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) 1999-2003 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 int debug; /* turn on/off debugging via syslog */
488 int debug_ovfl; /* turn on/off debug printk in overflow handler */
489 int fastctxsw; /* turn on/off fast (unsecure) ctxsw */
490 int expert_mode; /* turn on/off value checking */
495 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
496 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
497 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
498 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
499 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
500 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
501 unsigned long pfm_smpl_handler_calls;
502 unsigned long pfm_smpl_handler_cycles;
503 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
507 * perfmon internal variables
509 static pfm_stats_t pfm_stats[NR_CPUS];
510 static pfm_session_t pfm_sessions; /* global sessions information */
512 static struct proc_dir_entry *perfmon_dir;
513 static pfm_uuid_t pfm_null_uuid = {0,};
515 static spinlock_t pfm_buffer_fmt_lock;
516 static LIST_HEAD(pfm_buffer_fmt_list);
518 static pmu_config_t *pmu_conf;
520 /* sysctl() controls */
521 static pfm_sysctl_t pfm_sysctl;
524 static ctl_table pfm_ctl_table[]={
525 {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
526 {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
527 {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
528 {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
531 static ctl_table pfm_sysctl_dir[] = {
532 {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
535 static ctl_table pfm_sysctl_root[] = {
536 {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
539 static struct ctl_table_header *pfm_sysctl_header;
541 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
542 static int pfm_flush(struct file *filp);
544 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
545 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
548 pfm_put_task(struct task_struct *task)
550 if (task != current) put_task_struct(task);
554 pfm_set_task_notify(struct task_struct *task)
556 struct thread_info *info;
558 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
559 set_bit(TIF_NOTIFY_RESUME, &info->flags);
563 pfm_clear_task_notify(void)
565 clear_thread_flag(TIF_NOTIFY_RESUME);
569 pfm_reserve_page(unsigned long a)
571 SetPageReserved(vmalloc_to_page((void *)a));
574 pfm_unreserve_page(unsigned long a)
576 ClearPageReserved(vmalloc_to_page((void*)a));
579 static inline unsigned long
580 pfm_protect_ctx_ctxsw(pfm_context_t *x)
582 spin_lock(&(x)->ctx_lock);
586 static inline unsigned long
587 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
589 spin_unlock(&(x)->ctx_lock);
592 static inline unsigned int
593 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
595 return do_munmap(mm, addr, len);
598 static inline unsigned long
599 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
601 return get_unmapped_area(file, addr, len, pgoff, flags);
605 static struct super_block *
606 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
608 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
611 static struct file_system_type pfm_fs_type = {
613 .get_sb = pfmfs_get_sb,
614 .kill_sb = kill_anon_super,
617 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
618 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
619 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
620 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
623 /* forward declaration */
624 static struct file_operations pfm_file_ops;
627 * forward declarations
630 static void pfm_lazy_save_regs (struct task_struct *ta);
633 void dump_pmu_state(const char *);
634 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
636 #include "perfmon_itanium.h"
637 #include "perfmon_mckinley.h"
638 #include "perfmon_generic.h"
640 static pmu_config_t *pmu_confs[]={
643 &pmu_conf_gen, /* must be last */
648 static int pfm_end_notify_user(pfm_context_t *ctx);
651 pfm_clear_psr_pp(void)
653 ia64_rsm(IA64_PSR_PP);
660 ia64_ssm(IA64_PSR_PP);
665 pfm_clear_psr_up(void)
667 ia64_rsm(IA64_PSR_UP);
674 ia64_ssm(IA64_PSR_UP);
678 static inline unsigned long
682 tmp = ia64_getreg(_IA64_REG_PSR);
688 pfm_set_psr_l(unsigned long val)
690 ia64_setreg(_IA64_REG_PSR_L, val);
702 pfm_unfreeze_pmu(void)
709 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
713 for (i=0; i < nibrs; i++) {
714 ia64_set_ibr(i, ibrs[i]);
715 ia64_dv_serialize_instruction();
721 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
725 for (i=0; i < ndbrs; i++) {
726 ia64_set_dbr(i, dbrs[i]);
727 ia64_dv_serialize_data();
733 * PMD[i] must be a counter. no check is made
735 static inline unsigned long
736 pfm_read_soft_counter(pfm_context_t *ctx, int i)
738 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
742 * PMD[i] must be a counter. no check is made
745 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
747 unsigned long ovfl_val = pmu_conf->ovfl_val;
749 ctx->ctx_pmds[i].val = val & ~ovfl_val;
751 * writing to unimplemented part is ignore, so we do not need to
754 ia64_set_pmd(i, val & ovfl_val);
758 pfm_get_new_msg(pfm_context_t *ctx)
762 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
764 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
765 if (next == ctx->ctx_msgq_head) return NULL;
767 idx = ctx->ctx_msgq_tail;
768 ctx->ctx_msgq_tail = next;
770 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
772 return ctx->ctx_msgq+idx;
776 pfm_get_next_msg(pfm_context_t *ctx)
780 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
782 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
787 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
792 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
794 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));
800 pfm_reset_msgq(pfm_context_t *ctx)
802 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
803 DPRINT(("ctx=%p msgq reset\n", ctx));
807 pfm_rvmalloc(unsigned long size)
812 size = PAGE_ALIGN(size);
815 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
816 memset(mem, 0, size);
817 addr = (unsigned long)mem;
819 pfm_reserve_page(addr);
828 pfm_rvfree(void *mem, unsigned long size)
833 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
834 addr = (unsigned long) mem;
835 while ((long) size > 0) {
836 pfm_unreserve_page(addr);
845 static pfm_context_t *
846 pfm_context_alloc(void)
851 * allocate context descriptor
852 * must be able to free with interrupts disabled
854 ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
856 memset(ctx, 0, sizeof(pfm_context_t));
857 DPRINT(("alloc ctx @%p\n", ctx));
863 pfm_context_free(pfm_context_t *ctx)
866 DPRINT(("free ctx @%p\n", ctx));
872 pfm_mask_monitoring(struct task_struct *task)
874 pfm_context_t *ctx = PFM_GET_CTX(task);
875 struct thread_struct *th = &task->thread;
876 unsigned long mask, val, ovfl_mask;
879 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
881 ovfl_mask = pmu_conf->ovfl_val;
883 * monitoring can only be masked as a result of a valid
884 * counter overflow. In UP, it means that the PMU still
885 * has an owner. Note that the owner can be different
886 * from the current task. However the PMU state belongs
888 * In SMP, a valid overflow only happens when task is
889 * current. Therefore if we come here, we know that
890 * the PMU state belongs to the current task, therefore
891 * we can access the live registers.
893 * So in both cases, the live register contains the owner's
894 * state. We can ONLY touch the PMU registers and NOT the PSR.
896 * As a consequence to this call, the thread->pmds[] array
897 * contains stale information which must be ignored
898 * when context is reloaded AND monitoring is active (see
901 mask = ctx->ctx_used_pmds[0];
902 for (i = 0; mask; i++, mask>>=1) {
903 /* skip non used pmds */
904 if ((mask & 0x1) == 0) continue;
905 val = ia64_get_pmd(i);
907 if (PMD_IS_COUNTING(i)) {
909 * we rebuild the full 64 bit value of the counter
911 ctx->ctx_pmds[i].val += (val & ovfl_mask);
913 ctx->ctx_pmds[i].val = val;
915 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
917 ctx->ctx_pmds[i].val,
921 * mask monitoring by setting the privilege level to 0
922 * we cannot use psr.pp/psr.up for this, it is controlled by
925 * if task is current, modify actual registers, otherwise modify
926 * thread save state, i.e., what will be restored in pfm_load_regs()
928 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
929 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
930 if ((mask & 0x1) == 0UL) continue;
931 ia64_set_pmc(i, th->pmcs[i] & ~0xfUL);
932 th->pmcs[i] &= ~0xfUL;
933 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i]));
936 * make all of this visible
942 * must always be done with task == current
944 * context must be in MASKED state when calling
947 pfm_restore_monitoring(struct task_struct *task)
949 pfm_context_t *ctx = PFM_GET_CTX(task);
950 struct thread_struct *th = &task->thread;
951 unsigned long mask, ovfl_mask;
952 unsigned long psr, val;
955 is_system = ctx->ctx_fl_system;
956 ovfl_mask = pmu_conf->ovfl_val;
958 if (task != current) {
959 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
962 if (ctx->ctx_state != PFM_CTX_MASKED) {
963 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
964 task->pid, current->pid, ctx->ctx_state);
969 * monitoring is masked via the PMC.
970 * As we restore their value, we do not want each counter to
971 * restart right away. We stop monitoring using the PSR,
972 * restore the PMC (and PMD) and then re-establish the psr
973 * as it was. Note that there can be no pending overflow at
974 * this point, because monitoring was MASKED.
976 * system-wide session are pinned and self-monitoring
978 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
980 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
986 * first, we restore the PMD
988 mask = ctx->ctx_used_pmds[0];
989 for (i = 0; mask; i++, mask>>=1) {
990 /* skip non used pmds */
991 if ((mask & 0x1) == 0) continue;
993 if (PMD_IS_COUNTING(i)) {
995 * we split the 64bit value according to
998 val = ctx->ctx_pmds[i].val & ovfl_mask;
999 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1001 val = ctx->ctx_pmds[i].val;
1003 ia64_set_pmd(i, val);
1005 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1007 ctx->ctx_pmds[i].val,
1013 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1014 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1015 if ((mask & 0x1) == 0UL) continue;
1016 th->pmcs[i] = ctx->ctx_pmcs[i];
1017 ia64_set_pmc(i, th->pmcs[i]);
1018 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
1023 * must restore DBR/IBR because could be modified while masked
1024 * XXX: need to optimize
1026 if (ctx->ctx_fl_using_dbreg) {
1027 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1028 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1034 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1036 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1043 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1049 for (i=0; mask; i++, mask>>=1) {
1050 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1055 * reload from thread state (used for ctxw only)
1058 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1061 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1063 for (i=0; mask; i++, mask>>=1) {
1064 if ((mask & 0x1) == 0) continue;
1065 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1066 ia64_set_pmd(i, val);
1072 * propagate PMD from context to thread-state
1075 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1077 struct thread_struct *thread = &task->thread;
1078 unsigned long ovfl_val = pmu_conf->ovfl_val;
1079 unsigned long mask = ctx->ctx_all_pmds[0];
1083 DPRINT(("mask=0x%lx\n", mask));
1085 for (i=0; mask; i++, mask>>=1) {
1087 val = ctx->ctx_pmds[i].val;
1090 * We break up the 64 bit value into 2 pieces
1091 * the lower bits go to the machine state in the
1092 * thread (will be reloaded on ctxsw in).
1093 * The upper part stays in the soft-counter.
1095 if (PMD_IS_COUNTING(i)) {
1096 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1099 thread->pmds[i] = val;
1101 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1104 ctx->ctx_pmds[i].val));
1109 * propagate PMC from context to thread-state
1112 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1114 struct thread_struct *thread = &task->thread;
1115 unsigned long mask = ctx->ctx_all_pmcs[0];
1118 DPRINT(("mask=0x%lx\n", mask));
1120 for (i=0; mask; i++, mask>>=1) {
1121 /* masking 0 with ovfl_val yields 0 */
1122 thread->pmcs[i] = ctx->ctx_pmcs[i];
1123 DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
1130 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1134 for (i=0; mask; i++, mask>>=1) {
1135 if ((mask & 0x1) == 0) continue;
1136 ia64_set_pmc(i, pmcs[i]);
1142 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1144 return memcmp(a, b, sizeof(pfm_uuid_t));
1148 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1151 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1156 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1159 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1165 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1169 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1174 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1178 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1183 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1186 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1191 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)
1194 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1198 static pfm_buffer_fmt_t *
1199 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1201 struct list_head * pos;
1202 pfm_buffer_fmt_t * entry;
1204 list_for_each(pos, &pfm_buffer_fmt_list) {
1205 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1206 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1213 * find a buffer format based on its uuid
1215 static pfm_buffer_fmt_t *
1216 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1218 pfm_buffer_fmt_t * fmt;
1219 spin_lock(&pfm_buffer_fmt_lock);
1220 fmt = __pfm_find_buffer_fmt(uuid);
1221 spin_unlock(&pfm_buffer_fmt_lock);
1226 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1230 /* some sanity checks */
1231 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1233 /* we need at least a handler */
1234 if (fmt->fmt_handler == NULL) return -EINVAL;
1237 * XXX: need check validity of fmt_arg_size
1240 spin_lock(&pfm_buffer_fmt_lock);
1242 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1243 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1247 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1248 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1251 spin_unlock(&pfm_buffer_fmt_lock);
1254 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1257 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1259 pfm_buffer_fmt_t *fmt;
1262 spin_lock(&pfm_buffer_fmt_lock);
1264 fmt = __pfm_find_buffer_fmt(uuid);
1266 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1270 list_del_init(&fmt->fmt_list);
1271 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1274 spin_unlock(&pfm_buffer_fmt_lock);
1278 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1281 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1283 unsigned long flags;
1285 * validy checks on cpu_mask have been done upstream
1289 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1290 pfm_sessions.pfs_sys_sessions,
1291 pfm_sessions.pfs_task_sessions,
1292 pfm_sessions.pfs_sys_use_dbregs,
1298 * cannot mix system wide and per-task sessions
1300 if (pfm_sessions.pfs_task_sessions > 0UL) {
1301 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1302 pfm_sessions.pfs_task_sessions));
1306 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1308 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1310 pfm_sessions.pfs_sys_session[cpu] = task;
1312 pfm_sessions.pfs_sys_sessions++ ;
1315 if (pfm_sessions.pfs_sys_sessions) goto abort;
1316 pfm_sessions.pfs_task_sessions++;
1319 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1320 pfm_sessions.pfs_sys_sessions,
1321 pfm_sessions.pfs_task_sessions,
1322 pfm_sessions.pfs_sys_use_dbregs,
1331 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1332 pfm_sessions.pfs_sys_session[cpu]->pid,
1333 smp_processor_id()));
1342 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1344 unsigned long flags;
1346 * validy checks on cpu_mask have been done upstream
1350 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1351 pfm_sessions.pfs_sys_sessions,
1352 pfm_sessions.pfs_task_sessions,
1353 pfm_sessions.pfs_sys_use_dbregs,
1359 pfm_sessions.pfs_sys_session[cpu] = NULL;
1361 * would not work with perfmon+more than one bit in cpu_mask
1363 if (ctx && ctx->ctx_fl_using_dbreg) {
1364 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1365 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1367 pfm_sessions.pfs_sys_use_dbregs--;
1370 pfm_sessions.pfs_sys_sessions--;
1372 pfm_sessions.pfs_task_sessions--;
1374 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1375 pfm_sessions.pfs_sys_sessions,
1376 pfm_sessions.pfs_task_sessions,
1377 pfm_sessions.pfs_sys_use_dbregs,
1387 * removes virtual mapping of the sampling buffer.
