2 * This file implements the perfmon-2 subsystem which is used
3 * to program the IA-64 Performance Monitoring Unit (PMU).
5 * The initial version of perfmon.c was written by
6 * Ganesh Venkitachalam, IBM Corp.
8 * Then it was modified for perfmon-1.x by Stephane Eranian and
9 * David Mosberger, Hewlett Packard Co.
11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
12 * by Stephane Eranian, Hewlett Packard Co.
14 * Copyright (C) 1999-2005 Hewlett Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 * David Mosberger-Tang <davidm@hpl.hp.com>
18 * More information about perfmon available at:
19 * http://www.hpl.hp.com/research/linux/perfmon
22 #include <linux/config.h>
23 #include <linux/module.h>
24 #include <linux/kernel.h>
25 #include <linux/sched.h>
26 #include <linux/interrupt.h>
27 #include <linux/smp_lock.h>
28 #include <linux/proc_fs.h>
29 #include <linux/seq_file.h>
30 #include <linux/init.h>
31 #include <linux/vmalloc.h>
33 #include <linux/sysctl.h>
34 #include <linux/list.h>
35 #include <linux/file.h>
36 #include <linux/poll.h>
37 #include <linux/vfs.h>
38 #include <linux/pagemap.h>
39 #include <linux/mount.h>
40 #include <linux/bitops.h>
41 #include <linux/capability.h>
42 #include <linux/rcupdate.h>
43 #include <linux/completion.h>
44 #include <linux/vs_memory.h>
45 #include <linux/vs_cvirt.h>
47 #include <asm/errno.h>
48 #include <asm/intrinsics.h>
50 #include <asm/perfmon.h>
51 #include <asm/processor.h>
52 #include <asm/signal.h>
53 #include <asm/system.h>
54 #include <asm/uaccess.h>
55 #include <asm/delay.h>
59 * perfmon context state
61 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
62 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
63 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
64 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
66 #define PFM_INVALID_ACTIVATION (~0UL)
69 * depth of message queue
71 #define PFM_MAX_MSGS 32
72 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
75 * type of a PMU register (bitmask).
77 * bit0 : register implemented
80 * bit4 : pmc has pmc.pm
81 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
82 * bit6-7 : register type
85 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
86 #define PFM_REG_IMPL 0x1 /* register implemented */
87 #define PFM_REG_END 0x2 /* end marker */
88 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
89 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
90 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
91 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
92 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
94 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
95 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
97 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
99 /* i assumed unsigned */
100 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
101 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
103 /* XXX: these assume that register i is implemented */
104 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
105 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
106 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
107 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
109 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
110 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
111 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
112 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
114 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
115 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
117 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
118 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
119 #define PFM_CTX_TASK(h) (h)->ctx_task
121 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
123 /* XXX: does not support more than 64 PMDs */
124 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
125 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
127 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
129 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
130 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
131 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
132 #define PFM_CODE_RR 0 /* requesting code range restriction */
133 #define PFM_DATA_RR 1 /* requestion data range restriction */
135 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
136 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
137 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
139 #define RDEP(x) (1UL<<(x))
142 * context protection macros
144 * - we need to protect against CPU concurrency (spin_lock)
145 * - we need to protect against PMU overflow interrupts (local_irq_disable)
147 * - we need to protect against PMU overflow interrupts (local_irq_disable)
149 * spin_lock_irqsave()/spin_lock_irqrestore():
150 * in SMP: local_irq_disable + spin_lock
151 * in UP : local_irq_disable
153 * spin_lock()/spin_lock():
154 * in UP : removed automatically
155 * in SMP: protect against context accesses from other CPU. interrupts
156 * are not masked. This is useful for the PMU interrupt handler
157 * because we know we will not get PMU concurrency in that code.
159 #define PROTECT_CTX(c, f) \
161 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \
162 spin_lock_irqsave(&(c)->ctx_lock, f); \
163 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
166 #define UNPROTECT_CTX(c, f) \
168 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
169 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
172 #define PROTECT_CTX_NOPRINT(c, f) \
174 spin_lock_irqsave(&(c)->ctx_lock, f); \
178 #define UNPROTECT_CTX_NOPRINT(c, f) \
180 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
184 #define PROTECT_CTX_NOIRQ(c) \
186 spin_lock(&(c)->ctx_lock); \
189 #define UNPROTECT_CTX_NOIRQ(c) \
191 spin_unlock(&(c)->ctx_lock); \
197 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
198 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
199 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
201 #else /* !CONFIG_SMP */
202 #define SET_ACTIVATION(t) do {} while(0)
203 #define GET_ACTIVATION(t) do {} while(0)
204 #define INC_ACTIVATION(t) do {} while(0)
205 #endif /* CONFIG_SMP */
207 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
208 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
209 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
211 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
212 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
214 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
217 * cmp0 must be the value of pmc0
219 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
221 #define PFMFS_MAGIC 0xa0b4d889
226 #define PFM_DEBUGGING 1
230 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
233 #define DPRINT_ovfl(a) \
235 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
240 * 64-bit software counter structure
242 * the next_reset_type is applied to the next call to pfm_reset_regs()
245 unsigned long val; /* virtual 64bit counter value */
246 unsigned long lval; /* last reset value */
247 unsigned long long_reset; /* reset value on sampling overflow */
248 unsigned long short_reset; /* reset value on overflow */
249 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
250 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
251 unsigned long seed; /* seed for random-number generator */
252 unsigned long mask; /* mask for random-number generator */
253 unsigned int flags; /* notify/do not notify */
254 unsigned long eventid; /* overflow event identifier */
261 unsigned int block:1; /* when 1, task will blocked on user notifications */
262 unsigned int system:1; /* do system wide monitoring */
263 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
264 unsigned int is_sampling:1; /* true if using a custom format */
265 unsigned int excl_idle:1; /* exclude idle task in system wide session */
266 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
267 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
268 unsigned int no_msg:1; /* no message sent on overflow */
269 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
270 unsigned int reserved:22;
271 } pfm_context_flags_t;
273 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
274 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
275 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
279 * perfmon context: encapsulates all the state of a monitoring session
282 typedef struct pfm_context {
283 spinlock_t ctx_lock; /* context protection */
285 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
286 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
288 struct task_struct *ctx_task; /* task to which context is attached */
290 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
292 struct completion ctx_restart_done; /* use for blocking notification mode */
294 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
295 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
296 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
298 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
299 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
300 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
302 unsigned long ctx_pmcs[IA64_NUM_PMC_REGS]; /* saved copies of PMC values */
304 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
305 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
306 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
307 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
309 pfm_counter_t ctx_pmds[IA64_NUM_PMD_REGS]; /* software state for PMDS */
311 u64 ctx_saved_psr_up; /* only contains psr.up value */
313 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
314 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
315 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
317 int ctx_fd; /* file descriptor used my this context */
318 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
320 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
321 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
322 unsigned long ctx_smpl_size; /* size of sampling buffer */
323 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
325 wait_queue_head_t ctx_msgq_wait;
326 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
329 struct fasync_struct *ctx_async_queue;
331 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
335 * magic number used to verify that structure is really
338 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
340 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
343 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
344 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
346 #define SET_LAST_CPU(ctx, v) do {} while(0)
347 #define GET_LAST_CPU(ctx) do {} while(0)
351 #define ctx_fl_block ctx_flags.block
352 #define ctx_fl_system ctx_flags.system
353 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
354 #define ctx_fl_is_sampling ctx_flags.is_sampling
355 #define ctx_fl_excl_idle ctx_flags.excl_idle
356 #define ctx_fl_going_zombie ctx_flags.going_zombie
357 #define ctx_fl_trap_reason ctx_flags.trap_reason
358 #define ctx_fl_no_msg ctx_flags.no_msg
359 #define ctx_fl_can_restart ctx_flags.can_restart
361 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
362 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
365 * global information about all sessions
366 * mostly used to synchronize between system wide and per-process
369 spinlock_t pfs_lock; /* lock the structure */
371 unsigned int pfs_task_sessions; /* number of per task sessions */
372 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
373 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
374 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
375 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
379 * information about a PMC or PMD.
380 * dep_pmd[]: a bitmask of dependent PMD registers
381 * dep_pmc[]: a bitmask of dependent PMC registers
383 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
387 unsigned long default_value; /* power-on default value */
388 unsigned long reserved_mask; /* bitmask of reserved bits */
389 pfm_reg_check_t read_check;
390 pfm_reg_check_t write_check;
391 unsigned long dep_pmd[4];
392 unsigned long dep_pmc[4];
395 /* assume cnum is a valid monitor */
396 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
399 * This structure is initialized at boot time and contains
400 * a description of the PMU main characteristics.
402 * If the probe function is defined, detection is based
403 * on its return value:
404 * - 0 means recognized PMU
405 * - anything else means not supported
406 * When the probe function is not defined, then the pmu_family field
407 * is used and it must match the host CPU family such that:
408 * - cpu->family & config->pmu_family != 0
411 unsigned long ovfl_val; /* overflow value for counters */
413 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
414 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
416 unsigned int num_pmcs; /* number of PMCS: computed at init time */
417 unsigned int num_pmds; /* number of PMDS: computed at init time */
418 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
419 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
421 char *pmu_name; /* PMU family name */
422 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
423 unsigned int flags; /* pmu specific flags */
424 unsigned int num_ibrs; /* number of IBRS: computed at init time */
425 unsigned int num_dbrs; /* number of DBRS: computed at init time */
426 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
427 int (*probe)(void); /* customized probe routine */
428 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
433 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
436 * debug register related type definitions
439 unsigned long ibr_mask:56;
440 unsigned long ibr_plm:4;
441 unsigned long ibr_ig:3;
442 unsigned long ibr_x:1;
446 unsigned long dbr_mask:56;
447 unsigned long dbr_plm:4;
448 unsigned long dbr_ig:2;
449 unsigned long dbr_w:1;
450 unsigned long dbr_r:1;
461 * perfmon command descriptions
464 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
467 unsigned int cmd_narg;
469 int (*cmd_getsize)(void *arg, size_t *sz);
472 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
473 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
474 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
475 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
478 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
479 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
480 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
481 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
482 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
484 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
487 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
488 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
489 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
490 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
491 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
492 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
493 unsigned long pfm_smpl_handler_calls;
494 unsigned long pfm_smpl_handler_cycles;
495 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
499 * perfmon internal variables
501 static pfm_stats_t pfm_stats[NR_CPUS];
502 static pfm_session_t pfm_sessions; /* global sessions information */
504 static DEFINE_SPINLOCK(pfm_alt_install_check);
505 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
507 static struct proc_dir_entry *perfmon_dir;
508 static pfm_uuid_t pfm_null_uuid = {0,};
510 static spinlock_t pfm_buffer_fmt_lock;
511 static LIST_HEAD(pfm_buffer_fmt_list);
513 static pmu_config_t *pmu_conf;
515 /* sysctl() controls */
516 pfm_sysctl_t pfm_sysctl;
517 EXPORT_SYMBOL(pfm_sysctl);
519 static ctl_table pfm_ctl_table[]={
520 {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
521 {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
522 {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
523 {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
526 static ctl_table pfm_sysctl_dir[] = {
527 {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
530 static ctl_table pfm_sysctl_root[] = {
531 {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
534 static struct ctl_table_header *pfm_sysctl_header;
536 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
537 static int pfm_flush(struct file *filp);
539 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
540 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
543 pfm_put_task(struct task_struct *task)
545 if (task != current) put_task_struct(task);
549 pfm_set_task_notify(struct task_struct *task)
551 struct thread_info *info;
553 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
554 set_bit(TIF_NOTIFY_RESUME, &info->flags);
558 pfm_clear_task_notify(void)
560 clear_thread_flag(TIF_NOTIFY_RESUME);
564 pfm_reserve_page(unsigned long a)
566 SetPageReserved(vmalloc_to_page((void *)a));
569 pfm_unreserve_page(unsigned long a)
571 ClearPageReserved(vmalloc_to_page((void*)a));
574 static inline unsigned long
575 pfm_protect_ctx_ctxsw(pfm_context_t *x)
577 spin_lock(&(x)->ctx_lock);
582 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
584 spin_unlock(&(x)->ctx_lock);
587 static inline unsigned int
588 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
590 return do_munmap(mm, addr, len);
593 static inline unsigned long
594 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
596 return get_unmapped_area(file, addr, len, pgoff, flags);
600 static struct super_block *
601 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
603 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
606 static struct file_system_type pfm_fs_type = {
608 .get_sb = pfmfs_get_sb,
609 .kill_sb = kill_anon_super,
612 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
613 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
614 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
615 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
616 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
619 /* forward declaration */
620 static struct file_operations pfm_file_ops;
623 * forward declarations
626 static void pfm_lazy_save_regs (struct task_struct *ta);
629 void dump_pmu_state(const char *);
630 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
632 #include "perfmon_itanium.h"
633 #include "perfmon_mckinley.h"
634 #include "perfmon_montecito.h"
635 #include "perfmon_generic.h"
637 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);
1278 extern void update_pal_halt_status(int);
1281 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1283 unsigned long flags;
1285 * validy checks on cpu_mask have been done upstream
1289 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1290 pfm_sessions.pfs_sys_sessions,
1291 pfm_sessions.pfs_task_sessions,
1292 pfm_sessions.pfs_sys_use_dbregs,
1298 * cannot mix system wide and per-task sessions
1300 if (pfm_sessions.pfs_task_sessions > 0UL) {
1301 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1302 pfm_sessions.pfs_task_sessions));
1306 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1308 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1310 pfm_sessions.pfs_sys_session[cpu] = task;
1312 pfm_sessions.pfs_sys_sessions++ ;
1315 if (pfm_sessions.pfs_sys_sessions) goto abort;
1316 pfm_sessions.pfs_task_sessions++;
1319 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1320 pfm_sessions.pfs_sys_sessions,
1321 pfm_sessions.pfs_task_sessions,
1322 pfm_sessions.pfs_sys_use_dbregs,
1327 * disable default_idle() to go to PAL_HALT
1329 update_pal_halt_status(0);
1336 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1337 pfm_sessions.pfs_sys_session[cpu]->pid,
1347 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1349 unsigned long flags;
1351 * validy checks on cpu_mask have been done upstream
1355 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1356 pfm_sessions.pfs_sys_sessions,
1357 pfm_sessions.pfs_task_sessions,
1358 pfm_sessions.pfs_sys_use_dbregs,
1364 pfm_sessions.pfs_sys_session[cpu] = NULL;
1366 * would not work with perfmon+more than one bit in cpu_mask
1368 if (ctx && ctx->ctx_fl_using_dbreg) {
1369 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1370 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1372 pfm_sessions.pfs_sys_use_dbregs--;
1375 pfm_sessions.pfs_sys_sessions--;
1377 pfm_sessions.pfs_task_sessions--;
1379 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1380 pfm_sessions.pfs_sys_sessions,
1381 pfm_sessions.pfs_task_sessions,
1382 pfm_sessions.pfs_sys_use_dbregs,
1387 * if possible, enable default_idle() to go into PAL_HALT
1389 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1390 update_pal_halt_status(1);
1398 * removes virtual mapping of the sampling buffer.
