2 * This file implements the perfmon-2 subsystem which is used
3 * to program the IA-64 Performance Monitoring Unit (PMU).
5 * The initial version of perfmon.c was written by
6 * Ganesh Venkitachalam, IBM Corp.
8 * Then it was modified for perfmon-1.x by Stephane Eranian and
9 * David Mosberger, Hewlett Packard Co.
11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
12 * by Stephane Eranian, Hewlett Packard Co.
14 * Copyright (C) 1999-2005 Hewlett Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 * David Mosberger-Tang <davidm@hpl.hp.com>
18 * More information about perfmon available at:
19 * http://www.hpl.hp.com/research/linux/perfmon
22 #include <linux/module.h>
23 #include <linux/kernel.h>
24 #include <linux/sched.h>
25 #include <linux/interrupt.h>
26 #include <linux/smp_lock.h>
27 #include <linux/proc_fs.h>
28 #include <linux/seq_file.h>
29 #include <linux/init.h>
30 #include <linux/vmalloc.h>
32 #include <linux/sysctl.h>
33 #include <linux/list.h>
34 #include <linux/file.h>
35 #include <linux/poll.h>
36 #include <linux/vfs.h>
37 #include <linux/smp.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>
45 #include <asm/errno.h>
46 #include <asm/intrinsics.h>
48 #include <asm/perfmon.h>
49 #include <asm/processor.h>
50 #include <asm/signal.h>
51 #include <asm/system.h>
52 #include <asm/uaccess.h>
53 #include <asm/delay.h>
57 * perfmon context state
59 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
60 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
61 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
62 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
64 #define PFM_INVALID_ACTIVATION (~0UL)
66 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
67 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
70 * depth of message queue
72 #define PFM_MAX_MSGS 32
73 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
76 * type of a PMU register (bitmask).
78 * bit0 : register implemented
81 * bit4 : pmc has pmc.pm
82 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
83 * bit6-7 : register type
86 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
87 #define PFM_REG_IMPL 0x1 /* register implemented */
88 #define PFM_REG_END 0x2 /* end marker */
89 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
90 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
91 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
92 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
93 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
95 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
96 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
98 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
100 /* i assumed unsigned */
101 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
102 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
104 /* XXX: these assume that register i is implemented */
105 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
106 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
107 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
108 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
110 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
111 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
112 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
113 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
115 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
116 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
118 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
119 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
120 #define PFM_CTX_TASK(h) (h)->ctx_task
122 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
124 /* XXX: does not support more than 64 PMDs */
125 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
126 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
128 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
130 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
131 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
132 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
133 #define PFM_CODE_RR 0 /* requesting code range restriction */
134 #define PFM_DATA_RR 1 /* requestion data range restriction */
136 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
137 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
138 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
140 #define RDEP(x) (1UL<<(x))
143 * context protection macros
145 * - we need to protect against CPU concurrency (spin_lock)
146 * - we need to protect against PMU overflow interrupts (local_irq_disable)
148 * - we need to protect against PMU overflow interrupts (local_irq_disable)
150 * spin_lock_irqsave()/spin_lock_irqrestore():
151 * in SMP: local_irq_disable + spin_lock
152 * in UP : local_irq_disable
154 * spin_lock()/spin_lock():
155 * in UP : removed automatically
156 * in SMP: protect against context accesses from other CPU. interrupts
157 * are not masked. This is useful for the PMU interrupt handler
158 * because we know we will not get PMU concurrency in that code.