1388 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1389 * a PROTECT_CTX() section.
1392 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1397 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1398 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1402 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1405 * does the actual unmapping
1407 down_write(&task->mm->mmap_sem);
1409 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1411 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1413 up_write(&task->mm->mmap_sem);
1415 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1418 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1424 * free actual physical storage used by sampling buffer
1428 pfm_free_smpl_buffer(pfm_context_t *ctx)
1430 pfm_buffer_fmt_t *fmt;
1432 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1435 * we won't use the buffer format anymore
1437 fmt = ctx->ctx_buf_fmt;
1439 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1442 ctx->ctx_smpl_vaddr));
1444 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1449 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1451 ctx->ctx_smpl_hdr = NULL;
1452 ctx->ctx_smpl_size = 0UL;
1457 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1463 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1465 if (fmt == NULL) return;
1467 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1472 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1473 * no real gain from having the whole whorehouse mounted. So we don't need
1474 * any operations on the root directory. However, we need a non-trivial
1475 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1477 static struct vfsmount *pfmfs_mnt;
1482 int err = register_filesystem(&pfm_fs_type);
1484 pfmfs_mnt = kern_mount(&pfm_fs_type);
1485 err = PTR_ERR(pfmfs_mnt);
1486 if (IS_ERR(pfmfs_mnt))
1487 unregister_filesystem(&pfm_fs_type);
1497 unregister_filesystem(&pfm_fs_type);
1502 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1507 unsigned long flags;
1508 DECLARE_WAITQUEUE(wait, current);
1509 if (PFM_IS_FILE(filp) == 0) {
1510 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1514 ctx = (pfm_context_t *)filp->private_data;
1516 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1521 * check even when there is no message
1523 if (size < sizeof(pfm_msg_t)) {
1524 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1528 PROTECT_CTX(ctx, flags);
1531 * put ourselves on the wait queue
1533 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1541 set_current_state(TASK_INTERRUPTIBLE);
1543 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1546 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1548 UNPROTECT_CTX(ctx, flags);
1551 * check non-blocking read
1554 if(filp->f_flags & O_NONBLOCK) break;
1557 * check pending signals
1559 if(signal_pending(current)) {
1564 * no message, so wait
1568 PROTECT_CTX(ctx, flags);
1570 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1571 set_current_state(TASK_RUNNING);
1572 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1574 if (ret < 0) goto abort;
1577 msg = pfm_get_next_msg(ctx);
1579 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1583 DPRINT(("[%d] fd=%d type=%d\n", current->pid, msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1586 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1589 UNPROTECT_CTX(ctx, flags);
1595 pfm_write(struct file *file, const char __user *ubuf,
1596 size_t size, loff_t *ppos)
1598 DPRINT(("pfm_write called\n"));
1603 pfm_poll(struct file *filp, poll_table * wait)
1606 unsigned long flags;
1607 unsigned int mask = 0;
1609 if (PFM_IS_FILE(filp) == 0) {
1610 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1614 ctx = (pfm_context_t *)filp->private_data;
1616 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1621 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1623 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1625 PROTECT_CTX(ctx, flags);
1627 if (PFM_CTXQ_EMPTY(ctx) == 0)
1628 mask = POLLIN | POLLRDNORM;
1630 UNPROTECT_CTX(ctx, flags);
1632 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1638 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1640 DPRINT(("pfm_ioctl called\n"));
1645 * interrupt cannot be masked when coming here
1648 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1652 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1654 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1658 ctx->ctx_async_queue, ret));
1664 pfm_fasync(int fd, struct file *filp, int on)
1669 if (PFM_IS_FILE(filp) == 0) {
1670 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1674 ctx = (pfm_context_t *)filp->private_data;
1676 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1680 * we cannot mask interrupts during this call because this may
1681 * may go to sleep if memory is not readily avalaible.
1683 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1684 * done in caller. Serialization of this function is ensured by caller.
1686 ret = pfm_do_fasync(fd, filp, ctx, on);
1689 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1692 ctx->ctx_async_queue, ret));
1699 * this function is exclusively called from pfm_close().
1700 * The context is not protected at that time, nor are interrupts
1701 * on the remote CPU. That's necessary to avoid deadlocks.
1704 pfm_syswide_force_stop(void *info)
1706 pfm_context_t *ctx = (pfm_context_t *)info;
1707 struct pt_regs *regs = ia64_task_regs(current);
1708 struct task_struct *owner;
1709 unsigned long flags;
1712 if (ctx->ctx_cpu != smp_processor_id()) {
1713 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1715 smp_processor_id());
1718 owner = GET_PMU_OWNER();
1719 if (owner != ctx->ctx_task) {
1720 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1722 owner->pid, ctx->ctx_task->pid);
1725 if (GET_PMU_CTX() != ctx) {
1726 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1728 GET_PMU_CTX(), ctx);
1732 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1734 * the context is already protected in pfm_close(), we simply
1735 * need to mask interrupts to avoid a PMU interrupt race on
1738 local_irq_save(flags);
1740 ret = pfm_context_unload(ctx, NULL, 0, regs);
1742 DPRINT(("context_unload returned %d\n", ret));
1746 * unmask interrupts, PMU interrupts are now spurious here
1748 local_irq_restore(flags);
1752 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1756 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1757 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1758 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1760 #endif /* CONFIG_SMP */
1763 * called for each close(). Partially free resources.
1764 * When caller is self-monitoring, the context is unloaded.
1767 pfm_flush(struct file *filp)
1770 struct task_struct *task;
1771 struct pt_regs *regs;
1772 unsigned long flags;
1773 unsigned long smpl_buf_size = 0UL;
1774 void *smpl_buf_vaddr = NULL;
1775 int state, is_system;
1777 if (PFM_IS_FILE(filp) == 0) {
1778 DPRINT(("bad magic for\n"));
1782 ctx = (pfm_context_t *)filp->private_data;
1784 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1789 * remove our file from the async queue, if we use this mode.
1790 * This can be done without the context being protected. We come
1791 * here when the context has become unreacheable by other tasks.
1793 * We may still have active monitoring at this point and we may
1794 * end up in pfm_overflow_handler(). However, fasync_helper()
1795 * operates with interrupts disabled and it cleans up the
1796 * queue. If the PMU handler is called prior to entering
1797 * fasync_helper() then it will send a signal. If it is
1798 * invoked after, it will find an empty queue and no
1799 * signal will be sent. In both case, we are safe
1801 if (filp->f_flags & FASYNC) {
1802 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1803 pfm_do_fasync (-1, filp, ctx, 0);
1806 PROTECT_CTX(ctx, flags);
1808 state = ctx->ctx_state;
1809 is_system = ctx->ctx_fl_system;
1811 task = PFM_CTX_TASK(ctx);
1812 regs = ia64_task_regs(task);
1814 DPRINT(("ctx_state=%d is_current=%d\n",
1816 task == current ? 1 : 0));
1819 * if state == UNLOADED, then task is NULL
1823 * we must stop and unload because we are losing access to the context.
1825 if (task == current) {
1828 * the task IS the owner but it migrated to another CPU: that's bad
1829 * but we must handle this cleanly. Unfortunately, the kernel does
1830 * not provide a mechanism to block migration (while the context is loaded).
1832 * We need to release the resource on the ORIGINAL cpu.
1834 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1836 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1838 * keep context protected but unmask interrupt for IPI
1840 local_irq_restore(flags);
1842 pfm_syswide_cleanup_other_cpu(ctx);
1845 * restore interrupt masking
1847 local_irq_save(flags);
1850 * context is unloaded at this point
1853 #endif /* CONFIG_SMP */
1856 DPRINT(("forcing unload\n"));
1858 * stop and unload, returning with state UNLOADED
1859 * and session unreserved.
1861 pfm_context_unload(ctx, NULL, 0, regs);
1863 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1868 * remove virtual mapping, if any, for the calling task.
1869 * cannot reset ctx field until last user is calling close().
1871 * ctx_smpl_vaddr must never be cleared because it is needed
1872 * by every task with access to the context
1874 * When called from do_exit(), the mm context is gone already, therefore
1875 * mm is NULL, i.e., the VMA is already gone and we do not have to
1878 if (ctx->ctx_smpl_vaddr && current->mm) {
1879 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1880 smpl_buf_size = ctx->ctx_smpl_size;
1883 UNPROTECT_CTX(ctx, flags);
1886 * if there was a mapping, then we systematically remove it
1887 * at this point. Cannot be done inside critical section
1888 * because some VM function reenables interrupts.
1891 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1896 * called either on explicit close() or from exit_files().
1897 * Only the LAST user of the file gets to this point, i.e., it is
1900 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1901 * (fput()),i.e, last task to access the file. Nobody else can access the
1902 * file at this point.
1904 * When called from exit_files(), the VMA has been freed because exit_mm()
1905 * is executed before exit_files().
1907 * When called from exit_files(), the current task is not yet ZOMBIE but we
1908 * flush the PMU state to the context.
1911 pfm_close(struct inode *inode, struct file *filp)
1914 struct task_struct *task;
1915 struct pt_regs *regs;
1916 DECLARE_WAITQUEUE(wait, current);
1917 unsigned long flags;
1918 unsigned long smpl_buf_size = 0UL;
1919 void *smpl_buf_addr = NULL;
1920 int free_possible = 1;
1921 int state, is_system;
1923 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1925 if (PFM_IS_FILE(filp) == 0) {
1926 DPRINT(("bad magic\n"));
1930 ctx = (pfm_context_t *)filp->private_data;
1932 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1936 PROTECT_CTX(ctx, flags);
1938 state = ctx->ctx_state;
1939 is_system = ctx->ctx_fl_system;
1941 task = PFM_CTX_TASK(ctx);
1942 regs = ia64_task_regs(task);
1944 DPRINT(("ctx_state=%d is_current=%d\n",
1946 task == current ? 1 : 0));
1949 * if task == current, then pfm_flush() unloaded the context
1951 if (state == PFM_CTX_UNLOADED) goto doit;
1954 * context is loaded/masked and task != current, we need to
1955 * either force an unload or go zombie
1959 * The task is currently blocked or will block after an overflow.
1960 * we must force it to wakeup to get out of the
1961 * MASKED state and transition to the unloaded state by itself.
1963 * This situation is only possible for per-task mode
1965 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1968 * set a "partial" zombie state to be checked
1969 * upon return from down() in pfm_handle_work().
1971 * We cannot use the ZOMBIE state, because it is checked
1972 * by pfm_load_regs() which is called upon wakeup from down().
1973 * In such case, it would free the context and then we would
1974 * return to pfm_handle_work() which would access the
1975 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1976 * but visible to pfm_handle_work().
1978 * For some window of time, we have a zombie context with
1979 * ctx_state = MASKED and not ZOMBIE
1981 ctx->ctx_fl_going_zombie = 1;
1984 * force task to wake up from MASKED state
1986 up(&ctx->ctx_restart_sem);
1988 DPRINT(("waking up ctx_state=%d\n", state));
1991 * put ourself to sleep waiting for the other
1992 * task to report completion
1994 * the context is protected by mutex, therefore there
1995 * is no risk of being notified of completion before
1996 * begin actually on the waitq.
1998 set_current_state(TASK_INTERRUPTIBLE);
1999 add_wait_queue(&ctx->ctx_zombieq, &wait);
2001 UNPROTECT_CTX(ctx, flags);
2004 * XXX: check for signals :
2005 * - ok for explicit close
2006 * - not ok when coming from exit_files()
2011 PROTECT_CTX(ctx, flags);
2014 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2015 set_current_state(TASK_RUNNING);
2018 * context is unloaded at this point
2020 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2022 else if (task != current) {
2025 * switch context to zombie state
2027 ctx->ctx_state = PFM_CTX_ZOMBIE;
2029 DPRINT(("zombie ctx for [%d]\n", task->pid));
2031 * cannot free the context on the spot. deferred until
2032 * the task notices the ZOMBIE state
2036 pfm_context_unload(ctx, NULL, 0, regs);
2041 /* reload state, may have changed during opening of critical section */
2042 state = ctx->ctx_state;
2045 * the context is still attached to a task (possibly current)
2046 * we cannot destroy it right now
2050 * we must free the sampling buffer right here because
2051 * we cannot rely on it being cleaned up later by the
2052 * monitored task. It is not possible to free vmalloc'ed
2053 * memory in pfm_load_regs(). Instead, we remove the buffer
2054 * now. should there be subsequent PMU overflow originally
2055 * meant for sampling, the will be converted to spurious
2056 * and that's fine because the monitoring tools is gone anyway.
2058 if (ctx->ctx_smpl_hdr) {
2059 smpl_buf_addr = ctx->ctx_smpl_hdr;
2060 smpl_buf_size = ctx->ctx_smpl_size;
2061 /* no more sampling */
2062 ctx->ctx_smpl_hdr = NULL;
2063 ctx->ctx_fl_is_sampling = 0;
2066 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2072 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2075 * UNLOADED that the session has already been unreserved.
2077 if (state == PFM_CTX_ZOMBIE) {
2078 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2082 * disconnect file descriptor from context must be done
2085 filp->private_data = NULL;
2088 * if we free on the spot, the context is now completely unreacheable
2089 * from the callers side. The monitored task side is also cut, so we
2092 * If we have a deferred free, only the caller side is disconnected.
2094 UNPROTECT_CTX(ctx, flags);
2097 * All memory free operations (especially for vmalloc'ed memory)
2098 * MUST be done with interrupts ENABLED.
