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>
45 #include <asm/errno.h>
46 #include <asm/intrinsics.h>
48 #include <asm/perfmon.h>
49 #include <asm/processor.h>
50 #include <asm/signal.h>
51 #include <asm/system.h>
52 #include <asm/uaccess.h>
53 #include <asm/delay.h>
57 * perfmon context state
59 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
60 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
61 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
62 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
64 #define PFM_INVALID_ACTIVATION (~0UL)
67 * depth of message queue
69 #define PFM_MAX_MSGS 32
70 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
73 * type of a PMU register (bitmask).
75 * bit0 : register implemented
78 * bit4 : pmc has pmc.pm
79 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
80 * bit6-7 : register type
83 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
84 #define PFM_REG_IMPL 0x1 /* register implemented */
85 #define PFM_REG_END 0x2 /* end marker */
86 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
87 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
88 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
89 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
90 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
92 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
93 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
95 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
97 /* i assumed unsigned */
98 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
99 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
101 /* XXX: these assume that register i is implemented */
102 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
103 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
104 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
105 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
107 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
108 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
109 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
110 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
112 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
113 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
115 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
116 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
117 #define PFM_CTX_TASK(h) (h)->ctx_task
119 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
121 /* XXX: does not support more than 64 PMDs */
122 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
123 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
125 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
127 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
128 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
129 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
130 #define PFM_CODE_RR 0 /* requesting code range restriction */
131 #define PFM_DATA_RR 1 /* requestion data range restriction */
133 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
134 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
135 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
137 #define RDEP(x) (1UL<<(x))
140 * context protection macros
142 * - we need to protect against CPU concurrency (spin_lock)
143 * - we need to protect against PMU overflow interrupts (local_irq_disable)
145 * - we need to protect against PMU overflow interrupts (local_irq_disable)
147 * spin_lock_irqsave()/spin_lock_irqrestore():
148 * in SMP: local_irq_disable + spin_lock
149 * in UP : local_irq_disable
151 * spin_lock()/spin_lock():
152 * in UP : removed automatically
153 * in SMP: protect against context accesses from other CPU. interrupts
154 * are not masked. This is useful for the PMU interrupt handler
155 * because we know we will not get PMU concurrency in that code.
157 #define PROTECT_CTX(c, f) \
159 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \
160 spin_lock_irqsave(&(c)->ctx_lock, f); \
161 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
164 #define UNPROTECT_CTX(c, f) \
166 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
167 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
170 #define PROTECT_CTX_NOPRINT(c, f) \
172 spin_lock_irqsave(&(c)->ctx_lock, f); \
176 #define UNPROTECT_CTX_NOPRINT(c, f) \
178 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
182 #define PROTECT_CTX_NOIRQ(c) \
184 spin_lock(&(c)->ctx_lock); \
187 #define UNPROTECT_CTX_NOIRQ(c) \
189 spin_unlock(&(c)->ctx_lock); \
195 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
196 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
197 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
199 #else /* !CONFIG_SMP */
200 #define SET_ACTIVATION(t) do {} while(0)
201 #define GET_ACTIVATION(t) do {} while(0)
202 #define INC_ACTIVATION(t) do {} while(0)
203 #endif /* CONFIG_SMP */
205 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
206 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
207 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
209 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
210 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
212 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
215 * cmp0 must be the value of pmc0
217 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
219 #define PFMFS_MAGIC 0xa0b4d889
224 #define PFM_DEBUGGING 1
228 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
231 #define DPRINT_ovfl(a) \
233 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; } \
238 * 64-bit software counter structure
240 * the next_reset_type is applied to the next call to pfm_reset_regs()
243 unsigned long val; /* virtual 64bit counter value */
244 unsigned long lval; /* last reset value */
245 unsigned long long_reset; /* reset value on sampling overflow */
246 unsigned long short_reset; /* reset value on overflow */
247 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
248 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
249 unsigned long seed; /* seed for random-number generator */
250 unsigned long mask; /* mask for random-number generator */
251 unsigned int flags; /* notify/do not notify */
252 unsigned long eventid; /* overflow event identifier */
259 unsigned int block:1; /* when 1, task will blocked on user notifications */
260 unsigned int system:1; /* do system wide monitoring */
261 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
262 unsigned int is_sampling:1; /* true if using a custom format */
263 unsigned int excl_idle:1; /* exclude idle task in system wide session */
264 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
265 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
266 unsigned int no_msg:1; /* no message sent on overflow */
267 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
268 unsigned int reserved:22;
269 } pfm_context_flags_t;
271 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
272 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
273 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
277 * perfmon context: encapsulates all the state of a monitoring session
280 typedef struct pfm_context {
281 spinlock_t ctx_lock; /* context protection */
283 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
284 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
286 struct task_struct *ctx_task; /* task to which context is attached */
288 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
290 struct semaphore ctx_restart_sem; /* use for blocking notification mode */
292 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
293 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
294 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
296 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
297 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
298 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
300 unsigned long ctx_pmcs[IA64_NUM_PMC_REGS]; /* saved copies of PMC values */
302 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
303 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
304 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
305 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
307 pfm_counter_t ctx_pmds[IA64_NUM_PMD_REGS]; /* software state for PMDS */
309 u64 ctx_saved_psr_up; /* only contains psr.up value */
311 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
312 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
313 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
315 int ctx_fd; /* file descriptor used my this context */
316 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
318 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
319 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
320 unsigned long ctx_smpl_size; /* size of sampling buffer */
321 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
323 wait_queue_head_t ctx_msgq_wait;
324 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
327 struct fasync_struct *ctx_async_queue;
329 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
333 * magic number used to verify that structure is really
336 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
338 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
341 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
342 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
344 #define SET_LAST_CPU(ctx, v) do {} while(0)
345 #define GET_LAST_CPU(ctx) do {} while(0)
349 #define ctx_fl_block ctx_flags.block
350 #define ctx_fl_system ctx_flags.system
351 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
352 #define ctx_fl_is_sampling ctx_flags.is_sampling
353 #define ctx_fl_excl_idle ctx_flags.excl_idle
354 #define ctx_fl_going_zombie ctx_flags.going_zombie
355 #define ctx_fl_trap_reason ctx_flags.trap_reason
356 #define ctx_fl_no_msg ctx_flags.no_msg
357 #define ctx_fl_can_restart ctx_flags.can_restart
359 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
360 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
363 * global information about all sessions
364 * mostly used to synchronize between system wide and per-process
367 spinlock_t pfs_lock; /* lock the structure */
369 unsigned int pfs_task_sessions; /* number of per task sessions */
370 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
371 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
372 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
373 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
377 * information about a PMC or PMD.
378 * dep_pmd[]: a bitmask of dependent PMD registers
379 * dep_pmc[]: a bitmask of dependent PMC registers
381 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
385 unsigned long default_value; /* power-on default value */
386 unsigned long reserved_mask; /* bitmask of reserved bits */
387 pfm_reg_check_t read_check;
388 pfm_reg_check_t write_check;
389 unsigned long dep_pmd[4];
390 unsigned long dep_pmc[4];
393 /* assume cnum is a valid monitor */
394 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
397 * This structure is initialized at boot time and contains
398 * a description of the PMU main characteristics.
400 * If the probe function is defined, detection is based
401 * on its return value:
402 * - 0 means recognized PMU
403 * - anything else means not supported
404 * When the probe function is not defined, then the pmu_family field
405 * is used and it must match the host CPU family such that:
406 * - cpu->family & config->pmu_family != 0
409 unsigned long ovfl_val; /* overflow value for counters */
411 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
412 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
414 unsigned int num_pmcs; /* number of PMCS: computed at init time */
415 unsigned int num_pmds; /* number of PMDS: computed at init time */
416 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
417 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
419 char *pmu_name; /* PMU family name */
420 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
421 unsigned int flags; /* pmu specific flags */
422 unsigned int num_ibrs; /* number of IBRS: computed at init time */
423 unsigned int num_dbrs; /* number of DBRS: computed at init time */
424 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
425 int (*probe)(void); /* customized probe routine */
426 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
431 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
434 * debug register related type definitions
437 unsigned long ibr_mask:56;
438 unsigned long ibr_plm:4;
439 unsigned long ibr_ig:3;
440 unsigned long ibr_x:1;
444 unsigned long dbr_mask:56;
445 unsigned long dbr_plm:4;
446 unsigned long dbr_ig:2;
447 unsigned long dbr_w:1;
448 unsigned long dbr_r:1;
459 * perfmon command descriptions
462 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
465 unsigned int cmd_narg;
467 int (*cmd_getsize)(void *arg, size_t *sz);
470 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
471 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
472 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
473 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
476 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
477 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
478 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
479 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
480 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
482 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
485 int debug; /* turn on/off debugging via syslog */
486 int debug_ovfl; /* turn on/off debug printk in overflow handler */
487 int fastctxsw; /* turn on/off fast (unsecure) ctxsw */
488 int expert_mode; /* turn on/off value checking */
493 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
494 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
495 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
496 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
497 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
498 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
499 unsigned long pfm_smpl_handler_calls;
500 unsigned long pfm_smpl_handler_cycles;
501 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
505 * perfmon internal variables
507 static pfm_stats_t pfm_stats[NR_CPUS];
508 static pfm_session_t pfm_sessions; /* global sessions information */
510 static struct proc_dir_entry *perfmon_dir;
511 static pfm_uuid_t pfm_null_uuid = {0,};
513 static spinlock_t pfm_buffer_fmt_lock;
514 static LIST_HEAD(pfm_buffer_fmt_list);
516 static pmu_config_t *pmu_conf;
518 /* sysctl() controls */
519 static pfm_sysctl_t pfm_sysctl;
522 static ctl_table pfm_ctl_table[]={
523 {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
524 {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
525 {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
526 {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
529 static ctl_table pfm_sysctl_dir[] = {
530 {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
533 static ctl_table pfm_sysctl_root[] = {
534 {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
537 static struct ctl_table_header *pfm_sysctl_header;
539 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
540 static int pfm_flush(struct file *filp);
542 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
543 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
546 pfm_put_task(struct task_struct *task)
548 if (task != current) put_task_struct(task);
552 pfm_set_task_notify(struct task_struct *task)
554 struct thread_info *info;
556 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
557 set_bit(TIF_NOTIFY_RESUME, &info->flags);
561 pfm_clear_task_notify(void)
563 clear_thread_flag(TIF_NOTIFY_RESUME);
567 pfm_reserve_page(unsigned long a)
569 SetPageReserved(vmalloc_to_page((void *)a));
572 pfm_unreserve_page(unsigned long a)
574 ClearPageReserved(vmalloc_to_page((void*)a));
577 static inline unsigned long
578 pfm_protect_ctx_ctxsw(pfm_context_t *x)
580 spin_lock(&(x)->ctx_lock);
584 static inline unsigned long
585 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
587 spin_unlock(&(x)->ctx_lock);
590 static inline unsigned int
591 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
593 return do_munmap(mm, addr, len);
596 static inline unsigned long
597 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
599 return get_unmapped_area(file, addr, len, pgoff, flags);
603 static struct super_block *
604 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
606 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
609 static struct file_system_type pfm_fs_type = {
611 .get_sb = pfmfs_get_sb,
612 .kill_sb = kill_anon_super,
615 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
616 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
617 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
618 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
621 /* forward declaration */
622 static struct file_operations pfm_file_ops;
625 * forward declarations
628 static void pfm_lazy_save_regs (struct task_struct *ta);
631 void dump_pmu_state(const char *);
632 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
634 #include "perfmon_itanium.h"
635 #include "perfmon_mckinley.h"
636 #include "perfmon_generic.h"
638 static pmu_config_t *pmu_confs[]={
641 &pmu_conf_gen, /* must be last */
646 static int pfm_end_notify_user(pfm_context_t *ctx);
649 pfm_clear_psr_pp(void)
651 ia64_rsm(IA64_PSR_PP);
658 ia64_ssm(IA64_PSR_PP);
663 pfm_clear_psr_up(void)
665 ia64_rsm(IA64_PSR_UP);
672 ia64_ssm(IA64_PSR_UP);
676 static inline unsigned long
680 tmp = ia64_getreg(_IA64_REG_PSR);
686 pfm_set_psr_l(unsigned long val)
688 ia64_setreg(_IA64_REG_PSR_L, val);
700 pfm_unfreeze_pmu(void)
707 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
711 for (i=0; i < nibrs; i++) {
712 ia64_set_ibr(i, ibrs[i]);
713 ia64_dv_serialize_instruction();
719 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
723 for (i=0; i < ndbrs; i++) {
724 ia64_set_dbr(i, dbrs[i]);
725 ia64_dv_serialize_data();
731 * PMD[i] must be a counter. no check is made
733 static inline unsigned long
734 pfm_read_soft_counter(pfm_context_t *ctx, int i)
736 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
740 * PMD[i] must be a counter. no check is made
743 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
745 unsigned long ovfl_val = pmu_conf->ovfl_val;
747 ctx->ctx_pmds[i].val = val & ~ovfl_val;
749 * writing to unimplemented part is ignore, so we do not need to
752 ia64_set_pmd(i, val & ovfl_val);
756 pfm_get_new_msg(pfm_context_t *ctx)
760 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
762 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
763 if (next == ctx->ctx_msgq_head) return NULL;
765 idx = ctx->ctx_msgq_tail;
766 ctx->ctx_msgq_tail = next;
768 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
770 return ctx->ctx_msgq+idx;
774 pfm_get_next_msg(pfm_context_t *ctx)
778 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
780 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
785 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
790 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
792 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));
798 pfm_reset_msgq(pfm_context_t *ctx)
800 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
801 DPRINT(("ctx=%p msgq reset\n", ctx));
805 pfm_rvmalloc(unsigned long size)
810 size = PAGE_ALIGN(size);
813 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
814 memset(mem, 0, size);
815 addr = (unsigned long)mem;
817 pfm_reserve_page(addr);
826 pfm_rvfree(void *mem, unsigned long size)
831 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
832 addr = (unsigned long) mem;
833 while ((long) size > 0) {
834 pfm_unreserve_page(addr);
843 static pfm_context_t *
844 pfm_context_alloc(void)
849 * allocate context descriptor
850 * must be able to free with interrupts disabled
852 ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
854 memset(ctx, 0, sizeof(pfm_context_t));
855 DPRINT(("alloc ctx @%p\n", ctx));
861 pfm_context_free(pfm_context_t *ctx)
864 DPRINT(("free ctx @%p\n", ctx));
870 pfm_mask_monitoring(struct task_struct *task)
872 pfm_context_t *ctx = PFM_GET_CTX(task);
873 struct thread_struct *th = &task->thread;
874 unsigned long mask, val, ovfl_mask;
877 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
879 ovfl_mask = pmu_conf->ovfl_val;
881 * monitoring can only be masked as a result of a valid
882 * counter overflow. In UP, it means that the PMU still
883 * has an owner. Note that the owner can be different
884 * from the current task. However the PMU state belongs
886 * In SMP, a valid overflow only happens when task is
887 * current. Therefore if we come here, we know that
888 * the PMU state belongs to the current task, therefore
889 * we can access the live registers.
891 * So in both cases, the live register contains the owner's
892 * state. We can ONLY touch the PMU registers and NOT the PSR.
