procname = "kernel",
566 .child = pfm_sysctl_dir,
570 static struct ctl_table_header *pfm_sysctl_header;
572 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
574 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
575 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
578 pfm_put_task(struct task_struct *task)
580 if (task != current) put_task_struct(task);
584 pfm_reserve_page(unsigned long a)
586 SetPageReserved(vmalloc_to_page((void *)a));
589 pfm_unreserve_page(unsigned long a)
591 ClearPageReserved(vmalloc_to_page((void*)a));
594 static inline unsigned long
595 pfm_protect_ctx_ctxsw(pfm_context_t *x)
597 spin_lock(&(x)->ctx_lock);
602 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
604 spin_unlock(&(x)->ctx_lock);
607 /* forward declaration */
608 static const struct dentry_operations pfmfs_dentry_operations;
610 static struct dentry *
611 pfmfs_mount(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
613 return mount_pseudo(fs_type, "pfm:", NULL, &pfmfs_dentry_operations,
617 static struct file_system_type pfm_fs_type = {
619 .mount = pfmfs_mount,
620 .kill_sb = kill_anon_super,
622 MODULE_ALIAS_FS("pfmfs");
624 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
625 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
626 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
627 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
628 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
631 /* forward declaration */
632 static const struct file_operations pfm_file_ops;
635 * forward declarations
638 static void pfm_lazy_save_regs (struct task_struct *ta);
641 void dump_pmu_state(const char *);
642 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
644 #include "perfmon_itanium.h"
645 #include "perfmon_mckinley.h"
646 #include "perfmon_montecito.h"
647 #include "perfmon_generic.h"
649 static pmu_config_t *pmu_confs[]={
653 &pmu_conf_gen, /* must be last */
658 static int pfm_end_notify_user(pfm_context_t *ctx);
661 pfm_clear_psr_pp(void)
663 ia64_rsm(IA64_PSR_PP);
670 ia64_ssm(IA64_PSR_PP);
675 pfm_clear_psr_up(void)
677 ia64_rsm(IA64_PSR_UP);
684 ia64_ssm(IA64_PSR_UP);
688 static inline unsigned long
692 tmp = ia64_getreg(_IA64_REG_PSR);
698 pfm_set_psr_l(unsigned long val)
700 ia64_setreg(_IA64_REG_PSR_L, val);
712 pfm_unfreeze_pmu(void)
719 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
723 for (i=0; i < nibrs; i++) {
724 ia64_set_ibr(i, ibrs[i]);
725 ia64_dv_serialize_instruction();
731 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
735 for (i=0; i < ndbrs; i++) {
736 ia64_set_dbr(i, dbrs[i]);
737 ia64_dv_serialize_data();
743 * PMD[i] must be a counter. no check is made
745 static inline unsigned long
746 pfm_read_soft_counter(pfm_context_t *ctx, int i)
748 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
752 * PMD[i] must be a counter. no check is made
755 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
757 unsigned long ovfl_val = pmu_conf->ovfl_val;
759 ctx->ctx_pmds[i].val = val & ~ovfl_val;
761 * writing to unimplemented part is ignore, so we do not need to
764 ia64_set_pmd(i, val & ovfl_val);
768 pfm_get_new_msg(pfm_context_t *ctx)
772 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
774 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
775 if (next == ctx->ctx_msgq_head) return NULL;
777 idx = ctx->ctx_msgq_tail;
778 ctx->ctx_msgq_tail = next;
780 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
782 return ctx->ctx_msgq+idx;
786 pfm_get_next_msg(pfm_context_t *ctx)
790 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
792 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
797 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
802 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
804 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));
810 pfm_reset_msgq(pfm_context_t *ctx)
812 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
813 DPRINT(("ctx=%p msgq reset\n", ctx));
817 pfm_rvmalloc(unsigned long size)
822 size = PAGE_ALIGN(size);
825 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
826 addr = (unsigned long)mem;
828 pfm_reserve_page(addr);
837 pfm_rvfree(void *mem, unsigned long size)
842 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
843 addr = (unsigned long) mem;
844 while ((long) size > 0) {
845 pfm_unreserve_page(addr);
854 static pfm_context_t *
855 pfm_context_alloc(int ctx_flags)
860 * allocate context descriptor
861 * must be able to free with interrupts disabled
863 ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
865 DPRINT(("alloc ctx @%p\n", ctx));
868 * init context protection lock
870 spin_lock_init(&ctx->ctx_lock);
873 * context is unloaded
875 ctx->ctx_state = PFM_CTX_UNLOADED;
878 * initialization of context's flags
880 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
881 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
882 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
884 * will move to set properties
885 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
889 * init restart semaphore to locked
891 init_completion(&ctx->ctx_restart_done);
894 * activation is used in SMP only
896 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
897 SET_LAST_CPU(ctx, -1);
900 * initialize notification message queue
902 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
903 init_waitqueue_head(&ctx->ctx_msgq_wait);
904 init_waitqueue_head(&ctx->ctx_zombieq);
911 pfm_context_free(pfm_context_t *ctx)
914 DPRINT(("free ctx @%p\n", ctx));
920 pfm_mask_monitoring(struct task_struct *task)
922 pfm_context_t *ctx = PFM_GET_CTX(task);
923 unsigned long mask, val, ovfl_mask;
926 DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
928 ovfl_mask = pmu_conf->ovfl_val;
930 * monitoring can only be masked as a result of a valid
931 * counter overflow. In UP, it means that the PMU still
932 * has an owner. Note that the owner can be different
933 * from the current task. However the PMU state belongs
935 * In SMP, a valid overflow only happens when task is
936 * current. Therefore if we come here, we know that
937 * the PMU state belongs to the current task, therefore
938 * we can access the live registers.
940 * So in both cases, the live register contains the owner's
941 * state. We can ONLY touch the PMU registers and NOT the PSR.
943 * As a consequence to this call, the ctx->th_pmds[] array
944 * contains stale information which must be ignored
945 * when context is reloaded AND monitoring is active (see
948 mask = ctx->ctx_used_pmds[0];
949 for (i = 0; mask; i++, mask>>=1) {
950 /* skip non used pmds */
951 if ((mask & 0x1) == 0) continue;
952 val = ia64_get_pmd(i);
954 if (PMD_IS_COUNTING(i)) {
956 * we rebuild the full 64 bit value of the counter
958 ctx->ctx_pmds[i].val += (val & ovfl_mask);
960 ctx->ctx_pmds[i].val = val;
962 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
964 ctx->ctx_pmds[i].val,
968 * mask monitoring by setting the privilege level to 0
969 * we cannot use psr.pp/psr.up for this, it is controlled by
972 * if task is current, modify actual registers, otherwise modify
973 * thread save state, i.e., what will be restored in pfm_load_regs()
975 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
976 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
977 if ((mask & 0x1) == 0UL) continue;
978 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
979 ctx->th_pmcs[i] &= ~0xfUL;
980 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
983 * make all of this visible
989 * must always be done with task == current
991 * context must be in MASKED state when calling
994 pfm_restore_monitoring(struct task_struct *task)
996 pfm_context_t *ctx = PFM_GET_CTX(task);
997 unsigned long mask, ovfl_mask;
998 unsigned long psr, val;
1001 is_system = ctx->ctx_fl_system;
1002 ovfl_mask = pmu_conf->ovfl_val;
1004 if (task != current) {
1005 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1008 if (ctx->ctx_state != PFM_CTX_MASKED) {
1009 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1010 task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1013 psr = pfm_get_psr();
1015 * monitoring is masked via the PMC.
1016 * As we restore their value, we do not want each counter to
1017 * restart right away. We stop monitoring using the PSR,
1018 * restore the PMC (and PMD) and then re-establish the psr
1019 * as it was. Note that there can be no pending overflow at
1020 * this point, because monitoring was MASKED.
1022 * system-wide session are pinned and self-monitoring
1024 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1025 /* disable dcr pp */
1026 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1032 * first, we restore the PMD
1034 mask = ctx->ctx_used_pmds[0];
1035 for (i = 0; mask; i++, mask>>=1) {
1036 /* skip non used pmds */
1037 if ((mask & 0x1) == 0) continue;
1039 if (PMD_IS_COUNTING(i)) {
1041 * we split the 64bit value according to
1044 val = ctx->ctx_pmds[i].val & ovfl_mask;
1045 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1047 val = ctx->ctx_pmds[i].val;
1049 ia64_set_pmd(i, val);
1051 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1053 ctx->ctx_pmds[i].val,
1059 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1060 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1061 if ((mask & 0x1) == 0UL) continue;
1062 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1063 ia64_set_pmc(i, ctx->th_pmcs[i]);
1064 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1065 task_pid_nr(task), i, ctx->th_pmcs[i]));
1070 * must restore DBR/IBR because could be modified while masked
1071 * XXX: need to optimize
1073 if (ctx->ctx_fl_using_dbreg) {
1074 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1075 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1081 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1083 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1090 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1096 for (i=0; mask; i++, mask>>=1) {
1097 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1102 * reload from thread state (used for ctxw only)
1105 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1108 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1110 for (i=0; mask; i++, mask>>=1) {
1111 if ((mask & 0x1) == 0) continue;
1112 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1113 ia64_set_pmd(i, val);
1119 * propagate PMD from context to thread-state
1122 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1124 unsigned long ovfl_val = pmu_conf->ovfl_val;
1125 unsigned long mask = ctx->ctx_all_pmds[0];
1129 DPRINT(("mask=0x%lx\n", mask));
1131 for (i=0; mask; i++, mask>>=1) {
1133 val = ctx->ctx_pmds[i].val;
1136 * We break up the 64 bit value into 2 pieces
1137 * the lower bits go to the machine state in the
1138 * thread (will be reloaded on ctxsw in).
1139 * The upper part stays in the soft-counter.
1141 if (PMD_IS_COUNTING(i)) {
1142 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1145 ctx->th_pmds[i] = val;
1147 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1150 ctx->ctx_pmds[i].val));
1155 * propagate PMC from context to thread-state
1158 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1160 unsigned long mask = ctx->ctx_all_pmcs[0];
1163 DPRINT(("mask=0x%lx\n", mask));
1165 for (i=0; mask; i++, mask>>=1) {
1166 /* masking 0 with ovfl_val yields 0 */
1167 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1168 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1175 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1179 for (i=0; mask; i++, mask>>=1) {
1180 if ((mask & 0x1) == 0) continue;
1181 ia64_set_pmc(i, pmcs[i]);
1187 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1189 return memcmp(a, b, sizeof(pfm_uuid_t));
1193 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1196 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1201 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1204 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1210 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1214 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1219 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1223 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1228 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1231 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1236 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)
1239 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1243 static pfm_buffer_fmt_t *
1244 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1246 struct list_head * pos;
1247 pfm_buffer_fmt_t * entry;
1249 list_for_each(pos, &pfm_buffer_fmt_list) {
1250 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1251 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1258 * find a buffer format based on its uuid
1260 static pfm_buffer_fmt_t *
1261 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1263 pfm_buffer_fmt_t * fmt;
1264 spin_lock(&pfm_buffer_fmt_lock);
1265 fmt = __pfm_find_buffer_fmt(uuid);
1266 spin_unlock(&pfm_buffer_fmt_lock);
1271 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1275 /* some sanity checks */
1276 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1278 /* we need at least a handler */
1279 if (fmt->fmt_handler == NULL) return -EINVAL;
1282 * XXX: need check validity of fmt_arg_size
1285 spin_lock(&pfm_buffer_fmt_lock);
1287 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1288 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1292 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1293 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1296 spin_unlock(&pfm_buffer_fmt_lock);
1299 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1302 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1304 pfm_buffer_fmt_t *fmt;
1307 spin_lock(&pfm_buffer_fmt_lock);
1309 fmt = __pfm_find_buffer_fmt(uuid);
1311 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1315 list_del_init(&fmt->fmt_list);
1316 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1319 spin_unlock(&pfm_buffer_fmt_lock);
1323 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1325 extern void update_pal_halt_status(int);
1328 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1330 unsigned long flags;
1332 * validity checks on cpu_mask have been done upstream
1336 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1337 pfm_sessions.pfs_sys_sessions,
1338 pfm_sessions.pfs_task_sessions,
1339 pfm_sessions.pfs_sys_use_dbregs,
1345 * cannot mix system wide and per-task sessions
1347 if (pfm_sessions.pfs_task_sessions > 0UL) {
1348 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1349 pfm_sessions.pfs_task_sessions));
1353 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1355 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1357 pfm_sessions.pfs_sys_session[cpu] = task;
1359 pfm_sessions.pfs_sys_sessions++ ;
1362 if (pfm_sessions.pfs_sys_sessions) goto abort;
1363 pfm_sessions.pfs_task_sessions++;
1366 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1367 pfm_sessions.pfs_sys_sessions,
1368 pfm_sessions.pfs_task_sessions,
1369 pfm_sessions.pfs_sys_use_dbregs,
1374 * disable default_idle() to go to PAL_HALT
1376 update_pal_halt_status(0);
1383 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1384 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1394 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1396 unsigned long flags;
1398 * validity checks on cpu_mask have been done upstream
1402 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1403 pfm_sessions.pfs_sys_sessions,
1404 pfm_sessions.pfs_task_sessions,
1405 pfm_sessions.pfs_sys_use_dbregs,
1411 pfm_sessions.pfs_sys_session[cpu] = NULL;
1413 * would not work with perfmon+more than one bit in cpu_mask
1415 if (ctx && ctx->ctx_fl_using_dbreg) {
1416 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1417 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1419 pfm_sessions.pfs_sys_use_dbregs--;
1422 pfm_sessions.pfs_sys_sessions--;
1424 pfm_sessions.pfs_task_sessions--;
1426 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1427 pfm_sessions.pfs_sys_sessions,
1428 pfm_sessions.pfs_task_sessions,
1429 pfm_sessions.pfs_sys_use_dbregs,
1434 * if possible, enable default_idle() to go into PAL_HALT
1436 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1437 update_pal_halt_status(1);
1445 * removes virtual mapping of the sampling buffer.
1446 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1447 * a PROTECT_CTX() section.
