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/config.h>
23 #include <linux/module.h>
24 #include <linux/kernel.h>
25 #include <linux/sched.h>
26 #include <linux/interrupt.h>
27 #include <linux/smp_lock.h>
28 #include <linux/proc_fs.h>
29 #include <linux/seq_file.h>
30 #include <linux/init.h>
31 #include <linux/vmalloc.h>
33 #include <linux/sysctl.h>
34 #include <linux/list.h>
35 #include <linux/file.h>
36 #include <linux/poll.h>
37 #include <linux/vfs.h>
38 #include <linux/pagemap.h>
39 #include <linux/mount.h>
40 #include <linux/version.h>
41 #include <linux/vs_memory.h>
42 #include <linux/vs_cvirt.h>
43 #include <linux/bitops.h>
44 #include <linux/vs_memory.h>
45 #include <linux/vs_cvirt.h>
47 #include <asm/errno.h>
48 #include <asm/intrinsics.h>
50 #include <asm/perfmon.h>
51 #include <asm/processor.h>
52 #include <asm/signal.h>
53 #include <asm/system.h>
54 #include <asm/uaccess.h>
55 #include <asm/delay.h>
59 * perfmon context state
61 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
62 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
63 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
64 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
66 #define PFM_INVALID_ACTIVATION (~0UL)
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_lock_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, current->pid)); \
162 spin_lock_irqsave(&(c)->ctx_lock, f); \
163 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
166 #define UNPROTECT_CTX(c, f) \
168 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
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] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); 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] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); 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 semaphore ctx_restart_sem; /* 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[IA64_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[IA64_NUM_PMD_REGS]; /* software state for PMDS */
311 u64 ctx_saved_psr_up; /* only contains psr.up value */
313 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
314 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
315 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
317 int ctx_fd; /* file descriptor used my this context */
318 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
320 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
321 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
322 unsigned long ctx_smpl_size; /* size of sampling buffer */
323 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
325 wait_queue_head_t ctx_msgq_wait;
326 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
329 struct fasync_struct *ctx_async_queue;
331 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
335 * magic number used to verify that structure is really
338 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
340 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
343 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
344 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
346 #define SET_LAST_CPU(ctx, v) do {} while(0)
347 #define GET_LAST_CPU(ctx) do {} while(0)
351 #define ctx_fl_block ctx_flags.block
352 #define ctx_fl_system ctx_flags.system
353 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
354 #define ctx_fl_is_sampling ctx_flags.is_sampling
355 #define ctx_fl_excl_idle ctx_flags.excl_idle
356 #define ctx_fl_going_zombie ctx_flags.going_zombie
357 #define ctx_fl_trap_reason ctx_flags.trap_reason
358 #define ctx_fl_no_msg ctx_flags.no_msg
359 #define ctx_fl_can_restart ctx_flags.can_restart
361 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
362 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
365 * global information about all sessions
366 * mostly used to synchronize between system wide and per-process
369 spinlock_t pfs_lock; /* lock the structure */
371 unsigned int pfs_task_sessions; /* number of per task sessions */
372 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
373 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
374 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
375 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
379 * information about a PMC or PMD.
380 * dep_pmd[]: a bitmask of dependent PMD registers
381 * dep_pmc[]: a bitmask of dependent PMC registers
383 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
387 unsigned long default_value; /* power-on default value */
388 unsigned long reserved_mask; /* bitmask of reserved bits */
389 pfm_reg_check_t read_check;
390 pfm_reg_check_t write_check;
391 unsigned long dep_pmd[4];
392 unsigned long dep_pmc[4];
395 /* assume cnum is a valid monitor */
396 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
399 * This structure is initialized at boot time and contains
400 * a description of the PMU main characteristics.
402 * If the probe function is defined, detection is based
403 * on its return value:
404 * - 0 means recognized PMU
405 * - anything else means not supported
406 * When the probe function is not defined, then the pmu_family field
407 * is used and it must match the host CPU family such that:
408 * - cpu->family & config->pmu_family != 0
411 unsigned long ovfl_val; /* overflow value for counters */
413 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
414 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
416 unsigned int num_pmcs; /* number of PMCS: computed at init time */
417 unsigned int num_pmds; /* number of PMDS: computed at init time */
418 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
419 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
421 char *pmu_name; /* PMU family name */
422 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
423 unsigned int flags; /* pmu specific flags */
424 unsigned int num_ibrs; /* number of IBRS: computed at init time */
425 unsigned int num_dbrs; /* number of DBRS: computed at init time */
426 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
427 int (*probe)(void); /* customized probe routine */
428 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
433 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
436 * debug register related type definitions
439 unsigned long ibr_mask:56;
440 unsigned long ibr_plm:4;
441 unsigned long ibr_ig:3;
442 unsigned long ibr_x:1;
446 unsigned long dbr_mask:56;
447 unsigned long dbr_plm:4;
448 unsigned long dbr_ig:2;
449 unsigned long dbr_w:1;
450 unsigned long dbr_r:1;
461 * perfmon command descriptions
464 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
467 unsigned int cmd_narg;
469 int (*cmd_getsize)(void *arg, size_t *sz);
472 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
473 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
474 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
475 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
478 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
479 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
480 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
481 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
482 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
484 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
487 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
488 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
489 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
490 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
491 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
492 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
493 unsigned long pfm_smpl_handler_calls;
494 unsigned long pfm_smpl_handler_cycles;
495 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
499 * perfmon internal variables
501 static pfm_stats_t pfm_stats[NR_CPUS];
502 static pfm_session_t pfm_sessions; /* global sessions information */
504 static spinlock_t pfm_alt_install_check = SPIN_LOCK_UNLOCKED;
505 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
507 static struct proc_dir_entry *perfmon_dir;
508 static pfm_uuid_t pfm_null_uuid = {0,};
510 static spinlock_t pfm_buffer_fmt_lock;
511 static LIST_HEAD(pfm_buffer_fmt_list);
513 static pmu_config_t *pmu_conf;
515 /* sysctl() controls */
516 pfm_sysctl_t pfm_sysctl;
517 EXPORT_SYMBOL(pfm_sysctl);
519 static ctl_table pfm_ctl_table[]={
520 {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
521 {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
522 {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
523 {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
526 static ctl_table pfm_sysctl_dir[] = {
527 {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
530 static ctl_table pfm_sysctl_root[] = {
531 {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
534 static struct ctl_table_header *pfm_sysctl_header;
536 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
537 static int pfm_flush(struct file *filp);
539 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
540 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
543 pfm_put_task(struct task_struct *task)
545 if (task != current) put_task_struct(task);
549 pfm_set_task_notify(struct task_struct *task)
551 struct thread_info *info;
553 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
554 set_bit(TIF_NOTIFY_RESUME, &info->flags);
558 pfm_clear_task_notify(void)
560 clear_thread_flag(TIF_NOTIFY_RESUME);
564 pfm_reserve_page(unsigned long a)
566 SetPageReserved(vmalloc_to_page((void *)a));
569 pfm_unreserve_page(unsigned long a)
571 ClearPageReserved(vmalloc_to_page((void*)a));
574 static inline unsigned long
575 pfm_protect_ctx_ctxsw(pfm_context_t *x)
577 spin_lock(&(x)->ctx_lock);
581 static inline unsigned long
582 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
584 spin_unlock(&(x)->ctx_lock);
587 static inline unsigned int
588 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
590 return do_munmap(mm, addr, len);
593 static inline unsigned long
594 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
596 return get_unmapped_area(file, addr, len, pgoff, flags);
600 static struct super_block *
601 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
603 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
606 static struct file_system_type pfm_fs_type = {
608 .get_sb = pfmfs_get_sb,
609 .kill_sb = kill_anon_super,
612 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
613 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
614 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
615 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
616 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
619 /* forward declaration */
620 static struct file_operations pfm_file_ops;
623 * forward declarations
626 static void pfm_lazy_save_regs (struct task_struct *ta);
629 void dump_pmu_state(const char *);
630 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
632 #include "perfmon_itanium.h"
633 #include "perfmon_mckinley.h"
634 #include "perfmon_generic.h"
636 static pmu_config_t *pmu_confs[]={
639 &pmu_conf_gen, /* must be last */
644 static int pfm_end_notify_user(pfm_context_t *ctx);
647 pfm_clear_psr_pp(void)
649 ia64_rsm(IA64_PSR_PP);
656 ia64_ssm(IA64_PSR_PP);
661 pfm_clear_psr_up(void)
663 ia64_rsm(IA64_PSR_UP);
670 ia64_ssm(IA64_PSR_UP);
674 static inline unsigned long
678 tmp = ia64_getreg(_IA64_REG_PSR);
684 pfm_set_psr_l(unsigned long val)
686 ia64_setreg(_IA64_REG_PSR_L, val);
698 pfm_unfreeze_pmu(void)
705 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
709 for (i=0; i < nibrs; i++) {
710 ia64_set_ibr(i, ibrs[i]);
711 ia64_dv_serialize_instruction();
717 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
721 for (i=0; i < ndbrs; i++) {
722 ia64_set_dbr(i, dbrs[i]);
723 ia64_dv_serialize_data();
729 * PMD[i] must be a counter. no check is made
731 static inline unsigned long
732 pfm_read_soft_counter(pfm_context_t *ctx, int i)
734 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
738 * PMD[i] must be a counter. no check is made
741 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
743 unsigned long ovfl_val = pmu_conf->ovfl_val;
745 ctx->ctx_pmds[i].val = val & ~ovfl_val;
747 * writing to unimplemented part is ignore, so we do not need to
750 ia64_set_pmd(i, val & ovfl_val);
754 pfm_get_new_msg(pfm_context_t *ctx)
758 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
760 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
761 if (next == ctx->ctx_msgq_head) return NULL;
763 idx = ctx->ctx_msgq_tail;
764 ctx->ctx_msgq_tail = next;
766 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
768 return ctx->ctx_msgq+idx;
772 pfm_get_next_msg(pfm_context_t *ctx)
776 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
778 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
783 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
788 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
790 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));
796 pfm_reset_msgq(pfm_context_t *ctx)
798 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
799 DPRINT(("ctx=%p msgq reset\n", ctx));
803 pfm_rvmalloc(unsigned long size)
808 size = PAGE_ALIGN(size);
811 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
812 memset(mem, 0, size);
813 addr = (unsigned long)mem;
815 pfm_reserve_page(addr);
824 pfm_rvfree(void *mem, unsigned long size)
829 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
830 addr = (unsigned long) mem;
831 while ((long) size > 0) {
832 pfm_unreserve_page(addr);
841 static pfm_context_t *
842 pfm_context_alloc(void)
847 * allocate context descriptor
848 * must be able to free with interrupts disabled
850 ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
852 memset(ctx, 0, sizeof(pfm_context_t));
853 DPRINT(("alloc ctx @%p\n", ctx));
859 pfm_context_free(pfm_context_t *ctx)
862 DPRINT(("free ctx @%p\n", ctx));
868 pfm_mask_monitoring(struct task_struct *task)
870 pfm_context_t *ctx = PFM_GET_CTX(task);
871 struct thread_struct *th = &task->thread;
872 unsigned long mask, val, ovfl_mask;
875 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
877 ovfl_mask = pmu_conf->ovfl_val;
879 * monitoring can only be masked as a result of a valid
880 * counter overflow. In UP, it means that the PMU still
881 * has an owner. Note that the owner can be different
882 * from the current task. However the PMU state belongs
884 * In SMP, a valid overflow only happens when task is
885 * current. Therefore if we come here, we know that
886 * the PMU state belongs to the current task, therefore
887 * we can access the live registers.
889 * So in both cases, the live register contains the owner's
890 * state. We can ONLY touch the PMU registers and NOT the PSR.
892 * As a consequence to this call, the thread->pmds[] array
893 * contains stale information which must be ignored
894 * when context is reloaded AND monitoring is active (see
897 mask = ctx->ctx_used_pmds[0];
898 for (i = 0; mask; i++, mask>>=1) {
899 /* skip non used pmds */
900 if ((mask & 0x1) == 0) continue;
901 val = ia64_get_pmd(i);
903 if (PMD_IS_COUNTING(i)) {
905 * we rebuild the full 64 bit value of the counter
907 ctx->ctx_pmds[i].val += (val & ovfl_mask);
909 ctx->ctx_pmds[i].val = val;
911 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
913 ctx->ctx_pmds[i].val,
917 * mask monitoring by setting the privilege level to 0
918 * we cannot use psr.pp/psr.up for this, it is controlled by
921 * if task is current, modify actual registers, otherwise modify
922 * thread save state, i.e., what will be restored in pfm_load_regs()
924 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
925 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
926 if ((mask & 0x1) == 0UL) continue;
927 ia64_set_pmc(i, th->pmcs[i] & ~0xfUL);
928 th->pmcs[i] &= ~0xfUL;
929 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i]));
932 * make all of this visible
938 * must always be done with task == current
940 * context must be in MASKED state when calling
943 pfm_restore_monitoring(struct task_struct *task)
945 pfm_context_t *ctx = PFM_GET_CTX(task);
946 struct thread_struct *th = &task->thread;
947 unsigned long mask, ovfl_mask;
948 unsigned long psr, val;
951 is_system = ctx->ctx_fl_system;
952 ovfl_mask = pmu_conf->ovfl_val;
954 if (task != current) {
955 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
958 if (ctx->ctx_state != PFM_CTX_MASKED) {
959 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
960 task->pid, current->pid, ctx->ctx_state);
965 * monitoring is masked via the PMC.
966 * As we restore their value, we do not want each counter to
967 * restart right away. We stop monitoring using the PSR,
968 * restore the PMC (and PMD) and then re-establish the psr
969 * as it was. Note that there can be no pending overflow at
970 * this point, because monitoring was MASKED.
972 * system-wide session are pinned and self-monitoring
974 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
976 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
982 * first, we restore the PMD
984 mask = ctx->ctx_used_pmds[0];
985 for (i = 0; mask; i++, mask>>=1) {
986 /* skip non used pmds */
987 if ((mask & 0x1) == 0) continue;
989 if (PMD_IS_COUNTING(i)) {
991 * we split the 64bit value according to
994 val = ctx->ctx_pmds[i].val & ovfl_mask;
995 ctx->ctx_pmds[i].val &= ~ovfl_mask;
997 val = ctx->ctx_pmds[i].val;
999 ia64_set_pmd(i, val);
1001 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1003 ctx->ctx_pmds[i].val,
1009 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1010 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1011 if ((mask & 0x1) == 0UL) continue;
1012 th->pmcs[i] = ctx->ctx_pmcs[i];
1013 ia64_set_pmc(i, th->pmcs[i]);
1014 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
1019 * must restore DBR/IBR because could be modified while masked
1020 * XXX: need to optimize
1022 if (ctx->ctx_fl_using_dbreg) {
1023 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1024 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1030 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1032 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1039 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1045 for (i=0; mask; i++, mask>>=1) {
1046 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1051 * reload from thread state (used for ctxw only)
1054 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1057 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1059 for (i=0; mask; i++, mask>>=1) {
1060 if ((mask & 0x1) == 0) continue;
1061 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1062 ia64_set_pmd(i, val);
1068 * propagate PMD from context to thread-state
1071 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1073 struct thread_struct *thread = &task->thread;
1074 unsigned long ovfl_val = pmu_conf->ovfl_val;
1075 unsigned long mask = ctx->ctx_all_pmds[0];
1079 DPRINT(("mask=0x%lx\n", mask));
1081 for (i=0; mask; i++, mask>>=1) {
1083 val = ctx->ctx_pmds[i].val;
1086 * We break up the 64 bit value into 2 pieces
1087 * the lower bits go to the machine state in the
1088 * thread (will be reloaded on ctxsw in).