1399 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1400 * a PROTECT_CTX() section.
1403 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1408 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1409 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1413 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1416 * does the actual unmapping
1418 down_write(&task->mm->mmap_sem);
1420 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1422 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1424 up_write(&task->mm->mmap_sem);
1426 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1429 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1435 * free actual physical storage used by sampling buffer
1439 pfm_free_smpl_buffer(pfm_context_t *ctx)
1441 pfm_buffer_fmt_t *fmt;
1443 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1446 * we won't use the buffer format anymore
1448 fmt = ctx->ctx_buf_fmt;
1450 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1453 ctx->ctx_smpl_vaddr));
1455 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1460 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1462 ctx->ctx_smpl_hdr = NULL;
1463 ctx->ctx_smpl_size = 0UL;
1468 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1474 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1476 if (fmt == NULL) return;
1478 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1483 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1484 * no real gain from having the whole whorehouse mounted. So we don't need
1485 * any operations on the root directory. However, we need a non-trivial
1486 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1488 static struct vfsmount *pfmfs_mnt;
1493 int err = register_filesystem(&pfm_fs_type);
1495 pfmfs_mnt = kern_mount(&pfm_fs_type);
1496 err = PTR_ERR(pfmfs_mnt);
1497 if (IS_ERR(pfmfs_mnt))
1498 unregister_filesystem(&pfm_fs_type);
1508 unregister_filesystem(&pfm_fs_type);
1513 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1518 unsigned long flags;
1519 DECLARE_WAITQUEUE(wait, current);
1520 if (PFM_IS_FILE(filp) == 0) {
1521 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1525 ctx = (pfm_context_t *)filp->private_data;
1527 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1532 * check even when there is no message
1534 if (size < sizeof(pfm_msg_t)) {
1535 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1539 PROTECT_CTX(ctx, flags);
1542 * put ourselves on the wait queue
1544 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1552 set_current_state(TASK_INTERRUPTIBLE);
1554 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1557 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1559 UNPROTECT_CTX(ctx, flags);
1562 * check non-blocking read
1565 if(filp->f_flags & O_NONBLOCK) break;
1568 * check pending signals
1570 if(signal_pending(current)) {
1575 * no message, so wait
1579 PROTECT_CTX(ctx, flags);
1581 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1582 set_current_state(TASK_RUNNING);
1583 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1585 if (ret < 0) goto abort;
1588 msg = pfm_get_next_msg(ctx);
1590 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1594 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1597 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1600 UNPROTECT_CTX(ctx, flags);
1606 pfm_write(struct file *file, const char __user *ubuf,
1607 size_t size, loff_t *ppos)
1609 DPRINT(("pfm_write called\n"));
1614 pfm_poll(struct file *filp, poll_table * wait)
1617 unsigned long flags;
1618 unsigned int mask = 0;
1620 if (PFM_IS_FILE(filp) == 0) {
1621 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1625 ctx = (pfm_context_t *)filp->private_data;
1627 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1632 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1634 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1636 PROTECT_CTX(ctx, flags);
1638 if (PFM_CTXQ_EMPTY(ctx) == 0)
1639 mask = POLLIN | POLLRDNORM;
1641 UNPROTECT_CTX(ctx, flags);
1643 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1649 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1651 DPRINT(("pfm_ioctl called\n"));
1656 * interrupt cannot be masked when coming here
1659 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1663 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1665 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1669 ctx->ctx_async_queue, ret));
1675 pfm_fasync(int fd, struct file *filp, int on)
1680 if (PFM_IS_FILE(filp) == 0) {
1681 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1685 ctx = (pfm_context_t *)filp->private_data;
1687 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1691 * we cannot mask interrupts during this call because this may
1692 * may go to sleep if memory is not readily avalaible.
1694 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1695 * done in caller. Serialization of this function is ensured by caller.
1697 ret = pfm_do_fasync(fd, filp, ctx, on);
1700 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1703 ctx->ctx_async_queue, ret));
1710 * this function is exclusively called from pfm_close().
1711 * The context is not protected at that time, nor are interrupts
1712 * on the remote CPU. That's necessary to avoid deadlocks.
1715 pfm_syswide_force_stop(void *info)
1717 pfm_context_t *ctx = (pfm_context_t *)info;
1718 struct pt_regs *regs = task_pt_regs(current);
1719 struct task_struct *owner;
1720 unsigned long flags;
1723 if (ctx->ctx_cpu != smp_processor_id()) {
1724 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1726 smp_processor_id());
1729 owner = GET_PMU_OWNER();
1730 if (owner != ctx->ctx_task) {
1731 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1733 owner->pid, ctx->ctx_task->pid);
1736 if (GET_PMU_CTX() != ctx) {
1737 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1739 GET_PMU_CTX(), ctx);
1743 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1745 * the context is already protected in pfm_close(), we simply
1746 * need to mask interrupts to avoid a PMU interrupt race on
1749 local_irq_save(flags);
1751 ret = pfm_context_unload(ctx, NULL, 0, regs);
1753 DPRINT(("context_unload returned %d\n", ret));
1757 * unmask interrupts, PMU interrupts are now spurious here
1759 local_irq_restore(flags);
1763 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1767 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1768 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1769 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1771 #endif /* CONFIG_SMP */
1774 * called for each close(). Partially free resources.
1775 * When caller is self-monitoring, the context is unloaded.
1778 pfm_flush(struct file *filp)
1781 struct task_struct *task;
1782 struct pt_regs *regs;
1783 unsigned long flags;
1784 unsigned long smpl_buf_size = 0UL;
1785 void *smpl_buf_vaddr = NULL;
1786 int state, is_system;
1788 if (PFM_IS_FILE(filp) == 0) {
1789 DPRINT(("bad magic for\n"));
1793 ctx = (pfm_context_t *)filp->private_data;
1795 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1800 * remove our file from the async queue, if we use this mode.
1801 * This can be done without the context being protected. We come
1802 * here when the context has become unreacheable by other tasks.
1804 * We may still have active monitoring at this point and we may
1805 * end up in pfm_overflow_handler(). However, fasync_helper()
1806 * operates with interrupts disabled and it cleans up the
1807 * queue. If the PMU handler is called prior to entering
1808 * fasync_helper() then it will send a signal. If it is
1809 * invoked after, it will find an empty queue and no
1810 * signal will be sent. In both case, we are safe
1812 if (filp->f_flags & FASYNC) {
1813 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1814 pfm_do_fasync (-1, filp, ctx, 0);
1817 PROTECT_CTX(ctx, flags);
1819 state = ctx->ctx_state;
1820 is_system = ctx->ctx_fl_system;
1822 task = PFM_CTX_TASK(ctx);
1823 regs = task_pt_regs(task);
1825 DPRINT(("ctx_state=%d is_current=%d\n",
1827 task == current ? 1 : 0));
1830 * if state == UNLOADED, then task is NULL
1834 * we must stop and unload because we are losing access to the context.
1836 if (task == current) {
1839 * the task IS the owner but it migrated to another CPU: that's bad
1840 * but we must handle this cleanly. Unfortunately, the kernel does
1841 * not provide a mechanism to block migration (while the context is loaded).
1843 * We need to release the resource on the ORIGINAL cpu.
1845 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1847 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1849 * keep context protected but unmask interrupt for IPI
1851 local_irq_restore(flags);
1853 pfm_syswide_cleanup_other_cpu(ctx);
1856 * restore interrupt masking
1858 local_irq_save(flags);
1861 * context is unloaded at this point
1864 #endif /* CONFIG_SMP */
1867 DPRINT(("forcing unload\n"));
1869 * stop and unload, returning with state UNLOADED
1870 * and session unreserved.
1872 pfm_context_unload(ctx, NULL, 0, regs);
1874 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1879 * remove virtual mapping, if any, for the calling task.
1880 * cannot reset ctx field until last user is calling close().
1882 * ctx_smpl_vaddr must never be cleared because it is needed
1883 * by every task with access to the context
1885 * When called from do_exit(), the mm context is gone already, therefore
1886 * mm is NULL, i.e., the VMA is already gone and we do not have to
1889 if (ctx->ctx_smpl_vaddr && current->mm) {
1890 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1891 smpl_buf_size = ctx->ctx_smpl_size;
1894 UNPROTECT_CTX(ctx, flags);
1897 * if there was a mapping, then we systematically remove it
1898 * at this point. Cannot be done inside critical section
1899 * because some VM function reenables interrupts.
1902 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1907 * called either on explicit close() or from exit_files().
1908 * Only the LAST user of the file gets to this point, i.e., it is
1911 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1912 * (fput()),i.e, last task to access the file. Nobody else can access the
1913 * file at this point.
1915 * When called from exit_files(), the VMA has been freed because exit_mm()
1916 * is executed before exit_files().
1918 * When called from exit_files(), the current task is not yet ZOMBIE but we
1919 * flush the PMU state to the context.
1922 pfm_close(struct inode *inode, struct file *filp)
1925 struct task_struct *task;
1926 struct pt_regs *regs;
1927 DECLARE_WAITQUEUE(wait, current);
1928 unsigned long flags;
1929 unsigned long smpl_buf_size = 0UL;
1930 void *smpl_buf_addr = NULL;
1931 int free_possible = 1;
1932 int state, is_system;
1934 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1936 if (PFM_IS_FILE(filp) == 0) {
1937 DPRINT(("bad magic\n"));
1941 ctx = (pfm_context_t *)filp->private_data;
1943 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1947 PROTECT_CTX(ctx, flags);
1949 state = ctx->ctx_state;
1950 is_system = ctx->ctx_fl_system;
1952 task = PFM_CTX_TASK(ctx);
1953 regs = task_pt_regs(task);
1955 DPRINT(("ctx_state=%d is_current=%d\n",
1957 task == current ? 1 : 0));
1960 * if task == current, then pfm_flush() unloaded the context
1962 if (state == PFM_CTX_UNLOADED) goto doit;
1965 * context is loaded/masked and task != current, we need to
1966 * either force an unload or go zombie
1970 * The task is currently blocked or will block after an overflow.
1971 * we must force it to wakeup to get out of the
1972 * MASKED state and transition to the unloaded state by itself.
1974 * This situation is only possible for per-task mode
1976 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1979 * set a "partial" zombie state to be checked
1980 * upon return from down() in pfm_handle_work().
1982 * We cannot use the ZOMBIE state, because it is checked
1983 * by pfm_load_regs() which is called upon wakeup from down().
1984 * In such case, it would free the context and then we would
1985 * return to pfm_handle_work() which would access the
1986 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1987 * but visible to pfm_handle_work().
1989 * For some window of time, we have a zombie context with
1990 * ctx_state = MASKED and not ZOMBIE
1992 ctx->ctx_fl_going_zombie = 1;
1995 * force task to wake up from MASKED state
1997 complete(&ctx->ctx_restart_done);
1999 DPRINT(("waking up ctx_state=%d\n", state));
2002 * put ourself to sleep waiting for the other
2003 * task to report completion
2005 * the context is protected by mutex, therefore there
2006 * is no risk of being notified of completion before
2007 * begin actually on the waitq.
2009 set_current_state(TASK_INTERRUPTIBLE);
2010 add_wait_queue(&ctx->ctx_zombieq, &wait);
2012 UNPROTECT_CTX(ctx, flags);
2015 * XXX: check for signals :
2016 * - ok for explicit close
2017 * - not ok when coming from exit_files()
2022 PROTECT_CTX(ctx, flags);
2025 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2026 set_current_state(TASK_RUNNING);
2029 * context is unloaded at this point
2031 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2033 else if (task != current) {
2036 * switch context to zombie state
2038 ctx->ctx_state = PFM_CTX_ZOMBIE;
2040 DPRINT(("zombie ctx for [%d]\n", task->pid));
2042 * cannot free the context on the spot. deferred until
2043 * the task notices the ZOMBIE state
2047 pfm_context_unload(ctx, NULL, 0, regs);
2052 /* reload state, may have changed during opening of critical section */
2053 state = ctx->ctx_state;
2056 * the context is still attached to a task (possibly current)
2057 * we cannot destroy it right now
2061 * we must free the sampling buffer right here because
2062 * we cannot rely on it being cleaned up later by the
2063 * monitored task. It is not possible to free vmalloc'ed
2064 * memory in pfm_load_regs(). Instead, we remove the buffer
2065 * now. should there be subsequent PMU overflow originally
2066 * meant for sampling, the will be converted to spurious
2067 * and that's fine because the monitoring tools is gone anyway.
2069 if (ctx->ctx_smpl_hdr) {
2070 smpl_buf_addr = ctx->ctx_smpl_hdr;
2071 smpl_buf_size = ctx->ctx_smpl_size;
2072 /* no more sampling */
2073 ctx->ctx_smpl_hdr = NULL;
2074 ctx->ctx_fl_is_sampling = 0;
2077 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2083 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2086 * UNLOADED that the session has already been unreserved.
2088 if (state == PFM_CTX_ZOMBIE) {
2089 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2093 * disconnect file descriptor from context must be done
2096 filp->private_data = NULL;
2099 * if we free on the spot, the context is now completely unreacheable
2100 * from the callers side. The monitored task side is also cut, so we
2103 * If we have a deferred free, only the caller side is disconnected.
2105 UNPROTECT_CTX(ctx, flags);
2108 * All memory free operations (especially for vmalloc'ed memory)
2109 * MUST be done with interrupts ENABLED.