894 * As a consequence to this call, the thread->pmds[] array
895 * contains stale information which must be ignored
896 * when context is reloaded AND monitoring is active (see
899 mask = ctx->ctx_used_pmds[0];
900 for (i = 0; mask; i++, mask>>=1) {
901 /* skip non used pmds */
902 if ((mask & 0x1) == 0) continue;
903 val = ia64_get_pmd(i);
905 if (PMD_IS_COUNTING(i)) {
907 * we rebuild the full 64 bit value of the counter
909 ctx->ctx_pmds[i].val += (val & ovfl_mask);
911 ctx->ctx_pmds[i].val = val;
913 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
915 ctx->ctx_pmds[i].val,
919 * mask monitoring by setting the privilege level to 0
920 * we cannot use psr.pp/psr.up for this, it is controlled by
923 * if task is current, modify actual registers, otherwise modify
924 * thread save state, i.e., what will be restored in pfm_load_regs()
926 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
927 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
928 if ((mask & 0x1) == 0UL) continue;
929 ia64_set_pmc(i, th->pmcs[i] & ~0xfUL);
930 th->pmcs[i] &= ~0xfUL;
931 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i]));
934 * make all of this visible
940 * must always be done with task == current
942 * context must be in MASKED state when calling
945 pfm_restore_monitoring(struct task_struct *task)
947 pfm_context_t *ctx = PFM_GET_CTX(task);
948 struct thread_struct *th = &task->thread;
949 unsigned long mask, ovfl_mask;
950 unsigned long psr, val;
953 is_system = ctx->ctx_fl_system;
954 ovfl_mask = pmu_conf->ovfl_val;
956 if (task != current) {
957 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
960 if (ctx->ctx_state != PFM_CTX_MASKED) {
961 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
962 task->pid, current->pid, ctx->ctx_state);
967 * monitoring is masked via the PMC.
968 * As we restore their value, we do not want each counter to
969 * restart right away. We stop monitoring using the PSR,
970 * restore the PMC (and PMD) and then re-establish the psr
971 * as it was. Note that there can be no pending overflow at
972 * this point, because monitoring was MASKED.
974 * system-wide session are pinned and self-monitoring
976 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
978 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
984 * first, we restore the PMD
986 mask = ctx->ctx_used_pmds[0];
987 for (i = 0; mask; i++, mask>>=1) {
988 /* skip non used pmds */
989 if ((mask & 0x1) == 0) continue;
991 if (PMD_IS_COUNTING(i)) {
993 * we split the 64bit value according to
996 val = ctx->ctx_pmds[i].val & ovfl_mask;
997 ctx->ctx_pmds[i].val &= ~ovfl_mask;
999 val = ctx->ctx_pmds[i].val;
1001 ia64_set_pmd(i, val);
1003 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1005 ctx->ctx_pmds[i].val,
1011 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1012 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1013 if ((mask & 0x1) == 0UL) continue;
1014 th->pmcs[i] = ctx->ctx_pmcs[i];
1015 ia64_set_pmc(i, th->pmcs[i]);
1016 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
1021 * must restore DBR/IBR because could be modified while masked
1022 * XXX: need to optimize
1024 if (ctx->ctx_fl_using_dbreg) {
1025 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1026 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1032 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1034 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1041 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1047 for (i=0; mask; i++, mask>>=1) {
1048 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1053 * reload from thread state (used for ctxw only)
1056 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1059 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1061 for (i=0; mask; i++, mask>>=1) {
1062 if ((mask & 0x1) == 0) continue;
1063 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1064 ia64_set_pmd(i, val);
1070 * propagate PMD from context to thread-state
1073 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1075 struct thread_struct *thread = &task->thread;
1076 unsigned long ovfl_val = pmu_conf->ovfl_val;
1077 unsigned long mask = ctx->ctx_all_pmds[0];
1081 DPRINT(("mask=0x%lx\n", mask));
1083 for (i=0; mask; i++, mask>>=1) {
1085 val = ctx->ctx_pmds[i].val;
1088 * We break up the 64 bit value into 2 pieces
1089 * the lower bits go to the machine state in the
1090 * thread (will be reloaded on ctxsw in).
1091 * The upper part stays in the soft-counter.
1093 if (PMD_IS_COUNTING(i)) {
1094 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1097 thread->pmds[i] = val;
1099 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1102 ctx->ctx_pmds[i].val));
1107 * propagate PMC from context to thread-state
1110 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1112 struct thread_struct *thread = &task->thread;
1113 unsigned long mask = ctx->ctx_all_pmcs[0];
1116 DPRINT(("mask=0x%lx\n", mask));
1118 for (i=0; mask; i++, mask>>=1) {
1119 /* masking 0 with ovfl_val yields 0 */
1120 thread->pmcs[i] = ctx->ctx_pmcs[i];
1121 DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
1128 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1132 for (i=0; mask; i++, mask>>=1) {
1133 if ((mask & 0x1) == 0) continue;
1134 ia64_set_pmc(i, pmcs[i]);
1140 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1142 return memcmp(a, b, sizeof(pfm_uuid_t));
1146 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1149 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1154 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1157 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1163 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1167 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1172 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1176 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1181 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1184 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1189 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)
1192 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1196 static pfm_buffer_fmt_t *
1197 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1199 struct list_head * pos;
1200 pfm_buffer_fmt_t * entry;
1202 list_for_each(pos, &pfm_buffer_fmt_list) {
1203 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1204 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1211 * find a buffer format based on its uuid
1213 static pfm_buffer_fmt_t *
1214 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1216 pfm_buffer_fmt_t * fmt;
1217 spin_lock(&pfm_buffer_fmt_lock);
1218 fmt = __pfm_find_buffer_fmt(uuid);
1219 spin_unlock(&pfm_buffer_fmt_lock);
1224 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1228 /* some sanity checks */
1229 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1231 /* we need at least a handler */
1232 if (fmt->fmt_handler == NULL) return -EINVAL;
1235 * XXX: need check validity of fmt_arg_size
1238 spin_lock(&pfm_buffer_fmt_lock);
1240 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1241 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1245 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1246 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1249 spin_unlock(&pfm_buffer_fmt_lock);
1252 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1255 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1257 pfm_buffer_fmt_t *fmt;
1260 spin_lock(&pfm_buffer_fmt_lock);
1262 fmt = __pfm_find_buffer_fmt(uuid);
1264 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1268 list_del_init(&fmt->fmt_list);
1269 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1272 spin_unlock(&pfm_buffer_fmt_lock);
1276 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1279 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1281 unsigned long flags;
1283 * validy checks on cpu_mask have been done upstream
1287 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1288 pfm_sessions.pfs_sys_sessions,
1289 pfm_sessions.pfs_task_sessions,
1290 pfm_sessions.pfs_sys_use_dbregs,
1296 * cannot mix system wide and per-task sessions
1298 if (pfm_sessions.pfs_task_sessions > 0UL) {
1299 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1300 pfm_sessions.pfs_task_sessions));
1304 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1306 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1308 pfm_sessions.pfs_sys_session[cpu] = task;
1310 pfm_sessions.pfs_sys_sessions++ ;
1313 if (pfm_sessions.pfs_sys_sessions) goto abort;
1314 pfm_sessions.pfs_task_sessions++;
1317 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1318 pfm_sessions.pfs_sys_sessions,
1319 pfm_sessions.pfs_task_sessions,
1320 pfm_sessions.pfs_sys_use_dbregs,
1329 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1330 pfm_sessions.pfs_sys_session[cpu]->pid,
1331 smp_processor_id()));
1340 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1342 unsigned long flags;
1344 * validy checks on cpu_mask have been done upstream
1348 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1349 pfm_sessions.pfs_sys_sessions,
1350 pfm_sessions.pfs_task_sessions,
1351 pfm_sessions.pfs_sys_use_dbregs,
1357 pfm_sessions.pfs_sys_session[cpu] = NULL;
1359 * would not work with perfmon+more than one bit in cpu_mask
1361 if (ctx && ctx->ctx_fl_using_dbreg) {
1362 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1363 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1365 pfm_sessions.pfs_sys_use_dbregs--;
1368 pfm_sessions.pfs_sys_sessions--;
1370 pfm_sessions.pfs_task_sessions--;
1372 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1373 pfm_sessions.pfs_sys_sessions,
1374 pfm_sessions.pfs_task_sessions,
1375 pfm_sessions.pfs_sys_use_dbregs,
1385 * removes virtual mapping of the sampling buffer.
1386 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1387 * a PROTECT_CTX() section.
1390 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1395 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1396 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1400 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1403 * does the actual unmapping
1405 down_write(&task->mm->mmap_sem);
1407 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1409 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1411 up_write(&task->mm->mmap_sem);
1413 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1416 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1422 * free actual physical storage used by sampling buffer
1426 pfm_free_smpl_buffer(pfm_context_t *ctx)
1428 pfm_buffer_fmt_t *fmt;
1430 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1433 * we won't use the buffer format anymore
1435 fmt = ctx->ctx_buf_fmt;
1437 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1440 ctx->ctx_smpl_vaddr));
1442 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1447 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1449 ctx->ctx_smpl_hdr = NULL;
1450 ctx->ctx_smpl_size = 0UL;
1455 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1461 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1463 if (fmt == NULL) return;
1465 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1470 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1471 * no real gain from having the whole whorehouse mounted. So we don't need
1472 * any operations on the root directory. However, we need a non-trivial
1473 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1475 static struct vfsmount *pfmfs_mnt;
1480 int err = register_filesystem(&pfm_fs_type);
1482 pfmfs_mnt = kern_mount(&pfm_fs_type);
1483 err = PTR_ERR(pfmfs_mnt);
1484 if (IS_ERR(pfmfs_mnt))
1485 unregister_filesystem(&pfm_fs_type);
1495 unregister_filesystem(&pfm_fs_type);
1500 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1505 unsigned long flags;
1506 DECLARE_WAITQUEUE(wait, current);
1507 if (PFM_IS_FILE(filp) == 0) {
1508 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1512 ctx = (pfm_context_t *)filp->private_data;
1514 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1519 * check even when there is no message
1521 if (size < sizeof(pfm_msg_t)) {
1522 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1526 PROTECT_CTX(ctx, flags);
1529 * put ourselves on the wait queue
1531 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1539 set_current_state(TASK_INTERRUPTIBLE);
1541 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1544 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1546 UNPROTECT_CTX(ctx, flags);
1549 * check non-blocking read
1552 if(filp->f_flags & O_NONBLOCK) break;
1555 * check pending signals
1557 if(signal_pending(current)) {
1562 * no message, so wait
1566 PROTECT_CTX(ctx, flags);
1568 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1569 set_current_state(TASK_RUNNING);
1570 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1572 if (ret < 0) goto abort;
1575 msg = pfm_get_next_msg(ctx);
1577 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1581 DPRINT(("[%d] fd=%d type=%d\n", current->pid, msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1584 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1587 UNPROTECT_CTX(ctx, flags);
1593 pfm_write(struct file *file, const char __user *ubuf,
1594 size_t size, loff_t *ppos)
1596 DPRINT(("pfm_write called\n"));
1601 pfm_poll(struct file *filp, poll_table * wait)
1604 unsigned long flags;
1605 unsigned int mask = 0;
1607 if (PFM_IS_FILE(filp) == 0) {
1608 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1612 ctx = (pfm_context_t *)filp->private_data;
1614 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1619 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1621 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1623 PROTECT_CTX(ctx, flags);
1625 if (PFM_CTXQ_EMPTY(ctx) == 0)
1626 mask = POLLIN | POLLRDNORM;
1628 UNPROTECT_CTX(ctx, flags);
1630 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1636 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1638 DPRINT(("pfm_ioctl called\n"));
1643 * interrupt cannot be masked when coming here
1646 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1650 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1652 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1656 ctx->ctx_async_queue, ret));
1662 pfm_fasync(int fd, struct file *filp, int on)
1667 if (PFM_IS_FILE(filp) == 0) {
1668 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1672 ctx = (pfm_context_t *)filp->private_data;
1674 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1678 * we cannot mask interrupts during this call because this may
1679 * may go to sleep if memory is not readily avalaible.
1681 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1682 * done in caller. Serialization of this function is ensured by caller.
1684 ret = pfm_do_fasync(fd, filp, ctx, on);
1687 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1690 ctx->ctx_async_queue, ret));
1697 * this function is exclusively called from pfm_close().
1698 * The context is not protected at that time, nor are interrupts
1699 * on the remote CPU. That's necessary to avoid deadlocks.
1702 pfm_syswide_force_stop(void *info)
1704 pfm_context_t *ctx = (pfm_context_t *)info;
1705 struct pt_regs *regs = ia64_task_regs(current);
1706 struct task_struct *owner;
1707 unsigned long flags;
1710 if (ctx->ctx_cpu != smp_processor_id()) {
1711 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1713 smp_processor_id());
1716 owner = GET_PMU_OWNER();
1717 if (owner != ctx->ctx_task) {
1718 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1720 owner->pid, ctx->ctx_task->pid);
1723 if (GET_PMU_CTX() != ctx) {
1724 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1726 GET_PMU_CTX(), ctx);
1730 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1732 * the context is already protected in pfm_close(), we simply
1733 * need to mask interrupts to avoid a PMU interrupt race on
1736 local_irq_save(flags);
1738 ret = pfm_context_unload(ctx, NULL, 0, regs);
1740 DPRINT(("context_unload returned %d\n", ret));
1744 * unmask interrupts, PMU interrupts are now spurious here
1746 local_irq_restore(flags);
1750 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1754 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1755 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1756 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1758 #endif /* CONFIG_SMP */
1761 * called for each close(). Partially free resources.
1762 * When caller is self-monitoring, the context is unloaded.
1765 pfm_flush(struct file *filp)
1768 struct task_struct *task;
1769 struct pt_regs *regs;
1770 unsigned long flags;
1771 unsigned long smpl_buf_size = 0UL;
1772 void *smpl_buf_vaddr = NULL;
1773 int state, is_system;
1775 if (PFM_IS_FILE(filp) == 0) {
1776 DPRINT(("bad magic for\n"));
1780 ctx = (pfm_context_t *)filp->private_data;
1782 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1787 * remove our file from the async queue, if we use this mode.
1788 * This can be done without the context being protected. We come
1789 * here when the context has become unreacheable by other tasks.
1791 * We may still have active monitoring at this point and we may
1792 * end up in pfm_overflow_handler(). However, fasync_helper()
1793 * operates with interrupts disabled and it cleans up the
1794 * queue. If the PMU handler is called prior to entering
1795 * fasync_helper() then it will send a signal. If it is
1796 * invoked after, it will find an empty queue and no
1797 * signal will be sent. In both case, we are safe
1799 if (filp->f_flags & FASYNC) {
1800 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1801 pfm_do_fasync (-1, filp, ctx, 0);
1804 PROTECT_CTX(ctx, flags);
1806 state = ctx->ctx_state;
1807 is_system = ctx->ctx_fl_system;
1809 task = PFM_CTX_TASK(ctx);
1810 regs = ia64_task_regs(task);
1812 DPRINT(("ctx_state=%d is_current=%d\n",
1814 task == current ? 1 : 0));
1817 * if state == UNLOADED, then task is NULL
1821 * we must stop and unload because we are losing access to the context.