1089 * The upper part stays in the soft-counter.
1091 if (PMD_IS_COUNTING(i)) {
1092 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1095 thread->pmds[i] = val;
1097 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1100 ctx->ctx_pmds[i].val));
1105 * propagate PMC from context to thread-state
1108 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1110 struct thread_struct *thread = &task->thread;
1111 unsigned long mask = ctx->ctx_all_pmcs[0];
1114 DPRINT(("mask=0x%lx\n", mask));
1116 for (i=0; mask; i++, mask>>=1) {
1117 /* masking 0 with ovfl_val yields 0 */
1118 thread->pmcs[i] = ctx->ctx_pmcs[i];
1119 DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
1126 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1130 for (i=0; mask; i++, mask>>=1) {
1131 if ((mask & 0x1) == 0) continue;
1132 ia64_set_pmc(i, pmcs[i]);
1138 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1140 return memcmp(a, b, sizeof(pfm_uuid_t));
1144 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1147 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1152 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1155 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1161 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1165 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1170 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1174 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1179 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1182 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1187 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)
1190 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1194 static pfm_buffer_fmt_t *
1195 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1197 struct list_head * pos;
1198 pfm_buffer_fmt_t * entry;
1200 list_for_each(pos, &pfm_buffer_fmt_list) {
1201 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1202 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1209 * find a buffer format based on its uuid
1211 static pfm_buffer_fmt_t *
1212 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1214 pfm_buffer_fmt_t * fmt;
1215 spin_lock(&pfm_buffer_fmt_lock);
1216 fmt = __pfm_find_buffer_fmt(uuid);
1217 spin_unlock(&pfm_buffer_fmt_lock);
1222 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1226 /* some sanity checks */
1227 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1229 /* we need at least a handler */
1230 if (fmt->fmt_handler == NULL) return -EINVAL;
1233 * XXX: need check validity of fmt_arg_size
1236 spin_lock(&pfm_buffer_fmt_lock);
1238 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1239 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1243 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1244 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1247 spin_unlock(&pfm_buffer_fmt_lock);
1250 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1253 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1255 pfm_buffer_fmt_t *fmt;
1258 spin_lock(&pfm_buffer_fmt_lock);
1260 fmt = __pfm_find_buffer_fmt(uuid);
1262 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1266 list_del_init(&fmt->fmt_list);
1267 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1270 spin_unlock(&pfm_buffer_fmt_lock);
1274 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1276 extern void update_pal_halt_status(int);
1279 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1281 unsigned long flags;
1283 * validy checks on cpu_mask have been done upstream
1287 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1288 pfm_sessions.pfs_sys_sessions,
1289 pfm_sessions.pfs_task_sessions,
1290 pfm_sessions.pfs_sys_use_dbregs,
1296 * cannot mix system wide and per-task sessions
1298 if (pfm_sessions.pfs_task_sessions > 0UL) {
1299 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1300 pfm_sessions.pfs_task_sessions));
1304 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1306 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1308 pfm_sessions.pfs_sys_session[cpu] = task;
1310 pfm_sessions.pfs_sys_sessions++ ;
1313 if (pfm_sessions.pfs_sys_sessions) goto abort;
1314 pfm_sessions.pfs_task_sessions++;
1317 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1318 pfm_sessions.pfs_sys_sessions,
1319 pfm_sessions.pfs_task_sessions,
1320 pfm_sessions.pfs_sys_use_dbregs,
1325 * disable default_idle() to go to PAL_HALT
1327 update_pal_halt_status(0);
1334 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1335 pfm_sessions.pfs_sys_session[cpu]->pid,
1345 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1347 unsigned long flags;
1349 * validy checks on cpu_mask have been done upstream
1353 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1354 pfm_sessions.pfs_sys_sessions,
1355 pfm_sessions.pfs_task_sessions,
1356 pfm_sessions.pfs_sys_use_dbregs,
1362 pfm_sessions.pfs_sys_session[cpu] = NULL;
1364 * would not work with perfmon+more than one bit in cpu_mask
1366 if (ctx && ctx->ctx_fl_using_dbreg) {
1367 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1368 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1370 pfm_sessions.pfs_sys_use_dbregs--;
1373 pfm_sessions.pfs_sys_sessions--;
1375 pfm_sessions.pfs_task_sessions--;
1377 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1378 pfm_sessions.pfs_sys_sessions,
1379 pfm_sessions.pfs_task_sessions,
1380 pfm_sessions.pfs_sys_use_dbregs,
1385 * if possible, enable default_idle() to go into PAL_HALT
1387 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1388 update_pal_halt_status(1);
1396 * removes virtual mapping of the sampling buffer.
1397 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1398 * a PROTECT_CTX() section.
1401 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1406 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1407 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1411 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1414 * does the actual unmapping
1416 down_write(&task->mm->mmap_sem);
1418 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1420 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1422 up_write(&task->mm->mmap_sem);
1424 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1427 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1433 * free actual physical storage used by sampling buffer
1437 pfm_free_smpl_buffer(pfm_context_t *ctx)
1439 pfm_buffer_fmt_t *fmt;
1441 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1444 * we won't use the buffer format anymore
1446 fmt = ctx->ctx_buf_fmt;
1448 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1451 ctx->ctx_smpl_vaddr));
1453 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1458 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1460 ctx->ctx_smpl_hdr = NULL;
1461 ctx->ctx_smpl_size = 0UL;
1466 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1472 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1474 if (fmt == NULL) return;
1476 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1481 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1482 * no real gain from having the whole whorehouse mounted. So we don't need
1483 * any operations on the root directory. However, we need a non-trivial
1484 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1486 static struct vfsmount *pfmfs_mnt;
1491 int err = register_filesystem(&pfm_fs_type);
1493 pfmfs_mnt = kern_mount(&pfm_fs_type);
1494 err = PTR_ERR(pfmfs_mnt);
1495 if (IS_ERR(pfmfs_mnt))
1496 unregister_filesystem(&pfm_fs_type);
1506 unregister_filesystem(&pfm_fs_type);
1511 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1516 unsigned long flags;
1517 DECLARE_WAITQUEUE(wait, current);
1518 if (PFM_IS_FILE(filp) == 0) {
1519 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1523 ctx = (pfm_context_t *)filp->private_data;
1525 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1530 * check even when there is no message
1532 if (size < sizeof(pfm_msg_t)) {
1533 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1537 PROTECT_CTX(ctx, flags);
1540 * put ourselves on the wait queue
1542 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1550 set_current_state(TASK_INTERRUPTIBLE);
1552 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1555 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1557 UNPROTECT_CTX(ctx, flags);
1560 * check non-blocking read
1563 if(filp->f_flags & O_NONBLOCK) break;
1566 * check pending signals
1568 if(signal_pending(current)) {
1573 * no message, so wait
1577 PROTECT_CTX(ctx, flags);
1579 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1580 set_current_state(TASK_RUNNING);
1581 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1583 if (ret < 0) goto abort;
1586 msg = pfm_get_next_msg(ctx);
1588 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1592 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1595 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1598 UNPROTECT_CTX(ctx, flags);
1604 pfm_write(struct file *file, const char __user *ubuf,
1605 size_t size, loff_t *ppos)
1607 DPRINT(("pfm_write called\n"));
1612 pfm_poll(struct file *filp, poll_table * wait)
1615 unsigned long flags;
1616 unsigned int mask = 0;
1618 if (PFM_IS_FILE(filp) == 0) {
1619 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1623 ctx = (pfm_context_t *)filp->private_data;
1625 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1630 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1632 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1634 PROTECT_CTX(ctx, flags);
1636 if (PFM_CTXQ_EMPTY(ctx) == 0)
1637 mask = POLLIN | POLLRDNORM;
1639 UNPROTECT_CTX(ctx, flags);
1641 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1647 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1649 DPRINT(("pfm_ioctl called\n"));
1654 * interrupt cannot be masked when coming here
1657 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1661 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1663 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1667 ctx->ctx_async_queue, ret));
1673 pfm_fasync(int fd, struct file *filp, int on)
1678 if (PFM_IS_FILE(filp) == 0) {
1679 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1683 ctx = (pfm_context_t *)filp->private_data;
1685 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1689 * we cannot mask interrupts during this call because this may
1690 * may go to sleep if memory is not readily avalaible.
1692 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1693 * done in caller. Serialization of this function is ensured by caller.
1695 ret = pfm_do_fasync(fd, filp, ctx, on);
1698 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1701 ctx->ctx_async_queue, ret));
1708 * this function is exclusively called from pfm_close().
1709 * The context is not protected at that time, nor are interrupts
1710 * on the remote CPU. That's necessary to avoid deadlocks.
1713 pfm_syswide_force_stop(void *info)
1715 pfm_context_t *ctx = (pfm_context_t *)info;
1716 struct pt_regs *regs = ia64_task_regs(current);
1717 struct task_struct *owner;
1718 unsigned long flags;
1721 if (ctx->ctx_cpu != smp_processor_id()) {
1722 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1724 smp_processor_id());
1727 owner = GET_PMU_OWNER();
1728 if (owner != ctx->ctx_task) {
1729 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1731 owner->pid, ctx->ctx_task->pid);
1734 if (GET_PMU_CTX() != ctx) {
1735 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1737 GET_PMU_CTX(), ctx);
1741 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1743 * the context is already protected in pfm_close(), we simply
1744 * need to mask interrupts to avoid a PMU interrupt race on
1747 local_irq_save(flags);
1749 ret = pfm_context_unload(ctx, NULL, 0, regs);
1751 DPRINT(("context_unload returned %d\n", ret));
1755 * unmask interrupts, PMU interrupts are now spurious here
1757 local_irq_restore(flags);
1761 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1765 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1766 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1767 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1769 #endif /* CONFIG_SMP */
1772 * called for each close(). Partially free resources.
1773 * When caller is self-monitoring, the context is unloaded.
1776 pfm_flush(struct file *filp)
1779 struct task_struct *task;
1780 struct pt_regs *regs;
1781 unsigned long flags;
1782 unsigned long smpl_buf_size = 0UL;
1783 void *smpl_buf_vaddr = NULL;
1784 int state, is_system;
1786 if (PFM_IS_FILE(filp) == 0) {
1787 DPRINT(("bad magic for\n"));
1791 ctx = (pfm_context_t *)filp->private_data;
1793 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1798 * remove our file from the async queue, if we use this mode.
1799 * This can be done without the context being protected. We come
1800 * here when the context has become unreacheable by other tasks.
1802 * We may still have active monitoring at this point and we may
1803 * end up in pfm_overflow_handler(). However, fasync_helper()
1804 * operates with interrupts disabled and it cleans up the
1805 * queue. If the PMU handler is called prior to entering
1806 * fasync_helper() then it will send a signal. If it is
1807 * invoked after, it will find an empty queue and no
1808 * signal will be sent. In both case, we are safe
1810 if (filp->f_flags & FASYNC) {
1811 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1812 pfm_do_fasync (-1, filp, ctx, 0);
1815 PROTECT_CTX(ctx, flags);
1817 state = ctx->ctx_state;
1818 is_system = ctx->ctx_fl_system;
1820 task = PFM_CTX_TASK(ctx);
1821 regs = ia64_task_regs(task);
1823 DPRINT(("ctx_state=%d is_current=%d\n",
1825 task == current ? 1 : 0));
1828 * if state == UNLOADED, then task is NULL
1832 * we must stop and unload because we are losing access to the context.
1834 if (task == current) {
1837 * the task IS the owner but it migrated to another CPU: that's bad
1838 * but we must handle this cleanly. Unfortunately, the kernel does
1839 * not provide a mechanism to block migration (while the context is loaded).
1841 * We need to release the resource on the ORIGINAL cpu.
1843 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1845 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1847 * keep context protected but unmask interrupt for IPI
1849 local_irq_restore(flags);
1851 pfm_syswide_cleanup_other_cpu(ctx);
1854 * restore interrupt masking
1856 local_irq_save(flags);
1859 * context is unloaded at this point
1862 #endif /* CONFIG_SMP */
1865 DPRINT(("forcing unload\n"));
1867 * stop and unload, returning with state UNLOADED
1868 * and session unreserved.
1870 pfm_context_unload(ctx, NULL, 0, regs);
1872 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1877 * remove virtual mapping, if any, for the calling task.
1878 * cannot reset ctx field until last user is calling close().
1880 * ctx_smpl_vaddr must never be cleared because it is needed
1881 * by every task with access to the context
1883 * When called from do_exit(), the mm context is gone already, therefore
1884 * mm is NULL, i.e., the VMA is already gone and we do not have to
1887 if (ctx->ctx_smpl_vaddr && current->mm) {
1888 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1889 smpl_buf_size = ctx->ctx_smpl_size;
1892 UNPROTECT_CTX(ctx, flags);
1895 * if there was a mapping, then we systematically remove it
1896 * at this point. Cannot be done inside critical section
1897 * because some VM function reenables interrupts.
1900 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1905 * called either on explicit close() or from exit_files().
1906 * Only the LAST user of the file gets to this point, i.e., it is
1909 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1910 * (fput()),i.e, last task to access the file. Nobody else can access the
1911 * file at this point.
1913 * When called from exit_files(), the VMA has been freed because exit_mm()
1914 * is executed before exit_files().
1916 * When called from exit_files(), the current task is not yet ZOMBIE but we
1917 * flush the PMU state to the context.
1920 pfm_close(struct inode *inode, struct file *filp)
1923 struct task_struct *task;
1924 struct pt_regs *regs;
1925 DECLARE_WAITQUEUE(wait, current);
1926 unsigned long flags;
1927 unsigned long smpl_buf_size = 0UL;
1928 void *smpl_buf_addr = NULL;
1929 int free_possible = 1;
1930 int state, is_system;
1932 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1934 if (PFM_IS_FILE(filp) == 0) {
1935 DPRINT(("bad magic\n"));
1939 ctx = (pfm_context_t *)filp->private_data;
1941 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1945 PROTECT_CTX(ctx, flags);
1947 state = ctx->ctx_state;
1948 is_system = ctx->ctx_fl_system;
1950 task = PFM_CTX_TASK(ctx);
1951 regs = ia64_task_regs(task);
1953 DPRINT(("ctx_state=%d is_current=%d\n",
1955 task == current ? 1 : 0));
1958 * if task == current, then pfm_flush() unloaded the context
1960 if (state == PFM_CTX_UNLOADED) goto doit;
1963 * context is loaded/masked and task != current, we need to
1964 * either force an unload or go zombie
1968 * The task is currently blocked or will block after an overflow.
