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>
44 #include <asm/errno.h>
45 #include <asm/intrinsics.h>
47 #include <asm/perfmon.h>
48 #include <asm/processor.h>
49 #include <asm/signal.h>
50 #include <asm/system.h>
51 #include <asm/uaccess.h>
52 #include <asm/delay.h>
56 * perfmon context state
58 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
59 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
60 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
61 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
63 #define PFM_INVALID_ACTIVATION (~0UL)
65 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
66 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
69 * depth of message queue
71 #define PFM_MAX_MSGS 32
72 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
75 * type of a PMU register (bitmask).
77 * bit0 : register implemented
80 * bit4 : pmc has pmc.pm
81 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
82 * bit6-7 : register type
85 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
86 #define PFM_REG_IMPL 0x1 /* register implemented */
87 #define PFM_REG_END 0x2 /* end marker */
88 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
89 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
90 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
91 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
92 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
94 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
95 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
97 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
99 /* i assumed unsigned */
100 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
101 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
103 /* XXX: these assume that register i is implemented */
104 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
105 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
106 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
107 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
109 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
110 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
111 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
112 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
114 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
115 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
117 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
118 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
119 #define PFM_CTX_TASK(h) (h)->ctx_task
121 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
123 /* XXX: does not support more than 64 PMDs */
124 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
125 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
127 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
129 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
130 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
131 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
132 #define PFM_CODE_RR 0 /* requesting code range restriction */
133 #define PFM_DATA_RR 1 /* requestion data range restriction */
135 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
136 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
137 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
139 #define RDEP(x) (1UL<<(x))
142 * context protection macros
144 * - we need to protect against CPU concurrency (spin_lock)
145 * - we need to protect against PMU overflow interrupts (local_irq_disable)
147 * - we need to protect against PMU overflow interrupts (local_irq_disable)
149 * spin_lock_irqsave()/spin_unlock_irqrestore():
150 * in SMP: local_irq_disable + spin_lock
151 * in UP : local_irq_disable
153 * spin_lock()/spin_lock():
154 * in UP : removed automatically
155 * in SMP: protect against context accesses from other CPU. interrupts
156 * are not masked. This is useful for the PMU interrupt handler
157 * because we know we will not get PMU concurrency in that code.
159 #define PROTECT_CTX(c, f) \
161 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
162 spin_lock_irqsave(&(c)->ctx_lock, f); \
163 DPRINT(("spinlocked ctx %p by [%d]\n", c, task_pid_nr(current))); \
166 #define UNPROTECT_CTX(c, f) \
168 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
169 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
172 #define PROTECT_CTX_NOPRINT(c, f) \
174 spin_lock_irqsave(&(c)->ctx_lock, f); \
178 #define UNPROTECT_CTX_NOPRINT(c, f) \
180 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
184 #define PROTECT_CTX_NOIRQ(c) \
186 spin_lock(&(c)->ctx_lock); \
189 #define UNPROTECT_CTX_NOIRQ(c) \
191 spin_unlock(&(c)->ctx_lock); \
197 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
198 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
199 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
201 #else /* !CONFIG_SMP */
202 #define SET_ACTIVATION(t) do {} while(0)
203 #define GET_ACTIVATION(t) do {} while(0)
204 #define INC_ACTIVATION(t) do {} while(0)
205 #endif /* CONFIG_SMP */
207 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
208 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
209 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
211 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
212 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
214 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
217 * cmp0 must be the value of pmc0
219 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
221 #define PFMFS_MAGIC 0xa0b4d889
226 #define PFM_DEBUGGING 1
230 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
233 #define DPRINT_ovfl(a) \
235 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; } \
240 * 64-bit software counter structure
242 * the next_reset_type is applied to the next call to pfm_reset_regs()
245 unsigned long val; /* virtual 64bit counter value */
246 unsigned long lval; /* last reset value */
247 unsigned long long_reset; /* reset value on sampling overflow */
248 unsigned long short_reset; /* reset value on overflow */
249 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
250 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
251 unsigned long seed; /* seed for random-number generator */
252 unsigned long mask; /* mask for random-number generator */
253 unsigned int flags; /* notify/do not notify */
254 unsigned long eventid; /* overflow event identifier */
261 unsigned int block:1; /* when 1, task will blocked on user notifications */
262 unsigned int system:1; /* do system wide monitoring */
263 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
264 unsigned int is_sampling:1; /* true if using a custom format */
265 unsigned int excl_idle:1; /* exclude idle task in system wide session */
266 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
267 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
268 unsigned int no_msg:1; /* no message sent on overflow */
269 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
270 unsigned int reserved:22;
271 } pfm_context_flags_t;
273 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
274 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
275 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
279 * perfmon context: encapsulates all the state of a monitoring session
282 typedef struct pfm_context {
283 spinlock_t ctx_lock; /* context protection */
285 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
286 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
288 struct task_struct *ctx_task; /* task to which context is attached */
290 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
292 struct completion ctx_restart_done; /* use for blocking notification mode */
294 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
295 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
296 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
298 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
299 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
300 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
302 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
304 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
305 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
306 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
307 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
309 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
311 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
312 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
314 u64 ctx_saved_psr_up; /* only contains psr.up value */
316 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
317 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
318 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
320 int ctx_fd; /* file descriptor used my this context */
321 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
323 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
324 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
325 unsigned long ctx_smpl_size; /* size of sampling buffer */
326 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
328 wait_queue_head_t ctx_msgq_wait;
329 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
332 struct fasync_struct *ctx_async_queue;
334 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
338 * magic number used to verify that structure is really
341 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
343 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
346 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
347 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
349 #define SET_LAST_CPU(ctx, v) do {} while(0)
350 #define GET_LAST_CPU(ctx) do {} while(0)
354 #define ctx_fl_block ctx_flags.block
355 #define ctx_fl_system ctx_flags.system
356 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
357 #define ctx_fl_is_sampling ctx_flags.is_sampling
358 #define ctx_fl_excl_idle ctx_flags.excl_idle
359 #define ctx_fl_going_zombie ctx_flags.going_zombie
360 #define ctx_fl_trap_reason ctx_flags.trap_reason
361 #define ctx_fl_no_msg ctx_flags.no_msg
362 #define ctx_fl_can_restart ctx_flags.can_restart
364 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
365 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
368 * global information about all sessions
369 * mostly used to synchronize between system wide and per-process
372 spinlock_t pfs_lock; /* lock the structure */
374 unsigned int pfs_task_sessions; /* number of per task sessions */
375 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
376 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
377 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
378 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
382 * information about a PMC or PMD.
383 * dep_pmd[]: a bitmask of dependent PMD registers
384 * dep_pmc[]: a bitmask of dependent PMC registers
386 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
390 unsigned long default_value; /* power-on default value */
391 unsigned long reserved_mask; /* bitmask of reserved bits */
392 pfm_reg_check_t read_check;
393 pfm_reg_check_t write_check;
394 unsigned long dep_pmd[4];
395 unsigned long dep_pmc[4];
398 /* assume cnum is a valid monitor */
399 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
402 * This structure is initialized at boot time and contains
403 * a description of the PMU main characteristics.
405 * If the probe function is defined, detection is based
406 * on its return value:
407 * - 0 means recognized PMU
408 * - anything else means not supported
409 * When the probe function is not defined, then the pmu_family field
410 * is used and it must match the host CPU family such that:
411 * - cpu->family & config->pmu_family != 0
414 unsigned long ovfl_val; /* overflow value for counters */
416 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
417 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
419 unsigned int num_pmcs; /* number of PMCS: computed at init time */
420 unsigned int num_pmds; /* number of PMDS: computed at init time */
421 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
422 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
424 char *pmu_name; /* PMU family name */
425 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
426 unsigned int flags; /* pmu specific flags */
427 unsigned int num_ibrs; /* number of IBRS: computed at init time */
428 unsigned int num_dbrs; /* number of DBRS: computed at init time */
429 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
430 int (*probe)(void); /* customized probe routine */
431 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
436 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
439 * debug register related type definitions
442 unsigned long ibr_mask:56;
443 unsigned long ibr_plm:4;
444 unsigned long ibr_ig:3;
445 unsigned long ibr_x:1;
449 unsigned long dbr_mask:56;
450 unsigned long dbr_plm:4;
451 unsigned long dbr_ig:2;
452 unsigned long dbr_w:1;
453 unsigned long dbr_r:1;
464 * perfmon command descriptions
467 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
470 unsigned int cmd_narg;
472 int (*cmd_getsize)(void *arg, size_t *sz);
475 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
476 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
477 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
478 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
481 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
482 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
483 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
484 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
485 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
487 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
490 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
491 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
492 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
493 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
496 unsigned long pfm_smpl_handler_calls;
497 unsigned long pfm_smpl_handler_cycles;
498 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
502 * perfmon internal variables
504 static pfm_stats_t pfm_stats[NR_CPUS];
505 static pfm_session_t pfm_sessions; /* global sessions information */
507 static DEFINE_SPINLOCK(pfm_alt_install_check);
508 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
510 static struct proc_dir_entry *perfmon_dir;
511 static pfm_uuid_t pfm_null_uuid = {0,};
513 static spinlock_t pfm_buffer_fmt_lock;
514 static LIST_HEAD(pfm_buffer_fmt_list);
516 static pmu_config_t *pmu_conf;
518 /* sysctl() controls */
519 pfm_sysctl_t pfm_sysctl;
520 EXPORT_SYMBOL(pfm_sysctl);
522 static ctl_table pfm_ctl_table[]={
524 .ctl_name = CTL_UNNUMBERED,
526 .data = &pfm_sysctl.debug,
527 .maxlen = sizeof(int),
529 .proc_handler = &proc_dointvec,
532 .ctl_name = CTL_UNNUMBERED,
533 .procname = "debug_ovfl",
534 .data = &pfm_sysctl.debug_ovfl,
535 .maxlen = sizeof(int),
537 .proc_handler = &proc_dointvec,
540 .ctl_name = CTL_UNNUMBERED,
541 .procname = "fastctxsw",
542 .data = &pfm_sysctl.fastctxsw,
543 .maxlen = sizeof(int),
545 .proc_handler = &proc_dointvec,
548 .ctl_name = CTL_UNNUMBERED,
549 .procname = "expert_mode",
550 .data = &pfm_sysctl.expert_mode,
551 .maxlen = sizeof(int),
553 .proc_handler = &proc_dointvec,
557 static ctl_table pfm_sysctl_dir[] = {
559 .ctl_name = CTL_UNNUMBERED,
560 .procname = "perfmon",
562 .child = pfm_ctl_table,
566 static ctl_table pfm_sysctl_root[] = {
568 .ctl_name = CTL_KERN,
569 .procname = "kernel",
571 .child = pfm_sysctl_dir,
575 static struct ctl_table_header *pfm_sysctl_header;
577 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
579 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
580 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
583 pfm_put_task(struct task_struct *task)
585 if (task != current) put_task_struct(task);
589 pfm_reserve_page(unsigned long a)
591 SetPageReserved(vmalloc_to_page((void *)a));
594 pfm_unreserve_page(unsigned long a)
596 ClearPageReserved(vmalloc_to_page((void*)a));
599 static inline unsigned long
600 pfm_protect_ctx_ctxsw(pfm_context_t *x)
602 spin_lock(&(x)->ctx_lock);
607 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
609 spin_unlock(&(x)->ctx_lock);
612 static inline unsigned int
613 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
615 return do_munmap(mm, addr, len);
618 static inline unsigned long
619 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
621 return get_unmapped_area(file, addr, len, pgoff, flags);
626 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data,
627 struct vfsmount *mnt)
629 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
632 static struct file_system_type pfm_fs_type = {
634 .get_sb = pfmfs_get_sb,
635 .kill_sb = kill_anon_super,
638 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
639 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
640 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
641 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
642 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
645 /* forward declaration */
646 static const struct file_operations pfm_file_ops;
649 * forward declarations
652 static void pfm_lazy_save_regs (struct task_struct *ta);
655 void dump_pmu_state(const char *);
656 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
658 #include "perfmon_itanium.h"
659 #include "perfmon_mckinley.h"
660 #include "perfmon_montecito.h"
661 #include "perfmon_generic.h"
663 static pmu_config_t *pmu_confs[]={
667 &pmu_conf_gen, /* must be last */
672 static int pfm_end_notify_user(pfm_context_t *ctx);
675 pfm_clear_psr_pp(void)
677 ia64_rsm(IA64_PSR_PP);
684 ia64_ssm(IA64_PSR_PP);
689 pfm_clear_psr_up(void)
691 ia64_rsm(IA64_PSR_UP);
698 ia64_ssm(IA64_PSR_UP);
702 static inline unsigned long
706 tmp = ia64_getreg(_IA64_REG_PSR);
712 pfm_set_psr_l(unsigned long val)
714 ia64_setreg(_IA64_REG_PSR_L, val);
726 pfm_unfreeze_pmu(void)
733 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
737 for (i=0; i < nibrs; i++) {
738 ia64_set_ibr(i, ibrs[i]);
739 ia64_dv_serialize_instruction();
745 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
749 for (i=0; i < ndbrs; i++) {
750 ia64_set_dbr(i, dbrs[i]);
751 ia64_dv_serialize_data();
757 * PMD[i] must be a counter. no check is made
759 static inline unsigned long
760 pfm_read_soft_counter(pfm_context_t *ctx, int i)
762 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
766 * PMD[i] must be a counter. no check is made
769 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
771 unsigned long ovfl_val = pmu_conf->ovfl_val;
773 ctx->ctx_pmds[i].val = val & ~ovfl_val;
775 * writing to unimplemented part is ignore, so we do not need to
778 ia64_set_pmd(i, val & ovfl_val);
782 pfm_get_new_msg(pfm_context_t *ctx)
786 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
788 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
789 if (next == ctx->ctx_msgq_head) return NULL;
791 idx = ctx->ctx_msgq_tail;
792 ctx->ctx_msgq_tail = next;
794 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
796 return ctx->ctx_msgq+idx;
800 pfm_get_next_msg(pfm_context_t *ctx)
804 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
806 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
811 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
816 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
818 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));
824 pfm_reset_msgq(pfm_context_t *ctx)
826 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
827 DPRINT(("ctx=%p msgq reset\n", ctx));
831 pfm_rvmalloc(unsigned long size)
836 size = PAGE_ALIGN(size);
839 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
840 memset(mem, 0, size);
841 addr = (unsigned long)mem;
843 pfm_reserve_page(addr);
852 pfm_rvfree(void *mem, unsigned long size)
857 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
858 addr = (unsigned long) mem;
859 while ((long) size > 0) {
860 pfm_unreserve_page(addr);
869 static pfm_context_t *
870 pfm_context_alloc(int ctx_flags)
875 * allocate context descriptor
876 * must be able to free with interrupts disabled
878 ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
880 DPRINT(("alloc ctx @%p\n", ctx));
883 * init context protection lock
885 spin_lock_init(&ctx->ctx_lock);
888 * context is unloaded
890 ctx->ctx_state = PFM_CTX_UNLOADED;
893 * initialization of context's flags
895 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
896 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
897 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
899 * will move to set properties
900 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
904 * init restart semaphore to locked
906 init_completion(&ctx->ctx_restart_done);
909 * activation is used in SMP only
911 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
912 SET_LAST_CPU(ctx, -1);
915 * initialize notification message queue
917 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
918 init_waitqueue_head(&ctx->ctx_msgq_wait);
919 init_waitqueue_head(&ctx->ctx_zombieq);
926 pfm_context_free(pfm_context_t *ctx)
929 DPRINT(("free ctx @%p\n", ctx));
935 pfm_mask_monitoring(struct task_struct *task)
937 pfm_context_t *ctx = PFM_GET_CTX(task);
938 unsigned long mask, val, ovfl_mask;
941 DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
943 ovfl_mask = pmu_conf->ovfl_val;
945 * monitoring can only be masked as a result of a valid
946 * counter overflow. In UP, it means that the PMU still
947 * has an owner. Note that the owner can be different
948 * from the current task. However the PMU state belongs
950 * In SMP, a valid overflow only happens when task is
951 * current. Therefore if we come here, we know that
952 * the PMU state belongs to the current task, therefore
953 * we can access the live registers.