2111 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2114 * return the memory used by the context
2116 if (free_possible) pfm_context_free(ctx);
2122 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2124 DPRINT(("pfm_no_open called\n"));
2130 static struct file_operations pfm_file_ops = {
2131 .llseek = no_llseek,
2136 .open = pfm_no_open, /* special open code to disallow open via /proc */
2137 .fasync = pfm_fasync,
2138 .release = pfm_close,
2143 pfmfs_delete_dentry(struct dentry *dentry)
2148 static struct dentry_operations pfmfs_dentry_operations = {
2149 .d_delete = pfmfs_delete_dentry,
2154 pfm_alloc_fd(struct file **cfile)
2157 struct file *file = NULL;
2158 struct inode * inode;
2162 fd = get_unused_fd();
2163 if (fd < 0) return -ENFILE;
2167 file = get_empty_filp();
2168 if (!file) goto out;
2171 * allocate a new inode
2173 inode = new_inode(pfmfs_mnt->mnt_sb);
2174 if (!inode) goto out;
2176 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2178 inode->i_mode = S_IFCHR|S_IRUGO;
2179 inode->i_uid = current->fsuid;
2180 inode->i_gid = current->fsgid;
2182 sprintf(name, "[%lu]", inode->i_ino);
2184 this.len = strlen(name);
2185 this.hash = inode->i_ino;
2190 * allocate a new dcache entry
2192 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2193 if (!file->f_dentry) goto out;
2195 file->f_dentry->d_op = &pfmfs_dentry_operations;
2197 d_add(file->f_dentry, inode);
2198 file->f_vfsmnt = mntget(pfmfs_mnt);
2199 file->f_mapping = inode->i_mapping;
2201 file->f_op = &pfm_file_ops;
2202 file->f_mode = FMODE_READ;
2203 file->f_flags = O_RDONLY;
2207 * may have to delay until context is attached?
2209 fd_install(fd, file);
2212 * the file structure we will use
2218 if (file) put_filp(file);
2224 pfm_free_fd(int fd, struct file *file)
2226 struct files_struct *files = current->files;
2227 struct fdtable *fdt;
2230 * there ie no fd_uninstall(), so we do it here
2232 spin_lock(&files->file_lock);
2233 fdt = files_fdtable(files);
2234 rcu_assign_pointer(fdt->fd[fd], NULL);
2235 spin_unlock(&files->file_lock);
2243 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2245 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2248 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2251 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2262 * allocate a sampling buffer and remaps it into the user address space of the task
2265 pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2267 struct mm_struct *mm = task->mm;
2268 struct vm_area_struct *vma = NULL;
2274 * the fixed header + requested size and align to page boundary
2276 size = PAGE_ALIGN(rsize);
2278 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2281 * check requested size to avoid Denial-of-service attacks
2282 * XXX: may have to refine this test
2283 * Check against address space limit.
2285 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2288 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2292 * We do the easy to undo allocations first.
2294 * pfm_rvmalloc(), clears the buffer, so there is no leak
2296 smpl_buf = pfm_rvmalloc(size);
2297 if (smpl_buf == NULL) {
2298 DPRINT(("Can't allocate sampling buffer\n"));
2302 DPRINT(("smpl_buf @%p\n", smpl_buf));
2305 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2307 DPRINT(("Cannot allocate vma\n"));
2310 memset(vma, 0, sizeof(*vma));
2313 * partially initialize the vma for the sampling buffer
2316 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2317 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2320 * Now we have everything we need and we can initialize
2321 * and connect all the data structures
2324 ctx->ctx_smpl_hdr = smpl_buf;
2325 ctx->ctx_smpl_size = size; /* aligned size */
2328 * Let's do the difficult operations next.
2330 * now we atomically find some area in the address space and
2331 * remap the buffer in it.
2333 down_write(&task->mm->mmap_sem);
2335 /* find some free area in address space, must have mmap sem held */
2336 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2337 if (vma->vm_start == 0UL) {
2338 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2339 up_write(&task->mm->mmap_sem);
2342 vma->vm_end = vma->vm_start + size;
2343 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2345 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2347 /* can only be applied to current task, need to have the mm semaphore held when called */
2348 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2349 DPRINT(("Can't remap buffer\n"));
2350 up_write(&task->mm->mmap_sem);
2355 * now insert the vma in the vm list for the process, must be
2356 * done with mmap lock held
2358 insert_vm_struct(mm, vma);
2360 vx_vmpages_add(mm, size >> PAGE_SHIFT);
2361 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2363 up_write(&task->mm->mmap_sem);
2366 * keep track of user level virtual address
2368 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2369 *(unsigned long *)user_vaddr = vma->vm_start;
2374 kmem_cache_free(vm_area_cachep, vma);
2376 pfm_rvfree(smpl_buf, size);
2382 * XXX: do something better here
2385 pfm_bad_permissions(struct task_struct *task)
2387 /* inspired by ptrace_attach() */
2388 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2397 return ((current->uid != task->euid)
2398 || (current->uid != task->suid)
2399 || (current->uid != task->uid)
2400 || (current->gid != task->egid)
2401 || (current->gid != task->sgid)
2402 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2406 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2412 ctx_flags = pfx->ctx_flags;
2414 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2417 * cannot block in this mode
2419 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2420 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2425 /* probably more to add here */
2431 pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2432 unsigned int cpu, pfarg_context_t *arg)
2434 pfm_buffer_fmt_t *fmt = NULL;
2435 unsigned long size = 0UL;
2437 void *fmt_arg = NULL;
2439 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2441 /* invoke and lock buffer format, if found */
2442 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2444 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2449 * buffer argument MUST be contiguous to pfarg_context_t
2451 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2453 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2455 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2457 if (ret) goto error;
2459 /* link buffer format and context */
2460 ctx->ctx_buf_fmt = fmt;
2463 * check if buffer format wants to use perfmon buffer allocation/mapping service
2465 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2466 if (ret) goto error;
2470 * buffer is always remapped into the caller's address space
2472 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2473 if (ret) goto error;
2475 /* keep track of user address of buffer */
2476 arg->ctx_smpl_vaddr = uaddr;
2478 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2485 pfm_reset_pmu_state(pfm_context_t *ctx)
2490 * install reset values for PMC.
2492 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2493 if (PMC_IS_IMPL(i) == 0) continue;
2494 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2495 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2498 * PMD registers are set to 0UL when the context in memset()
2502 * On context switched restore, we must restore ALL pmc and ALL pmd even
2503 * when they are not actively used by the task. In UP, the incoming process
2504 * may otherwise pick up left over PMC, PMD state from the previous process.
2505 * As opposed to PMD, stale PMC can cause harm to the incoming
2506 * process because they may change what is being measured.
2507 * Therefore, we must systematically reinstall the entire
2508 * PMC state. In SMP, the same thing is possible on the
2509 * same CPU but also on between 2 CPUs.
2511 * The problem with PMD is information leaking especially
2512 * to user level when psr.sp=0
2514 * There is unfortunately no easy way to avoid this problem
2515 * on either UP or SMP. This definitively slows down the
2516 * pfm_load_regs() function.
2520 * bitmask of all PMCs accessible to this context
2522 * PMC0 is treated differently.
2524 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2527 * bitmask of all PMDs that are accesible to this context
2529 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2531 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2534 * useful in case of re-enable after disable
2536 ctx->ctx_used_ibrs[0] = 0UL;
2537 ctx->ctx_used_dbrs[0] = 0UL;
2541 pfm_ctx_getsize(void *arg, size_t *sz)
2543 pfarg_context_t *req = (pfarg_context_t *)arg;
2544 pfm_buffer_fmt_t *fmt;
2548 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2550 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2552 DPRINT(("cannot find buffer format\n"));
2555 /* get just enough to copy in user parameters */
2556 *sz = fmt->fmt_arg_size;
2557 DPRINT(("arg_size=%lu\n", *sz));
2565 * cannot attach if :
2567 * - task not owned by caller
2568 * - task incompatible with context mode
2571 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2574 * no kernel task or task not owner by caller
2576 if (task->mm == NULL) {
2577 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2580 if (pfm_bad_permissions(task)) {
2581 DPRINT(("no permission to attach to [%d]\n", task->pid));
2585 * cannot block in self-monitoring mode
2587 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2588 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2592 if (task->exit_state == EXIT_ZOMBIE) {
2593 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2598 * always ok for self
2600 if (task == current) return 0;
2602 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
2603 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2607 * make sure the task is off any CPU
2609 wait_task_inactive(task);
2611 /* more to come... */
2617 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2619 struct task_struct *p = current;
2622 /* XXX: need to add more checks here */
2623 if (pid < 2) return -EPERM;
2625 if (pid != current->pid) {
2627 read_lock(&tasklist_lock);
2629 p = find_task_by_pid(pid);
2631 /* make sure task cannot go away while we operate on it */
2632 if (p) get_task_struct(p);
2634 read_unlock(&tasklist_lock);
2636 if (p == NULL) return -ESRCH;
2639 ret = pfm_task_incompatible(ctx, p);
2642 } else if (p != current) {
2651 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2653 pfarg_context_t *req = (pfarg_context_t *)arg;
2658 /* let's check the arguments first */
2659 ret = pfarg_is_sane(current, req);
2660 if (ret < 0) return ret;
2662 ctx_flags = req->ctx_flags;
2666 ctx = pfm_context_alloc();
2667 if (!ctx) goto error;
2669 ret = pfm_alloc_fd(&filp);
2670 if (ret < 0) goto error_file;
2672 req->ctx_fd = ctx->ctx_fd = ret;
2675 * attach context to file
2677 filp->private_data = ctx;
2680 * does the user want to sample?
2682 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2683 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2684 if (ret) goto buffer_error;
2688 * init context protection lock
2690 spin_lock_init(&ctx->ctx_lock);
2693 * context is unloaded
2695 ctx->ctx_state = PFM_CTX_UNLOADED;
2698 * initialization of context's flags
2700 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2701 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2702 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2703 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2705 * will move to set properties
2706 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2710 * init restart semaphore to locked
2712 init_completion(&ctx->ctx_restart_done);
2715 * activation is used in SMP only
2717 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2718 SET_LAST_CPU(ctx, -1);
2721 * initialize notification message queue
2723 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2724 init_waitqueue_head(&ctx->ctx_msgq_wait);
2725 init_waitqueue_head(&ctx->ctx_zombieq);
2727 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2732 ctx->ctx_fl_excl_idle,
2737 * initialize soft PMU state
2739 pfm_reset_pmu_state(ctx);
2744 pfm_free_fd(ctx->ctx_fd, filp);
2746 if (ctx->ctx_buf_fmt) {
2747 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2750 pfm_context_free(ctx);
2756 static inline unsigned long
2757 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2759 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2760 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2761 extern unsigned long carta_random32 (unsigned long seed);
2763 if (reg->flags & PFM_REGFL_RANDOM) {
2764 new_seed = carta_random32(old_seed);
2765 val -= (old_seed & mask); /* counter values are negative numbers! */
2766 if ((mask >> 32) != 0)
2767 /* construct a full 64-bit random value: */
2768 new_seed |= carta_random32(old_seed >> 32) << 32;
2769 reg->seed = new_seed;
2776 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2778 unsigned long mask = ovfl_regs[0];
2779 unsigned long reset_others = 0UL;
2784 * now restore reset value on sampling overflowed counters
2786 mask >>= PMU_FIRST_COUNTER;
2787 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2789 if ((mask & 0x1UL) == 0UL) continue;
2791 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2792 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2794 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2798 * Now take care of resetting the other registers
2800 for(i = 0; reset_others; i++, reset_others >>= 1) {
2802 if ((reset_others & 0x1) == 0) continue;
2804 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2806 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2807 is_long_reset ? "long" : "short", i, val));
2812 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2814 unsigned long mask = ovfl_regs[0];
2815 unsigned long reset_others = 0UL;
2819 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2821 if (ctx->ctx_state == PFM_CTX_MASKED) {
2822 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2827 * now restore reset value on sampling overflowed counters
2829 mask >>= PMU_FIRST_COUNTER;
2830 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2832 if ((mask & 0x1UL) == 0UL) continue;
2834 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2835 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2837 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2839 pfm_write_soft_counter(ctx, i, val);
2843 * Now take care of resetting the other registers
2845 for(i = 0; reset_others; i++, reset_others >>= 1) {
2847 if ((reset_others & 0x1) == 0) continue;
2849 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2851 if (PMD_IS_COUNTING(i)) {
2852 pfm_write_soft_counter(ctx, i, val);
2854 ia64_set_pmd(i, val);
2856 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2857 is_long_reset ? "long" : "short", i, val));
2863 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2865 struct thread_struct *thread = NULL;
2866 struct task_struct *task;
2867 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2868 unsigned long value, pmc_pm;
2869 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2870 unsigned int cnum, reg_flags, flags, pmc_type;
2871 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2872 int is_monitor, is_counting, state;
2874 pfm_reg_check_t wr_func;
2875 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2877 state = ctx->ctx_state;
2878 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2879 is_system = ctx->ctx_fl_system;
2880 task = ctx->ctx_task;
2881 impl_pmds = pmu_conf->impl_pmds[0];
2883 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2886 thread = &task->thread;
2888 * In system wide and when the context is loaded, access can only happen
2889 * when the caller is running on the CPU being monitored by the session.
2890 * It does not have to be the owner (ctx_task) of the context per se.
2892 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2893 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2896 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2898 expert_mode = pfm_sysctl.expert_mode;
2900 for (i = 0; i < count; i++, req++) {
2902 cnum = req->reg_num;
2903 reg_flags = req->reg_flags;
2904 value = req->reg_value;
2905 smpl_pmds = req->reg_smpl_pmds[0];
2906 reset_pmds = req->reg_reset_pmds[0];
2910 if (cnum >= PMU_MAX_PMCS) {
2911 DPRINT(("pmc%u is invalid\n", cnum));
2915 pmc_type = pmu_conf->pmc_desc[cnum].type;
2916 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2917 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2918 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2921 * we reject all non implemented PMC as well
2922 * as attempts to modify PMC[0-3] which are used
2923 * as status registers by the PMU
2925 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2926 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2929 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2931 * If the PMC is a monitor, then if the value is not the default:
2932 * - system-wide session: PMCx.pm=1 (privileged monitor)
2933 * - per-task : PMCx.pm=0 (user monitor)
2935 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2936 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2945 * enforce generation of overflow interrupt. Necessary on all
2948 value |= 1 << PMU_PMC_OI;
2950 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2951 flags |= PFM_REGFL_OVFL_NOTIFY;
2954 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2956 /* verify validity of smpl_pmds */
2957 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2958 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2962 /* verify validity of reset_pmds */
2963 if ((reset_pmds & impl_pmds) != reset_pmds) {
2964 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2968 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2969 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2972 /* eventid on non-counting monitors are ignored */
2976 * execute write checker, if any
2978 if (likely(expert_mode == 0 && wr_func)) {
2979 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2980 if (ret) goto error;
2985 * no error on this register
2987 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2990 * Now we commit the changes to the software state
2994 * update overflow information
2998 * full flag update each time a register is programmed
3000 ctx->ctx_pmds[cnum].flags = flags;
3002 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
3003 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
3004 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
3007 * Mark all PMDS to be accessed as used.