2100 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2103 * return the memory used by the context
2105 if (free_possible) pfm_context_free(ctx);
2111 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2113 DPRINT(("pfm_no_open called\n"));
2119 static struct file_operations pfm_file_ops = {
2120 .llseek = no_llseek,
2125 .open = pfm_no_open, /* special open code to disallow open via /proc */
2126 .fasync = pfm_fasync,
2127 .release = pfm_close,
2132 pfmfs_delete_dentry(struct dentry *dentry)
2137 static struct dentry_operations pfmfs_dentry_operations = {
2138 .d_delete = pfmfs_delete_dentry,
2143 pfm_alloc_fd(struct file **cfile)
2146 struct file *file = NULL;
2147 struct inode * inode;
2151 fd = get_unused_fd();
2152 if (fd < 0) return -ENFILE;
2156 file = get_empty_filp();
2157 if (!file) goto out;
2160 * allocate a new inode
2162 inode = new_inode(pfmfs_mnt->mnt_sb);
2163 if (!inode) goto out;
2165 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2167 inode->i_sb = pfmfs_mnt->mnt_sb;
2168 inode->i_mode = S_IFCHR|S_IRUGO;
2170 inode->i_uid = current->fsuid;
2171 inode->i_gid = current->fsgid;
2173 sprintf(name, "[%lu]", inode->i_ino);
2175 this.len = strlen(name);
2176 this.hash = inode->i_ino;
2181 * allocate a new dcache entry
2183 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2184 if (!file->f_dentry) goto out;
2186 file->f_dentry->d_op = &pfmfs_dentry_operations;
2188 d_add(file->f_dentry, inode);
2189 file->f_vfsmnt = mntget(pfmfs_mnt);
2190 file->f_mapping = inode->i_mapping;
2192 file->f_op = &pfm_file_ops;
2193 file->f_mode = FMODE_READ;
2194 file->f_flags = O_RDONLY;
2198 * may have to delay until context is attached?
2200 fd_install(fd, file);
2203 * the file structure we will use
2209 if (file) put_filp(file);
2215 pfm_free_fd(int fd, struct file *file)
2217 struct files_struct *files = current->files;
2220 * there ie no fd_uninstall(), so we do it here
2222 spin_lock(&files->file_lock);
2223 files->fd[fd] = NULL;
2224 spin_unlock(&files->file_lock);
2226 if (file) put_filp(file);
2231 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2233 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2236 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2239 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2250 * allocate a sampling buffer and remaps it into the user address space of the task
2253 pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2255 struct mm_struct *mm = task->mm;
2256 struct vm_area_struct *vma = NULL;
2262 * the fixed header + requested size and align to page boundary
2264 size = PAGE_ALIGN(rsize);
2266 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2269 * check requested size to avoid Denial-of-service attacks
2270 * XXX: may have to refine this test
2271 * Check against address space limit.
2273 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2276 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2280 * We do the easy to undo allocations first.
2282 * pfm_rvmalloc(), clears the buffer, so there is no leak
2284 smpl_buf = pfm_rvmalloc(size);
2285 if (smpl_buf == NULL) {
2286 DPRINT(("Can't allocate sampling buffer\n"));
2290 DPRINT(("smpl_buf @%p\n", smpl_buf));
2293 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2295 DPRINT(("Cannot allocate vma\n"));
2298 memset(vma, 0, sizeof(*vma));
2301 * partially initialize the vma for the sampling buffer
2304 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2305 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2308 * Now we have everything we need and we can initialize
2309 * and connect all the data structures
2312 ctx->ctx_smpl_hdr = smpl_buf;
2313 ctx->ctx_smpl_size = size; /* aligned size */
2316 * Let's do the difficult operations next.
2318 * now we atomically find some area in the address space and
2319 * remap the buffer in it.
2321 down_write(&task->mm->mmap_sem);
2323 /* find some free area in address space, must have mmap sem held */
2324 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2325 if (vma->vm_start == 0UL) {
2326 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2327 up_write(&task->mm->mmap_sem);
2330 vma->vm_end = vma->vm_start + size;
2331 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2333 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2335 /* can only be applied to current task, need to have the mm semaphore held when called */
2336 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2337 DPRINT(("Can't remap buffer\n"));
2338 up_write(&task->mm->mmap_sem);
2343 * now insert the vma in the vm list for the process, must be
2344 * done with mmap lock held
2346 insert_vm_struct(mm, vma);
2348 // mm->total_vm += size >> PAGE_SHIFT;
2349 vx_vmpages_add(mm, size >> PAGE_SHIFT);
2350 vm_stat_account(vma);
2351 up_write(&task->mm->mmap_sem);
2354 * keep track of user level virtual address
2356 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2357 *(unsigned long *)user_vaddr = vma->vm_start;
2362 kmem_cache_free(vm_area_cachep, vma);
2364 pfm_rvfree(smpl_buf, size);
2370 * XXX: do something better here
2373 pfm_bad_permissions(struct task_struct *task)
2375 /* inspired by ptrace_attach() */
2376 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2385 return ((current->uid != task->euid)
2386 || (current->uid != task->suid)
2387 || (current->uid != task->uid)
2388 || (current->gid != task->egid)
2389 || (current->gid != task->sgid)
2390 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2394 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2400 ctx_flags = pfx->ctx_flags;
2402 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2405 * cannot block in this mode
2407 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2408 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2413 /* probably more to add here */
2419 pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2420 unsigned int cpu, pfarg_context_t *arg)
2422 pfm_buffer_fmt_t *fmt = NULL;
2423 unsigned long size = 0UL;
2425 void *fmt_arg = NULL;
2427 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2429 /* invoke and lock buffer format, if found */
2430 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2432 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2437 * buffer argument MUST be contiguous to pfarg_context_t
2439 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2441 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2443 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2445 if (ret) goto error;
2447 /* link buffer format and context */
2448 ctx->ctx_buf_fmt = fmt;
2451 * check if buffer format wants to use perfmon buffer allocation/mapping service
2453 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2454 if (ret) goto error;
2458 * buffer is always remapped into the caller's address space
2460 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2461 if (ret) goto error;
2463 /* keep track of user address of buffer */
2464 arg->ctx_smpl_vaddr = uaddr;
2466 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2473 pfm_reset_pmu_state(pfm_context_t *ctx)
2478 * install reset values for PMC.
2480 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2481 if (PMC_IS_IMPL(i) == 0) continue;
2482 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2483 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2486 * PMD registers are set to 0UL when the context in memset()
2490 * On context switched restore, we must restore ALL pmc and ALL pmd even
2491 * when they are not actively used by the task. In UP, the incoming process
2492 * may otherwise pick up left over PMC, PMD state from the previous process.
2493 * As opposed to PMD, stale PMC can cause harm to the incoming
2494 * process because they may change what is being measured.
2495 * Therefore, we must systematically reinstall the entire
2496 * PMC state. In SMP, the same thing is possible on the
2497 * same CPU but also on between 2 CPUs.
2499 * The problem with PMD is information leaking especially
2500 * to user level when psr.sp=0
2502 * There is unfortunately no easy way to avoid this problem
2503 * on either UP or SMP. This definitively slows down the
2504 * pfm_load_regs() function.
2508 * bitmask of all PMCs accessible to this context
2510 * PMC0 is treated differently.
2512 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2515 * bitmask of all PMDs that are accesible to this context
2517 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2519 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2522 * useful in case of re-enable after disable
2524 ctx->ctx_used_ibrs[0] = 0UL;
2525 ctx->ctx_used_dbrs[0] = 0UL;
2529 pfm_ctx_getsize(void *arg, size_t *sz)
2531 pfarg_context_t *req = (pfarg_context_t *)arg;
2532 pfm_buffer_fmt_t *fmt;
2536 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2538 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2540 DPRINT(("cannot find buffer format\n"));
2543 /* get just enough to copy in user parameters */
2544 *sz = fmt->fmt_arg_size;
2545 DPRINT(("arg_size=%lu\n", *sz));
2553 * cannot attach if :
2555 * - task not owned by caller
2556 * - task incompatible with context mode
2559 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2562 * no kernel task or task not owner by caller
2564 if (task->mm == NULL) {
2565 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2568 if (pfm_bad_permissions(task)) {
2569 DPRINT(("no permission to attach to [%d]\n", task->pid));
2573 * cannot block in self-monitoring mode
2575 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2576 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2580 if (task->exit_state == EXIT_ZOMBIE) {
2581 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2586 * always ok for self
2588 if (task == current) return 0;
2590 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
2591 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2595 * make sure the task is off any CPU
2597 wait_task_inactive(task);
2599 /* more to come... */
2605 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2607 struct task_struct *p = current;
2610 /* XXX: need to add more checks here */
2611 if (pid < 2) return -EPERM;
2613 if (pid != current->pid) {
2615 read_lock(&tasklist_lock);
2617 p = find_task_by_pid(pid);
2619 /* make sure task cannot go away while we operate on it */
2620 if (p) get_task_struct(p);
2622 read_unlock(&tasklist_lock);
2624 if (p == NULL) return -ESRCH;
2627 ret = pfm_task_incompatible(ctx, p);
2630 } else if (p != current) {
2639 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2641 pfarg_context_t *req = (pfarg_context_t *)arg;
2646 /* let's check the arguments first */
2647 ret = pfarg_is_sane(current, req);
2648 if (ret < 0) return ret;
2650 ctx_flags = req->ctx_flags;
2654 ctx = pfm_context_alloc();
2655 if (!ctx) goto error;
2657 ret = pfm_alloc_fd(&filp);
2658 if (ret < 0) goto error_file;
2660 req->ctx_fd = ctx->ctx_fd = ret;
2663 * attach context to file
2665 filp->private_data = ctx;
2668 * does the user want to sample?
2670 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2671 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2672 if (ret) goto buffer_error;
2676 * init context protection lock
2678 spin_lock_init(&ctx->ctx_lock);
2681 * context is unloaded
2683 ctx->ctx_state = PFM_CTX_UNLOADED;
2686 * initialization of context's flags
2688 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2689 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2690 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2691 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2693 * will move to set properties
2694 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2698 * init restart semaphore to locked
2700 sema_init(&ctx->ctx_restart_sem, 0);
2703 * activation is used in SMP only
2705 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2706 SET_LAST_CPU(ctx, -1);
2709 * initialize notification message queue
2711 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2712 init_waitqueue_head(&ctx->ctx_msgq_wait);
2713 init_waitqueue_head(&ctx->ctx_zombieq);
2715 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2720 ctx->ctx_fl_excl_idle,
2725 * initialize soft PMU state
2727 pfm_reset_pmu_state(ctx);
2732 pfm_free_fd(ctx->ctx_fd, filp);
2734 if (ctx->ctx_buf_fmt) {
2735 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2738 pfm_context_free(ctx);
2744 static inline unsigned long
2745 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2747 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2748 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2749 extern unsigned long carta_random32 (unsigned long seed);
2751 if (reg->flags & PFM_REGFL_RANDOM) {
2752 new_seed = carta_random32(old_seed);
2753 val -= (old_seed & mask); /* counter values are negative numbers! */
2754 if ((mask >> 32) != 0)
2755 /* construct a full 64-bit random value: */
2756 new_seed |= carta_random32(old_seed >> 32) << 32;
2757 reg->seed = new_seed;
2764 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2766 unsigned long mask = ovfl_regs[0];
2767 unsigned long reset_others = 0UL;
2772 * now restore reset value on sampling overflowed counters
2774 mask >>= PMU_FIRST_COUNTER;
2775 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2777 if ((mask & 0x1UL) == 0UL) continue;
2779 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2780 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2782 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2786 * Now take care of resetting the other registers
2788 for(i = 0; reset_others; i++, reset_others >>= 1) {
2790 if ((reset_others & 0x1) == 0) continue;
2792 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2794 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2795 is_long_reset ? "long" : "short", i, val));
2800 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2802 unsigned long mask = ovfl_regs[0];
2803 unsigned long reset_others = 0UL;
2807 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2809 if (ctx->ctx_state == PFM_CTX_MASKED) {
2810 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2815 * now restore reset value on sampling overflowed counters
2817 mask >>= PMU_FIRST_COUNTER;
2818 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2820 if ((mask & 0x1UL) == 0UL) continue;
2822 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2823 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2825 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2827 pfm_write_soft_counter(ctx, i, val);
2831 * Now take care of resetting the other registers
2833 for(i = 0; reset_others; i++, reset_others >>= 1) {
2835 if ((reset_others & 0x1) == 0) continue;
2837 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2839 if (PMD_IS_COUNTING(i)) {
2840 pfm_write_soft_counter(ctx, i, val);
2842 ia64_set_pmd(i, val);
2844 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2845 is_long_reset ? "long" : "short", i, val));
2851 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2853 struct thread_struct *thread = NULL;
2854 struct task_struct *task;
2855 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2856 unsigned long value, pmc_pm;
2857 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2858 unsigned int cnum, reg_flags, flags, pmc_type;
2859 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2860 int is_monitor, is_counting, state;
2862 pfm_reg_check_t wr_func;
2863 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2865 state = ctx->ctx_state;
2866 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2867 is_system = ctx->ctx_fl_system;
2868 task = ctx->ctx_task;
2869 impl_pmds = pmu_conf->impl_pmds[0];
2871 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2874 thread = &task->thread;
2876 * In system wide and when the context is loaded, access can only happen
2877 * when the caller is running on the CPU being monitored by the session.
2878 * It does not have to be the owner (ctx_task) of the context per se.