955 * So in both cases, the live register contains the owner's
956 * state. We can ONLY touch the PMU registers and NOT the PSR.
958 * As a consequence to this call, the ctx->th_pmds[] array
959 * contains stale information which must be ignored
960 * when context is reloaded AND monitoring is active (see
963 mask = ctx->ctx_used_pmds[0];
964 for (i = 0; mask; i++, mask>>=1) {
965 /* skip non used pmds */
966 if ((mask & 0x1) == 0) continue;
967 val = ia64_get_pmd(i);
969 if (PMD_IS_COUNTING(i)) {
971 * we rebuild the full 64 bit value of the counter
973 ctx->ctx_pmds[i].val += (val & ovfl_mask);
975 ctx->ctx_pmds[i].val = val;
977 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
979 ctx->ctx_pmds[i].val,
983 * mask monitoring by setting the privilege level to 0
984 * we cannot use psr.pp/psr.up for this, it is controlled by
987 * if task is current, modify actual registers, otherwise modify
988 * thread save state, i.e., what will be restored in pfm_load_regs()
990 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
991 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
992 if ((mask & 0x1) == 0UL) continue;
993 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
994 ctx->th_pmcs[i] &= ~0xfUL;
995 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
998 * make all of this visible
1004 * must always be done with task == current
1006 * context must be in MASKED state when calling
1009 pfm_restore_monitoring(struct task_struct *task)
1011 pfm_context_t *ctx = PFM_GET_CTX(task);
1012 unsigned long mask, ovfl_mask;
1013 unsigned long psr, val;
1016 is_system = ctx->ctx_fl_system;
1017 ovfl_mask = pmu_conf->ovfl_val;
1019 if (task != current) {
1020 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1023 if (ctx->ctx_state != PFM_CTX_MASKED) {
1024 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1025 task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1028 psr = pfm_get_psr();
1030 * monitoring is masked via the PMC.
1031 * As we restore their value, we do not want each counter to
1032 * restart right away. We stop monitoring using the PSR,
1033 * restore the PMC (and PMD) and then re-establish the psr
1034 * as it was. Note that there can be no pending overflow at
1035 * this point, because monitoring was MASKED.
1037 * system-wide session are pinned and self-monitoring
1039 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1040 /* disable dcr pp */
1041 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1047 * first, we restore the PMD
1049 mask = ctx->ctx_used_pmds[0];
1050 for (i = 0; mask; i++, mask>>=1) {
1051 /* skip non used pmds */
1052 if ((mask & 0x1) == 0) continue;
1054 if (PMD_IS_COUNTING(i)) {
1056 * we split the 64bit value according to
1059 val = ctx->ctx_pmds[i].val & ovfl_mask;
1060 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1062 val = ctx->ctx_pmds[i].val;
1064 ia64_set_pmd(i, val);
1066 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1068 ctx->ctx_pmds[i].val,
1074 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1075 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1076 if ((mask & 0x1) == 0UL) continue;
1077 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1078 ia64_set_pmc(i, ctx->th_pmcs[i]);
1079 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1080 task_pid_nr(task), i, ctx->th_pmcs[i]));
1085 * must restore DBR/IBR because could be modified while masked
1086 * XXX: need to optimize
1088 if (ctx->ctx_fl_using_dbreg) {
1089 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1090 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1096 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1098 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1105 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1111 for (i=0; mask; i++, mask>>=1) {
1112 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1117 * reload from thread state (used for ctxw only)
1120 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1123 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1125 for (i=0; mask; i++, mask>>=1) {
1126 if ((mask & 0x1) == 0) continue;
1127 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1128 ia64_set_pmd(i, val);
1134 * propagate PMD from context to thread-state
1137 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1139 unsigned long ovfl_val = pmu_conf->ovfl_val;
1140 unsigned long mask = ctx->ctx_all_pmds[0];
1144 DPRINT(("mask=0x%lx\n", mask));
1146 for (i=0; mask; i++, mask>>=1) {
1148 val = ctx->ctx_pmds[i].val;
1151 * We break up the 64 bit value into 2 pieces
1152 * the lower bits go to the machine state in the
1153 * thread (will be reloaded on ctxsw in).
1154 * The upper part stays in the soft-counter.
1156 if (PMD_IS_COUNTING(i)) {
1157 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1160 ctx->th_pmds[i] = val;
1162 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1165 ctx->ctx_pmds[i].val));
1170 * propagate PMC from context to thread-state
1173 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1175 unsigned long mask = ctx->ctx_all_pmcs[0];
1178 DPRINT(("mask=0x%lx\n", mask));
1180 for (i=0; mask; i++, mask>>=1) {
1181 /* masking 0 with ovfl_val yields 0 */
1182 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1183 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1190 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1194 for (i=0; mask; i++, mask>>=1) {
1195 if ((mask & 0x1) == 0) continue;
1196 ia64_set_pmc(i, pmcs[i]);
1202 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1204 return memcmp(a, b, sizeof(pfm_uuid_t));
1208 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1211 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1216 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1219 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1225 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1229 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1234 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1238 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1243 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1246 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1251 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)
1254 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1258 static pfm_buffer_fmt_t *
1259 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1261 struct list_head * pos;
1262 pfm_buffer_fmt_t * entry;
1264 list_for_each(pos, &pfm_buffer_fmt_list) {
1265 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1266 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1273 * find a buffer format based on its uuid
1275 static pfm_buffer_fmt_t *
1276 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1278 pfm_buffer_fmt_t * fmt;
1279 spin_lock(&pfm_buffer_fmt_lock);
1280 fmt = __pfm_find_buffer_fmt(uuid);
1281 spin_unlock(&pfm_buffer_fmt_lock);
1286 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1290 /* some sanity checks */
1291 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1293 /* we need at least a handler */
1294 if (fmt->fmt_handler == NULL) return -EINVAL;
1297 * XXX: need check validity of fmt_arg_size
1300 spin_lock(&pfm_buffer_fmt_lock);
1302 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1303 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1307 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1308 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1311 spin_unlock(&pfm_buffer_fmt_lock);
1314 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1317 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1319 pfm_buffer_fmt_t *fmt;
1322 spin_lock(&pfm_buffer_fmt_lock);
1324 fmt = __pfm_find_buffer_fmt(uuid);
1326 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1330 list_del_init(&fmt->fmt_list);
1331 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1334 spin_unlock(&pfm_buffer_fmt_lock);
1338 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1340 extern void update_pal_halt_status(int);
1343 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1345 unsigned long flags;
1347 * validity checks on cpu_mask have been done upstream
1351 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1352 pfm_sessions.pfs_sys_sessions,
1353 pfm_sessions.pfs_task_sessions,
1354 pfm_sessions.pfs_sys_use_dbregs,
1360 * cannot mix system wide and per-task sessions
1362 if (pfm_sessions.pfs_task_sessions > 0UL) {
1363 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1364 pfm_sessions.pfs_task_sessions));
1368 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1370 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1372 pfm_sessions.pfs_sys_session[cpu] = task;
1374 pfm_sessions.pfs_sys_sessions++ ;
1377 if (pfm_sessions.pfs_sys_sessions) goto abort;
1378 pfm_sessions.pfs_task_sessions++;
1381 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1382 pfm_sessions.pfs_sys_sessions,
1383 pfm_sessions.pfs_task_sessions,
1384 pfm_sessions.pfs_sys_use_dbregs,
1389 * disable default_idle() to go to PAL_HALT
1391 update_pal_halt_status(0);
1398 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1399 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1409 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1411 unsigned long flags;
1413 * validity checks on cpu_mask have been done upstream
1417 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1418 pfm_sessions.pfs_sys_sessions,
1419 pfm_sessions.pfs_task_sessions,
1420 pfm_sessions.pfs_sys_use_dbregs,
1426 pfm_sessions.pfs_sys_session[cpu] = NULL;
1428 * would not work with perfmon+more than one bit in cpu_mask
1430 if (ctx && ctx->ctx_fl_using_dbreg) {
1431 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1432 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1434 pfm_sessions.pfs_sys_use_dbregs--;
1437 pfm_sessions.pfs_sys_sessions--;
1439 pfm_sessions.pfs_task_sessions--;
1441 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1442 pfm_sessions.pfs_sys_sessions,
1443 pfm_sessions.pfs_task_sessions,
1444 pfm_sessions.pfs_sys_use_dbregs,
1449 * if possible, enable default_idle() to go into PAL_HALT
1451 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1452 update_pal_halt_status(1);
1460 * removes virtual mapping of the sampling buffer.
1461 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1462 * a PROTECT_CTX() section.
1465 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1470 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1471 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1475 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1478 * does the actual unmapping
1480 down_write(&task->mm->mmap_sem);
1482 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1484 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1486 up_write(&task->mm->mmap_sem);
1488 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1491 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1497 * free actual physical storage used by sampling buffer
1501 pfm_free_smpl_buffer(pfm_context_t *ctx)
1503 pfm_buffer_fmt_t *fmt;
1505 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1508 * we won't use the buffer format anymore
1510 fmt = ctx->ctx_buf_fmt;
1512 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1515 ctx->ctx_smpl_vaddr));
1517 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1522 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1524 ctx->ctx_smpl_hdr = NULL;
1525 ctx->ctx_smpl_size = 0UL;
1530 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1536 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1538 if (fmt == NULL) return;
1540 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1545 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1546 * no real gain from having the whole whorehouse mounted. So we don't need
1547 * any operations on the root directory. However, we need a non-trivial
1548 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1550 static struct vfsmount *pfmfs_mnt;
1555 int err = register_filesystem(&pfm_fs_type);
1557 pfmfs_mnt = kern_mount(&pfm_fs_type);
1558 err = PTR_ERR(pfmfs_mnt);
1559 if (IS_ERR(pfmfs_mnt))
1560 unregister_filesystem(&pfm_fs_type);
1568 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1573 unsigned long flags;
1574 DECLARE_WAITQUEUE(wait, current);
1575 if (PFM_IS_FILE(filp) == 0) {
1576 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1580 ctx = (pfm_context_t *)filp->private_data;
1582 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1587 * check even when there is no message
1589 if (size < sizeof(pfm_msg_t)) {
1590 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1594 PROTECT_CTX(ctx, flags);
1597 * put ourselves on the wait queue
1599 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1607 set_current_state(TASK_INTERRUPTIBLE);
1609 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1612 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1614 UNPROTECT_CTX(ctx, flags);
1617 * check non-blocking read
1620 if(filp->f_flags & O_NONBLOCK) break;
1623 * check pending signals
1625 if(signal_pending(current)) {
1630 * no message, so wait
1634 PROTECT_CTX(ctx, flags);
1636 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1637 set_current_state(TASK_RUNNING);
1638 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1640 if (ret < 0) goto abort;
1643 msg = pfm_get_next_msg(ctx);
1645 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1649 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1652 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1655 UNPROTECT_CTX(ctx, flags);
1661 pfm_write(struct file *file, const char __user *ubuf,
1662 size_t size, loff_t *ppos)
1664 DPRINT(("pfm_write called\n"));
1669 pfm_poll(struct file *filp, poll_table * wait)
1672 unsigned long flags;
1673 unsigned int mask = 0;
1675 if (PFM_IS_FILE(filp) == 0) {
1676 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1680 ctx = (pfm_context_t *)filp->private_data;
1682 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1687 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1689 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1691 PROTECT_CTX(ctx, flags);
1693 if (PFM_CTXQ_EMPTY(ctx) == 0)
1694 mask = POLLIN | POLLRDNORM;
1696 UNPROTECT_CTX(ctx, flags);
1698 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1704 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1706 DPRINT(("pfm_ioctl called\n"));
1711 * interrupt cannot be masked when coming here
1714 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1718 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1720 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1721 task_pid_nr(current),
1724 ctx->ctx_async_queue, ret));
1730 pfm_fasync(int fd, struct file *filp, int on)
1735 if (PFM_IS_FILE(filp) == 0) {
1736 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1740 ctx = (pfm_context_t *)filp->private_data;
1742 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1746 * we cannot mask interrupts during this call because this may
1747 * may go to sleep if memory is not readily avalaible.
1749 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1750 * done in caller. Serialization of this function is ensured by caller.
1752 ret = pfm_do_fasync(fd, filp, ctx, on);
1755 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1758 ctx->ctx_async_queue, ret));
1765 * this function is exclusively called from pfm_close().
1766 * The context is not protected at that time, nor are interrupts
1767 * on the remote CPU. That's necessary to avoid deadlocks.
1770 pfm_syswide_force_stop(void *info)
1772 pfm_context_t *ctx = (pfm_context_t *)info;
1773 struct pt_regs *regs = task_pt_regs(current);
1774 struct task_struct *owner;
1775 unsigned long flags;
1778 if (ctx->ctx_cpu != smp_processor_id()) {
1779 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1781 smp_processor_id());
1784 owner = GET_PMU_OWNER();
1785 if (owner != ctx->ctx_task) {
1786 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1788 task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1791 if (GET_PMU_CTX() != ctx) {
1792 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1794 GET_PMU_CTX(), ctx);
1798 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1800 * the context is already protected in pfm_close(), we simply
1801 * need to mask interrupts to avoid a PMU interrupt race on
1804 local_irq_save(flags);
1806 ret = pfm_context_unload(ctx, NULL, 0, regs);
1808 DPRINT(("context_unload returned %d\n", ret));
1812 * unmask interrupts, PMU interrupts are now spurious here
1814 local_irq_restore(flags);
1818 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1822 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1823 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1824 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1826 #endif /* CONFIG_SMP */
1829 * called for each close(). Partially free resources.
1830 * When caller is self-monitoring, the context is unloaded.
1833 pfm_flush(struct file *filp, fl_owner_t id)
1836 struct task_struct *task;
1837 struct pt_regs *regs;
1838 unsigned long flags;
1839 unsigned long smpl_buf_size = 0UL;
1840 void *smpl_buf_vaddr = NULL;
1841 int state, is_system;
1843 if (PFM_IS_FILE(filp) == 0) {
1844 DPRINT(("bad magic for\n"));
1848 ctx = (pfm_context_t *)filp->private_data;
1850 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1855 * remove our file from the async queue, if we use this mode.
1856 * This can be done without the context being protected. We come
1857 * here when the context has become unreachable by other tasks.