3009 * We do not keep track of PMC because we have to
3010 * systematically restore ALL of them.
3012 * We do not update the used_monitors mask, because
3013 * if we have not programmed them, then will be in
3014 * a quiescent state, therefore we will not need to
3015 * mask/restore then when context is MASKED.
3017 CTX_USED_PMD(ctx, reset_pmds);
3018 CTX_USED_PMD(ctx, smpl_pmds);
3020 * make sure we do not try to reset on
3021 * restart because we have established new values
3023 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3026 * Needed in case the user does not initialize the equivalent
3027 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3028 * possible leak here.
3030 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3033 * keep track of the monitor PMC that we are using.
3034 * we save the value of the pmc in ctx_pmcs[] and if
3035 * the monitoring is not stopped for the context we also
3036 * place it in the saved state area so that it will be
3037 * picked up later by the context switch code.
3039 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3041 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3042 * monitoring needs to be stopped.
3044 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3047 * update context state
3049 ctx->ctx_pmcs[cnum] = value;
3053 * write thread state
3055 if (is_system == 0) thread->pmcs[cnum] = value;
3058 * write hardware register if we can
3060 if (can_access_pmu) {
3061 ia64_set_pmc(cnum, value);
3066 * per-task SMP only here
3068 * we are guaranteed that the task is not running on the other CPU,
3069 * we indicate that this PMD will need to be reloaded if the task
3070 * is rescheduled on the CPU it ran last on.
3072 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3077 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",
3083 ctx->ctx_all_pmcs[0],
3084 ctx->ctx_used_pmds[0],
3085 ctx->ctx_pmds[cnum].eventid,
3088 ctx->ctx_reload_pmcs[0],
3089 ctx->ctx_used_monitors[0],
3090 ctx->ctx_ovfl_regs[0]));
3094 * make sure the changes are visible
3096 if (can_access_pmu) ia64_srlz_d();
3100 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3105 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3107 struct thread_struct *thread = NULL;
3108 struct task_struct *task;
3109 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3110 unsigned long value, hw_value, ovfl_mask;
3112 int i, can_access_pmu = 0, state;
3113 int is_counting, is_loaded, is_system, expert_mode;
3115 pfm_reg_check_t wr_func;
3118 state = ctx->ctx_state;
3119 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3120 is_system = ctx->ctx_fl_system;
3121 ovfl_mask = pmu_conf->ovfl_val;
3122 task = ctx->ctx_task;
3124 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3127 * on both UP and SMP, we can only write to the PMC when the task is
3128 * the owner of the local PMU.
3130 if (likely(is_loaded)) {
3131 thread = &task->thread;
3133 * In system wide and when the context is loaded, access can only happen
3134 * when the caller is running on the CPU being monitored by the session.
3135 * It does not have to be the owner (ctx_task) of the context per se.
3137 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3138 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3141 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3143 expert_mode = pfm_sysctl.expert_mode;
3145 for (i = 0; i < count; i++, req++) {
3147 cnum = req->reg_num;
3148 value = req->reg_value;
3150 if (!PMD_IS_IMPL(cnum)) {
3151 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3154 is_counting = PMD_IS_COUNTING(cnum);
3155 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3158 * execute write checker, if any
3160 if (unlikely(expert_mode == 0 && wr_func)) {
3161 unsigned long v = value;
3163 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3164 if (ret) goto abort_mission;
3171 * no error on this register
3173 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3176 * now commit changes to software state
3181 * update virtualized (64bits) counter
3185 * write context state
3187 ctx->ctx_pmds[cnum].lval = value;
3190 * when context is load we use the split value
3193 hw_value = value & ovfl_mask;
3194 value = value & ~ovfl_mask;
3198 * update reset values (not just for counters)
3200 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3201 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3204 * update randomization parameters (not just for counters)
3206 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3207 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3210 * update context value
3212 ctx->ctx_pmds[cnum].val = value;
3215 * Keep track of what we use
3217 * We do not keep track of PMC because we have to
3218 * systematically restore ALL of them.
3220 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3223 * mark this PMD register used as well
3225 CTX_USED_PMD(ctx, RDEP(cnum));
3228 * make sure we do not try to reset on
3229 * restart because we have established new values
3231 if (is_counting && state == PFM_CTX_MASKED) {
3232 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3237 * write thread state
3239 if (is_system == 0) thread->pmds[cnum] = hw_value;
3242 * write hardware register if we can
3244 if (can_access_pmu) {
3245 ia64_set_pmd(cnum, hw_value);
3249 * we are guaranteed that the task is not running on the other CPU,
3250 * we indicate that this PMD will need to be reloaded if the task
3251 * is rescheduled on the CPU it ran last on.
3253 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3258 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3259 "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",
3265 ctx->ctx_pmds[cnum].val,
3266 ctx->ctx_pmds[cnum].short_reset,
3267 ctx->ctx_pmds[cnum].long_reset,
3268 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3269 ctx->ctx_pmds[cnum].seed,
3270 ctx->ctx_pmds[cnum].mask,
3271 ctx->ctx_used_pmds[0],
3272 ctx->ctx_pmds[cnum].reset_pmds[0],
3273 ctx->ctx_reload_pmds[0],
3274 ctx->ctx_all_pmds[0],
3275 ctx->ctx_ovfl_regs[0]));
3279 * make changes visible
3281 if (can_access_pmu) ia64_srlz_d();
3287 * for now, we have only one possibility for error
3289 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3294 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3295 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3296 * interrupt is delivered during the call, it will be kept pending until we leave, making
3297 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3298 * guaranteed to return consistent data to the user, it may simply be old. It is not
3299 * trivial to treat the overflow while inside the call because you may end up in
3300 * some module sampling buffer code causing deadlocks.
3303 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3305 struct thread_struct *thread = NULL;
3306 struct task_struct *task;
3307 unsigned long val = 0UL, lval, ovfl_mask, sval;
3308 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3309 unsigned int cnum, reg_flags = 0;
3310 int i, can_access_pmu = 0, state;
3311 int is_loaded, is_system, is_counting, expert_mode;
3313 pfm_reg_check_t rd_func;
3316 * access is possible when loaded only for
3317 * self-monitoring tasks or in UP mode
3320 state = ctx->ctx_state;
3321 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3322 is_system = ctx->ctx_fl_system;
3323 ovfl_mask = pmu_conf->ovfl_val;
3324 task = ctx->ctx_task;
3326 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3328 if (likely(is_loaded)) {
3329 thread = &task->thread;
3331 * In system wide and when the context is loaded, access can only happen
3332 * when the caller is running on the CPU being monitored by the session.
3333 * It does not have to be the owner (ctx_task) of the context per se.
3335 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3336 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3340 * this can be true when not self-monitoring only in UP
3342 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3344 if (can_access_pmu) ia64_srlz_d();
3346 expert_mode = pfm_sysctl.expert_mode;
3348 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3354 * on both UP and SMP, we can only read the PMD from the hardware register when
3355 * the task is the owner of the local PMU.
3358 for (i = 0; i < count; i++, req++) {
3360 cnum = req->reg_num;
3361 reg_flags = req->reg_flags;
3363 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3365 * we can only read the register that we use. That includes
3366 * the one we explicitely initialize AND the one we want included
3367 * in the sampling buffer (smpl_regs).
3369 * Having this restriction allows optimization in the ctxsw routine
3370 * without compromising security (leaks)
3372 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3374 sval = ctx->ctx_pmds[cnum].val;
3375 lval = ctx->ctx_pmds[cnum].lval;
3376 is_counting = PMD_IS_COUNTING(cnum);
3379 * If the task is not the current one, then we check if the
3380 * PMU state is still in the local live register due to lazy ctxsw.
3381 * If true, then we read directly from the registers.
3383 if (can_access_pmu){
3384 val = ia64_get_pmd(cnum);
3387 * context has been saved
3388 * if context is zombie, then task does not exist anymore.
3389 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3391 val = is_loaded ? thread->pmds[cnum] : 0UL;
3393 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3397 * XXX: need to check for overflow when loaded
3404 * execute read checker, if any
3406 if (unlikely(expert_mode == 0 && rd_func)) {
3407 unsigned long v = val;
3408 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3409 if (ret) goto error;
3414 PFM_REG_RETFLAG_SET(reg_flags, 0);
3416 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3419 * update register return value, abort all if problem during copy.
3420 * we only modify the reg_flags field. no check mode is fine because
3421 * access has been verified upfront in sys_perfmonctl().
3423 req->reg_value = val;
3424 req->reg_flags = reg_flags;
3425 req->reg_last_reset_val = lval;
3431 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3436 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3440 if (req == NULL) return -EINVAL;
3442 ctx = GET_PMU_CTX();
3444 if (ctx == NULL) return -EINVAL;
3447 * for now limit to current task, which is enough when calling
3448 * from overflow handler
3450 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3452 return pfm_write_pmcs(ctx, req, nreq, regs);
3454 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3457 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3461 if (req == NULL) return -EINVAL;
3463 ctx = GET_PMU_CTX();
3465 if (ctx == NULL) return -EINVAL;
3468 * for now limit to current task, which is enough when calling
3469 * from overflow handler
3471 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3473 return pfm_read_pmds(ctx, req, nreq, regs);
3475 EXPORT_SYMBOL(pfm_mod_read_pmds);
3478 * Only call this function when a process it trying to
3479 * write the debug registers (reading is always allowed)
3482 pfm_use_debug_registers(struct task_struct *task)
3484 pfm_context_t *ctx = task->thread.pfm_context;
3485 unsigned long flags;
3488 if (pmu_conf->use_rr_dbregs == 0) return 0;
3490 DPRINT(("called for [%d]\n", task->pid));
3495 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3498 * Even on SMP, we do not need to use an atomic here because
3499 * the only way in is via ptrace() and this is possible only when the
3500 * process is stopped. Even in the case where the ctxsw out is not totally
3501 * completed by the time we come here, there is no way the 'stopped' process
3502 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3503 * So this is always safe.
3505 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3510 * We cannot allow setting breakpoints when system wide monitoring
3511 * sessions are using the debug registers.
3513 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3516 pfm_sessions.pfs_ptrace_use_dbregs++;
3518 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3519 pfm_sessions.pfs_ptrace_use_dbregs,
3520 pfm_sessions.pfs_sys_use_dbregs,
3529 * This function is called for every task that exits with the
3530 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3531 * able to use the debug registers for debugging purposes via
3532 * ptrace(). Therefore we know it was not using them for
3533 * perfmormance monitoring, so we only decrement the number
3534 * of "ptraced" debug register users to keep the count up to date
3537 pfm_release_debug_registers(struct task_struct *task)
3539 unsigned long flags;
3542 if (pmu_conf->use_rr_dbregs == 0) return 0;
3545 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3546 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3549 pfm_sessions.pfs_ptrace_use_dbregs--;
3558 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3560 struct task_struct *task;
3561 pfm_buffer_fmt_t *fmt;
3562 pfm_ovfl_ctrl_t rst_ctrl;
3563 int state, is_system;
3566 state = ctx->ctx_state;
3567 fmt = ctx->ctx_buf_fmt;
3568 is_system = ctx->ctx_fl_system;
3569 task = PFM_CTX_TASK(ctx);
3572 case PFM_CTX_MASKED:
3574 case PFM_CTX_LOADED:
3575 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3577 case PFM_CTX_UNLOADED:
3578 case PFM_CTX_ZOMBIE:
3579 DPRINT(("invalid state=%d\n", state));
3582 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3587 * In system wide and when the context is loaded, access can only happen
3588 * when the caller is running on the CPU being monitored by the session.
3589 * It does not have to be the owner (ctx_task) of the context per se.
3591 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3592 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3597 if (unlikely(task == NULL)) {
3598 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3602 if (task == current || is_system) {
3604 fmt = ctx->ctx_buf_fmt;
3606 DPRINT(("restarting self %d ovfl=0x%lx\n",
3608 ctx->ctx_ovfl_regs[0]));
3610 if (CTX_HAS_SMPL(ctx)) {
3612 prefetch(ctx->ctx_smpl_hdr);
3614 rst_ctrl.bits.mask_monitoring = 0;
3615 rst_ctrl.bits.reset_ovfl_pmds = 0;
3617 if (state == PFM_CTX_LOADED)
3618 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3620 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3622 rst_ctrl.bits.mask_monitoring = 0;
3623 rst_ctrl.bits.reset_ovfl_pmds = 1;
3627 if (rst_ctrl.bits.reset_ovfl_pmds)
3628 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3630 if (rst_ctrl.bits.mask_monitoring == 0) {
3631 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3633 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3635 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3637 // cannot use pfm_stop_monitoring(task, regs);
3641 * clear overflowed PMD mask to remove any stale information
3643 ctx->ctx_ovfl_regs[0] = 0UL;
3646 * back to LOADED state
3648 ctx->ctx_state = PFM_CTX_LOADED;
3651 * XXX: not really useful for self monitoring
3653 ctx->ctx_fl_can_restart = 0;
3659 * restart another task
3663 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3664 * one is seen by the task.
3666 if (state == PFM_CTX_MASKED) {
3667 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3669 * will prevent subsequent restart before this one is
3670 * seen by other task
3672 ctx->ctx_fl_can_restart = 0;
3676 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3677 * the task is blocked or on its way to block. That's the normal
3678 * restart path. If the monitoring is not masked, then the task
3679 * can be actively monitoring and we cannot directly intervene.
3680 * Therefore we use the trap mechanism to catch the task and
3681 * force it to reset the buffer/reset PMDs.
3683 * if non-blocking, then we ensure that the task will go into
3684 * pfm_handle_work() before returning to user mode.
3686 * We cannot explicitely reset another task, it MUST always
3687 * be done by the task itself. This works for system wide because
3688 * the tool that is controlling the session is logically doing
3689 * "self-monitoring".