2880 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2881 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2884 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2886 expert_mode = pfm_sysctl.expert_mode;
2888 for (i = 0; i < count; i++, req++) {
2890 cnum = req->reg_num;
2891 reg_flags = req->reg_flags;
2892 value = req->reg_value;
2893 smpl_pmds = req->reg_smpl_pmds[0];
2894 reset_pmds = req->reg_reset_pmds[0];
2898 if (cnum >= PMU_MAX_PMCS) {
2899 DPRINT(("pmc%u is invalid\n", cnum));
2903 pmc_type = pmu_conf->pmc_desc[cnum].type;
2904 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2905 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2906 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2909 * we reject all non implemented PMC as well
2910 * as attempts to modify PMC[0-3] which are used
2911 * as status registers by the PMU
2913 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2914 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2917 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2919 * If the PMC is a monitor, then if the value is not the default:
2920 * - system-wide session: PMCx.pm=1 (privileged monitor)
2921 * - per-task : PMCx.pm=0 (user monitor)
2923 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2924 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2933 * enforce generation of overflow interrupt. Necessary on all
2936 value |= 1 << PMU_PMC_OI;
2938 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2939 flags |= PFM_REGFL_OVFL_NOTIFY;
2942 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2944 /* verify validity of smpl_pmds */
2945 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2946 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2950 /* verify validity of reset_pmds */
2951 if ((reset_pmds & impl_pmds) != reset_pmds) {
2952 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2956 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2957 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2960 /* eventid on non-counting monitors are ignored */
2964 * execute write checker, if any
2966 if (likely(expert_mode == 0 && wr_func)) {
2967 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2968 if (ret) goto error;
2973 * no error on this register
2975 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2978 * Now we commit the changes to the software state
2982 * update overflow information
2986 * full flag update each time a register is programmed
2988 ctx->ctx_pmds[cnum].flags = flags;
2990 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2991 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2992 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2995 * Mark all PMDS to be accessed as used.
2997 * We do not keep track of PMC because we have to
2998 * systematically restore ALL of them.
3000 * We do not update the used_monitors mask, because
3001 * if we have not programmed them, then will be in
3002 * a quiescent state, therefore we will not need to
3003 * mask/restore then when context is MASKED.
3005 CTX_USED_PMD(ctx, reset_pmds);
3006 CTX_USED_PMD(ctx, smpl_pmds);
3008 * make sure we do not try to reset on
3009 * restart because we have established new values
3011 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3014 * Needed in case the user does not initialize the equivalent
3015 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3016 * possible leak here.
3018 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3021 * keep track of the monitor PMC that we are using.
3022 * we save the value of the pmc in ctx_pmcs[] and if
3023 * the monitoring is not stopped for the context we also
3024 * place it in the saved state area so that it will be
3025 * picked up later by the context switch code.
3027 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3029 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3030 * monitoring needs to be stopped.
3032 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3035 * update context state
3037 ctx->ctx_pmcs[cnum] = value;
3041 * write thread state
3043 if (is_system == 0) thread->pmcs[cnum] = value;
3046 * write hardware register if we can
3048 if (can_access_pmu) {
3049 ia64_set_pmc(cnum, value);
3054 * per-task SMP only here
3056 * we are guaranteed that the task is not running on the other CPU,
3057 * we indicate that this PMD will need to be reloaded if the task
3058 * is rescheduled on the CPU it ran last on.
3060 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3065 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",
3071 ctx->ctx_all_pmcs[0],
3072 ctx->ctx_used_pmds[0],
3073 ctx->ctx_pmds[cnum].eventid,
3076 ctx->ctx_reload_pmcs[0],
3077 ctx->ctx_used_monitors[0],
3078 ctx->ctx_ovfl_regs[0]));
3082 * make sure the changes are visible
3084 if (can_access_pmu) ia64_srlz_d();
3088 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3093 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3095 struct thread_struct *thread = NULL;
3096 struct task_struct *task;
3097 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3098 unsigned long value, hw_value, ovfl_mask;
3100 int i, can_access_pmu = 0, state;
3101 int is_counting, is_loaded, is_system, expert_mode;
3103 pfm_reg_check_t wr_func;
3106 state = ctx->ctx_state;
3107 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3108 is_system = ctx->ctx_fl_system;
3109 ovfl_mask = pmu_conf->ovfl_val;
3110 task = ctx->ctx_task;
3112 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3115 * on both UP and SMP, we can only write to the PMC when the task is
3116 * the owner of the local PMU.
3118 if (likely(is_loaded)) {
3119 thread = &task->thread;
3121 * In system wide and when the context is loaded, access can only happen
3122 * when the caller is running on the CPU being monitored by the session.
3123 * It does not have to be the owner (ctx_task) of the context per se.
3125 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3126 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3129 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3131 expert_mode = pfm_sysctl.expert_mode;
3133 for (i = 0; i < count; i++, req++) {
3135 cnum = req->reg_num;
3136 value = req->reg_value;
3138 if (!PMD_IS_IMPL(cnum)) {
3139 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3142 is_counting = PMD_IS_COUNTING(cnum);
3143 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3146 * execute write checker, if any
3148 if (unlikely(expert_mode == 0 && wr_func)) {
3149 unsigned long v = value;
3151 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3152 if (ret) goto abort_mission;
3159 * no error on this register
3161 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3164 * now commit changes to software state
3169 * update virtualized (64bits) counter
3173 * write context state
3175 ctx->ctx_pmds[cnum].lval = value;
3178 * when context is load we use the split value
3181 hw_value = value & ovfl_mask;
3182 value = value & ~ovfl_mask;
3186 * update reset values (not just for counters)
3188 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3189 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3192 * update randomization parameters (not just for counters)
3194 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3195 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3198 * update context value
3200 ctx->ctx_pmds[cnum].val = value;
3203 * Keep track of what we use
3205 * We do not keep track of PMC because we have to
3206 * systematically restore ALL of them.
3208 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3211 * mark this PMD register used as well
3213 CTX_USED_PMD(ctx, RDEP(cnum));
3216 * make sure we do not try to reset on
3217 * restart because we have established new values
3219 if (is_counting && state == PFM_CTX_MASKED) {
3220 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3225 * write thread state
3227 if (is_system == 0) thread->pmds[cnum] = hw_value;
3230 * write hardware register if we can
3232 if (can_access_pmu) {
3233 ia64_set_pmd(cnum, hw_value);
3237 * we are guaranteed that the task is not running on the other CPU,
3238 * we indicate that this PMD will need to be reloaded if the task
3239 * is rescheduled on the CPU it ran last on.
3241 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3246 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3247 "long_reset=0x%lx notify=%c seed=0x%lx mask=0x%lx used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3253 ctx->ctx_pmds[cnum].val,
3254 ctx->ctx_pmds[cnum].short_reset,
3255 ctx->ctx_pmds[cnum].long_reset,
3256 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3257 ctx->ctx_pmds[cnum].seed,
3258 ctx->ctx_pmds[cnum].mask,
3259 ctx->ctx_used_pmds[0],
3260 ctx->ctx_pmds[cnum].reset_pmds[0],
3261 ctx->ctx_reload_pmds[0],
3262 ctx->ctx_all_pmds[0],
3263 ctx->ctx_ovfl_regs[0]));
3267 * make changes visible
3269 if (can_access_pmu) ia64_srlz_d();
3275 * for now, we have only one possibility for error
3277 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3282 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3283 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3284 * interrupt is delivered during the call, it will be kept pending until we leave, making
3285 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3286 * guaranteed to return consistent data to the user, it may simply be old. It is not
3287 * trivial to treat the overflow while inside the call because you may end up in
3288 * some module sampling buffer code causing deadlocks.
3291 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3293 struct thread_struct *thread = NULL;
3294 struct task_struct *task;
3295 unsigned long val = 0UL, lval, ovfl_mask, sval;
3296 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3297 unsigned int cnum, reg_flags = 0;
3298 int i, can_access_pmu = 0, state;
3299 int is_loaded, is_system, is_counting, expert_mode;
3301 pfm_reg_check_t rd_func;
3304 * access is possible when loaded only for
3305 * self-monitoring tasks or in UP mode
3308 state = ctx->ctx_state;
3309 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3310 is_system = ctx->ctx_fl_system;
3311 ovfl_mask = pmu_conf->ovfl_val;
3312 task = ctx->ctx_task;
3314 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3316 if (likely(is_loaded)) {
3317 thread = &task->thread;
3319 * In system wide and when the context is loaded, access can only happen
3320 * when the caller is running on the CPU being monitored by the session.
3321 * It does not have to be the owner (ctx_task) of the context per se.
3323 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3324 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3328 * this can be true when not self-monitoring only in UP
3330 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3332 if (can_access_pmu) ia64_srlz_d();
3334 expert_mode = pfm_sysctl.expert_mode;
3336 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3342 * on both UP and SMP, we can only read the PMD from the hardware register when
3343 * the task is the owner of the local PMU.
3346 for (i = 0; i < count; i++, req++) {
3348 cnum = req->reg_num;
3349 reg_flags = req->reg_flags;
3351 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3353 * we can only read the register that we use. That includes
3354 * the one we explicitely initialize AND the one we want included
3355 * in the sampling buffer (smpl_regs).
3357 * Having this restriction allows optimization in the ctxsw routine
3358 * without compromising security (leaks)
3360 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3362 sval = ctx->ctx_pmds[cnum].val;
3363 lval = ctx->ctx_pmds[cnum].lval;
3364 is_counting = PMD_IS_COUNTING(cnum);
3367 * If the task is not the current one, then we check if the
3368 * PMU state is still in the local live register due to lazy ctxsw.
3369 * If true, then we read directly from the registers.
3371 if (can_access_pmu){
3372 val = ia64_get_pmd(cnum);
3375 * context has been saved
3376 * if context is zombie, then task does not exist anymore.
3377 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3379 val = is_loaded ? thread->pmds[cnum] : 0UL;
3381 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3385 * XXX: need to check for overflow when loaded
3392 * execute read checker, if any
3394 if (unlikely(expert_mode == 0 && rd_func)) {
3395 unsigned long v = val;
3396 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3397 if (ret) goto error;
3402 PFM_REG_RETFLAG_SET(reg_flags, 0);
3404 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3407 * update register return value, abort all if problem during copy.
3408 * we only modify the reg_flags field. no check mode is fine because
3409 * access has been verified upfront in sys_perfmonctl().
3411 req->reg_value = val;
3412 req->reg_flags = reg_flags;
3413 req->reg_last_reset_val = lval;
3419 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3424 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3428 if (req == NULL) return -EINVAL;
3430 ctx = GET_PMU_CTX();
3432 if (ctx == NULL) return -EINVAL;
3435 * for now limit to current task, which is enough when calling
3436 * from overflow handler
3438 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3440 return pfm_write_pmcs(ctx, req, nreq, regs);
3442 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3445 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3449 if (req == NULL) return -EINVAL;
3451 ctx = GET_PMU_CTX();
3453 if (ctx == NULL) return -EINVAL;
3456 * for now limit to current task, which is enough when calling
3457 * from overflow handler
3459 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3461 return pfm_read_pmds(ctx, req, nreq, regs);
3463 EXPORT_SYMBOL(pfm_mod_read_pmds);
3466 * Only call this function when a process it trying to
3467 * write the debug registers (reading is always allowed)
3470 pfm_use_debug_registers(struct task_struct *task)
3472 pfm_context_t *ctx = task->thread.pfm_context;
3473 unsigned long flags;
3476 if (pmu_conf->use_rr_dbregs == 0) return 0;
3478 DPRINT(("called for [%d]\n", task->pid));
3483 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3486 * Even on SMP, we do not need to use an atomic here because
3487 * the only way in is via ptrace() and this is possible only when the
3488 * process is stopped. Even in the case where the ctxsw out is not totally
3489 * completed by the time we come here, there is no way the 'stopped' process
3490 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3491 * So this is always safe.
3493 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3498 * We cannot allow setting breakpoints when system wide monitoring
3499 * sessions are using the debug registers.
3501 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3504 pfm_sessions.pfs_ptrace_use_dbregs++;
3506 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3507 pfm_sessions.pfs_ptrace_use_dbregs,
3508 pfm_sessions.pfs_sys_use_dbregs,
3517 * This function is called for every task that exits with the
3518 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3519 * able to use the debug registers for debugging purposes via
3520 * ptrace(). Therefore we know it was not using them for
3521 * perfmormance monitoring, so we only decrement the number
3522 * of "ptraced" debug register users to keep the count up to date
3525 pfm_release_debug_registers(struct task_struct *task)
3527 unsigned long flags;
3530 if (pmu_conf->use_rr_dbregs == 0) return 0;
3533 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3534 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3537 pfm_sessions.pfs_ptrace_use_dbregs--;
3546 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3548 struct task_struct *task;
3549 pfm_buffer_fmt_t *fmt;
3550 pfm_ovfl_ctrl_t rst_ctrl;
3551 int state, is_system;
3554 state = ctx->ctx_state;
3555 fmt = ctx->ctx_buf_fmt;
3556 is_system = ctx->ctx_fl_system;
3557 task = PFM_CTX_TASK(ctx);
3560 case PFM_CTX_MASKED:
3562 case PFM_CTX_LOADED:
3563 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3565 case PFM_CTX_UNLOADED:
3566 case PFM_CTX_ZOMBIE:
3567 DPRINT(("invalid state=%d\n", state));
3570 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3575 * In system wide and when the context is loaded, access can only happen
3576 * when the caller is running on the CPU being monitored by the session.
3577 * It does not have to be the owner (ctx_task) of the context per se.
3579 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3580 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3585 if (unlikely(task == NULL)) {
3586 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3590 if (task == current || is_system) {
3592 fmt = ctx->ctx_buf_fmt;
3594 DPRINT(("restarting self %d ovfl=0x%lx\n",
3596 ctx->ctx_ovfl_regs[0]));
3598 if (CTX_HAS_SMPL(ctx)) {
3600 prefetch(ctx->ctx_smpl_hdr);
3602 rst_ctrl.bits.mask_monitoring = 0;
3603 rst_ctrl.bits.reset_ovfl_pmds = 0;
3605 if (state == PFM_CTX_LOADED)
3606 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3608 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3610 rst_ctrl.bits.mask_monitoring = 0;
3611 rst_ctrl.bits.reset_ovfl_pmds = 1;
3615 if (rst_ctrl.bits.reset_ovfl_pmds)
3616 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3618 if (rst_ctrl.bits.mask_monitoring == 0) {
3619 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3621 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3623 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3625 // cannot use pfm_stop_monitoring(task, regs);
3629 * clear overflowed PMD mask to remove any stale information
3631 ctx->ctx_ovfl_regs[0] = 0UL;
3634 * back to LOADED state
3636 ctx->ctx_state = PFM_CTX_LOADED;
3639 * XXX: not really useful for self monitoring
3641 ctx->ctx_fl_can_restart = 0;
3647 * restart another task
3651 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3652 * one is seen by the task.