1450 pfm_remove_smpl_mapping(void *vaddr, unsigned long size)
1452 struct task_struct *task = current;
1456 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1457 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1461 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1464 * does the actual unmapping
1466 r = vm_munmap((unsigned long)vaddr, size);
1469 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1472 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1478 * free actual physical storage used by sampling buffer
1482 pfm_free_smpl_buffer(pfm_context_t *ctx)
1484 pfm_buffer_fmt_t *fmt;
1486 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1489 * we won't use the buffer format anymore
1491 fmt = ctx->ctx_buf_fmt;
1493 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1496 ctx->ctx_smpl_vaddr));
1498 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1503 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1505 ctx->ctx_smpl_hdr = NULL;
1506 ctx->ctx_smpl_size = 0UL;
1511 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1517 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1519 if (fmt == NULL) return;
1521 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1526 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1527 * no real gain from having the whole whorehouse mounted. So we don't need
1528 * any operations on the root directory. However, we need a non-trivial
1529 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1531 static struct vfsmount *pfmfs_mnt __read_mostly;
1536 int err = register_filesystem(&pfm_fs_type);
1538 pfmfs_mnt = kern_mount(&pfm_fs_type);
1539 err = PTR_ERR(pfmfs_mnt);
1540 if (IS_ERR(pfmfs_mnt))
1541 unregister_filesystem(&pfm_fs_type);
1549 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1554 unsigned long flags;
1555 DECLARE_WAITQUEUE(wait, current);
1556 if (PFM_IS_FILE(filp) == 0) {
1557 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1561 ctx = filp->private_data;
1563 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1568 * check even when there is no message
1570 if (size < sizeof(pfm_msg_t)) {
1571 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1575 PROTECT_CTX(ctx, flags);
1578 * put ourselves on the wait queue
1580 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1588 set_current_state(TASK_INTERRUPTIBLE);
1590 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1593 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1595 UNPROTECT_CTX(ctx, flags);
1598 * check non-blocking read
1601 if(filp->f_flags & O_NONBLOCK) break;
1604 * check pending signals
1606 if(signal_pending(current)) {
1611 * no message, so wait
1615 PROTECT_CTX(ctx, flags);
1617 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1618 set_current_state(TASK_RUNNING);
1619 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1621 if (ret < 0) goto abort;
1624 msg = pfm_get_next_msg(ctx);
1626 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1630 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1633 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1636 UNPROTECT_CTX(ctx, flags);
1642 pfm_write(struct file *file, const char __user *ubuf,
1643 size_t size, loff_t *ppos)
1645 DPRINT(("pfm_write called\n"));
1650 pfm_poll(struct file *filp, poll_table * wait)
1653 unsigned long flags;
1654 unsigned int mask = 0;
1656 if (PFM_IS_FILE(filp) == 0) {
1657 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1661 ctx = filp->private_data;
1663 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1668 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1670 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1672 PROTECT_CTX(ctx, flags);
1674 if (PFM_CTXQ_EMPTY(ctx) == 0)
1675 mask = POLLIN | POLLRDNORM;
1677 UNPROTECT_CTX(ctx, flags);
1679 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1685 pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1687 DPRINT(("pfm_ioctl called\n"));
1692 * interrupt cannot be masked when coming here
1695 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1699 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1701 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1702 task_pid_nr(current),
1705 ctx->ctx_async_queue, ret));
1711 pfm_fasync(int fd, struct file *filp, int on)
1716 if (PFM_IS_FILE(filp) == 0) {
1717 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1721 ctx = filp->private_data;
1723 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1727 * we cannot mask interrupts during this call because this may
1728 * may go to sleep if memory is not readily avalaible.
1730 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1731 * done in caller. Serialization of this function is ensured by caller.
1733 ret = pfm_do_fasync(fd, filp, ctx, on);
1736 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1739 ctx->ctx_async_queue, ret));
1746 * this function is exclusively called from pfm_close().
1747 * The context is not protected at that time, nor are interrupts
1748 * on the remote CPU. That's necessary to avoid deadlocks.
1751 pfm_syswide_force_stop(void *info)
1753 pfm_context_t *ctx = (pfm_context_t *)info;
1754 struct pt_regs *regs = task_pt_regs(current);
1755 struct task_struct *owner;
1756 unsigned long flags;
1759 if (ctx->ctx_cpu != smp_processor_id()) {
1760 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1762 smp_processor_id());
1765 owner = GET_PMU_OWNER();
1766 if (owner != ctx->ctx_task) {
1767 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1769 task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1772 if (GET_PMU_CTX() != ctx) {
1773 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1775 GET_PMU_CTX(), ctx);
1779 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1781 * the context is already protected in pfm_close(), we simply
1782 * need to mask interrupts to avoid a PMU interrupt race on
1785 local_irq_save(flags);
1787 ret = pfm_context_unload(ctx, NULL, 0, regs);
1789 DPRINT(("context_unload returned %d\n", ret));
1793 * unmask interrupts, PMU interrupts are now spurious here
1795 local_irq_restore(flags);
1799 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1803 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1804 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1805 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1807 #endif /* CONFIG_SMP */
1810 * called for each close(). Partially free resources.
1811 * When caller is self-monitoring, the context is unloaded.
1814 pfm_flush(struct file *filp, fl_owner_t id)
1817 struct task_struct *task;
1818 struct pt_regs *regs;
1819 unsigned long flags;
1820 unsigned long smpl_buf_size = 0UL;
1821 void *smpl_buf_vaddr = NULL;
1822 int state, is_system;
1824 if (PFM_IS_FILE(filp) == 0) {
1825 DPRINT(("bad magic for\n"));
1829 ctx = filp->private_data;
1831 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1836 * remove our file from the async queue, if we use this mode.
1837 * This can be done without the context being protected. We come
1838 * here when the context has become unreachable by other tasks.
1840 * We may still have active monitoring at this point and we may
1841 * end up in pfm_overflow_handler(). However, fasync_helper()
1842 * operates with interrupts disabled and it cleans up the
1843 * queue. If the PMU handler is called prior to entering
1844 * fasync_helper() then it will send a signal. If it is
1845 * invoked after, it will find an empty queue and no
1846 * signal will be sent. In both case, we are safe
1848 PROTECT_CTX(ctx, flags);
1850 state = ctx->ctx_state;
1851 is_system = ctx->ctx_fl_system;
1853 task = PFM_CTX_TASK(ctx);
1854 regs = task_pt_regs(task);
1856 DPRINT(("ctx_state=%d is_current=%d\n",
1858 task == current ? 1 : 0));
1861 * if state == UNLOADED, then task is NULL
1865 * we must stop and unload because we are losing access to the context.
1867 if (task == current) {
1870 * the task IS the owner but it migrated to another CPU: that's bad
1871 * but we must handle this cleanly. Unfortunately, the kernel does
1872 * not provide a mechanism to block migration (while the context is loaded).
1874 * We need to release the resource on the ORIGINAL cpu.
1876 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1878 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1880 * keep context protected but unmask interrupt for IPI
1882 local_irq_restore(flags);
1884 pfm_syswide_cleanup_other_cpu(ctx);
1887 * restore interrupt masking
1889 local_irq_save(flags);
1892 * context is unloaded at this point
1895 #endif /* CONFIG_SMP */
1898 DPRINT(("forcing unload\n"));
1900 * stop and unload, returning with state UNLOADED
1901 * and session unreserved.
1903 pfm_context_unload(ctx, NULL, 0, regs);
1905 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1910 * remove virtual mapping, if any, for the calling task.
1911 * cannot reset ctx field until last user is calling close().
1913 * ctx_smpl_vaddr must never be cleared because it is needed
1914 * by every task with access to the context
1916 * When called from do_exit(), the mm context is gone already, therefore
1917 * mm is NULL, i.e., the VMA is already gone and we do not have to
1920 if (ctx->ctx_smpl_vaddr && current->mm) {
1921 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1922 smpl_buf_size = ctx->ctx_smpl_size;
1925 UNPROTECT_CTX(ctx, flags);
1928 * if there was a mapping, then we systematically remove it
1929 * at this point. Cannot be done inside critical section
1930 * because some VM function reenables interrupts.
1933 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(smpl_buf_vaddr, smpl_buf_size);
1938 * called either on explicit close() or from exit_files().
1939 * Only the LAST user of the file gets to this point, i.e., it is
1942 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1943 * (fput()),i.e, last task to access the file. Nobody else can access the
1944 * file at this point.
1946 * When called from exit_files(), the VMA has been freed because exit_mm()
1947 * is executed before exit_files().
1949 * When called from exit_files(), the current task is not yet ZOMBIE but we
1950 * flush the PMU state to the context.
1953 pfm_close(struct inode *inode, struct file *filp)
1956 struct task_struct *task;
1957 struct pt_regs *regs;
1958 DECLARE_WAITQUEUE(wait, current);
1959 unsigned long flags;
1960 unsigned long smpl_buf_size = 0UL;
1961 void *smpl_buf_addr = NULL;
1962 int free_possible = 1;
1963 int state, is_system;
1965 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1967 if (PFM_IS_FILE(filp) == 0) {
1968 DPRINT(("bad magic\n"));
1972 ctx = filp->private_data;
1974 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1978 PROTECT_CTX(ctx, flags);
1980 state = ctx->ctx_state;
1981 is_system = ctx->ctx_fl_system;
1983 task = PFM_CTX_TASK(ctx);
1984 regs = task_pt_regs(task);
1986 DPRINT(("ctx_state=%d is_current=%d\n",
1988 task == current ? 1 : 0));
1991 * if task == current, then pfm_flush() unloaded the context
1993 if (state == PFM_CTX_UNLOADED) goto doit;
1996 * context is loaded/masked and task != current, we need to
1997 * either force an unload or go zombie
2001 * The task is currently blocked or will block after an overflow.
2002 * we must force it to wakeup to get out of the
2003 * MASKED state and transition to the unloaded state by itself.
2005 * This situation is only possible for per-task mode
2007 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2010 * set a "partial" zombie state to be checked
2011 * upon return from down() in pfm_handle_work().
2013 * We cannot use the ZOMBIE state, because it is checked
2014 * by pfm_load_regs() which is called upon wakeup from down().
2015 * In such case, it would free the context and then we would
2016 * return to pfm_handle_work() which would access the
2017 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2018 * but visible to pfm_handle_work().
2020 * For some window of time, we have a zombie context with
2021 * ctx_state = MASKED and not ZOMBIE
2023 ctx->ctx_fl_going_zombie = 1;
2026 * force task to wake up from MASKED state
2028 complete(&ctx->ctx_restart_done);
2030 DPRINT(("waking up ctx_state=%d\n", state));
2033 * put ourself to sleep waiting for the other
2034 * task to report completion
2036 * the context is protected by mutex, therefore there
2037 * is no risk of being notified of completion before
2038 * begin actually on the waitq.
2040 set_current_state(TASK_INTERRUPTIBLE);
2041 add_wait_queue(&ctx->ctx_zombieq, &wait);
2043 UNPROTECT_CTX(ctx, flags);
2046 * XXX: check for signals :
2047 * - ok for explicit close
2048 * - not ok when coming from exit_files()
2053 PROTECT_CTX(ctx, flags);
2056 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2057 set_current_state(TASK_RUNNING);
2060 * context is unloaded at this point
2062 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2064 else if (task != current) {
2067 * switch context to zombie state
2069 ctx->ctx_state = PFM_CTX_ZOMBIE;
2071 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2073 * cannot free the context on the spot. deferred until
2074 * the task notices the ZOMBIE state
2078 pfm_context_unload(ctx, NULL, 0, regs);
2083 /* reload state, may have changed during opening of critical section */
2084 state = ctx->ctx_state;
2087 * the context is still attached to a task (possibly current)
2088 * we cannot destroy it right now
2092 * we must free the sampling buffer right here because
2093 * we cannot rely on it being cleaned up later by the
2094 * monitored task. It is not possible to free vmalloc'ed
2095 * memory in pfm_load_regs(). Instead, we remove the buffer
2096 * now. should there be subsequent PMU overflow originally
2097 * meant for sampling, the will be converted to spurious
2098 * and that's fine because the monitoring tools is gone anyway.
2100 if (ctx->ctx_smpl_hdr) {
2101 smpl_buf_addr = ctx->ctx_smpl_hdr;
2102 smpl_buf_size = ctx->ctx_smpl_size;
2103 /* no more sampling */
2104 ctx->ctx_smpl_hdr = NULL;
2105 ctx->ctx_fl_is_sampling = 0;
2108 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2114 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2117 * UNLOADED that the session has already been unreserved.
2119 if (state == PFM_CTX_ZOMBIE) {
2120 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2124 * disconnect file descriptor from context must be done
2127 filp->private_data = NULL;
2130 * if we free on the spot, the context is now completely unreachable
2131 * from the callers side. The monitored task side is also cut, so we
2134 * If we have a deferred free, only the caller side is disconnected.
2136 UNPROTECT_CTX(ctx, flags);
2139 * All memory free operations (especially for vmalloc'ed memory)
2140 * MUST be done with interrupts ENABLED.
2142 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2145 * return the memory used by the context
2147 if (free_possible) pfm_context_free(ctx);
2153 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2155 DPRINT(("pfm_no_open called\n"));
2161 static const struct file_operations pfm_file_ops = {
2162 .llseek = no_llseek,
2166 .unlocked_ioctl = pfm_ioctl,
2167 .open = pfm_no_open, /* special open code to disallow open via /proc */
2168 .fasync = pfm_fasync,
2169 .release = pfm_close,
2174 pfmfs_delete_dentry(const struct dentry *dentry)
2179 static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2181 return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2182 dentry->d_inode->i_ino);
2185 static const struct dentry_operations pfmfs_dentry_operations = {
2186 .d_delete = pfmfs_delete_dentry,
2187 .d_dname = pfmfs_dname,
2191 static struct file *
2192 pfm_alloc_file(pfm_context_t *ctx)
2195 struct inode *inode;
2197 struct qstr this = { .name = "" };
2200 * allocate a new inode
2202 inode = new_inode(pfmfs_mnt->mnt_sb);
2204 return ERR_PTR(-ENOMEM);
2206 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2208 inode->i_mode = S_IFCHR|S_IRUGO;
2209 inode->i_uid = current_fsuid();
2210 inode->i_gid = current_fsgid();
2213 * allocate a new dcache entry
2215 path.dentry = d_alloc(pfmfs_mnt->mnt_root, &this);
2218 return ERR_PTR(-ENOMEM);
2220 path.mnt = mntget(pfmfs_mnt);
2222 d_add(path.dentry, inode);
2224 file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2230 file->f_flags = O_RDONLY;
2231 file->private_data = ctx;
2237 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2239 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2242 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2245 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2256 * allocate a sampling buffer and remaps it into the user address space of the task
2259 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2261 struct mm_struct *mm = task->mm;
2262 struct vm_area_struct *vma = NULL;
2268 * the fixed header + requested size and align to page boundary
2270 size = PAGE_ALIGN(rsize);
2272 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2275 * check requested size to avoid Denial-of-service attacks
2276 * XXX: may have to refine this test
2277 * Check against address space limit.