1859 * We may still have active monitoring at this point and we may
1860 * end up in pfm_overflow_handler(). However, fasync_helper()
1861 * operates with interrupts disabled and it cleans up the
1862 * queue. If the PMU handler is called prior to entering
1863 * fasync_helper() then it will send a signal. If it is
1864 * invoked after, it will find an empty queue and no
1865 * signal will be sent. In both case, we are safe
1867 if (filp->f_flags & FASYNC) {
1868 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1869 pfm_do_fasync (-1, filp, ctx, 0);
1872 PROTECT_CTX(ctx, flags);
1874 state = ctx->ctx_state;
1875 is_system = ctx->ctx_fl_system;
1877 task = PFM_CTX_TASK(ctx);
1878 regs = task_pt_regs(task);
1880 DPRINT(("ctx_state=%d is_current=%d\n",
1882 task == current ? 1 : 0));
1885 * if state == UNLOADED, then task is NULL
1889 * we must stop and unload because we are losing access to the context.
1891 if (task == current) {
1894 * the task IS the owner but it migrated to another CPU: that's bad
1895 * but we must handle this cleanly. Unfortunately, the kernel does
1896 * not provide a mechanism to block migration (while the context is loaded).
1898 * We need to release the resource on the ORIGINAL cpu.
1900 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1902 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1904 * keep context protected but unmask interrupt for IPI
1906 local_irq_restore(flags);
1908 pfm_syswide_cleanup_other_cpu(ctx);
1911 * restore interrupt masking
1913 local_irq_save(flags);
1916 * context is unloaded at this point
1919 #endif /* CONFIG_SMP */
1922 DPRINT(("forcing unload\n"));
1924 * stop and unload, returning with state UNLOADED
1925 * and session unreserved.
1927 pfm_context_unload(ctx, NULL, 0, regs);
1929 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1934 * remove virtual mapping, if any, for the calling task.
1935 * cannot reset ctx field until last user is calling close().
1937 * ctx_smpl_vaddr must never be cleared because it is needed
1938 * by every task with access to the context
1940 * When called from do_exit(), the mm context is gone already, therefore
1941 * mm is NULL, i.e., the VMA is already gone and we do not have to
1944 if (ctx->ctx_smpl_vaddr && current->mm) {
1945 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1946 smpl_buf_size = ctx->ctx_smpl_size;
1949 UNPROTECT_CTX(ctx, flags);
1952 * if there was a mapping, then we systematically remove it
1953 * at this point. Cannot be done inside critical section
1954 * because some VM function reenables interrupts.
1957 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1962 * called either on explicit close() or from exit_files().
1963 * Only the LAST user of the file gets to this point, i.e., it is
1966 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1967 * (fput()),i.e, last task to access the file. Nobody else can access the
1968 * file at this point.
1970 * When called from exit_files(), the VMA has been freed because exit_mm()
1971 * is executed before exit_files().
1973 * When called from exit_files(), the current task is not yet ZOMBIE but we
1974 * flush the PMU state to the context.
1977 pfm_close(struct inode *inode, struct file *filp)
1980 struct task_struct *task;
1981 struct pt_regs *regs;
1982 DECLARE_WAITQUEUE(wait, current);
1983 unsigned long flags;
1984 unsigned long smpl_buf_size = 0UL;
1985 void *smpl_buf_addr = NULL;
1986 int free_possible = 1;
1987 int state, is_system;
1989 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1991 if (PFM_IS_FILE(filp) == 0) {
1992 DPRINT(("bad magic\n"));
1996 ctx = (pfm_context_t *)filp->private_data;
1998 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
2002 PROTECT_CTX(ctx, flags);
2004 state = ctx->ctx_state;
2005 is_system = ctx->ctx_fl_system;
2007 task = PFM_CTX_TASK(ctx);
2008 regs = task_pt_regs(task);
2010 DPRINT(("ctx_state=%d is_current=%d\n",
2012 task == current ? 1 : 0));
2015 * if task == current, then pfm_flush() unloaded the context
2017 if (state == PFM_CTX_UNLOADED) goto doit;
2020 * context is loaded/masked and task != current, we need to
2021 * either force an unload or go zombie
2025 * The task is currently blocked or will block after an overflow.
2026 * we must force it to wakeup to get out of the
2027 * MASKED state and transition to the unloaded state by itself.
2029 * This situation is only possible for per-task mode
2031 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2034 * set a "partial" zombie state to be checked
2035 * upon return from down() in pfm_handle_work().
2037 * We cannot use the ZOMBIE state, because it is checked
2038 * by pfm_load_regs() which is called upon wakeup from down().
2039 * In such case, it would free the context and then we would
2040 * return to pfm_handle_work() which would access the
2041 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2042 * but visible to pfm_handle_work().
2044 * For some window of time, we have a zombie context with
2045 * ctx_state = MASKED and not ZOMBIE
2047 ctx->ctx_fl_going_zombie = 1;
2050 * force task to wake up from MASKED state
2052 complete(&ctx->ctx_restart_done);
2054 DPRINT(("waking up ctx_state=%d\n", state));
2057 * put ourself to sleep waiting for the other
2058 * task to report completion
2060 * the context is protected by mutex, therefore there
2061 * is no risk of being notified of completion before
2062 * begin actually on the waitq.
2064 set_current_state(TASK_INTERRUPTIBLE);
2065 add_wait_queue(&ctx->ctx_zombieq, &wait);
2067 UNPROTECT_CTX(ctx, flags);
2070 * XXX: check for signals :
2071 * - ok for explicit close
2072 * - not ok when coming from exit_files()
2077 PROTECT_CTX(ctx, flags);
2080 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2081 set_current_state(TASK_RUNNING);
2084 * context is unloaded at this point
2086 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2088 else if (task != current) {
2091 * switch context to zombie state
2093 ctx->ctx_state = PFM_CTX_ZOMBIE;
2095 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2097 * cannot free the context on the spot. deferred until
2098 * the task notices the ZOMBIE state
2102 pfm_context_unload(ctx, NULL, 0, regs);
2107 /* reload state, may have changed during opening of critical section */
2108 state = ctx->ctx_state;
2111 * the context is still attached to a task (possibly current)
2112 * we cannot destroy it right now
2116 * we must free the sampling buffer right here because
2117 * we cannot rely on it being cleaned up later by the
2118 * monitored task. It is not possible to free vmalloc'ed
2119 * memory in pfm_load_regs(). Instead, we remove the buffer
2120 * now. should there be subsequent PMU overflow originally
2121 * meant for sampling, the will be converted to spurious
2122 * and that's fine because the monitoring tools is gone anyway.
2124 if (ctx->ctx_smpl_hdr) {
2125 smpl_buf_addr = ctx->ctx_smpl_hdr;
2126 smpl_buf_size = ctx->ctx_smpl_size;
2127 /* no more sampling */
2128 ctx->ctx_smpl_hdr = NULL;
2129 ctx->ctx_fl_is_sampling = 0;
2132 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2138 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2141 * UNLOADED that the session has already been unreserved.
2143 if (state == PFM_CTX_ZOMBIE) {
2144 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2148 * disconnect file descriptor from context must be done
2151 filp->private_data = NULL;
2154 * if we free on the spot, the context is now completely unreachable
2155 * from the callers side. The monitored task side is also cut, so we
2158 * If we have a deferred free, only the caller side is disconnected.
2160 UNPROTECT_CTX(ctx, flags);
2163 * All memory free operations (especially for vmalloc'ed memory)
2164 * MUST be done with interrupts ENABLED.
2166 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2169 * return the memory used by the context
2171 if (free_possible) pfm_context_free(ctx);
2177 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2179 DPRINT(("pfm_no_open called\n"));
2185 static const struct file_operations pfm_file_ops = {
2186 .llseek = no_llseek,
2191 .open = pfm_no_open, /* special open code to disallow open via /proc */
2192 .fasync = pfm_fasync,
2193 .release = pfm_close,
2198 pfmfs_delete_dentry(struct dentry *dentry)
2203 static struct dentry_operations pfmfs_dentry_operations = {
2204 .d_delete = pfmfs_delete_dentry,
2208 static struct file *
2209 pfm_alloc_file(pfm_context_t *ctx)
2212 struct inode *inode;
2213 struct dentry *dentry;
2218 * allocate a new inode
2220 inode = new_inode(pfmfs_mnt->mnt_sb);
2222 return ERR_PTR(-ENOMEM);
2224 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2226 inode->i_mode = S_IFCHR|S_IRUGO;
2227 inode->i_uid = current->fsuid;
2228 inode->i_gid = current->fsgid;
2230 sprintf(name, "[%lu]", inode->i_ino);
2232 this.len = strlen(name);
2233 this.hash = inode->i_ino;
2236 * allocate a new dcache entry
2238 dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2241 return ERR_PTR(-ENOMEM);
2244 dentry->d_op = &pfmfs_dentry_operations;
2245 d_add(dentry, inode);
2247 file = alloc_file(pfmfs_mnt, dentry, FMODE_READ, &pfm_file_ops);
2250 return ERR_PTR(-ENFILE);
2253 file->f_flags = O_RDONLY;
2254 file->private_data = ctx;
2260 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2262 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2265 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2268 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2279 * allocate a sampling buffer and remaps it into the user address space of the task
2282 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2284 struct mm_struct *mm = task->mm;
2285 struct vm_area_struct *vma = NULL;
2291 * the fixed header + requested size and align to page boundary
2293 size = PAGE_ALIGN(rsize);
2295 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2298 * check requested size to avoid Denial-of-service attacks
2299 * XXX: may have to refine this test
2300 * Check against address space limit.
2302 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2305 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2309 * We do the easy to undo allocations first.
2311 * pfm_rvmalloc(), clears the buffer, so there is no leak
2313 smpl_buf = pfm_rvmalloc(size);
2314 if (smpl_buf == NULL) {
2315 DPRINT(("Can't allocate sampling buffer\n"));
2319 DPRINT(("smpl_buf @%p\n", smpl_buf));
2322 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2324 DPRINT(("Cannot allocate vma\n"));
2329 * partially initialize the vma for the sampling buffer
2332 vma->vm_file = filp;
2333 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2334 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2337 * Now we have everything we need and we can initialize
2338 * and connect all the data structures
2341 ctx->ctx_smpl_hdr = smpl_buf;
2342 ctx->ctx_smpl_size = size; /* aligned size */
2345 * Let's do the difficult operations next.
2347 * now we atomically find some area in the address space and
2348 * remap the buffer in it.
2350 down_write(&task->mm->mmap_sem);
2352 /* find some free area in address space, must have mmap sem held */
2353 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2354 if (vma->vm_start == 0UL) {
2355 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2356 up_write(&task->mm->mmap_sem);
2359 vma->vm_end = vma->vm_start + size;
2360 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2362 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2364 /* can only be applied to current task, need to have the mm semaphore held when called */
2365 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2366 DPRINT(("Can't remap buffer\n"));
2367 up_write(&task->mm->mmap_sem);
2374 * now insert the vma in the vm list for the process, must be
2375 * done with mmap lock held
2377 insert_vm_struct(mm, vma);
2379 mm->total_vm += size >> PAGE_SHIFT;
2380 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2382 up_write(&task->mm->mmap_sem);
2385 * keep track of user level virtual address
2387 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2388 *(unsigned long *)user_vaddr = vma->vm_start;
2393 kmem_cache_free(vm_area_cachep, vma);
2395 pfm_rvfree(smpl_buf, size);
2401 * XXX: do something better here
2404 pfm_bad_permissions(struct task_struct *task)
2406 /* inspired by ptrace_attach() */
2407 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2416 return ((current->uid != task->euid)
2417 || (current->uid != task->suid)
2418 || (current->uid != task->uid)
2419 || (current->gid != task->egid)
2420 || (current->gid != task->sgid)
2421 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2425 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2431 ctx_flags = pfx->ctx_flags;
2433 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2436 * cannot block in this mode
2438 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2439 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2444 /* probably more to add here */
2450 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2451 unsigned int cpu, pfarg_context_t *arg)
2453 pfm_buffer_fmt_t *fmt = NULL;
2454 unsigned long size = 0UL;
2456 void *fmt_arg = NULL;
2458 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2460 /* invoke and lock buffer format, if found */
2461 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2463 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2468 * buffer argument MUST be contiguous to pfarg_context_t
2470 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2472 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2474 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2476 if (ret) goto error;
2478 /* link buffer format and context */
2479 ctx->ctx_buf_fmt = fmt;
2480 ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2483 * check if buffer format wants to use perfmon buffer allocation/mapping service
2485 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2486 if (ret) goto error;
2490 * buffer is always remapped into the caller's address space
2492 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2493 if (ret) goto error;
2495 /* keep track of user address of buffer */
2496 arg->ctx_smpl_vaddr = uaddr;
2498 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2505 pfm_reset_pmu_state(pfm_context_t *ctx)
2510 * install reset values for PMC.
2512 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2513 if (PMC_IS_IMPL(i) == 0) continue;
2514 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2515 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2518 * PMD registers are set to 0UL when the context in memset()
2522 * On context switched restore, we must restore ALL pmc and ALL pmd even
2523 * when they are not actively used by the task. In UP, the incoming process
2524 * may otherwise pick up left over PMC, PMD state from the previous process.
2525 * As opposed to PMD, stale PMC can cause harm to the incoming
2526 * process because they may change what is being measured.
2527 * Therefore, we must systematically reinstall the entire
2528 * PMC state. In SMP, the same thing is possible on the
2529 * same CPU but also on between 2 CPUs.
2531 * The problem with PMD is information leaking especially
2532 * to user level when psr.sp=0
2534 * There is unfortunately no easy way to avoid this problem
2535 * on either UP or SMP. This definitively slows down the
2536 * pfm_load_regs() function.
2540 * bitmask of all PMCs accessible to this context
2542 * PMC0 is treated differently.
2544 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2547 * bitmask of all PMDs that are accessible to this context
2549 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2551 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2554 * useful in case of re-enable after disable
2556 ctx->ctx_used_ibrs[0] = 0UL;
2557 ctx->ctx_used_dbrs[0] = 0UL;
2561 pfm_ctx_getsize(void *arg, size_t *sz)
2563 pfarg_context_t *req = (pfarg_context_t *)arg;
2564 pfm_buffer_fmt_t *fmt;
2568 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2570 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2572 DPRINT(("cannot find buffer format\n"));
2575 /* get just enough to copy in user parameters */
2576 *sz = fmt->fmt_arg_size;
2577 DPRINT(("arg_size=%lu\n", *sz));
2585 * cannot attach if :
2587 * - task not owned by caller
2588 * - task incompatible with context mode
2591 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2594 * no kernel task or task not owner by caller
2596 if (task->mm == NULL) {
2597 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2600 if (pfm_bad_permissions(task)) {
2601 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task)));
2605 * cannot block in self-monitoring mode
2607 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2608 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2612 if (task->exit_state == EXIT_ZOMBIE) {
2613 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task)));
2618 * always ok for self
2620 if (task == current) return 0;
2622 if (!task_is_stopped_or_traced(task)) {
2623 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2627 * make sure the task is off any CPU
2629 wait_task_inactive(task);
2631 /* more to come... */
2637 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2639 struct task_struct *p = current;
2642 /* XXX: need to add more checks here */
2643 if (pid < 2) return -EPERM;
2645 if (pid != task_pid_vnr(current)) {
2647 read_lock(&tasklist_lock);
2649 p = find_task_by_vpid(pid);
2651 /* make sure task cannot go away while we operate on it */
2652 if (p) get_task_struct(p);
2654 read_unlock(&tasklist_lock);
2656 if (p == NULL) return -ESRCH;
2659 ret = pfm_task_incompatible(ctx, p);
2662 } else if (p != current) {
2671 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2673 pfarg_context_t *req = (pfarg_context_t *)arg;
2680 /* let's check the arguments first */
2681 ret = pfarg_is_sane(current, req);
2685 ctx_flags = req->ctx_flags;
2689 fd = get_unused_fd();
2693 ctx = pfm_context_alloc(ctx_flags);
2697 filp = pfm_alloc_file(ctx);
2699 ret = PTR_ERR(filp);
2703 req->ctx_fd = ctx->ctx_fd = fd;
2706 * does the user want to sample?