3654 if (state == PFM_CTX_MASKED) {
3655 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3657 * will prevent subsequent restart before this one is
3658 * seen by other task
3660 ctx->ctx_fl_can_restart = 0;
3664 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3665 * the task is blocked or on its way to block. That's the normal
3666 * restart path. If the monitoring is not masked, then the task
3667 * can be actively monitoring and we cannot directly intervene.
3668 * Therefore we use the trap mechanism to catch the task and
3669 * force it to reset the buffer/reset PMDs.
3671 * if non-blocking, then we ensure that the task will go into
3672 * pfm_handle_work() before returning to user mode.
3674 * We cannot explicitely reset another task, it MUST always
3675 * be done by the task itself. This works for system wide because
3676 * the tool that is controlling the session is logically doing
3677 * "self-monitoring".
3679 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3680 DPRINT(("unblocking [%d] \n", task->pid));
3681 up(&ctx->ctx_restart_sem);
3683 DPRINT(("[%d] armed exit trap\n", task->pid));
3685 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3687 PFM_SET_WORK_PENDING(task, 1);
3689 pfm_set_task_notify(task);
3692 * XXX: send reschedule if task runs on another CPU
3699 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3701 unsigned int m = *(unsigned int *)arg;
3703 pfm_sysctl.debug = m == 0 ? 0 : 1;
3705 pfm_debug_var = pfm_sysctl.debug;
3707 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3710 memset(pfm_stats, 0, sizeof(pfm_stats));
3711 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3717 * arg can be NULL and count can be zero for this function
3720 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3722 struct thread_struct *thread = NULL;
3723 struct task_struct *task;
3724 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3725 unsigned long flags;
3730 int i, can_access_pmu = 0;
3731 int is_system, is_loaded;
3733 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3735 state = ctx->ctx_state;
3736 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3737 is_system = ctx->ctx_fl_system;
3738 task = ctx->ctx_task;
3740 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3743 * on both UP and SMP, we can only write to the PMC when the task is
3744 * the owner of the local PMU.
3747 thread = &task->thread;
3749 * In system wide and when the context is loaded, access can only happen
3750 * when the caller is running on the CPU being monitored by the session.
3751 * It does not have to be the owner (ctx_task) of the context per se.
3753 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3754 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3757 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3761 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3762 * ensuring that no real breakpoint can be installed via this call.
3764 * IMPORTANT: regs can be NULL in this function
3767 first_time = ctx->ctx_fl_using_dbreg == 0;
3770 * don't bother if we are loaded and task is being debugged
3772 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3773 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3778 * check for debug registers in system wide mode
3780 * If though a check is done in pfm_context_load(),
3781 * we must repeat it here, in case the registers are
3782 * written after the context is loaded
3787 if (first_time && is_system) {
3788 if (pfm_sessions.pfs_ptrace_use_dbregs)
3791 pfm_sessions.pfs_sys_use_dbregs++;
3796 if (ret != 0) return ret;
3799 * mark ourself as user of the debug registers for
3802 ctx->ctx_fl_using_dbreg = 1;
3805 * clear hardware registers to make sure we don't
3806 * pick up stale state.
3808 * for a system wide session, we do not use
3809 * thread.dbr, thread.ibr because this process
3810 * never leaves the current CPU and the state
3811 * is shared by all processes running on it
3813 if (first_time && can_access_pmu) {
3814 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3815 for (i=0; i < pmu_conf->num_ibrs; i++) {
3816 ia64_set_ibr(i, 0UL);
3817 ia64_dv_serialize_instruction();
3820 for (i=0; i < pmu_conf->num_dbrs; i++) {
3821 ia64_set_dbr(i, 0UL);
3822 ia64_dv_serialize_data();
3828 * Now install the values into the registers
3830 for (i = 0; i < count; i++, req++) {
3832 rnum = req->dbreg_num;
3833 dbreg.val = req->dbreg_value;
3837 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3838 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3839 rnum, dbreg.val, mode, i, count));
3845 * make sure we do not install enabled breakpoint
3848 if (mode == PFM_CODE_RR)
3849 dbreg.ibr.ibr_x = 0;
3851 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3854 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3857 * Debug registers, just like PMC, can only be modified
3858 * by a kernel call. Moreover, perfmon() access to those
3859 * registers are centralized in this routine. The hardware
3860 * does not modify the value of these registers, therefore,
3861 * if we save them as they are written, we can avoid having
3862 * to save them on context switch out. This is made possible
3863 * by the fact that when perfmon uses debug registers, ptrace()
3864 * won't be able to modify them concurrently.
3866 if (mode == PFM_CODE_RR) {
3867 CTX_USED_IBR(ctx, rnum);
3869 if (can_access_pmu) {
3870 ia64_set_ibr(rnum, dbreg.val);
3871 ia64_dv_serialize_instruction();
3874 ctx->ctx_ibrs[rnum] = dbreg.val;
3876 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3877 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3879 CTX_USED_DBR(ctx, rnum);
3881 if (can_access_pmu) {
3882 ia64_set_dbr(rnum, dbreg.val);
3883 ia64_dv_serialize_data();
3885 ctx->ctx_dbrs[rnum] = dbreg.val;
3887 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3888 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3896 * in case it was our first attempt, we undo the global modifications
3900 if (ctx->ctx_fl_system) {
3901 pfm_sessions.pfs_sys_use_dbregs--;
3904 ctx->ctx_fl_using_dbreg = 0;
3907 * install error return flag
3909 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3915 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3917 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3921 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3923 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3927 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3931 if (req == NULL) return -EINVAL;
3933 ctx = GET_PMU_CTX();
3935 if (ctx == NULL) return -EINVAL;
3938 * for now limit to current task, which is enough when calling
3939 * from overflow handler
3941 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3943 return pfm_write_ibrs(ctx, req, nreq, regs);
3945 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3948 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3952 if (req == NULL) return -EINVAL;
3954 ctx = GET_PMU_CTX();
3956 if (ctx == NULL) return -EINVAL;
3959 * for now limit to current task, which is enough when calling
3960 * from overflow handler
3962 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3964 return pfm_write_dbrs(ctx, req, nreq, regs);
3966 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3970 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3972 pfarg_features_t *req = (pfarg_features_t *)arg;
3974 req->ft_version = PFM_VERSION;
3979 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3981 struct pt_regs *tregs;
3982 struct task_struct *task = PFM_CTX_TASK(ctx);
3983 int state, is_system;
3985 state = ctx->ctx_state;
3986 is_system = ctx->ctx_fl_system;
3989 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3991 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3994 * In system wide and when the context is loaded, access can only happen
3995 * when the caller is running on the CPU being monitored by the session.
3996 * It does not have to be the owner (ctx_task) of the context per se.
3998 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3999 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4002 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4003 PFM_CTX_TASK(ctx)->pid,
4007 * in system mode, we need to update the PMU directly
4008 * and the user level state of the caller, which may not
4009 * necessarily be the creator of the context.
4013 * Update local PMU first
4017 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4021 * update local cpuinfo
4023 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4026 * stop monitoring, does srlz.i
4031 * stop monitoring in the caller
4033 ia64_psr(regs)->pp = 0;
4041 if (task == current) {
4042 /* stop monitoring at kernel level */
4046 * stop monitoring at the user level
4048 ia64_psr(regs)->up = 0;
4050 tregs = ia64_task_regs(task);
4053 * stop monitoring at the user level
4055 ia64_psr(tregs)->up = 0;
4058 * monitoring disabled in kernel at next reschedule
4060 ctx->ctx_saved_psr_up = 0;
4061 DPRINT(("task=[%d]\n", task->pid));
4068 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4070 struct pt_regs *tregs;
4071 int state, is_system;
4073 state = ctx->ctx_state;
4074 is_system = ctx->ctx_fl_system;
4076 if (state != PFM_CTX_LOADED) return -EINVAL;
4079 * In system wide and when the context is loaded, access can only happen
4080 * when the caller is running on the CPU being monitored by the session.
4081 * It does not have to be the owner (ctx_task) of the context per se.
4083 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4084 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4089 * in system mode, we need to update the PMU directly
4090 * and the user level state of the caller, which may not
4091 * necessarily be the creator of the context.
4096 * set user level psr.pp for the caller
4098 ia64_psr(regs)->pp = 1;
4101 * now update the local PMU and cpuinfo
4103 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4106 * start monitoring at kernel level
4111 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4121 if (ctx->ctx_task == current) {
4123 /* start monitoring at kernel level */
4127 * activate monitoring at user level
4129 ia64_psr(regs)->up = 1;
4132 tregs = ia64_task_regs(ctx->ctx_task);
4135 * start monitoring at the kernel level the next
4136 * time the task is scheduled
4138 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4141 * activate monitoring at user level
4143 ia64_psr(tregs)->up = 1;
4149 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4151 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4156 for (i = 0; i < count; i++, req++) {
4158 cnum = req->reg_num;
4160 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4162 req->reg_value = PMC_DFL_VAL(cnum);
4164 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4166 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4171 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4176 pfm_check_task_exist(pfm_context_t *ctx)
4178 struct task_struct *g, *t;
4181 read_lock(&tasklist_lock);
4183 do_each_thread (g, t) {
4184 if (t->thread.pfm_context == ctx) {
4188 } while_each_thread (g, t);
4190 read_unlock(&tasklist_lock);
4192 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4198 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4200 struct task_struct *task;
4201 struct thread_struct *thread;
4202 struct pfm_context_t *old;
4203 unsigned long flags;
4205 struct task_struct *owner_task = NULL;
4207 pfarg_load_t *req = (pfarg_load_t *)arg;
4208 unsigned long *pmcs_source, *pmds_source;
4211 int state, is_system, set_dbregs = 0;
4213 state = ctx->ctx_state;
4214 is_system = ctx->ctx_fl_system;
4216 * can only load from unloaded or terminated state
4218 if (state != PFM_CTX_UNLOADED) {
4219 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4225 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4227 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4228 DPRINT(("cannot use blocking mode on self\n"));
4232 ret = pfm_get_task(ctx, req->load_pid, &task);
4234 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4241 * system wide is self monitoring only
4243 if (is_system && task != current) {
4244 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4249 thread = &task->thread;
4253 * cannot load a context which is using range restrictions,
4254 * into a task that is being debugged.
4256 if (ctx->ctx_fl_using_dbreg) {
4257 if (thread->flags & IA64_THREAD_DBG_VALID) {
4259 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4265 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4266 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4269 pfm_sessions.pfs_sys_use_dbregs++;
4270 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4277 if (ret) goto error;
4281 * SMP system-wide monitoring implies self-monitoring.
4283 * The programming model expects the task to
4284 * be pinned on a CPU throughout the session.
4285 * Here we take note of the current CPU at the
4286 * time the context is loaded. No call from
4287 * another CPU will be allowed.
4289 * The pinning via shed_setaffinity()
4290 * must be done by the calling task prior
4293 * systemwide: keep track of CPU this session is supposed to run on
4295 the_cpu = ctx->ctx_cpu = smp_processor_id();
4299 * now reserve the session
4301 ret = pfm_reserve_session(current, is_system, the_cpu);
4302 if (ret) goto error;
4305 * task is necessarily stopped at this point.
4307 * If the previous context was zombie, then it got removed in
4308 * pfm_save_regs(). Therefore we should not see it here.
4309 * If we see a context, then this is an active context
4311 * XXX: needs to be atomic
4313 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4314 thread->pfm_context, ctx));
4316 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4318 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4322 pfm_reset_msgq(ctx);
4324 ctx->ctx_state = PFM_CTX_LOADED;
4327 * link context to task
4329 ctx->ctx_task = task;
4333 * we load as stopped
4335 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4336 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4338 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4340 thread->flags |= IA64_THREAD_PM_VALID;
4344 * propagate into thread-state
4346 pfm_copy_pmds(task, ctx);
4347 pfm_copy_pmcs(task, ctx);
4349 pmcs_source = thread->pmcs;
4350 pmds_source = thread->pmds;
4353 * always the case for system-wide
4355 if (task == current) {
4357 if (is_system == 0) {
4359 /* allow user level control */
4360 ia64_psr(regs)->sp = 0;
4361 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4363 SET_LAST_CPU(ctx, smp_processor_id());
4365 SET_ACTIVATION(ctx);
4368 * push the other task out, if any
4370 owner_task = GET_PMU_OWNER();
4371 if (owner_task) pfm_lazy_save_regs(owner_task);
4375 * load all PMD from ctx to PMU (as opposed to thread state)
4376 * restore all PMC from ctx to PMU
4378 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4379 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4381 ctx->ctx_reload_pmcs[0] = 0UL;
4382 ctx->ctx_reload_pmds[0] = 0UL;
4385 * guaranteed safe by earlier check against DBG_VALID
4387 if (ctx->ctx_fl_using_dbreg) {
4388 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4389 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4394 SET_PMU_OWNER(task, ctx);
4396 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4399 * when not current, task MUST be stopped, so this is safe
4401 regs = ia64_task_regs(task);
4403 /* force a full reload */
4404 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4405 SET_LAST_CPU(ctx, -1);
4407 /* initial saved psr (stopped) */
4408 ctx->ctx_saved_psr_up = 0UL;
4409 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4415 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4418 * we must undo the dbregs setting (for system-wide)
4420 if (ret && set_dbregs) {
4422 pfm_sessions.pfs_sys_use_dbregs--;
4426 * release task, there is now a link with the context
4428 if (is_system == 0 && task != current) {
4432 ret = pfm_check_task_exist(ctx);
4434 ctx->ctx_state = PFM_CTX_UNLOADED;
4435 ctx->ctx_task = NULL;
4443 * in this function, we do not need to increase the use count
4444 * for the task via get_task_struct(), because we hold the
4445 * context lock. If the task were to disappear while having
4446 * a context attached, it would go through pfm_exit_thread()
4447 * which also grabs the context lock and would therefore be blocked
4448 * until we are here.