1969 * we must force it to wakeup to get out of the
1970 * MASKED state and transition to the unloaded state by itself.
1972 * This situation is only possible for per-task mode
1974 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1977 * set a "partial" zombie state to be checked
1978 * upon return from down() in pfm_handle_work().
1980 * We cannot use the ZOMBIE state, because it is checked
1981 * by pfm_load_regs() which is called upon wakeup from down().
1982 * In such case, it would free the context and then we would
1983 * return to pfm_handle_work() which would access the
1984 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1985 * but visible to pfm_handle_work().
1987 * For some window of time, we have a zombie context with
1988 * ctx_state = MASKED and not ZOMBIE
1990 ctx->ctx_fl_going_zombie = 1;
1993 * force task to wake up from MASKED state
1995 up(&ctx->ctx_restart_sem);
1997 DPRINT(("waking up ctx_state=%d\n", state));
2000 * put ourself to sleep waiting for the other
2001 * task to report completion
2003 * the context is protected by mutex, therefore there
2004 * is no risk of being notified of completion before
2005 * begin actually on the waitq.
2007 set_current_state(TASK_INTERRUPTIBLE);
2008 add_wait_queue(&ctx->ctx_zombieq, &wait);
2010 UNPROTECT_CTX(ctx, flags);
2013 * XXX: check for signals :
2014 * - ok for explicit close
2015 * - not ok when coming from exit_files()
2020 PROTECT_CTX(ctx, flags);
2023 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2024 set_current_state(TASK_RUNNING);
2027 * context is unloaded at this point
2029 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2031 else if (task != current) {
2034 * switch context to zombie state
2036 ctx->ctx_state = PFM_CTX_ZOMBIE;
2038 DPRINT(("zombie ctx for [%d]\n", task->pid));
2040 * cannot free the context on the spot. deferred until
2041 * the task notices the ZOMBIE state
2045 pfm_context_unload(ctx, NULL, 0, regs);
2050 /* reload state, may have changed during opening of critical section */
2051 state = ctx->ctx_state;
2054 * the context is still attached to a task (possibly current)
2055 * we cannot destroy it right now
2059 * we must free the sampling buffer right here because
2060 * we cannot rely on it being cleaned up later by the
2061 * monitored task. It is not possible to free vmalloc'ed
2062 * memory in pfm_load_regs(). Instead, we remove the buffer
2063 * now. should there be subsequent PMU overflow originally
2064 * meant for sampling, the will be converted to spurious
2065 * and that's fine because the monitoring tools is gone anyway.
2067 if (ctx->ctx_smpl_hdr) {
2068 smpl_buf_addr = ctx->ctx_smpl_hdr;
2069 smpl_buf_size = ctx->ctx_smpl_size;
2070 /* no more sampling */
2071 ctx->ctx_smpl_hdr = NULL;
2072 ctx->ctx_fl_is_sampling = 0;
2075 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2081 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2084 * UNLOADED that the session has already been unreserved.
2086 if (state == PFM_CTX_ZOMBIE) {
2087 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2091 * disconnect file descriptor from context must be done
2094 filp->private_data = NULL;
2097 * if we free on the spot, the context is now completely unreacheable
2098 * from the callers side. The monitored task side is also cut, so we
2101 * If we have a deferred free, only the caller side is disconnected.
2103 UNPROTECT_CTX(ctx, flags);
2106 * All memory free operations (especially for vmalloc'ed memory)
2107 * MUST be done with interrupts ENABLED.
2109 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2112 * return the memory used by the context
2114 if (free_possible) pfm_context_free(ctx);
2120 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2122 DPRINT(("pfm_no_open called\n"));
2128 static struct file_operations pfm_file_ops = {
2129 .llseek = no_llseek,
2134 .open = pfm_no_open, /* special open code to disallow open via /proc */
2135 .fasync = pfm_fasync,
2136 .release = pfm_close,
2141 pfmfs_delete_dentry(struct dentry *dentry)
2146 static struct dentry_operations pfmfs_dentry_operations = {
2147 .d_delete = pfmfs_delete_dentry,
2152 pfm_alloc_fd(struct file **cfile)
2155 struct file *file = NULL;
2156 struct inode * inode;
2160 fd = get_unused_fd();
2161 if (fd < 0) return -ENFILE;
2165 file = get_empty_filp();
2166 if (!file) goto out;
2169 * allocate a new inode
2171 inode = new_inode(pfmfs_mnt->mnt_sb);
2172 if (!inode) goto out;
2174 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2176 inode->i_mode = S_IFCHR|S_IRUGO;
2177 inode->i_uid = current->fsuid;
2178 inode->i_gid = current->fsgid;
2180 sprintf(name, "[%lu]", inode->i_ino);
2182 this.len = strlen(name);
2183 this.hash = inode->i_ino;
2188 * allocate a new dcache entry
2190 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2191 if (!file->f_dentry) goto out;
2193 file->f_dentry->d_op = &pfmfs_dentry_operations;
2195 d_add(file->f_dentry, inode);
2196 file->f_vfsmnt = mntget(pfmfs_mnt);
2197 file->f_mapping = inode->i_mapping;
2199 file->f_op = &pfm_file_ops;
2200 file->f_mode = FMODE_READ;
2201 file->f_flags = O_RDONLY;
2205 * may have to delay until context is attached?
2207 fd_install(fd, file);
2210 * the file structure we will use
2216 if (file) put_filp(file);
2222 pfm_free_fd(int fd, struct file *file)
2224 struct files_struct *files = current->files;
2227 * there ie no fd_uninstall(), so we do it here
2229 spin_lock(&files->file_lock);
2230 files->fd[fd] = NULL;
2231 spin_unlock(&files->file_lock);
2233 if (file) put_filp(file);
2238 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2240 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2243 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2246 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2257 * allocate a sampling buffer and remaps it into the user address space of the task
2260 pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2262 struct mm_struct *mm = task->mm;
2263 struct vm_area_struct *vma = NULL;
2269 * the fixed header + requested size and align to page boundary
2271 size = PAGE_ALIGN(rsize);
2273 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2276 * check requested size to avoid Denial-of-service attacks
2277 * XXX: may have to refine this test
2278 * Check against address space limit.
2280 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2283 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2287 * We do the easy to undo allocations first.
2289 * pfm_rvmalloc(), clears the buffer, so there is no leak
2291 smpl_buf = pfm_rvmalloc(size);
2292 if (smpl_buf == NULL) {
2293 DPRINT(("Can't allocate sampling buffer\n"));
2297 DPRINT(("smpl_buf @%p\n", smpl_buf));
2300 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2302 DPRINT(("Cannot allocate vma\n"));
2305 memset(vma, 0, sizeof(*vma));
2308 * partially initialize the vma for the sampling buffer
2311 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2312 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2315 * Now we have everything we need and we can initialize
2316 * and connect all the data structures
2319 ctx->ctx_smpl_hdr = smpl_buf;
2320 ctx->ctx_smpl_size = size; /* aligned size */
2323 * Let's do the difficult operations next.
2325 * now we atomically find some area in the address space and
2326 * remap the buffer in it.
2328 down_write(&task->mm->mmap_sem);
2330 /* find some free area in address space, must have mmap sem held */
2331 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2332 if (vma->vm_start == 0UL) {
2333 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2334 up_write(&task->mm->mmap_sem);
2337 vma->vm_end = vma->vm_start + size;
2338 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2340 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2342 /* can only be applied to current task, need to have the mm semaphore held when called */
2343 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2344 DPRINT(("Can't remap buffer\n"));
2345 up_write(&task->mm->mmap_sem);
2350 * now insert the vma in the vm list for the process, must be
2351 * done with mmap lock held
2353 insert_vm_struct(mm, vma);
2355 vx_vmpages_add(mm, size >> PAGE_SHIFT);
2356 vm_stat_account(vma);
2357 up_write(&task->mm->mmap_sem);
2360 * keep track of user level virtual address
2362 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2363 *(unsigned long *)user_vaddr = vma->vm_start;
2368 kmem_cache_free(vm_area_cachep, vma);
2370 pfm_rvfree(smpl_buf, size);
2376 * XXX: do something better here
2379 pfm_bad_permissions(struct task_struct *task)
2381 /* inspired by ptrace_attach() */
2382 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2391 return ((current->uid != task->euid)
2392 || (current->uid != task->suid)
2393 || (current->uid != task->uid)
2394 || (current->gid != task->egid)
2395 || (current->gid != task->sgid)
2396 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2400 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2406 ctx_flags = pfx->ctx_flags;
2408 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2411 * cannot block in this mode
2413 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2414 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2419 /* probably more to add here */
2425 pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2426 unsigned int cpu, pfarg_context_t *arg)
2428 pfm_buffer_fmt_t *fmt = NULL;
2429 unsigned long size = 0UL;
2431 void *fmt_arg = NULL;
2433 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2435 /* invoke and lock buffer format, if found */
2436 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2438 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2443 * buffer argument MUST be contiguous to pfarg_context_t
2445 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2447 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2449 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2451 if (ret) goto error;
2453 /* link buffer format and context */
2454 ctx->ctx_buf_fmt = fmt;
2457 * check if buffer format wants to use perfmon buffer allocation/mapping service
2459 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2460 if (ret) goto error;
2464 * buffer is always remapped into the caller's address space
2466 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2467 if (ret) goto error;
2469 /* keep track of user address of buffer */
2470 arg->ctx_smpl_vaddr = uaddr;
2472 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2479 pfm_reset_pmu_state(pfm_context_t *ctx)
2484 * install reset values for PMC.
2486 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2487 if (PMC_IS_IMPL(i) == 0) continue;
2488 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2489 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2492 * PMD registers are set to 0UL when the context in memset()
2496 * On context switched restore, we must restore ALL pmc and ALL pmd even
2497 * when they are not actively used by the task. In UP, the incoming process
2498 * may otherwise pick up left over PMC, PMD state from the previous process.
2499 * As opposed to PMD, stale PMC can cause harm to the incoming
2500 * process because they may change what is being measured.
2501 * Therefore, we must systematically reinstall the entire
2502 * PMC state. In SMP, the same thing is possible on the
2503 * same CPU but also on between 2 CPUs.
2505 * The problem with PMD is information leaking especially
2506 * to user level when psr.sp=0
2508 * There is unfortunately no easy way to avoid this problem
2509 * on either UP or SMP. This definitively slows down the
2510 * pfm_load_regs() function.
2514 * bitmask of all PMCs accessible to this context
2516 * PMC0 is treated differently.
2518 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2521 * bitmask of all PMDs that are accesible to this context
2523 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2525 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2528 * useful in case of re-enable after disable
2530 ctx->ctx_used_ibrs[0] = 0UL;
2531 ctx->ctx_used_dbrs[0] = 0UL;
2535 pfm_ctx_getsize(void *arg, size_t *sz)
2537 pfarg_context_t *req = (pfarg_context_t *)arg;
2538 pfm_buffer_fmt_t *fmt;
2542 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2544 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2546 DPRINT(("cannot find buffer format\n"));
2549 /* get just enough to copy in user parameters */
2550 *sz = fmt->fmt_arg_size;
2551 DPRINT(("arg_size=%lu\n", *sz));
2559 * cannot attach if :
2561 * - task not owned by caller
2562 * - task incompatible with context mode
2565 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2568 * no kernel task or task not owner by caller
2570 if (task->mm == NULL) {
2571 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2574 if (pfm_bad_permissions(task)) {
2575 DPRINT(("no permission to attach to [%d]\n", task->pid));
2579 * cannot block in self-monitoring mode
2581 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2582 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2586 if (task->exit_state == EXIT_ZOMBIE) {
2587 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2592 * always ok for self
2594 if (task == current) return 0;
2596 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
2597 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2601 * make sure the task is off any CPU
2603 wait_task_inactive(task);
2605 /* more to come... */
2611 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2613 struct task_struct *p = current;
2616 /* XXX: need to add more checks here */
2617 if (pid < 2) return -EPERM;
2619 if (pid != current->pid) {
2621 read_lock(&tasklist_lock);
2623 p = find_task_by_pid(pid);
2625 /* make sure task cannot go away while we operate on it */
2626 if (p) get_task_struct(p);
2628 read_unlock(&tasklist_lock);
2630 if (p == NULL) return -ESRCH;
2633 ret = pfm_task_incompatible(ctx, p);
2636 } else if (p != current) {
2645 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2647 pfarg_context_t *req = (pfarg_context_t *)arg;
2652 /* let's check the arguments first */
2653 ret = pfarg_is_sane(current, req);
2654 if (ret < 0) return ret;
2656 ctx_flags = req->ctx_flags;
2660 ctx = pfm_context_alloc();
2661 if (!ctx) goto error;
2663 ret = pfm_alloc_fd(&filp);
2664 if (ret < 0) goto error_file;
2666 req->ctx_fd = ctx->ctx_fd = ret;
2669 * attach context to file
2671 filp->private_data = ctx;
2674 * does the user want to sample?
2676 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2677 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2678 if (ret) goto buffer_error;
2682 * init context protection lock
2684 spin_lock_init(&ctx->ctx_lock);
2687 * context is unloaded
2689 ctx->ctx_state = PFM_CTX_UNLOADED;
2692 * initialization of context's flags
2694 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2695 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2696 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2697 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2699 * will move to set properties
2700 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2704 * init restart semaphore to locked
2706 sema_init(&ctx->ctx_restart_sem, 0);
2709 * activation is used in SMP only
2711 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2712 SET_LAST_CPU(ctx, -1);
2715 * initialize notification message queue
2717 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2718 init_waitqueue_head(&ctx->ctx_msgq_wait);
2719 init_waitqueue_head(&ctx->ctx_zombieq);
2721 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2726 ctx->ctx_fl_excl_idle,
2731 * initialize soft PMU state
2733 pfm_reset_pmu_state(ctx);
2738 pfm_free_fd(ctx->ctx_fd, filp);
2740 if (ctx->ctx_buf_fmt) {
2741 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2744 pfm_context_free(ctx);
2750 static inline unsigned long
2751 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2753 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2754 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2755 extern unsigned long carta_random32 (unsigned long seed);
2757 if (reg->flags & PFM_REGFL_RANDOM) {
2758 new_seed = carta_random32(old_seed);
2759 val -= (old_seed & mask); /* counter values are negative numbers! */
2760 if ((mask >> 32) != 0)
2761 /* construct a full 64-bit random value: */
2762 new_seed |= carta_random32(old_seed >> 32) << 32;
2763 reg->seed = new_seed;
2770 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2772 unsigned long mask = ovfl_regs[0];
2773 unsigned long reset_others = 0UL;
2778 * now restore reset value on sampling overflowed counters
2780 mask >>= PMU_FIRST_COUNTER;
2781 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2783 if ((mask & 0x1UL) == 0UL) continue;
2785 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2786 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2788 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2792 * Now take care of resetting the other registers
2794 for(i = 0; reset_others; i++, reset_others >>= 1) {
2796 if ((reset_others & 0x1) == 0) continue;
2798 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2800 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2801 is_long_reset ? "long" : "short", i, val));
2806 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2808 unsigned long mask = ovfl_regs[0];
2809 unsigned long reset_others = 0UL;
2813 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2815 if (ctx->ctx_state == PFM_CTX_MASKED) {
2816 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2821 * now restore reset value on sampling overflowed counters
2823 mask >>= PMU_FIRST_COUNTER;
2824 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2826 if ((mask & 0x1UL) == 0UL) continue;
2828 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2829 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2831 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2833 pfm_write_soft_counter(ctx, i, val);
2837 * Now take care of resetting the other registers
2839 for(i = 0; reset_others; i++, reset_others >>= 1) {
2841 if ((reset_others & 0x1) == 0) continue;
2843 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2845 if (PMD_IS_COUNTING(i)) {
2846 pfm_write_soft_counter(ctx, i, val);
2848 ia64_set_pmd(i, val);
2850 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2851 is_long_reset ? "long" : "short", i, val));
2857 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2859 struct thread_struct *thread = NULL;
2860 struct task_struct *task;
2861 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2862 unsigned long value, pmc_pm;
2863 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2864 unsigned int cnum, reg_flags, flags, pmc_type;
2865 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2866 int is_monitor, is_counting, state;
2868 pfm_reg_check_t wr_func;
2869 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2871 state = ctx->ctx_state;
2872 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2873 is_system = ctx->ctx_fl_system;
2874 task = ctx->ctx_task;
2875 impl_pmds = pmu_conf->impl_pmds[0];
2877 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2880 thread = &task->thread;
2882 * In system wide and when the context is loaded, access can only happen
2883 * when the caller is running on the CPU being monitored by the session.