2279 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2282 if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2286 * We do the easy to undo allocations first.
2288 * pfm_rvmalloc(), clears the buffer, so there is no leak
2290 smpl_buf = pfm_rvmalloc(size);
2291 if (smpl_buf == NULL) {
2292 DPRINT(("Can't allocate sampling buffer\n"));
2296 DPRINT(("smpl_buf @%p\n", smpl_buf));
2299 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2301 DPRINT(("Cannot allocate vma\n"));
2304 INIT_LIST_HEAD(&vma->anon_vma_chain);
2307 * partially initialize the vma for the sampling buffer
2310 vma->vm_file = get_file(filp);
2311 vma->vm_flags = VM_READ|VM_MAYREAD|VM_DONTEXPAND|VM_DONTDUMP;
2312 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2315 * Now we have everything we need and we can initialize
2316 * and connect all the data structures
2319 ctx->ctx_smpl_hdr = smpl_buf;
2320 ctx->ctx_smpl_size = size; /* aligned size */
2323 * Let's do the difficult operations next.
2325 * now we atomically find some area in the address space and
2326 * remap the buffer in it.
2328 down_write(&task->mm->mmap_sem);
2330 /* find some free area in address space, must have mmap sem held */
2331 vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2332 if (IS_ERR_VALUE(vma->vm_start)) {
2333 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2334 up_write(&task->mm->mmap_sem);
2337 vma->vm_end = vma->vm_start + size;
2338 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2340 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2342 /* can only be applied to current task, need to have the mm semaphore held when called */
2343 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2344 DPRINT(("Can't remap buffer\n"));
2345 up_write(&task->mm->mmap_sem);
2350 * now insert the vma in the vm list for the process, must be
2351 * done with mmap lock held
2353 insert_vm_struct(mm, vma);
2355 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2357 up_write(&task->mm->mmap_sem);
2360 * keep track of user level virtual address
2362 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2363 *(unsigned long *)user_vaddr = vma->vm_start;
2368 kmem_cache_free(vm_area_cachep, vma);
2370 pfm_rvfree(smpl_buf, size);
2376 * XXX: do something better here
2379 pfm_bad_permissions(struct task_struct *task)
2381 const struct cred *tcred;
2382 kuid_t uid = current_uid();
2383 kgid_t gid = current_gid();
2387 tcred = __task_cred(task);
2389 /* inspired by ptrace_attach() */
2390 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2391 from_kuid(&init_user_ns, uid),
2392 from_kgid(&init_user_ns, gid),
2393 from_kuid(&init_user_ns, tcred->euid),
2394 from_kuid(&init_user_ns, tcred->suid),
2395 from_kuid(&init_user_ns, tcred->uid),
2396 from_kgid(&init_user_ns, tcred->egid),
2397 from_kgid(&init_user_ns, tcred->sgid)));
2399 ret = ((!uid_eq(uid, tcred->euid))
2400 || (!uid_eq(uid, tcred->suid))
2401 || (!uid_eq(uid, tcred->uid))
2402 || (!gid_eq(gid, tcred->egid))
2403 || (!gid_eq(gid, tcred->sgid))
2404 || (!gid_eq(gid, tcred->gid))) && !capable(CAP_SYS_PTRACE);
2411 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2417 ctx_flags = pfx->ctx_flags;
2419 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2422 * cannot block in this mode
2424 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2425 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2430 /* probably more to add here */
2436 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2437 unsigned int cpu, pfarg_context_t *arg)
2439 pfm_buffer_fmt_t *fmt = NULL;
2440 unsigned long size = 0UL;
2442 void *fmt_arg = NULL;
2444 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2446 /* invoke and lock buffer format, if found */
2447 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2449 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2454 * buffer argument MUST be contiguous to pfarg_context_t
2456 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2458 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2460 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2462 if (ret) goto error;
2464 /* link buffer format and context */
2465 ctx->ctx_buf_fmt = fmt;
2466 ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2469 * check if buffer format wants to use perfmon buffer allocation/mapping service
2471 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2472 if (ret) goto error;
2476 * buffer is always remapped into the caller's address space
2478 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2479 if (ret) goto error;
2481 /* keep track of user address of buffer */
2482 arg->ctx_smpl_vaddr = uaddr;
2484 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2491 pfm_reset_pmu_state(pfm_context_t *ctx)
2496 * install reset values for PMC.
2498 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2499 if (PMC_IS_IMPL(i) == 0) continue;
2500 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2501 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2504 * PMD registers are set to 0UL when the context in memset()
2508 * On context switched restore, we must restore ALL pmc and ALL pmd even
2509 * when they are not actively used by the task. In UP, the incoming process
2510 * may otherwise pick up left over PMC, PMD state from the previous process.
2511 * As opposed to PMD, stale PMC can cause harm to the incoming
2512 * process because they may change what is being measured.
2513 * Therefore, we must systematically reinstall the entire
2514 * PMC state. In SMP, the same thing is possible on the
2515 * same CPU but also on between 2 CPUs.
2517 * The problem with PMD is information leaking especially
2518 * to user level when psr.sp=0
2520 * There is unfortunately no easy way to avoid this problem
2521 * on either UP or SMP. This definitively slows down the
2522 * pfm_load_regs() function.
2526 * bitmask of all PMCs accessible to this context
2528 * PMC0 is treated differently.
2530 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2533 * bitmask of all PMDs that are accessible to this context
2535 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2537 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2540 * useful in case of re-enable after disable
2542 ctx->ctx_used_ibrs[0] = 0UL;
2543 ctx->ctx_used_dbrs[0] = 0UL;
2547 pfm_ctx_getsize(void *arg, size_t *sz)
2549 pfarg_context_t *req = (pfarg_context_t *)arg;
2550 pfm_buffer_fmt_t *fmt;
2554 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2556 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2558 DPRINT(("cannot find buffer format\n"));
2561 /* get just enough to copy in user parameters */
2562 *sz = fmt->fmt_arg_size;
2563 DPRINT(("arg_size=%lu\n", *sz));
2571 * cannot attach if :
2573 * - task not owned by caller
2574 * - task incompatible with context mode
2577 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2580 * no kernel task or task not owner by caller
2582 if (task->mm == NULL) {
2583 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2586 if (pfm_bad_permissions(task)) {
2587 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task)));
2591 * cannot block in self-monitoring mode
2593 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2594 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2598 if (task->exit_state == EXIT_ZOMBIE) {
2599 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task)));
2604 * always ok for self
2606 if (task == current) return 0;
2608 if (!task_is_stopped_or_traced(task)) {
2609 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2613 * make sure the task is off any CPU
2615 wait_task_inactive(task, 0);
2617 /* more to come... */
2623 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2625 struct task_struct *p = current;
2628 /* XXX: need to add more checks here */
2629 if (pid < 2) return -EPERM;
2631 if (pid != task_pid_vnr(current)) {
2633 read_lock(&tasklist_lock);
2635 p = find_task_by_vpid(pid);
2637 /* make sure task cannot go away while we operate on it */
2638 if (p) get_task_struct(p);
2640 read_unlock(&tasklist_lock);
2642 if (p == NULL) return -ESRCH;
2645 ret = pfm_task_incompatible(ctx, p);
2648 } else if (p != current) {
2657 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2659 pfarg_context_t *req = (pfarg_context_t *)arg;
2666 /* let's check the arguments first */
2667 ret = pfarg_is_sane(current, req);
2671 ctx_flags = req->ctx_flags;
2675 fd = get_unused_fd();
2679 ctx = pfm_context_alloc(ctx_flags);
2683 filp = pfm_alloc_file(ctx);
2685 ret = PTR_ERR(filp);
2689 req->ctx_fd = ctx->ctx_fd = fd;
2692 * does the user want to sample?
2694 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2695 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2700 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2705 ctx->ctx_fl_excl_idle,
2710 * initialize soft PMU state
2712 pfm_reset_pmu_state(ctx);
2714 fd_install(fd, filp);
2719 path = filp->f_path;
2723 if (ctx->ctx_buf_fmt) {
2724 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2727 pfm_context_free(ctx);
2734 static inline unsigned long
2735 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2737 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2738 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2739 extern unsigned long carta_random32 (unsigned long seed);
2741 if (reg->flags & PFM_REGFL_RANDOM) {
2742 new_seed = carta_random32(old_seed);
2743 val -= (old_seed & mask); /* counter values are negative numbers! */
2744 if ((mask >> 32) != 0)
2745 /* construct a full 64-bit random value: */
2746 new_seed |= carta_random32(old_seed >> 32) << 32;
2747 reg->seed = new_seed;
2754 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2756 unsigned long mask = ovfl_regs[0];
2757 unsigned long reset_others = 0UL;
2762 * now restore reset value on sampling overflowed counters
2764 mask >>= PMU_FIRST_COUNTER;
2765 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2767 if ((mask & 0x1UL) == 0UL) continue;
2769 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2770 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2772 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2776 * Now take care of resetting the other registers
2778 for(i = 0; reset_others; i++, reset_others >>= 1) {
2780 if ((reset_others & 0x1) == 0) continue;
2782 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2784 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2785 is_long_reset ? "long" : "short", i, val));
2790 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2792 unsigned long mask = ovfl_regs[0];
2793 unsigned long reset_others = 0UL;
2797 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2799 if (ctx->ctx_state == PFM_CTX_MASKED) {
2800 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2805 * now restore reset value on sampling overflowed counters
2807 mask >>= PMU_FIRST_COUNTER;
2808 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2810 if ((mask & 0x1UL) == 0UL) continue;
2812 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2813 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2815 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2817 pfm_write_soft_counter(ctx, i, val);
2821 * Now take care of resetting the other registers
2823 for(i = 0; reset_others; i++, reset_others >>= 1) {
2825 if ((reset_others & 0x1) == 0) continue;
2827 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2829 if (PMD_IS_COUNTING(i)) {
2830 pfm_write_soft_counter(ctx, i, val);
2832 ia64_set_pmd(i, val);
2834 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2835 is_long_reset ? "long" : "short", i, val));
2841 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2843 struct task_struct *task;
2844 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2845 unsigned long value, pmc_pm;
2846 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2847 unsigned int cnum, reg_flags, flags, pmc_type;
2848 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2849 int is_monitor, is_counting, state;
2851 pfm_reg_check_t wr_func;
2852 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2854 state = ctx->ctx_state;
2855 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2856 is_system = ctx->ctx_fl_system;
2857 task = ctx->ctx_task;
2858 impl_pmds = pmu_conf->impl_pmds[0];
2860 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2864 * In system wide and when the context is loaded, access can only happen
2865 * when the caller is running on the CPU being monitored by the session.
2866 * It does not have to be the owner (ctx_task) of the context per se.
2868 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2869 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2872 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2874 expert_mode = pfm_sysctl.expert_mode;
2876 for (i = 0; i < count; i++, req++) {
2878 cnum = req->reg_num;
2879 reg_flags = req->reg_flags;
2880 value = req->reg_value;
2881 smpl_pmds = req->reg_smpl_pmds[0];
2882 reset_pmds = req->reg_reset_pmds[0];
2886 if (cnum >= PMU_MAX_PMCS) {
2887 DPRINT(("pmc%u is invalid\n", cnum));
2891 pmc_type = pmu_conf->pmc_desc[cnum].type;
2892 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2893 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2894 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2897 * we reject all non implemented PMC as well
2898 * as attempts to modify PMC[0-3] which are used
2899 * as status registers by the PMU
2901 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2902 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2905 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2907 * If the PMC is a monitor, then if the value is not the default:
2908 * - system-wide session: PMCx.pm=1 (privileged monitor)
2909 * - per-task : PMCx.pm=0 (user monitor)
2911 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2912 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2921 * enforce generation of overflow interrupt. Necessary on all
2924 value |= 1 << PMU_PMC_OI;
2926 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2927 flags |= PFM_REGFL_OVFL_NOTIFY;
2930 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2932 /* verify validity of smpl_pmds */
2933 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2934 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2938 /* verify validity of reset_pmds */
2939 if ((reset_pmds & impl_pmds) != reset_pmds) {
2940 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2944 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2945 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2948 /* eventid on non-counting monitors are ignored */
2952 * execute write checker, if any
2954 if (likely(expert_mode == 0 && wr_func)) {
2955 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2956 if (ret) goto error;
2961 * no error on this register
2963 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2966 * Now we commit the changes to the software state
2970 * update overflow information
2974 * full flag update each time a register is programmed
2976 ctx->ctx_pmds[cnum].flags = flags;
2978 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2979 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2980 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2983 * Mark all PMDS to be accessed as used.
2985 * We do not keep track of PMC because we have to
2986 * systematically restore ALL of them.
2988 * We do not update the used_monitors mask, because
2989 * if we have not programmed them, then will be in
2990 * a quiescent state, therefore we will not need to
2991 * mask/restore then when context is MASKED.
2993 CTX_USED_PMD(ctx, reset_pmds);
2994 CTX_USED_PMD(ctx, smpl_pmds);
2996 * make sure we do not try to reset on
2997 * restart because we have established new values
2999 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3002 * Needed in case the user does not initialize the equivalent
3003 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3004 * possible leak here.
3006 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3009 * keep track of the monitor PMC that we are using.