4450 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4453 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4455 struct task_struct *task = PFM_CTX_TASK(ctx);
4456 struct pt_regs *tregs;
4457 int prev_state, is_system;
4460 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4462 prev_state = ctx->ctx_state;
4463 is_system = ctx->ctx_fl_system;
4466 * unload only when necessary
4468 if (prev_state == PFM_CTX_UNLOADED) {
4469 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4474 * clear psr and dcr bits
4476 ret = pfm_stop(ctx, NULL, 0, regs);
4477 if (ret) return ret;
4479 ctx->ctx_state = PFM_CTX_UNLOADED;
4482 * in system mode, we need to update the PMU directly
4483 * and the user level state of the caller, which may not
4484 * necessarily be the creator of the context.
4491 * local PMU is taken care of in pfm_stop()
4493 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4494 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4497 * save PMDs in context
4500 pfm_flush_pmds(current, ctx);
4503 * at this point we are done with the PMU
4504 * so we can unreserve the resource.
4506 if (prev_state != PFM_CTX_ZOMBIE)
4507 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4510 * disconnect context from task
4512 task->thread.pfm_context = NULL;
4514 * disconnect task from context
4516 ctx->ctx_task = NULL;
4519 * There is nothing more to cleanup here.
4527 tregs = task == current ? regs : ia64_task_regs(task);
4529 if (task == current) {
4531 * cancel user level control
4533 ia64_psr(regs)->sp = 1;
4535 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4538 * save PMDs to context
4541 pfm_flush_pmds(task, ctx);
4544 * at this point we are done with the PMU
4545 * so we can unreserve the resource.
4547 * when state was ZOMBIE, we have already unreserved.
4549 if (prev_state != PFM_CTX_ZOMBIE)
4550 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4553 * reset activation counter and psr
4555 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4556 SET_LAST_CPU(ctx, -1);
4559 * PMU state will not be restored
4561 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4564 * break links between context and task
4566 task->thread.pfm_context = NULL;
4567 ctx->ctx_task = NULL;
4569 PFM_SET_WORK_PENDING(task, 0);
4571 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4572 ctx->ctx_fl_can_restart = 0;
4573 ctx->ctx_fl_going_zombie = 0;
4575 DPRINT(("disconnected [%d] from context\n", task->pid));
4582 * called only from exit_thread(): task == current
4583 * we come here only if current has a context attached (loaded or masked)
4586 pfm_exit_thread(struct task_struct *task)
4589 unsigned long flags;
4590 struct pt_regs *regs = ia64_task_regs(task);
4594 ctx = PFM_GET_CTX(task);
4596 PROTECT_CTX(ctx, flags);
4598 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4600 state = ctx->ctx_state;
4602 case PFM_CTX_UNLOADED:
4604 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4605 * be in unloaded state
4607 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4609 case PFM_CTX_LOADED:
4610 case PFM_CTX_MASKED:
4611 ret = pfm_context_unload(ctx, NULL, 0, regs);
4613 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4615 DPRINT(("ctx unloaded for current state was %d\n", state));
4617 pfm_end_notify_user(ctx);
4619 case PFM_CTX_ZOMBIE:
4620 ret = pfm_context_unload(ctx, NULL, 0, regs);
4622 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4627 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4630 UNPROTECT_CTX(ctx, flags);
4632 { u64 psr = pfm_get_psr();
4633 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4634 BUG_ON(GET_PMU_OWNER());
4635 BUG_ON(ia64_psr(regs)->up);
4636 BUG_ON(ia64_psr(regs)->pp);
4640 * All memory free operations (especially for vmalloc'ed memory)
4641 * MUST be done with interrupts ENABLED.
4643 if (free_ok) pfm_context_free(ctx);
4647 * functions MUST be listed in the increasing order of their index (see permfon.h)
4649 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4650 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4651 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4652 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4653 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4655 static pfm_cmd_desc_t pfm_cmd_tab[]={
4656 /* 0 */PFM_CMD_NONE,
4657 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4658 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4659 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4660 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4661 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4662 /* 6 */PFM_CMD_NONE,
4663 /* 7 */PFM_CMD_NONE,
4664 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4665 /* 9 */PFM_CMD_NONE,
4666 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4667 /* 11 */PFM_CMD_NONE,
4668 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4669 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4670 /* 14 */PFM_CMD_NONE,
4671 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4672 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4673 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4674 /* 18 */PFM_CMD_NONE,
4675 /* 19 */PFM_CMD_NONE,
4676 /* 20 */PFM_CMD_NONE,
4677 /* 21 */PFM_CMD_NONE,
4678 /* 22 */PFM_CMD_NONE,
4679 /* 23 */PFM_CMD_NONE,
4680 /* 24 */PFM_CMD_NONE,
4681 /* 25 */PFM_CMD_NONE,
4682 /* 26 */PFM_CMD_NONE,
4683 /* 27 */PFM_CMD_NONE,
4684 /* 28 */PFM_CMD_NONE,
4685 /* 29 */PFM_CMD_NONE,
4686 /* 30 */PFM_CMD_NONE,
4687 /* 31 */PFM_CMD_NONE,
4688 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4689 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4691 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4694 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4696 struct task_struct *task;
4697 int state, old_state;
4700 state = ctx->ctx_state;
4701 task = ctx->ctx_task;
4704 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4708 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4712 task->state, PFM_CMD_STOPPED(cmd)));
4715 * self-monitoring always ok.
4717 * for system-wide the caller can either be the creator of the
4718 * context (to one to which the context is attached to) OR
4719 * a task running on the same CPU as the session.
4721 if (task == current || ctx->ctx_fl_system) return 0;
4724 * if context is UNLOADED we are safe to go
4726 if (state == PFM_CTX_UNLOADED) return 0;
4729 * no command can operate on a zombie context
4731 if (state == PFM_CTX_ZOMBIE) {
4732 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4737 * context is LOADED or MASKED. Some commands may need to have
4740 * We could lift this restriction for UP but it would mean that
4741 * the user has no guarantee the task would not run between
4742 * two successive calls to perfmonctl(). That's probably OK.
4743 * If this user wants to ensure the task does not run, then
4744 * the task must be stopped.
4746 if (PFM_CMD_STOPPED(cmd)) {
4747 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
4748 DPRINT(("[%d] task not in stopped state\n", task->pid));
4752 * task is now stopped, wait for ctxsw out
4754 * This is an interesting point in the code.
4755 * We need to unprotect the context because
4756 * the pfm_save_regs() routines needs to grab
4757 * the same lock. There are danger in doing
4758 * this because it leaves a window open for
4759 * another task to get access to the context
4760 * and possibly change its state. The one thing
4761 * that is not possible is for the context to disappear
4762 * because we are protected by the VFS layer, i.e.,
4763 * get_fd()/put_fd().
4767 UNPROTECT_CTX(ctx, flags);
4769 wait_task_inactive(task);
4771 PROTECT_CTX(ctx, flags);
4774 * we must recheck to verify if state has changed
4776 if (ctx->ctx_state != old_state) {
4777 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4785 * system-call entry point (must return long)
4788 sys_perfmonctl (int fd, int cmd, void __user *arg, int count, long arg5, long arg6, long arg7,
4789 long arg8, long stack)
4791 struct pt_regs *regs = (struct pt_regs *)&stack;
4792 struct file *file = NULL;
4793 pfm_context_t *ctx = NULL;
4794 unsigned long flags = 0UL;
4795 void *args_k = NULL;
4796 long ret; /* will expand int return types */
4797 size_t base_sz, sz, xtra_sz = 0;
4798 int narg, completed_args = 0, call_made = 0, cmd_flags;
4799 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4800 int (*getsize)(void *arg, size_t *sz);
4801 #define PFM_MAX_ARGSIZE 4096
4804 * reject any call if perfmon was disabled at initialization
4806 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4808 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4809 DPRINT(("invalid cmd=%d\n", cmd));
4813 func = pfm_cmd_tab[cmd].cmd_func;
4814 narg = pfm_cmd_tab[cmd].cmd_narg;
4815 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4816 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4817 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4819 if (unlikely(func == NULL)) {
4820 DPRINT(("invalid cmd=%d\n", cmd));
4824 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4832 * check if number of arguments matches what the command expects
4834 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4838 sz = xtra_sz + base_sz*count;
4840 * limit abuse to min page size
4842 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4843 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4848 * allocate default-sized argument buffer
4850 if (likely(count && args_k == NULL)) {
4851 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4852 if (args_k == NULL) return -ENOMEM;
4860 * assume sz = 0 for command without parameters
4862 if (sz && copy_from_user(args_k, arg, sz)) {
4863 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4868 * check if command supports extra parameters
4870 if (completed_args == 0 && getsize) {
4872 * get extra parameters size (based on main argument)
4874 ret = (*getsize)(args_k, &xtra_sz);
4875 if (ret) goto error_args;
4879 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4881 /* retry if necessary */
4882 if (likely(xtra_sz)) goto restart_args;
4885 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4890 if (unlikely(file == NULL)) {
4891 DPRINT(("invalid fd %d\n", fd));
4894 if (unlikely(PFM_IS_FILE(file) == 0)) {
4895 DPRINT(("fd %d not related to perfmon\n", fd));
4899 ctx = (pfm_context_t *)file->private_data;
4900 if (unlikely(ctx == NULL)) {
4901 DPRINT(("no context for fd %d\n", fd));
4904 prefetch(&ctx->ctx_state);
4906 PROTECT_CTX(ctx, flags);
4909 * check task is stopped
4911 ret = pfm_check_task_state(ctx, cmd, flags);
4912 if (unlikely(ret)) goto abort_locked;
4915 ret = (*func)(ctx, args_k, count, regs);
4921 DPRINT(("context unlocked\n"));
4922 UNPROTECT_CTX(ctx, flags);
4926 /* copy argument back to user, if needed */
4927 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4930 if (args_k) kfree(args_k);
4932 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4938 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4940 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4941 pfm_ovfl_ctrl_t rst_ctrl;
4945 state = ctx->ctx_state;
4947 * Unlock sampling buffer and reset index atomically
4948 * XXX: not really needed when blocking
4950 if (CTX_HAS_SMPL(ctx)) {
4952 rst_ctrl.bits.mask_monitoring = 0;
4953 rst_ctrl.bits.reset_ovfl_pmds = 0;
4955 if (state == PFM_CTX_LOADED)
4956 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4958 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4960 rst_ctrl.bits.mask_monitoring = 0;
4961 rst_ctrl.bits.reset_ovfl_pmds = 1;
4965 if (rst_ctrl.bits.reset_ovfl_pmds) {
4966 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4968 if (rst_ctrl.bits.mask_monitoring == 0) {
4969 DPRINT(("resuming monitoring\n"));
4970 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4972 DPRINT(("stopping monitoring\n"));
4973 //pfm_stop_monitoring(current, regs);
4975 ctx->ctx_state = PFM_CTX_LOADED;
4980 * context MUST BE LOCKED when calling
4981 * can only be called for current
4984 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4988 DPRINT(("entering for [%d]\n", current->pid));
4990 ret = pfm_context_unload(ctx, NULL, 0, regs);
4992 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", current->pid, ret);
4996 * and wakeup controlling task, indicating we are now disconnected
4998 wake_up_interruptible(&ctx->ctx_zombieq);
5001 * given that context is still locked, the controlling
5002 * task will only get access when we return from
5003 * pfm_handle_work().
5007 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5010 pfm_handle_work(void)
5013 struct pt_regs *regs;
5014 unsigned long flags;
5015 unsigned long ovfl_regs;
5016 unsigned int reason;
5019 ctx = PFM_GET_CTX(current);
5021 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5025 PROTECT_CTX(ctx, flags);
5027 PFM_SET_WORK_PENDING(current, 0);
5029 pfm_clear_task_notify();
5031 regs = ia64_task_regs(current);
5034 * extract reason for being here and clear
5036 reason = ctx->ctx_fl_trap_reason;
5037 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5038 ovfl_regs = ctx->ctx_ovfl_regs[0];
5040 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5043 * must be done before we check for simple-reset mode
5045 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5048 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5049 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5051 UNPROTECT_CTX(ctx, flags);
5054 * pfm_handle_work() is currently called with interrupts disabled.
5055 * The down_interruptible call may sleep, therefore we
5056 * must re-enable interrupts to avoid deadlocks. It is
5057 * safe to do so because this function is called ONLY
5058 * when returning to user level (PUStk=1), in which case
5059 * there is no risk of kernel stack overflow due to deep
5060 * interrupt nesting.
5062 BUG_ON(flags & IA64_PSR_I);
5065 DPRINT(("before block sleeping\n"));
5068 * may go through without blocking on SMP systems
5069 * if restart has been received already by the time we call down()
5071 ret = down_interruptible(&ctx->ctx_restart_sem);
5073 DPRINT(("after block sleeping ret=%d\n", ret));
5076 * disable interrupts to restore state we had upon entering
5079 local_irq_disable();
5081 PROTECT_CTX(ctx, flags);
5084 * we need to read the ovfl_regs only after wake-up
5085 * because we may have had pfm_write_pmds() in between
5086 * and that can changed PMD values and therefore
5087 * ovfl_regs is reset for these new PMD values.
5089 ovfl_regs = ctx->ctx_ovfl_regs[0];
5091 if (ctx->ctx_fl_going_zombie) {
5093 DPRINT(("context is zombie, bailing out\n"));
5094 pfm_context_force_terminate(ctx, regs);
5098 * in case of interruption of down() we don't restart anything
5100 if (ret < 0) goto nothing_to_do;
5103 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5104 ctx->ctx_ovfl_regs[0] = 0UL;
5108 UNPROTECT_CTX(ctx, flags);
5112 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5114 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5115 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5119 DPRINT(("waking up somebody\n"));
5121 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5124 * safe, we are not in intr handler, nor in ctxsw when
5127 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5133 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5135 pfm_msg_t *msg = NULL;
5137 if (ctx->ctx_fl_no_msg == 0) {
5138 msg = pfm_get_new_msg(ctx);
5140 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5144 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5145 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5146 msg->pfm_ovfl_msg.msg_active_set = 0;
5147 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5148 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5149 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5150 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5151 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5154 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5160 return pfm_notify_user(ctx, msg);
5164 pfm_end_notify_user(pfm_context_t *ctx)
5168 msg = pfm_get_new_msg(ctx);
5170 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5174 memset(msg, 0, sizeof(*msg));
5176 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5177 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5178 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5180 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5185 return pfm_notify_user(ctx, msg);
5189 * main overflow processing routine.