2708 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2709 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2714 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2719 ctx->ctx_fl_excl_idle,
2724 * initialize soft PMU state
2726 pfm_reset_pmu_state(ctx);
2728 fd_install(fd, filp);
2733 path = filp->f_path;
2737 if (ctx->ctx_buf_fmt) {
2738 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2741 pfm_context_free(ctx);
2748 static inline unsigned long
2749 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2751 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2752 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2753 extern unsigned long carta_random32 (unsigned long seed);
2755 if (reg->flags & PFM_REGFL_RANDOM) {
2756 new_seed = carta_random32(old_seed);
2757 val -= (old_seed & mask); /* counter values are negative numbers! */
2758 if ((mask >> 32) != 0)
2759 /* construct a full 64-bit random value: */
2760 new_seed |= carta_random32(old_seed >> 32) << 32;
2761 reg->seed = new_seed;
2768 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2770 unsigned long mask = ovfl_regs[0];
2771 unsigned long reset_others = 0UL;
2776 * now restore reset value on sampling overflowed counters
2778 mask >>= PMU_FIRST_COUNTER;
2779 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2781 if ((mask & 0x1UL) == 0UL) continue;
2783 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2784 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2786 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2790 * Now take care of resetting the other registers
2792 for(i = 0; reset_others; i++, reset_others >>= 1) {
2794 if ((reset_others & 0x1) == 0) continue;
2796 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2798 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2799 is_long_reset ? "long" : "short", i, val));
2804 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2806 unsigned long mask = ovfl_regs[0];
2807 unsigned long reset_others = 0UL;
2811 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2813 if (ctx->ctx_state == PFM_CTX_MASKED) {
2814 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2819 * now restore reset value on sampling overflowed counters
2821 mask >>= PMU_FIRST_COUNTER;
2822 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2824 if ((mask & 0x1UL) == 0UL) continue;
2826 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2827 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2829 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2831 pfm_write_soft_counter(ctx, i, val);
2835 * Now take care of resetting the other registers
2837 for(i = 0; reset_others; i++, reset_others >>= 1) {
2839 if ((reset_others & 0x1) == 0) continue;
2841 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2843 if (PMD_IS_COUNTING(i)) {
2844 pfm_write_soft_counter(ctx, i, val);
2846 ia64_set_pmd(i, val);
2848 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2849 is_long_reset ? "long" : "short", i, val));
2855 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2857 struct task_struct *task;
2858 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2859 unsigned long value, pmc_pm;
2860 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2861 unsigned int cnum, reg_flags, flags, pmc_type;
2862 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2863 int is_monitor, is_counting, state;
2865 pfm_reg_check_t wr_func;
2866 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2868 state = ctx->ctx_state;
2869 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2870 is_system = ctx->ctx_fl_system;
2871 task = ctx->ctx_task;
2872 impl_pmds = pmu_conf->impl_pmds[0];
2874 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2878 * In system wide and when the context is loaded, access can only happen
2879 * when the caller is running on the CPU being monitored by the session.
2880 * It does not have to be the owner (ctx_task) of the context per se.
2882 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2883 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2886 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2888 expert_mode = pfm_sysctl.expert_mode;
2890 for (i = 0; i < count; i++, req++) {
2892 cnum = req->reg_num;
2893 reg_flags = req->reg_flags;
2894 value = req->reg_value;
2895 smpl_pmds = req->reg_smpl_pmds[0];
2896 reset_pmds = req->reg_reset_pmds[0];
2900 if (cnum >= PMU_MAX_PMCS) {
2901 DPRINT(("pmc%u is invalid\n", cnum));
2905 pmc_type = pmu_conf->pmc_desc[cnum].type;
2906 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2907 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2908 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2911 * we reject all non implemented PMC as well
2912 * as attempts to modify PMC[0-3] which are used
2913 * as status registers by the PMU
2915 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2916 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2919 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2921 * If the PMC is a monitor, then if the value is not the default:
2922 * - system-wide session: PMCx.pm=1 (privileged monitor)
2923 * - per-task : PMCx.pm=0 (user monitor)
2925 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2926 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2935 * enforce generation of overflow interrupt. Necessary on all
2938 value |= 1 << PMU_PMC_OI;
2940 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2941 flags |= PFM_REGFL_OVFL_NOTIFY;
2944 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2946 /* verify validity of smpl_pmds */
2947 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2948 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2952 /* verify validity of reset_pmds */
2953 if ((reset_pmds & impl_pmds) != reset_pmds) {
2954 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2958 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2959 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2962 /* eventid on non-counting monitors are ignored */
2966 * execute write checker, if any
2968 if (likely(expert_mode == 0 && wr_func)) {
2969 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2970 if (ret) goto error;
2975 * no error on this register
2977 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2980 * Now we commit the changes to the software state
2984 * update overflow information
2988 * full flag update each time a register is programmed
2990 ctx->ctx_pmds[cnum].flags = flags;
2992 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2993 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2994 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2997 * Mark all PMDS to be accessed as used.
2999 * We do not keep track of PMC because we have to
3000 * systematically restore ALL of them.
3002 * We do not update the used_monitors mask, because
3003 * if we have not programmed them, then will be in
3004 * a quiescent state, therefore we will not need to
3005 * mask/restore then when context is MASKED.
3007 CTX_USED_PMD(ctx, reset_pmds);
3008 CTX_USED_PMD(ctx, smpl_pmds);
3010 * make sure we do not try to reset on
3011 * restart because we have established new values
3013 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3016 * Needed in case the user does not initialize the equivalent
3017 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3018 * possible leak here.
3020 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3023 * keep track of the monitor PMC that we are using.
3024 * we save the value of the pmc in ctx_pmcs[] and if
3025 * the monitoring is not stopped for the context we also
3026 * place it in the saved state area so that it will be
3027 * picked up later by the context switch code.
3029 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3031 * The value in th_pmcs[] may be modified on overflow, i.e., when
3032 * monitoring needs to be stopped.
3034 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3037 * update context state
3039 ctx->ctx_pmcs[cnum] = value;
3043 * write thread state
3045 if (is_system == 0) ctx->th_pmcs[cnum] = value;
3048 * write hardware register if we can
3050 if (can_access_pmu) {
3051 ia64_set_pmc(cnum, value);
3056 * per-task SMP only here
3058 * we are guaranteed that the task is not running on the other CPU,
3059 * we indicate that this PMD will need to be reloaded if the task
3060 * is rescheduled on the CPU it ran last on.
3062 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3067 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",
3073 ctx->ctx_all_pmcs[0],
3074 ctx->ctx_used_pmds[0],
3075 ctx->ctx_pmds[cnum].eventid,
3078 ctx->ctx_reload_pmcs[0],
3079 ctx->ctx_used_monitors[0],
3080 ctx->ctx_ovfl_regs[0]));
3084 * make sure the changes are visible
3086 if (can_access_pmu) ia64_srlz_d();
3090 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3095 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3097 struct task_struct *task;
3098 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3099 unsigned long value, hw_value, ovfl_mask;
3101 int i, can_access_pmu = 0, state;
3102 int is_counting, is_loaded, is_system, expert_mode;
3104 pfm_reg_check_t wr_func;
3107 state = ctx->ctx_state;
3108 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3109 is_system = ctx->ctx_fl_system;
3110 ovfl_mask = pmu_conf->ovfl_val;
3111 task = ctx->ctx_task;
3113 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3116 * on both UP and SMP, we can only write to the PMC when the task is
3117 * the owner of the local PMU.
3119 if (likely(is_loaded)) {
3121 * In system wide and when the context is loaded, access can only happen
3122 * when the caller is running on the CPU being monitored by the session.
3123 * It does not have to be the owner (ctx_task) of the context per se.
3125 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3126 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3129 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3131 expert_mode = pfm_sysctl.expert_mode;
3133 for (i = 0; i < count; i++, req++) {
3135 cnum = req->reg_num;
3136 value = req->reg_value;
3138 if (!PMD_IS_IMPL(cnum)) {
3139 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3142 is_counting = PMD_IS_COUNTING(cnum);
3143 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3146 * execute write checker, if any
3148 if (unlikely(expert_mode == 0 && wr_func)) {
3149 unsigned long v = value;
3151 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3152 if (ret) goto abort_mission;
3159 * no error on this register
3161 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3164 * now commit changes to software state
3169 * update virtualized (64bits) counter
3173 * write context state
3175 ctx->ctx_pmds[cnum].lval = value;
3178 * when context is load we use the split value
3181 hw_value = value & ovfl_mask;
3182 value = value & ~ovfl_mask;
3186 * update reset values (not just for counters)
3188 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3189 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3192 * update randomization parameters (not just for counters)
3194 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3195 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3198 * update context value
3200 ctx->ctx_pmds[cnum].val = value;
3203 * Keep track of what we use
3205 * We do not keep track of PMC because we have to
3206 * systematically restore ALL of them.
3208 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3211 * mark this PMD register used as well
3213 CTX_USED_PMD(ctx, RDEP(cnum));
3216 * make sure we do not try to reset on
3217 * restart because we have established new values
3219 if (is_counting && state == PFM_CTX_MASKED) {
3220 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3225 * write thread state
3227 if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3230 * write hardware register if we can
3232 if (can_access_pmu) {
3233 ia64_set_pmd(cnum, hw_value);
3237 * we are guaranteed that the task is not running on the other CPU,
3238 * we indicate that this PMD will need to be reloaded if the task
3239 * is rescheduled on the CPU it ran last on.
3241 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3246 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3247 "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",
3253 ctx->ctx_pmds[cnum].val,
3254 ctx->ctx_pmds[cnum].short_reset,
3255 ctx->ctx_pmds[cnum].long_reset,
3256 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3257 ctx->ctx_pmds[cnum].seed,
3258 ctx->ctx_pmds[cnum].mask,
3259 ctx->ctx_used_pmds[0],
3260 ctx->ctx_pmds[cnum].reset_pmds[0],
3261 ctx->ctx_reload_pmds[0],
3262 ctx->ctx_all_pmds[0],
3263 ctx->ctx_ovfl_regs[0]));
3267 * make changes visible
3269 if (can_access_pmu) ia64_srlz_d();
3275 * for now, we have only one possibility for error
3277 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3282 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3283 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3284 * interrupt is delivered during the call, it will be kept pending until we leave, making
3285 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3286 * guaranteed to return consistent data to the user, it may simply be old. It is not
3287 * trivial to treat the overflow while inside the call because you may end up in
3288 * some module sampling buffer code causing deadlocks.
3291 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3293 struct task_struct *task;
3294 unsigned long val = 0UL, lval, ovfl_mask, sval;
3295 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3296 unsigned int cnum, reg_flags = 0;
3297 int i, can_access_pmu = 0, state;
3298 int is_loaded, is_system, is_counting, expert_mode;
3300 pfm_reg_check_t rd_func;
3303 * access is possible when loaded only for
3304 * self-monitoring tasks or in UP mode
3307 state = ctx->ctx_state;
3308 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3309 is_system = ctx->ctx_fl_system;
3310 ovfl_mask = pmu_conf->ovfl_val;
3311 task = ctx->ctx_task;
3313 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3315 if (likely(is_loaded)) {
3317 * In system wide and when the context is loaded, access can only happen
3318 * when the caller is running on the CPU being monitored by the session.
3319 * It does not have to be the owner (ctx_task) of the context per se.
3321 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3322 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3326 * this can be true when not self-monitoring only in UP
3328 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3330 if (can_access_pmu) ia64_srlz_d();
3332 expert_mode = pfm_sysctl.expert_mode;
3334 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3340 * on both UP and SMP, we can only read the PMD from the hardware register when
3341 * the task is the owner of the local PMU.
3344 for (i = 0; i < count; i++, req++) {
3346 cnum = req->reg_num;
3347 reg_flags = req->reg_flags;
3349 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3351 * we can only read the register that we use. That includes
3352 * the one we explicitly initialize AND the one we want included
3353 * in the sampling buffer (smpl_regs).
3355 * Having this restriction allows optimization in the ctxsw routine
3356 * without compromising security (leaks)
3358 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3360 sval = ctx->ctx_pmds[cnum].val;
3361 lval = ctx->ctx_pmds[cnum].lval;
3362 is_counting = PMD_IS_COUNTING(cnum);
3365 * If the task is not the current one, then we check if the
3366 * PMU state is still in the local live register due to lazy ctxsw.
3367 * If true, then we read directly from the registers.
3369 if (can_access_pmu){
3370 val = ia64_get_pmd(cnum);
3373 * context has been saved
3374 * if context is zombie, then task does not exist anymore.
3375 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3377 val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3379 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3383 * XXX: need to check for overflow when loaded
3390 * execute read checker, if any
3392 if (unlikely(expert_mode == 0 && rd_func)) {
3393 unsigned long v = val;
3394 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3395 if (ret) goto error;
3400 PFM_REG_RETFLAG_SET(reg_flags, 0);
3402 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3405 * update register return value, abort all if problem during copy.
3406 * we only modify the reg_flags field. no check mode is fine because
3407 * access has been verified upfront in sys_perfmonctl().
3409 req->reg_value = val;
3410 req->reg_flags = reg_flags;
3411 req->reg_last_reset_val = lval;
3417 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3422 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3426 if (req == NULL) return -EINVAL;
3428 ctx = GET_PMU_CTX();
3430 if (ctx == NULL) return -EINVAL;
3433 * for now limit to current task, which is enough when calling
3434 * from overflow handler
3436 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3438 return pfm_write_pmcs(ctx, req, nreq, regs);
3440 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3443 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3447 if (req == NULL) return -EINVAL;
3449 ctx = GET_PMU_CTX();
3451 if (ctx == NULL) return -EINVAL;
3454 * for now limit to current task, which is enough when calling
3455 * from overflow handler
3457 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3459 return pfm_read_pmds(ctx, req, nreq, regs);
3461 EXPORT_SYMBOL(pfm_mod_read_pmds);
3464 * Only call this function when a process it trying to
3465 * write the debug registers (reading is always allowed)
3468 pfm_use_debug_registers(struct task_struct *task)
3470 pfm_context_t *ctx = task->thread.pfm_context;
3471 unsigned long flags;
3474 if (pmu_conf->use_rr_dbregs == 0) return 0;
3476 DPRINT(("called for [%d]\n", task_pid_nr(task)));
3481 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3484 * Even on SMP, we do not need to use an atomic here because
3485 * the only way in is via ptrace() and this is possible only when the
3486 * process is stopped. Even in the case where the ctxsw out is not totally
3487 * completed by the time we come here, there is no way the 'stopped' process
3488 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3489 * So this is always safe.
3491 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3496 * We cannot allow setting breakpoints when system wide monitoring
3497 * sessions are using the debug registers.