1823 if (task == current) {
1826 * the task IS the owner but it migrated to another CPU: that's bad
1827 * but we must handle this cleanly. Unfortunately, the kernel does
1828 * not provide a mechanism to block migration (while the context is loaded).
1830 * We need to release the resource on the ORIGINAL cpu.
1832 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1834 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1836 * keep context protected but unmask interrupt for IPI
1838 local_irq_restore(flags);
1840 pfm_syswide_cleanup_other_cpu(ctx);
1843 * restore interrupt masking
1845 local_irq_save(flags);
1848 * context is unloaded at this point
1851 #endif /* CONFIG_SMP */
1854 DPRINT(("forcing unload\n"));
1856 * stop and unload, returning with state UNLOADED
1857 * and session unreserved.
1859 pfm_context_unload(ctx, NULL, 0, regs);
1861 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1866 * remove virtual mapping, if any, for the calling task.
1867 * cannot reset ctx field until last user is calling close().
1869 * ctx_smpl_vaddr must never be cleared because it is needed
1870 * by every task with access to the context
1872 * When called from do_exit(), the mm context is gone already, therefore
1873 * mm is NULL, i.e., the VMA is already gone and we do not have to
1876 if (ctx->ctx_smpl_vaddr && current->mm) {
1877 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1878 smpl_buf_size = ctx->ctx_smpl_size;
1881 UNPROTECT_CTX(ctx, flags);
1884 * if there was a mapping, then we systematically remove it
1885 * at this point. Cannot be done inside critical section
1886 * because some VM function reenables interrupts.
1889 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1894 * called either on explicit close() or from exit_files().
1895 * Only the LAST user of the file gets to this point, i.e., it is
1898 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1899 * (fput()),i.e, last task to access the file. Nobody else can access the
1900 * file at this point.
1902 * When called from exit_files(), the VMA has been freed because exit_mm()
1903 * is executed before exit_files().
1905 * When called from exit_files(), the current task is not yet ZOMBIE but we
1906 * flush the PMU state to the context.
1909 pfm_close(struct inode *inode, struct file *filp)
1912 struct task_struct *task;
1913 struct pt_regs *regs;
1914 DECLARE_WAITQUEUE(wait, current);
1915 unsigned long flags;
1916 unsigned long smpl_buf_size = 0UL;
1917 void *smpl_buf_addr = NULL;
1918 int free_possible = 1;
1919 int state, is_system;
1921 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1923 if (PFM_IS_FILE(filp) == 0) {
1924 DPRINT(("bad magic\n"));
1928 ctx = (pfm_context_t *)filp->private_data;
1930 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1934 PROTECT_CTX(ctx, flags);
1936 state = ctx->ctx_state;
1937 is_system = ctx->ctx_fl_system;
1939 task = PFM_CTX_TASK(ctx);
1940 regs = ia64_task_regs(task);
1942 DPRINT(("ctx_state=%d is_current=%d\n",
1944 task == current ? 1 : 0));
1947 * if task == current, then pfm_flush() unloaded the context
1949 if (state == PFM_CTX_UNLOADED) goto doit;
1952 * context is loaded/masked and task != current, we need to
1953 * either force an unload or go zombie
1957 * The task is currently blocked or will block after an overflow.
1958 * we must force it to wakeup to get out of the
1959 * MASKED state and transition to the unloaded state by itself.
1961 * This situation is only possible for per-task mode
1963 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1966 * set a "partial" zombie state to be checked
1967 * upon return from down() in pfm_handle_work().
1969 * We cannot use the ZOMBIE state, because it is checked
1970 * by pfm_load_regs() which is called upon wakeup from down().
1971 * In such case, it would free the context and then we would
1972 * return to pfm_handle_work() which would access the
1973 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1974 * but visible to pfm_handle_work().
1976 * For some window of time, we have a zombie context with
1977 * ctx_state = MASKED and not ZOMBIE
1979 ctx->ctx_fl_going_zombie = 1;
1982 * force task to wake up from MASKED state
1984 up(&ctx->ctx_restart_sem);
1986 DPRINT(("waking up ctx_state=%d\n", state));
1989 * put ourself to sleep waiting for the other
1990 * task to report completion
1992 * the context is protected by mutex, therefore there
1993 * is no risk of being notified of completion before
1994 * begin actually on the waitq.
1996 set_current_state(TASK_INTERRUPTIBLE);
1997 add_wait_queue(&ctx->ctx_zombieq, &wait);
1999 UNPROTECT_CTX(ctx, flags);
2002 * XXX: check for signals :
2003 * - ok for explicit close
2004 * - not ok when coming from exit_files()
2009 PROTECT_CTX(ctx, flags);
2012 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2013 set_current_state(TASK_RUNNING);
2016 * context is unloaded at this point
2018 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2020 else if (task != current) {
2023 * switch context to zombie state
2025 ctx->ctx_state = PFM_CTX_ZOMBIE;
2027 DPRINT(("zombie ctx for [%d]\n", task->pid));
2029 * cannot free the context on the spot. deferred until
2030 * the task notices the ZOMBIE state
2034 pfm_context_unload(ctx, NULL, 0, regs);
2039 /* reload state, may have changed during opening of critical section */
2040 state = ctx->ctx_state;
2043 * the context is still attached to a task (possibly current)
2044 * we cannot destroy it right now
2048 * we must free the sampling buffer right here because
2049 * we cannot rely on it being cleaned up later by the
2050 * monitored task. It is not possible to free vmalloc'ed
2051 * memory in pfm_load_regs(). Instead, we remove the buffer
2052 * now. should there be subsequent PMU overflow originally
2053 * meant for sampling, the will be converted to spurious
2054 * and that's fine because the monitoring tools is gone anyway.
2056 if (ctx->ctx_smpl_hdr) {
2057 smpl_buf_addr = ctx->ctx_smpl_hdr;
2058 smpl_buf_size = ctx->ctx_smpl_size;
2059 /* no more sampling */
2060 ctx->ctx_smpl_hdr = NULL;
2061 ctx->ctx_fl_is_sampling = 0;
2064 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2070 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2073 * UNLOADED that the session has already been unreserved.
2075 if (state == PFM_CTX_ZOMBIE) {
2076 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2080 * disconnect file descriptor from context must be done
2083 filp->private_data = NULL;
2086 * if we free on the spot, the context is now completely unreacheable
2087 * from the callers side. The monitored task side is also cut, so we
2090 * If we have a deferred free, only the caller side is disconnected.
2092 UNPROTECT_CTX(ctx, flags);
2095 * All memory free operations (especially for vmalloc'ed memory)
2096 * MUST be done with interrupts ENABLED.
2098 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2101 * return the memory used by the context
2103 if (free_possible) pfm_context_free(ctx);
2109 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2111 DPRINT(("pfm_no_open called\n"));
2117 static struct file_operations pfm_file_ops = {
2118 .llseek = no_llseek,
2123 .open = pfm_no_open, /* special open code to disallow open via /proc */
2124 .fasync = pfm_fasync,
2125 .release = pfm_close,
2130 pfmfs_delete_dentry(struct dentry *dentry)
2135 static struct dentry_operations pfmfs_dentry_operations = {
2136 .d_delete = pfmfs_delete_dentry,
2141 pfm_alloc_fd(struct file **cfile)
2144 struct file *file = NULL;
2145 struct inode * inode;
2149 fd = get_unused_fd();
2150 if (fd < 0) return -ENFILE;
2154 file = get_empty_filp();
2155 if (!file) goto out;
2158 * allocate a new inode
2160 inode = new_inode(pfmfs_mnt->mnt_sb);
2161 if (!inode) goto out;
2163 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2165 inode->i_sb = pfmfs_mnt->mnt_sb;
2166 inode->i_mode = S_IFCHR|S_IRUGO;
2168 inode->i_uid = current->fsuid;
2169 inode->i_gid = current->fsgid;
2171 sprintf(name, "[%lu]", inode->i_ino);
2173 this.len = strlen(name);
2174 this.hash = inode->i_ino;
2179 * allocate a new dcache entry
2181 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2182 if (!file->f_dentry) goto out;
2184 file->f_dentry->d_op = &pfmfs_dentry_operations;
2186 d_add(file->f_dentry, inode);
2187 file->f_vfsmnt = mntget(pfmfs_mnt);
2188 file->f_mapping = inode->i_mapping;
2190 file->f_op = &pfm_file_ops;
2191 file->f_mode = FMODE_READ;
2192 file->f_flags = O_RDONLY;
2196 * may have to delay until context is attached?
2198 fd_install(fd, file);
2201 * the file structure we will use
2207 if (file) put_filp(file);
2213 pfm_free_fd(int fd, struct file *file)
2215 struct files_struct *files = current->files;
2218 * there ie no fd_uninstall(), so we do it here
2220 spin_lock(&files->file_lock);
2221 files->fd[fd] = NULL;
2222 spin_unlock(&files->file_lock);
2224 if (file) put_filp(file);
2229 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2231 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2234 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2237 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2248 * allocate a sampling buffer and remaps it into the user address space of the task
2251 pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2253 struct mm_struct *mm = task->mm;
2254 struct vm_area_struct *vma = NULL;
2260 * the fixed header + requested size and align to page boundary
2262 size = PAGE_ALIGN(rsize);
2264 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2267 * check requested size to avoid Denial-of-service attacks
2268 * XXX: may have to refine this test
2269 * Check against address space limit.
2271 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2274 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2278 * We do the easy to undo allocations first.
2280 * pfm_rvmalloc(), clears the buffer, so there is no leak
2282 smpl_buf = pfm_rvmalloc(size);
2283 if (smpl_buf == NULL) {
2284 DPRINT(("Can't allocate sampling buffer\n"));
2288 DPRINT(("smpl_buf @%p\n", smpl_buf));
2291 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2293 DPRINT(("Cannot allocate vma\n"));
2296 memset(vma, 0, sizeof(*vma));
2299 * partially initialize the vma for the sampling buffer
2302 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2303 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2306 * Now we have everything we need and we can initialize
2307 * and connect all the data structures
2310 ctx->ctx_smpl_hdr = smpl_buf;
2311 ctx->ctx_smpl_size = size; /* aligned size */
2314 * Let's do the difficult operations next.
2316 * now we atomically find some area in the address space and
2317 * remap the buffer in it.
2319 down_write(&task->mm->mmap_sem);
2321 /* find some free area in address space, must have mmap sem held */
2322 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2323 if (vma->vm_start == 0UL) {
2324 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2325 up_write(&task->mm->mmap_sem);
2328 vma->vm_end = vma->vm_start + size;
2329 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2331 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2333 /* can only be applied to current task, need to have the mm semaphore held when called */
2334 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2335 DPRINT(("Can't remap buffer\n"));
2336 up_write(&task->mm->mmap_sem);
2341 * now insert the vma in the vm list for the process, must be
2342 * done with mmap lock held
2344 insert_vm_struct(mm, vma);
2346 // mm->total_vm += size >> PAGE_SHIFT;
2347 vx_vmpages_add(mm, size >> PAGE_SHIFT);
2348 vm_stat_account(vma);
2349 up_write(&task->mm->mmap_sem);
2352 * keep track of user level virtual address
2354 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2355 *(unsigned long *)user_vaddr = vma->vm_start;
2360 kmem_cache_free(vm_area_cachep, vma);
2362 pfm_rvfree(smpl_buf, size);
2368 * XXX: do something better here
2371 pfm_bad_permissions(struct task_struct *task)
2373 /* inspired by ptrace_attach() */
2374 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2383 return ((current->uid != task->euid)
2384 || (current->uid != task->suid)
2385 || (current->uid != task->uid)
2386 || (current->gid != task->egid)
2387 || (current->gid != task->sgid)
2388 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2392 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2398 ctx_flags = pfx->ctx_flags;
2400 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2403 * cannot block in this mode
2405 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2406 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2411 /* probably more to add here */
2417 pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2418 unsigned int cpu, pfarg_context_t *arg)
2420 pfm_buffer_fmt_t *fmt = NULL;
2421 unsigned long size = 0UL;
2423 void *fmt_arg = NULL;
2425 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2427 /* invoke and lock buffer format, if found */
2428 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2430 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2435 * buffer argument MUST be contiguous to pfarg_context_t
2437 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2439 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2441 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2443 if (ret) goto error;
2445 /* link buffer format and context */
2446 ctx->ctx_buf_fmt = fmt;
2449 * check if buffer format wants to use perfmon buffer allocation/mapping service
2451 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2452 if (ret) goto error;
2456 * buffer is always remapped into the caller's address space
2458 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2459 if (ret) goto error;
2461 /* keep track of user address of buffer */
2462 arg->ctx_smpl_vaddr = uaddr;
2464 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2471 pfm_reset_pmu_state(pfm_context_t *ctx)
2476 * install reset values for PMC.
2478 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2479 if (PMC_IS_IMPL(i) == 0) continue;
2480 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2481 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2484 * PMD registers are set to 0UL when the context in memset()
2488 * On context switched restore, we must restore ALL pmc and ALL pmd even
2489 * when they are not actively used by the task. In UP, the incoming process
2490 * may otherwise pick up left over PMC, PMD state from the previous process.
2491 * As opposed to PMD, stale PMC can cause harm to the incoming
2492 * process because they may change what is being measured.
2493 * Therefore, we must systematically reinstall the entire
2494 * PMC state. In SMP, the same thing is possible on the
2495 * same CPU but also on between 2 CPUs.
2497 * The problem with PMD is information leaking especially
2498 * to user level when psr.sp=0
2500 * There is unfortunately no easy way to avoid this problem
2501 * on either UP or SMP. This definitively slows down the
2502 * pfm_load_regs() function.
2506 * bitmask of all PMCs accessible to this context
2508 * PMC0 is treated differently.
2510 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2513 * bitmask of all PMDs that are accesible to this context
2515 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2517 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2520 * useful in case of re-enable after disable
2522 ctx->ctx_used_ibrs[0] = 0UL;
2523 ctx->ctx_used_dbrs[0] = 0UL;
2527 pfm_ctx_getsize(void *arg, size_t *sz)
2529 pfarg_context_t *req = (pfarg_context_t *)arg;
2530 pfm_buffer_fmt_t *fmt;
2534 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2536 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2538 DPRINT(("cannot find buffer format\n"));
2541 /* get just enough to copy in user parameters */
2542 *sz = fmt->fmt_arg_size;
2543 DPRINT(("arg_size=%lu\n", *sz));
2551 * cannot attach if :
2553 * - task not owned by caller
2554 * - task incompatible with context mode
2557 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2560 * no kernel task or task not owner by caller
2562 if (task->mm == NULL) {
2563 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2566 if (pfm_bad_permissions(task)) {
2567 DPRINT(("no permission to attach to [%d]\n", task->pid));
2571 * cannot block in self-monitoring mode
2573 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2574 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2578 if (task->exit_state == EXIT_ZOMBIE) {
2579 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2584 * always ok for self
2586 if (task == current) return 0;
2588 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
2589 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2593 * make sure the task is off any CPU
2595 wait_task_inactive(task);
2597 /* more to come... */
2603 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2605 struct task_struct *p = current;
2608 /* XXX: need to add more checks here */
2609 if (pid < 2) return -EPERM;
2611 if (pid != current->pid) {
2613 read_lock(&tasklist_lock);
2615 p = find_task_by_pid(pid);
2617 /* make sure task cannot go away while we operate on it */
2618 if (p) get_task_struct(p);
2620 read_unlock(&tasklist_lock);
2622 if (p == NULL) return -ESRCH;
2625 ret = pfm_task_incompatible(ctx, p);
2628 } else if (p != current) {
2637 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2639 pfarg_context_t *req = (pfarg_context_t *)arg;
2644 /* let's check the arguments first */
2645 ret = pfarg_is_sane(current, req);
2646 if (ret < 0) return ret;
2648 ctx_flags = req->ctx_flags;
2652 ctx = pfm_context_alloc();
2653 if (!ctx) goto error;
2655 ret = pfm_alloc_fd(&filp);
2656 if (ret < 0) goto error_file;
2658 req->ctx_fd = ctx->ctx_fd = ret;
2661 * attach context to file
2663 filp->private_data = ctx;
2666 * does the user want to sample?