5190 * it can be called from the interrupt path or explicitely during the context switch code
5193 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5195 pfm_ovfl_arg_t *ovfl_arg;
5197 unsigned long old_val, ovfl_val, new_val;
5198 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5199 unsigned long tstamp;
5200 pfm_ovfl_ctrl_t ovfl_ctrl;
5201 unsigned int i, has_smpl;
5202 int must_notify = 0;
5204 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5207 * sanity test. Should never happen
5209 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5211 tstamp = ia64_get_itc();
5212 mask = pmc0 >> PMU_FIRST_COUNTER;
5213 ovfl_val = pmu_conf->ovfl_val;
5214 has_smpl = CTX_HAS_SMPL(ctx);
5216 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5217 "used_pmds=0x%lx\n",
5219 task ? task->pid: -1,
5220 (regs ? regs->cr_iip : 0),
5221 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5222 ctx->ctx_used_pmds[0]));
5226 * first we update the virtual counters
5227 * assume there was a prior ia64_srlz_d() issued
5229 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5231 /* skip pmd which did not overflow */
5232 if ((mask & 0x1) == 0) continue;
5235 * Note that the pmd is not necessarily 0 at this point as qualified events
5236 * may have happened before the PMU was frozen. The residual count is not
5237 * taken into consideration here but will be with any read of the pmd via
5240 old_val = new_val = ctx->ctx_pmds[i].val;
5241 new_val += 1 + ovfl_val;
5242 ctx->ctx_pmds[i].val = new_val;
5245 * check for overflow condition
5247 if (likely(old_val > new_val)) {
5248 ovfl_pmds |= 1UL << i;
5249 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5252 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5256 ia64_get_pmd(i) & ovfl_val,
5262 * there was no 64-bit overflow, nothing else to do
5264 if (ovfl_pmds == 0UL) return;
5267 * reset all control bits
5273 * if a sampling format module exists, then we "cache" the overflow by
5274 * calling the module's handler() routine.
5277 unsigned long start_cycles, end_cycles;
5278 unsigned long pmd_mask;
5280 int this_cpu = smp_processor_id();
5282 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5283 ovfl_arg = &ctx->ctx_ovfl_arg;
5285 prefetch(ctx->ctx_smpl_hdr);
5287 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5291 if ((pmd_mask & 0x1) == 0) continue;
5293 ovfl_arg->ovfl_pmd = (unsigned char )i;
5294 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5295 ovfl_arg->active_set = 0;
5296 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5297 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5299 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5300 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5301 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5304 * copy values of pmds of interest. Sampling format may copy them
5305 * into sampling buffer.
5308 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5309 if ((smpl_pmds & 0x1) == 0) continue;
5310 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5311 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5315 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5317 start_cycles = ia64_get_itc();
5320 * call custom buffer format record (handler) routine
5322 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5324 end_cycles = ia64_get_itc();
5327 * For those controls, we take the union because they have
5328 * an all or nothing behavior.
5330 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5331 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5332 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5334 * build the bitmask of pmds to reset now
5336 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5338 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5341 * when the module cannot handle the rest of the overflows, we abort right here
5343 if (ret && pmd_mask) {
5344 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5345 pmd_mask<<PMU_FIRST_COUNTER));
5348 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5350 ovfl_pmds &= ~reset_pmds;
5353 * when no sampling module is used, then the default
5354 * is to notify on overflow if requested by user
5356 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5357 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5358 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5359 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5361 * if needed, we reset all overflowed pmds
5363 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5366 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5369 * reset the requested PMD registers using the short reset values
5372 unsigned long bm = reset_pmds;
5373 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5376 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5378 * keep track of what to reset when unblocking
5380 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5383 * check for blocking context
5385 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5387 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5390 * set the perfmon specific checking pending work for the task
5392 PFM_SET_WORK_PENDING(task, 1);
5395 * when coming from ctxsw, current still points to the
5396 * previous task, therefore we must work with task and not current.
5398 pfm_set_task_notify(task);
5401 * defer until state is changed (shorten spin window). the context is locked
5402 * anyway, so the signal receiver would come spin for nothing.
5407 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5408 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5409 PFM_GET_WORK_PENDING(task),
5410 ctx->ctx_fl_trap_reason,
5413 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5415 * in case monitoring must be stopped, we toggle the psr bits
5417 if (ovfl_ctrl.bits.mask_monitoring) {
5418 pfm_mask_monitoring(task);
5419 ctx->ctx_state = PFM_CTX_MASKED;
5420 ctx->ctx_fl_can_restart = 1;
5424 * send notification now
5426 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5431 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5433 task ? task->pid : -1,
5439 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5440 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5441 * come here as zombie only if the task is the current task. In which case, we
5442 * can access the PMU hardware directly.
5444 * Note that zombies do have PM_VALID set. So here we do the minimal.
5446 * In case the context was zombified it could not be reclaimed at the time
5447 * the monitoring program exited. At this point, the PMU reservation has been
5448 * returned, the sampiing buffer has been freed. We must convert this call
5449 * into a spurious interrupt. However, we must also avoid infinite overflows
5450 * by stopping monitoring for this task. We can only come here for a per-task
5451 * context. All we need to do is to stop monitoring using the psr bits which
5452 * are always task private. By re-enabling secure montioring, we ensure that
5453 * the monitored task will not be able to re-activate monitoring.
5454 * The task will eventually be context switched out, at which point the context
5455 * will be reclaimed (that includes releasing ownership of the PMU).
5457 * So there might be a window of time where the number of per-task session is zero
5458 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5459 * context. This is safe because if a per-task session comes in, it will push this one
5460 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5461 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5462 * also push our zombie context out.
5464 * Overall pretty hairy stuff....
5466 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5468 ia64_psr(regs)->up = 0;
5469 ia64_psr(regs)->sp = 1;
5474 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5476 struct task_struct *task;
5478 unsigned long flags;
5480 int this_cpu = smp_processor_id();
5483 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5486 * srlz.d done before arriving here
5488 pmc0 = ia64_get_pmc(0);
5490 task = GET_PMU_OWNER();
5491 ctx = GET_PMU_CTX();
5494 * if we have some pending bits set
5495 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5497 if (PMC0_HAS_OVFL(pmc0) && task) {
5499 * we assume that pmc0.fr is always set here
5503 if (!ctx) goto report_spurious1;
5505 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5506 goto report_spurious2;
5508 PROTECT_CTX_NOPRINT(ctx, flags);
5510 pfm_overflow_handler(task, ctx, pmc0, regs);
5512 UNPROTECT_CTX_NOPRINT(ctx, flags);
5515 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5519 * keep it unfrozen at all times
5526 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5527 this_cpu, task->pid);
5531 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5539 pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5541 unsigned long start_cycles, total_cycles;
5542 unsigned long min, max;
5546 this_cpu = get_cpu();
5547 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5548 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5550 start_cycles = ia64_get_itc();
5552 ret = pfm_do_interrupt_handler(irq, arg, regs);
5554 total_cycles = ia64_get_itc();
5557 * don't measure spurious interrupts
5559 if (likely(ret == 0)) {
5560 total_cycles -= start_cycles;
5562 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5563 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5565 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5567 put_cpu_no_resched();
5572 * /proc/perfmon interface, for debug only
5575 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5578 pfm_proc_start(struct seq_file *m, loff_t *pos)
5581 return PFM_PROC_SHOW_HEADER;
5584 while (*pos <= NR_CPUS) {
5585 if (cpu_online(*pos - 1)) {
5586 return (void *)*pos;
5594 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5597 return pfm_proc_start(m, pos);
5601 pfm_proc_stop(struct seq_file *m, void *v)
5606 pfm_proc_show_header(struct seq_file *m)
5608 struct list_head * pos;
5609 pfm_buffer_fmt_t * entry;
5610 unsigned long flags;
5613 "perfmon version : %u.%u\n"
5616 "expert mode : %s\n"
5617 "ovfl_mask : 0x%lx\n"
5618 "PMU flags : 0x%x\n",
5619 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5621 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5622 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5629 "proc_sessions : %u\n"
5630 "sys_sessions : %u\n"
5631 "sys_use_dbregs : %u\n"
5632 "ptrace_use_dbregs : %u\n",
5633 pfm_sessions.pfs_task_sessions,
5634 pfm_sessions.pfs_sys_sessions,
5635 pfm_sessions.pfs_sys_use_dbregs,
5636 pfm_sessions.pfs_ptrace_use_dbregs);
5640 spin_lock(&pfm_buffer_fmt_lock);
5642 list_for_each(pos, &pfm_buffer_fmt_list) {
5643 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5644 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5655 entry->fmt_uuid[10],
5656 entry->fmt_uuid[11],
5657 entry->fmt_uuid[12],
5658 entry->fmt_uuid[13],
5659 entry->fmt_uuid[14],
5660 entry->fmt_uuid[15],
5663 spin_unlock(&pfm_buffer_fmt_lock);
5668 pfm_proc_show(struct seq_file *m, void *v)
5674 if (v == PFM_PROC_SHOW_HEADER) {
5675 pfm_proc_show_header(m);
5679 /* show info for CPU (v - 1) */
5683 "CPU%-2d overflow intrs : %lu\n"
5684 "CPU%-2d overflow cycles : %lu\n"
5685 "CPU%-2d overflow min : %lu\n"
5686 "CPU%-2d overflow max : %lu\n"
5687 "CPU%-2d smpl handler calls : %lu\n"
5688 "CPU%-2d smpl handler cycles : %lu\n"
5689 "CPU%-2d spurious intrs : %lu\n"
5690 "CPU%-2d replay intrs : %lu\n"
5691 "CPU%-2d syst_wide : %d\n"
5692 "CPU%-2d dcr_pp : %d\n"
5693 "CPU%-2d exclude idle : %d\n"
5694 "CPU%-2d owner : %d\n"
5695 "CPU%-2d context : %p\n"
5696 "CPU%-2d activations : %lu\n",
5697 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5698 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5699 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5700 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5701 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5702 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5703 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5704 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5705 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5706 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5707 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5708 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5709 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5710 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5712 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5714 psr = pfm_get_psr();
5719 "CPU%-2d psr : 0x%lx\n"
5720 "CPU%-2d pmc0 : 0x%lx\n",
5722 cpu, ia64_get_pmc(0));
5724 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5725 if (PMC_IS_COUNTING(i) == 0) continue;
5727 "CPU%-2d pmc%u : 0x%lx\n"
5728 "CPU%-2d pmd%u : 0x%lx\n",
5729 cpu, i, ia64_get_pmc(i),
5730 cpu, i, ia64_get_pmd(i));
5736 struct seq_operations pfm_seq_ops = {
5737 .start = pfm_proc_start,
5738 .next = pfm_proc_next,
5739 .stop = pfm_proc_stop,
5740 .show = pfm_proc_show
5744 pfm_proc_open(struct inode *inode, struct file *file)
5746 return seq_open(file, &pfm_seq_ops);
5751 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5752 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5753 * is active or inactive based on mode. We must rely on the value in
5754 * local_cpu_data->pfm_syst_info
5757 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5759 struct pt_regs *regs;
5761 unsigned long dcr_pp;
5763 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5766 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5767 * on every CPU, so we can rely on the pid to identify the idle task.
5769 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5770 regs = ia64_task_regs(task);
5771 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5775 * if monitoring has started
5778 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5780 * context switching in?
5783 /* mask monitoring for the idle task */
5784 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5790 * context switching out
5791 * restore monitoring for next task
5793 * Due to inlining this odd if-then-else construction generates
5796 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5805 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5807 struct task_struct *task = ctx->ctx_task;
5809 ia64_psr(regs)->up = 0;
5810 ia64_psr(regs)->sp = 1;
5812 if (GET_PMU_OWNER() == task) {
5813 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5814 SET_PMU_OWNER(NULL, NULL);
5818 * disconnect the task from the context and vice-versa
5820 PFM_SET_WORK_PENDING(task, 0);
5822 task->thread.pfm_context = NULL;
5823 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5825 DPRINT(("force cleanup for [%d]\n", task->pid));
5830 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5833 pfm_save_regs(struct task_struct *task)
5836 struct thread_struct *t;
5837 unsigned long flags;
5841 ctx = PFM_GET_CTX(task);
5842 if (ctx == NULL) return;
5846 * we always come here with interrupts ALREADY disabled by
5847 * the scheduler. So we simply need to protect against concurrent
5848 * access, not CPU concurrency.
5850 flags = pfm_protect_ctx_ctxsw(ctx);
5852 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5853 struct pt_regs *regs = ia64_task_regs(task);
5857 pfm_force_cleanup(ctx, regs);
5859 BUG_ON(ctx->ctx_smpl_hdr);
5861 pfm_unprotect_ctx_ctxsw(ctx, flags);
5863 pfm_context_free(ctx);
5868 * save current PSR: needed because we modify it
5871 psr = pfm_get_psr();
5873 BUG_ON(psr & (IA64_PSR_I));
5877 * This is the last instruction which may generate an overflow
5879 * We do not need to set psr.sp because, it is irrelevant in kernel.
5880 * It will be restored from ipsr when going back to user level
5885 * keep a copy of psr.up (for reload)
5887 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5890 * release ownership of this PMU.
5891 * PM interrupts are masked, so nothing
5894 SET_PMU_OWNER(NULL, NULL);
5897 * we systematically save the PMD as we have no
5898 * guarantee we will be schedule at that same
5901 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5904 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5905 * we will need it on the restore path to check
5906 * for pending overflow.
5908 t->pmcs[0] = ia64_get_pmc(0);
5911 * unfreeze PMU if had pending overflows
5913 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5916 * finally, allow context access.
5917 * interrupts will still be masked after this call.
5919 pfm_unprotect_ctx_ctxsw(ctx, flags);
5922 #else /* !CONFIG_SMP */
5924 pfm_save_regs(struct task_struct *task)
5929 ctx = PFM_GET_CTX(task);
5930 if (ctx == NULL) return;
5933 * save current PSR: needed because we modify it
5935 psr = pfm_get_psr();
5937 BUG_ON(psr & (IA64_PSR_I));
5941 * This is the last instruction which may generate an overflow
5943 * We do not need to set psr.sp because, it is irrelevant in kernel.