3691 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3692 DPRINT(("unblocking [%d] \n", task->pid));
3693 complete(&ctx->ctx_restart_done);
3695 DPRINT(("[%d] armed exit trap\n", task->pid));
3697 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3699 PFM_SET_WORK_PENDING(task, 1);
3701 pfm_set_task_notify(task);
3704 * XXX: send reschedule if task runs on another CPU
3711 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3713 unsigned int m = *(unsigned int *)arg;
3715 pfm_sysctl.debug = m == 0 ? 0 : 1;
3717 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3720 memset(pfm_stats, 0, sizeof(pfm_stats));
3721 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3727 * arg can be NULL and count can be zero for this function
3730 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3732 struct thread_struct *thread = NULL;
3733 struct task_struct *task;
3734 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3735 unsigned long flags;
3740 int i, can_access_pmu = 0;
3741 int is_system, is_loaded;
3743 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3745 state = ctx->ctx_state;
3746 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3747 is_system = ctx->ctx_fl_system;
3748 task = ctx->ctx_task;
3750 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3753 * on both UP and SMP, we can only write to the PMC when the task is
3754 * the owner of the local PMU.
3757 thread = &task->thread;
3759 * In system wide and when the context is loaded, access can only happen
3760 * when the caller is running on the CPU being monitored by the session.
3761 * It does not have to be the owner (ctx_task) of the context per se.
3763 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3764 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3767 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3771 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3772 * ensuring that no real breakpoint can be installed via this call.
3774 * IMPORTANT: regs can be NULL in this function
3777 first_time = ctx->ctx_fl_using_dbreg == 0;
3780 * don't bother if we are loaded and task is being debugged
3782 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3783 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3788 * check for debug registers in system wide mode
3790 * If though a check is done in pfm_context_load(),
3791 * we must repeat it here, in case the registers are
3792 * written after the context is loaded
3797 if (first_time && is_system) {
3798 if (pfm_sessions.pfs_ptrace_use_dbregs)
3801 pfm_sessions.pfs_sys_use_dbregs++;
3806 if (ret != 0) return ret;
3809 * mark ourself as user of the debug registers for
3812 ctx->ctx_fl_using_dbreg = 1;
3815 * clear hardware registers to make sure we don't
3816 * pick up stale state.
3818 * for a system wide session, we do not use
3819 * thread.dbr, thread.ibr because this process
3820 * never leaves the current CPU and the state
3821 * is shared by all processes running on it
3823 if (first_time && can_access_pmu) {
3824 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3825 for (i=0; i < pmu_conf->num_ibrs; i++) {
3826 ia64_set_ibr(i, 0UL);
3827 ia64_dv_serialize_instruction();
3830 for (i=0; i < pmu_conf->num_dbrs; i++) {
3831 ia64_set_dbr(i, 0UL);
3832 ia64_dv_serialize_data();
3838 * Now install the values into the registers
3840 for (i = 0; i < count; i++, req++) {
3842 rnum = req->dbreg_num;
3843 dbreg.val = req->dbreg_value;
3847 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3848 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3849 rnum, dbreg.val, mode, i, count));
3855 * make sure we do not install enabled breakpoint
3858 if (mode == PFM_CODE_RR)
3859 dbreg.ibr.ibr_x = 0;
3861 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3864 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3867 * Debug registers, just like PMC, can only be modified
3868 * by a kernel call. Moreover, perfmon() access to those
3869 * registers are centralized in this routine. The hardware
3870 * does not modify the value of these registers, therefore,
3871 * if we save them as they are written, we can avoid having
3872 * to save them on context switch out. This is made possible
3873 * by the fact that when perfmon uses debug registers, ptrace()
3874 * won't be able to modify them concurrently.
3876 if (mode == PFM_CODE_RR) {
3877 CTX_USED_IBR(ctx, rnum);
3879 if (can_access_pmu) {
3880 ia64_set_ibr(rnum, dbreg.val);
3881 ia64_dv_serialize_instruction();
3884 ctx->ctx_ibrs[rnum] = dbreg.val;
3886 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3887 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3889 CTX_USED_DBR(ctx, rnum);
3891 if (can_access_pmu) {
3892 ia64_set_dbr(rnum, dbreg.val);
3893 ia64_dv_serialize_data();
3895 ctx->ctx_dbrs[rnum] = dbreg.val;
3897 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3898 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3906 * in case it was our first attempt, we undo the global modifications
3910 if (ctx->ctx_fl_system) {
3911 pfm_sessions.pfs_sys_use_dbregs--;
3914 ctx->ctx_fl_using_dbreg = 0;
3917 * install error return flag
3919 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3925 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3927 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3931 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3933 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3937 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3941 if (req == NULL) return -EINVAL;
3943 ctx = GET_PMU_CTX();
3945 if (ctx == NULL) return -EINVAL;
3948 * for now limit to current task, which is enough when calling
3949 * from overflow handler
3951 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3953 return pfm_write_ibrs(ctx, req, nreq, regs);
3955 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3958 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3962 if (req == NULL) return -EINVAL;
3964 ctx = GET_PMU_CTX();
3966 if (ctx == NULL) return -EINVAL;
3969 * for now limit to current task, which is enough when calling
3970 * from overflow handler
3972 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3974 return pfm_write_dbrs(ctx, req, nreq, regs);
3976 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3980 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3982 pfarg_features_t *req = (pfarg_features_t *)arg;
3984 req->ft_version = PFM_VERSION;
3989 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3991 struct pt_regs *tregs;
3992 struct task_struct *task = PFM_CTX_TASK(ctx);
3993 int state, is_system;
3995 state = ctx->ctx_state;
3996 is_system = ctx->ctx_fl_system;
3999 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
4001 if (state == PFM_CTX_UNLOADED) return -EINVAL;
4004 * In system wide and when the context is loaded, access can only happen
4005 * when the caller is running on the CPU being monitored by the session.
4006 * It does not have to be the owner (ctx_task) of the context per se.
4008 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4009 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4012 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4013 PFM_CTX_TASK(ctx)->pid,
4017 * in system mode, we need to update the PMU directly
4018 * and the user level state of the caller, which may not
4019 * necessarily be the creator of the context.
4023 * Update local PMU first
4027 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4031 * update local cpuinfo
4033 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4036 * stop monitoring, does srlz.i
4041 * stop monitoring in the caller
4043 ia64_psr(regs)->pp = 0;
4051 if (task == current) {
4052 /* stop monitoring at kernel level */
4056 * stop monitoring at the user level
4058 ia64_psr(regs)->up = 0;
4060 tregs = task_pt_regs(task);
4063 * stop monitoring at the user level
4065 ia64_psr(tregs)->up = 0;
4068 * monitoring disabled in kernel at next reschedule
4070 ctx->ctx_saved_psr_up = 0;
4071 DPRINT(("task=[%d]\n", task->pid));
4078 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4080 struct pt_regs *tregs;
4081 int state, is_system;
4083 state = ctx->ctx_state;
4084 is_system = ctx->ctx_fl_system;
4086 if (state != PFM_CTX_LOADED) return -EINVAL;
4089 * In system wide and when the context is loaded, access can only happen
4090 * when the caller is running on the CPU being monitored by the session.
4091 * It does not have to be the owner (ctx_task) of the context per se.
4093 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4094 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4099 * in system mode, we need to update the PMU directly
4100 * and the user level state of the caller, which may not
4101 * necessarily be the creator of the context.
4106 * set user level psr.pp for the caller
4108 ia64_psr(regs)->pp = 1;
4111 * now update the local PMU and cpuinfo
4113 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4116 * start monitoring at kernel level
4121 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4131 if (ctx->ctx_task == current) {
4133 /* start monitoring at kernel level */
4137 * activate monitoring at user level
4139 ia64_psr(regs)->up = 1;
4142 tregs = task_pt_regs(ctx->ctx_task);
4145 * start monitoring at the kernel level the next
4146 * time the task is scheduled
4148 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4151 * activate monitoring at user level
4153 ia64_psr(tregs)->up = 1;
4159 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4161 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4166 for (i = 0; i < count; i++, req++) {
4168 cnum = req->reg_num;
4170 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4172 req->reg_value = PMC_DFL_VAL(cnum);
4174 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4176 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4181 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4186 pfm_check_task_exist(pfm_context_t *ctx)
4188 struct task_struct *g, *t;
4191 read_lock(&tasklist_lock);
4193 do_each_thread (g, t) {
4194 if (t->thread.pfm_context == ctx) {
4198 } while_each_thread (g, t);
4200 read_unlock(&tasklist_lock);
4202 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4208 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4210 struct task_struct *task;
4211 struct thread_struct *thread;
4212 struct pfm_context_t *old;
4213 unsigned long flags;
4215 struct task_struct *owner_task = NULL;
4217 pfarg_load_t *req = (pfarg_load_t *)arg;
4218 unsigned long *pmcs_source, *pmds_source;
4221 int state, is_system, set_dbregs = 0;
4223 state = ctx->ctx_state;
4224 is_system = ctx->ctx_fl_system;
4226 * can only load from unloaded or terminated state
4228 if (state != PFM_CTX_UNLOADED) {
4229 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4235 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4237 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4238 DPRINT(("cannot use blocking mode on self\n"));
4242 ret = pfm_get_task(ctx, req->load_pid, &task);
4244 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4251 * system wide is self monitoring only
4253 if (is_system && task != current) {
4254 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4259 thread = &task->thread;
4263 * cannot load a context which is using range restrictions,
4264 * into a task that is being debugged.
4266 if (ctx->ctx_fl_using_dbreg) {
4267 if (thread->flags & IA64_THREAD_DBG_VALID) {
4269 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4275 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4276 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4279 pfm_sessions.pfs_sys_use_dbregs++;
4280 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4287 if (ret) goto error;
4291 * SMP system-wide monitoring implies self-monitoring.
4293 * The programming model expects the task to
4294 * be pinned on a CPU throughout the session.
4295 * Here we take note of the current CPU at the
4296 * time the context is loaded. No call from
4297 * another CPU will be allowed.
4299 * The pinning via shed_setaffinity()
4300 * must be done by the calling task prior
4303 * systemwide: keep track of CPU this session is supposed to run on
4305 the_cpu = ctx->ctx_cpu = smp_processor_id();
4309 * now reserve the session
4311 ret = pfm_reserve_session(current, is_system, the_cpu);
4312 if (ret) goto error;
4315 * task is necessarily stopped at this point.
4317 * If the previous context was zombie, then it got removed in
4318 * pfm_save_regs(). Therefore we should not see it here.
4319 * If we see a context, then this is an active context
4321 * XXX: needs to be atomic
4323 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4324 thread->pfm_context, ctx));
4327 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4329 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4333 pfm_reset_msgq(ctx);
4335 ctx->ctx_state = PFM_CTX_LOADED;
4338 * link context to task
4340 ctx->ctx_task = task;
4344 * we load as stopped
4346 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4347 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4349 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4351 thread->flags |= IA64_THREAD_PM_VALID;
4355 * propagate into thread-state
4357 pfm_copy_pmds(task, ctx);
4358 pfm_copy_pmcs(task, ctx);
4360 pmcs_source = thread->pmcs;
4361 pmds_source = thread->pmds;
4364 * always the case for system-wide
4366 if (task == current) {
4368 if (is_system == 0) {
4370 /* allow user level control */
4371 ia64_psr(regs)->sp = 0;
4372 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4374 SET_LAST_CPU(ctx, smp_processor_id());
4376 SET_ACTIVATION(ctx);
4379 * push the other task out, if any
4381 owner_task = GET_PMU_OWNER();
4382 if (owner_task) pfm_lazy_save_regs(owner_task);
4386 * load all PMD from ctx to PMU (as opposed to thread state)
4387 * restore all PMC from ctx to PMU
4389 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4390 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4392 ctx->ctx_reload_pmcs[0] = 0UL;
4393 ctx->ctx_reload_pmds[0] = 0UL;
4396 * guaranteed safe by earlier check against DBG_VALID
4398 if (ctx->ctx_fl_using_dbreg) {
4399 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4400 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4405 SET_PMU_OWNER(task, ctx);
4407 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4410 * when not current, task MUST be stopped, so this is safe
4412 regs = task_pt_regs(task);
4414 /* force a full reload */
4415 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4416 SET_LAST_CPU(ctx, -1);
4418 /* initial saved psr (stopped) */
4419 ctx->ctx_saved_psr_up = 0UL;
4420 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4426 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4429 * we must undo the dbregs setting (for system-wide)
4431 if (ret && set_dbregs) {
4433 pfm_sessions.pfs_sys_use_dbregs--;
4437 * release task, there is now a link with the context
4439 if (is_system == 0 && task != current) {
4443 ret = pfm_check_task_exist(ctx);
4445 ctx->ctx_state = PFM_CTX_UNLOADED;
4446 ctx->ctx_task = NULL;
4454 * in this function, we do not need to increase the use count
4455 * for the task via get_task_struct(), because we hold the
4456 * context lock. If the task were to disappear while having
4457 * a context attached, it would go through pfm_exit_thread()
4458 * which also grabs the context lock and would therefore be blocked
4459 * until we are here.
4461 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4464 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4466 struct task_struct *task = PFM_CTX_TASK(ctx);
4467 struct pt_regs *tregs;
4468 int prev_state, is_system;
4471 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4473 prev_state = ctx->ctx_state;
4474 is_system = ctx->ctx_fl_system;
4477 * unload only when necessary
4479 if (prev_state == PFM_CTX_UNLOADED) {
4480 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4485 * clear psr and dcr bits
4487 ret = pfm_stop(ctx, NULL, 0, regs);
4488 if (ret) return ret;
4490 ctx->ctx_state = PFM_CTX_UNLOADED;
4493 * in system mode, we need to update the PMU directly
4494 * and the user level state of the caller, which may not
4495 * necessarily be the creator of the context.
4502 * local PMU is taken care of in pfm_stop()
4504 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4505 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4508 * save PMDs in context
4511 pfm_flush_pmds(current, ctx);
4514 * at this point we are done with the PMU
4515 * so we can unreserve the resource.
4517 if (prev_state != PFM_CTX_ZOMBIE)
4518 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4521 * disconnect context from task
4523 task->thread.pfm_context = NULL;
4525 * disconnect task from context
4527 ctx->ctx_task = NULL;
4530 * There is nothing more to cleanup here.
4538 tregs = task == current ? regs : task_pt_regs(task);
4540 if (task == current) {
4542 * cancel user level control
4544 ia64_psr(regs)->sp = 1;
4546 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4549 * save PMDs to context
4552 pfm_flush_pmds(task, ctx);
4555 * at this point we are done with the PMU
4556 * so we can unreserve the resource.
4558 * when state was ZOMBIE, we have already unreserved.