3499 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3502 pfm_sessions.pfs_ptrace_use_dbregs++;
3504 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3505 pfm_sessions.pfs_ptrace_use_dbregs,
3506 pfm_sessions.pfs_sys_use_dbregs,
3507 task_pid_nr(task), ret));
3515 * This function is called for every task that exits with the
3516 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3517 * able to use the debug registers for debugging purposes via
3518 * ptrace(). Therefore we know it was not using them for
3519 * perfmormance monitoring, so we only decrement the number
3520 * of "ptraced" debug register users to keep the count up to date
3523 pfm_release_debug_registers(struct task_struct *task)
3525 unsigned long flags;
3528 if (pmu_conf->use_rr_dbregs == 0) return 0;
3531 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3532 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3535 pfm_sessions.pfs_ptrace_use_dbregs--;
3544 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3546 struct task_struct *task;
3547 pfm_buffer_fmt_t *fmt;
3548 pfm_ovfl_ctrl_t rst_ctrl;
3549 int state, is_system;
3552 state = ctx->ctx_state;
3553 fmt = ctx->ctx_buf_fmt;
3554 is_system = ctx->ctx_fl_system;
3555 task = PFM_CTX_TASK(ctx);
3558 case PFM_CTX_MASKED:
3560 case PFM_CTX_LOADED:
3561 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3563 case PFM_CTX_UNLOADED:
3564 case PFM_CTX_ZOMBIE:
3565 DPRINT(("invalid state=%d\n", state));
3568 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3573 * In system wide and when the context is loaded, access can only happen
3574 * when the caller is running on the CPU being monitored by the session.
3575 * It does not have to be the owner (ctx_task) of the context per se.
3577 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3578 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3583 if (unlikely(task == NULL)) {
3584 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3588 if (task == current || is_system) {
3590 fmt = ctx->ctx_buf_fmt;
3592 DPRINT(("restarting self %d ovfl=0x%lx\n",
3594 ctx->ctx_ovfl_regs[0]));
3596 if (CTX_HAS_SMPL(ctx)) {
3598 prefetch(ctx->ctx_smpl_hdr);
3600 rst_ctrl.bits.mask_monitoring = 0;
3601 rst_ctrl.bits.reset_ovfl_pmds = 0;
3603 if (state == PFM_CTX_LOADED)
3604 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3606 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3608 rst_ctrl.bits.mask_monitoring = 0;
3609 rst_ctrl.bits.reset_ovfl_pmds = 1;
3613 if (rst_ctrl.bits.reset_ovfl_pmds)
3614 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3616 if (rst_ctrl.bits.mask_monitoring == 0) {
3617 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3619 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3621 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3623 // cannot use pfm_stop_monitoring(task, regs);
3627 * clear overflowed PMD mask to remove any stale information
3629 ctx->ctx_ovfl_regs[0] = 0UL;
3632 * back to LOADED state
3634 ctx->ctx_state = PFM_CTX_LOADED;
3637 * XXX: not really useful for self monitoring
3639 ctx->ctx_fl_can_restart = 0;
3645 * restart another task
3649 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3650 * one is seen by the task.
3652 if (state == PFM_CTX_MASKED) {
3653 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3655 * will prevent subsequent restart before this one is
3656 * seen by other task
3658 ctx->ctx_fl_can_restart = 0;
3662 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3663 * the task is blocked or on its way to block. That's the normal
3664 * restart path. If the monitoring is not masked, then the task
3665 * can be actively monitoring and we cannot directly intervene.
3666 * Therefore we use the trap mechanism to catch the task and
3667 * force it to reset the buffer/reset PMDs.
3669 * if non-blocking, then we ensure that the task will go into
3670 * pfm_handle_work() before returning to user mode.
3672 * We cannot explicitly reset another task, it MUST always
3673 * be done by the task itself. This works for system wide because
3674 * the tool that is controlling the session is logically doing
3675 * "self-monitoring".
3677 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3678 DPRINT(("unblocking [%d] \n", task_pid_nr(task)));
3679 complete(&ctx->ctx_restart_done);
3681 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3683 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3685 PFM_SET_WORK_PENDING(task, 1);
3687 tsk_set_notify_resume(task);
3690 * XXX: send reschedule if task runs on another CPU
3697 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3699 unsigned int m = *(unsigned int *)arg;
3701 pfm_sysctl.debug = m == 0 ? 0 : 1;
3703 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3706 memset(pfm_stats, 0, sizeof(pfm_stats));
3707 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3713 * arg can be NULL and count can be zero for this function
3716 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3718 struct thread_struct *thread = NULL;
3719 struct task_struct *task;
3720 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3721 unsigned long flags;
3726 int i, can_access_pmu = 0;
3727 int is_system, is_loaded;
3729 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3731 state = ctx->ctx_state;
3732 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3733 is_system = ctx->ctx_fl_system;
3734 task = ctx->ctx_task;
3736 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3739 * on both UP and SMP, we can only write to the PMC when the task is
3740 * the owner of the local PMU.
3743 thread = &task->thread;
3745 * In system wide and when the context is loaded, access can only happen
3746 * when the caller is running on the CPU being monitored by the session.
3747 * It does not have to be the owner (ctx_task) of the context per se.
3749 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3750 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3753 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3757 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3758 * ensuring that no real breakpoint can be installed via this call.
3760 * IMPORTANT: regs can be NULL in this function
3763 first_time = ctx->ctx_fl_using_dbreg == 0;
3766 * don't bother if we are loaded and task is being debugged
3768 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3769 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3774 * check for debug registers in system wide mode
3776 * If though a check is done in pfm_context_load(),
3777 * we must repeat it here, in case the registers are
3778 * written after the context is loaded
3783 if (first_time && is_system) {
3784 if (pfm_sessions.pfs_ptrace_use_dbregs)
3787 pfm_sessions.pfs_sys_use_dbregs++;
3792 if (ret != 0) return ret;
3795 * mark ourself as user of the debug registers for
3798 ctx->ctx_fl_using_dbreg = 1;
3801 * clear hardware registers to make sure we don't
3802 * pick up stale state.
3804 * for a system wide session, we do not use
3805 * thread.dbr, thread.ibr because this process
3806 * never leaves the current CPU and the state
3807 * is shared by all processes running on it
3809 if (first_time && can_access_pmu) {
3810 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3811 for (i=0; i < pmu_conf->num_ibrs; i++) {
3812 ia64_set_ibr(i, 0UL);
3813 ia64_dv_serialize_instruction();
3816 for (i=0; i < pmu_conf->num_dbrs; i++) {
3817 ia64_set_dbr(i, 0UL);
3818 ia64_dv_serialize_data();
3824 * Now install the values into the registers
3826 for (i = 0; i < count; i++, req++) {
3828 rnum = req->dbreg_num;
3829 dbreg.val = req->dbreg_value;
3833 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3834 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3835 rnum, dbreg.val, mode, i, count));
3841 * make sure we do not install enabled breakpoint
3844 if (mode == PFM_CODE_RR)
3845 dbreg.ibr.ibr_x = 0;
3847 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3850 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3853 * Debug registers, just like PMC, can only be modified
3854 * by a kernel call. Moreover, perfmon() access to those
3855 * registers are centralized in this routine. The hardware
3856 * does not modify the value of these registers, therefore,
3857 * if we save them as they are written, we can avoid having
3858 * to save them on context switch out. This is made possible
3859 * by the fact that when perfmon uses debug registers, ptrace()
3860 * won't be able to modify them concurrently.
3862 if (mode == PFM_CODE_RR) {
3863 CTX_USED_IBR(ctx, rnum);
3865 if (can_access_pmu) {
3866 ia64_set_ibr(rnum, dbreg.val);
3867 ia64_dv_serialize_instruction();
3870 ctx->ctx_ibrs[rnum] = dbreg.val;
3872 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3873 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3875 CTX_USED_DBR(ctx, rnum);
3877 if (can_access_pmu) {
3878 ia64_set_dbr(rnum, dbreg.val);
3879 ia64_dv_serialize_data();
3881 ctx->ctx_dbrs[rnum] = dbreg.val;
3883 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3884 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3892 * in case it was our first attempt, we undo the global modifications
3896 if (ctx->ctx_fl_system) {
3897 pfm_sessions.pfs_sys_use_dbregs--;
3900 ctx->ctx_fl_using_dbreg = 0;
3903 * install error return flag
3905 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3911 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3913 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3917 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3919 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3923 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3927 if (req == NULL) return -EINVAL;
3929 ctx = GET_PMU_CTX();
3931 if (ctx == NULL) return -EINVAL;
3934 * for now limit to current task, which is enough when calling
3935 * from overflow handler
3937 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3939 return pfm_write_ibrs(ctx, req, nreq, regs);
3941 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3944 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3948 if (req == NULL) return -EINVAL;
3950 ctx = GET_PMU_CTX();
3952 if (ctx == NULL) return -EINVAL;
3955 * for now limit to current task, which is enough when calling
3956 * from overflow handler
3958 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3960 return pfm_write_dbrs(ctx, req, nreq, regs);
3962 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3966 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3968 pfarg_features_t *req = (pfarg_features_t *)arg;
3970 req->ft_version = PFM_VERSION;
3975 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3977 struct pt_regs *tregs;
3978 struct task_struct *task = PFM_CTX_TASK(ctx);
3979 int state, is_system;
3981 state = ctx->ctx_state;
3982 is_system = ctx->ctx_fl_system;
3985 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3987 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3990 * In system wide and when the context is loaded, access can only happen
3991 * when the caller is running on the CPU being monitored by the session.
3992 * It does not have to be the owner (ctx_task) of the context per se.
3994 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3995 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3998 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3999 task_pid_nr(PFM_CTX_TASK(ctx)),
4003 * in system mode, we need to update the PMU directly
4004 * and the user level state of the caller, which may not
4005 * necessarily be the creator of the context.
4009 * Update local PMU first
4013 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4017 * update local cpuinfo
4019 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4022 * stop monitoring, does srlz.i
4027 * stop monitoring in the caller
4029 ia64_psr(regs)->pp = 0;
4037 if (task == current) {
4038 /* stop monitoring at kernel level */
4042 * stop monitoring at the user level
4044 ia64_psr(regs)->up = 0;
4046 tregs = task_pt_regs(task);
4049 * stop monitoring at the user level
4051 ia64_psr(tregs)->up = 0;
4054 * monitoring disabled in kernel at next reschedule
4056 ctx->ctx_saved_psr_up = 0;
4057 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4064 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4066 struct pt_regs *tregs;
4067 int state, is_system;
4069 state = ctx->ctx_state;
4070 is_system = ctx->ctx_fl_system;
4072 if (state != PFM_CTX_LOADED) return -EINVAL;
4075 * In system wide and when the context is loaded, access can only happen
4076 * when the caller is running on the CPU being monitored by the session.
4077 * It does not have to be the owner (ctx_task) of the context per se.
4079 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4080 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4085 * in system mode, we need to update the PMU directly
4086 * and the user level state of the caller, which may not
4087 * necessarily be the creator of the context.
4092 * set user level psr.pp for the caller
4094 ia64_psr(regs)->pp = 1;
4097 * now update the local PMU and cpuinfo
4099 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4102 * start monitoring at kernel level
4107 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4117 if (ctx->ctx_task == current) {
4119 /* start monitoring at kernel level */
4123 * activate monitoring at user level
4125 ia64_psr(regs)->up = 1;
4128 tregs = task_pt_regs(ctx->ctx_task);
4131 * start monitoring at the kernel level the next
4132 * time the task is scheduled
4134 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4137 * activate monitoring at user level
4139 ia64_psr(tregs)->up = 1;
4145 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4147 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4152 for (i = 0; i < count; i++, req++) {
4154 cnum = req->reg_num;
4156 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4158 req->reg_value = PMC_DFL_VAL(cnum);
4160 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4162 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4167 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4172 pfm_check_task_exist(pfm_context_t *ctx)
4174 struct task_struct *g, *t;
4177 read_lock(&tasklist_lock);
4179 do_each_thread (g, t) {
4180 if (t->thread.pfm_context == ctx) {
4184 } while_each_thread (g, t);
4186 read_unlock(&tasklist_lock);
4188 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4194 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4196 struct task_struct *task;
4197 struct thread_struct *thread;
4198 struct pfm_context_t *old;
4199 unsigned long flags;
4201 struct task_struct *owner_task = NULL;
4203 pfarg_load_t *req = (pfarg_load_t *)arg;
4204 unsigned long *pmcs_source, *pmds_source;
4207 int state, is_system, set_dbregs = 0;
4209 state = ctx->ctx_state;
4210 is_system = ctx->ctx_fl_system;
4212 * can only load from unloaded or terminated state
4214 if (state != PFM_CTX_UNLOADED) {
4215 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4221 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4223 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4224 DPRINT(("cannot use blocking mode on self\n"));
4228 ret = pfm_get_task(ctx, req->load_pid, &task);
4230 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4237 * system wide is self monitoring only
4239 if (is_system && task != current) {
4240 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4245 thread = &task->thread;
4249 * cannot load a context which is using range restrictions,
4250 * into a task that is being debugged.
4252 if (ctx->ctx_fl_using_dbreg) {
4253 if (thread->flags & IA64_THREAD_DBG_VALID) {
4255 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4261 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4262 DPRINT(("cannot load [%d] dbregs in use\n",
4263 task_pid_nr(task)));
4266 pfm_sessions.pfs_sys_use_dbregs++;
4267 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4274 if (ret) goto error;
4278 * SMP system-wide monitoring implies self-monitoring.
4280 * The programming model expects the task to
4281 * be pinned on a CPU throughout the session.
4282 * Here we take note of the current CPU at the
4283 * time the context is loaded. No call from
4284 * another CPU will be allowed.
4286 * The pinning via shed_setaffinity()
4287 * must be done by the calling task prior
4290 * systemwide: keep track of CPU this session is supposed to run on
4292 the_cpu = ctx->ctx_cpu = smp_processor_id();
4296 * now reserve the session
4298 ret = pfm_reserve_session(current, is_system, the_cpu);
4299 if (ret) goto error;
4302 * task is necessarily stopped at this point.
4304 * If the previous context was zombie, then it got removed in
4305 * pfm_save_regs(). Therefore we should not see it here.
4306 * If we see a context, then this is an active context
4308 * XXX: needs to be atomic
4310 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4311 thread->pfm_context, ctx));
4314 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4316 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4320 pfm_reset_msgq(ctx);
4322 ctx->ctx_state = PFM_CTX_LOADED;
4325 * link context to task
4327 ctx->ctx_task = task;
4331 * we load as stopped
4333 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4334 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4336 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4338 thread->flags |= IA64_THREAD_PM_VALID;
4342 * propagate into thread-state
4344 pfm_copy_pmds(task, ctx);
4345 pfm_copy_pmcs(task, ctx);
4347 pmcs_source = ctx->th_pmcs;
4348 pmds_source = ctx->th_pmds;
4351 * always the case for system-wide
4353 if (task == current) {
4355 if (is_system == 0) {
4357 /* allow user level control */
4358 ia64_psr(regs)->sp = 0;
4359 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4361 SET_LAST_CPU(ctx, smp_processor_id());
4363 SET_ACTIVATION(ctx);
4366 * push the other task out, if any
4368 owner_task = GET_PMU_OWNER();
4369 if (owner_task) pfm_lazy_save_regs(owner_task);
4373 * load all PMD from ctx to PMU (as opposed to thread state)
4374 * restore all PMC from ctx to PMU
4376 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4377 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4379 ctx->ctx_reload_pmcs[0] = 0UL;
4380 ctx->ctx_reload_pmds[0] = 0UL;
4383 * guaranteed safe by earlier check against DBG_VALID
4385 if (ctx->ctx_fl_using_dbreg) {
4386 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4387 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4392 SET_PMU_OWNER(task, ctx);
4394 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4397 * when not current, task MUST be stopped, so this is safe
4399 regs = task_pt_regs(task);
4401 /* force a full reload */
4402 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4403 SET_LAST_CPU(ctx, -1);
4405 /* initial saved psr (stopped) */
4406 ctx->ctx_saved_psr_up = 0UL;
4407 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4413 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4416 * we must undo the dbregs setting (for system-wide)
4418 if (ret && set_dbregs) {
4420 pfm_sessions.pfs_sys_use_dbregs--;
4424 * release task, there is now a link with the context
4426 if (is_system == 0 && task != current) {
4430 ret = pfm_check_task_exist(ctx);
4432 ctx->ctx_state = PFM_CTX_UNLOADED;
4433 ctx->ctx_task = NULL;
4441 * in this function, we do not need to increase the use count
4442 * for the task via get_task_struct(), because we hold the
4443 * context lock. If the task were to disappear while having
4444 * a context attached, it would go through pfm_exit_thread()
4445 * which also grabs the context lock and would therefore be blocked
4446 * until we are here.