2884 * It does not have to be the owner (ctx_task) of the context per se.
2886 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2887 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2890 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2892 expert_mode = pfm_sysctl.expert_mode;
2894 for (i = 0; i < count; i++, req++) {
2896 cnum = req->reg_num;
2897 reg_flags = req->reg_flags;
2898 value = req->reg_value;
2899 smpl_pmds = req->reg_smpl_pmds[0];
2900 reset_pmds = req->reg_reset_pmds[0];
2904 if (cnum >= PMU_MAX_PMCS) {
2905 DPRINT(("pmc%u is invalid\n", cnum));
2909 pmc_type = pmu_conf->pmc_desc[cnum].type;
2910 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2911 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2912 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2915 * we reject all non implemented PMC as well
2916 * as attempts to modify PMC[0-3] which are used
2917 * as status registers by the PMU
2919 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2920 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2923 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2925 * If the PMC is a monitor, then if the value is not the default:
2926 * - system-wide session: PMCx.pm=1 (privileged monitor)
2927 * - per-task : PMCx.pm=0 (user monitor)
2929 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2930 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2939 * enforce generation of overflow interrupt. Necessary on all
2942 value |= 1 << PMU_PMC_OI;
2944 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2945 flags |= PFM_REGFL_OVFL_NOTIFY;
2948 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2950 /* verify validity of smpl_pmds */
2951 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2952 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2956 /* verify validity of reset_pmds */
2957 if ((reset_pmds & impl_pmds) != reset_pmds) {
2958 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2962 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2963 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2966 /* eventid on non-counting monitors are ignored */
2970 * execute write checker, if any
2972 if (likely(expert_mode == 0 && wr_func)) {
2973 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2974 if (ret) goto error;
2979 * no error on this register
2981 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2984 * Now we commit the changes to the software state
2988 * update overflow information
2992 * full flag update each time a register is programmed
2994 ctx->ctx_pmds[cnum].flags = flags;
2996 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2997 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2998 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
3001 * Mark all PMDS to be accessed as used.
3003 * We do not keep track of PMC because we have to
3004 * systematically restore ALL of them.
3006 * We do not update the used_monitors mask, because
3007 * if we have not programmed them, then will be in
3008 * a quiescent state, therefore we will not need to
3009 * mask/restore then when context is MASKED.
3011 CTX_USED_PMD(ctx, reset_pmds);
3012 CTX_USED_PMD(ctx, smpl_pmds);
3014 * make sure we do not try to reset on
3015 * restart because we have established new values
3017 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3020 * Needed in case the user does not initialize the equivalent
3021 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3022 * possible leak here.
3024 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3027 * keep track of the monitor PMC that we are using.
3028 * we save the value of the pmc in ctx_pmcs[] and if
3029 * the monitoring is not stopped for the context we also
3030 * place it in the saved state area so that it will be
3031 * picked up later by the context switch code.
3033 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3035 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3036 * monitoring needs to be stopped.
3038 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3041 * update context state
3043 ctx->ctx_pmcs[cnum] = value;
3047 * write thread state
3049 if (is_system == 0) thread->pmcs[cnum] = value;
3052 * write hardware register if we can
3054 if (can_access_pmu) {
3055 ia64_set_pmc(cnum, value);
3060 * per-task SMP only here
3062 * we are guaranteed that the task is not running on the other CPU,
3063 * we indicate that this PMD will need to be reloaded if the task
3064 * is rescheduled on the CPU it ran last on.
3066 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3071 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",
3077 ctx->ctx_all_pmcs[0],
3078 ctx->ctx_used_pmds[0],
3079 ctx->ctx_pmds[cnum].eventid,
3082 ctx->ctx_reload_pmcs[0],
3083 ctx->ctx_used_monitors[0],
3084 ctx->ctx_ovfl_regs[0]));
3088 * make sure the changes are visible
3090 if (can_access_pmu) ia64_srlz_d();
3094 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3099 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3101 struct thread_struct *thread = NULL;
3102 struct task_struct *task;
3103 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3104 unsigned long value, hw_value, ovfl_mask;
3106 int i, can_access_pmu = 0, state;
3107 int is_counting, is_loaded, is_system, expert_mode;
3109 pfm_reg_check_t wr_func;
3112 state = ctx->ctx_state;
3113 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3114 is_system = ctx->ctx_fl_system;
3115 ovfl_mask = pmu_conf->ovfl_val;
3116 task = ctx->ctx_task;
3118 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3121 * on both UP and SMP, we can only write to the PMC when the task is
3122 * the owner of the local PMU.
3124 if (likely(is_loaded)) {
3125 thread = &task->thread;
3127 * In system wide and when the context is loaded, access can only happen
3128 * when the caller is running on the CPU being monitored by the session.
3129 * It does not have to be the owner (ctx_task) of the context per se.
3131 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3132 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3135 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3137 expert_mode = pfm_sysctl.expert_mode;
3139 for (i = 0; i < count; i++, req++) {
3141 cnum = req->reg_num;
3142 value = req->reg_value;
3144 if (!PMD_IS_IMPL(cnum)) {
3145 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3148 is_counting = PMD_IS_COUNTING(cnum);
3149 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3152 * execute write checker, if any
3154 if (unlikely(expert_mode == 0 && wr_func)) {
3155 unsigned long v = value;
3157 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3158 if (ret) goto abort_mission;
3165 * no error on this register
3167 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3170 * now commit changes to software state
3175 * update virtualized (64bits) counter
3179 * write context state
3181 ctx->ctx_pmds[cnum].lval = value;
3184 * when context is load we use the split value
3187 hw_value = value & ovfl_mask;
3188 value = value & ~ovfl_mask;
3192 * update reset values (not just for counters)
3194 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3195 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3198 * update randomization parameters (not just for counters)
3200 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3201 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3204 * update context value
3206 ctx->ctx_pmds[cnum].val = value;
3209 * Keep track of what we use
3211 * We do not keep track of PMC because we have to
3212 * systematically restore ALL of them.
3214 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3217 * mark this PMD register used as well
3219 CTX_USED_PMD(ctx, RDEP(cnum));
3222 * make sure we do not try to reset on
3223 * restart because we have established new values
3225 if (is_counting && state == PFM_CTX_MASKED) {
3226 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3231 * write thread state
3233 if (is_system == 0) thread->pmds[cnum] = hw_value;
3236 * write hardware register if we can
3238 if (can_access_pmu) {
3239 ia64_set_pmd(cnum, hw_value);
3243 * we are guaranteed that the task is not running on the other CPU,
3244 * we indicate that this PMD will need to be reloaded if the task
3245 * is rescheduled on the CPU it ran last on.
3247 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3252 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3253 "long_reset=0x%lx notify=%c seed=0x%lx mask=0x%lx used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3259 ctx->ctx_pmds[cnum].val,
3260 ctx->ctx_pmds[cnum].short_reset,
3261 ctx->ctx_pmds[cnum].long_reset,
3262 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3263 ctx->ctx_pmds[cnum].seed,
3264 ctx->ctx_pmds[cnum].mask,
3265 ctx->ctx_used_pmds[0],
3266 ctx->ctx_pmds[cnum].reset_pmds[0],
3267 ctx->ctx_reload_pmds[0],
3268 ctx->ctx_all_pmds[0],
3269 ctx->ctx_ovfl_regs[0]));
3273 * make changes visible
3275 if (can_access_pmu) ia64_srlz_d();
3281 * for now, we have only one possibility for error
3283 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3288 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3289 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3290 * interrupt is delivered during the call, it will be kept pending until we leave, making
3291 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3292 * guaranteed to return consistent data to the user, it may simply be old. It is not
3293 * trivial to treat the overflow while inside the call because you may end up in
3294 * some module sampling buffer code causing deadlocks.
3297 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3299 struct thread_struct *thread = NULL;
3300 struct task_struct *task;
3301 unsigned long val = 0UL, lval, ovfl_mask, sval;
3302 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3303 unsigned int cnum, reg_flags = 0;
3304 int i, can_access_pmu = 0, state;
3305 int is_loaded, is_system, is_counting, expert_mode;
3307 pfm_reg_check_t rd_func;
3310 * access is possible when loaded only for
3311 * self-monitoring tasks or in UP mode
3314 state = ctx->ctx_state;
3315 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3316 is_system = ctx->ctx_fl_system;
3317 ovfl_mask = pmu_conf->ovfl_val;
3318 task = ctx->ctx_task;
3320 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3322 if (likely(is_loaded)) {
3323 thread = &task->thread;
3325 * In system wide and when the context is loaded, access can only happen
3326 * when the caller is running on the CPU being monitored by the session.
3327 * It does not have to be the owner (ctx_task) of the context per se.
3329 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3330 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3334 * this can be true when not self-monitoring only in UP
3336 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3338 if (can_access_pmu) ia64_srlz_d();
3340 expert_mode = pfm_sysctl.expert_mode;
3342 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3348 * on both UP and SMP, we can only read the PMD from the hardware register when
3349 * the task is the owner of the local PMU.
3352 for (i = 0; i < count; i++, req++) {
3354 cnum = req->reg_num;
3355 reg_flags = req->reg_flags;
3357 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3359 * we can only read the register that we use. That includes
3360 * the one we explicitely initialize AND the one we want included
3361 * in the sampling buffer (smpl_regs).
3363 * Having this restriction allows optimization in the ctxsw routine
3364 * without compromising security (leaks)
3366 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3368 sval = ctx->ctx_pmds[cnum].val;
3369 lval = ctx->ctx_pmds[cnum].lval;
3370 is_counting = PMD_IS_COUNTING(cnum);
3373 * If the task is not the current one, then we check if the
3374 * PMU state is still in the local live register due to lazy ctxsw.
3375 * If true, then we read directly from the registers.
3377 if (can_access_pmu){
3378 val = ia64_get_pmd(cnum);
3381 * context has been saved
3382 * if context is zombie, then task does not exist anymore.
3383 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3385 val = is_loaded ? thread->pmds[cnum] : 0UL;
3387 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3391 * XXX: need to check for overflow when loaded
3398 * execute read checker, if any
3400 if (unlikely(expert_mode == 0 && rd_func)) {
3401 unsigned long v = val;
3402 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3403 if (ret) goto error;
3408 PFM_REG_RETFLAG_SET(reg_flags, 0);
3410 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3413 * update register return value, abort all if problem during copy.
3414 * we only modify the reg_flags field. no check mode is fine because
3415 * access has been verified upfront in sys_perfmonctl().
3417 req->reg_value = val;
3418 req->reg_flags = reg_flags;
3419 req->reg_last_reset_val = lval;
3425 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3430 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3434 if (req == NULL) return -EINVAL;
3436 ctx = GET_PMU_CTX();
3438 if (ctx == NULL) return -EINVAL;
3441 * for now limit to current task, which is enough when calling
3442 * from overflow handler
3444 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3446 return pfm_write_pmcs(ctx, req, nreq, regs);
3448 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3451 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3455 if (req == NULL) return -EINVAL;
3457 ctx = GET_PMU_CTX();
3459 if (ctx == NULL) return -EINVAL;
3462 * for now limit to current task, which is enough when calling
3463 * from overflow handler
3465 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3467 return pfm_read_pmds(ctx, req, nreq, regs);
3469 EXPORT_SYMBOL(pfm_mod_read_pmds);
3472 * Only call this function when a process it trying to
3473 * write the debug registers (reading is always allowed)
3476 pfm_use_debug_registers(struct task_struct *task)
3478 pfm_context_t *ctx = task->thread.pfm_context;
3479 unsigned long flags;
3482 if (pmu_conf->use_rr_dbregs == 0) return 0;
3484 DPRINT(("called for [%d]\n", task->pid));
3489 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3492 * Even on SMP, we do not need to use an atomic here because
3493 * the only way in is via ptrace() and this is possible only when the
3494 * process is stopped. Even in the case where the ctxsw out is not totally
3495 * completed by the time we come here, there is no way the 'stopped' process
3496 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3497 * So this is always safe.
3499 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3504 * We cannot allow setting breakpoints when system wide monitoring
3505 * sessions are using the debug registers.
3507 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3510 pfm_sessions.pfs_ptrace_use_dbregs++;
3512 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3513 pfm_sessions.pfs_ptrace_use_dbregs,
3514 pfm_sessions.pfs_sys_use_dbregs,
3523 * This function is called for every task that exits with the
3524 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3525 * able to use the debug registers for debugging purposes via
3526 * ptrace(). Therefore we know it was not using them for
3527 * perfmormance monitoring, so we only decrement the number
3528 * of "ptraced" debug register users to keep the count up to date
3531 pfm_release_debug_registers(struct task_struct *task)
3533 unsigned long flags;
3536 if (pmu_conf->use_rr_dbregs == 0) return 0;
3539 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3540 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3543 pfm_sessions.pfs_ptrace_use_dbregs--;
3552 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3554 struct task_struct *task;
3555 pfm_buffer_fmt_t *fmt;
3556 pfm_ovfl_ctrl_t rst_ctrl;
3557 int state, is_system;
3560 state = ctx->ctx_state;
3561 fmt = ctx->ctx_buf_fmt;
3562 is_system = ctx->ctx_fl_system;
3563 task = PFM_CTX_TASK(ctx);
3566 case PFM_CTX_MASKED:
3568 case PFM_CTX_LOADED:
3569 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3571 case PFM_CTX_UNLOADED:
3572 case PFM_CTX_ZOMBIE:
3573 DPRINT(("invalid state=%d\n", state));
3576 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3581 * In system wide and when the context is loaded, access can only happen
3582 * when the caller is running on the CPU being monitored by the session.