3010 * we save the value of the pmc in ctx_pmcs[] and if
3011 * the monitoring is not stopped for the context we also
3012 * place it in the saved state area so that it will be
3013 * picked up later by the context switch code.
3015 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3017 * The value in th_pmcs[] may be modified on overflow, i.e., when
3018 * monitoring needs to be stopped.
3020 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3023 * update context state
3025 ctx->ctx_pmcs[cnum] = value;
3029 * write thread state
3031 if (is_system == 0) ctx->th_pmcs[cnum] = value;
3034 * write hardware register if we can
3036 if (can_access_pmu) {
3037 ia64_set_pmc(cnum, value);
3042 * per-task SMP only here
3044 * we are guaranteed that the task is not running on the other CPU,
3045 * we indicate that this PMD will need to be reloaded if the task
3046 * is rescheduled on the CPU it ran last on.
3048 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3053 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",
3059 ctx->ctx_all_pmcs[0],
3060 ctx->ctx_used_pmds[0],
3061 ctx->ctx_pmds[cnum].eventid,
3064 ctx->ctx_reload_pmcs[0],
3065 ctx->ctx_used_monitors[0],
3066 ctx->ctx_ovfl_regs[0]));
3070 * make sure the changes are visible
3072 if (can_access_pmu) ia64_srlz_d();
3076 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3081 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3083 struct task_struct *task;
3084 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3085 unsigned long value, hw_value, ovfl_mask;
3087 int i, can_access_pmu = 0, state;
3088 int is_counting, is_loaded, is_system, expert_mode;
3090 pfm_reg_check_t wr_func;
3093 state = ctx->ctx_state;
3094 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3095 is_system = ctx->ctx_fl_system;
3096 ovfl_mask = pmu_conf->ovfl_val;
3097 task = ctx->ctx_task;
3099 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3102 * on both UP and SMP, we can only write to the PMC when the task is
3103 * the owner of the local PMU.
3105 if (likely(is_loaded)) {
3107 * In system wide and when the context is loaded, access can only happen
3108 * when the caller is running on the CPU being monitored by the session.
3109 * It does not have to be the owner (ctx_task) of the context per se.
3111 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3112 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3115 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3117 expert_mode = pfm_sysctl.expert_mode;
3119 for (i = 0; i < count; i++, req++) {
3121 cnum = req->reg_num;
3122 value = req->reg_value;
3124 if (!PMD_IS_IMPL(cnum)) {
3125 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3128 is_counting = PMD_IS_COUNTING(cnum);
3129 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3132 * execute write checker, if any
3134 if (unlikely(expert_mode == 0 && wr_func)) {
3135 unsigned long v = value;
3137 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3138 if (ret) goto abort_mission;
3145 * no error on this register
3147 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3150 * now commit changes to software state
3155 * update virtualized (64bits) counter
3159 * write context state
3161 ctx->ctx_pmds[cnum].lval = value;
3164 * when context is load we use the split value
3167 hw_value = value & ovfl_mask;
3168 value = value & ~ovfl_mask;
3172 * update reset values (not just for counters)
3174 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3175 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3178 * update randomization parameters (not just for counters)
3180 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3181 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3184 * update context value
3186 ctx->ctx_pmds[cnum].val = value;
3189 * Keep track of what we use
3191 * We do not keep track of PMC because we have to
3192 * systematically restore ALL of them.
3194 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3197 * mark this PMD register used as well
3199 CTX_USED_PMD(ctx, RDEP(cnum));
3202 * make sure we do not try to reset on
3203 * restart because we have established new values
3205 if (is_counting && state == PFM_CTX_MASKED) {
3206 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3211 * write thread state
3213 if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3216 * write hardware register if we can
3218 if (can_access_pmu) {
3219 ia64_set_pmd(cnum, hw_value);
3223 * we are guaranteed that the task is not running on the other CPU,
3224 * we indicate that this PMD will need to be reloaded if the task
3225 * is rescheduled on the CPU it ran last on.
3227 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3232 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3233 "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",
3239 ctx->ctx_pmds[cnum].val,
3240 ctx->ctx_pmds[cnum].short_reset,
3241 ctx->ctx_pmds[cnum].long_reset,
3242 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3243 ctx->ctx_pmds[cnum].seed,
3244 ctx->ctx_pmds[cnum].mask,
3245 ctx->ctx_used_pmds[0],
3246 ctx->ctx_pmds[cnum].reset_pmds[0],
3247 ctx->ctx_reload_pmds[0],
3248 ctx->ctx_all_pmds[0],
3249 ctx->ctx_ovfl_regs[0]));
3253 * make changes visible
3255 if (can_access_pmu) ia64_srlz_d();
3261 * for now, we have only one possibility for error
3263 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3268 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3269 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3270 * interrupt is delivered during the call, it will be kept pending until we leave, making
3271 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3272 * guaranteed to return consistent data to the user, it may simply be old. It is not
3273 * trivial to treat the overflow while inside the call because you may end up in
3274 * some module sampling buffer code causing deadlocks.
3277 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3279 struct task_struct *task;
3280 unsigned long val = 0UL, lval, ovfl_mask, sval;
3281 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3282 unsigned int cnum, reg_flags = 0;
3283 int i, can_access_pmu = 0, state;
3284 int is_loaded, is_system, is_counting, expert_mode;
3286 pfm_reg_check_t rd_func;
3289 * access is possible when loaded only for
3290 * self-monitoring tasks or in UP mode
3293 state = ctx->ctx_state;
3294 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3295 is_system = ctx->ctx_fl_system;
3296 ovfl_mask = pmu_conf->ovfl_val;
3297 task = ctx->ctx_task;
3299 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3301 if (likely(is_loaded)) {
3303 * In system wide and when the context is loaded, access can only happen
3304 * when the caller is running on the CPU being monitored by the session.
3305 * It does not have to be the owner (ctx_task) of the context per se.
3307 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3308 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3312 * this can be true when not self-monitoring only in UP
3314 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3316 if (can_access_pmu) ia64_srlz_d();
3318 expert_mode = pfm_sysctl.expert_mode;
3320 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3326 * on both UP and SMP, we can only read the PMD from the hardware register when
3327 * the task is the owner of the local PMU.
3330 for (i = 0; i < count; i++, req++) {
3332 cnum = req->reg_num;
3333 reg_flags = req->reg_flags;
3335 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3337 * we can only read the register that we use. That includes
3338 * the one we explicitly initialize AND the one we want included
3339 * in the sampling buffer (smpl_regs).
3341 * Having this restriction allows optimization in the ctxsw routine
3342 * without compromising security (leaks)
3344 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3346 sval = ctx->ctx_pmds[cnum].val;
3347 lval = ctx->ctx_pmds[cnum].lval;
3348 is_counting = PMD_IS_COUNTING(cnum);
3351 * If the task is not the current one, then we check if the
3352 * PMU state is still in the local live register due to lazy ctxsw.
3353 * If true, then we read directly from the registers.
3355 if (can_access_pmu){
3356 val = ia64_get_pmd(cnum);
3359 * context has been saved
3360 * if context is zombie, then task does not exist anymore.
3361 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3363 val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3365 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3369 * XXX: need to check for overflow when loaded
3376 * execute read checker, if any
3378 if (unlikely(expert_mode == 0 && rd_func)) {
3379 unsigned long v = val;
3380 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3381 if (ret) goto error;
3386 PFM_REG_RETFLAG_SET(reg_flags, 0);
3388 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3391 * update register return value, abort all if problem during copy.
3392 * we only modify the reg_flags field. no check mode is fine because
3393 * access has been verified upfront in sys_perfmonctl().
3395 req->reg_value = val;
3396 req->reg_flags = reg_flags;
3397 req->reg_last_reset_val = lval;
3403 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3408 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3412 if (req == NULL) return -EINVAL;
3414 ctx = GET_PMU_CTX();
3416 if (ctx == NULL) return -EINVAL;
3419 * for now limit to current task, which is enough when calling
3420 * from overflow handler
3422 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3424 return pfm_write_pmcs(ctx, req, nreq, regs);
3426 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3429 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3433 if (req == NULL) return -EINVAL;
3435 ctx = GET_PMU_CTX();
3437 if (ctx == NULL) return -EINVAL;
3440 * for now limit to current task, which is enough when calling
3441 * from overflow handler
3443 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3445 return pfm_read_pmds(ctx, req, nreq, regs);
3447 EXPORT_SYMBOL(pfm_mod_read_pmds);
3450 * Only call this function when a process it trying to
3451 * write the debug registers (reading is always allowed)
3454 pfm_use_debug_registers(struct task_struct *task)
3456 pfm_context_t *ctx = task->thread.pfm_context;
3457 unsigned long flags;
3460 if (pmu_conf->use_rr_dbregs == 0) return 0;
3462 DPRINT(("called for [%d]\n", task_pid_nr(task)));
3467 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3470 * Even on SMP, we do not need to use an atomic here because
3471 * the only way in is via ptrace() and this is possible only when the
3472 * process is stopped. Even in the case where the ctxsw out is not totally
3473 * completed by the time we come here, there is no way the 'stopped' process
3474 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3475 * So this is always safe.
3477 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3482 * We cannot allow setting breakpoints when system wide monitoring
3483 * sessions are using the debug registers.
3485 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3488 pfm_sessions.pfs_ptrace_use_dbregs++;
3490 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3491 pfm_sessions.pfs_ptrace_use_dbregs,
3492 pfm_sessions.pfs_sys_use_dbregs,
3493 task_pid_nr(task), ret));
3501 * This function is called for every task that exits with the
3502 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3503 * able to use the debug registers for debugging purposes via
3504 * ptrace(). Therefore we know it was not using them for
3505 * performance monitoring, so we only decrement the number
3506 * of "ptraced" debug register users to keep the count up to date
3509 pfm_release_debug_registers(struct task_struct *task)
3511 unsigned long flags;
3514 if (pmu_conf->use_rr_dbregs == 0) return 0;
3517 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3518 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3521 pfm_sessions.pfs_ptrace_use_dbregs--;
3530 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3532 struct task_struct *task;
3533 pfm_buffer_fmt_t *fmt;
3534 pfm_ovfl_ctrl_t rst_ctrl;
3535 int state, is_system;
3538 state = ctx->ctx_state;
3539 fmt = ctx->ctx_buf_fmt;
3540 is_system = ctx->ctx_fl_system;
3541 task = PFM_CTX_TASK(ctx);
3544 case PFM_CTX_MASKED:
3546 case PFM_CTX_LOADED:
3547 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3549 case PFM_CTX_UNLOADED:
3550 case PFM_CTX_ZOMBIE:
3551 DPRINT(("invalid state=%d\n", state));
3554 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3559 * In system wide and when the context is loaded, access can only happen
3560 * when the caller is running on the CPU being monitored by the session.
3561 * It does not have to be the owner (ctx_task) of the context per se.
3563 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3564 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3569 if (unlikely(task == NULL)) {
3570 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3574 if (task == current || is_system) {
3576 fmt = ctx->ctx_buf_fmt;
3578 DPRINT(("restarting self %d ovfl=0x%lx\n",
3580 ctx->ctx_ovfl_regs[0]));
3582 if (CTX_HAS_SMPL(ctx)) {
3584 prefetch(ctx->ctx_smpl_hdr);
3586 rst_ctrl.bits.mask_monitoring = 0;
3587 rst_ctrl.bits.reset_ovfl_pmds = 0;
3589 if (state == PFM_CTX_LOADED)
3590 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3592 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3594 rst_ctrl.bits.mask_monitoring = 0;
3595 rst_ctrl.bits.reset_ovfl_pmds = 1;
3599 if (rst_ctrl.bits.reset_ovfl_pmds)
3600 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3602 if (rst_ctrl.bits.mask_monitoring == 0) {
3603 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3605 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3607 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3609 // cannot use pfm_stop_monitoring(task, regs);
3613 * clear overflowed PMD mask to remove any stale information
3615 ctx->ctx_ovfl_regs[0] = 0UL;
3618 * back to LOADED state
3620 ctx->ctx_state = PFM_CTX_LOADED;
3623 * XXX: not really useful for self monitoring
3625 ctx->ctx_fl_can_restart = 0;
3631 * restart another task
3635 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3636 * one is seen by the task.
3638 if (state == PFM_CTX_MASKED) {
3639 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3641 * will prevent subsequent restart before this one is
3642 * seen by other task
3644 ctx->ctx_fl_can_restart = 0;
3648 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3649 * the task is blocked or on its way to block. That's the normal
3650 * restart path. If the monitoring is not masked, then the task
3651 * can be actively monitoring and we cannot directly intervene.
3652 * Therefore we use the trap mechanism to catch the task and
3653 * force it to reset the buffer/reset PMDs.
3655 * if non-blocking, then we ensure that the task will go into
3656 * pfm_handle_work() before returning to user mode.
3658 * We cannot explicitly reset another task, it MUST always
3659 * be done by the task itself. This works for system wide because
3660 * the tool that is controlling the session is logically doing
3661 * "self-monitoring".
3663 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3664 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3665 complete(&ctx->ctx_restart_done);
3667 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3669 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3671 PFM_SET_WORK_PENDING(task, 1);
3673 set_notify_resume(task);
3676 * XXX: send reschedule if task runs on another CPU
3683 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3685 unsigned int m = *(unsigned int *)arg;
3687 pfm_sysctl.debug = m == 0 ? 0 : 1;
3689 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3692 memset(pfm_stats, 0, sizeof(pfm_stats));
3693 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3699 * arg can be NULL and count can be zero for this function
3702 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3704 struct thread_struct *thread = NULL;
3705 struct task_struct *task;
3706 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3707 unsigned long flags;
3712 int i, can_access_pmu = 0;
3713 int is_system, is_loaded;
3715 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3717 state = ctx->ctx_state;
3718 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3719 is_system = ctx->ctx_fl_system;
3720 task = ctx->ctx_task;
3722 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3725 * on both UP and SMP, we can only write to the PMC when the task is
3726 * the owner of the local PMU.
3729 thread = &task->thread;
3731 * In system wide and when the context is loaded, access can only happen
3732 * when the caller is running on the CPU being monitored by the session.
3733 * It does not have to be the owner (ctx_task) of the context per se.