4448 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4451 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4453 struct task_struct *task = PFM_CTX_TASK(ctx);
4454 struct pt_regs *tregs;
4455 int prev_state, is_system;
4458 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4460 prev_state = ctx->ctx_state;
4461 is_system = ctx->ctx_fl_system;
4464 * unload only when necessary
4466 if (prev_state == PFM_CTX_UNLOADED) {
4467 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4472 * clear psr and dcr bits
4474 ret = pfm_stop(ctx, NULL, 0, regs);
4475 if (ret) return ret;
4477 ctx->ctx_state = PFM_CTX_UNLOADED;
4480 * in system mode, we need to update the PMU directly
4481 * and the user level state of the caller, which may not
4482 * necessarily be the creator of the context.
4489 * local PMU is taken care of in pfm_stop()
4491 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4492 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4495 * save PMDs in context
4498 pfm_flush_pmds(current, ctx);
4501 * at this point we are done with the PMU
4502 * so we can unreserve the resource.
4504 if (prev_state != PFM_CTX_ZOMBIE)
4505 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4508 * disconnect context from task
4510 task->thread.pfm_context = NULL;
4512 * disconnect task from context
4514 ctx->ctx_task = NULL;
4517 * There is nothing more to cleanup here.
4525 tregs = task == current ? regs : task_pt_regs(task);
4527 if (task == current) {
4529 * cancel user level control
4531 ia64_psr(regs)->sp = 1;
4533 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4536 * save PMDs to context
4539 pfm_flush_pmds(task, ctx);
4542 * at this point we are done with the PMU
4543 * so we can unreserve the resource.
4545 * when state was ZOMBIE, we have already unreserved.
4547 if (prev_state != PFM_CTX_ZOMBIE)
4548 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4551 * reset activation counter and psr
4553 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4554 SET_LAST_CPU(ctx, -1);
4557 * PMU state will not be restored
4559 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4562 * break links between context and task
4564 task->thread.pfm_context = NULL;
4565 ctx->ctx_task = NULL;
4567 PFM_SET_WORK_PENDING(task, 0);
4569 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4570 ctx->ctx_fl_can_restart = 0;
4571 ctx->ctx_fl_going_zombie = 0;
4573 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4580 * called only from exit_thread(): task == current
4581 * we come here only if current has a context attached (loaded or masked)
4584 pfm_exit_thread(struct task_struct *task)
4587 unsigned long flags;
4588 struct pt_regs *regs = task_pt_regs(task);
4592 ctx = PFM_GET_CTX(task);
4594 PROTECT_CTX(ctx, flags);
4596 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4598 state = ctx->ctx_state;
4600 case PFM_CTX_UNLOADED:
4602 * only comes to this function if pfm_context is not NULL, i.e., cannot
4603 * be in unloaded state
4605 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4607 case PFM_CTX_LOADED:
4608 case PFM_CTX_MASKED:
4609 ret = pfm_context_unload(ctx, NULL, 0, regs);
4611 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4613 DPRINT(("ctx unloaded for current state was %d\n", state));
4615 pfm_end_notify_user(ctx);
4617 case PFM_CTX_ZOMBIE:
4618 ret = pfm_context_unload(ctx, NULL, 0, regs);
4620 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4625 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4628 UNPROTECT_CTX(ctx, flags);
4630 { u64 psr = pfm_get_psr();
4631 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4632 BUG_ON(GET_PMU_OWNER());
4633 BUG_ON(ia64_psr(regs)->up);
4634 BUG_ON(ia64_psr(regs)->pp);
4638 * All memory free operations (especially for vmalloc'ed memory)
4639 * MUST be done with interrupts ENABLED.
4641 if (free_ok) pfm_context_free(ctx);
4645 * functions MUST be listed in the increasing order of their index (see permfon.h)
4647 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4648 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4649 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4650 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4651 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4653 static pfm_cmd_desc_t pfm_cmd_tab[]={
4654 /* 0 */PFM_CMD_NONE,
4655 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4656 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4657 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4658 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4659 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4660 /* 6 */PFM_CMD_NONE,
4661 /* 7 */PFM_CMD_NONE,
4662 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4663 /* 9 */PFM_CMD_NONE,
4664 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4665 /* 11 */PFM_CMD_NONE,
4666 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4667 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4668 /* 14 */PFM_CMD_NONE,
4669 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4670 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4671 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4672 /* 18 */PFM_CMD_NONE,
4673 /* 19 */PFM_CMD_NONE,
4674 /* 20 */PFM_CMD_NONE,
4675 /* 21 */PFM_CMD_NONE,
4676 /* 22 */PFM_CMD_NONE,
4677 /* 23 */PFM_CMD_NONE,
4678 /* 24 */PFM_CMD_NONE,
4679 /* 25 */PFM_CMD_NONE,
4680 /* 26 */PFM_CMD_NONE,
4681 /* 27 */PFM_CMD_NONE,
4682 /* 28 */PFM_CMD_NONE,
4683 /* 29 */PFM_CMD_NONE,
4684 /* 30 */PFM_CMD_NONE,
4685 /* 31 */PFM_CMD_NONE,
4686 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4687 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4689 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4692 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4694 struct task_struct *task;
4695 int state, old_state;
4698 state = ctx->ctx_state;
4699 task = ctx->ctx_task;
4702 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4706 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4710 task->state, PFM_CMD_STOPPED(cmd)));
4713 * self-monitoring always ok.
4715 * for system-wide the caller can either be the creator of the
4716 * context (to one to which the context is attached to) OR
4717 * a task running on the same CPU as the session.
4719 if (task == current || ctx->ctx_fl_system) return 0;
4722 * we are monitoring another thread
4725 case PFM_CTX_UNLOADED:
4727 * if context is UNLOADED we are safe to go
4730 case PFM_CTX_ZOMBIE:
4732 * no command can operate on a zombie context
4734 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4736 case PFM_CTX_MASKED:
4738 * PMU state has been saved to software even though
4739 * the thread may still be running.
4741 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4745 * context is LOADED or MASKED. Some commands may need to have
4748 * We could lift this restriction for UP but it would mean that
4749 * the user has no guarantee the task would not run between
4750 * two successive calls to perfmonctl(). That's probably OK.
4751 * If this user wants to ensure the task does not run, then
4752 * the task must be stopped.
4754 if (PFM_CMD_STOPPED(cmd)) {
4755 if (!task_is_stopped_or_traced(task)) {
4756 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4760 * task is now stopped, wait for ctxsw out
4762 * This is an interesting point in the code.
4763 * We need to unprotect the context because
4764 * the pfm_save_regs() routines needs to grab
4765 * the same lock. There are danger in doing
4766 * this because it leaves a window open for
4767 * another task to get access to the context
4768 * and possibly change its state. The one thing
4769 * that is not possible is for the context to disappear
4770 * because we are protected by the VFS layer, i.e.,
4771 * get_fd()/put_fd().
4775 UNPROTECT_CTX(ctx, flags);
4777 wait_task_inactive(task);
4779 PROTECT_CTX(ctx, flags);
4782 * we must recheck to verify if state has changed
4784 if (ctx->ctx_state != old_state) {
4785 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4793 * system-call entry point (must return long)
4796 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4798 struct file *file = NULL;
4799 pfm_context_t *ctx = NULL;
4800 unsigned long flags = 0UL;
4801 void *args_k = NULL;
4802 long ret; /* will expand int return types */
4803 size_t base_sz, sz, xtra_sz = 0;
4804 int narg, completed_args = 0, call_made = 0, cmd_flags;
4805 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4806 int (*getsize)(void *arg, size_t *sz);
4807 #define PFM_MAX_ARGSIZE 4096
4810 * reject any call if perfmon was disabled at initialization
4812 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4814 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4815 DPRINT(("invalid cmd=%d\n", cmd));
4819 func = pfm_cmd_tab[cmd].cmd_func;
4820 narg = pfm_cmd_tab[cmd].cmd_narg;
4821 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4822 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4823 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4825 if (unlikely(func == NULL)) {
4826 DPRINT(("invalid cmd=%d\n", cmd));
4830 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4838 * check if number of arguments matches what the command expects
4840 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4844 sz = xtra_sz + base_sz*count;
4846 * limit abuse to min page size
4848 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4849 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4854 * allocate default-sized argument buffer
4856 if (likely(count && args_k == NULL)) {
4857 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4858 if (args_k == NULL) return -ENOMEM;
4866 * assume sz = 0 for command without parameters
4868 if (sz && copy_from_user(args_k, arg, sz)) {
4869 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4874 * check if command supports extra parameters
4876 if (completed_args == 0 && getsize) {
4878 * get extra parameters size (based on main argument)
4880 ret = (*getsize)(args_k, &xtra_sz);
4881 if (ret) goto error_args;
4885 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4887 /* retry if necessary */
4888 if (likely(xtra_sz)) goto restart_args;
4891 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4896 if (unlikely(file == NULL)) {
4897 DPRINT(("invalid fd %d\n", fd));
4900 if (unlikely(PFM_IS_FILE(file) == 0)) {
4901 DPRINT(("fd %d not related to perfmon\n", fd));
4905 ctx = (pfm_context_t *)file->private_data;
4906 if (unlikely(ctx == NULL)) {
4907 DPRINT(("no context for fd %d\n", fd));
4910 prefetch(&ctx->ctx_state);
4912 PROTECT_CTX(ctx, flags);
4915 * check task is stopped
4917 ret = pfm_check_task_state(ctx, cmd, flags);
4918 if (unlikely(ret)) goto abort_locked;
4921 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4927 DPRINT(("context unlocked\n"));
4928 UNPROTECT_CTX(ctx, flags);
4931 /* copy argument back to user, if needed */
4932 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4940 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4946 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4948 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4949 pfm_ovfl_ctrl_t rst_ctrl;
4953 state = ctx->ctx_state;
4955 * Unlock sampling buffer and reset index atomically
4956 * XXX: not really needed when blocking
4958 if (CTX_HAS_SMPL(ctx)) {
4960 rst_ctrl.bits.mask_monitoring = 0;
4961 rst_ctrl.bits.reset_ovfl_pmds = 0;
4963 if (state == PFM_CTX_LOADED)
4964 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4966 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4968 rst_ctrl.bits.mask_monitoring = 0;
4969 rst_ctrl.bits.reset_ovfl_pmds = 1;
4973 if (rst_ctrl.bits.reset_ovfl_pmds) {
4974 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4976 if (rst_ctrl.bits.mask_monitoring == 0) {
4977 DPRINT(("resuming monitoring\n"));
4978 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4980 DPRINT(("stopping monitoring\n"));
4981 //pfm_stop_monitoring(current, regs);
4983 ctx->ctx_state = PFM_CTX_LOADED;
4988 * context MUST BE LOCKED when calling
4989 * can only be called for current
4992 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4996 DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4998 ret = pfm_context_unload(ctx, NULL, 0, regs);
5000 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
5004 * and wakeup controlling task, indicating we are now disconnected
5006 wake_up_interruptible(&ctx->ctx_zombieq);
5009 * given that context is still locked, the controlling
5010 * task will only get access when we return from
5011 * pfm_handle_work().
5015 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5017 * pfm_handle_work() can be called with interrupts enabled
5018 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5019 * call may sleep, therefore we must re-enable interrupts
5020 * to avoid deadlocks. It is safe to do so because this function
5021 * is called ONLY when returning to user level (PUStk=1), in which case
5022 * there is no risk of kernel stack overflow due to deep
5023 * interrupt nesting.
5026 pfm_handle_work(void)
5029 struct pt_regs *regs;
5030 unsigned long flags, dummy_flags;
5031 unsigned long ovfl_regs;
5032 unsigned int reason;
5035 ctx = PFM_GET_CTX(current);
5037 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", task_pid_nr(current));
5041 PROTECT_CTX(ctx, flags);
5043 PFM_SET_WORK_PENDING(current, 0);
5045 tsk_clear_notify_resume(current);
5047 regs = task_pt_regs(current);
5050 * extract reason for being here and clear
5052 reason = ctx->ctx_fl_trap_reason;
5053 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5054 ovfl_regs = ctx->ctx_ovfl_regs[0];
5056 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5059 * must be done before we check for simple-reset mode
5061 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5064 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5065 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5068 * restore interrupt mask to what it was on entry.
5069 * Could be enabled/diasbled.
5071 UNPROTECT_CTX(ctx, flags);
5074 * force interrupt enable because of down_interruptible()
5078 DPRINT(("before block sleeping\n"));
5081 * may go through without blocking on SMP systems
5082 * if restart has been received already by the time we call down()
5084 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5086 DPRINT(("after block sleeping ret=%d\n", ret));
5089 * lock context and mask interrupts again
5090 * We save flags into a dummy because we may have
5091 * altered interrupts mask compared to entry in this
5094 PROTECT_CTX(ctx, dummy_flags);
5097 * we need to read the ovfl_regs only after wake-up
5098 * because we may have had pfm_write_pmds() in between
5099 * and that can changed PMD values and therefore
5100 * ovfl_regs is reset for these new PMD values.
5102 ovfl_regs = ctx->ctx_ovfl_regs[0];
5104 if (ctx->ctx_fl_going_zombie) {
5106 DPRINT(("context is zombie, bailing out\n"));
5107 pfm_context_force_terminate(ctx, regs);
5111 * in case of interruption of down() we don't restart anything
5113 if (ret < 0) goto nothing_to_do;
5116 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5117 ctx->ctx_ovfl_regs[0] = 0UL;
5121 * restore flags as they were upon entry
5123 UNPROTECT_CTX(ctx, flags);
5127 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5129 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5130 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5134 DPRINT(("waking up somebody\n"));
5136 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5139 * safe, we are not in intr handler, nor in ctxsw when
5142 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5148 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5150 pfm_msg_t *msg = NULL;
5152 if (ctx->ctx_fl_no_msg == 0) {
5153 msg = pfm_get_new_msg(ctx);
5155 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5159 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5160 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5161 msg->pfm_ovfl_msg.msg_active_set = 0;
5162 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5163 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5164 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5165 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5166 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5169 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5175 return pfm_notify_user(ctx, msg);
5179 pfm_end_notify_user(pfm_context_t *ctx)
5183 msg = pfm_get_new_msg(ctx);
5185 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5189 memset(msg, 0, sizeof(*msg));
5191 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5192 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5193 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5195 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5200 return pfm_notify_user(ctx, msg);
5204 * main overflow processing routine.