2668 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2669 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2670 if (ret) goto buffer_error;
2674 * init context protection lock
2676 spin_lock_init(&ctx->ctx_lock);
2679 * context is unloaded
2681 ctx->ctx_state = PFM_CTX_UNLOADED;
2684 * initialization of context's flags
2686 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2687 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2688 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2689 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2691 * will move to set properties
2692 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2696 * init restart semaphore to locked
2698 sema_init(&ctx->ctx_restart_sem, 0);
2701 * activation is used in SMP only
2703 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2704 SET_LAST_CPU(ctx, -1);
2707 * initialize notification message queue
2709 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2710 init_waitqueue_head(&ctx->ctx_msgq_wait);
2711 init_waitqueue_head(&ctx->ctx_zombieq);
2713 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2718 ctx->ctx_fl_excl_idle,
2723 * initialize soft PMU state
2725 pfm_reset_pmu_state(ctx);
2730 pfm_free_fd(ctx->ctx_fd, filp);
2732 if (ctx->ctx_buf_fmt) {
2733 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2736 pfm_context_free(ctx);
2742 static inline unsigned long
2743 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2745 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2746 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2747 extern unsigned long carta_random32 (unsigned long seed);
2749 if (reg->flags & PFM_REGFL_RANDOM) {
2750 new_seed = carta_random32(old_seed);
2751 val -= (old_seed & mask); /* counter values are negative numbers! */
2752 if ((mask >> 32) != 0)
2753 /* construct a full 64-bit random value: */
2754 new_seed |= carta_random32(old_seed >> 32) << 32;
2755 reg->seed = new_seed;
2762 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2764 unsigned long mask = ovfl_regs[0];
2765 unsigned long reset_others = 0UL;
2770 * now restore reset value on sampling overflowed counters
2772 mask >>= PMU_FIRST_COUNTER;
2773 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2775 if ((mask & 0x1UL) == 0UL) continue;
2777 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2778 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2780 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2784 * Now take care of resetting the other registers
2786 for(i = 0; reset_others; i++, reset_others >>= 1) {
2788 if ((reset_others & 0x1) == 0) continue;
2790 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2792 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2793 is_long_reset ? "long" : "short", i, val));
2798 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2800 unsigned long mask = ovfl_regs[0];
2801 unsigned long reset_others = 0UL;
2805 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2807 if (ctx->ctx_state == PFM_CTX_MASKED) {
2808 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2813 * now restore reset value on sampling overflowed counters
2815 mask >>= PMU_FIRST_COUNTER;
2816 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2818 if ((mask & 0x1UL) == 0UL) continue;
2820 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2821 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2823 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2825 pfm_write_soft_counter(ctx, i, val);
2829 * Now take care of resetting the other registers
2831 for(i = 0; reset_others; i++, reset_others >>= 1) {
2833 if ((reset_others & 0x1) == 0) continue;
2835 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2837 if (PMD_IS_COUNTING(i)) {
2838 pfm_write_soft_counter(ctx, i, val);
2840 ia64_set_pmd(i, val);
2842 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2843 is_long_reset ? "long" : "short", i, val));
2849 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2851 struct thread_struct *thread = NULL;
2852 struct task_struct *task;
2853 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2854 unsigned long value, pmc_pm;
2855 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2856 unsigned int cnum, reg_flags, flags, pmc_type;
2857 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2858 int is_monitor, is_counting, state;
2860 pfm_reg_check_t wr_func;
2861 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2863 state = ctx->ctx_state;
2864 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2865 is_system = ctx->ctx_fl_system;
2866 task = ctx->ctx_task;
2867 impl_pmds = pmu_conf->impl_pmds[0];
2869 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2872 thread = &task->thread;
2874 * In system wide and when the context is loaded, access can only happen
2875 * when the caller is running on the CPU being monitored by the session.
2876 * It does not have to be the owner (ctx_task) of the context per se.
2878 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2879 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2882 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2884 expert_mode = pfm_sysctl.expert_mode;
2886 for (i = 0; i < count; i++, req++) {
2888 cnum = req->reg_num;
2889 reg_flags = req->reg_flags;
2890 value = req->reg_value;
2891 smpl_pmds = req->reg_smpl_pmds[0];
2892 reset_pmds = req->reg_reset_pmds[0];
2896 if (cnum >= PMU_MAX_PMCS) {
2897 DPRINT(("pmc%u is invalid\n", cnum));
2901 pmc_type = pmu_conf->pmc_desc[cnum].type;
2902 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2903 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2904 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2907 * we reject all non implemented PMC as well
2908 * as attempts to modify PMC[0-3] which are used
2909 * as status registers by the PMU
2911 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2912 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2915 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2917 * If the PMC is a monitor, then if the value is not the default:
2918 * - system-wide session: PMCx.pm=1 (privileged monitor)
2919 * - per-task : PMCx.pm=0 (user monitor)
2921 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2922 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2931 * enforce generation of overflow interrupt. Necessary on all
2934 value |= 1 << PMU_PMC_OI;
2936 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2937 flags |= PFM_REGFL_OVFL_NOTIFY;
2940 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2942 /* verify validity of smpl_pmds */
2943 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2944 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2948 /* verify validity of reset_pmds */
2949 if ((reset_pmds & impl_pmds) != reset_pmds) {
2950 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2954 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2955 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2958 /* eventid on non-counting monitors are ignored */
2962 * execute write checker, if any
2964 if (likely(expert_mode == 0 && wr_func)) {
2965 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2966 if (ret) goto error;
2971 * no error on this register
2973 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2976 * Now we commit the changes to the software state
2980 * update overflow information
2984 * full flag update each time a register is programmed
2986 ctx->ctx_pmds[cnum].flags = flags;
2988 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2989 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2990 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2993 * Mark all PMDS to be accessed as used.
2995 * We do not keep track of PMC because we have to
2996 * systematically restore ALL of them.
2998 * We do not update the used_monitors mask, because
2999 * if we have not programmed them, then will be in
3000 * a quiescent state, therefore we will not need to
3001 * mask/restore then when context is MASKED.
3003 CTX_USED_PMD(ctx, reset_pmds);
3004 CTX_USED_PMD(ctx, smpl_pmds);
3006 * make sure we do not try to reset on
3007 * restart because we have established new values
3009 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3012 * Needed in case the user does not initialize the equivalent
3013 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3014 * possible leak here.
3016 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3019 * keep track of the monitor PMC that we are using.
3020 * we save the value of the pmc in ctx_pmcs[] and if
3021 * the monitoring is not stopped for the context we also
3022 * place it in the saved state area so that it will be
3023 * picked up later by the context switch code.
3025 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3027 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3028 * monitoring needs to be stopped.
3030 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3033 * update context state
3035 ctx->ctx_pmcs[cnum] = value;
3039 * write thread state
3041 if (is_system == 0) thread->pmcs[cnum] = value;
3044 * write hardware register if we can
3046 if (can_access_pmu) {
3047 ia64_set_pmc(cnum, value);
3052 * per-task SMP only here
3054 * we are guaranteed that the task is not running on the other CPU,
3055 * we indicate that this PMD will need to be reloaded if the task
3056 * is rescheduled on the CPU it ran last on.
3058 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3063 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",
3069 ctx->ctx_all_pmcs[0],
3070 ctx->ctx_used_pmds[0],
3071 ctx->ctx_pmds[cnum].eventid,
3074 ctx->ctx_reload_pmcs[0],
3075 ctx->ctx_used_monitors[0],
3076 ctx->ctx_ovfl_regs[0]));
3080 * make sure the changes are visible
3082 if (can_access_pmu) ia64_srlz_d();
3086 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3091 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3093 struct thread_struct *thread = NULL;
3094 struct task_struct *task;
3095 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3096 unsigned long value, hw_value, ovfl_mask;
3098 int i, can_access_pmu = 0, state;
3099 int is_counting, is_loaded, is_system, expert_mode;
3101 pfm_reg_check_t wr_func;
3104 state = ctx->ctx_state;
3105 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3106 is_system = ctx->ctx_fl_system;
3107 ovfl_mask = pmu_conf->ovfl_val;
3108 task = ctx->ctx_task;
3110 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3113 * on both UP and SMP, we can only write to the PMC when the task is
3114 * the owner of the local PMU.
3116 if (likely(is_loaded)) {
3117 thread = &task->thread;
3119 * In system wide and when the context is loaded, access can only happen
3120 * when the caller is running on the CPU being monitored by the session.
3121 * It does not have to be the owner (ctx_task) of the context per se.
3123 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3124 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3127 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3129 expert_mode = pfm_sysctl.expert_mode;
3131 for (i = 0; i < count; i++, req++) {
3133 cnum = req->reg_num;
3134 value = req->reg_value;
3136 if (!PMD_IS_IMPL(cnum)) {
3137 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3140 is_counting = PMD_IS_COUNTING(cnum);
3141 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3144 * execute write checker, if any
3146 if (unlikely(expert_mode == 0 && wr_func)) {
3147 unsigned long v = value;
3149 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3150 if (ret) goto abort_mission;
3157 * no error on this register
3159 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3162 * now commit changes to software state
3167 * update virtualized (64bits) counter
3171 * write context state
3173 ctx->ctx_pmds[cnum].lval = value;
3176 * when context is load we use the split value
3179 hw_value = value & ovfl_mask;
3180 value = value & ~ovfl_mask;
3184 * update reset values (not just for counters)
3186 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3187 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3190 * update randomization parameters (not just for counters)
3192 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3193 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3196 * update context value
3198 ctx->ctx_pmds[cnum].val = value;
3201 * Keep track of what we use
3203 * We do not keep track of PMC because we have to
3204 * systematically restore ALL of them.
3206 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3209 * mark this PMD register used as well
3211 CTX_USED_PMD(ctx, RDEP(cnum));
3214 * make sure we do not try to reset on
3215 * restart because we have established new values
3217 if (is_counting && state == PFM_CTX_MASKED) {
3218 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3223 * write thread state
3225 if (is_system == 0) thread->pmds[cnum] = hw_value;
3228 * write hardware register if we can
3230 if (can_access_pmu) {
3231 ia64_set_pmd(cnum, hw_value);
3235 * we are guaranteed that the task is not running on the other CPU,
3236 * we indicate that this PMD will need to be reloaded if the task
3237 * is rescheduled on the CPU it ran last on.
3239 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3244 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3245 "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",
3251 ctx->ctx_pmds[cnum].val,
3252 ctx->ctx_pmds[cnum].short_reset,
3253 ctx->ctx_pmds[cnum].long_reset,
3254 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3255 ctx->ctx_pmds[cnum].seed,
3256 ctx->ctx_pmds[cnum].mask,
3257 ctx->ctx_used_pmds[0],
3258 ctx->ctx_pmds[cnum].reset_pmds[0],
3259 ctx->ctx_reload_pmds[0],
3260 ctx->ctx_all_pmds[0],
3261 ctx->ctx_ovfl_regs[0]));
3265 * make changes visible
3267 if (can_access_pmu) ia64_srlz_d();
3273 * for now, we have only one possibility for error
3275 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3280 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3281 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3282 * interrupt is delivered during the call, it will be kept pending until we leave, making
3283 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3284 * guaranteed to return consistent data to the user, it may simply be old. It is not
3285 * trivial to treat the overflow while inside the call because you may end up in
3286 * some module sampling buffer code causing deadlocks.
3289 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3291 struct thread_struct *thread = NULL;
3292 struct task_struct *task;
3293 unsigned long val = 0UL, lval, ovfl_mask, sval;
3294 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3295 unsigned int cnum, reg_flags = 0;
3296 int i, can_access_pmu = 0, state;
3297 int is_loaded, is_system, is_counting, expert_mode;
3299 pfm_reg_check_t rd_func;
3302 * access is possible when loaded only for
3303 * self-monitoring tasks or in UP mode
3306 state = ctx->ctx_state;
3307 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3308 is_system = ctx->ctx_fl_system;
3309 ovfl_mask = pmu_conf->ovfl_val;
3310 task = ctx->ctx_task;
3312 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3314 if (likely(is_loaded)) {
3315 thread = &task->thread;
3317 * In system wide and when the context is loaded, access can only happen
3318 * when the caller is running on the CPU being monitored by the session.
3319 * It does not have to be the owner (ctx_task) of the context per se.
3321 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3322 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3326 * this can be true when not self-monitoring only in UP
3328 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3330 if (can_access_pmu) ia64_srlz_d();
3332 expert_mode = pfm_sysctl.expert_mode;
3334 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3340 * on both UP and SMP, we can only read the PMD from the hardware register when
3341 * the task is the owner of the local PMU.
3344 for (i = 0; i < count; i++, req++) {
3346 cnum = req->reg_num;
3347 reg_flags = req->reg_flags;
3349 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3351 * we can only read the register that we use. That includes
3352 * the one we explicitely initialize AND the one we want included
3353 * in the sampling buffer (smpl_regs).
3355 * Having this restriction allows optimization in the ctxsw routine
3356 * without compromising security (leaks)
3358 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3360 sval = ctx->ctx_pmds[cnum].val;
3361 lval = ctx->ctx_pmds[cnum].lval;
3362 is_counting = PMD_IS_COUNTING(cnum);
3365 * If the task is not the current one, then we check if the
3366 * PMU state is still in the local live register due to lazy ctxsw.
3367 * If true, then we read directly from the registers.
3369 if (can_access_pmu){
3370 val = ia64_get_pmd(cnum);
3373 * context has been saved
3374 * if context is zombie, then task does not exist anymore.
3375 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3377 val = is_loaded ? thread->pmds[cnum] : 0UL;
3379 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3383 * XXX: need to check for overflow when loaded
3390 * execute read checker, if any
3392 if (unlikely(expert_mode == 0 && rd_func)) {
3393 unsigned long v = val;
3394 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3395 if (ret) goto error;
3400 PFM_REG_RETFLAG_SET(reg_flags, 0);
3402 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3405 * update register return value, abort all if problem during copy.
3406 * we only modify the reg_flags field. no check mode is fine because
3407 * access has been verified upfront in sys_perfmonctl().