160 #define PROTECT_CTX(c, f) \
162 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \
163 spin_lock_irqsave(&(c)->ctx_lock, f); \
164 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
167 #define UNPROTECT_CTX(c, f) \
169 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
170 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
173 #define PROTECT_CTX_NOPRINT(c, f) \
175 spin_lock_irqsave(&(c)->ctx_lock, f); \
179 #define UNPROTECT_CTX_NOPRINT(c, f) \
181 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
185 #define PROTECT_CTX_NOIRQ(c) \
187 spin_lock(&(c)->ctx_lock); \
190 #define UNPROTECT_CTX_NOIRQ(c) \
192 spin_unlock(&(c)->ctx_lock); \
198 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
199 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
200 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
202 #else /* !CONFIG_SMP */
203 #define SET_ACTIVATION(t) do {} while(0)
204 #define GET_ACTIVATION(t) do {} while(0)
205 #define INC_ACTIVATION(t) do {} while(0)
206 #endif /* CONFIG_SMP */
208 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
209 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
210 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
212 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
213 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
215 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
218 * cmp0 must be the value of pmc0
220 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
222 #define PFMFS_MAGIC 0xa0b4d889
227 #define PFM_DEBUGGING 1
231 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
234 #define DPRINT_ovfl(a) \
236 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; } \
241 * 64-bit software counter structure
243 * the next_reset_type is applied to the next call to pfm_reset_regs()
246 unsigned long val; /* virtual 64bit counter value */
247 unsigned long lval; /* last reset value */
248 unsigned long long_reset; /* reset value on sampling overflow */
249 unsigned long short_reset; /* reset value on overflow */
250 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
251 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
252 unsigned long seed; /* seed for random-number generator */
253 unsigned long mask; /* mask for random-number generator */
254 unsigned int flags; /* notify/do not notify */
255 unsigned long eventid; /* overflow event identifier */
262 unsigned int block:1; /* when 1, task will blocked on user notifications */
263 unsigned int system:1; /* do system wide monitoring */
264 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
265 unsigned int is_sampling:1; /* true if using a custom format */
266 unsigned int excl_idle:1; /* exclude idle task in system wide session */
267 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
268 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
269 unsigned int no_msg:1; /* no message sent on overflow */
270 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
271 unsigned int reserved:22;
272 } pfm_context_flags_t;
274 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
275 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
276 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
280 * perfmon context: encapsulates all the state of a monitoring session
283 typedef struct pfm_context {
284 spinlock_t ctx_lock; /* context protection */
286 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
287 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
289 struct task_struct *ctx_task; /* task to which context is attached */
291 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
293 struct completion ctx_restart_done; /* use for blocking notification mode */
295 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
296 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
297 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
299 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
300 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
301 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
303 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
305 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
306 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
307 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
308 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
310 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
312 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
313 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
315 u64 ctx_saved_psr_up; /* only contains psr.up value */
317 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
318 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
319 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
321 int ctx_fd; /* file descriptor used my this context */
322 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
324 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
325 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
326 unsigned long ctx_smpl_size; /* size of sampling buffer */
327 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
329 wait_queue_head_t ctx_msgq_wait;
330 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
333 struct fasync_struct *ctx_async_queue;
335 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
339 * magic number used to verify that structure is really
342 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
344 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
347 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
348 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
350 #define SET_LAST_CPU(ctx, v) do {} while(0)
351 #define GET_LAST_CPU(ctx) do {} while(0)
355 #define ctx_fl_block ctx_flags.block
356 #define ctx_fl_system ctx_flags.system
357 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
358 #define ctx_fl_is_sampling ctx_flags.is_sampling
359 #define ctx_fl_excl_idle ctx_flags.excl_idle
360 #define ctx_fl_going_zombie ctx_flags.going_zombie
361 #define ctx_fl_trap_reason ctx_flags.trap_reason
362 #define ctx_fl_no_msg ctx_flags.no_msg
363 #define ctx_fl_can_restart ctx_flags.can_restart
365 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
366 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
369 * global information about all sessions
370 * mostly used to synchronize between system wide and per-process
373 spinlock_t pfs_lock; /* lock the structure */
375 unsigned int pfs_task_sessions; /* number of per task sessions */
376 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
377 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
378 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
379 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
383 * information about a PMC or PMD.
384 * dep_pmd[]: a bitmask of dependent PMD registers
385 * dep_pmc[]: a bitmask of dependent PMC registers
387 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
391 unsigned long default_value; /* power-on default value */
392 unsigned long reserved_mask; /* bitmask of reserved bits */
393 pfm_reg_check_t read_check;
394 pfm_reg_check_t write_check;
395 unsigned long dep_pmd[4];
396 unsigned long dep_pmc[4];
399 /* assume cnum is a valid monitor */
400 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
403 * This structure is initialized at boot time and contains
404 * a description of the PMU main characteristics.