4560 if (prev_state != PFM_CTX_ZOMBIE)
4561 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4564 * reset activation counter and psr
4566 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4567 SET_LAST_CPU(ctx, -1);
4570 * PMU state will not be restored
4572 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4575 * break links between context and task
4577 task->thread.pfm_context = NULL;
4578 ctx->ctx_task = NULL;
4580 PFM_SET_WORK_PENDING(task, 0);
4582 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4583 ctx->ctx_fl_can_restart = 0;
4584 ctx->ctx_fl_going_zombie = 0;
4586 DPRINT(("disconnected [%d] from context\n", task->pid));
4593 * called only from exit_thread(): task == current
4594 * we come here only if current has a context attached (loaded or masked)
4597 pfm_exit_thread(struct task_struct *task)
4600 unsigned long flags;
4601 struct pt_regs *regs = task_pt_regs(task);
4605 ctx = PFM_GET_CTX(task);
4607 PROTECT_CTX(ctx, flags);
4609 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4611 state = ctx->ctx_state;
4613 case PFM_CTX_UNLOADED:
4615 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4616 * be in unloaded state
4618 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4620 case PFM_CTX_LOADED:
4621 case PFM_CTX_MASKED:
4622 ret = pfm_context_unload(ctx, NULL, 0, regs);
4624 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4626 DPRINT(("ctx unloaded for current state was %d\n", state));
4628 pfm_end_notify_user(ctx);
4630 case PFM_CTX_ZOMBIE:
4631 ret = pfm_context_unload(ctx, NULL, 0, regs);
4633 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4638 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4641 UNPROTECT_CTX(ctx, flags);
4643 { u64 psr = pfm_get_psr();
4644 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4645 BUG_ON(GET_PMU_OWNER());
4646 BUG_ON(ia64_psr(regs)->up);
4647 BUG_ON(ia64_psr(regs)->pp);
4651 * All memory free operations (especially for vmalloc'ed memory)
4652 * MUST be done with interrupts ENABLED.
4654 if (free_ok) pfm_context_free(ctx);
4658 * functions MUST be listed in the increasing order of their index (see permfon.h)
4660 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4661 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4662 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4663 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4664 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4666 static pfm_cmd_desc_t pfm_cmd_tab[]={
4667 /* 0 */PFM_CMD_NONE,
4668 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4669 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4670 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4671 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4672 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4673 /* 6 */PFM_CMD_NONE,
4674 /* 7 */PFM_CMD_NONE,
4675 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4676 /* 9 */PFM_CMD_NONE,
4677 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4678 /* 11 */PFM_CMD_NONE,
4679 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4680 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4681 /* 14 */PFM_CMD_NONE,
4682 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4683 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4684 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4685 /* 18 */PFM_CMD_NONE,
4686 /* 19 */PFM_CMD_NONE,
4687 /* 20 */PFM_CMD_NONE,
4688 /* 21 */PFM_CMD_NONE,
4689 /* 22 */PFM_CMD_NONE,
4690 /* 23 */PFM_CMD_NONE,
4691 /* 24 */PFM_CMD_NONE,
4692 /* 25 */PFM_CMD_NONE,
4693 /* 26 */PFM_CMD_NONE,
4694 /* 27 */PFM_CMD_NONE,
4695 /* 28 */PFM_CMD_NONE,
4696 /* 29 */PFM_CMD_NONE,
4697 /* 30 */PFM_CMD_NONE,
4698 /* 31 */PFM_CMD_NONE,
4699 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4700 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4702 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4705 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4707 struct task_struct *task;
4708 int state, old_state;
4711 state = ctx->ctx_state;
4712 task = ctx->ctx_task;
4715 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4719 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4723 task->state, PFM_CMD_STOPPED(cmd)));
4726 * self-monitoring always ok.
4728 * for system-wide the caller can either be the creator of the
4729 * context (to one to which the context is attached to) OR
4730 * a task running on the same CPU as the session.
4732 if (task == current || ctx->ctx_fl_system) return 0;
4735 * we are monitoring another thread
4738 case PFM_CTX_UNLOADED:
4740 * if context is UNLOADED we are safe to go
4743 case PFM_CTX_ZOMBIE:
4745 * no command can operate on a zombie context
4747 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4749 case PFM_CTX_MASKED:
4751 * PMU state has been saved to software even though
4752 * the thread may still be running.
4754 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4758 * context is LOADED or MASKED. Some commands may need to have
4761 * We could lift this restriction for UP but it would mean that
4762 * the user has no guarantee the task would not run between
4763 * two successive calls to perfmonctl(). That's probably OK.
4764 * If this user wants to ensure the task does not run, then
4765 * the task must be stopped.
4767 if (PFM_CMD_STOPPED(cmd)) {
4768 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
4769 DPRINT(("[%d] task not in stopped state\n", task->pid));
4773 * task is now stopped, wait for ctxsw out
4775 * This is an interesting point in the code.
4776 * We need to unprotect the context because
4777 * the pfm_save_regs() routines needs to grab
4778 * the same lock. There are danger in doing
4779 * this because it leaves a window open for
4780 * another task to get access to the context
4781 * and possibly change its state. The one thing
4782 * that is not possible is for the context to disappear
4783 * because we are protected by the VFS layer, i.e.,
4784 * get_fd()/put_fd().
4788 UNPROTECT_CTX(ctx, flags);
4790 wait_task_inactive(task);
4792 PROTECT_CTX(ctx, flags);
4795 * we must recheck to verify if state has changed
4797 if (ctx->ctx_state != old_state) {
4798 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4806 * system-call entry point (must return long)
4809 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4811 struct file *file = NULL;
4812 pfm_context_t *ctx = NULL;
4813 unsigned long flags = 0UL;
4814 void *args_k = NULL;
4815 long ret; /* will expand int return types */
4816 size_t base_sz, sz, xtra_sz = 0;
4817 int narg, completed_args = 0, call_made = 0, cmd_flags;
4818 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4819 int (*getsize)(void *arg, size_t *sz);
4820 #define PFM_MAX_ARGSIZE 4096
4823 * reject any call if perfmon was disabled at initialization
4825 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4827 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4828 DPRINT(("invalid cmd=%d\n", cmd));
4832 func = pfm_cmd_tab[cmd].cmd_func;
4833 narg = pfm_cmd_tab[cmd].cmd_narg;
4834 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4835 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4836 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4838 if (unlikely(func == NULL)) {
4839 DPRINT(("invalid cmd=%d\n", cmd));
4843 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4851 * check if number of arguments matches what the command expects
4853 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4857 sz = xtra_sz + base_sz*count;
4859 * limit abuse to min page size
4861 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4862 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4867 * allocate default-sized argument buffer
4869 if (likely(count && args_k == NULL)) {
4870 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4871 if (args_k == NULL) return -ENOMEM;
4879 * assume sz = 0 for command without parameters
4881 if (sz && copy_from_user(args_k, arg, sz)) {
4882 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4887 * check if command supports extra parameters
4889 if (completed_args == 0 && getsize) {
4891 * get extra parameters size (based on main argument)
4893 ret = (*getsize)(args_k, &xtra_sz);
4894 if (ret) goto error_args;
4898 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4900 /* retry if necessary */
4901 if (likely(xtra_sz)) goto restart_args;
4904 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4909 if (unlikely(file == NULL)) {
4910 DPRINT(("invalid fd %d\n", fd));
4913 if (unlikely(PFM_IS_FILE(file) == 0)) {
4914 DPRINT(("fd %d not related to perfmon\n", fd));
4918 ctx = (pfm_context_t *)file->private_data;
4919 if (unlikely(ctx == NULL)) {
4920 DPRINT(("no context for fd %d\n", fd));
4923 prefetch(&ctx->ctx_state);
4925 PROTECT_CTX(ctx, flags);
4928 * check task is stopped
4930 ret = pfm_check_task_state(ctx, cmd, flags);
4931 if (unlikely(ret)) goto abort_locked;
4934 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4940 DPRINT(("context unlocked\n"));
4941 UNPROTECT_CTX(ctx, flags);
4944 /* copy argument back to user, if needed */
4945 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4953 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4959 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4961 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4962 pfm_ovfl_ctrl_t rst_ctrl;
4966 state = ctx->ctx_state;
4968 * Unlock sampling buffer and reset index atomically
4969 * XXX: not really needed when blocking
4971 if (CTX_HAS_SMPL(ctx)) {
4973 rst_ctrl.bits.mask_monitoring = 0;
4974 rst_ctrl.bits.reset_ovfl_pmds = 0;
4976 if (state == PFM_CTX_LOADED)
4977 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4979 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4981 rst_ctrl.bits.mask_monitoring = 0;
4982 rst_ctrl.bits.reset_ovfl_pmds = 1;
4986 if (rst_ctrl.bits.reset_ovfl_pmds) {
4987 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4989 if (rst_ctrl.bits.mask_monitoring == 0) {
4990 DPRINT(("resuming monitoring\n"));
4991 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4993 DPRINT(("stopping monitoring\n"));
4994 //pfm_stop_monitoring(current, regs);
4996 ctx->ctx_state = PFM_CTX_LOADED;
5001 * context MUST BE LOCKED when calling
5002 * can only be called for current
5005 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
5009 DPRINT(("entering for [%d]\n", current->pid));
5011 ret = pfm_context_unload(ctx, NULL, 0, regs);
5013 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", current->pid, ret);
5017 * and wakeup controlling task, indicating we are now disconnected
5019 wake_up_interruptible(&ctx->ctx_zombieq);
5022 * given that context is still locked, the controlling
5023 * task will only get access when we return from
5024 * pfm_handle_work().
5028 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5030 * pfm_handle_work() can be called with interrupts enabled
5031 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5032 * call may sleep, therefore we must re-enable interrupts
5033 * to avoid deadlocks. It is safe to do so because this function
5034 * is called ONLY when returning to user level (PUStk=1), in which case
5035 * there is no risk of kernel stack overflow due to deep
5036 * interrupt nesting.
5039 pfm_handle_work(void)
5042 struct pt_regs *regs;
5043 unsigned long flags, dummy_flags;
5044 unsigned long ovfl_regs;
5045 unsigned int reason;
5048 ctx = PFM_GET_CTX(current);
5050 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5054 PROTECT_CTX(ctx, flags);
5056 PFM_SET_WORK_PENDING(current, 0);
5058 pfm_clear_task_notify();
5060 regs = task_pt_regs(current);
5063 * extract reason for being here and clear
5065 reason = ctx->ctx_fl_trap_reason;
5066 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5067 ovfl_regs = ctx->ctx_ovfl_regs[0];
5069 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5072 * must be done before we check for simple-reset mode
5074 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5077 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5078 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5081 * restore interrupt mask to what it was on entry.
5082 * Could be enabled/diasbled.
5084 UNPROTECT_CTX(ctx, flags);
5087 * force interrupt enable because of down_interruptible()
5091 DPRINT(("before block sleeping\n"));
5094 * may go through without blocking on SMP systems
5095 * if restart has been received already by the time we call down()
5097 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5099 DPRINT(("after block sleeping ret=%d\n", ret));
5102 * lock context and mask interrupts again
5103 * We save flags into a dummy because we may have
5104 * altered interrupts mask compared to entry in this
5107 PROTECT_CTX(ctx, dummy_flags);
5110 * we need to read the ovfl_regs only after wake-up
5111 * because we may have had pfm_write_pmds() in between
5112 * and that can changed PMD values and therefore
5113 * ovfl_regs is reset for these new PMD values.
5115 ovfl_regs = ctx->ctx_ovfl_regs[0];
5117 if (ctx->ctx_fl_going_zombie) {
5119 DPRINT(("context is zombie, bailing out\n"));
5120 pfm_context_force_terminate(ctx, regs);
5124 * in case of interruption of down() we don't restart anything
5126 if (ret < 0) goto nothing_to_do;
5129 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5130 ctx->ctx_ovfl_regs[0] = 0UL;
5134 * restore flags as they were upon entry
5136 UNPROTECT_CTX(ctx, flags);
5140 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5142 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5143 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5147 DPRINT(("waking up somebody\n"));
5149 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5152 * safe, we are not in intr handler, nor in ctxsw when
5155 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5161 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5163 pfm_msg_t *msg = NULL;
5165 if (ctx->ctx_fl_no_msg == 0) {
5166 msg = pfm_get_new_msg(ctx);
5168 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5172 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5173 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5174 msg->pfm_ovfl_msg.msg_active_set = 0;
5175 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5176 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5177 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5178 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5179 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5182 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5188 return pfm_notify_user(ctx, msg);
5192 pfm_end_notify_user(pfm_context_t *ctx)
5196 msg = pfm_get_new_msg(ctx);
5198 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5202 memset(msg, 0, sizeof(*msg));
5204 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5205 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5206 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5208 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5213 return pfm_notify_user(ctx, msg);
5217 * main overflow processing routine.
5218 * it can be called from the interrupt path or explicitely during the context switch code
5221 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5223 pfm_ovfl_arg_t *ovfl_arg;
5225 unsigned long old_val, ovfl_val, new_val;
5226 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5227 unsigned long tstamp;
5228 pfm_ovfl_ctrl_t ovfl_ctrl;
5229 unsigned int i, has_smpl;
5230 int must_notify = 0;
5232 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5235 * sanity test. Should never happen
5237 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5239 tstamp = ia64_get_itc();
5240 mask = pmc0 >> PMU_FIRST_COUNTER;
5241 ovfl_val = pmu_conf->ovfl_val;
5242 has_smpl = CTX_HAS_SMPL(ctx);
5244 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5245 "used_pmds=0x%lx\n",
5247 task ? task->pid: -1,
5248 (regs ? regs->cr_iip : 0),
5249 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5250 ctx->ctx_used_pmds[0]));
5254 * first we update the virtual counters
5255 * assume there was a prior ia64_srlz_d() issued
5257 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5259 /* skip pmd which did not overflow */
5260 if ((mask & 0x1) == 0) continue;
5263 * Note that the pmd is not necessarily 0 at this point as qualified events
5264 * may have happened before the PMU was frozen. The residual count is not
5265 * taken into consideration here but will be with any read of the pmd via
5268 old_val = new_val = ctx->ctx_pmds[i].val;
5269 new_val += 1 + ovfl_val;
5270 ctx->ctx_pmds[i].val = new_val;
5273 * check for overflow condition
5275 if (likely(old_val > new_val)) {
5276 ovfl_pmds |= 1UL << i;
5277 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5280 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5284 ia64_get_pmd(i) & ovfl_val,
5290 * there was no 64-bit overflow, nothing else to do
5292 if (ovfl_pmds == 0UL) return;
5295 * reset all control bits
5301 * if a sampling format module exists, then we "cache" the overflow by
5302 * calling the module's handler() routine.