5944 * It will be restored from ipsr when going back to user level
5949 * keep a copy of psr.up (for reload)
5951 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5955 pfm_lazy_save_regs (struct task_struct *task)
5958 struct thread_struct *t;
5959 unsigned long flags;
5961 { u64 psr = pfm_get_psr();
5962 BUG_ON(psr & IA64_PSR_UP);
5965 ctx = PFM_GET_CTX(task);
5969 * we need to mask PMU overflow here to
5970 * make sure that we maintain pmc0 until
5971 * we save it. overflow interrupts are
5972 * treated as spurious if there is no
5975 * XXX: I don't think this is necessary
5977 PROTECT_CTX(ctx,flags);
5980 * release ownership of this PMU.
5981 * must be done before we save the registers.
5983 * after this call any PMU interrupt is treated
5986 SET_PMU_OWNER(NULL, NULL);
5989 * save all the pmds we use
5991 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5994 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5995 * it is needed to check for pended overflow
5996 * on the restore path
5998 t->pmcs[0] = ia64_get_pmc(0);
6001 * unfreeze PMU if had pending overflows
6003 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6006 * now get can unmask PMU interrupts, they will
6007 * be treated as purely spurious and we will not
6008 * lose any information
6010 UNPROTECT_CTX(ctx,flags);
6012 #endif /* CONFIG_SMP */
6016 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6019 pfm_load_regs (struct task_struct *task)
6022 struct thread_struct *t;
6023 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6024 unsigned long flags;
6026 int need_irq_resend;
6028 ctx = PFM_GET_CTX(task);
6029 if (unlikely(ctx == NULL)) return;
6031 BUG_ON(GET_PMU_OWNER());
6035 * possible on unload
6037 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6040 * we always come here with interrupts ALREADY disabled by
6041 * the scheduler. So we simply need to protect against concurrent
6042 * access, not CPU concurrency.
6044 flags = pfm_protect_ctx_ctxsw(ctx);
6045 psr = pfm_get_psr();
6047 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6049 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6050 BUG_ON(psr & IA64_PSR_I);
6052 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6053 struct pt_regs *regs = ia64_task_regs(task);
6055 BUG_ON(ctx->ctx_smpl_hdr);
6057 pfm_force_cleanup(ctx, regs);
6059 pfm_unprotect_ctx_ctxsw(ctx, flags);
6062 * this one (kmalloc'ed) is fine with interrupts disabled
6064 pfm_context_free(ctx);
6070 * we restore ALL the debug registers to avoid picking up
6073 if (ctx->ctx_fl_using_dbreg) {
6074 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6075 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6078 * retrieve saved psr.up
6080 psr_up = ctx->ctx_saved_psr_up;
6083 * if we were the last user of the PMU on that CPU,
6084 * then nothing to do except restore psr
6086 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6089 * retrieve partial reload masks (due to user modifications)
6091 pmc_mask = ctx->ctx_reload_pmcs[0];
6092 pmd_mask = ctx->ctx_reload_pmds[0];
6096 * To avoid leaking information to the user level when psr.sp=0,
6097 * we must reload ALL implemented pmds (even the ones we don't use).
6098 * In the kernel we only allow PFM_READ_PMDS on registers which
6099 * we initialized or requested (sampling) so there is no risk there.
6101 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6104 * ALL accessible PMCs are systematically reloaded, unused registers
6105 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6106 * up stale configuration.
6108 * PMC0 is never in the mask. It is always restored separately.
6110 pmc_mask = ctx->ctx_all_pmcs[0];
6113 * when context is MASKED, we will restore PMC with plm=0
6114 * and PMD with stale information, but that's ok, nothing
6117 * XXX: optimize here
6119 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6120 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6123 * check for pending overflow at the time the state
6126 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6128 * reload pmc0 with the overflow information
6129 * On McKinley PMU, this will trigger a PMU interrupt
6131 ia64_set_pmc(0, t->pmcs[0]);
6136 * will replay the PMU interrupt
6138 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6140 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6144 * we just did a reload, so we reset the partial reload fields
6146 ctx->ctx_reload_pmcs[0] = 0UL;
6147 ctx->ctx_reload_pmds[0] = 0UL;
6149 SET_LAST_CPU(ctx, smp_processor_id());
6152 * dump activation value for this PMU
6156 * record current activation for this context
6158 SET_ACTIVATION(ctx);
6161 * establish new ownership.
6163 SET_PMU_OWNER(task, ctx);
6166 * restore the psr.up bit. measurement
6168 * no PMU interrupt can happen at this point
6169 * because we still have interrupts disabled.
6171 if (likely(psr_up)) pfm_set_psr_up();
6174 * allow concurrent access to context
6176 pfm_unprotect_ctx_ctxsw(ctx, flags);
6178 #else /* !CONFIG_SMP */
6180 * reload PMU state for UP kernels
6181 * in 2.5 we come here with interrupts disabled
6184 pfm_load_regs (struct task_struct *task)
6186 struct thread_struct *t;
6188 struct task_struct *owner;
6189 unsigned long pmd_mask, pmc_mask;
6191 int need_irq_resend;
6193 owner = GET_PMU_OWNER();
6194 ctx = PFM_GET_CTX(task);
6196 psr = pfm_get_psr();
6198 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6199 BUG_ON(psr & IA64_PSR_I);
6202 * we restore ALL the debug registers to avoid picking up
6205 * This must be done even when the task is still the owner
6206 * as the registers may have been modified via ptrace()
6207 * (not perfmon) by the previous task.
6209 if (ctx->ctx_fl_using_dbreg) {
6210 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6211 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6215 * retrieved saved psr.up
6217 psr_up = ctx->ctx_saved_psr_up;
6218 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6221 * short path, our state is still there, just
6222 * need to restore psr and we go
6224 * we do not touch either PMC nor PMD. the psr is not touched
6225 * by the overflow_handler. So we are safe w.r.t. to interrupt
6226 * concurrency even without interrupt masking.
6228 if (likely(owner == task)) {
6229 if (likely(psr_up)) pfm_set_psr_up();
6234 * someone else is still using the PMU, first push it out and
6235 * then we'll be able to install our stuff !
6237 * Upon return, there will be no owner for the current PMU
6239 if (owner) pfm_lazy_save_regs(owner);
6242 * To avoid leaking information to the user level when psr.sp=0,
6243 * we must reload ALL implemented pmds (even the ones we don't use).
6244 * In the kernel we only allow PFM_READ_PMDS on registers which
6245 * we initialized or requested (sampling) so there is no risk there.
6247 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6250 * ALL accessible PMCs are systematically reloaded, unused registers
6251 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6252 * up stale configuration.
6254 * PMC0 is never in the mask. It is always restored separately
6256 pmc_mask = ctx->ctx_all_pmcs[0];
6258 pfm_restore_pmds(t->pmds, pmd_mask);
6259 pfm_restore_pmcs(t->pmcs, pmc_mask);
6262 * check for pending overflow at the time the state
6265 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6267 * reload pmc0 with the overflow information
6268 * On McKinley PMU, this will trigger a PMU interrupt
6270 ia64_set_pmc(0, t->pmcs[0]);
6276 * will replay the PMU interrupt
6278 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6280 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6284 * establish new ownership.
6286 SET_PMU_OWNER(task, ctx);
6289 * restore the psr.up bit. measurement
6291 * no PMU interrupt can happen at this point
6292 * because we still have interrupts disabled.
6294 if (likely(psr_up)) pfm_set_psr_up();
6296 #endif /* CONFIG_SMP */
6299 * this function assumes monitoring is stopped
6302 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6305 unsigned long mask2, val, pmd_val, ovfl_val;
6306 int i, can_access_pmu = 0;
6310 * is the caller the task being monitored (or which initiated the
6311 * session for system wide measurements)
6313 is_self = ctx->ctx_task == task ? 1 : 0;
6316 * can access PMU is task is the owner of the PMU state on the current CPU
6317 * or if we are running on the CPU bound to the context in system-wide mode
6318 * (that is not necessarily the task the context is attached to in this mode).
6319 * In system-wide we always have can_access_pmu true because a task running on an
6320 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6322 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6323 if (can_access_pmu) {
6325 * Mark the PMU as not owned
6326 * This will cause the interrupt handler to do nothing in case an overflow
6327 * interrupt was in-flight
6328 * This also guarantees that pmc0 will contain the final state
6329 * It virtually gives us full control on overflow processing from that point
6332 SET_PMU_OWNER(NULL, NULL);
6333 DPRINT(("releasing ownership\n"));
6336 * read current overflow status:
6338 * we are guaranteed to read the final stable state
6341 pmc0 = ia64_get_pmc(0); /* slow */
6344 * reset freeze bit, overflow status information destroyed
6348 pmc0 = task->thread.pmcs[0];
6350 * clear whatever overflow status bits there were
6352 task->thread.pmcs[0] = 0;
6354 ovfl_val = pmu_conf->ovfl_val;
6356 * we save all the used pmds
6357 * we take care of overflows for counting PMDs
6359 * XXX: sampling situation is not taken into account here
6361 mask2 = ctx->ctx_used_pmds[0];
6363 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6365 for (i = 0; mask2; i++, mask2>>=1) {
6367 /* skip non used pmds */
6368 if ((mask2 & 0x1) == 0) continue;
6371 * can access PMU always true in system wide mode
6373 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6375 if (PMD_IS_COUNTING(i)) {
6376 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6379 ctx->ctx_pmds[i].val,
6383 * we rebuild the full 64 bit value of the counter
6385 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6388 * now everything is in ctx_pmds[] and we need
6389 * to clear the saved context from save_regs() such that
6390 * pfm_read_pmds() gets the correct value
6395 * take care of overflow inline
6397 if (pmc0 & (1UL << i)) {
6398 val += 1 + ovfl_val;
6399 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6403 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, i, val, pmd_val));
6405 if (is_self) task->thread.pmds[i] = pmd_val;
6407 ctx->ctx_pmds[i].val = val;
6411 static struct irqaction perfmon_irqaction = {
6412 .handler = pfm_interrupt_handler,
6413 .flags = SA_INTERRUPT,
6418 * perfmon initialization routine, called from the initcall() table
6420 static int init_pfm_fs(void);
6428 family = local_cpu_data->family;
6433 if ((*p)->probe() == 0) goto found;
6434 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6445 static struct file_operations pfm_proc_fops = {
6446 .open = pfm_proc_open,
6448 .llseek = seq_lseek,
6449 .release = seq_release,
6455 unsigned int n, n_counters, i;
6457 printk("perfmon: version %u.%u IRQ %u\n",
6460 IA64_PERFMON_VECTOR);
6462 if (pfm_probe_pmu()) {
6463 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6464 local_cpu_data->family);
6469 * compute the number of implemented PMD/PMC from the
6470 * description tables
6473 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6474 if (PMC_IS_IMPL(i) == 0) continue;
6475 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6478 pmu_conf->num_pmcs = n;
6480 n = 0; n_counters = 0;
6481 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6482 if (PMD_IS_IMPL(i) == 0) continue;
6483 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6485 if (PMD_IS_COUNTING(i)) n_counters++;
6487 pmu_conf->num_pmds = n;
6488 pmu_conf->num_counters = n_counters;
6491 * sanity checks on the number of debug registers
6493 if (pmu_conf->use_rr_dbregs) {
6494 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6495 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6499 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6500 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6506 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6510 pmu_conf->num_counters,
6511 ffz(pmu_conf->ovfl_val));
6514 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6515 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6521 * create /proc/perfmon (mostly for debugging purposes)
6523 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6524 if (perfmon_dir == NULL) {
6525 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6530 * install customized file operations for /proc/perfmon entry
6532 perfmon_dir->proc_fops = &pfm_proc_fops;
6535 * create /proc/sys/kernel/perfmon (for debugging purposes)
6537 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6540 * initialize all our spinlocks
6542 spin_lock_init(&pfm_sessions.pfs_lock);
6543 spin_lock_init(&pfm_buffer_fmt_lock);
6547 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6552 __initcall(pfm_init);
6555 * this function is called before pfm_init()
6558 pfm_init_percpu (void)
6561 * make sure no measurement is active
6562 * (may inherit programmed PMCs from EFI).
6568 * we run with the PMU not frozen at all times
6572 if (smp_processor_id() == 0)
6573 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6575 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6580 * used for debug purposes only
6583 dump_pmu_state(const char *from)
6585 struct task_struct *task;
6586 struct thread_struct *t;
6587 struct pt_regs *regs;
6589 unsigned long psr, dcr, info, flags;
6592 local_irq_save(flags);
6594 this_cpu = smp_processor_id();
6595 regs = ia64_task_regs(current);
6596 info = PFM_CPUINFO_GET();
6597 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6599 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6600 local_irq_restore(flags);
6604 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6611 task = GET_PMU_OWNER();
6612 ctx = GET_PMU_CTX();
6614 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6616 psr = pfm_get_psr();
6618 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",
6621 psr & IA64_PSR_PP ? 1 : 0,
6622 psr & IA64_PSR_UP ? 1 : 0,
6623 dcr & IA64_DCR_PP ? 1 : 0,
6626 ia64_psr(regs)->pp);
6628 ia64_psr(regs)->up = 0;
6629 ia64_psr(regs)->pp = 0;
6631 t = ¤t->thread;
6633 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6634 if (PMC_IS_IMPL(i) == 0) continue;
6635 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6638 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6639 if (PMD_IS_IMPL(i) == 0) continue;
6640 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6644 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6647 ctx->ctx_smpl_vaddr,
6651 ctx->ctx_saved_psr_up);
6653 local_irq_restore(flags);
6657 * called from process.c:copy_thread(). task is new child.
6660 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6662 struct thread_struct *thread;
6664 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6666 thread = &task->thread;
6669 * cut links inherited from parent (current)
6671 thread->pfm_context = NULL;
6673 PFM_SET_WORK_PENDING(task, 0);
6676 * the psr bits are already set properly in copy_threads()
6679 #else /* !CONFIG_PERFMON */
6681 sys_perfmonctl (int fd, int cmd, void *arg, int count, long arg5, long arg6, long arg7,
6682 long arg8, long stack)
6686 #endif /* CONFIG_PERFMON */