3409 req->reg_value = val;
3410 req->reg_flags = reg_flags;
3411 req->reg_last_reset_val = lval;
3417 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3422 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3426 if (req == NULL) return -EINVAL;
3428 ctx = GET_PMU_CTX();
3430 if (ctx == NULL) return -EINVAL;
3433 * for now limit to current task, which is enough when calling
3434 * from overflow handler
3436 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3438 return pfm_write_pmcs(ctx, req, nreq, regs);
3440 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3443 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3447 if (req == NULL) return -EINVAL;
3449 ctx = GET_PMU_CTX();
3451 if (ctx == NULL) return -EINVAL;
3454 * for now limit to current task, which is enough when calling
3455 * from overflow handler
3457 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3459 return pfm_read_pmds(ctx, req, nreq, regs);
3461 EXPORT_SYMBOL(pfm_mod_read_pmds);
3464 * Only call this function when a process it trying to
3465 * write the debug registers (reading is always allowed)
3468 pfm_use_debug_registers(struct task_struct *task)
3470 pfm_context_t *ctx = task->thread.pfm_context;
3471 unsigned long flags;
3474 if (pmu_conf->use_rr_dbregs == 0) return 0;
3476 DPRINT(("called for [%d]\n", task->pid));
3481 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3484 * Even on SMP, we do not need to use an atomic here because
3485 * the only way in is via ptrace() and this is possible only when the
3486 * process is stopped. Even in the case where the ctxsw out is not totally
3487 * completed by the time we come here, there is no way the 'stopped' process
3488 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3489 * So this is always safe.
3491 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3496 * We cannot allow setting breakpoints when system wide monitoring
3497 * sessions are using the debug registers.
3499 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3502 pfm_sessions.pfs_ptrace_use_dbregs++;
3504 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3505 pfm_sessions.pfs_ptrace_use_dbregs,
3506 pfm_sessions.pfs_sys_use_dbregs,
3515 * This function is called for every task that exits with the
3516 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3517 * able to use the debug registers for debugging purposes via
3518 * ptrace(). Therefore we know it was not using them for
3519 * perfmormance monitoring, so we only decrement the number
3520 * of "ptraced" debug register users to keep the count up to date
3523 pfm_release_debug_registers(struct task_struct *task)
3525 unsigned long flags;
3528 if (pmu_conf->use_rr_dbregs == 0) return 0;
3531 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3532 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3535 pfm_sessions.pfs_ptrace_use_dbregs--;
3544 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3546 struct task_struct *task;
3547 pfm_buffer_fmt_t *fmt;
3548 pfm_ovfl_ctrl_t rst_ctrl;
3549 int state, is_system;
3552 state = ctx->ctx_state;
3553 fmt = ctx->ctx_buf_fmt;
3554 is_system = ctx->ctx_fl_system;
3555 task = PFM_CTX_TASK(ctx);
3558 case PFM_CTX_MASKED:
3560 case PFM_CTX_LOADED:
3561 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3563 case PFM_CTX_UNLOADED:
3564 case PFM_CTX_ZOMBIE:
3565 DPRINT(("invalid state=%d\n", state));
3568 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3573 * In system wide and when the context is loaded, access can only happen
3574 * when the caller is running on the CPU being monitored by the session.
3575 * It does not have to be the owner (ctx_task) of the context per se.
3577 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3578 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3583 if (unlikely(task == NULL)) {
3584 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3588 if (task == current || is_system) {
3590 fmt = ctx->ctx_buf_fmt;
3592 DPRINT(("restarting self %d ovfl=0x%lx\n",
3594 ctx->ctx_ovfl_regs[0]));
3596 if (CTX_HAS_SMPL(ctx)) {
3598 prefetch(ctx->ctx_smpl_hdr);
3600 rst_ctrl.bits.mask_monitoring = 0;
3601 rst_ctrl.bits.reset_ovfl_pmds = 0;
3603 if (state == PFM_CTX_LOADED)
3604 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3606 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3608 rst_ctrl.bits.mask_monitoring = 0;
3609 rst_ctrl.bits.reset_ovfl_pmds = 1;
3613 if (rst_ctrl.bits.reset_ovfl_pmds)
3614 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3616 if (rst_ctrl.bits.mask_monitoring == 0) {
3617 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3619 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3621 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3623 // cannot use pfm_stop_monitoring(task, regs);
3627 * clear overflowed PMD mask to remove any stale information
3629 ctx->ctx_ovfl_regs[0] = 0UL;
3632 * back to LOADED state
3634 ctx->ctx_state = PFM_CTX_LOADED;
3637 * XXX: not really useful for self monitoring
3639 ctx->ctx_fl_can_restart = 0;
3645 * restart another task
3649 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3650 * one is seen by the task.
3652 if (state == PFM_CTX_MASKED) {
3653 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3655 * will prevent subsequent restart before this one is
3656 * seen by other task
3658 ctx->ctx_fl_can_restart = 0;
3662 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3663 * the task is blocked or on its way to block. That's the normal
3664 * restart path. If the monitoring is not masked, then the task
3665 * can be actively monitoring and we cannot directly intervene.
3666 * Therefore we use the trap mechanism to catch the task and
3667 * force it to reset the buffer/reset PMDs.
3669 * if non-blocking, then we ensure that the task will go into
3670 * pfm_handle_work() before returning to user mode.
3672 * We cannot explicitely reset another task, it MUST always
3673 * be done by the task itself. This works for system wide because
3674 * the tool that is controlling the session is logically doing
3675 * "self-monitoring".
3677 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3678 DPRINT(("unblocking [%d] \n", task->pid));
3679 up(&ctx->ctx_restart_sem);
3681 DPRINT(("[%d] armed exit trap\n", task->pid));
3683 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3685 PFM_SET_WORK_PENDING(task, 1);
3687 pfm_set_task_notify(task);
3690 * XXX: send reschedule if task runs on another CPU
3697 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3699 unsigned int m = *(unsigned int *)arg;
3701 pfm_sysctl.debug = m == 0 ? 0 : 1;
3703 pfm_debug_var = pfm_sysctl.debug;
3705 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3708 memset(pfm_stats, 0, sizeof(pfm_stats));
3709 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3715 * arg can be NULL and count can be zero for this function
3718 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3720 struct thread_struct *thread = NULL;
3721 struct task_struct *task;
3722 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3723 unsigned long flags;
3728 int i, can_access_pmu = 0;
3729 int is_system, is_loaded;
3731 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3733 state = ctx->ctx_state;
3734 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3735 is_system = ctx->ctx_fl_system;
3736 task = ctx->ctx_task;
3738 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3741 * on both UP and SMP, we can only write to the PMC when the task is
3742 * the owner of the local PMU.
3745 thread = &task->thread;
3747 * In system wide and when the context is loaded, access can only happen
3748 * when the caller is running on the CPU being monitored by the session.
3749 * It does not have to be the owner (ctx_task) of the context per se.
3751 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3752 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3755 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3759 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3760 * ensuring that no real breakpoint can be installed via this call.
3762 * IMPORTANT: regs can be NULL in this function
3765 first_time = ctx->ctx_fl_using_dbreg == 0;
3768 * don't bother if we are loaded and task is being debugged
3770 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3771 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3776 * check for debug registers in system wide mode
3778 * If though a check is done in pfm_context_load(),
3779 * we must repeat it here, in case the registers are
3780 * written after the context is loaded
3785 if (first_time && is_system) {
3786 if (pfm_sessions.pfs_ptrace_use_dbregs)
3789 pfm_sessions.pfs_sys_use_dbregs++;
3794 if (ret != 0) return ret;
3797 * mark ourself as user of the debug registers for
3800 ctx->ctx_fl_using_dbreg = 1;
3803 * clear hardware registers to make sure we don't
3804 * pick up stale state.
3806 * for a system wide session, we do not use
3807 * thread.dbr, thread.ibr because this process
3808 * never leaves the current CPU and the state
3809 * is shared by all processes running on it
3811 if (first_time && can_access_pmu) {
3812 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3813 for (i=0; i < pmu_conf->num_ibrs; i++) {
3814 ia64_set_ibr(i, 0UL);
3815 ia64_dv_serialize_instruction();
3818 for (i=0; i < pmu_conf->num_dbrs; i++) {
3819 ia64_set_dbr(i, 0UL);
3820 ia64_dv_serialize_data();
3826 * Now install the values into the registers
3828 for (i = 0; i < count; i++, req++) {
3830 rnum = req->dbreg_num;
3831 dbreg.val = req->dbreg_value;
3835 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3836 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3837 rnum, dbreg.val, mode, i, count));
3843 * make sure we do not install enabled breakpoint
3846 if (mode == PFM_CODE_RR)
3847 dbreg.ibr.ibr_x = 0;
3849 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3852 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3855 * Debug registers, just like PMC, can only be modified
3856 * by a kernel call. Moreover, perfmon() access to those
3857 * registers are centralized in this routine. The hardware
3858 * does not modify the value of these registers, therefore,
3859 * if we save them as they are written, we can avoid having
3860 * to save them on context switch out. This is made possible
3861 * by the fact that when perfmon uses debug registers, ptrace()
3862 * won't be able to modify them concurrently.
3864 if (mode == PFM_CODE_RR) {
3865 CTX_USED_IBR(ctx, rnum);
3867 if (can_access_pmu) {
3868 ia64_set_ibr(rnum, dbreg.val);
3869 ia64_dv_serialize_instruction();
3872 ctx->ctx_ibrs[rnum] = dbreg.val;
3874 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3875 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3877 CTX_USED_DBR(ctx, rnum);
3879 if (can_access_pmu) {
3880 ia64_set_dbr(rnum, dbreg.val);
3881 ia64_dv_serialize_data();
3883 ctx->ctx_dbrs[rnum] = dbreg.val;
3885 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3886 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3894 * in case it was our first attempt, we undo the global modifications
3898 if (ctx->ctx_fl_system) {
3899 pfm_sessions.pfs_sys_use_dbregs--;
3902 ctx->ctx_fl_using_dbreg = 0;
3905 * install error return flag
3907 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3913 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3915 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3919 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3921 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3925 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3929 if (req == NULL) return -EINVAL;
3931 ctx = GET_PMU_CTX();
3933 if (ctx == NULL) return -EINVAL;
3936 * for now limit to current task, which is enough when calling
3937 * from overflow handler
3939 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3941 return pfm_write_ibrs(ctx, req, nreq, regs);
3943 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3946 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3950 if (req == NULL) return -EINVAL;
3952 ctx = GET_PMU_CTX();
3954 if (ctx == NULL) return -EINVAL;
3957 * for now limit to current task, which is enough when calling
3958 * from overflow handler
3960 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3962 return pfm_write_dbrs(ctx, req, nreq, regs);
3964 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3968 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3970 pfarg_features_t *req = (pfarg_features_t *)arg;
3972 req->ft_version = PFM_VERSION;
3977 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3979 struct pt_regs *tregs;
3980 struct task_struct *task = PFM_CTX_TASK(ctx);
3981 int state, is_system;
3983 state = ctx->ctx_state;
3984 is_system = ctx->ctx_fl_system;
3987 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3989 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3992 * In system wide and when the context is loaded, access can only happen
3993 * when the caller is running on the CPU being monitored by the session.
3994 * It does not have to be the owner (ctx_task) of the context per se.
3996 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3997 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4000 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4001 PFM_CTX_TASK(ctx)->pid,
4005 * in system mode, we need to update the PMU directly
4006 * and the user level state of the caller, which may not
4007 * necessarily be the creator of the context.
4011 * Update local PMU first
4015 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4019 * update local cpuinfo
4021 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4024 * stop monitoring, does srlz.i
4029 * stop monitoring in the caller
4031 ia64_psr(regs)->pp = 0;
4039 if (task == current) {
4040 /* stop monitoring at kernel level */
4044 * stop monitoring at the user level
4046 ia64_psr(regs)->up = 0;
4048 tregs = ia64_task_regs(task);
4051 * stop monitoring at the user level
4053 ia64_psr(tregs)->up = 0;
4056 * monitoring disabled in kernel at next reschedule
4058 ctx->ctx_saved_psr_up = 0;
4059 DPRINT(("task=[%d]\n", task->pid));
4066 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4068 struct pt_regs *tregs;
4069 int state, is_system;
4071 state = ctx->ctx_state;
4072 is_system = ctx->ctx_fl_system;
4074 if (state != PFM_CTX_LOADED) return -EINVAL;
4077 * In system wide and when the context is loaded, access can only happen
4078 * when the caller is running on the CPU being monitored by the session.
4079 * It does not have to be the owner (ctx_task) of the context per se.
4081 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4082 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4087 * in system mode, we need to update the PMU directly
4088 * and the user level state of the caller, which may not
4089 * necessarily be the creator of the context.
4094 * set user level psr.pp for the caller
4096 ia64_psr(regs)->pp = 1;
4099 * now update the local PMU and cpuinfo
4101 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4104 * start monitoring at kernel level
4109 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4119 if (ctx->ctx_task == current) {
4121 /* start monitoring at kernel level */
4125 * activate monitoring at user level
4127 ia64_psr(regs)->up = 1;
4130 tregs = ia64_task_regs(ctx->ctx_task);
4133 * start monitoring at the kernel level the next
4134 * time the task is scheduled
4136 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4139 * activate monitoring at user level
4141 ia64_psr(tregs)->up = 1;
4147 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4149 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4154 for (i = 0; i < count; i++, req++) {
4156 cnum = req->reg_num;
4158 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4160 req->reg_value = PMC_DFL_VAL(cnum);
4162 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4164 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4169 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4174 pfm_check_task_exist(pfm_context_t *ctx)
4176 struct task_struct *g, *t;
4179 read_lock(&tasklist_lock);
4181 do_each_thread (g, t) {
4182 if (t->thread.pfm_context == ctx) {
4186 } while_each_thread (g, t);
4188 read_unlock(&tasklist_lock);
4190 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4196 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4198 struct task_struct *task;
4199 struct thread_struct *thread;
4200 struct pfm_context_t *old;
4201 unsigned long flags;
4203 struct task_struct *owner_task = NULL;
4205 pfarg_load_t *req = (pfarg_load_t *)arg;
4206 unsigned long *pmcs_source, *pmds_source;
4209 int state, is_system, set_dbregs = 0;
4211 state = ctx->ctx_state;
4212 is_system = ctx->ctx_fl_system;
4214 * can only load from unloaded or terminated state
4216 if (state != PFM_CTX_UNLOADED) {
4217 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4223 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4225 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4226 DPRINT(("cannot use blocking mode on self\n"));
4230 ret = pfm_get_task(ctx, req->load_pid, &task);
4232 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4239 * system wide is self monitoring only
4241 if (is_system && task != current) {
4242 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4247 thread = &task->thread;
4251 * cannot load a context which is using range restrictions,
4252 * into a task that is being debugged.
4254 if (ctx->ctx_fl_using_dbreg) {
4255 if (thread->flags & IA64_THREAD_DBG_VALID) {
4257 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4263 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4264 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4267 pfm_sessions.pfs_sys_use_dbregs++;
4268 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4275 if (ret) goto error;
4279 * SMP system-wide monitoring implies self-monitoring.
4281 * The programming model expects the task to
4282 * be pinned on a CPU throughout the session.
4283 * Here we take note of the current CPU at the
4284 * time the context is loaded. No call from
4285 * another CPU will be allowed.