3735 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3736 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3739 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3743 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3744 * ensuring that no real breakpoint can be installed via this call.
3746 * IMPORTANT: regs can be NULL in this function
3749 first_time = ctx->ctx_fl_using_dbreg == 0;
3752 * don't bother if we are loaded and task is being debugged
3754 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3755 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3760 * check for debug registers in system wide mode
3762 * If though a check is done in pfm_context_load(),
3763 * we must repeat it here, in case the registers are
3764 * written after the context is loaded
3769 if (first_time && is_system) {
3770 if (pfm_sessions.pfs_ptrace_use_dbregs)
3773 pfm_sessions.pfs_sys_use_dbregs++;
3778 if (ret != 0) return ret;
3781 * mark ourself as user of the debug registers for
3784 ctx->ctx_fl_using_dbreg = 1;
3787 * clear hardware registers to make sure we don't
3788 * pick up stale state.
3790 * for a system wide session, we do not use
3791 * thread.dbr, thread.ibr because this process
3792 * never leaves the current CPU and the state
3793 * is shared by all processes running on it
3795 if (first_time && can_access_pmu) {
3796 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3797 for (i=0; i < pmu_conf->num_ibrs; i++) {
3798 ia64_set_ibr(i, 0UL);
3799 ia64_dv_serialize_instruction();
3802 for (i=0; i < pmu_conf->num_dbrs; i++) {
3803 ia64_set_dbr(i, 0UL);
3804 ia64_dv_serialize_data();
3810 * Now install the values into the registers
3812 for (i = 0; i < count; i++, req++) {
3814 rnum = req->dbreg_num;
3815 dbreg.val = req->dbreg_value;
3819 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3820 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3821 rnum, dbreg.val, mode, i, count));
3827 * make sure we do not install enabled breakpoint
3830 if (mode == PFM_CODE_RR)
3831 dbreg.ibr.ibr_x = 0;
3833 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3836 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3839 * Debug registers, just like PMC, can only be modified
3840 * by a kernel call. Moreover, perfmon() access to those
3841 * registers are centralized in this routine. The hardware
3842 * does not modify the value of these registers, therefore,
3843 * if we save them as they are written, we can avoid having
3844 * to save them on context switch out. This is made possible
3845 * by the fact that when perfmon uses debug registers, ptrace()
3846 * won't be able to modify them concurrently.
3848 if (mode == PFM_CODE_RR) {
3849 CTX_USED_IBR(ctx, rnum);
3851 if (can_access_pmu) {
3852 ia64_set_ibr(rnum, dbreg.val);
3853 ia64_dv_serialize_instruction();
3856 ctx->ctx_ibrs[rnum] = dbreg.val;
3858 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3859 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3861 CTX_USED_DBR(ctx, rnum);
3863 if (can_access_pmu) {
3864 ia64_set_dbr(rnum, dbreg.val);
3865 ia64_dv_serialize_data();
3867 ctx->ctx_dbrs[rnum] = dbreg.val;
3869 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3870 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3878 * in case it was our first attempt, we undo the global modifications
3882 if (ctx->ctx_fl_system) {
3883 pfm_sessions.pfs_sys_use_dbregs--;
3886 ctx->ctx_fl_using_dbreg = 0;
3889 * install error return flag
3891 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3897 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3899 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3903 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3905 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3909 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3913 if (req == NULL) return -EINVAL;
3915 ctx = GET_PMU_CTX();
3917 if (ctx == NULL) return -EINVAL;
3920 * for now limit to current task, which is enough when calling
3921 * from overflow handler
3923 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3925 return pfm_write_ibrs(ctx, req, nreq, regs);
3927 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3930 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3934 if (req == NULL) return -EINVAL;
3936 ctx = GET_PMU_CTX();
3938 if (ctx == NULL) return -EINVAL;
3941 * for now limit to current task, which is enough when calling
3942 * from overflow handler
3944 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3946 return pfm_write_dbrs(ctx, req, nreq, regs);
3948 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3952 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3954 pfarg_features_t *req = (pfarg_features_t *)arg;
3956 req->ft_version = PFM_VERSION;
3961 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3963 struct pt_regs *tregs;
3964 struct task_struct *task = PFM_CTX_TASK(ctx);
3965 int state, is_system;
3967 state = ctx->ctx_state;
3968 is_system = ctx->ctx_fl_system;
3971 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3973 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3976 * In system wide and when the context is loaded, access can only happen
3977 * when the caller is running on the CPU being monitored by the session.
3978 * It does not have to be the owner (ctx_task) of the context per se.
3980 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3981 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3984 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3985 task_pid_nr(PFM_CTX_TASK(ctx)),
3989 * in system mode, we need to update the PMU directly
3990 * and the user level state of the caller, which may not
3991 * necessarily be the creator of the context.
3995 * Update local PMU first
3999 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4003 * update local cpuinfo
4005 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4008 * stop monitoring, does srlz.i
4013 * stop monitoring in the caller
4015 ia64_psr(regs)->pp = 0;
4023 if (task == current) {
4024 /* stop monitoring at kernel level */
4028 * stop monitoring at the user level
4030 ia64_psr(regs)->up = 0;
4032 tregs = task_pt_regs(task);
4035 * stop monitoring at the user level
4037 ia64_psr(tregs)->up = 0;
4040 * monitoring disabled in kernel at next reschedule
4042 ctx->ctx_saved_psr_up = 0;
4043 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4050 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4052 struct pt_regs *tregs;
4053 int state, is_system;
4055 state = ctx->ctx_state;
4056 is_system = ctx->ctx_fl_system;
4058 if (state != PFM_CTX_LOADED) return -EINVAL;
4061 * In system wide and when the context is loaded, access can only happen
4062 * when the caller is running on the CPU being monitored by the session.
4063 * It does not have to be the owner (ctx_task) of the context per se.
4065 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4066 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4071 * in system mode, we need to update the PMU directly
4072 * and the user level state of the caller, which may not
4073 * necessarily be the creator of the context.
4078 * set user level psr.pp for the caller
4080 ia64_psr(regs)->pp = 1;
4083 * now update the local PMU and cpuinfo
4085 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4088 * start monitoring at kernel level
4093 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4103 if (ctx->ctx_task == current) {
4105 /* start monitoring at kernel level */
4109 * activate monitoring at user level
4111 ia64_psr(regs)->up = 1;
4114 tregs = task_pt_regs(ctx->ctx_task);
4117 * start monitoring at the kernel level the next
4118 * time the task is scheduled
4120 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4123 * activate monitoring at user level
4125 ia64_psr(tregs)->up = 1;
4131 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4133 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4138 for (i = 0; i < count; i++, req++) {
4140 cnum = req->reg_num;
4142 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4144 req->reg_value = PMC_DFL_VAL(cnum);
4146 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4148 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4153 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4158 pfm_check_task_exist(pfm_context_t *ctx)
4160 struct task_struct *g, *t;
4163 read_lock(&tasklist_lock);
4165 do_each_thread (g, t) {
4166 if (t->thread.pfm_context == ctx) {
4170 } while_each_thread (g, t);
4172 read_unlock(&tasklist_lock);
4174 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4180 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4182 struct task_struct *task;
4183 struct thread_struct *thread;
4184 struct pfm_context_t *old;
4185 unsigned long flags;
4187 struct task_struct *owner_task = NULL;
4189 pfarg_load_t *req = (pfarg_load_t *)arg;
4190 unsigned long *pmcs_source, *pmds_source;
4193 int state, is_system, set_dbregs = 0;
4195 state = ctx->ctx_state;
4196 is_system = ctx->ctx_fl_system;
4198 * can only load from unloaded or terminated state
4200 if (state != PFM_CTX_UNLOADED) {
4201 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4207 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4209 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4210 DPRINT(("cannot use blocking mode on self\n"));
4214 ret = pfm_get_task(ctx, req->load_pid, &task);
4216 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4223 * system wide is self monitoring only
4225 if (is_system && task != current) {
4226 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4231 thread = &task->thread;
4235 * cannot load a context which is using range restrictions,
4236 * into a task that is being debugged.
4238 if (ctx->ctx_fl_using_dbreg) {
4239 if (thread->flags & IA64_THREAD_DBG_VALID) {
4241 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4247 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4248 DPRINT(("cannot load [%d] dbregs in use\n",
4249 task_pid_nr(task)));
4252 pfm_sessions.pfs_sys_use_dbregs++;
4253 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4260 if (ret) goto error;
4264 * SMP system-wide monitoring implies self-monitoring.
4266 * The programming model expects the task to
4267 * be pinned on a CPU throughout the session.
4268 * Here we take note of the current CPU at the
4269 * time the context is loaded. No call from
4270 * another CPU will be allowed.
4272 * The pinning via shed_setaffinity()
4273 * must be done by the calling task prior
4276 * systemwide: keep track of CPU this session is supposed to run on
4278 the_cpu = ctx->ctx_cpu = smp_processor_id();
4282 * now reserve the session
4284 ret = pfm_reserve_session(current, is_system, the_cpu);
4285 if (ret) goto error;
4288 * task is necessarily stopped at this point.
4290 * If the previous context was zombie, then it got removed in
4291 * pfm_save_regs(). Therefore we should not see it here.
4292 * If we see a context, then this is an active context
4294 * XXX: needs to be atomic
4296 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4297 thread->pfm_context, ctx));
4300 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4302 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4306 pfm_reset_msgq(ctx);
4308 ctx->ctx_state = PFM_CTX_LOADED;
4311 * link context to task
4313 ctx->ctx_task = task;
4317 * we load as stopped
4319 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4320 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4322 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4324 thread->flags |= IA64_THREAD_PM_VALID;
4328 * propagate into thread-state
4330 pfm_copy_pmds(task, ctx);
4331 pfm_copy_pmcs(task, ctx);
4333 pmcs_source = ctx->th_pmcs;
4334 pmds_source = ctx->th_pmds;
4337 * always the case for system-wide
4339 if (task == current) {
4341 if (is_system == 0) {
4343 /* allow user level control */
4344 ia64_psr(regs)->sp = 0;
4345 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4347 SET_LAST_CPU(ctx, smp_processor_id());
4349 SET_ACTIVATION(ctx);
4352 * push the other task out, if any
4354 owner_task = GET_PMU_OWNER();
4355 if (owner_task) pfm_lazy_save_regs(owner_task);
4359 * load all PMD from ctx to PMU (as opposed to thread state)
4360 * restore all PMC from ctx to PMU
4362 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4363 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4365 ctx->ctx_reload_pmcs[0] = 0UL;
4366 ctx->ctx_reload_pmds[0] = 0UL;
4369 * guaranteed safe by earlier check against DBG_VALID
4371 if (ctx->ctx_fl_using_dbreg) {
4372 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4373 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4378 SET_PMU_OWNER(task, ctx);
4380 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4383 * when not current, task MUST be stopped, so this is safe
4385 regs = task_pt_regs(task);
4387 /* force a full reload */
4388 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4389 SET_LAST_CPU(ctx, -1);
4391 /* initial saved psr (stopped) */
4392 ctx->ctx_saved_psr_up = 0UL;
4393 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4399 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4402 * we must undo the dbregs setting (for system-wide)
4404 if (ret && set_dbregs) {
4406 pfm_sessions.pfs_sys_use_dbregs--;
4410 * release task, there is now a link with the context
4412 if (is_system == 0 && task != current) {
4416 ret = pfm_check_task_exist(ctx);
4418 ctx->ctx_state = PFM_CTX_UNLOADED;
4419 ctx->ctx_task = NULL;
4427 * in this function, we do not need to increase the use count
4428 * for the task via get_task_struct(), because we hold the
4429 * context lock. If the task were to disappear while having
4430 * a context attached, it would go through pfm_exit_thread()
4431 * which also grabs the context lock and would therefore be blocked
4432 * until we are here.
4434 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4437 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4439 struct task_struct *task = PFM_CTX_TASK(ctx);
4440 struct pt_regs *tregs;
4441 int prev_state, is_system;
4444 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4446 prev_state = ctx->ctx_state;
4447 is_system = ctx->ctx_fl_system;
4450 * unload only when necessary
4452 if (prev_state == PFM_CTX_UNLOADED) {
4453 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4458 * clear psr and dcr bits
4460 ret = pfm_stop(ctx, NULL, 0, regs);
4461 if (ret) return ret;
4463 ctx->ctx_state = PFM_CTX_UNLOADED;
4466 * in system mode, we need to update the PMU directly
4467 * and the user level state of the caller, which may not
4468 * necessarily be the creator of the context.
4475 * local PMU is taken care of in pfm_stop()
4477 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4478 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4481 * save PMDs in context
4484 pfm_flush_pmds(current, ctx);
4487 * at this point we are done with the PMU
4488 * so we can unreserve the resource.
4490 if (prev_state != PFM_CTX_ZOMBIE)
4491 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4494 * disconnect context from task
4496 task->thread.pfm_context = NULL;
4498 * disconnect task from context
4500 ctx->ctx_task = NULL;
4503 * There is nothing more to cleanup here.
4511 tregs = task == current ? regs : task_pt_regs(task);
4513 if (task == current) {
4515 * cancel user level control
4517 ia64_psr(regs)->sp = 1;
4519 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4522 * save PMDs to context
4525 pfm_flush_pmds(task, ctx);
4528 * at this point we are done with the PMU
4529 * so we can unreserve the resource.
4531 * when state was ZOMBIE, we have already unreserved.