5205 * it can be called from the interrupt path or explicitly during the context switch code
5208 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5210 pfm_ovfl_arg_t *ovfl_arg;
5212 unsigned long old_val, ovfl_val, new_val;
5213 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5214 unsigned long tstamp;
5215 pfm_ovfl_ctrl_t ovfl_ctrl;
5216 unsigned int i, has_smpl;
5217 int must_notify = 0;
5219 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5222 * sanity test. Should never happen
5224 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5226 tstamp = ia64_get_itc();
5227 mask = pmc0 >> PMU_FIRST_COUNTER;
5228 ovfl_val = pmu_conf->ovfl_val;
5229 has_smpl = CTX_HAS_SMPL(ctx);
5231 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5232 "used_pmds=0x%lx\n",
5234 task ? task_pid_nr(task): -1,
5235 (regs ? regs->cr_iip : 0),
5236 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5237 ctx->ctx_used_pmds[0]));
5241 * first we update the virtual counters
5242 * assume there was a prior ia64_srlz_d() issued
5244 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5246 /* skip pmd which did not overflow */
5247 if ((mask & 0x1) == 0) continue;
5250 * Note that the pmd is not necessarily 0 at this point as qualified events
5251 * may have happened before the PMU was frozen. The residual count is not
5252 * taken into consideration here but will be with any read of the pmd via
5255 old_val = new_val = ctx->ctx_pmds[i].val;
5256 new_val += 1 + ovfl_val;
5257 ctx->ctx_pmds[i].val = new_val;
5260 * check for overflow condition
5262 if (likely(old_val > new_val)) {
5263 ovfl_pmds |= 1UL << i;
5264 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5267 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5271 ia64_get_pmd(i) & ovfl_val,
5277 * there was no 64-bit overflow, nothing else to do
5279 if (ovfl_pmds == 0UL) return;
5282 * reset all control bits
5288 * if a sampling format module exists, then we "cache" the overflow by
5289 * calling the module's handler() routine.
5292 unsigned long start_cycles, end_cycles;
5293 unsigned long pmd_mask;
5295 int this_cpu = smp_processor_id();
5297 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5298 ovfl_arg = &ctx->ctx_ovfl_arg;
5300 prefetch(ctx->ctx_smpl_hdr);
5302 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5306 if ((pmd_mask & 0x1) == 0) continue;
5308 ovfl_arg->ovfl_pmd = (unsigned char )i;
5309 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5310 ovfl_arg->active_set = 0;
5311 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5312 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5314 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5315 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5316 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5319 * copy values of pmds of interest. Sampling format may copy them
5320 * into sampling buffer.
5323 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5324 if ((smpl_pmds & 0x1) == 0) continue;
5325 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5326 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5330 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5332 start_cycles = ia64_get_itc();
5335 * call custom buffer format record (handler) routine
5337 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5339 end_cycles = ia64_get_itc();
5342 * For those controls, we take the union because they have
5343 * an all or nothing behavior.
5345 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5346 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5347 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5349 * build the bitmask of pmds to reset now
5351 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5353 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5356 * when the module cannot handle the rest of the overflows, we abort right here
5358 if (ret && pmd_mask) {
5359 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5360 pmd_mask<<PMU_FIRST_COUNTER));
5363 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5365 ovfl_pmds &= ~reset_pmds;
5368 * when no sampling module is used, then the default
5369 * is to notify on overflow if requested by user
5371 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5372 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5373 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5374 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5376 * if needed, we reset all overflowed pmds
5378 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5381 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5384 * reset the requested PMD registers using the short reset values
5387 unsigned long bm = reset_pmds;
5388 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5391 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5393 * keep track of what to reset when unblocking
5395 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5398 * check for blocking context
5400 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5402 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5405 * set the perfmon specific checking pending work for the task
5407 PFM_SET_WORK_PENDING(task, 1);
5410 * when coming from ctxsw, current still points to the
5411 * previous task, therefore we must work with task and not current.
5413 tsk_set_notify_resume(task);
5416 * defer until state is changed (shorten spin window). the context is locked
5417 * anyway, so the signal receiver would come spin for nothing.
5422 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5423 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5424 PFM_GET_WORK_PENDING(task),
5425 ctx->ctx_fl_trap_reason,
5428 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5430 * in case monitoring must be stopped, we toggle the psr bits
5432 if (ovfl_ctrl.bits.mask_monitoring) {
5433 pfm_mask_monitoring(task);
5434 ctx->ctx_state = PFM_CTX_MASKED;
5435 ctx->ctx_fl_can_restart = 1;
5439 * send notification now
5441 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5446 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5448 task ? task_pid_nr(task) : -1,
5454 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5455 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5456 * come here as zombie only if the task is the current task. In which case, we
5457 * can access the PMU hardware directly.
5459 * Note that zombies do have PM_VALID set. So here we do the minimal.
5461 * In case the context was zombified it could not be reclaimed at the time
5462 * the monitoring program exited. At this point, the PMU reservation has been
5463 * returned, the sampiing buffer has been freed. We must convert this call
5464 * into a spurious interrupt. However, we must also avoid infinite overflows
5465 * by stopping monitoring for this task. We can only come here for a per-task
5466 * context. All we need to do is to stop monitoring using the psr bits which
5467 * are always task private. By re-enabling secure montioring, we ensure that
5468 * the monitored task will not be able to re-activate monitoring.
5469 * The task will eventually be context switched out, at which point the context
5470 * will be reclaimed (that includes releasing ownership of the PMU).
5472 * So there might be a window of time where the number of per-task session is zero
5473 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5474 * context. This is safe because if a per-task session comes in, it will push this one
5475 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5476 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5477 * also push our zombie context out.
5479 * Overall pretty hairy stuff....
5481 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5483 ia64_psr(regs)->up = 0;
5484 ia64_psr(regs)->sp = 1;
5489 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5491 struct task_struct *task;
5493 unsigned long flags;
5495 int this_cpu = smp_processor_id();
5498 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5501 * srlz.d done before arriving here
5503 pmc0 = ia64_get_pmc(0);
5505 task = GET_PMU_OWNER();
5506 ctx = GET_PMU_CTX();
5509 * if we have some pending bits set
5510 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5512 if (PMC0_HAS_OVFL(pmc0) && task) {
5514 * we assume that pmc0.fr is always set here
5518 if (!ctx) goto report_spurious1;
5520 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5521 goto report_spurious2;
5523 PROTECT_CTX_NOPRINT(ctx, flags);
5525 pfm_overflow_handler(task, ctx, pmc0, regs);
5527 UNPROTECT_CTX_NOPRINT(ctx, flags);
5530 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5534 * keep it unfrozen at all times
5541 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5542 this_cpu, task_pid_nr(task));
5546 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5554 pfm_interrupt_handler(int irq, void *arg)
5556 unsigned long start_cycles, total_cycles;
5557 unsigned long min, max;
5560 struct pt_regs *regs = get_irq_regs();
5562 this_cpu = get_cpu();
5563 if (likely(!pfm_alt_intr_handler)) {
5564 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5565 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5567 start_cycles = ia64_get_itc();
5569 ret = pfm_do_interrupt_handler(arg, regs);
5571 total_cycles = ia64_get_itc();
5574 * don't measure spurious interrupts
5576 if (likely(ret == 0)) {
5577 total_cycles -= start_cycles;
5579 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5580 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5582 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5586 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5589 put_cpu_no_resched();
5594 * /proc/perfmon interface, for debug only
5597 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5600 pfm_proc_start(struct seq_file *m, loff_t *pos)
5603 return PFM_PROC_SHOW_HEADER;
5606 while (*pos <= NR_CPUS) {
5607 if (cpu_online(*pos - 1)) {
5608 return (void *)*pos;
5616 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5619 return pfm_proc_start(m, pos);
5623 pfm_proc_stop(struct seq_file *m, void *v)
5628 pfm_proc_show_header(struct seq_file *m)
5630 struct list_head * pos;
5631 pfm_buffer_fmt_t * entry;
5632 unsigned long flags;
5635 "perfmon version : %u.%u\n"
5638 "expert mode : %s\n"
5639 "ovfl_mask : 0x%lx\n"
5640 "PMU flags : 0x%x\n",
5641 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5643 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5644 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5651 "proc_sessions : %u\n"
5652 "sys_sessions : %u\n"
5653 "sys_use_dbregs : %u\n"
5654 "ptrace_use_dbregs : %u\n",
5655 pfm_sessions.pfs_task_sessions,
5656 pfm_sessions.pfs_sys_sessions,
5657 pfm_sessions.pfs_sys_use_dbregs,
5658 pfm_sessions.pfs_ptrace_use_dbregs);
5662 spin_lock(&pfm_buffer_fmt_lock);
5664 list_for_each(pos, &pfm_buffer_fmt_list) {
5665 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5666 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5677 entry->fmt_uuid[10],
5678 entry->fmt_uuid[11],
5679 entry->fmt_uuid[12],
5680 entry->fmt_uuid[13],
5681 entry->fmt_uuid[14],
5682 entry->fmt_uuid[15],
5685 spin_unlock(&pfm_buffer_fmt_lock);
5690 pfm_proc_show(struct seq_file *m, void *v)
5696 if (v == PFM_PROC_SHOW_HEADER) {
5697 pfm_proc_show_header(m);
5701 /* show info for CPU (v - 1) */
5705 "CPU%-2d overflow intrs : %lu\n"
5706 "CPU%-2d overflow cycles : %lu\n"
5707 "CPU%-2d overflow min : %lu\n"
5708 "CPU%-2d overflow max : %lu\n"
5709 "CPU%-2d smpl handler calls : %lu\n"
5710 "CPU%-2d smpl handler cycles : %lu\n"
5711 "CPU%-2d spurious intrs : %lu\n"
5712 "CPU%-2d replay intrs : %lu\n"
5713 "CPU%-2d syst_wide : %d\n"
5714 "CPU%-2d dcr_pp : %d\n"
5715 "CPU%-2d exclude idle : %d\n"
5716 "CPU%-2d owner : %d\n"
5717 "CPU%-2d context : %p\n"
5718 "CPU%-2d activations : %lu\n",
5719 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5720 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5721 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5722 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5723 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5724 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5725 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5726 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5727 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5728 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5729 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5730 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5731 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5732 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5734 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5736 psr = pfm_get_psr();
5741 "CPU%-2d psr : 0x%lx\n"
5742 "CPU%-2d pmc0 : 0x%lx\n",
5744 cpu, ia64_get_pmc(0));
5746 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5747 if (PMC_IS_COUNTING(i) == 0) continue;
5749 "CPU%-2d pmc%u : 0x%lx\n"
5750 "CPU%-2d pmd%u : 0x%lx\n",
5751 cpu, i, ia64_get_pmc(i),
5752 cpu, i, ia64_get_pmd(i));
5758 const struct seq_operations pfm_seq_ops = {
5759 .start = pfm_proc_start,
5760 .next = pfm_proc_next,
5761 .stop = pfm_proc_stop,
5762 .show = pfm_proc_show
5766 pfm_proc_open(struct inode *inode, struct file *file)
5768 return seq_open(file, &pfm_seq_ops);
5773 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5774 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5775 * is active or inactive based on mode. We must rely on the value in
5776 * local_cpu_data->pfm_syst_info
5779 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5781 struct pt_regs *regs;
5783 unsigned long dcr_pp;
5785 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5788 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5789 * on every CPU, so we can rely on the pid to identify the idle task.
5791 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5792 regs = task_pt_regs(task);
5793 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5797 * if monitoring has started
5800 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5802 * context switching in?
5805 /* mask monitoring for the idle task */
5806 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5812 * context switching out
5813 * restore monitoring for next task
5815 * Due to inlining this odd if-then-else construction generates
5818 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5827 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5829 struct task_struct *task = ctx->ctx_task;
5831 ia64_psr(regs)->up = 0;
5832 ia64_psr(regs)->sp = 1;
5834 if (GET_PMU_OWNER() == task) {
5835 DPRINT(("cleared ownership for [%d]\n",
5836 task_pid_nr(ctx->ctx_task)));
5837 SET_PMU_OWNER(NULL, NULL);
5841 * disconnect the task from the context and vice-versa
5843 PFM_SET_WORK_PENDING(task, 0);
5845 task->thread.pfm_context = NULL;
5846 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5848 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task)));
5853 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5856 pfm_save_regs(struct task_struct *task)
5859 unsigned long flags;
5863 ctx = PFM_GET_CTX(task);
5864 if (ctx == NULL) return;
5867 * we always come here with interrupts ALREADY disabled by
5868 * the scheduler. So we simply need to protect against concurrent
5869 * access, not CPU concurrency.
5871 flags = pfm_protect_ctx_ctxsw(ctx);
5873 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5874 struct pt_regs *regs = task_pt_regs(task);
5878 pfm_force_cleanup(ctx, regs);
5880 BUG_ON(ctx->ctx_smpl_hdr);
5882 pfm_unprotect_ctx_ctxsw(ctx, flags);
5884 pfm_context_free(ctx);
5889 * save current PSR: needed because we modify it
5892 psr = pfm_get_psr();
5894 BUG_ON(psr & (IA64_PSR_I));
5898 * This is the last instruction which may generate an overflow
5900 * We do not need to set psr.sp because, it is irrelevant in kernel.
5901 * It will be restored from ipsr when going back to user level
5906 * keep a copy of psr.up (for reload)
5908 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5911 * release ownership of this PMU.
5912 * PM interrupts are masked, so nothing
5915 SET_PMU_OWNER(NULL, NULL);
5918 * we systematically save the PMD as we have no
5919 * guarantee we will be schedule at that same
5922 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5925 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5926 * we will need it on the restore path to check
5927 * for pending overflow.
5929 ctx->th_pmcs[0] = ia64_get_pmc(0);
5932 * unfreeze PMU if had pending overflows
5934 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5937 * finally, allow context access.
5938 * interrupts will still be masked after this call.
5940 pfm_unprotect_ctx_ctxsw(ctx, flags);
5943 #else /* !CONFIG_SMP */
5945 pfm_save_regs(struct task_struct *task)
5950 ctx = PFM_GET_CTX(task);
5951 if (ctx == NULL) return;
5954 * save current PSR: needed because we modify it
5956 psr = pfm_get_psr();
5958 BUG_ON(psr & (IA64_PSR_I));
5962 * This is the last instruction which may generate an overflow
5964 * We do not need to set psr.sp because, it is irrelevant in kernel.
5965 * It will be restored from ipsr when going back to user level
5970 * keep a copy of psr.up (for reload)
5972 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5976 pfm_lazy_save_regs (struct task_struct *task)
5979 unsigned long flags;
5981 { u64 psr = pfm_get_psr();
5982 BUG_ON(psr & IA64_PSR_UP);
5985 ctx = PFM_GET_CTX(task);
5988 * we need to mask PMU overflow here to
5989 * make sure that we maintain pmc0 until
5990 * we save it. overflow interrupts are
5991 * treated as spurious if there is no
5994 * XXX: I don't think this is necessary
5996 PROTECT_CTX(ctx,flags);
5999 * release ownership of this PMU.