5305 unsigned long start_cycles, end_cycles;
5306 unsigned long pmd_mask;
5308 int this_cpu = smp_processor_id();
5310 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5311 ovfl_arg = &ctx->ctx_ovfl_arg;
5313 prefetch(ctx->ctx_smpl_hdr);
5315 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5319 if ((pmd_mask & 0x1) == 0) continue;
5321 ovfl_arg->ovfl_pmd = (unsigned char )i;
5322 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5323 ovfl_arg->active_set = 0;
5324 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5325 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5327 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5328 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5329 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5332 * copy values of pmds of interest. Sampling format may copy them
5333 * into sampling buffer.
5336 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5337 if ((smpl_pmds & 0x1) == 0) continue;
5338 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5339 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5343 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5345 start_cycles = ia64_get_itc();
5348 * call custom buffer format record (handler) routine
5350 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5352 end_cycles = ia64_get_itc();
5355 * For those controls, we take the union because they have
5356 * an all or nothing behavior.
5358 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5359 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5360 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5362 * build the bitmask of pmds to reset now
5364 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5366 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5369 * when the module cannot handle the rest of the overflows, we abort right here
5371 if (ret && pmd_mask) {
5372 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5373 pmd_mask<<PMU_FIRST_COUNTER));
5376 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5378 ovfl_pmds &= ~reset_pmds;
5381 * when no sampling module is used, then the default
5382 * is to notify on overflow if requested by user
5384 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5385 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5386 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5387 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5389 * if needed, we reset all overflowed pmds
5391 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5394 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5397 * reset the requested PMD registers using the short reset values
5400 unsigned long bm = reset_pmds;
5401 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5404 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5406 * keep track of what to reset when unblocking
5408 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5411 * check for blocking context
5413 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5415 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5418 * set the perfmon specific checking pending work for the task
5420 PFM_SET_WORK_PENDING(task, 1);
5423 * when coming from ctxsw, current still points to the
5424 * previous task, therefore we must work with task and not current.
5426 pfm_set_task_notify(task);
5429 * defer until state is changed (shorten spin window). the context is locked
5430 * anyway, so the signal receiver would come spin for nothing.
5435 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5436 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5437 PFM_GET_WORK_PENDING(task),
5438 ctx->ctx_fl_trap_reason,
5441 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5443 * in case monitoring must be stopped, we toggle the psr bits
5445 if (ovfl_ctrl.bits.mask_monitoring) {
5446 pfm_mask_monitoring(task);
5447 ctx->ctx_state = PFM_CTX_MASKED;
5448 ctx->ctx_fl_can_restart = 1;
5452 * send notification now
5454 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5459 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5461 task ? task->pid : -1,
5467 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5468 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5469 * come here as zombie only if the task is the current task. In which case, we
5470 * can access the PMU hardware directly.
5472 * Note that zombies do have PM_VALID set. So here we do the minimal.
5474 * In case the context was zombified it could not be reclaimed at the time
5475 * the monitoring program exited. At this point, the PMU reservation has been
5476 * returned, the sampiing buffer has been freed. We must convert this call
5477 * into a spurious interrupt. However, we must also avoid infinite overflows
5478 * by stopping monitoring for this task. We can only come here for a per-task
5479 * context. All we need to do is to stop monitoring using the psr bits which
5480 * are always task private. By re-enabling secure montioring, we ensure that
5481 * the monitored task will not be able to re-activate monitoring.
5482 * The task will eventually be context switched out, at which point the context
5483 * will be reclaimed (that includes releasing ownership of the PMU).
5485 * So there might be a window of time where the number of per-task session is zero
5486 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5487 * context. This is safe because if a per-task session comes in, it will push this one
5488 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5489 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5490 * also push our zombie context out.
5492 * Overall pretty hairy stuff....
5494 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5496 ia64_psr(regs)->up = 0;
5497 ia64_psr(regs)->sp = 1;
5502 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5504 struct task_struct *task;
5506 unsigned long flags;
5508 int this_cpu = smp_processor_id();
5511 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5514 * srlz.d done before arriving here
5516 pmc0 = ia64_get_pmc(0);
5518 task = GET_PMU_OWNER();
5519 ctx = GET_PMU_CTX();
5522 * if we have some pending bits set
5523 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5525 if (PMC0_HAS_OVFL(pmc0) && task) {
5527 * we assume that pmc0.fr is always set here
5531 if (!ctx) goto report_spurious1;
5533 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5534 goto report_spurious2;
5536 PROTECT_CTX_NOPRINT(ctx, flags);
5538 pfm_overflow_handler(task, ctx, pmc0, regs);
5540 UNPROTECT_CTX_NOPRINT(ctx, flags);
5543 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5547 * keep it unfrozen at all times
5554 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5555 this_cpu, task->pid);
5559 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5567 pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5569 unsigned long start_cycles, total_cycles;
5570 unsigned long min, max;
5574 this_cpu = get_cpu();
5575 if (likely(!pfm_alt_intr_handler)) {
5576 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5577 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5579 start_cycles = ia64_get_itc();
5581 ret = pfm_do_interrupt_handler(irq, arg, regs);
5583 total_cycles = ia64_get_itc();
5586 * don't measure spurious interrupts
5588 if (likely(ret == 0)) {
5589 total_cycles -= start_cycles;
5591 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5592 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5594 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5598 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5601 put_cpu_no_resched();
5606 * /proc/perfmon interface, for debug only
5609 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5612 pfm_proc_start(struct seq_file *m, loff_t *pos)
5615 return PFM_PROC_SHOW_HEADER;
5618 while (*pos <= NR_CPUS) {
5619 if (cpu_online(*pos - 1)) {
5620 return (void *)*pos;
5628 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5631 return pfm_proc_start(m, pos);
5635 pfm_proc_stop(struct seq_file *m, void *v)
5640 pfm_proc_show_header(struct seq_file *m)
5642 struct list_head * pos;
5643 pfm_buffer_fmt_t * entry;
5644 unsigned long flags;
5647 "perfmon version : %u.%u\n"
5650 "expert mode : %s\n"
5651 "ovfl_mask : 0x%lx\n"
5652 "PMU flags : 0x%x\n",
5653 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5655 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5656 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5663 "proc_sessions : %u\n"
5664 "sys_sessions : %u\n"
5665 "sys_use_dbregs : %u\n"
5666 "ptrace_use_dbregs : %u\n",
5667 pfm_sessions.pfs_task_sessions,
5668 pfm_sessions.pfs_sys_sessions,
5669 pfm_sessions.pfs_sys_use_dbregs,
5670 pfm_sessions.pfs_ptrace_use_dbregs);
5674 spin_lock(&pfm_buffer_fmt_lock);
5676 list_for_each(pos, &pfm_buffer_fmt_list) {
5677 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5678 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5689 entry->fmt_uuid[10],
5690 entry->fmt_uuid[11],
5691 entry->fmt_uuid[12],
5692 entry->fmt_uuid[13],
5693 entry->fmt_uuid[14],
5694 entry->fmt_uuid[15],
5697 spin_unlock(&pfm_buffer_fmt_lock);
5702 pfm_proc_show(struct seq_file *m, void *v)
5708 if (v == PFM_PROC_SHOW_HEADER) {
5709 pfm_proc_show_header(m);
5713 /* show info for CPU (v - 1) */
5717 "CPU%-2d overflow intrs : %lu\n"
5718 "CPU%-2d overflow cycles : %lu\n"
5719 "CPU%-2d overflow min : %lu\n"
5720 "CPU%-2d overflow max : %lu\n"
5721 "CPU%-2d smpl handler calls : %lu\n"
5722 "CPU%-2d smpl handler cycles : %lu\n"
5723 "CPU%-2d spurious intrs : %lu\n"
5724 "CPU%-2d replay intrs : %lu\n"
5725 "CPU%-2d syst_wide : %d\n"
5726 "CPU%-2d dcr_pp : %d\n"
5727 "CPU%-2d exclude idle : %d\n"
5728 "CPU%-2d owner : %d\n"
5729 "CPU%-2d context : %p\n"
5730 "CPU%-2d activations : %lu\n",
5731 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5732 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5733 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5734 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5735 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5736 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5737 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5738 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5739 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5740 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5741 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5742 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5743 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5744 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5746 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5748 psr = pfm_get_psr();
5753 "CPU%-2d psr : 0x%lx\n"
5754 "CPU%-2d pmc0 : 0x%lx\n",
5756 cpu, ia64_get_pmc(0));
5758 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5759 if (PMC_IS_COUNTING(i) == 0) continue;
5761 "CPU%-2d pmc%u : 0x%lx\n"
5762 "CPU%-2d pmd%u : 0x%lx\n",
5763 cpu, i, ia64_get_pmc(i),
5764 cpu, i, ia64_get_pmd(i));
5770 struct seq_operations pfm_seq_ops = {
5771 .start = pfm_proc_start,
5772 .next = pfm_proc_next,
5773 .stop = pfm_proc_stop,
5774 .show = pfm_proc_show
5778 pfm_proc_open(struct inode *inode, struct file *file)
5780 return seq_open(file, &pfm_seq_ops);
5785 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5786 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5787 * is active or inactive based on mode. We must rely on the value in
5788 * local_cpu_data->pfm_syst_info
5791 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5793 struct pt_regs *regs;
5795 unsigned long dcr_pp;
5797 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5800 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5801 * on every CPU, so we can rely on the pid to identify the idle task.
5803 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5804 regs = task_pt_regs(task);
5805 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5809 * if monitoring has started
5812 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5814 * context switching in?
5817 /* mask monitoring for the idle task */
5818 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5824 * context switching out
5825 * restore monitoring for next task
5827 * Due to inlining this odd if-then-else construction generates
5830 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5839 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5841 struct task_struct *task = ctx->ctx_task;
5843 ia64_psr(regs)->up = 0;
5844 ia64_psr(regs)->sp = 1;
5846 if (GET_PMU_OWNER() == task) {
5847 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5848 SET_PMU_OWNER(NULL, NULL);
5852 * disconnect the task from the context and vice-versa
5854 PFM_SET_WORK_PENDING(task, 0);
5856 task->thread.pfm_context = NULL;
5857 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5859 DPRINT(("force cleanup for [%d]\n", task->pid));
5864 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5867 pfm_save_regs(struct task_struct *task)
5870 struct thread_struct *t;
5871 unsigned long flags;
5875 ctx = PFM_GET_CTX(task);
5876 if (ctx == NULL) return;
5880 * we always come here with interrupts ALREADY disabled by
5881 * the scheduler. So we simply need to protect against concurrent
5882 * access, not CPU concurrency.
5884 flags = pfm_protect_ctx_ctxsw(ctx);
5886 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5887 struct pt_regs *regs = task_pt_regs(task);
5891 pfm_force_cleanup(ctx, regs);
5893 BUG_ON(ctx->ctx_smpl_hdr);
5895 pfm_unprotect_ctx_ctxsw(ctx, flags);
5897 pfm_context_free(ctx);
5902 * save current PSR: needed because we modify it
5905 psr = pfm_get_psr();
5907 BUG_ON(psr & (IA64_PSR_I));
5911 * This is the last instruction which may generate an overflow
5913 * We do not need to set psr.sp because, it is irrelevant in kernel.
5914 * It will be restored from ipsr when going back to user level
5919 * keep a copy of psr.up (for reload)
5921 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5924 * release ownership of this PMU.
5925 * PM interrupts are masked, so nothing
5928 SET_PMU_OWNER(NULL, NULL);
5931 * we systematically save the PMD as we have no
5932 * guarantee we will be schedule at that same
5935 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5938 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5939 * we will need it on the restore path to check
5940 * for pending overflow.
5942 t->pmcs[0] = ia64_get_pmc(0);
5945 * unfreeze PMU if had pending overflows
5947 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5950 * finally, allow context access.
5951 * interrupts will still be masked after this call.
5953 pfm_unprotect_ctx_ctxsw(ctx, flags);
5956 #else /* !CONFIG_SMP */
5958 pfm_save_regs(struct task_struct *task)
5963 ctx = PFM_GET_CTX(task);
5964 if (ctx == NULL) return;
5967 * save current PSR: needed because we modify it
5969 psr = pfm_get_psr();
5971 BUG_ON(psr & (IA64_PSR_I));
5975 * This is the last instruction which may generate an overflow
5977 * We do not need to set psr.sp because, it is irrelevant in kernel.
5978 * It will be restored from ipsr when going back to user level
5983 * keep a copy of psr.up (for reload)
5985 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5989 pfm_lazy_save_regs (struct task_struct *task)
5992 struct thread_struct *t;
5993 unsigned long flags;
5995 { u64 psr = pfm_get_psr();
5996 BUG_ON(psr & IA64_PSR_UP);
5999 ctx = PFM_GET_CTX(task);
6003 * we need to mask PMU overflow here to
6004 * make sure that we maintain pmc0 until
6005 * we save it. overflow interrupts are
6006 * treated as spurious if there is no
6009 * XXX: I don't think this is necessary
6011 PROTECT_CTX(ctx,flags);
6014 * release ownership of this PMU.
6015 * must be done before we save the registers.
6017 * after this call any PMU interrupt is treated
6020 SET_PMU_OWNER(NULL, NULL);
6023 * save all the pmds we use
6025 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
6028 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6029 * it is needed to check for pended overflow
6030 * on the restore path
6032 t->pmcs[0] = ia64_get_pmc(0);
6035 * unfreeze PMU if had pending overflows
6037 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6040 * now get can unmask PMU interrupts, they will
6041 * be treated as purely spurious and we will not
6042 * lose any information
6044 UNPROTECT_CTX(ctx,flags);
6046 #endif /* CONFIG_SMP */
6050 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6053 pfm_load_regs (struct task_struct *task)
6056 struct thread_struct *t;
6057 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6058 unsigned long flags;
6060 int need_irq_resend;
6062 ctx = PFM_GET_CTX(task);
6063 if (unlikely(ctx == NULL)) return;
6065 BUG_ON(GET_PMU_OWNER());
6069 * possible on unload
6071 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6074 * we always come here with interrupts ALREADY disabled by
6075 * the scheduler. So we simply need to protect against concurrent
6076 * access, not CPU concurrency.