3583 * It does not have to be the owner (ctx_task) of the context per se.
3585 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3586 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3591 if (unlikely(task == NULL)) {
3592 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3596 if (task == current || is_system) {
3598 fmt = ctx->ctx_buf_fmt;
3600 DPRINT(("restarting self %d ovfl=0x%lx\n",
3602 ctx->ctx_ovfl_regs[0]));
3604 if (CTX_HAS_SMPL(ctx)) {
3606 prefetch(ctx->ctx_smpl_hdr);
3608 rst_ctrl.bits.mask_monitoring = 0;
3609 rst_ctrl.bits.reset_ovfl_pmds = 0;
3611 if (state == PFM_CTX_LOADED)
3612 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3614 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3616 rst_ctrl.bits.mask_monitoring = 0;
3617 rst_ctrl.bits.reset_ovfl_pmds = 1;
3621 if (rst_ctrl.bits.reset_ovfl_pmds)
3622 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3624 if (rst_ctrl.bits.mask_monitoring == 0) {
3625 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3627 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3629 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3631 // cannot use pfm_stop_monitoring(task, regs);
3635 * clear overflowed PMD mask to remove any stale information
3637 ctx->ctx_ovfl_regs[0] = 0UL;
3640 * back to LOADED state
3642 ctx->ctx_state = PFM_CTX_LOADED;
3645 * XXX: not really useful for self monitoring
3647 ctx->ctx_fl_can_restart = 0;
3653 * restart another task
3657 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3658 * one is seen by the task.
3660 if (state == PFM_CTX_MASKED) {
3661 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3663 * will prevent subsequent restart before this one is
3664 * seen by other task
3666 ctx->ctx_fl_can_restart = 0;
3670 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3671 * the task is blocked or on its way to block. That's the normal
3672 * restart path. If the monitoring is not masked, then the task
3673 * can be actively monitoring and we cannot directly intervene.
3674 * Therefore we use the trap mechanism to catch the task and
3675 * force it to reset the buffer/reset PMDs.
3677 * if non-blocking, then we ensure that the task will go into
3678 * pfm_handle_work() before returning to user mode.
3680 * We cannot explicitely reset another task, it MUST always
3681 * be done by the task itself. This works for system wide because
3682 * the tool that is controlling the session is logically doing
3683 * "self-monitoring".
3685 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3686 DPRINT(("unblocking [%d] \n", task->pid));
3687 up(&ctx->ctx_restart_sem);
3689 DPRINT(("[%d] armed exit trap\n", task->pid));
3691 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3693 PFM_SET_WORK_PENDING(task, 1);
3695 pfm_set_task_notify(task);
3698 * XXX: send reschedule if task runs on another CPU
3705 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3707 unsigned int m = *(unsigned int *)arg;
3709 pfm_sysctl.debug = m == 0 ? 0 : 1;
3711 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3714 memset(pfm_stats, 0, sizeof(pfm_stats));
3715 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3721 * arg can be NULL and count can be zero for this function
3724 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3726 struct thread_struct *thread = NULL;
3727 struct task_struct *task;
3728 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3729 unsigned long flags;
3734 int i, can_access_pmu = 0;
3735 int is_system, is_loaded;
3737 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3739 state = ctx->ctx_state;
3740 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3741 is_system = ctx->ctx_fl_system;
3742 task = ctx->ctx_task;
3744 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3747 * on both UP and SMP, we can only write to the PMC when the task is
3748 * the owner of the local PMU.
3751 thread = &task->thread;
3753 * In system wide and when the context is loaded, access can only happen
3754 * when the caller is running on the CPU being monitored by the session.
3755 * It does not have to be the owner (ctx_task) of the context per se.
3757 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3758 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3761 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3765 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3766 * ensuring that no real breakpoint can be installed via this call.
3768 * IMPORTANT: regs can be NULL in this function
3771 first_time = ctx->ctx_fl_using_dbreg == 0;
3774 * don't bother if we are loaded and task is being debugged
3776 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3777 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3782 * check for debug registers in system wide mode
3784 * If though a check is done in pfm_context_load(),
3785 * we must repeat it here, in case the registers are
3786 * written after the context is loaded
3791 if (first_time && is_system) {
3792 if (pfm_sessions.pfs_ptrace_use_dbregs)
3795 pfm_sessions.pfs_sys_use_dbregs++;
3800 if (ret != 0) return ret;
3803 * mark ourself as user of the debug registers for
3806 ctx->ctx_fl_using_dbreg = 1;
3809 * clear hardware registers to make sure we don't
3810 * pick up stale state.
3812 * for a system wide session, we do not use
3813 * thread.dbr, thread.ibr because this process
3814 * never leaves the current CPU and the state
3815 * is shared by all processes running on it
3817 if (first_time && can_access_pmu) {
3818 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3819 for (i=0; i < pmu_conf->num_ibrs; i++) {
3820 ia64_set_ibr(i, 0UL);
3821 ia64_dv_serialize_instruction();
3824 for (i=0; i < pmu_conf->num_dbrs; i++) {
3825 ia64_set_dbr(i, 0UL);
3826 ia64_dv_serialize_data();
3832 * Now install the values into the registers
3834 for (i = 0; i < count; i++, req++) {
3836 rnum = req->dbreg_num;
3837 dbreg.val = req->dbreg_value;
3841 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3842 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3843 rnum, dbreg.val, mode, i, count));
3849 * make sure we do not install enabled breakpoint
3852 if (mode == PFM_CODE_RR)
3853 dbreg.ibr.ibr_x = 0;
3855 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3858 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3861 * Debug registers, just like PMC, can only be modified
3862 * by a kernel call. Moreover, perfmon() access to those
3863 * registers are centralized in this routine. The hardware
3864 * does not modify the value of these registers, therefore,
3865 * if we save them as they are written, we can avoid having
3866 * to save them on context switch out. This is made possible
3867 * by the fact that when perfmon uses debug registers, ptrace()
3868 * won't be able to modify them concurrently.
3870 if (mode == PFM_CODE_RR) {
3871 CTX_USED_IBR(ctx, rnum);
3873 if (can_access_pmu) {
3874 ia64_set_ibr(rnum, dbreg.val);
3875 ia64_dv_serialize_instruction();
3878 ctx->ctx_ibrs[rnum] = dbreg.val;
3880 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3881 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3883 CTX_USED_DBR(ctx, rnum);
3885 if (can_access_pmu) {
3886 ia64_set_dbr(rnum, dbreg.val);
3887 ia64_dv_serialize_data();
3889 ctx->ctx_dbrs[rnum] = dbreg.val;
3891 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3892 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3900 * in case it was our first attempt, we undo the global modifications
3904 if (ctx->ctx_fl_system) {
3905 pfm_sessions.pfs_sys_use_dbregs--;
3908 ctx->ctx_fl_using_dbreg = 0;
3911 * install error return flag
3913 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3919 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3921 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3925 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3927 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3931 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3935 if (req == NULL) return -EINVAL;
3937 ctx = GET_PMU_CTX();
3939 if (ctx == NULL) return -EINVAL;
3942 * for now limit to current task, which is enough when calling
3943 * from overflow handler
3945 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3947 return pfm_write_ibrs(ctx, req, nreq, regs);
3949 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3952 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3956 if (req == NULL) return -EINVAL;
3958 ctx = GET_PMU_CTX();
3960 if (ctx == NULL) return -EINVAL;
3963 * for now limit to current task, which is enough when calling
3964 * from overflow handler
3966 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3968 return pfm_write_dbrs(ctx, req, nreq, regs);
3970 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3974 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3976 pfarg_features_t *req = (pfarg_features_t *)arg;
3978 req->ft_version = PFM_VERSION;
3983 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3985 struct pt_regs *tregs;
3986 struct task_struct *task = PFM_CTX_TASK(ctx);
3987 int state, is_system;
3989 state = ctx->ctx_state;
3990 is_system = ctx->ctx_fl_system;
3993 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3995 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3998 * In system wide and when the context is loaded, access can only happen
3999 * when the caller is running on the CPU being monitored by the session.
4000 * It does not have to be the owner (ctx_task) of the context per se.
4002 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4003 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4006 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4007 PFM_CTX_TASK(ctx)->pid,
4011 * in system mode, we need to update the PMU directly
4012 * and the user level state of the caller, which may not
4013 * necessarily be the creator of the context.
4017 * Update local PMU first
4021 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4025 * update local cpuinfo
4027 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4030 * stop monitoring, does srlz.i
4035 * stop monitoring in the caller
4037 ia64_psr(regs)->pp = 0;
4045 if (task == current) {
4046 /* stop monitoring at kernel level */
4050 * stop monitoring at the user level
4052 ia64_psr(regs)->up = 0;
4054 tregs = ia64_task_regs(task);
4057 * stop monitoring at the user level
4059 ia64_psr(tregs)->up = 0;
4062 * monitoring disabled in kernel at next reschedule
4064 ctx->ctx_saved_psr_up = 0;
4065 DPRINT(("task=[%d]\n", task->pid));
4072 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4074 struct pt_regs *tregs;
4075 int state, is_system;
4077 state = ctx->ctx_state;
4078 is_system = ctx->ctx_fl_system;
4080 if (state != PFM_CTX_LOADED) return -EINVAL;
4083 * In system wide and when the context is loaded, access can only happen
4084 * when the caller is running on the CPU being monitored by the session.
4085 * It does not have to be the owner (ctx_task) of the context per se.
4087 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4088 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4093 * in system mode, we need to update the PMU directly
4094 * and the user level state of the caller, which may not
4095 * necessarily be the creator of the context.
4100 * set user level psr.pp for the caller
4102 ia64_psr(regs)->pp = 1;
4105 * now update the local PMU and cpuinfo
4107 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4110 * start monitoring at kernel level
4115 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4125 if (ctx->ctx_task == current) {
4127 /* start monitoring at kernel level */
4131 * activate monitoring at user level
4133 ia64_psr(regs)->up = 1;
4136 tregs = ia64_task_regs(ctx->ctx_task);
4139 * start monitoring at the kernel level the next
4140 * time the task is scheduled
4142 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4145 * activate monitoring at user level
4147 ia64_psr(tregs)->up = 1;
4153 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4155 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4160 for (i = 0; i < count; i++, req++) {
4162 cnum = req->reg_num;
4164 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4166 req->reg_value = PMC_DFL_VAL(cnum);
4168 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4170 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4175 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4180 pfm_check_task_exist(pfm_context_t *ctx)
4182 struct task_struct *g, *t;
4185 read_lock(&tasklist_lock);
4187 do_each_thread (g, t) {
4188 if (t->thread.pfm_context == ctx) {
4192 } while_each_thread (g, t);
4194 read_unlock(&tasklist_lock);
4196 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4202 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4204 struct task_struct *task;
4205 struct thread_struct *thread;
4206 struct pfm_context_t *old;
4207 unsigned long flags;
4209 struct task_struct *owner_task = NULL;
4211 pfarg_load_t *req = (pfarg_load_t *)arg;
4212 unsigned long *pmcs_source, *pmds_source;
4215 int state, is_system, set_dbregs = 0;
4217 state = ctx->ctx_state;
4218 is_system = ctx->ctx_fl_system;
4220 * can only load from unloaded or terminated state
4222 if (state != PFM_CTX_UNLOADED) {
4223 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4229 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4231 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4232 DPRINT(("cannot use blocking mode on self\n"));
4236 ret = pfm_get_task(ctx, req->load_pid, &task);
4238 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4245 * system wide is self monitoring only
4247 if (is_system && task != current) {
4248 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4253 thread = &task->thread;
4257 * cannot load a context which is using range restrictions,
4258 * into a task that is being debugged.
4260 if (ctx->ctx_fl_using_dbreg) {
4261 if (thread->flags & IA64_THREAD_DBG_VALID) {
4263 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4269 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4270 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4273 pfm_sessions.pfs_sys_use_dbregs++;
4274 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4281 if (ret) goto error;
4285 * SMP system-wide monitoring implies self-monitoring.
4287 * The programming model expects the task to
4288 * be pinned on a CPU throughout the session.
4289 * Here we take note of the current CPU at the
4290 * time the context is loaded. No call from
4291 * another CPU will be allowed.
4293 * The pinning via shed_setaffinity()
4294 * must be done by the calling task prior
4297 * systemwide: keep track of CPU this session is supposed to run on
4299 the_cpu = ctx->ctx_cpu = smp_processor_id();
4303 * now reserve the session
4305 ret = pfm_reserve_session(current, is_system, the_cpu);
4306 if (ret) goto error;
4309 * task is necessarily stopped at this point.
4311 * If the previous context was zombie, then it got removed in
4312 * pfm_save_regs(). Therefore we should not see it here.
4313 * If we see a context, then this is an active context
4315 * XXX: needs to be atomic
4317 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4318 thread->pfm_context, ctx));
4320 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4322 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4326 pfm_reset_msgq(ctx);
4328 ctx->ctx_state = PFM_CTX_LOADED;
4331 * link context to task
4333 ctx->ctx_task = task;
4337 * we load as stopped
4339 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4340 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4342 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4344 thread->flags |= IA64_THREAD_PM_VALID;
4348 * propagate into thread-state
4350 pfm_copy_pmds(task, ctx);
4351 pfm_copy_pmcs(task, ctx);
4353 pmcs_source = thread->pmcs;
4354 pmds_source = thread->pmds;
4357 * always the case for system-wide
4359 if (task == current) {
4361 if (is_system == 0) {
4363 /* allow user level control */
4364 ia64_psr(regs)->sp = 0;
4365 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4367 SET_LAST_CPU(ctx, smp_processor_id());
4369 SET_ACTIVATION(ctx);
4372 * push the other task out, if any
4374 owner_task = GET_PMU_OWNER();
4375 if (owner_task) pfm_lazy_save_regs(owner_task);
4379 * load all PMD from ctx to PMU (as opposed to thread state)
4380 * restore all PMC from ctx to PMU
4382 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4383 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4385 ctx->ctx_reload_pmcs[0] = 0UL;
4386 ctx->ctx_reload_pmds[0] = 0UL;
4389 * guaranteed safe by earlier check against DBG_VALID
4391 if (ctx->ctx_fl_using_dbreg) {
4392 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4393 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4398 SET_PMU_OWNER(task, ctx);
4400 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4403 * when not current, task MUST be stopped, so this is safe
4405 regs = ia64_task_regs(task);
4407 /* force a full reload */
4408 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4409 SET_LAST_CPU(ctx, -1);
4411 /* initial saved psr (stopped) */
4412 ctx->ctx_saved_psr_up = 0UL;
4413 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4419 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4422 * we must undo the dbregs setting (for system-wide)
4424 if (ret && set_dbregs) {
4426 pfm_sessions.pfs_sys_use_dbregs--;
4430 * release task, there is now a link with the context
4432 if (is_system == 0 && task != current) {
4436 ret = pfm_check_task_exist(ctx);
4438 ctx->ctx_state = PFM_CTX_UNLOADED;
4439 ctx->ctx_task = NULL;
4447 * in this function, we do not need to increase the use count
4448 * for the task via get_task_struct(), because we hold the
4449 * context lock. If the task were to disappear while having
4450 * a context attached, it would go through pfm_exit_thread()
4451 * which also grabs the context lock and would therefore be blocked
4452 * until we are here.