4533 if (prev_state != PFM_CTX_ZOMBIE)
4534 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4537 * reset activation counter and psr
4539 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4540 SET_LAST_CPU(ctx, -1);
4543 * PMU state will not be restored
4545 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4548 * break links between context and task
4550 task->thread.pfm_context = NULL;
4551 ctx->ctx_task = NULL;
4553 PFM_SET_WORK_PENDING(task, 0);
4555 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4556 ctx->ctx_fl_can_restart = 0;
4557 ctx->ctx_fl_going_zombie = 0;
4559 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4566 * called only from exit_thread(): task == current
4567 * we come here only if current has a context attached (loaded or masked)
4570 pfm_exit_thread(struct task_struct *task)
4573 unsigned long flags;
4574 struct pt_regs *regs = task_pt_regs(task);
4578 ctx = PFM_GET_CTX(task);
4580 PROTECT_CTX(ctx, flags);
4582 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4584 state = ctx->ctx_state;
4586 case PFM_CTX_UNLOADED:
4588 * only comes to this function if pfm_context is not NULL, i.e., cannot
4589 * be in unloaded state
4591 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4593 case PFM_CTX_LOADED:
4594 case PFM_CTX_MASKED:
4595 ret = pfm_context_unload(ctx, NULL, 0, regs);
4597 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4599 DPRINT(("ctx unloaded for current state was %d\n", state));
4601 pfm_end_notify_user(ctx);
4603 case PFM_CTX_ZOMBIE:
4604 ret = pfm_context_unload(ctx, NULL, 0, regs);
4606 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4611 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4614 UNPROTECT_CTX(ctx, flags);
4616 { u64 psr = pfm_get_psr();
4617 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4618 BUG_ON(GET_PMU_OWNER());
4619 BUG_ON(ia64_psr(regs)->up);
4620 BUG_ON(ia64_psr(regs)->pp);
4624 * All memory free operations (especially for vmalloc'ed memory)
4625 * MUST be done with interrupts ENABLED.
4627 if (free_ok) pfm_context_free(ctx);
4631 * functions MUST be listed in the increasing order of their index (see permfon.h)
4633 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4634 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4635 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4636 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4637 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4639 static pfm_cmd_desc_t pfm_cmd_tab[]={
4640 /* 0 */PFM_CMD_NONE,
4641 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4642 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4643 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4644 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4645 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4646 /* 6 */PFM_CMD_NONE,
4647 /* 7 */PFM_CMD_NONE,
4648 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4649 /* 9 */PFM_CMD_NONE,
4650 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4651 /* 11 */PFM_CMD_NONE,
4652 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4653 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4654 /* 14 */PFM_CMD_NONE,
4655 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4656 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4657 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4658 /* 18 */PFM_CMD_NONE,
4659 /* 19 */PFM_CMD_NONE,
4660 /* 20 */PFM_CMD_NONE,
4661 /* 21 */PFM_CMD_NONE,
4662 /* 22 */PFM_CMD_NONE,
4663 /* 23 */PFM_CMD_NONE,
4664 /* 24 */PFM_CMD_NONE,
4665 /* 25 */PFM_CMD_NONE,
4666 /* 26 */PFM_CMD_NONE,
4667 /* 27 */PFM_CMD_NONE,
4668 /* 28 */PFM_CMD_NONE,
4669 /* 29 */PFM_CMD_NONE,
4670 /* 30 */PFM_CMD_NONE,
4671 /* 31 */PFM_CMD_NONE,
4672 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4673 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4675 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4678 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4680 struct task_struct *task;
4681 int state, old_state;
4684 state = ctx->ctx_state;
4685 task = ctx->ctx_task;
4688 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4692 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4696 task->state, PFM_CMD_STOPPED(cmd)));
4699 * self-monitoring always ok.
4701 * for system-wide the caller can either be the creator of the
4702 * context (to one to which the context is attached to) OR
4703 * a task running on the same CPU as the session.
4705 if (task == current || ctx->ctx_fl_system) return 0;
4708 * we are monitoring another thread
4711 case PFM_CTX_UNLOADED:
4713 * if context is UNLOADED we are safe to go
4716 case PFM_CTX_ZOMBIE:
4718 * no command can operate on a zombie context
4720 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4722 case PFM_CTX_MASKED:
4724 * PMU state has been saved to software even though
4725 * the thread may still be running.
4727 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4731 * context is LOADED or MASKED. Some commands may need to have
4734 * We could lift this restriction for UP but it would mean that
4735 * the user has no guarantee the task would not run between
4736 * two successive calls to perfmonctl(). That's probably OK.
4737 * If this user wants to ensure the task does not run, then
4738 * the task must be stopped.
4740 if (PFM_CMD_STOPPED(cmd)) {
4741 if (!task_is_stopped_or_traced(task)) {
4742 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4746 * task is now stopped, wait for ctxsw out
4748 * This is an interesting point in the code.
4749 * We need to unprotect the context because
4750 * the pfm_save_regs() routines needs to grab
4751 * the same lock. There are danger in doing
4752 * this because it leaves a window open for
4753 * another task to get access to the context
4754 * and possibly change its state. The one thing
4755 * that is not possible is for the context to disappear
4756 * because we are protected by the VFS layer, i.e.,
4757 * get_fd()/put_fd().
4761 UNPROTECT_CTX(ctx, flags);
4763 wait_task_inactive(task, 0);
4765 PROTECT_CTX(ctx, flags);
4768 * we must recheck to verify if state has changed
4770 if (ctx->ctx_state != old_state) {
4771 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4779 * system-call entry point (must return long)
4782 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4784 struct fd f = {NULL, 0};
4785 pfm_context_t *ctx = NULL;
4786 unsigned long flags = 0UL;
4787 void *args_k = NULL;
4788 long ret; /* will expand int return types */
4789 size_t base_sz, sz, xtra_sz = 0;
4790 int narg, completed_args = 0, call_made = 0, cmd_flags;
4791 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4792 int (*getsize)(void *arg, size_t *sz);
4793 #define PFM_MAX_ARGSIZE 4096
4796 * reject any call if perfmon was disabled at initialization
4798 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4800 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4801 DPRINT(("invalid cmd=%d\n", cmd));
4805 func = pfm_cmd_tab[cmd].cmd_func;
4806 narg = pfm_cmd_tab[cmd].cmd_narg;
4807 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4808 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4809 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4811 if (unlikely(func == NULL)) {
4812 DPRINT(("invalid cmd=%d\n", cmd));
4816 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4824 * check if number of arguments matches what the command expects
4826 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4830 sz = xtra_sz + base_sz*count;
4832 * limit abuse to min page size
4834 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4835 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4840 * allocate default-sized argument buffer
4842 if (likely(count && args_k == NULL)) {
4843 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4844 if (args_k == NULL) return -ENOMEM;
4852 * assume sz = 0 for command without parameters
4854 if (sz && copy_from_user(args_k, arg, sz)) {
4855 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4860 * check if command supports extra parameters
4862 if (completed_args == 0 && getsize) {
4864 * get extra parameters size (based on main argument)
4866 ret = (*getsize)(args_k, &xtra_sz);
4867 if (ret) goto error_args;
4871 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4873 /* retry if necessary */
4874 if (likely(xtra_sz)) goto restart_args;
4877 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4882 if (unlikely(f.file == NULL)) {
4883 DPRINT(("invalid fd %d\n", fd));
4886 if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4887 DPRINT(("fd %d not related to perfmon\n", fd));
4891 ctx = f.file->private_data;
4892 if (unlikely(ctx == NULL)) {
4893 DPRINT(("no context for fd %d\n", fd));
4896 prefetch(&ctx->ctx_state);
4898 PROTECT_CTX(ctx, flags);
4901 * check task is stopped
4903 ret = pfm_check_task_state(ctx, cmd, flags);
4904 if (unlikely(ret)) goto abort_locked;
4907 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4913 DPRINT(("context unlocked\n"));
4914 UNPROTECT_CTX(ctx, flags);
4917 /* copy argument back to user, if needed */
4918 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4926 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4932 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4934 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4935 pfm_ovfl_ctrl_t rst_ctrl;
4939 state = ctx->ctx_state;
4941 * Unlock sampling buffer and reset index atomically
4942 * XXX: not really needed when blocking
4944 if (CTX_HAS_SMPL(ctx)) {
4946 rst_ctrl.bits.mask_monitoring = 0;
4947 rst_ctrl.bits.reset_ovfl_pmds = 0;
4949 if (state == PFM_CTX_LOADED)
4950 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4952 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4954 rst_ctrl.bits.mask_monitoring = 0;
4955 rst_ctrl.bits.reset_ovfl_pmds = 1;
4959 if (rst_ctrl.bits.reset_ovfl_pmds) {
4960 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4962 if (rst_ctrl.bits.mask_monitoring == 0) {
4963 DPRINT(("resuming monitoring\n"));
4964 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4966 DPRINT(("stopping monitoring\n"));
4967 //pfm_stop_monitoring(current, regs);
4969 ctx->ctx_state = PFM_CTX_LOADED;
4974 * context MUST BE LOCKED when calling
4975 * can only be called for current
4978 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4982 DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4984 ret = pfm_context_unload(ctx, NULL, 0, regs);
4986 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4990 * and wakeup controlling task, indicating we are now disconnected
4992 wake_up_interruptible(&ctx->ctx_zombieq);
4995 * given that context is still locked, the controlling
4996 * task will only get access when we return from
4997 * pfm_handle_work().
5001 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5004 * pfm_handle_work() can be called with interrupts enabled
5005 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5006 * call may sleep, therefore we must re-enable interrupts
5007 * to avoid deadlocks. It is safe to do so because this function
5008 * is called ONLY when returning to user level (pUStk=1), in which case
5009 * there is no risk of kernel stack overflow due to deep
5010 * interrupt nesting.
5013 pfm_handle_work(void)
5016 struct pt_regs *regs;
5017 unsigned long flags, dummy_flags;
5018 unsigned long ovfl_regs;
5019 unsigned int reason;
5022 ctx = PFM_GET_CTX(current);
5024 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5025 task_pid_nr(current));
5029 PROTECT_CTX(ctx, flags);
5031 PFM_SET_WORK_PENDING(current, 0);
5033 regs = task_pt_regs(current);
5036 * extract reason for being here and clear
5038 reason = ctx->ctx_fl_trap_reason;
5039 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5040 ovfl_regs = ctx->ctx_ovfl_regs[0];
5042 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5045 * must be done before we check for simple-reset mode
5047 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5050 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5051 if (reason == PFM_TRAP_REASON_RESET)
5055 * restore interrupt mask to what it was on entry.
5056 * Could be enabled/diasbled.
5058 UNPROTECT_CTX(ctx, flags);
5061 * force interrupt enable because of down_interruptible()
5065 DPRINT(("before block sleeping\n"));
5068 * may go through without blocking on SMP systems
5069 * if restart has been received already by the time we call down()
5071 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5073 DPRINT(("after block sleeping ret=%d\n", ret));
5076 * lock context and mask interrupts again
5077 * We save flags into a dummy because we may have
5078 * altered interrupts mask compared to entry in this
5081 PROTECT_CTX(ctx, dummy_flags);
5084 * we need to read the ovfl_regs only after wake-up
5085 * because we may have had pfm_write_pmds() in between
5086 * and that can changed PMD values and therefore
5087 * ovfl_regs is reset for these new PMD values.
5089 ovfl_regs = ctx->ctx_ovfl_regs[0];
5091 if (ctx->ctx_fl_going_zombie) {
5093 DPRINT(("context is zombie, bailing out\n"));
5094 pfm_context_force_terminate(ctx, regs);
5098 * in case of interruption of down() we don't restart anything
5104 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5105 ctx->ctx_ovfl_regs[0] = 0UL;
5109 * restore flags as they were upon entry
5111 UNPROTECT_CTX(ctx, flags);
5115 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5117 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5118 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5122 DPRINT(("waking up somebody\n"));
5124 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5127 * safe, we are not in intr handler, nor in ctxsw when
5130 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5136 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5138 pfm_msg_t *msg = NULL;
5140 if (ctx->ctx_fl_no_msg == 0) {
5141 msg = pfm_get_new_msg(ctx);
5143 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5147 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5148 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5149 msg->pfm_ovfl_msg.msg_active_set = 0;
5150 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5151 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5152 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5153 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5154 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5157 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5163 return pfm_notify_user(ctx, msg);
5167 pfm_end_notify_user(pfm_context_t *ctx)
5171 msg = pfm_get_new_msg(ctx);
5173 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5177 memset(msg, 0, sizeof(*msg));
5179 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5180 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5181 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5183 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5188 return pfm_notify_user(ctx, msg);
5192 * main overflow processing routine.
5193 * it can be called from the interrupt path or explicitly during the context switch code
5195 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5196 unsigned long pmc0, struct pt_regs *regs)
5198 pfm_ovfl_arg_t *ovfl_arg;
5200 unsigned long old_val, ovfl_val, new_val;
5201 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5202 unsigned long tstamp;
5203 pfm_ovfl_ctrl_t ovfl_ctrl;
5204 unsigned int i, has_smpl;
5205 int must_notify = 0;
5207 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5210 * sanity test. Should never happen
5212 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5214 tstamp = ia64_get_itc();
5215 mask = pmc0 >> PMU_FIRST_COUNTER;
5216 ovfl_val = pmu_conf->ovfl_val;
5217 has_smpl = CTX_HAS_SMPL(ctx);
5219 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5220 "used_pmds=0x%lx\n",
5222 task ? task_pid_nr(task): -1,
5223 (regs ? regs->cr_iip : 0),
5224 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5225 ctx->ctx_used_pmds[0]));
5229 * first we update the virtual counters
5230 * assume there was a prior ia64_srlz_d() issued
5232 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5234 /* skip pmd which did not overflow */
5235 if ((mask & 0x1) == 0) continue;
5238 * Note that the pmd is not necessarily 0 at this point as qualified events
5239 * may have happened before the PMU was frozen. The residual count is not
5240 * taken into consideration here but will be with any read of the pmd via
5243 old_val = new_val = ctx->ctx_pmds[i].val;
5244 new_val += 1 + ovfl_val;
5245 ctx->ctx_pmds[i].val = new_val;
5248 * check for overflow condition
5250 if (likely(old_val > new_val)) {
5251 ovfl_pmds |= 1UL << i;
5252 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5255 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5259 ia64_get_pmd(i) & ovfl_val,
5265 * there was no 64-bit overflow, nothing else to do
5267 if (ovfl_pmds == 0UL) return;
5270 * reset all control bits
5276 * if a sampling format module exists, then we "cache" the overflow by
5277 * calling the module's handler() routine.