4287 * The pinning via shed_setaffinity()
4288 * must be done by the calling task prior
4291 * systemwide: keep track of CPU this session is supposed to run on
4293 the_cpu = ctx->ctx_cpu = smp_processor_id();
4297 * now reserve the session
4299 ret = pfm_reserve_session(current, is_system, the_cpu);
4300 if (ret) goto error;
4303 * task is necessarily stopped at this point.
4305 * If the previous context was zombie, then it got removed in
4306 * pfm_save_regs(). Therefore we should not see it here.
4307 * If we see a context, then this is an active context
4309 * XXX: needs to be atomic
4311 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4312 thread->pfm_context, ctx));
4314 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4316 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4320 pfm_reset_msgq(ctx);
4322 ctx->ctx_state = PFM_CTX_LOADED;
4325 * link context to task
4327 ctx->ctx_task = task;
4331 * we load as stopped
4333 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4334 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4336 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4338 thread->flags |= IA64_THREAD_PM_VALID;
4342 * propagate into thread-state
4344 pfm_copy_pmds(task, ctx);
4345 pfm_copy_pmcs(task, ctx);
4347 pmcs_source = thread->pmcs;
4348 pmds_source = thread->pmds;
4351 * always the case for system-wide
4353 if (task == current) {
4355 if (is_system == 0) {
4357 /* allow user level control */
4358 ia64_psr(regs)->sp = 0;
4359 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4361 SET_LAST_CPU(ctx, smp_processor_id());
4363 SET_ACTIVATION(ctx);
4366 * push the other task out, if any
4368 owner_task = GET_PMU_OWNER();
4369 if (owner_task) pfm_lazy_save_regs(owner_task);
4373 * load all PMD from ctx to PMU (as opposed to thread state)
4374 * restore all PMC from ctx to PMU
4376 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4377 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4379 ctx->ctx_reload_pmcs[0] = 0UL;
4380 ctx->ctx_reload_pmds[0] = 0UL;
4383 * guaranteed safe by earlier check against DBG_VALID
4385 if (ctx->ctx_fl_using_dbreg) {
4386 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4387 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4392 SET_PMU_OWNER(task, ctx);
4394 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4397 * when not current, task MUST be stopped, so this is safe
4399 regs = ia64_task_regs(task);
4401 /* force a full reload */
4402 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4403 SET_LAST_CPU(ctx, -1);
4405 /* initial saved psr (stopped) */
4406 ctx->ctx_saved_psr_up = 0UL;
4407 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4413 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4416 * we must undo the dbregs setting (for system-wide)
4418 if (ret && set_dbregs) {
4420 pfm_sessions.pfs_sys_use_dbregs--;
4424 * release task, there is now a link with the context
4426 if (is_system == 0 && task != current) {
4430 ret = pfm_check_task_exist(ctx);
4432 ctx->ctx_state = PFM_CTX_UNLOADED;
4433 ctx->ctx_task = NULL;
4441 * in this function, we do not need to increase the use count
4442 * for the task via get_task_struct(), because we hold the
4443 * context lock. If the task were to disappear while having
4444 * a context attached, it would go through pfm_exit_thread()
4445 * which also grabs the context lock and would therefore be blocked
4446 * until we are here.
4448 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4451 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4453 struct task_struct *task = PFM_CTX_TASK(ctx);
4454 struct pt_regs *tregs;
4455 int prev_state, is_system;
4458 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4460 prev_state = ctx->ctx_state;
4461 is_system = ctx->ctx_fl_system;
4464 * unload only when necessary
4466 if (prev_state == PFM_CTX_UNLOADED) {
4467 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4472 * clear psr and dcr bits
4474 ret = pfm_stop(ctx, NULL, 0, regs);
4475 if (ret) return ret;
4477 ctx->ctx_state = PFM_CTX_UNLOADED;
4480 * in system mode, we need to update the PMU directly
4481 * and the user level state of the caller, which may not
4482 * necessarily be the creator of the context.
4489 * local PMU is taken care of in pfm_stop()
4491 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4492 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4495 * save PMDs in context
4498 pfm_flush_pmds(current, ctx);
4501 * at this point we are done with the PMU
4502 * so we can unreserve the resource.
4504 if (prev_state != PFM_CTX_ZOMBIE)
4505 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4508 * disconnect context from task
4510 task->thread.pfm_context = NULL;
4512 * disconnect task from context
4514 ctx->ctx_task = NULL;
4517 * There is nothing more to cleanup here.
4525 tregs = task == current ? regs : ia64_task_regs(task);
4527 if (task == current) {
4529 * cancel user level control
4531 ia64_psr(regs)->sp = 1;
4533 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4536 * save PMDs to context
4539 pfm_flush_pmds(task, ctx);
4542 * at this point we are done with the PMU
4543 * so we can unreserve the resource.
4545 * when state was ZOMBIE, we have already unreserved.
4547 if (prev_state != PFM_CTX_ZOMBIE)
4548 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4551 * reset activation counter and psr
4553 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4554 SET_LAST_CPU(ctx, -1);
4557 * PMU state will not be restored
4559 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4562 * break links between context and task
4564 task->thread.pfm_context = NULL;
4565 ctx->ctx_task = NULL;
4567 PFM_SET_WORK_PENDING(task, 0);
4569 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4570 ctx->ctx_fl_can_restart = 0;
4571 ctx->ctx_fl_going_zombie = 0;
4573 DPRINT(("disconnected [%d] from context\n", task->pid));
4580 * called only from exit_thread(): task == current
4581 * we come here only if current has a context attached (loaded or masked)
4584 pfm_exit_thread(struct task_struct *task)
4587 unsigned long flags;
4588 struct pt_regs *regs = ia64_task_regs(task);
4592 ctx = PFM_GET_CTX(task);
4594 PROTECT_CTX(ctx, flags);
4596 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4598 state = ctx->ctx_state;
4600 case PFM_CTX_UNLOADED:
4602 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4603 * be in unloaded state
4605 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4607 case PFM_CTX_LOADED:
4608 case PFM_CTX_MASKED:
4609 ret = pfm_context_unload(ctx, NULL, 0, regs);
4611 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4613 DPRINT(("ctx unloaded for current state was %d\n", state));
4615 pfm_end_notify_user(ctx);
4617 case PFM_CTX_ZOMBIE:
4618 ret = pfm_context_unload(ctx, NULL, 0, regs);
4620 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4625 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4628 UNPROTECT_CTX(ctx, flags);
4630 { u64 psr = pfm_get_psr();
4631 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4632 BUG_ON(GET_PMU_OWNER());
4633 BUG_ON(ia64_psr(regs)->up);
4634 BUG_ON(ia64_psr(regs)->pp);
4638 * All memory free operations (especially for vmalloc'ed memory)
4639 * MUST be done with interrupts ENABLED.
4641 if (free_ok) pfm_context_free(ctx);
4645 * functions MUST be listed in the increasing order of their index (see permfon.h)
4647 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4648 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4649 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4650 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4651 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4653 static pfm_cmd_desc_t pfm_cmd_tab[]={
4654 /* 0 */PFM_CMD_NONE,
4655 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4656 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4657 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4658 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4659 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4660 /* 6 */PFM_CMD_NONE,
4661 /* 7 */PFM_CMD_NONE,
4662 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4663 /* 9 */PFM_CMD_NONE,
4664 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4665 /* 11 */PFM_CMD_NONE,
4666 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4667 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4668 /* 14 */PFM_CMD_NONE,
4669 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4670 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4671 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4672 /* 18 */PFM_CMD_NONE,
4673 /* 19 */PFM_CMD_NONE,
4674 /* 20 */PFM_CMD_NONE,
4675 /* 21 */PFM_CMD_NONE,
4676 /* 22 */PFM_CMD_NONE,
4677 /* 23 */PFM_CMD_NONE,
4678 /* 24 */PFM_CMD_NONE,
4679 /* 25 */PFM_CMD_NONE,
4680 /* 26 */PFM_CMD_NONE,
4681 /* 27 */PFM_CMD_NONE,
4682 /* 28 */PFM_CMD_NONE,
4683 /* 29 */PFM_CMD_NONE,
4684 /* 30 */PFM_CMD_NONE,
4685 /* 31 */PFM_CMD_NONE,
4686 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4687 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4689 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4692 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4694 struct task_struct *task;
4695 int state, old_state;
4698 state = ctx->ctx_state;
4699 task = ctx->ctx_task;
4702 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4706 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4710 task->state, PFM_CMD_STOPPED(cmd)));
4713 * self-monitoring always ok.
4715 * for system-wide the caller can either be the creator of the
4716 * context (to one to which the context is attached to) OR
4717 * a task running on the same CPU as the session.
4719 if (task == current || ctx->ctx_fl_system) return 0;
4722 * if context is UNLOADED we are safe to go
4724 if (state == PFM_CTX_UNLOADED) return 0;
4727 * no command can operate on a zombie context
4729 if (state == PFM_CTX_ZOMBIE) {
4730 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4735 * context is LOADED or MASKED. Some commands may need to have
4738 * We could lift this restriction for UP but it would mean that
4739 * the user has no guarantee the task would not run between
4740 * two successive calls to perfmonctl(). That's probably OK.
4741 * If this user wants to ensure the task does not run, then
4742 * the task must be stopped.
4744 if (PFM_CMD_STOPPED(cmd)) {
4745 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
4746 DPRINT(("[%d] task not in stopped state\n", task->pid));
4750 * task is now stopped, wait for ctxsw out
4752 * This is an interesting point in the code.
4753 * We need to unprotect the context because
4754 * the pfm_save_regs() routines needs to grab
4755 * the same lock. There are danger in doing
4756 * this because it leaves a window open for
4757 * another task to get access to the context
4758 * and possibly change its state. The one thing
4759 * that is not possible is for the context to disappear
4760 * because we are protected by the VFS layer, i.e.,
4761 * get_fd()/put_fd().
4765 UNPROTECT_CTX(ctx, flags);
4767 wait_task_inactive(task);
4769 PROTECT_CTX(ctx, flags);
4772 * we must recheck to verify if state has changed
4774 if (ctx->ctx_state != old_state) {
4775 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4783 * system-call entry point (must return long)
4786 sys_perfmonctl (int fd, int cmd, void __user *arg, int count, long arg5, long arg6, long arg7,
4787 long arg8, long stack)
4789 struct pt_regs *regs = (struct pt_regs *)&stack;
4790 struct file *file = NULL;
4791 pfm_context_t *ctx = NULL;
4792 unsigned long flags = 0UL;
4793 void *args_k = NULL;
4794 long ret; /* will expand int return types */
4795 size_t base_sz, sz, xtra_sz = 0;
4796 int narg, completed_args = 0, call_made = 0, cmd_flags;
4797 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4798 int (*getsize)(void *arg, size_t *sz);
4799 #define PFM_MAX_ARGSIZE 4096
4802 * reject any call if perfmon was disabled at initialization
4804 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4806 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4807 DPRINT(("invalid cmd=%d\n", cmd));
4811 func = pfm_cmd_tab[cmd].cmd_func;
4812 narg = pfm_cmd_tab[cmd].cmd_narg;
4813 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4814 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4815 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4817 if (unlikely(func == NULL)) {
4818 DPRINT(("invalid cmd=%d\n", cmd));
4822 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4830 * check if number of arguments matches what the command expects
4832 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4836 sz = xtra_sz + base_sz*count;
4838 * limit abuse to min page size
4840 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4841 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4846 * allocate default-sized argument buffer
4848 if (likely(count && args_k == NULL)) {
4849 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4850 if (args_k == NULL) return -ENOMEM;
4858 * assume sz = 0 for command without parameters
4860 if (sz && copy_from_user(args_k, arg, sz)) {
4861 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4866 * check if command supports extra parameters
4868 if (completed_args == 0 && getsize) {
4870 * get extra parameters size (based on main argument)
4872 ret = (*getsize)(args_k, &xtra_sz);
4873 if (ret) goto error_args;
4877 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4879 /* retry if necessary */
4880 if (likely(xtra_sz)) goto restart_args;
4883 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4888 if (unlikely(file == NULL)) {
4889 DPRINT(("invalid fd %d\n", fd));
4892 if (unlikely(PFM_IS_FILE(file) == 0)) {
4893 DPRINT(("fd %d not related to perfmon\n", fd));
4897 ctx = (pfm_context_t *)file->private_data;
4898 if (unlikely(ctx == NULL)) {
4899 DPRINT(("no context for fd %d\n", fd));
4902 prefetch(&ctx->ctx_state);
4904 PROTECT_CTX(ctx, flags);
4907 * check task is stopped
4909 ret = pfm_check_task_state(ctx, cmd, flags);
4910 if (unlikely(ret)) goto abort_locked;
4913 ret = (*func)(ctx, args_k, count, regs);
4919 DPRINT(("context unlocked\n"));
4920 UNPROTECT_CTX(ctx, flags);
4924 /* copy argument back to user, if needed */
4925 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4928 if (args_k) kfree(args_k);
4930 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4936 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4938 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4939 pfm_ovfl_ctrl_t rst_ctrl;
4943 state = ctx->ctx_state;
4945 * Unlock sampling buffer and reset index atomically
4946 * XXX: not really needed when blocking
4948 if (CTX_HAS_SMPL(ctx)) {
4950 rst_ctrl.bits.mask_monitoring = 0;
4951 rst_ctrl.bits.reset_ovfl_pmds = 0;
4953 if (state == PFM_CTX_LOADED)
4954 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4956 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4958 rst_ctrl.bits.mask_monitoring = 0;
4959 rst_ctrl.bits.reset_ovfl_pmds = 1;
4963 if (rst_ctrl.bits.reset_ovfl_pmds) {
4964 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4966 if (rst_ctrl.bits.mask_monitoring == 0) {
4967 DPRINT(("resuming monitoring\n"));
4968 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4970 DPRINT(("stopping monitoring\n"));
4971 //pfm_stop_monitoring(current, regs);
4973 ctx->ctx_state = PFM_CTX_LOADED;
4978 * context MUST BE LOCKED when calling
4979 * can only be called for current
4982 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4986 DPRINT(("entering for [%d]\n", current->pid));
4988 ret = pfm_context_unload(ctx, NULL, 0, regs);
4990 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", current->pid, ret);
4994 * and wakeup controlling task, indicating we are now disconnected
4996 wake_up_interruptible(&ctx->ctx_zombieq);
4999 * given that context is still locked, the controlling
5000 * task will only get access when we return from
5001 * pfm_handle_work().
5005 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5008 pfm_handle_work(void)
5011 struct pt_regs *regs;
5012 unsigned long flags;
5013 unsigned long ovfl_regs;
5014 unsigned int reason;
5017 ctx = PFM_GET_CTX(current);
5019 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5023 PROTECT_CTX(ctx, flags);
5025 PFM_SET_WORK_PENDING(current, 0);
5027 pfm_clear_task_notify();
5029 regs = ia64_task_regs(current);
5032 * extract reason for being here and clear
5034 reason = ctx->ctx_fl_trap_reason;
5035 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5036 ovfl_regs = ctx->ctx_ovfl_regs[0];
5038 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5041 * must be done before we check for simple-reset mode
5043 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5046 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5047 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5049 UNPROTECT_CTX(ctx, flags);
5052 * pfm_handle_work() is currently called with interrupts disabled.