4454 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4457 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4459 struct task_struct *task = PFM_CTX_TASK(ctx);
4460 struct pt_regs *tregs;
4461 int prev_state, is_system;
4464 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4466 prev_state = ctx->ctx_state;
4467 is_system = ctx->ctx_fl_system;
4470 * unload only when necessary
4472 if (prev_state == PFM_CTX_UNLOADED) {
4473 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4478 * clear psr and dcr bits
4480 ret = pfm_stop(ctx, NULL, 0, regs);
4481 if (ret) return ret;
4483 ctx->ctx_state = PFM_CTX_UNLOADED;
4486 * in system mode, we need to update the PMU directly
4487 * and the user level state of the caller, which may not
4488 * necessarily be the creator of the context.
4495 * local PMU is taken care of in pfm_stop()
4497 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4498 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4501 * save PMDs in context
4504 pfm_flush_pmds(current, ctx);
4507 * at this point we are done with the PMU
4508 * so we can unreserve the resource.
4510 if (prev_state != PFM_CTX_ZOMBIE)
4511 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4514 * disconnect context from task
4516 task->thread.pfm_context = NULL;
4518 * disconnect task from context
4520 ctx->ctx_task = NULL;
4523 * There is nothing more to cleanup here.
4531 tregs = task == current ? regs : ia64_task_regs(task);
4533 if (task == current) {
4535 * cancel user level control
4537 ia64_psr(regs)->sp = 1;
4539 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4542 * save PMDs to context
4545 pfm_flush_pmds(task, ctx);
4548 * at this point we are done with the PMU
4549 * so we can unreserve the resource.
4551 * when state was ZOMBIE, we have already unreserved.
4553 if (prev_state != PFM_CTX_ZOMBIE)
4554 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4557 * reset activation counter and psr
4559 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4560 SET_LAST_CPU(ctx, -1);
4563 * PMU state will not be restored
4565 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4568 * break links between context and task
4570 task->thread.pfm_context = NULL;
4571 ctx->ctx_task = NULL;
4573 PFM_SET_WORK_PENDING(task, 0);
4575 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4576 ctx->ctx_fl_can_restart = 0;
4577 ctx->ctx_fl_going_zombie = 0;
4579 DPRINT(("disconnected [%d] from context\n", task->pid));
4586 * called only from exit_thread(): task == current
4587 * we come here only if current has a context attached (loaded or masked)
4590 pfm_exit_thread(struct task_struct *task)
4593 unsigned long flags;
4594 struct pt_regs *regs = ia64_task_regs(task);
4598 ctx = PFM_GET_CTX(task);
4600 PROTECT_CTX(ctx, flags);
4602 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4604 state = ctx->ctx_state;
4606 case PFM_CTX_UNLOADED:
4608 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4609 * be in unloaded state
4611 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4613 case PFM_CTX_LOADED:
4614 case PFM_CTX_MASKED:
4615 ret = pfm_context_unload(ctx, NULL, 0, regs);
4617 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4619 DPRINT(("ctx unloaded for current state was %d\n", state));
4621 pfm_end_notify_user(ctx);
4623 case PFM_CTX_ZOMBIE:
4624 ret = pfm_context_unload(ctx, NULL, 0, regs);
4626 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4631 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4634 UNPROTECT_CTX(ctx, flags);
4636 { u64 psr = pfm_get_psr();
4637 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4638 BUG_ON(GET_PMU_OWNER());
4639 BUG_ON(ia64_psr(regs)->up);
4640 BUG_ON(ia64_psr(regs)->pp);
4644 * All memory free operations (especially for vmalloc'ed memory)
4645 * MUST be done with interrupts ENABLED.
4647 if (free_ok) pfm_context_free(ctx);
4651 * functions MUST be listed in the increasing order of their index (see permfon.h)
4653 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4654 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4655 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4656 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4657 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4659 static pfm_cmd_desc_t pfm_cmd_tab[]={
4660 /* 0 */PFM_CMD_NONE,
4661 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4662 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4663 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4664 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4665 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4666 /* 6 */PFM_CMD_NONE,
4667 /* 7 */PFM_CMD_NONE,
4668 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4669 /* 9 */PFM_CMD_NONE,
4670 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4671 /* 11 */PFM_CMD_NONE,
4672 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4673 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4674 /* 14 */PFM_CMD_NONE,
4675 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4676 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4677 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4678 /* 18 */PFM_CMD_NONE,
4679 /* 19 */PFM_CMD_NONE,
4680 /* 20 */PFM_CMD_NONE,
4681 /* 21 */PFM_CMD_NONE,
4682 /* 22 */PFM_CMD_NONE,
4683 /* 23 */PFM_CMD_NONE,
4684 /* 24 */PFM_CMD_NONE,
4685 /* 25 */PFM_CMD_NONE,
4686 /* 26 */PFM_CMD_NONE,
4687 /* 27 */PFM_CMD_NONE,
4688 /* 28 */PFM_CMD_NONE,
4689 /* 29 */PFM_CMD_NONE,
4690 /* 30 */PFM_CMD_NONE,
4691 /* 31 */PFM_CMD_NONE,
4692 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4693 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4695 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4698 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4700 struct task_struct *task;
4701 int state, old_state;
4704 state = ctx->ctx_state;
4705 task = ctx->ctx_task;
4708 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4712 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4716 task->state, PFM_CMD_STOPPED(cmd)));
4719 * self-monitoring always ok.
4721 * for system-wide the caller can either be the creator of the
4722 * context (to one to which the context is attached to) OR
4723 * a task running on the same CPU as the session.
4725 if (task == current || ctx->ctx_fl_system) return 0;
4728 * we are monitoring another thread
4731 case PFM_CTX_UNLOADED:
4733 * if context is UNLOADED we are safe to go
4736 case PFM_CTX_ZOMBIE:
4738 * no command can operate on a zombie context
4740 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4742 case PFM_CTX_MASKED:
4744 * PMU state has been saved to software even though
4745 * the thread may still be running.
4747 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4751 * context is LOADED or MASKED. Some commands may need to have
4754 * We could lift this restriction for UP but it would mean that
4755 * the user has no guarantee the task would not run between
4756 * two successive calls to perfmonctl(). That's probably OK.
4757 * If this user wants to ensure the task does not run, then
4758 * the task must be stopped.
4760 if (PFM_CMD_STOPPED(cmd)) {
4761 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
4762 DPRINT(("[%d] task not in stopped state\n", task->pid));
4766 * task is now stopped, wait for ctxsw out
4768 * This is an interesting point in the code.
4769 * We need to unprotect the context because
4770 * the pfm_save_regs() routines needs to grab
4771 * the same lock. There are danger in doing
4772 * this because it leaves a window open for
4773 * another task to get access to the context
4774 * and possibly change its state. The one thing
4775 * that is not possible is for the context to disappear
4776 * because we are protected by the VFS layer, i.e.,
4777 * get_fd()/put_fd().
4781 UNPROTECT_CTX(ctx, flags);
4783 wait_task_inactive(task);
4785 PROTECT_CTX(ctx, flags);
4788 * we must recheck to verify if state has changed
4790 if (ctx->ctx_state != old_state) {
4791 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4799 * system-call entry point (must return long)
4802 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4804 struct file *file = NULL;
4805 pfm_context_t *ctx = NULL;
4806 unsigned long flags = 0UL;
4807 void *args_k = NULL;
4808 long ret; /* will expand int return types */
4809 size_t base_sz, sz, xtra_sz = 0;
4810 int narg, completed_args = 0, call_made = 0, cmd_flags;
4811 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4812 int (*getsize)(void *arg, size_t *sz);
4813 #define PFM_MAX_ARGSIZE 4096
4816 * reject any call if perfmon was disabled at initialization
4818 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4820 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4821 DPRINT(("invalid cmd=%d\n", cmd));
4825 func = pfm_cmd_tab[cmd].cmd_func;
4826 narg = pfm_cmd_tab[cmd].cmd_narg;
4827 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4828 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4829 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4831 if (unlikely(func == NULL)) {
4832 DPRINT(("invalid cmd=%d\n", cmd));
4836 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4844 * check if number of arguments matches what the command expects
4846 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4850 sz = xtra_sz + base_sz*count;
4852 * limit abuse to min page size
4854 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4855 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4860 * allocate default-sized argument buffer
4862 if (likely(count && args_k == NULL)) {
4863 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4864 if (args_k == NULL) return -ENOMEM;
4872 * assume sz = 0 for command without parameters
4874 if (sz && copy_from_user(args_k, arg, sz)) {
4875 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4880 * check if command supports extra parameters
4882 if (completed_args == 0 && getsize) {
4884 * get extra parameters size (based on main argument)
4886 ret = (*getsize)(args_k, &xtra_sz);
4887 if (ret) goto error_args;
4891 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4893 /* retry if necessary */
4894 if (likely(xtra_sz)) goto restart_args;
4897 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4902 if (unlikely(file == NULL)) {
4903 DPRINT(("invalid fd %d\n", fd));
4906 if (unlikely(PFM_IS_FILE(file) == 0)) {
4907 DPRINT(("fd %d not related to perfmon\n", fd));
4911 ctx = (pfm_context_t *)file->private_data;
4912 if (unlikely(ctx == NULL)) {
4913 DPRINT(("no context for fd %d\n", fd));
4916 prefetch(&ctx->ctx_state);
4918 PROTECT_CTX(ctx, flags);
4921 * check task is stopped
4923 ret = pfm_check_task_state(ctx, cmd, flags);
4924 if (unlikely(ret)) goto abort_locked;
4927 ret = (*func)(ctx, args_k, count, ia64_task_regs(current));
4933 DPRINT(("context unlocked\n"));
4934 UNPROTECT_CTX(ctx, flags);
4938 /* copy argument back to user, if needed */
4939 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4942 if (args_k) kfree(args_k);
4944 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4950 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4952 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4953 pfm_ovfl_ctrl_t rst_ctrl;
4957 state = ctx->ctx_state;
4959 * Unlock sampling buffer and reset index atomically
4960 * XXX: not really needed when blocking
4962 if (CTX_HAS_SMPL(ctx)) {
4964 rst_ctrl.bits.mask_monitoring = 0;
4965 rst_ctrl.bits.reset_ovfl_pmds = 0;
4967 if (state == PFM_CTX_LOADED)
4968 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4970 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4972 rst_ctrl.bits.mask_monitoring = 0;
4973 rst_ctrl.bits.reset_ovfl_pmds = 1;
4977 if (rst_ctrl.bits.reset_ovfl_pmds) {
4978 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4980 if (rst_ctrl.bits.mask_monitoring == 0) {
4981 DPRINT(("resuming monitoring\n"));
4982 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4984 DPRINT(("stopping monitoring\n"));
4985 //pfm_stop_monitoring(current, regs);
4987 ctx->ctx_state = PFM_CTX_LOADED;
4992 * context MUST BE LOCKED when calling
4993 * can only be called for current
4996 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
5000 DPRINT(("entering for [%d]\n", current->pid));
5002 ret = pfm_context_unload(ctx, NULL, 0, regs);
5004 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", current->pid, ret);
5008 * and wakeup controlling task, indicating we are now disconnected
5010 wake_up_interruptible(&ctx->ctx_zombieq);
5013 * given that context is still locked, the controlling
5014 * task will only get access when we return from
5015 * pfm_handle_work().
5019 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5021 * pfm_handle_work() can be called with interrupts enabled
5022 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5023 * call may sleep, therefore we must re-enable interrupts
5024 * to avoid deadlocks. It is safe to do so because this function
5025 * is called ONLY when returning to user level (PUStk=1), in which case
5026 * there is no risk of kernel stack overflow due to deep
5027 * interrupt nesting.
5030 pfm_handle_work(void)
5033 struct pt_regs *regs;
5034 unsigned long flags, dummy_flags;
5035 unsigned long ovfl_regs;
5036 unsigned int reason;
5039 ctx = PFM_GET_CTX(current);
5041 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5045 PROTECT_CTX(ctx, flags);
5047 PFM_SET_WORK_PENDING(current, 0);
5049 pfm_clear_task_notify();
5051 regs = ia64_task_regs(current);
5054 * extract reason for being here and clear
5056 reason = ctx->ctx_fl_trap_reason;
5057 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5058 ovfl_regs = ctx->ctx_ovfl_regs[0];
5060 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5063 * must be done before we check for simple-reset mode
5065 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5068 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5069 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5072 * restore interrupt mask to what it was on entry.
5073 * Could be enabled/diasbled.
5075 UNPROTECT_CTX(ctx, flags);
5078 * force interrupt enable because of down_interruptible()
5082 DPRINT(("before block sleeping\n"));
5085 * may go through without blocking on SMP systems
5086 * if restart has been received already by the time we call down()
5088 ret = down_interruptible(&ctx->ctx_restart_sem);
5090 DPRINT(("after block sleeping ret=%d\n", ret));
5093 * lock context and mask interrupts again
5094 * We save flags into a dummy because we may have
5095 * altered interrupts mask compared to entry in this
5098 PROTECT_CTX(ctx, dummy_flags);
5101 * we need to read the ovfl_regs only after wake-up
5102 * because we may have had pfm_write_pmds() in between
5103 * and that can changed PMD values and therefore
5104 * ovfl_regs is reset for these new PMD values.
5106 ovfl_regs = ctx->ctx_ovfl_regs[0];
5108 if (ctx->ctx_fl_going_zombie) {
5110 DPRINT(("context is zombie, bailing out\n"));
5111 pfm_context_force_terminate(ctx, regs);
5115 * in case of interruption of down() we don't restart anything
5117 if (ret < 0) goto nothing_to_do;
5120 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5121 ctx->ctx_ovfl_regs[0] = 0UL;
5125 * restore flags as they were upon entry
5127 UNPROTECT_CTX(ctx, flags);
5131 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5133 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5134 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5138 DPRINT(("waking up somebody\n"));
5140 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5143 * safe, we are not in intr handler, nor in ctxsw when
5146 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5152 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5154 pfm_msg_t *msg = NULL;
5156 if (ctx->ctx_fl_no_msg == 0) {
5157 msg = pfm_get_new_msg(ctx);
5159 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5163 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5164 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5165 msg->pfm_ovfl_msg.msg_active_set = 0;
5166 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5167 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5168 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5169 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5170 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5173 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5179 return pfm_notify_user(ctx, msg);
5183 pfm_end_notify_user(pfm_context_t *ctx)
5187 msg = pfm_get_new_msg(ctx);
5189 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5193 memset(msg, 0, sizeof(*msg));
5195 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5196 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5197 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5199 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5204 return pfm_notify_user(ctx, msg);
5208 * main overflow processing routine.