6078 flags = pfm_protect_ctx_ctxsw(ctx);
6079 psr = pfm_get_psr();
6081 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6083 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6084 BUG_ON(psr & IA64_PSR_I);
6086 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6087 struct pt_regs *regs = task_pt_regs(task);
6089 BUG_ON(ctx->ctx_smpl_hdr);
6091 pfm_force_cleanup(ctx, regs);
6093 pfm_unprotect_ctx_ctxsw(ctx, flags);
6096 * this one (kmalloc'ed) is fine with interrupts disabled
6098 pfm_context_free(ctx);
6104 * we restore ALL the debug registers to avoid picking up
6107 if (ctx->ctx_fl_using_dbreg) {
6108 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6109 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6112 * retrieve saved psr.up
6114 psr_up = ctx->ctx_saved_psr_up;
6117 * if we were the last user of the PMU on that CPU,
6118 * then nothing to do except restore psr
6120 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6123 * retrieve partial reload masks (due to user modifications)
6125 pmc_mask = ctx->ctx_reload_pmcs[0];
6126 pmd_mask = ctx->ctx_reload_pmds[0];
6130 * To avoid leaking information to the user level when psr.sp=0,
6131 * we must reload ALL implemented pmds (even the ones we don't use).
6132 * In the kernel we only allow PFM_READ_PMDS on registers which
6133 * we initialized or requested (sampling) so there is no risk there.
6135 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6138 * ALL accessible PMCs are systematically reloaded, unused registers
6139 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6140 * up stale configuration.
6142 * PMC0 is never in the mask. It is always restored separately.
6144 pmc_mask = ctx->ctx_all_pmcs[0];
6147 * when context is MASKED, we will restore PMC with plm=0
6148 * and PMD with stale information, but that's ok, nothing
6151 * XXX: optimize here
6153 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6154 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6157 * check for pending overflow at the time the state
6160 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6162 * reload pmc0 with the overflow information
6163 * On McKinley PMU, this will trigger a PMU interrupt
6165 ia64_set_pmc(0, t->pmcs[0]);
6170 * will replay the PMU interrupt
6172 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6174 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6178 * we just did a reload, so we reset the partial reload fields
6180 ctx->ctx_reload_pmcs[0] = 0UL;
6181 ctx->ctx_reload_pmds[0] = 0UL;
6183 SET_LAST_CPU(ctx, smp_processor_id());
6186 * dump activation value for this PMU
6190 * record current activation for this context
6192 SET_ACTIVATION(ctx);
6195 * establish new ownership.
6197 SET_PMU_OWNER(task, ctx);
6200 * restore the psr.up bit. measurement
6202 * no PMU interrupt can happen at this point
6203 * because we still have interrupts disabled.
6205 if (likely(psr_up)) pfm_set_psr_up();
6208 * allow concurrent access to context
6210 pfm_unprotect_ctx_ctxsw(ctx, flags);
6212 #else /* !CONFIG_SMP */
6214 * reload PMU state for UP kernels
6215 * in 2.5 we come here with interrupts disabled
6218 pfm_load_regs (struct task_struct *task)
6220 struct thread_struct *t;
6222 struct task_struct *owner;
6223 unsigned long pmd_mask, pmc_mask;
6225 int need_irq_resend;
6227 owner = GET_PMU_OWNER();
6228 ctx = PFM_GET_CTX(task);
6230 psr = pfm_get_psr();
6232 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6233 BUG_ON(psr & IA64_PSR_I);
6236 * we restore ALL the debug registers to avoid picking up
6239 * This must be done even when the task is still the owner
6240 * as the registers may have been modified via ptrace()
6241 * (not perfmon) by the previous task.
6243 if (ctx->ctx_fl_using_dbreg) {
6244 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6245 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6249 * retrieved saved psr.up
6251 psr_up = ctx->ctx_saved_psr_up;
6252 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6255 * short path, our state is still there, just
6256 * need to restore psr and we go
6258 * we do not touch either PMC nor PMD. the psr is not touched
6259 * by the overflow_handler. So we are safe w.r.t. to interrupt
6260 * concurrency even without interrupt masking.
6262 if (likely(owner == task)) {
6263 if (likely(psr_up)) pfm_set_psr_up();
6268 * someone else is still using the PMU, first push it out and
6269 * then we'll be able to install our stuff !
6271 * Upon return, there will be no owner for the current PMU
6273 if (owner) pfm_lazy_save_regs(owner);
6276 * To avoid leaking information to the user level when psr.sp=0,
6277 * we must reload ALL implemented pmds (even the ones we don't use).
6278 * In the kernel we only allow PFM_READ_PMDS on registers which
6279 * we initialized or requested (sampling) so there is no risk there.
6281 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6284 * ALL accessible PMCs are systematically reloaded, unused registers
6285 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6286 * up stale configuration.
6288 * PMC0 is never in the mask. It is always restored separately
6290 pmc_mask = ctx->ctx_all_pmcs[0];
6292 pfm_restore_pmds(t->pmds, pmd_mask);
6293 pfm_restore_pmcs(t->pmcs, pmc_mask);
6296 * check for pending overflow at the time the state
6299 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6301 * reload pmc0 with the overflow information
6302 * On McKinley PMU, this will trigger a PMU interrupt
6304 ia64_set_pmc(0, t->pmcs[0]);
6310 * will replay the PMU interrupt
6312 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6314 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6318 * establish new ownership.
6320 SET_PMU_OWNER(task, ctx);
6323 * restore the psr.up bit. measurement
6325 * no PMU interrupt can happen at this point
6326 * because we still have interrupts disabled.
6328 if (likely(psr_up)) pfm_set_psr_up();
6330 #endif /* CONFIG_SMP */
6333 * this function assumes monitoring is stopped
6336 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6339 unsigned long mask2, val, pmd_val, ovfl_val;
6340 int i, can_access_pmu = 0;
6344 * is the caller the task being monitored (or which initiated the
6345 * session for system wide measurements)
6347 is_self = ctx->ctx_task == task ? 1 : 0;
6350 * can access PMU is task is the owner of the PMU state on the current CPU
6351 * or if we are running on the CPU bound to the context in system-wide mode
6352 * (that is not necessarily the task the context is attached to in this mode).
6353 * In system-wide we always have can_access_pmu true because a task running on an
6354 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6356 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6357 if (can_access_pmu) {
6359 * Mark the PMU as not owned
6360 * This will cause the interrupt handler to do nothing in case an overflow
6361 * interrupt was in-flight
6362 * This also guarantees that pmc0 will contain the final state
6363 * It virtually gives us full control on overflow processing from that point
6366 SET_PMU_OWNER(NULL, NULL);
6367 DPRINT(("releasing ownership\n"));
6370 * read current overflow status:
6372 * we are guaranteed to read the final stable state
6375 pmc0 = ia64_get_pmc(0); /* slow */
6378 * reset freeze bit, overflow status information destroyed
6382 pmc0 = task->thread.pmcs[0];
6384 * clear whatever overflow status bits there were
6386 task->thread.pmcs[0] = 0;
6388 ovfl_val = pmu_conf->ovfl_val;
6390 * we save all the used pmds
6391 * we take care of overflows for counting PMDs
6393 * XXX: sampling situation is not taken into account here
6395 mask2 = ctx->ctx_used_pmds[0];
6397 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6399 for (i = 0; mask2; i++, mask2>>=1) {
6401 /* skip non used pmds */
6402 if ((mask2 & 0x1) == 0) continue;
6405 * can access PMU always true in system wide mode
6407 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6409 if (PMD_IS_COUNTING(i)) {
6410 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6413 ctx->ctx_pmds[i].val,
6417 * we rebuild the full 64 bit value of the counter
6419 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6422 * now everything is in ctx_pmds[] and we need
6423 * to clear the saved context from save_regs() such that
6424 * pfm_read_pmds() gets the correct value
6429 * take care of overflow inline
6431 if (pmc0 & (1UL << i)) {
6432 val += 1 + ovfl_val;
6433 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6437 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, i, val, pmd_val));
6439 if (is_self) task->thread.pmds[i] = pmd_val;
6441 ctx->ctx_pmds[i].val = val;
6445 static struct irqaction perfmon_irqaction = {
6446 .handler = pfm_interrupt_handler,
6447 .flags = SA_INTERRUPT,
6452 pfm_alt_save_pmu_state(void *data)
6454 struct pt_regs *regs;
6456 regs = task_pt_regs(current);
6458 DPRINT(("called\n"));
6461 * should not be necessary but
6462 * let's take not risk
6466 ia64_psr(regs)->pp = 0;
6469 * This call is required
6470 * May cause a spurious interrupt on some processors
6478 pfm_alt_restore_pmu_state(void *data)
6480 struct pt_regs *regs;
6482 regs = task_pt_regs(current);
6484 DPRINT(("called\n"));
6487 * put PMU back in state expected
6492 ia64_psr(regs)->pp = 0;
6495 * perfmon runs with PMU unfrozen at all times
6503 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6508 /* some sanity checks */
6509 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6511 /* do the easy test first */
6512 if (pfm_alt_intr_handler) return -EBUSY;
6514 /* one at a time in the install or remove, just fail the others */
6515 if (!spin_trylock(&pfm_alt_install_check)) {
6519 /* reserve our session */
6520 for_each_online_cpu(reserve_cpu) {
6521 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6522 if (ret) goto cleanup_reserve;
6525 /* save the current system wide pmu states */
6526 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 0, 1);
6528 DPRINT(("on_each_cpu() failed: %d\n", ret));
6529 goto cleanup_reserve;
6532 /* officially change to the alternate interrupt handler */
6533 pfm_alt_intr_handler = hdl;
6535 spin_unlock(&pfm_alt_install_check);
6540 for_each_online_cpu(i) {
6541 /* don't unreserve more than we reserved */
6542 if (i >= reserve_cpu) break;
6544 pfm_unreserve_session(NULL, 1, i);
6547 spin_unlock(&pfm_alt_install_check);
6551 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6554 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6559 if (hdl == NULL) return -EINVAL;
6561 /* cannot remove someone else's handler! */
6562 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6564 /* one at a time in the install or remove, just fail the others */
6565 if (!spin_trylock(&pfm_alt_install_check)) {
6569 pfm_alt_intr_handler = NULL;
6571 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 0, 1);
6573 DPRINT(("on_each_cpu() failed: %d\n", ret));
6576 for_each_online_cpu(i) {
6577 pfm_unreserve_session(NULL, 1, i);
6580 spin_unlock(&pfm_alt_install_check);
6584 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6587 * perfmon initialization routine, called from the initcall() table
6589 static int init_pfm_fs(void);
6597 family = local_cpu_data->family;
6602 if ((*p)->probe() == 0) goto found;
6603 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6614 static struct file_operations pfm_proc_fops = {
6615 .open = pfm_proc_open,
6617 .llseek = seq_lseek,
6618 .release = seq_release,
6624 unsigned int n, n_counters, i;
6626 printk("perfmon: version %u.%u IRQ %u\n",
6629 IA64_PERFMON_VECTOR);
6631 if (pfm_probe_pmu()) {
6632 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6633 local_cpu_data->family);
6638 * compute the number of implemented PMD/PMC from the
6639 * description tables
6642 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6643 if (PMC_IS_IMPL(i) == 0) continue;
6644 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6647 pmu_conf->num_pmcs = n;
6649 n = 0; n_counters = 0;
6650 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6651 if (PMD_IS_IMPL(i) == 0) continue;
6652 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6654 if (PMD_IS_COUNTING(i)) n_counters++;
6656 pmu_conf->num_pmds = n;
6657 pmu_conf->num_counters = n_counters;
6660 * sanity checks on the number of debug registers
6662 if (pmu_conf->use_rr_dbregs) {
6663 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6664 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6668 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6669 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6675 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6679 pmu_conf->num_counters,
6680 ffz(pmu_conf->ovfl_val));
6683 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6684 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6690 * create /proc/perfmon (mostly for debugging purposes)
6692 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6693 if (perfmon_dir == NULL) {
6694 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6699 * install customized file operations for /proc/perfmon entry
6701 perfmon_dir->proc_fops = &pfm_proc_fops;
6704 * create /proc/sys/kernel/perfmon (for debugging purposes)
6706 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6709 * initialize all our spinlocks
6711 spin_lock_init(&pfm_sessions.pfs_lock);
6712 spin_lock_init(&pfm_buffer_fmt_lock);
6716 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6721 __initcall(pfm_init);
6724 * this function is called before pfm_init()
6727 pfm_init_percpu (void)
6730 * make sure no measurement is active
6731 * (may inherit programmed PMCs from EFI).
6737 * we run with the PMU not frozen at all times
6741 if (smp_processor_id() == 0)
6742 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6744 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6749 * used for debug purposes only
6752 dump_pmu_state(const char *from)
6754 struct task_struct *task;
6755 struct thread_struct *t;
6756 struct pt_regs *regs;
6758 unsigned long psr, dcr, info, flags;
6761 local_irq_save(flags);
6763 this_cpu = smp_processor_id();
6764 regs = task_pt_regs(current);
6765 info = PFM_CPUINFO_GET();
6766 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6768 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6769 local_irq_restore(flags);
6773 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6780 task = GET_PMU_OWNER();
6781 ctx = GET_PMU_CTX();
6783 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6785 psr = pfm_get_psr();
6787 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",
6790 psr & IA64_PSR_PP ? 1 : 0,
6791 psr & IA64_PSR_UP ? 1 : 0,
6792 dcr & IA64_DCR_PP ? 1 : 0,
6795 ia64_psr(regs)->pp);
6797 ia64_psr(regs)->up = 0;
6798 ia64_psr(regs)->pp = 0;
6800 t = ¤t->thread;
6802 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6803 if (PMC_IS_IMPL(i) == 0) continue;
6804 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6807 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6808 if (PMD_IS_IMPL(i) == 0) continue;
6809 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6813 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6816 ctx->ctx_smpl_vaddr,
6820 ctx->ctx_saved_psr_up);
6822 local_irq_restore(flags);
6826 * called from process.c:copy_thread(). task is new child.
6829 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6831 struct thread_struct *thread;
6833 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6835 thread = &task->thread;
6838 * cut links inherited from parent (current)
6840 thread->pfm_context = NULL;
6842 PFM_SET_WORK_PENDING(task, 0);
6845 * the psr bits are already set properly in copy_threads()
6848 #else /* !CONFIG_PERFMON */
6850 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6854 #endif /* CONFIG_PERFMON */