5053 * The down_interruptible call may sleep, therefore we
5054 * must re-enable interrupts to avoid deadlocks. It is
5055 * safe to do so because this function is called ONLY
5056 * when returning to user level (PUStk=1), in which case
5057 * there is no risk of kernel stack overflow due to deep
5058 * interrupt nesting.
5060 BUG_ON(flags & IA64_PSR_I);
5063 DPRINT(("before block sleeping\n"));
5066 * may go through without blocking on SMP systems
5067 * if restart has been received already by the time we call down()
5069 ret = down_interruptible(&ctx->ctx_restart_sem);
5071 DPRINT(("after block sleeping ret=%d\n", ret));
5074 * disable interrupts to restore state we had upon entering
5077 local_irq_disable();
5079 PROTECT_CTX(ctx, flags);
5082 * we need to read the ovfl_regs only after wake-up
5083 * because we may have had pfm_write_pmds() in between
5084 * and that can changed PMD values and therefore
5085 * ovfl_regs is reset for these new PMD values.
5087 ovfl_regs = ctx->ctx_ovfl_regs[0];
5089 if (ctx->ctx_fl_going_zombie) {
5091 DPRINT(("context is zombie, bailing out\n"));
5092 pfm_context_force_terminate(ctx, regs);
5096 * in case of interruption of down() we don't restart anything
5098 if (ret < 0) goto nothing_to_do;
5101 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5102 ctx->ctx_ovfl_regs[0] = 0UL;
5106 UNPROTECT_CTX(ctx, flags);
5110 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5112 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5113 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5117 DPRINT(("waking up somebody\n"));
5119 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5122 * safe, we are not in intr handler, nor in ctxsw when
5125 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5131 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5133 pfm_msg_t *msg = NULL;
5135 if (ctx->ctx_fl_no_msg == 0) {
5136 msg = pfm_get_new_msg(ctx);
5138 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5142 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5143 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5144 msg->pfm_ovfl_msg.msg_active_set = 0;
5145 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5146 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5147 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5148 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5149 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5152 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5158 return pfm_notify_user(ctx, msg);
5162 pfm_end_notify_user(pfm_context_t *ctx)
5166 msg = pfm_get_new_msg(ctx);
5168 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5172 memset(msg, 0, sizeof(*msg));
5174 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5175 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5176 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5178 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5183 return pfm_notify_user(ctx, msg);
5187 * main overflow processing routine.
5188 * it can be called from the interrupt path or explicitely during the context switch code
5191 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5193 pfm_ovfl_arg_t *ovfl_arg;
5195 unsigned long old_val, ovfl_val, new_val;
5196 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5197 unsigned long tstamp;
5198 pfm_ovfl_ctrl_t ovfl_ctrl;
5199 unsigned int i, has_smpl;
5200 int must_notify = 0;
5202 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5205 * sanity test. Should never happen
5207 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5209 tstamp = ia64_get_itc();
5210 mask = pmc0 >> PMU_FIRST_COUNTER;
5211 ovfl_val = pmu_conf->ovfl_val;
5212 has_smpl = CTX_HAS_SMPL(ctx);
5214 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5215 "used_pmds=0x%lx\n",
5217 task ? task->pid: -1,
5218 (regs ? regs->cr_iip : 0),
5219 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5220 ctx->ctx_used_pmds[0]));
5224 * first we update the virtual counters
5225 * assume there was a prior ia64_srlz_d() issued
5227 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5229 /* skip pmd which did not overflow */
5230 if ((mask & 0x1) == 0) continue;
5233 * Note that the pmd is not necessarily 0 at this point as qualified events
5234 * may have happened before the PMU was frozen. The residual count is not
5235 * taken into consideration here but will be with any read of the pmd via
5238 old_val = new_val = ctx->ctx_pmds[i].val;
5239 new_val += 1 + ovfl_val;
5240 ctx->ctx_pmds[i].val = new_val;
5243 * check for overflow condition
5245 if (likely(old_val > new_val)) {
5246 ovfl_pmds |= 1UL << i;
5247 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5250 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5254 ia64_get_pmd(i) & ovfl_val,
5260 * there was no 64-bit overflow, nothing else to do
5262 if (ovfl_pmds == 0UL) return;
5265 * reset all control bits
5271 * if a sampling format module exists, then we "cache" the overflow by
5272 * calling the module's handler() routine.
5275 unsigned long start_cycles, end_cycles;
5276 unsigned long pmd_mask;
5278 int this_cpu = smp_processor_id();
5280 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5281 ovfl_arg = &ctx->ctx_ovfl_arg;
5283 prefetch(ctx->ctx_smpl_hdr);
5285 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5289 if ((pmd_mask & 0x1) == 0) continue;
5291 ovfl_arg->ovfl_pmd = (unsigned char )i;
5292 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5293 ovfl_arg->active_set = 0;
5294 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5295 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5297 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5298 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5299 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5302 * copy values of pmds of interest. Sampling format may copy them
5303 * into sampling buffer.
5306 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5307 if ((smpl_pmds & 0x1) == 0) continue;
5308 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5309 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5313 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5315 start_cycles = ia64_get_itc();
5318 * call custom buffer format record (handler) routine
5320 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5322 end_cycles = ia64_get_itc();
5325 * For those controls, we take the union because they have
5326 * an all or nothing behavior.
5328 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5329 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5330 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5332 * build the bitmask of pmds to reset now
5334 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5336 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5339 * when the module cannot handle the rest of the overflows, we abort right here
5341 if (ret && pmd_mask) {
5342 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5343 pmd_mask<<PMU_FIRST_COUNTER));
5346 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5348 ovfl_pmds &= ~reset_pmds;
5351 * when no sampling module is used, then the default
5352 * is to notify on overflow if requested by user
5354 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5355 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5356 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5357 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5359 * if needed, we reset all overflowed pmds
5361 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5364 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5367 * reset the requested PMD registers using the short reset values
5370 unsigned long bm = reset_pmds;
5371 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5374 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5376 * keep track of what to reset when unblocking
5378 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5381 * check for blocking context
5383 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5385 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5388 * set the perfmon specific checking pending work for the task
5390 PFM_SET_WORK_PENDING(task, 1);
5393 * when coming from ctxsw, current still points to the
5394 * previous task, therefore we must work with task and not current.
5396 pfm_set_task_notify(task);
5399 * defer until state is changed (shorten spin window). the context is locked
5400 * anyway, so the signal receiver would come spin for nothing.
5405 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5406 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5407 PFM_GET_WORK_PENDING(task),
5408 ctx->ctx_fl_trap_reason,
5411 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5413 * in case monitoring must be stopped, we toggle the psr bits
5415 if (ovfl_ctrl.bits.mask_monitoring) {
5416 pfm_mask_monitoring(task);
5417 ctx->ctx_state = PFM_CTX_MASKED;
5418 ctx->ctx_fl_can_restart = 1;
5422 * send notification now
5424 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5429 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5431 task ? task->pid : -1,
5437 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5438 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5439 * come here as zombie only if the task is the current task. In which case, we
5440 * can access the PMU hardware directly.
5442 * Note that zombies do have PM_VALID set. So here we do the minimal.
5444 * In case the context was zombified it could not be reclaimed at the time
5445 * the monitoring program exited. At this point, the PMU reservation has been
5446 * returned, the sampiing buffer has been freed. We must convert this call
5447 * into a spurious interrupt. However, we must also avoid infinite overflows
5448 * by stopping monitoring for this task. We can only come here for a per-task
5449 * context. All we need to do is to stop monitoring using the psr bits which
5450 * are always task private. By re-enabling secure montioring, we ensure that
5451 * the monitored task will not be able to re-activate monitoring.
5452 * The task will eventually be context switched out, at which point the context
5453 * will be reclaimed (that includes releasing ownership of the PMU).
5455 * So there might be a window of time where the number of per-task session is zero
5456 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5457 * context. This is safe because if a per-task session comes in, it will push this one
5458 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5459 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5460 * also push our zombie context out.
5462 * Overall pretty hairy stuff....
5464 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5466 ia64_psr(regs)->up = 0;
5467 ia64_psr(regs)->sp = 1;
5472 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5474 struct task_struct *task;
5476 unsigned long flags;
5478 int this_cpu = smp_processor_id();
5481 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5484 * srlz.d done before arriving here
5486 pmc0 = ia64_get_pmc(0);
5488 task = GET_PMU_OWNER();
5489 ctx = GET_PMU_CTX();
5492 * if we have some pending bits set
5493 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5495 if (PMC0_HAS_OVFL(pmc0) && task) {
5497 * we assume that pmc0.fr is always set here
5501 if (!ctx) goto report_spurious1;
5503 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5504 goto report_spurious2;
5506 PROTECT_CTX_NOPRINT(ctx, flags);
5508 pfm_overflow_handler(task, ctx, pmc0, regs);
5510 UNPROTECT_CTX_NOPRINT(ctx, flags);
5513 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5517 * keep it unfrozen at all times
5524 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5525 this_cpu, task->pid);
5529 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5537 pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5539 unsigned long start_cycles, total_cycles;
5540 unsigned long min, max;
5544 this_cpu = get_cpu();
5545 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5546 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5548 start_cycles = ia64_get_itc();
5550 ret = pfm_do_interrupt_handler(irq, arg, regs);
5552 total_cycles = ia64_get_itc();
5555 * don't measure spurious interrupts
5557 if (likely(ret == 0)) {
5558 total_cycles -= start_cycles;
5560 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5561 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5563 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5565 put_cpu_no_resched();
5570 * /proc/perfmon interface, for debug only
5573 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5576 pfm_proc_start(struct seq_file *m, loff_t *pos)
5579 return PFM_PROC_SHOW_HEADER;
5582 while (*pos <= NR_CPUS) {
5583 if (cpu_online(*pos - 1)) {
5584 return (void *)*pos;
5592 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5595 return pfm_proc_start(m, pos);
5599 pfm_proc_stop(struct seq_file *m, void *v)
5604 pfm_proc_show_header(struct seq_file *m)
5606 struct list_head * pos;
5607 pfm_buffer_fmt_t * entry;
5608 unsigned long flags;
5611 "perfmon version : %u.%u\n"
5614 "expert mode : %s\n"
5615 "ovfl_mask : 0x%lx\n"
5616 "PMU flags : 0x%x\n",
5617 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5619 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5620 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5627 "proc_sessions : %u\n"
5628 "sys_sessions : %u\n"
5629 "sys_use_dbregs : %u\n"
5630 "ptrace_use_dbregs : %u\n",
5631 pfm_sessions.pfs_task_sessions,
5632 pfm_sessions.pfs_sys_sessions,
5633 pfm_sessions.pfs_sys_use_dbregs,
5634 pfm_sessions.pfs_ptrace_use_dbregs);
5638 spin_lock(&pfm_buffer_fmt_lock);
5640 list_for_each(pos, &pfm_buffer_fmt_list) {
5641 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5642 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5653 entry->fmt_uuid[10],
5654 entry->fmt_uuid[11],
5655 entry->fmt_uuid[12],
5656 entry->fmt_uuid[13],
5657 entry->fmt_uuid[14],
5658 entry->fmt_uuid[15],
5661 spin_unlock(&pfm_buffer_fmt_lock);
5666 pfm_proc_show(struct seq_file *m, void *v)
5672 if (v == PFM_PROC_SHOW_HEADER) {
5673 pfm_proc_show_header(m);
5677 /* show info for CPU (v - 1) */
5681 "CPU%-2d overflow intrs : %lu\n"
5682 "CPU%-2d overflow cycles : %lu\n"
5683 "CPU%-2d overflow min : %lu\n"
5684 "CPU%-2d overflow max : %lu\n"
5685 "CPU%-2d smpl handler calls : %lu\n"
5686 "CPU%-2d smpl handler cycles : %lu\n"
5687 "CPU%-2d spurious intrs : %lu\n"
5688 "CPU%-2d replay intrs : %lu\n"
5689 "CPU%-2d syst_wide : %d\n"
5690 "CPU%-2d dcr_pp : %d\n"
5691 "CPU%-2d exclude idle : %d\n"
5692 "CPU%-2d owner : %d\n"
5693 "CPU%-2d context : %p\n"
5694 "CPU%-2d activations : %lu\n",
5695 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5696 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5697 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5698 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5699 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5700 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5701 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5702 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5703 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5704 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5705 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5706 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5707 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5708 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5710 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5712 psr = pfm_get_psr();
5717 "CPU%-2d psr : 0x%lx\n"
5718 "CPU%-2d pmc0 : 0x%lx\n",
5720 cpu, ia64_get_pmc(0));
5722 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5723 if (PMC_IS_COUNTING(i) == 0) continue;
5725 "CPU%-2d pmc%u : 0x%lx\n"
5726 "CPU%-2d pmd%u : 0x%lx\n",
5727 cpu, i, ia64_get_pmc(i),
5728 cpu, i, ia64_get_pmd(i));
5734 struct seq_operations pfm_seq_ops = {
5735 .start = pfm_proc_start,
5736 .next = pfm_proc_next,
5737 .stop = pfm_proc_stop,
5738 .show = pfm_proc_show
5742 pfm_proc_open(struct inode *inode, struct file *file)
5744 return seq_open(file, &pfm_seq_ops);
5749 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5750 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5751 * is active or inactive based on mode. We must rely on the value in
5752 * local_cpu_data->pfm_syst_info
5755 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5757 struct pt_regs *regs;
5759 unsigned long dcr_pp;
5761 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5764 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5765 * on every CPU, so we can rely on the pid to identify the idle task.
5767 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5768 regs = ia64_task_regs(task);
5769 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5773 * if monitoring has started
5776 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5778 * context switching in?
5781 /* mask monitoring for the idle task */
5782 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5788 * context switching out
5789 * restore monitoring for next task
5791 * Due to inlining this odd if-then-else construction generates
5794 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5803 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5805 struct task_struct *task = ctx->ctx_task;
5807 ia64_psr(regs)->up = 0;
5808 ia64_psr(regs)->sp = 1;
5810 if (GET_PMU_OWNER() == task) {
5811 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5812 SET_PMU_OWNER(NULL, NULL);
5816 * disconnect the task from the context and vice-versa
5818 PFM_SET_WORK_PENDING(task, 0);
5820 task->thread.pfm_context = NULL;
5821 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5823 DPRINT(("force cleanup for [%d]\n", task->pid));
5828 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5831 pfm_save_regs(struct task_struct *task)
5834 struct thread_struct *t;
5835 unsigned long flags;
5839 ctx = PFM_GET_CTX(task);
5840 if (ctx == NULL) return;
5844 * we always come here with interrupts ALREADY disabled by
5845 * the scheduler. So we simply need to protect against concurrent
5846 * access, not CPU concurrency.
5848 flags = pfm_protect_ctx_ctxsw(ctx);
5850 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5851 struct pt_regs *regs = ia64_task_regs(task);
5855 pfm_force_cleanup(ctx, regs);
5857 BUG_ON(ctx->ctx_smpl_hdr);
5859 pfm_unprotect_ctx_ctxsw(ctx, flags);
5861 pfm_context_free(ctx);
5866 * save current PSR: needed because we modify it
5869 psr = pfm_get_psr();
5871 BUG_ON(psr & (IA64_PSR_I));
5875 * This is the last instruction which may generate an overflow
5877 * We do not need to set psr.sp because, it is irrelevant in kernel.