5209 * it can be called from the interrupt path or explicitely during the context switch code
5212 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5214 pfm_ovfl_arg_t *ovfl_arg;
5216 unsigned long old_val, ovfl_val, new_val;
5217 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5218 unsigned long tstamp;
5219 pfm_ovfl_ctrl_t ovfl_ctrl;
5220 unsigned int i, has_smpl;
5221 int must_notify = 0;
5223 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5226 * sanity test. Should never happen
5228 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5230 tstamp = ia64_get_itc();
5231 mask = pmc0 >> PMU_FIRST_COUNTER;
5232 ovfl_val = pmu_conf->ovfl_val;
5233 has_smpl = CTX_HAS_SMPL(ctx);
5235 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5236 "used_pmds=0x%lx\n",
5238 task ? task->pid: -1,
5239 (regs ? regs->cr_iip : 0),
5240 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5241 ctx->ctx_used_pmds[0]));
5245 * first we update the virtual counters
5246 * assume there was a prior ia64_srlz_d() issued
5248 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5250 /* skip pmd which did not overflow */
5251 if ((mask & 0x1) == 0) continue;
5254 * Note that the pmd is not necessarily 0 at this point as qualified events
5255 * may have happened before the PMU was frozen. The residual count is not
5256 * taken into consideration here but will be with any read of the pmd via
5259 old_val = new_val = ctx->ctx_pmds[i].val;
5260 new_val += 1 + ovfl_val;
5261 ctx->ctx_pmds[i].val = new_val;
5264 * check for overflow condition
5266 if (likely(old_val > new_val)) {
5267 ovfl_pmds |= 1UL << i;
5268 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5271 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5275 ia64_get_pmd(i) & ovfl_val,
5281 * there was no 64-bit overflow, nothing else to do
5283 if (ovfl_pmds == 0UL) return;
5286 * reset all control bits
5292 * if a sampling format module exists, then we "cache" the overflow by
5293 * calling the module's handler() routine.
5296 unsigned long start_cycles, end_cycles;
5297 unsigned long pmd_mask;
5299 int this_cpu = smp_processor_id();
5301 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5302 ovfl_arg = &ctx->ctx_ovfl_arg;
5304 prefetch(ctx->ctx_smpl_hdr);
5306 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5310 if ((pmd_mask & 0x1) == 0) continue;
5312 ovfl_arg->ovfl_pmd = (unsigned char )i;
5313 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5314 ovfl_arg->active_set = 0;
5315 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5316 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5318 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5319 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5320 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5323 * copy values of pmds of interest. Sampling format may copy them
5324 * into sampling buffer.
5327 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5328 if ((smpl_pmds & 0x1) == 0) continue;
5329 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5330 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5334 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5336 start_cycles = ia64_get_itc();
5339 * call custom buffer format record (handler) routine
5341 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5343 end_cycles = ia64_get_itc();
5346 * For those controls, we take the union because they have
5347 * an all or nothing behavior.
5349 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5350 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5351 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5353 * build the bitmask of pmds to reset now
5355 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5357 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5360 * when the module cannot handle the rest of the overflows, we abort right here
5362 if (ret && pmd_mask) {
5363 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5364 pmd_mask<<PMU_FIRST_COUNTER));
5367 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5369 ovfl_pmds &= ~reset_pmds;
5372 * when no sampling module is used, then the default
5373 * is to notify on overflow if requested by user
5375 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5376 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5377 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5378 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5380 * if needed, we reset all overflowed pmds
5382 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5385 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5388 * reset the requested PMD registers using the short reset values
5391 unsigned long bm = reset_pmds;
5392 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5395 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5397 * keep track of what to reset when unblocking
5399 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5402 * check for blocking context
5404 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5406 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5409 * set the perfmon specific checking pending work for the task
5411 PFM_SET_WORK_PENDING(task, 1);
5414 * when coming from ctxsw, current still points to the
5415 * previous task, therefore we must work with task and not current.
5417 pfm_set_task_notify(task);
5420 * defer until state is changed (shorten spin window). the context is locked
5421 * anyway, so the signal receiver would come spin for nothing.
5426 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5427 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5428 PFM_GET_WORK_PENDING(task),
5429 ctx->ctx_fl_trap_reason,
5432 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5434 * in case monitoring must be stopped, we toggle the psr bits
5436 if (ovfl_ctrl.bits.mask_monitoring) {
5437 pfm_mask_monitoring(task);
5438 ctx->ctx_state = PFM_CTX_MASKED;
5439 ctx->ctx_fl_can_restart = 1;
5443 * send notification now
5445 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5450 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5452 task ? task->pid : -1,
5458 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5459 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5460 * come here as zombie only if the task is the current task. In which case, we
5461 * can access the PMU hardware directly.
5463 * Note that zombies do have PM_VALID set. So here we do the minimal.
5465 * In case the context was zombified it could not be reclaimed at the time
5466 * the monitoring program exited. At this point, the PMU reservation has been
5467 * returned, the sampiing buffer has been freed. We must convert this call
5468 * into a spurious interrupt. However, we must also avoid infinite overflows
5469 * by stopping monitoring for this task. We can only come here for a per-task
5470 * context. All we need to do is to stop monitoring using the psr bits which
5471 * are always task private. By re-enabling secure montioring, we ensure that
5472 * the monitored task will not be able to re-activate monitoring.
5473 * The task will eventually be context switched out, at which point the context
5474 * will be reclaimed (that includes releasing ownership of the PMU).
5476 * So there might be a window of time where the number of per-task session is zero
5477 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5478 * context. This is safe because if a per-task session comes in, it will push this one
5479 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5480 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5481 * also push our zombie context out.
5483 * Overall pretty hairy stuff....
5485 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5487 ia64_psr(regs)->up = 0;
5488 ia64_psr(regs)->sp = 1;
5493 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5495 struct task_struct *task;
5497 unsigned long flags;
5499 int this_cpu = smp_processor_id();
5502 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5505 * srlz.d done before arriving here
5507 pmc0 = ia64_get_pmc(0);
5509 task = GET_PMU_OWNER();
5510 ctx = GET_PMU_CTX();
5513 * if we have some pending bits set
5514 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5516 if (PMC0_HAS_OVFL(pmc0) && task) {
5518 * we assume that pmc0.fr is always set here
5522 if (!ctx) goto report_spurious1;
5524 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5525 goto report_spurious2;
5527 PROTECT_CTX_NOPRINT(ctx, flags);
5529 pfm_overflow_handler(task, ctx, pmc0, regs);
5531 UNPROTECT_CTX_NOPRINT(ctx, flags);
5534 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5538 * keep it unfrozen at all times
5545 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5546 this_cpu, task->pid);
5550 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5558 pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5560 unsigned long start_cycles, total_cycles;
5561 unsigned long min, max;
5565 this_cpu = get_cpu();
5566 if (likely(!pfm_alt_intr_handler)) {
5567 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5568 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5570 start_cycles = ia64_get_itc();
5572 ret = pfm_do_interrupt_handler(irq, arg, regs);
5574 total_cycles = ia64_get_itc();
5577 * don't measure spurious interrupts
5579 if (likely(ret == 0)) {
5580 total_cycles -= start_cycles;
5582 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5583 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5585 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5589 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5592 put_cpu_no_resched();
5597 * /proc/perfmon interface, for debug only
5600 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5603 pfm_proc_start(struct seq_file *m, loff_t *pos)
5606 return PFM_PROC_SHOW_HEADER;
5609 while (*pos <= NR_CPUS) {
5610 if (cpu_online(*pos - 1)) {
5611 return (void *)*pos;
5619 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5622 return pfm_proc_start(m, pos);
5626 pfm_proc_stop(struct seq_file *m, void *v)
5631 pfm_proc_show_header(struct seq_file *m)
5633 struct list_head * pos;
5634 pfm_buffer_fmt_t * entry;
5635 unsigned long flags;
5638 "perfmon version : %u.%u\n"
5641 "expert mode : %s\n"
5642 "ovfl_mask : 0x%lx\n"
5643 "PMU flags : 0x%x\n",
5644 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5646 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5647 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5654 "proc_sessions : %u\n"
5655 "sys_sessions : %u\n"
5656 "sys_use_dbregs : %u\n"
5657 "ptrace_use_dbregs : %u\n",
5658 pfm_sessions.pfs_task_sessions,
5659 pfm_sessions.pfs_sys_sessions,
5660 pfm_sessions.pfs_sys_use_dbregs,
5661 pfm_sessions.pfs_ptrace_use_dbregs);
5665 spin_lock(&pfm_buffer_fmt_lock);
5667 list_for_each(pos, &pfm_buffer_fmt_list) {
5668 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5669 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5680 entry->fmt_uuid[10],
5681 entry->fmt_uuid[11],
5682 entry->fmt_uuid[12],
5683 entry->fmt_uuid[13],
5684 entry->fmt_uuid[14],
5685 entry->fmt_uuid[15],
5688 spin_unlock(&pfm_buffer_fmt_lock);
5693 pfm_proc_show(struct seq_file *m, void *v)
5699 if (v == PFM_PROC_SHOW_HEADER) {
5700 pfm_proc_show_header(m);
5704 /* show info for CPU (v - 1) */
5708 "CPU%-2d overflow intrs : %lu\n"
5709 "CPU%-2d overflow cycles : %lu\n"
5710 "CPU%-2d overflow min : %lu\n"
5711 "CPU%-2d overflow max : %lu\n"
5712 "CPU%-2d smpl handler calls : %lu\n"
5713 "CPU%-2d smpl handler cycles : %lu\n"
5714 "CPU%-2d spurious intrs : %lu\n"
5715 "CPU%-2d replay intrs : %lu\n"
5716 "CPU%-2d syst_wide : %d\n"
5717 "CPU%-2d dcr_pp : %d\n"
5718 "CPU%-2d exclude idle : %d\n"
5719 "CPU%-2d owner : %d\n"
5720 "CPU%-2d context : %p\n"
5721 "CPU%-2d activations : %lu\n",
5722 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5723 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5724 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5725 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5726 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5727 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5728 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5729 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5730 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5731 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5732 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5733 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5734 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5735 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5737 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5739 psr = pfm_get_psr();
5744 "CPU%-2d psr : 0x%lx\n"
5745 "CPU%-2d pmc0 : 0x%lx\n",
5747 cpu, ia64_get_pmc(0));
5749 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5750 if (PMC_IS_COUNTING(i) == 0) continue;
5752 "CPU%-2d pmc%u : 0x%lx\n"
5753 "CPU%-2d pmd%u : 0x%lx\n",
5754 cpu, i, ia64_get_pmc(i),
5755 cpu, i, ia64_get_pmd(i));
5761 struct seq_operations pfm_seq_ops = {
5762 .start = pfm_proc_start,
5763 .next = pfm_proc_next,
5764 .stop = pfm_proc_stop,
5765 .show = pfm_proc_show
5769 pfm_proc_open(struct inode *inode, struct file *file)
5771 return seq_open(file, &pfm_seq_ops);
5776 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5777 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5778 * is active or inactive based on mode. We must rely on the value in
5779 * local_cpu_data->pfm_syst_info
5782 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5784 struct pt_regs *regs;
5786 unsigned long dcr_pp;
5788 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5791 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5792 * on every CPU, so we can rely on the pid to identify the idle task.
5794 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5795 regs = ia64_task_regs(task);
5796 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5800 * if monitoring has started
5803 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5805 * context switching in?
5808 /* mask monitoring for the idle task */
5809 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5815 * context switching out
5816 * restore monitoring for next task
5818 * Due to inlining this odd if-then-else construction generates
5821 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5830 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5832 struct task_struct *task = ctx->ctx_task;
5834 ia64_psr(regs)->up = 0;
5835 ia64_psr(regs)->sp = 1;
5837 if (GET_PMU_OWNER() == task) {
5838 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5839 SET_PMU_OWNER(NULL, NULL);
5843 * disconnect the task from the context and vice-versa
5845 PFM_SET_WORK_PENDING(task, 0);
5847 task->thread.pfm_context = NULL;
5848 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5850 DPRINT(("force cleanup for [%d]\n", task->pid));
5855 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5858 pfm_save_regs(struct task_struct *task)
5861 struct thread_struct *t;
5862 unsigned long flags;
5866 ctx = PFM_GET_CTX(task);
5867 if (ctx == NULL) return;
5871 * we always come here with interrupts ALREADY disabled by
5872 * the scheduler. So we simply need to protect against concurrent
5873 * access, not CPU concurrency.
5875 flags = pfm_protect_ctx_ctxsw(ctx);
5877 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5878 struct pt_regs *regs = ia64_task_regs(task);
5882 pfm_force_cleanup(ctx, regs);
5884 BUG_ON(ctx->ctx_smpl_hdr);
5886 pfm_unprotect_ctx_ctxsw(ctx, flags);
5888 pfm_context_free(ctx);
5893 * save current PSR: needed because we modify it
5896 psr = pfm_get_psr();
5898 BUG_ON(psr & (IA64_PSR_I));
5902 * This is the last instruction which may generate an overflow
5904 * We do not need to set psr.sp because, it is irrelevant in kernel.
5905 * It will be restored from ipsr when going back to user level
5910 * keep a copy of psr.up (for reload)
5912 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5915 * release ownership of this PMU.
5916 * PM interrupts are masked, so nothing
5919 SET_PMU_OWNER(NULL, NULL);
5922 * we systematically save the PMD as we have no
5923 * guarantee we will be schedule at that same
5926 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5929 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5930 * we will need it on the restore path to check
5931 * for pending overflow.
5933 t->pmcs[0] = ia64_get_pmc(0);
5936 * unfreeze PMU if had pending overflows
5938 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5941 * finally, allow context access.
5942 * interrupts will still be masked after this call.
5944 pfm_unprotect_ctx_ctxsw(ctx, flags);
5947 #else /* !CONFIG_SMP */
5949 pfm_save_regs(struct task_struct *task)
5954 ctx = PFM_GET_CTX(task);
5955 if (ctx == NULL) return;
5958 * save current PSR: needed because we modify it
5960 psr = pfm_get_psr();
5962 BUG_ON(psr & (IA64_PSR_I));
5966 * This is the last instruction which may generate an overflow
5968 * We do not need to set psr.sp because, it is irrelevant in kernel.
5969 * It will be restored from ipsr when going back to user level
5974 * keep a copy of psr.up (for reload)
5976 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5980 pfm_lazy_save_regs (struct task_struct *task)
5983 struct thread_struct *t;
5984 unsigned long flags;
5986 { u64 psr = pfm_get_psr();
5987 BUG_ON(psr & IA64_PSR_UP);
5990 ctx = PFM_GET_CTX(task);
5994 * we need to mask PMU overflow here to
5995 * make sure that we maintain pmc0 until
5996 * we save it. overflow interrupts are
5997 * treated as spurious if there is no
6000 * XXX: I don't think this is necessary
6002 PROTECT_CTX(ctx,flags);
6005 * release ownership of this PMU.
6006 * must be done before we save the registers.