5280 unsigned long start_cycles, end_cycles;
5281 unsigned long pmd_mask;
5283 int this_cpu = smp_processor_id();
5285 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5286 ovfl_arg = &ctx->ctx_ovfl_arg;
5288 prefetch(ctx->ctx_smpl_hdr);
5290 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5294 if ((pmd_mask & 0x1) == 0) continue;
5296 ovfl_arg->ovfl_pmd = (unsigned char )i;
5297 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5298 ovfl_arg->active_set = 0;
5299 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5300 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5302 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5303 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5304 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5307 * copy values of pmds of interest. Sampling format may copy them
5308 * into sampling buffer.
5311 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5312 if ((smpl_pmds & 0x1) == 0) continue;
5313 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5314 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5318 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5320 start_cycles = ia64_get_itc();
5323 * call custom buffer format record (handler) routine
5325 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5327 end_cycles = ia64_get_itc();
5330 * For those controls, we take the union because they have
5331 * an all or nothing behavior.
5333 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5334 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5335 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5337 * build the bitmask of pmds to reset now
5339 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5341 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5344 * when the module cannot handle the rest of the overflows, we abort right here
5346 if (ret && pmd_mask) {
5347 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5348 pmd_mask<<PMU_FIRST_COUNTER));
5351 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5353 ovfl_pmds &= ~reset_pmds;
5356 * when no sampling module is used, then the default
5357 * is to notify on overflow if requested by user
5359 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5360 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5361 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5362 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5364 * if needed, we reset all overflowed pmds
5366 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5369 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5372 * reset the requested PMD registers using the short reset values
5375 unsigned long bm = reset_pmds;
5376 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5379 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5381 * keep track of what to reset when unblocking
5383 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5386 * check for blocking context
5388 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5390 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5393 * set the perfmon specific checking pending work for the task
5395 PFM_SET_WORK_PENDING(task, 1);
5398 * when coming from ctxsw, current still points to the
5399 * previous task, therefore we must work with task and not current.
5401 set_notify_resume(task);
5404 * defer until state is changed (shorten spin window). the context is locked
5405 * anyway, so the signal receiver would come spin for nothing.
5410 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5411 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5412 PFM_GET_WORK_PENDING(task),
5413 ctx->ctx_fl_trap_reason,
5416 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5418 * in case monitoring must be stopped, we toggle the psr bits
5420 if (ovfl_ctrl.bits.mask_monitoring) {
5421 pfm_mask_monitoring(task);
5422 ctx->ctx_state = PFM_CTX_MASKED;
5423 ctx->ctx_fl_can_restart = 1;
5427 * send notification now
5429 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5434 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5436 task ? task_pid_nr(task) : -1,
5442 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5443 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5444 * come here as zombie only if the task is the current task. In which case, we
5445 * can access the PMU hardware directly.
5447 * Note that zombies do have PM_VALID set. So here we do the minimal.
5449 * In case the context was zombified it could not be reclaimed at the time
5450 * the monitoring program exited. At this point, the PMU reservation has been
5451 * returned, the sampiing buffer has been freed. We must convert this call
5452 * into a spurious interrupt. However, we must also avoid infinite overflows
5453 * by stopping monitoring for this task. We can only come here for a per-task
5454 * context. All we need to do is to stop monitoring using the psr bits which
5455 * are always task private. By re-enabling secure montioring, we ensure that
5456 * the monitored task will not be able to re-activate monitoring.
5457 * The task will eventually be context switched out, at which point the context
5458 * will be reclaimed (that includes releasing ownership of the PMU).
5460 * So there might be a window of time where the number of per-task session is zero
5461 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5462 * context. This is safe because if a per-task session comes in, it will push this one
5463 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5464 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5465 * also push our zombie context out.
5467 * Overall pretty hairy stuff....
5469 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5471 ia64_psr(regs)->up = 0;
5472 ia64_psr(regs)->sp = 1;
5477 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5479 struct task_struct *task;
5481 unsigned long flags;
5483 int this_cpu = smp_processor_id();
5486 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5489 * srlz.d done before arriving here
5491 pmc0 = ia64_get_pmc(0);
5493 task = GET_PMU_OWNER();
5494 ctx = GET_PMU_CTX();
5497 * if we have some pending bits set
5498 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5500 if (PMC0_HAS_OVFL(pmc0) && task) {
5502 * we assume that pmc0.fr is always set here
5506 if (!ctx) goto report_spurious1;
5508 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5509 goto report_spurious2;
5511 PROTECT_CTX_NOPRINT(ctx, flags);
5513 pfm_overflow_handler(task, ctx, pmc0, regs);
5515 UNPROTECT_CTX_NOPRINT(ctx, flags);
5518 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5522 * keep it unfrozen at all times
5529 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5530 this_cpu, task_pid_nr(task));
5534 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5542 pfm_interrupt_handler(int irq, void *arg)
5544 unsigned long start_cycles, total_cycles;
5545 unsigned long min, max;
5548 struct pt_regs *regs = get_irq_regs();
5550 this_cpu = get_cpu();
5551 if (likely(!pfm_alt_intr_handler)) {
5552 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5553 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5555 start_cycles = ia64_get_itc();
5557 ret = pfm_do_interrupt_handler(arg, regs);
5559 total_cycles = ia64_get_itc();
5562 * don't measure spurious interrupts
5564 if (likely(ret == 0)) {
5565 total_cycles -= start_cycles;
5567 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5568 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5570 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5574 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5582 * /proc/perfmon interface, for debug only
5585 #define PFM_PROC_SHOW_HEADER ((void *)(long)nr_cpu_ids+1)
5588 pfm_proc_start(struct seq_file *m, loff_t *pos)
5591 return PFM_PROC_SHOW_HEADER;
5594 while (*pos <= nr_cpu_ids) {
5595 if (cpu_online(*pos - 1)) {
5596 return (void *)*pos;
5604 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5607 return pfm_proc_start(m, pos);
5611 pfm_proc_stop(struct seq_file *m, void *v)
5616 pfm_proc_show_header(struct seq_file *m)
5618 struct list_head * pos;
5619 pfm_buffer_fmt_t * entry;
5620 unsigned long flags;
5623 "perfmon version : %u.%u\n"
5626 "expert mode : %s\n"
5627 "ovfl_mask : 0x%lx\n"
5628 "PMU flags : 0x%x\n",
5629 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5631 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5632 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5639 "proc_sessions : %u\n"
5640 "sys_sessions : %u\n"
5641 "sys_use_dbregs : %u\n"
5642 "ptrace_use_dbregs : %u\n",
5643 pfm_sessions.pfs_task_sessions,
5644 pfm_sessions.pfs_sys_sessions,
5645 pfm_sessions.pfs_sys_use_dbregs,
5646 pfm_sessions.pfs_ptrace_use_dbregs);
5650 spin_lock(&pfm_buffer_fmt_lock);
5652 list_for_each(pos, &pfm_buffer_fmt_list) {
5653 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5654 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5665 entry->fmt_uuid[10],
5666 entry->fmt_uuid[11],
5667 entry->fmt_uuid[12],
5668 entry->fmt_uuid[13],
5669 entry->fmt_uuid[14],
5670 entry->fmt_uuid[15],
5673 spin_unlock(&pfm_buffer_fmt_lock);
5678 pfm_proc_show(struct seq_file *m, void *v)
5684 if (v == PFM_PROC_SHOW_HEADER) {
5685 pfm_proc_show_header(m);
5689 /* show info for CPU (v - 1) */
5693 "CPU%-2d overflow intrs : %lu\n"
5694 "CPU%-2d overflow cycles : %lu\n"
5695 "CPU%-2d overflow min : %lu\n"
5696 "CPU%-2d overflow max : %lu\n"
5697 "CPU%-2d smpl handler calls : %lu\n"
5698 "CPU%-2d smpl handler cycles : %lu\n"
5699 "CPU%-2d spurious intrs : %lu\n"
5700 "CPU%-2d replay intrs : %lu\n"
5701 "CPU%-2d syst_wide : %d\n"
5702 "CPU%-2d dcr_pp : %d\n"
5703 "CPU%-2d exclude idle : %d\n"
5704 "CPU%-2d owner : %d\n"
5705 "CPU%-2d context : %p\n"
5706 "CPU%-2d activations : %lu\n",
5707 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5708 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5709 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5710 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5711 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5712 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5713 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5714 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5715 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5716 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5717 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5718 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5719 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5720 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5722 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5724 psr = pfm_get_psr();
5729 "CPU%-2d psr : 0x%lx\n"
5730 "CPU%-2d pmc0 : 0x%lx\n",
5732 cpu, ia64_get_pmc(0));
5734 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5735 if (PMC_IS_COUNTING(i) == 0) continue;
5737 "CPU%-2d pmc%u : 0x%lx\n"
5738 "CPU%-2d pmd%u : 0x%lx\n",
5739 cpu, i, ia64_get_pmc(i),
5740 cpu, i, ia64_get_pmd(i));
5746 const struct seq_operations pfm_seq_ops = {
5747 .start = pfm_proc_start,
5748 .next = pfm_proc_next,
5749 .stop = pfm_proc_stop,
5750 .show = pfm_proc_show
5754 pfm_proc_open(struct inode *inode, struct file *file)
5756 return seq_open(file, &pfm_seq_ops);
5761 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5762 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5763 * is active or inactive based on mode. We must rely on the value in
5764 * local_cpu_data->pfm_syst_info
5767 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5769 struct pt_regs *regs;
5771 unsigned long dcr_pp;
5773 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5776 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5777 * on every CPU, so we can rely on the pid to identify the idle task.
5779 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5780 regs = task_pt_regs(task);
5781 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5785 * if monitoring has started
5788 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5790 * context switching in?
5793 /* mask monitoring for the idle task */
5794 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5800 * context switching out
5801 * restore monitoring for next task
5803 * Due to inlining this odd if-then-else construction generates
5806 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5815 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5817 struct task_struct *task = ctx->ctx_task;
5819 ia64_psr(regs)->up = 0;
5820 ia64_psr(regs)->sp = 1;
5822 if (GET_PMU_OWNER() == task) {
5823 DPRINT(("cleared ownership for [%d]\n",
5824 task_pid_nr(ctx->ctx_task)));
5825 SET_PMU_OWNER(NULL, NULL);
5829 * disconnect the task from the context and vice-versa
5831 PFM_SET_WORK_PENDING(task, 0);
5833 task->thread.pfm_context = NULL;
5834 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5836 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task)));
5841 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5844 pfm_save_regs(struct task_struct *task)
5847 unsigned long flags;
5851 ctx = PFM_GET_CTX(task);
5852 if (ctx == NULL) return;
5855 * we always come here with interrupts ALREADY disabled by
5856 * the scheduler. So we simply need to protect against concurrent
5857 * access, not CPU concurrency.
5859 flags = pfm_protect_ctx_ctxsw(ctx);
5861 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5862 struct pt_regs *regs = task_pt_regs(task);
5866 pfm_force_cleanup(ctx, regs);
5868 BUG_ON(ctx->ctx_smpl_hdr);
5870 pfm_unprotect_ctx_ctxsw(ctx, flags);
5872 pfm_context_free(ctx);
5877 * save current PSR: needed because we modify it
5880 psr = pfm_get_psr();
5882 BUG_ON(psr & (IA64_PSR_I));
5886 * This is the last instruction which may generate an overflow
5888 * We do not need to set psr.sp because, it is irrelevant in kernel.
5889 * It will be restored from ipsr when going back to user level
5894 * keep a copy of psr.up (for reload)
5896 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5899 * release ownership of this PMU.
5900 * PM interrupts are masked, so nothing
5903 SET_PMU_OWNER(NULL, NULL);
5906 * we systematically save the PMD as we have no
5907 * guarantee we will be schedule at that same
5910 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5913 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5914 * we will need it on the restore path to check
5915 * for pending overflow.
5917 ctx->th_pmcs[0] = ia64_get_pmc(0);
5920 * unfreeze PMU if had pending overflows
5922 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5925 * finally, allow context access.
5926 * interrupts will still be masked after this call.
5928 pfm_unprotect_ctx_ctxsw(ctx, flags);
5931 #else /* !CONFIG_SMP */
5933 pfm_save_regs(struct task_struct *task)
5938 ctx = PFM_GET_CTX(task);
5939 if (ctx == NULL) return;
5942 * save current PSR: needed because we modify it
5944 psr = pfm_get_psr();
5946 BUG_ON(psr & (IA64_PSR_I));
5950 * This is the last instruction which may generate an overflow
5952 * We do not need to set psr.sp because, it is irrelevant in kernel.
5953 * It will be restored from ipsr when going back to user level
5958 * keep a copy of psr.up (for reload)
5960 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5964 pfm_lazy_save_regs (struct task_struct *task)
5967 unsigned long flags;
5969 { u64 psr = pfm_get_psr();
5970 BUG_ON(psr & IA64_PSR_UP);
5973 ctx = PFM_GET_CTX(task);
5976 * we need to mask PMU overflow here to
5977 * make sure that we maintain pmc0 until
5978 * we save it. overflow interrupts are
5979 * treated as spurious if there is no
5982 * XXX: I don't think this is necessary
5984 PROTECT_CTX(ctx,flags);
5987 * release ownership of this PMU.
5988 * must be done before we save the registers.
5990 * after this call any PMU interrupt is treated
5993 SET_PMU_OWNER(NULL, NULL);
5996 * save all the pmds we use
5998 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6001 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6002 * it is needed to check for pended overflow
6003 * on the restore path
6005 ctx->th_pmcs[0] = ia64_get_pmc(0);
6008 * unfreeze PMU if had pending overflows
6010 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6013 * now get can unmask PMU interrupts, they will
6014 * be treated as purely spurious and we will not
6015 * lose any information
6017 UNPROTECT_CTX(ctx,flags);
6019 #endif /* CONFIG_SMP */
6023 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6026 pfm_load_regs (struct task_struct *task)
6029 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6030 unsigned long flags;
6032 int need_irq_resend;
6034 ctx = PFM_GET_CTX(task);
6035 if (unlikely(ctx == NULL)) return;
6037 BUG_ON(GET_PMU_OWNER());
6040 * possible on unload
6042 if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6045 * we always come here with interrupts ALREADY disabled by
6046 * the scheduler. So we simply need to protect against concurrent
6047 * access, not CPU concurrency.