5878 * It will be restored from ipsr when going back to user level
5883 * keep a copy of psr.up (for reload)
5885 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5888 * release ownership of this PMU.
5889 * PM interrupts are masked, so nothing
5892 SET_PMU_OWNER(NULL, NULL);
5895 * we systematically save the PMD as we have no
5896 * guarantee we will be schedule at that same
5899 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5902 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5903 * we will need it on the restore path to check
5904 * for pending overflow.
5906 t->pmcs[0] = ia64_get_pmc(0);
5909 * unfreeze PMU if had pending overflows
5911 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5914 * finally, allow context access.
5915 * interrupts will still be masked after this call.
5917 pfm_unprotect_ctx_ctxsw(ctx, flags);
5920 #else /* !CONFIG_SMP */
5922 pfm_save_regs(struct task_struct *task)
5927 ctx = PFM_GET_CTX(task);
5928 if (ctx == NULL) return;
5931 * save current PSR: needed because we modify it
5933 psr = pfm_get_psr();
5935 BUG_ON(psr & (IA64_PSR_I));
5939 * This is the last instruction which may generate an overflow
5941 * We do not need to set psr.sp because, it is irrelevant in kernel.
5942 * It will be restored from ipsr when going back to user level
5947 * keep a copy of psr.up (for reload)
5949 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5953 pfm_lazy_save_regs (struct task_struct *task)
5956 struct thread_struct *t;
5957 unsigned long flags;
5959 { u64 psr = pfm_get_psr();
5960 BUG_ON(psr & IA64_PSR_UP);
5963 ctx = PFM_GET_CTX(task);
5967 * we need to mask PMU overflow here to
5968 * make sure that we maintain pmc0 until
5969 * we save it. overflow interrupts are
5970 * treated as spurious if there is no
5973 * XXX: I don't think this is necessary
5975 PROTECT_CTX(ctx,flags);
5978 * release ownership of this PMU.
5979 * must be done before we save the registers.
5981 * after this call any PMU interrupt is treated
5984 SET_PMU_OWNER(NULL, NULL);
5987 * save all the pmds we use
5989 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5992 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5993 * it is needed to check for pended overflow
5994 * on the restore path
5996 t->pmcs[0] = ia64_get_pmc(0);
5999 * unfreeze PMU if had pending overflows
6001 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6004 * now get can unmask PMU interrupts, they will
6005 * be treated as purely spurious and we will not
6006 * lose any information
6008 UNPROTECT_CTX(ctx,flags);
6010 #endif /* CONFIG_SMP */
6014 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6017 pfm_load_regs (struct task_struct *task)
6020 struct thread_struct *t;
6021 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6022 unsigned long flags;
6024 int need_irq_resend;
6026 ctx = PFM_GET_CTX(task);
6027 if (unlikely(ctx == NULL)) return;
6029 BUG_ON(GET_PMU_OWNER());
6033 * possible on unload
6035 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6038 * we always come here with interrupts ALREADY disabled by
6039 * the scheduler. So we simply need to protect against concurrent
6040 * access, not CPU concurrency.
6042 flags = pfm_protect_ctx_ctxsw(ctx);
6043 psr = pfm_get_psr();
6045 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6047 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6048 BUG_ON(psr & IA64_PSR_I);
6050 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6051 struct pt_regs *regs = ia64_task_regs(task);
6053 BUG_ON(ctx->ctx_smpl_hdr);
6055 pfm_force_cleanup(ctx, regs);
6057 pfm_unprotect_ctx_ctxsw(ctx, flags);
6060 * this one (kmalloc'ed) is fine with interrupts disabled
6062 pfm_context_free(ctx);
6068 * we restore ALL the debug registers to avoid picking up
6071 if (ctx->ctx_fl_using_dbreg) {
6072 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6073 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6076 * retrieve saved psr.up
6078 psr_up = ctx->ctx_saved_psr_up;
6081 * if we were the last user of the PMU on that CPU,
6082 * then nothing to do except restore psr
6084 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6087 * retrieve partial reload masks (due to user modifications)
6089 pmc_mask = ctx->ctx_reload_pmcs[0];
6090 pmd_mask = ctx->ctx_reload_pmds[0];
6094 * To avoid leaking information to the user level when psr.sp=0,
6095 * we must reload ALL implemented pmds (even the ones we don't use).
6096 * In the kernel we only allow PFM_READ_PMDS on registers which
6097 * we initialized or requested (sampling) so there is no risk there.
6099 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6102 * ALL accessible PMCs are systematically reloaded, unused registers
6103 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6104 * up stale configuration.
6106 * PMC0 is never in the mask. It is always restored separately.
6108 pmc_mask = ctx->ctx_all_pmcs[0];
6111 * when context is MASKED, we will restore PMC with plm=0
6112 * and PMD with stale information, but that's ok, nothing
6115 * XXX: optimize here
6117 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6118 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6121 * check for pending overflow at the time the state
6124 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6126 * reload pmc0 with the overflow information
6127 * On McKinley PMU, this will trigger a PMU interrupt
6129 ia64_set_pmc(0, t->pmcs[0]);
6134 * will replay the PMU interrupt
6136 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6138 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6142 * we just did a reload, so we reset the partial reload fields
6144 ctx->ctx_reload_pmcs[0] = 0UL;
6145 ctx->ctx_reload_pmds[0] = 0UL;
6147 SET_LAST_CPU(ctx, smp_processor_id());
6150 * dump activation value for this PMU
6154 * record current activation for this context
6156 SET_ACTIVATION(ctx);
6159 * establish new ownership.
6161 SET_PMU_OWNER(task, ctx);
6164 * restore the psr.up bit. measurement
6166 * no PMU interrupt can happen at this point
6167 * because we still have interrupts disabled.
6169 if (likely(psr_up)) pfm_set_psr_up();
6172 * allow concurrent access to context
6174 pfm_unprotect_ctx_ctxsw(ctx, flags);
6176 #else /* !CONFIG_SMP */
6178 * reload PMU state for UP kernels
6179 * in 2.5 we come here with interrupts disabled
6182 pfm_load_regs (struct task_struct *task)
6184 struct thread_struct *t;
6186 struct task_struct *owner;
6187 unsigned long pmd_mask, pmc_mask;
6189 int need_irq_resend;
6191 owner = GET_PMU_OWNER();
6192 ctx = PFM_GET_CTX(task);
6194 psr = pfm_get_psr();
6196 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6197 BUG_ON(psr & IA64_PSR_I);
6200 * we restore ALL the debug registers to avoid picking up
6203 * This must be done even when the task is still the owner
6204 * as the registers may have been modified via ptrace()
6205 * (not perfmon) by the previous task.
6207 if (ctx->ctx_fl_using_dbreg) {
6208 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6209 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6213 * retrieved saved psr.up
6215 psr_up = ctx->ctx_saved_psr_up;
6216 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6219 * short path, our state is still there, just
6220 * need to restore psr and we go
6222 * we do not touch either PMC nor PMD. the psr is not touched
6223 * by the overflow_handler. So we are safe w.r.t. to interrupt
6224 * concurrency even without interrupt masking.
6226 if (likely(owner == task)) {
6227 if (likely(psr_up)) pfm_set_psr_up();
6232 * someone else is still using the PMU, first push it out and
6233 * then we'll be able to install our stuff !
6235 * Upon return, there will be no owner for the current PMU
6237 if (owner) pfm_lazy_save_regs(owner);
6240 * To avoid leaking information to the user level when psr.sp=0,
6241 * we must reload ALL implemented pmds (even the ones we don't use).
6242 * In the kernel we only allow PFM_READ_PMDS on registers which
6243 * we initialized or requested (sampling) so there is no risk there.
6245 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6248 * ALL accessible PMCs are systematically reloaded, unused registers
6249 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6250 * up stale configuration.
6252 * PMC0 is never in the mask. It is always restored separately
6254 pmc_mask = ctx->ctx_all_pmcs[0];
6256 pfm_restore_pmds(t->pmds, pmd_mask);
6257 pfm_restore_pmcs(t->pmcs, pmc_mask);
6260 * check for pending overflow at the time the state
6263 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6265 * reload pmc0 with the overflow information
6266 * On McKinley PMU, this will trigger a PMU interrupt
6268 ia64_set_pmc(0, t->pmcs[0]);
6274 * will replay the PMU interrupt
6276 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6278 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6282 * establish new ownership.
6284 SET_PMU_OWNER(task, ctx);
6287 * restore the psr.up bit. measurement
6289 * no PMU interrupt can happen at this point
6290 * because we still have interrupts disabled.
6292 if (likely(psr_up)) pfm_set_psr_up();
6294 #endif /* CONFIG_SMP */
6297 * this function assumes monitoring is stopped
6300 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6303 unsigned long mask2, val, pmd_val, ovfl_val;
6304 int i, can_access_pmu = 0;
6308 * is the caller the task being monitored (or which initiated the
6309 * session for system wide measurements)
6311 is_self = ctx->ctx_task == task ? 1 : 0;
6314 * can access PMU is task is the owner of the PMU state on the current CPU
6315 * or if we are running on the CPU bound to the context in system-wide mode
6316 * (that is not necessarily the task the context is attached to in this mode).
6317 * In system-wide we always have can_access_pmu true because a task running on an
6318 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6320 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6321 if (can_access_pmu) {
6323 * Mark the PMU as not owned
6324 * This will cause the interrupt handler to do nothing in case an overflow
6325 * interrupt was in-flight
6326 * This also guarantees that pmc0 will contain the final state
6327 * It virtually gives us full control on overflow processing from that point
6330 SET_PMU_OWNER(NULL, NULL);
6331 DPRINT(("releasing ownership\n"));
6334 * read current overflow status:
6336 * we are guaranteed to read the final stable state
6339 pmc0 = ia64_get_pmc(0); /* slow */
6342 * reset freeze bit, overflow status information destroyed
6346 pmc0 = task->thread.pmcs[0];
6348 * clear whatever overflow status bits there were
6350 task->thread.pmcs[0] = 0;
6352 ovfl_val = pmu_conf->ovfl_val;
6354 * we save all the used pmds
6355 * we take care of overflows for counting PMDs
6357 * XXX: sampling situation is not taken into account here
6359 mask2 = ctx->ctx_used_pmds[0];
6361 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6363 for (i = 0; mask2; i++, mask2>>=1) {
6365 /* skip non used pmds */
6366 if ((mask2 & 0x1) == 0) continue;
6369 * can access PMU always true in system wide mode
6371 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6373 if (PMD_IS_COUNTING(i)) {
6374 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6377 ctx->ctx_pmds[i].val,
6381 * we rebuild the full 64 bit value of the counter
6383 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6386 * now everything is in ctx_pmds[] and we need
6387 * to clear the saved context from save_regs() such that
6388 * pfm_read_pmds() gets the correct value
6393 * take care of overflow inline
6395 if (pmc0 & (1UL << i)) {
6396 val += 1 + ovfl_val;
6397 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6401 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, i, val, pmd_val));
6403 if (is_self) task->thread.pmds[i] = pmd_val;
6405 ctx->ctx_pmds[i].val = val;
6409 static struct irqaction perfmon_irqaction = {
6410 .handler = pfm_interrupt_handler,
6411 .flags = SA_INTERRUPT,
6416 * perfmon initialization routine, called from the initcall() table
6418 static int init_pfm_fs(void);
6426 family = local_cpu_data->family;
6431 if ((*p)->probe() == 0) goto found;
6432 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6443 static struct file_operations pfm_proc_fops = {
6444 .open = pfm_proc_open,
6446 .llseek = seq_lseek,
6447 .release = seq_release,
6453 unsigned int n, n_counters, i;
6455 printk("perfmon: version %u.%u IRQ %u\n",
6458 IA64_PERFMON_VECTOR);
6460 if (pfm_probe_pmu()) {
6461 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6462 local_cpu_data->family);
6467 * compute the number of implemented PMD/PMC from the
6468 * description tables
6471 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6472 if (PMC_IS_IMPL(i) == 0) continue;
6473 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6476 pmu_conf->num_pmcs = n;
6478 n = 0; n_counters = 0;
6479 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6480 if (PMD_IS_IMPL(i) == 0) continue;
6481 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6483 if (PMD_IS_COUNTING(i)) n_counters++;
6485 pmu_conf->num_pmds = n;
6486 pmu_conf->num_counters = n_counters;
6489 * sanity checks on the number of debug registers
6491 if (pmu_conf->use_rr_dbregs) {
6492 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6493 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6497 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6498 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6504 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6508 pmu_conf->num_counters,
6509 ffz(pmu_conf->ovfl_val));
6512 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6513 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6519 * create /proc/perfmon (mostly for debugging purposes)
6521 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6522 if (perfmon_dir == NULL) {
6523 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6528 * install customized file operations for /proc/perfmon entry
6530 perfmon_dir->proc_fops = &pfm_proc_fops;
6533 * create /proc/sys/kernel/perfmon (for debugging purposes)
6535 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6538 * initialize all our spinlocks
6540 spin_lock_init(&pfm_sessions.pfs_lock);
6541 spin_lock_init(&pfm_buffer_fmt_lock);
6545 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6550 __initcall(pfm_init);
6553 * this function is called before pfm_init()
6556 pfm_init_percpu (void)
6559 * make sure no measurement is active
6560 * (may inherit programmed PMCs from EFI).
6566 * we run with the PMU not frozen at all times
6570 if (smp_processor_id() == 0)
6571 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6573 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6578 * used for debug purposes only
6581 dump_pmu_state(const char *from)
6583 struct task_struct *task;
6584 struct thread_struct *t;
6585 struct pt_regs *regs;
6587 unsigned long psr, dcr, info, flags;
6590 local_irq_save(flags);
6592 this_cpu = smp_processor_id();
6593 regs = ia64_task_regs(current);
6594 info = PFM_CPUINFO_GET();
6595 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6597 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6598 local_irq_restore(flags);
6602 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6609 task = GET_PMU_OWNER();
6610 ctx = GET_PMU_CTX();
6612 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6614 psr = pfm_get_psr();
6616 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",
6619 psr & IA64_PSR_PP ? 1 : 0,
6620 psr & IA64_PSR_UP ? 1 : 0,
6621 dcr & IA64_DCR_PP ? 1 : 0,
6624 ia64_psr(regs)->pp);
6626 ia64_psr(regs)->up = 0;
6627 ia64_psr(regs)->pp = 0;
6629 t = ¤t->thread;
6631 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6632 if (PMC_IS_IMPL(i) == 0) continue;
6633 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6636 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6637 if (PMD_IS_IMPL(i) == 0) continue;
6638 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6642 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6645 ctx->ctx_smpl_vaddr,
6649 ctx->ctx_saved_psr_up);
6651 local_irq_restore(flags);
6655 * called from process.c:copy_thread(). task is new child.
6658 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6660 struct thread_struct *thread;
6662 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6664 thread = &task->thread;
6667 * cut links inherited from parent (current)
6669 thread->pfm_context = NULL;
6671 PFM_SET_WORK_PENDING(task, 0);
6674 * the psr bits are already set properly in copy_threads()
6677 #else /* !CONFIG_PERFMON */
6679 sys_perfmonctl (int fd, int cmd, void *arg, int count, long arg5, long arg6, long arg7,
6680 long arg8, long stack)
6684 #endif /* CONFIG_PERFMON */