6000 * must be done before we save the registers.
6002 * after this call any PMU interrupt is treated
6005 SET_PMU_OWNER(NULL, NULL);
6008 * save all the pmds we use
6010 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6013 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6014 * it is needed to check for pended overflow
6015 * on the restore path
6017 ctx->th_pmcs[0] = ia64_get_pmc(0);
6020 * unfreeze PMU if had pending overflows
6022 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6025 * now get can unmask PMU interrupts, they will
6026 * be treated as purely spurious and we will not
6027 * lose any information
6029 UNPROTECT_CTX(ctx,flags);
6031 #endif /* CONFIG_SMP */
6035 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6038 pfm_load_regs (struct task_struct *task)
6041 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6042 unsigned long flags;
6044 int need_irq_resend;
6046 ctx = PFM_GET_CTX(task);
6047 if (unlikely(ctx == NULL)) return;
6049 BUG_ON(GET_PMU_OWNER());
6052 * possible on unload
6054 if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6057 * we always come here with interrupts ALREADY disabled by
6058 * the scheduler. So we simply need to protect against concurrent
6059 * access, not CPU concurrency.
6061 flags = pfm_protect_ctx_ctxsw(ctx);
6062 psr = pfm_get_psr();
6064 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6066 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6067 BUG_ON(psr & IA64_PSR_I);
6069 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6070 struct pt_regs *regs = task_pt_regs(task);
6072 BUG_ON(ctx->ctx_smpl_hdr);
6074 pfm_force_cleanup(ctx, regs);
6076 pfm_unprotect_ctx_ctxsw(ctx, flags);
6079 * this one (kmalloc'ed) is fine with interrupts disabled
6081 pfm_context_free(ctx);
6087 * we restore ALL the debug registers to avoid picking up
6090 if (ctx->ctx_fl_using_dbreg) {
6091 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6092 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6095 * retrieve saved psr.up
6097 psr_up = ctx->ctx_saved_psr_up;
6100 * if we were the last user of the PMU on that CPU,
6101 * then nothing to do except restore psr
6103 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6106 * retrieve partial reload masks (due to user modifications)
6108 pmc_mask = ctx->ctx_reload_pmcs[0];
6109 pmd_mask = ctx->ctx_reload_pmds[0];
6113 * To avoid leaking information to the user level when psr.sp=0,
6114 * we must reload ALL implemented pmds (even the ones we don't use).
6115 * In the kernel we only allow PFM_READ_PMDS on registers which
6116 * we initialized or requested (sampling) so there is no risk there.
6118 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6121 * ALL accessible PMCs are systematically reloaded, unused registers
6122 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6123 * up stale configuration.
6125 * PMC0 is never in the mask. It is always restored separately.
6127 pmc_mask = ctx->ctx_all_pmcs[0];
6130 * when context is MASKED, we will restore PMC with plm=0
6131 * and PMD with stale information, but that's ok, nothing
6134 * XXX: optimize here
6136 if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6137 if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6140 * check for pending overflow at the time the state
6143 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6145 * reload pmc0 with the overflow information
6146 * On McKinley PMU, this will trigger a PMU interrupt
6148 ia64_set_pmc(0, ctx->th_pmcs[0]);
6150 ctx->th_pmcs[0] = 0UL;
6153 * will replay the PMU interrupt
6155 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6157 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6161 * we just did a reload, so we reset the partial reload fields
6163 ctx->ctx_reload_pmcs[0] = 0UL;
6164 ctx->ctx_reload_pmds[0] = 0UL;
6166 SET_LAST_CPU(ctx, smp_processor_id());
6169 * dump activation value for this PMU
6173 * record current activation for this context
6175 SET_ACTIVATION(ctx);
6178 * establish new ownership.
6180 SET_PMU_OWNER(task, ctx);
6183 * restore the psr.up bit. measurement
6185 * no PMU interrupt can happen at this point
6186 * because we still have interrupts disabled.
6188 if (likely(psr_up)) pfm_set_psr_up();
6191 * allow concurrent access to context
6193 pfm_unprotect_ctx_ctxsw(ctx, flags);
6195 #else /* !CONFIG_SMP */
6197 * reload PMU state for UP kernels
6198 * in 2.5 we come here with interrupts disabled
6201 pfm_load_regs (struct task_struct *task)
6204 struct task_struct *owner;
6205 unsigned long pmd_mask, pmc_mask;
6207 int need_irq_resend;
6209 owner = GET_PMU_OWNER();
6210 ctx = PFM_GET_CTX(task);
6211 psr = pfm_get_psr();
6213 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6214 BUG_ON(psr & IA64_PSR_I);
6217 * we restore ALL the debug registers to avoid picking up
6220 * This must be done even when the task is still the owner
6221 * as the registers may have been modified via ptrace()
6222 * (not perfmon) by the previous task.
6224 if (ctx->ctx_fl_using_dbreg) {
6225 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6226 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6230 * retrieved saved psr.up
6232 psr_up = ctx->ctx_saved_psr_up;
6233 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6236 * short path, our state is still there, just
6237 * need to restore psr and we go
6239 * we do not touch either PMC nor PMD. the psr is not touched
6240 * by the overflow_handler. So we are safe w.r.t. to interrupt
6241 * concurrency even without interrupt masking.
6243 if (likely(owner == task)) {
6244 if (likely(psr_up)) pfm_set_psr_up();
6249 * someone else is still using the PMU, first push it out and
6250 * then we'll be able to install our stuff !
6252 * Upon return, there will be no owner for the current PMU
6254 if (owner) pfm_lazy_save_regs(owner);
6257 * To avoid leaking information to the user level when psr.sp=0,
6258 * we must reload ALL implemented pmds (even the ones we don't use).
6259 * In the kernel we only allow PFM_READ_PMDS on registers which
6260 * we initialized or requested (sampling) so there is no risk there.
6262 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6265 * ALL accessible PMCs are systematically reloaded, unused registers
6266 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6267 * up stale configuration.
6269 * PMC0 is never in the mask. It is always restored separately
6271 pmc_mask = ctx->ctx_all_pmcs[0];
6273 pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6274 pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6277 * check for pending overflow at the time the state
6280 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6282 * reload pmc0 with the overflow information
6283 * On McKinley PMU, this will trigger a PMU interrupt
6285 ia64_set_pmc(0, ctx->th_pmcs[0]);
6288 ctx->th_pmcs[0] = 0UL;
6291 * will replay the PMU interrupt
6293 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6295 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6299 * establish new ownership.
6301 SET_PMU_OWNER(task, ctx);
6304 * restore the psr.up bit. measurement
6306 * no PMU interrupt can happen at this point
6307 * because we still have interrupts disabled.
6309 if (likely(psr_up)) pfm_set_psr_up();
6311 #endif /* CONFIG_SMP */
6314 * this function assumes monitoring is stopped
6317 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6320 unsigned long mask2, val, pmd_val, ovfl_val;
6321 int i, can_access_pmu = 0;
6325 * is the caller the task being monitored (or which initiated the
6326 * session for system wide measurements)
6328 is_self = ctx->ctx_task == task ? 1 : 0;
6331 * can access PMU is task is the owner of the PMU state on the current CPU
6332 * or if we are running on the CPU bound to the context in system-wide mode
6333 * (that is not necessarily the task the context is attached to in this mode).
6334 * In system-wide we always have can_access_pmu true because a task running on an
6335 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6337 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6338 if (can_access_pmu) {
6340 * Mark the PMU as not owned
6341 * This will cause the interrupt handler to do nothing in case an overflow
6342 * interrupt was in-flight
6343 * This also guarantees that pmc0 will contain the final state
6344 * It virtually gives us full control on overflow processing from that point
6347 SET_PMU_OWNER(NULL, NULL);
6348 DPRINT(("releasing ownership\n"));
6351 * read current overflow status:
6353 * we are guaranteed to read the final stable state
6356 pmc0 = ia64_get_pmc(0); /* slow */
6359 * reset freeze bit, overflow status information destroyed
6363 pmc0 = ctx->th_pmcs[0];
6365 * clear whatever overflow status bits there were
6367 ctx->th_pmcs[0] = 0;
6369 ovfl_val = pmu_conf->ovfl_val;
6371 * we save all the used pmds
6372 * we take care of overflows for counting PMDs
6374 * XXX: sampling situation is not taken into account here
6376 mask2 = ctx->ctx_used_pmds[0];
6378 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6380 for (i = 0; mask2; i++, mask2>>=1) {
6382 /* skip non used pmds */
6383 if ((mask2 & 0x1) == 0) continue;
6386 * can access PMU always true in system wide mode
6388 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6390 if (PMD_IS_COUNTING(i)) {
6391 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6394 ctx->ctx_pmds[i].val,
6398 * we rebuild the full 64 bit value of the counter
6400 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6403 * now everything is in ctx_pmds[] and we need
6404 * to clear the saved context from save_regs() such that
6405 * pfm_read_pmds() gets the correct value
6410 * take care of overflow inline
6412 if (pmc0 & (1UL << i)) {
6413 val += 1 + ovfl_val;
6414 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6418 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6420 if (is_self) ctx->th_pmds[i] = pmd_val;
6422 ctx->ctx_pmds[i].val = val;
6426 static struct irqaction perfmon_irqaction = {
6427 .handler = pfm_interrupt_handler,
6428 .flags = IRQF_DISABLED,
6433 pfm_alt_save_pmu_state(void *data)
6435 struct pt_regs *regs;
6437 regs = task_pt_regs(current);
6439 DPRINT(("called\n"));
6442 * should not be necessary but
6443 * let's take not risk
6447 ia64_psr(regs)->pp = 0;
6450 * This call is required
6451 * May cause a spurious interrupt on some processors
6459 pfm_alt_restore_pmu_state(void *data)
6461 struct pt_regs *regs;
6463 regs = task_pt_regs(current);
6465 DPRINT(("called\n"));
6468 * put PMU back in state expected
6473 ia64_psr(regs)->pp = 0;
6476 * perfmon runs with PMU unfrozen at all times
6484 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6489 /* some sanity checks */
6490 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6492 /* do the easy test first */
6493 if (pfm_alt_intr_handler) return -EBUSY;
6495 /* one at a time in the install or remove, just fail the others */
6496 if (!spin_trylock(&pfm_alt_install_check)) {
6500 /* reserve our session */
6501 for_each_online_cpu(reserve_cpu) {
6502 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6503 if (ret) goto cleanup_reserve;
6506 /* save the current system wide pmu states */
6507 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 0, 1);
6509 DPRINT(("on_each_cpu() failed: %d\n", ret));
6510 goto cleanup_reserve;
6513 /* officially change to the alternate interrupt handler */
6514 pfm_alt_intr_handler = hdl;
6516 spin_unlock(&pfm_alt_install_check);
6521 for_each_online_cpu(i) {
6522 /* don't unreserve more than we reserved */
6523 if (i >= reserve_cpu) break;
6525 pfm_unreserve_session(NULL, 1, i);
6528 spin_unlock(&pfm_alt_install_check);
6532 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6535 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6540 if (hdl == NULL) return -EINVAL;
6542 /* cannot remove someone else's handler! */
6543 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6545 /* one at a time in the install or remove, just fail the others */
6546 if (!spin_trylock(&pfm_alt_install_check)) {
6550 pfm_alt_intr_handler = NULL;
6552 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 0, 1);
6554 DPRINT(("on_each_cpu() failed: %d\n", ret));
6557 for_each_online_cpu(i) {
6558 pfm_unreserve_session(NULL, 1, i);
6561 spin_unlock(&pfm_alt_install_check);
6565 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6568 * perfmon initialization routine, called from the initcall() table
6570 static int init_pfm_fs(void);
6578 family = local_cpu_data->family;
6583 if ((*p)->probe() == 0) goto found;
6584 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6595 static const struct file_operations pfm_proc_fops = {
6596 .open = pfm_proc_open,
6598 .llseek = seq_lseek,
6599 .release = seq_release,
6605 unsigned int n, n_counters, i;
6607 printk("perfmon: version %u.%u IRQ %u\n",
6610 IA64_PERFMON_VECTOR);
6612 if (pfm_probe_pmu()) {
6613 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6614 local_cpu_data->family);
6619 * compute the number of implemented PMD/PMC from the
6620 * description tables
6623 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6624 if (PMC_IS_IMPL(i) == 0) continue;
6625 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6628 pmu_conf->num_pmcs = n;
6630 n = 0; n_counters = 0;
6631 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6632 if (PMD_IS_IMPL(i) == 0) continue;
6633 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6635 if (PMD_IS_COUNTING(i)) n_counters++;
6637 pmu_conf->num_pmds = n;
6638 pmu_conf->num_counters = n_counters;
6641 * sanity checks on the number of debug registers
6643 if (pmu_conf->use_rr_dbregs) {
6644 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6645 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6649 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6650 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6656 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6660 pmu_conf->num_counters,
6661 ffz(pmu_conf->ovfl_val));
6664 if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6665 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6671 * create /proc/perfmon (mostly for debugging purposes)
6673 perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6674 if (perfmon_dir == NULL) {
6675 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6681 * create /proc/sys/kernel/perfmon (for debugging purposes)
6683 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6686 * initialize all our spinlocks
6688 spin_lock_init(&pfm_sessions.pfs_lock);
6689 spin_lock_init(&pfm_buffer_fmt_lock);
6693 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6698 __initcall(pfm_init);
6701 * this function is called before pfm_init()
6704 pfm_init_percpu (void)
6706 static int first_time=1;
6708 * make sure no measurement is active
6709 * (may inherit programmed PMCs from EFI).
6715 * we run with the PMU not frozen at all times
6720 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6724 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6729 * used for debug purposes only
6732 dump_pmu_state(const char *from)
6734 struct task_struct *task;
6735 struct pt_regs *regs;
6737 unsigned long psr, dcr, info, flags;
6740 local_irq_save(flags);
6742 this_cpu = smp_processor_id();
6743 regs = task_pt_regs(current);
6744 info = PFM_CPUINFO_GET();
6745 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6747 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6748 local_irq_restore(flags);
6752 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6755 task_pid_nr(current),
6759 task = GET_PMU_OWNER();
6760 ctx = GET_PMU_CTX();
6762 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6764 psr = pfm_get_psr();
6766 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",
6769 psr & IA64_PSR_PP ? 1 : 0,
6770 psr & IA64_PSR_UP ? 1 : 0,
6771 dcr & IA64_DCR_PP ? 1 : 0,
6774 ia64_psr(regs)->pp);
6776 ia64_psr(regs)->up = 0;
6777 ia64_psr(regs)->pp = 0;
6779 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6780 if (PMC_IS_IMPL(i) == 0) continue;
6781 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]);
6784 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6785 if (PMD_IS_IMPL(i) == 0) continue;
6786 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]);
6790 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6793 ctx->ctx_smpl_vaddr,
6797 ctx->ctx_saved_psr_up);
6799 local_irq_restore(flags);
6803 * called from process.c:copy_thread(). task is new child.
6806 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6808 struct thread_struct *thread;
6810 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6812 thread = &task->thread;
6815 * cut links inherited from parent (current)
6817 thread->pfm_context = NULL;
6819 PFM_SET_WORK_PENDING(task, 0);
6822 * the psr bits are already set properly in copy_threads()
6825 #else /* !CONFIG_PERFMON */
6827 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6831 #endif /* CONFIG_PERFMON */
projects / linux-2.6 / blob