6049 flags = pfm_protect_ctx_ctxsw(ctx);
6050 psr = pfm_get_psr();
6052 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6054 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6055 BUG_ON(psr & IA64_PSR_I);
6057 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6058 struct pt_regs *regs = task_pt_regs(task);
6060 BUG_ON(ctx->ctx_smpl_hdr);
6062 pfm_force_cleanup(ctx, regs);
6064 pfm_unprotect_ctx_ctxsw(ctx, flags);
6067 * this one (kmalloc'ed) is fine with interrupts disabled
6069 pfm_context_free(ctx);
6075 * we restore ALL the debug registers to avoid picking up
6078 if (ctx->ctx_fl_using_dbreg) {
6079 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6080 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6083 * retrieve saved psr.up
6085 psr_up = ctx->ctx_saved_psr_up;
6088 * if we were the last user of the PMU on that CPU,
6089 * then nothing to do except restore psr
6091 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6094 * retrieve partial reload masks (due to user modifications)
6096 pmc_mask = ctx->ctx_reload_pmcs[0];
6097 pmd_mask = ctx->ctx_reload_pmds[0];
6101 * To avoid leaking information to the user level when psr.sp=0,
6102 * we must reload ALL implemented pmds (even the ones we don't use).
6103 * In the kernel we only allow PFM_READ_PMDS on registers which
6104 * we initialized or requested (sampling) so there is no risk there.
6106 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6109 * ALL accessible PMCs are systematically reloaded, unused registers
6110 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6111 * up stale configuration.
6113 * PMC0 is never in the mask. It is always restored separately.
6115 pmc_mask = ctx->ctx_all_pmcs[0];
6118 * when context is MASKED, we will restore PMC with plm=0
6119 * and PMD with stale information, but that's ok, nothing
6122 * XXX: optimize here
6124 if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6125 if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6128 * check for pending overflow at the time the state
6131 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6133 * reload pmc0 with the overflow information
6134 * On McKinley PMU, this will trigger a PMU interrupt
6136 ia64_set_pmc(0, ctx->th_pmcs[0]);
6138 ctx->th_pmcs[0] = 0UL;
6141 * will replay the PMU interrupt
6143 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6145 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6149 * we just did a reload, so we reset the partial reload fields
6151 ctx->ctx_reload_pmcs[0] = 0UL;
6152 ctx->ctx_reload_pmds[0] = 0UL;
6154 SET_LAST_CPU(ctx, smp_processor_id());
6157 * dump activation value for this PMU
6161 * record current activation for this context
6163 SET_ACTIVATION(ctx);
6166 * establish new ownership.
6168 SET_PMU_OWNER(task, ctx);
6171 * restore the psr.up bit. measurement
6173 * no PMU interrupt can happen at this point
6174 * because we still have interrupts disabled.
6176 if (likely(psr_up)) pfm_set_psr_up();
6179 * allow concurrent access to context
6181 pfm_unprotect_ctx_ctxsw(ctx, flags);
6183 #else /* !CONFIG_SMP */
6185 * reload PMU state for UP kernels
6186 * in 2.5 we come here with interrupts disabled
6189 pfm_load_regs (struct task_struct *task)
6192 struct task_struct *owner;
6193 unsigned long pmd_mask, pmc_mask;
6195 int need_irq_resend;
6197 owner = GET_PMU_OWNER();
6198 ctx = PFM_GET_CTX(task);
6199 psr = pfm_get_psr();
6201 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6202 BUG_ON(psr & IA64_PSR_I);
6205 * we restore ALL the debug registers to avoid picking up
6208 * This must be done even when the task is still the owner
6209 * as the registers may have been modified via ptrace()
6210 * (not perfmon) by the previous task.
6212 if (ctx->ctx_fl_using_dbreg) {
6213 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6214 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6218 * retrieved saved psr.up
6220 psr_up = ctx->ctx_saved_psr_up;
6221 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6224 * short path, our state is still there, just
6225 * need to restore psr and we go
6227 * we do not touch either PMC nor PMD. the psr is not touched
6228 * by the overflow_handler. So we are safe w.r.t. to interrupt
6229 * concurrency even without interrupt masking.
6231 if (likely(owner == task)) {
6232 if (likely(psr_up)) pfm_set_psr_up();
6237 * someone else is still using the PMU, first push it out and
6238 * then we'll be able to install our stuff !
6240 * Upon return, there will be no owner for the current PMU
6242 if (owner) pfm_lazy_save_regs(owner);
6245 * To avoid leaking information to the user level when psr.sp=0,
6246 * we must reload ALL implemented pmds (even the ones we don't use).
6247 * In the kernel we only allow PFM_READ_PMDS on registers which
6248 * we initialized or requested (sampling) so there is no risk there.
6250 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6253 * ALL accessible PMCs are systematically reloaded, unused registers
6254 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6255 * up stale configuration.
6257 * PMC0 is never in the mask. It is always restored separately
6259 pmc_mask = ctx->ctx_all_pmcs[0];
6261 pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6262 pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6265 * check for pending overflow at the time the state
6268 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6270 * reload pmc0 with the overflow information
6271 * On McKinley PMU, this will trigger a PMU interrupt
6273 ia64_set_pmc(0, ctx->th_pmcs[0]);
6276 ctx->th_pmcs[0] = 0UL;
6279 * will replay the PMU interrupt
6281 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6283 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6287 * establish new ownership.
6289 SET_PMU_OWNER(task, ctx);
6292 * restore the psr.up bit. measurement
6294 * no PMU interrupt can happen at this point
6295 * because we still have interrupts disabled.
6297 if (likely(psr_up)) pfm_set_psr_up();
6299 #endif /* CONFIG_SMP */
6302 * this function assumes monitoring is stopped
6305 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6308 unsigned long mask2, val, pmd_val, ovfl_val;
6309 int i, can_access_pmu = 0;
6313 * is the caller the task being monitored (or which initiated the
6314 * session for system wide measurements)
6316 is_self = ctx->ctx_task == task ? 1 : 0;
6319 * can access PMU is task is the owner of the PMU state on the current CPU
6320 * or if we are running on the CPU bound to the context in system-wide mode
6321 * (that is not necessarily the task the context is attached to in this mode).
6322 * In system-wide we always have can_access_pmu true because a task running on an
6323 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6325 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6326 if (can_access_pmu) {
6328 * Mark the PMU as not owned
6329 * This will cause the interrupt handler to do nothing in case an overflow
6330 * interrupt was in-flight
6331 * This also guarantees that pmc0 will contain the final state
6332 * It virtually gives us full control on overflow processing from that point
6335 SET_PMU_OWNER(NULL, NULL);
6336 DPRINT(("releasing ownership\n"));
6339 * read current overflow status:
6341 * we are guaranteed to read the final stable state
6344 pmc0 = ia64_get_pmc(0); /* slow */
6347 * reset freeze bit, overflow status information destroyed
6351 pmc0 = ctx->th_pmcs[0];
6353 * clear whatever overflow status bits there were
6355 ctx->th_pmcs[0] = 0;
6357 ovfl_val = pmu_conf->ovfl_val;
6359 * we save all the used pmds
6360 * we take care of overflows for counting PMDs
6362 * XXX: sampling situation is not taken into account here
6364 mask2 = ctx->ctx_used_pmds[0];
6366 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6368 for (i = 0; mask2; i++, mask2>>=1) {
6370 /* skip non used pmds */
6371 if ((mask2 & 0x1) == 0) continue;
6374 * can access PMU always true in system wide mode
6376 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6378 if (PMD_IS_COUNTING(i)) {
6379 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6382 ctx->ctx_pmds[i].val,
6386 * we rebuild the full 64 bit value of the counter
6388 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6391 * now everything is in ctx_pmds[] and we need
6392 * to clear the saved context from save_regs() such that
6393 * pfm_read_pmds() gets the correct value
6398 * take care of overflow inline
6400 if (pmc0 & (1UL << i)) {
6401 val += 1 + ovfl_val;
6402 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6406 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6408 if (is_self) ctx->th_pmds[i] = pmd_val;
6410 ctx->ctx_pmds[i].val = val;
6414 static struct irqaction perfmon_irqaction = {
6415 .handler = pfm_interrupt_handler,
6416 .flags = IRQF_DISABLED,
6421 pfm_alt_save_pmu_state(void *data)
6423 struct pt_regs *regs;
6425 regs = task_pt_regs(current);
6427 DPRINT(("called\n"));
6430 * should not be necessary but
6431 * let's take not risk
6435 ia64_psr(regs)->pp = 0;
6438 * This call is required
6439 * May cause a spurious interrupt on some processors
6447 pfm_alt_restore_pmu_state(void *data)
6449 struct pt_regs *regs;
6451 regs = task_pt_regs(current);
6453 DPRINT(("called\n"));
6456 * put PMU back in state expected
6461 ia64_psr(regs)->pp = 0;
6464 * perfmon runs with PMU unfrozen at all times
6472 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6477 /* some sanity checks */
6478 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6480 /* do the easy test first */
6481 if (pfm_alt_intr_handler) return -EBUSY;
6483 /* one at a time in the install or remove, just fail the others */
6484 if (!spin_trylock(&pfm_alt_install_check)) {
6488 /* reserve our session */
6489 for_each_online_cpu(reserve_cpu) {
6490 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6491 if (ret) goto cleanup_reserve;
6494 /* save the current system wide pmu states */
6495 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6497 DPRINT(("on_each_cpu() failed: %d\n", ret));
6498 goto cleanup_reserve;
6501 /* officially change to the alternate interrupt handler */
6502 pfm_alt_intr_handler = hdl;
6504 spin_unlock(&pfm_alt_install_check);
6509 for_each_online_cpu(i) {
6510 /* don't unreserve more than we reserved */
6511 if (i >= reserve_cpu) break;
6513 pfm_unreserve_session(NULL, 1, i);
6516 spin_unlock(&pfm_alt_install_check);
6520 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6523 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6528 if (hdl == NULL) return -EINVAL;
6530 /* cannot remove someone else's handler! */
6531 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6533 /* one at a time in the install or remove, just fail the others */
6534 if (!spin_trylock(&pfm_alt_install_check)) {
6538 pfm_alt_intr_handler = NULL;
6540 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6542 DPRINT(("on_each_cpu() failed: %d\n", ret));
6545 for_each_online_cpu(i) {
6546 pfm_unreserve_session(NULL, 1, i);
6549 spin_unlock(&pfm_alt_install_check);
6553 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6556 * perfmon initialization routine, called from the initcall() table
6558 static int init_pfm_fs(void);
6566 family = local_cpu_data->family;
6571 if ((*p)->probe() == 0) goto found;
6572 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6583 static const struct file_operations pfm_proc_fops = {
6584 .open = pfm_proc_open,
6586 .llseek = seq_lseek,
6587 .release = seq_release,
6593 unsigned int n, n_counters, i;
6595 printk("perfmon: version %u.%u IRQ %u\n",
6598 IA64_PERFMON_VECTOR);
6600 if (pfm_probe_pmu()) {
6601 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6602 local_cpu_data->family);
6607 * compute the number of implemented PMD/PMC from the
6608 * description tables
6611 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6612 if (PMC_IS_IMPL(i) == 0) continue;
6613 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6616 pmu_conf->num_pmcs = n;
6618 n = 0; n_counters = 0;
6619 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6620 if (PMD_IS_IMPL(i) == 0) continue;
6621 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6623 if (PMD_IS_COUNTING(i)) n_counters++;
6625 pmu_conf->num_pmds = n;
6626 pmu_conf->num_counters = n_counters;
6629 * sanity checks on the number of debug registers
6631 if (pmu_conf->use_rr_dbregs) {
6632 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6633 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6637 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6638 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6644 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6648 pmu_conf->num_counters,
6649 ffz(pmu_conf->ovfl_val));
6652 if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6653 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6659 * create /proc/perfmon (mostly for debugging purposes)
6661 perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6662 if (perfmon_dir == NULL) {
6663 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6669 * create /proc/sys/kernel/perfmon (for debugging purposes)
6671 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6674 * initialize all our spinlocks
6676 spin_lock_init(&pfm_sessions.pfs_lock);
6677 spin_lock_init(&pfm_buffer_fmt_lock);
6681 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6686 __initcall(pfm_init);
6689 * this function is called before pfm_init()
6692 pfm_init_percpu (void)
6694 static int first_time=1;
6696 * make sure no measurement is active
6697 * (may inherit programmed PMCs from EFI).
6703 * we run with the PMU not frozen at all times
6708 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6712 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6717 * used for debug purposes only
6720 dump_pmu_state(const char *from)
6722 struct task_struct *task;
6723 struct pt_regs *regs;
6725 unsigned long psr, dcr, info, flags;
6728 local_irq_save(flags);
6730 this_cpu = smp_processor_id();
6731 regs = task_pt_regs(current);
6732 info = PFM_CPUINFO_GET();
6733 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6735 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6736 local_irq_restore(flags);
6740 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6743 task_pid_nr(current),
6747 task = GET_PMU_OWNER();
6748 ctx = GET_PMU_CTX();
6750 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6752 psr = pfm_get_psr();
6754 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",
6757 psr & IA64_PSR_PP ? 1 : 0,
6758 psr & IA64_PSR_UP ? 1 : 0,
6759 dcr & IA64_DCR_PP ? 1 : 0,
6762 ia64_psr(regs)->pp);
6764 ia64_psr(regs)->up = 0;
6765 ia64_psr(regs)->pp = 0;
6767 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6768 if (PMC_IS_IMPL(i) == 0) continue;
6769 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]);
6772 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6773 if (PMD_IS_IMPL(i) == 0) continue;
6774 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]);
6778 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6781 ctx->ctx_smpl_vaddr,
6785 ctx->ctx_saved_psr_up);
6787 local_irq_restore(flags);
6791 * called from process.c:copy_thread(). task is new child.
6794 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6796 struct thread_struct *thread;
6798 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6800 thread = &task->thread;
6803 * cut links inherited from parent (current)
6805 thread->pfm_context = NULL;
6807 PFM_SET_WORK_PENDING(task, 0);
6810 * the psr bits are already set properly in copy_threads()
6813 #else /* !CONFIG_PERFMON */
6815 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6819 #endif /* CONFIG_PERFMON */