6008 * after this call any PMU interrupt is treated
6011 SET_PMU_OWNER(NULL, NULL);
6014 * save all the pmds we use
6016 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
6019 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6020 * it is needed to check for pended overflow
6021 * on the restore path
6023 t->pmcs[0] = ia64_get_pmc(0);
6026 * unfreeze PMU if had pending overflows
6028 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6031 * now get can unmask PMU interrupts, they will
6032 * be treated as purely spurious and we will not
6033 * lose any information
6035 UNPROTECT_CTX(ctx,flags);
6037 #endif /* CONFIG_SMP */
6041 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6044 pfm_load_regs (struct task_struct *task)
6047 struct thread_struct *t;
6048 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6049 unsigned long flags;
6051 int need_irq_resend;
6053 ctx = PFM_GET_CTX(task);
6054 if (unlikely(ctx == NULL)) return;
6056 BUG_ON(GET_PMU_OWNER());
6060 * possible on unload
6062 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6065 * we always come here with interrupts ALREADY disabled by
6066 * the scheduler. So we simply need to protect against concurrent
6067 * access, not CPU concurrency.
6069 flags = pfm_protect_ctx_ctxsw(ctx);
6070 psr = pfm_get_psr();
6072 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6074 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6075 BUG_ON(psr & IA64_PSR_I);
6077 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6078 struct pt_regs *regs = ia64_task_regs(task);
6080 BUG_ON(ctx->ctx_smpl_hdr);
6082 pfm_force_cleanup(ctx, regs);
6084 pfm_unprotect_ctx_ctxsw(ctx, flags);
6087 * this one (kmalloc'ed) is fine with interrupts disabled
6089 pfm_context_free(ctx);
6095 * we restore ALL the debug registers to avoid picking up
6098 if (ctx->ctx_fl_using_dbreg) {
6099 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6100 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6103 * retrieve saved psr.up
6105 psr_up = ctx->ctx_saved_psr_up;
6108 * if we were the last user of the PMU on that CPU,
6109 * then nothing to do except restore psr
6111 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6114 * retrieve partial reload masks (due to user modifications)
6116 pmc_mask = ctx->ctx_reload_pmcs[0];
6117 pmd_mask = ctx->ctx_reload_pmds[0];
6121 * To avoid leaking information to the user level when psr.sp=0,
6122 * we must reload ALL implemented pmds (even the ones we don't use).
6123 * In the kernel we only allow PFM_READ_PMDS on registers which
6124 * we initialized or requested (sampling) so there is no risk there.
6126 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6129 * ALL accessible PMCs are systematically reloaded, unused registers
6130 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6131 * up stale configuration.
6133 * PMC0 is never in the mask. It is always restored separately.
6135 pmc_mask = ctx->ctx_all_pmcs[0];
6138 * when context is MASKED, we will restore PMC with plm=0
6139 * and PMD with stale information, but that's ok, nothing
6142 * XXX: optimize here
6144 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6145 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6148 * check for pending overflow at the time the state
6151 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6153 * reload pmc0 with the overflow information
6154 * On McKinley PMU, this will trigger a PMU interrupt
6156 ia64_set_pmc(0, t->pmcs[0]);
6161 * will replay the PMU interrupt
6163 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6165 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6169 * we just did a reload, so we reset the partial reload fields
6171 ctx->ctx_reload_pmcs[0] = 0UL;
6172 ctx->ctx_reload_pmds[0] = 0UL;
6174 SET_LAST_CPU(ctx, smp_processor_id());
6177 * dump activation value for this PMU
6181 * record current activation for this context
6183 SET_ACTIVATION(ctx);
6186 * establish new ownership.
6188 SET_PMU_OWNER(task, ctx);
6191 * restore the psr.up bit. measurement
6193 * no PMU interrupt can happen at this point
6194 * because we still have interrupts disabled.
6196 if (likely(psr_up)) pfm_set_psr_up();
6199 * allow concurrent access to context
6201 pfm_unprotect_ctx_ctxsw(ctx, flags);
6203 #else /* !CONFIG_SMP */
6205 * reload PMU state for UP kernels
6206 * in 2.5 we come here with interrupts disabled
6209 pfm_load_regs (struct task_struct *task)
6211 struct thread_struct *t;
6213 struct task_struct *owner;
6214 unsigned long pmd_mask, pmc_mask;
6216 int need_irq_resend;
6218 owner = GET_PMU_OWNER();
6219 ctx = PFM_GET_CTX(task);
6221 psr = pfm_get_psr();
6223 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6224 BUG_ON(psr & IA64_PSR_I);
6227 * we restore ALL the debug registers to avoid picking up
6230 * This must be done even when the task is still the owner
6231 * as the registers may have been modified via ptrace()
6232 * (not perfmon) by the previous task.
6234 if (ctx->ctx_fl_using_dbreg) {
6235 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6236 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6240 * retrieved saved psr.up
6242 psr_up = ctx->ctx_saved_psr_up;
6243 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6246 * short path, our state is still there, just
6247 * need to restore psr and we go
6249 * we do not touch either PMC nor PMD. the psr is not touched
6250 * by the overflow_handler. So we are safe w.r.t. to interrupt
6251 * concurrency even without interrupt masking.
6253 if (likely(owner == task)) {
6254 if (likely(psr_up)) pfm_set_psr_up();
6259 * someone else is still using the PMU, first push it out and
6260 * then we'll be able to install our stuff !
6262 * Upon return, there will be no owner for the current PMU
6264 if (owner) pfm_lazy_save_regs(owner);
6267 * To avoid leaking information to the user level when psr.sp=0,
6268 * we must reload ALL implemented pmds (even the ones we don't use).
6269 * In the kernel we only allow PFM_READ_PMDS on registers which
6270 * we initialized or requested (sampling) so there is no risk there.
6272 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6275 * ALL accessible PMCs are systematically reloaded, unused registers
6276 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6277 * up stale configuration.
6279 * PMC0 is never in the mask. It is always restored separately
6281 pmc_mask = ctx->ctx_all_pmcs[0];
6283 pfm_restore_pmds(t->pmds, pmd_mask);
6284 pfm_restore_pmcs(t->pmcs, pmc_mask);
6287 * check for pending overflow at the time the state
6290 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6292 * reload pmc0 with the overflow information
6293 * On McKinley PMU, this will trigger a PMU interrupt
6295 ia64_set_pmc(0, t->pmcs[0]);
6301 * will replay the PMU interrupt
6303 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6305 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6309 * establish new ownership.
6311 SET_PMU_OWNER(task, ctx);
6314 * restore the psr.up bit. measurement
6316 * no PMU interrupt can happen at this point
6317 * because we still have interrupts disabled.
6319 if (likely(psr_up)) pfm_set_psr_up();
6321 #endif /* CONFIG_SMP */
6324 * this function assumes monitoring is stopped
6327 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6330 unsigned long mask2, val, pmd_val, ovfl_val;
6331 int i, can_access_pmu = 0;
6335 * is the caller the task being monitored (or which initiated the
6336 * session for system wide measurements)
6338 is_self = ctx->ctx_task == task ? 1 : 0;
6341 * can access PMU is task is the owner of the PMU state on the current CPU
6342 * or if we are running on the CPU bound to the context in system-wide mode
6343 * (that is not necessarily the task the context is attached to in this mode).
6344 * In system-wide we always have can_access_pmu true because a task running on an
6345 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6347 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6348 if (can_access_pmu) {
6350 * Mark the PMU as not owned
6351 * This will cause the interrupt handler to do nothing in case an overflow
6352 * interrupt was in-flight
6353 * This also guarantees that pmc0 will contain the final state
6354 * It virtually gives us full control on overflow processing from that point
6357 SET_PMU_OWNER(NULL, NULL);
6358 DPRINT(("releasing ownership\n"));
6361 * read current overflow status:
6363 * we are guaranteed to read the final stable state
6366 pmc0 = ia64_get_pmc(0); /* slow */
6369 * reset freeze bit, overflow status information destroyed
6373 pmc0 = task->thread.pmcs[0];
6375 * clear whatever overflow status bits there were
6377 task->thread.pmcs[0] = 0;
6379 ovfl_val = pmu_conf->ovfl_val;
6381 * we save all the used pmds
6382 * we take care of overflows for counting PMDs
6384 * XXX: sampling situation is not taken into account here
6386 mask2 = ctx->ctx_used_pmds[0];
6388 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6390 for (i = 0; mask2; i++, mask2>>=1) {
6392 /* skip non used pmds */
6393 if ((mask2 & 0x1) == 0) continue;
6396 * can access PMU always true in system wide mode
6398 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6400 if (PMD_IS_COUNTING(i)) {
6401 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6404 ctx->ctx_pmds[i].val,
6408 * we rebuild the full 64 bit value of the counter
6410 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6413 * now everything is in ctx_pmds[] and we need
6414 * to clear the saved context from save_regs() such that
6415 * pfm_read_pmds() gets the correct value
6420 * take care of overflow inline
6422 if (pmc0 & (1UL << i)) {
6423 val += 1 + ovfl_val;
6424 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6428 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, i, val, pmd_val));
6430 if (is_self) task->thread.pmds[i] = pmd_val;
6432 ctx->ctx_pmds[i].val = val;
6436 static struct irqaction perfmon_irqaction = {
6437 .handler = pfm_interrupt_handler,
6438 .flags = SA_INTERRUPT,
6443 pfm_alt_save_pmu_state(void *data)
6445 struct pt_regs *regs;
6447 regs = ia64_task_regs(current);
6449 DPRINT(("called\n"));
6452 * should not be necessary but
6453 * let's take not risk
6457 ia64_psr(regs)->pp = 0;
6460 * This call is required
6461 * May cause a spurious interrupt on some processors
6469 pfm_alt_restore_pmu_state(void *data)
6471 struct pt_regs *regs;
6473 regs = ia64_task_regs(current);
6475 DPRINT(("called\n"));
6478 * put PMU back in state expected
6483 ia64_psr(regs)->pp = 0;
6486 * perfmon runs with PMU unfrozen at all times
6494 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6499 /* some sanity checks */
6500 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6502 /* do the easy test first */
6503 if (pfm_alt_intr_handler) return -EBUSY;
6505 /* one at a time in the install or remove, just fail the others */
6506 if (!spin_trylock(&pfm_alt_install_check)) {
6510 /* reserve our session */
6511 for_each_online_cpu(reserve_cpu) {
6512 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6513 if (ret) goto cleanup_reserve;
6516 /* save the current system wide pmu states */
6517 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 0, 1);
6519 DPRINT(("on_each_cpu() failed: %d\n", ret));
6520 goto cleanup_reserve;
6523 /* officially change to the alternate interrupt handler */
6524 pfm_alt_intr_handler = hdl;
6526 spin_unlock(&pfm_alt_install_check);
6531 for_each_online_cpu(i) {
6532 /* don't unreserve more than we reserved */
6533 if (i >= reserve_cpu) break;
6535 pfm_unreserve_session(NULL, 1, i);
6538 spin_unlock(&pfm_alt_install_check);
6542 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6545 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6550 if (hdl == NULL) return -EINVAL;
6552 /* cannot remove someone else's handler! */
6553 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6555 /* one at a time in the install or remove, just fail the others */
6556 if (!spin_trylock(&pfm_alt_install_check)) {
6560 pfm_alt_intr_handler = NULL;
6562 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 0, 1);
6564 DPRINT(("on_each_cpu() failed: %d\n", ret));
6567 for_each_online_cpu(i) {
6568 pfm_unreserve_session(NULL, 1, i);
6571 spin_unlock(&pfm_alt_install_check);
6575 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6578 * perfmon initialization routine, called from the initcall() table
6580 static int init_pfm_fs(void);
6588 family = local_cpu_data->family;
6593 if ((*p)->probe() == 0) goto found;
6594 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6605 static struct file_operations pfm_proc_fops = {
6606 .open = pfm_proc_open,
6608 .llseek = seq_lseek,
6609 .release = seq_release,
6615 unsigned int n, n_counters, i;
6617 printk("perfmon: version %u.%u IRQ %u\n",
6620 IA64_PERFMON_VECTOR);
6622 if (pfm_probe_pmu()) {
6623 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6624 local_cpu_data->family);
6629 * compute the number of implemented PMD/PMC from the
6630 * description tables
6633 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6634 if (PMC_IS_IMPL(i) == 0) continue;
6635 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6638 pmu_conf->num_pmcs = n;
6640 n = 0; n_counters = 0;
6641 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6642 if (PMD_IS_IMPL(i) == 0) continue;
6643 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6645 if (PMD_IS_COUNTING(i)) n_counters++;
6647 pmu_conf->num_pmds = n;
6648 pmu_conf->num_counters = n_counters;
6651 * sanity checks on the number of debug registers
6653 if (pmu_conf->use_rr_dbregs) {
6654 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6655 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6659 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6660 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6666 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6670 pmu_conf->num_counters,
6671 ffz(pmu_conf->ovfl_val));
6674 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6675 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6681 * create /proc/perfmon (mostly for debugging purposes)
6683 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6684 if (perfmon_dir == NULL) {
6685 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6690 * install customized file operations for /proc/perfmon entry
6692 perfmon_dir->proc_fops = &pfm_proc_fops;
6695 * create /proc/sys/kernel/perfmon (for debugging purposes)
6697 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6700 * initialize all our spinlocks
6702 spin_lock_init(&pfm_sessions.pfs_lock);
6703 spin_lock_init(&pfm_buffer_fmt_lock);
6707 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6712 __initcall(pfm_init);
6715 * this function is called before pfm_init()
6718 pfm_init_percpu (void)
6721 * make sure no measurement is active
6722 * (may inherit programmed PMCs from EFI).
6728 * we run with the PMU not frozen at all times
6732 if (smp_processor_id() == 0)
6733 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6735 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6740 * used for debug purposes only
6743 dump_pmu_state(const char *from)
6745 struct task_struct *task;
6746 struct thread_struct *t;
6747 struct pt_regs *regs;
6749 unsigned long psr, dcr, info, flags;
6752 local_irq_save(flags);
6754 this_cpu = smp_processor_id();
6755 regs = ia64_task_regs(current);
6756 info = PFM_CPUINFO_GET();
6757 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6759 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6760 local_irq_restore(flags);
6764 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6771 task = GET_PMU_OWNER();
6772 ctx = GET_PMU_CTX();
6774 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6776 psr = pfm_get_psr();
6778 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",
6781 psr & IA64_PSR_PP ? 1 : 0,
6782 psr & IA64_PSR_UP ? 1 : 0,
6783 dcr & IA64_DCR_PP ? 1 : 0,
6786 ia64_psr(regs)->pp);
6788 ia64_psr(regs)->up = 0;
6789 ia64_psr(regs)->pp = 0;
6791 t = ¤t->thread;
6793 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6794 if (PMC_IS_IMPL(i) == 0) continue;
6795 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6798 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6799 if (PMD_IS_IMPL(i) == 0) continue;
6800 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6804 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6807 ctx->ctx_smpl_vaddr,
6811 ctx->ctx_saved_psr_up);
6813 local_irq_restore(flags);
6817 * called from process.c:copy_thread(). task is new child.
6820 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6822 struct thread_struct *thread;
6824 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6826 thread = &task->thread;
6829 * cut links inherited from parent (current)
6831 thread->pfm_context = NULL;
6833 PFM_SET_WORK_PENDING(task, 0);
6836 * the psr bits are already set properly in copy_threads()
6839 #else /* !CONFIG_PERFMON */
6841 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6845 #endif /* CONFIG_PERFMON */