406 * If the probe function is defined, detection is based
407 * on its return value:
408 * - 0 means recognized PMU
409 * - anything else means not supported
410 * When the probe function is not defined, then the pmu_family field
411 * is used and it must match the host CPU family such that:
412 * - cpu->family & config->pmu_family != 0
415 unsigned long ovfl_val; /* overflow value for counters */
417 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
418 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
420 unsigned int num_pmcs; /* number of PMCS: computed at init time */
421 unsigned int num_pmds; /* number of PMDS: computed at init time */
422 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
423 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
425 char *pmu_name; /* PMU family name */
426 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
427 unsigned int flags; /* pmu specific flags */
428 unsigned int num_ibrs; /* number of IBRS: computed at init time */
429 unsigned int num_dbrs; /* number of DBRS: computed at init time */
430 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
431 int (*probe)(void); /* customized probe routine */
432 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
437 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
440 * debug register related type definitions
443 unsigned long ibr_mask:56;
444 unsigned long ibr_plm:4;
445 unsigned long ibr_ig:3;
446 unsigned long ibr_x:1;
450 unsigned long dbr_mask:56;
451 unsigned long dbr_plm:4;
452 unsigned long dbr_ig:2;
453 unsigned long dbr_w:1;
454 unsigned long dbr_r:1;
465 * perfmon command descriptions
468 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
471 unsigned int cmd_narg;
473 int (*cmd_getsize)(void *arg, size_t *sz);
476 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
477 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
478 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
479 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
482 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
483 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
484 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
485 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
486 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
488 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
491 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
492 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
493 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
496 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
497 unsigned long pfm_smpl_handler_calls;
498 unsigned long pfm_smpl_handler_cycles;
499 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
503 * perfmon internal variables
505 static pfm_stats_t pfm_stats[NR_CPUS];
506 static pfm_session_t pfm_sessions; /* global sessions information */
508 static DEFINE_SPINLOCK(pfm_alt_install_check);
509 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
511 static struct proc_dir_entry *perfmon_dir;
512 static pfm_uuid_t pfm_null_uuid = {0,};
514 static spinlock_t pfm_buffer_fmt_lock;
515 static LIST_HEAD(pfm_buffer_fmt_list);
517 static pmu_config_t *pmu_conf;
519 /* sysctl() controls */
520 pfm_sysctl_t pfm_sysctl;
521 EXPORT_SYMBOL(pfm_sysctl);
523 static ctl_table pfm_ctl_table[]={
524 {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
525 {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
526 {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
527 {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
530 static ctl_table pfm_sysctl_dir[] = {
531 {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
534 static ctl_table pfm_sysctl_root[] = {
535 {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
538 static struct ctl_table_header *pfm_sysctl_header;
540 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
542 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
543 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
546 pfm_put_task(struct task_struct *task)
548 if (task != current) put_task_struct(task);
552 pfm_set_task_notify(struct task_struct *task)
554 struct thread_info *info;
556 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
557 set_bit(TIF_NOTIFY_RESUME, &info->flags);
561 pfm_clear_task_notify(void)
563 clear_thread_flag(TIF_NOTIFY_RESUME);
567 pfm_reserve_page(unsigned long a)
569 SetPageReserved(vmalloc_to_page((void *)a));
572 pfm_unreserve_page(unsigned long a)
574 ClearPageReserved(vmalloc_to_page((void*)a));
577 static inline unsigned long
578 pfm_protect_ctx_ctxsw(pfm_context_t *x)
580 spin_lock(&(x)->ctx_lock);
585 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
587 spin_unlock(&(x)->ctx_lock);
590 static inline unsigned int
591 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
593 return do_munmap(mm, addr, len);
596 static inline unsigned long
597 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
599 return get_unmapped_area(file, addr, len, pgoff, flags);
604 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data,
605 struct vfsmount *mnt)
607 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
610 static struct file_system_type pfm_fs_type = {
612 .get_sb = pfmfs_get_sb,
613 .kill_sb = kill_anon_super,
616 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
617 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
618 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
619 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
620 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
623 /* forward declaration */
624 static struct file_operations pfm_file_ops;
627 * forward declarations
630 static void pfm_lazy_save_regs (struct task_struct *ta);
633 void dump_pmu_state(const char *);
634 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
636 #include "perfmon_itanium.h"
637 #include "perfmon_mckinley.h"
638 #include "perfmon_montecito.h"
639 #include "perfmon_generic.h"
641 static pmu_config_t *pmu_confs[]={
645 &pmu_conf_gen, /* must be last */
650 static int pfm_end_notify_user(pfm_context_t *ctx);
653 pfm_clear_psr_pp(void)
655 ia64_rsm(IA64_PSR_PP);
662 ia64_ssm(IA64_PSR_PP);
667 pfm_clear_psr_up(void)
669 ia64_rsm(IA64_PSR_UP);
676 ia64_ssm(IA64_PSR_UP);
680 static inline unsigned long
684 tmp = ia64_getreg(_IA64_REG_PSR);
690 pfm_set_psr_l(unsigned long val)
692 ia64_setreg(_IA64_REG_PSR_L, val);
704 pfm_unfreeze_pmu(void)
711 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
715 for (i=0; i < nibrs; i++) {
716 ia64_set_ibr(i, ibrs[i]);
717 ia64_dv_serialize_instruction();
723 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
727 for (i=0; i < ndbrs; i++) {
728 ia64_set_dbr(i, dbrs[i]);
729 ia64_dv_serialize_data();
735 * PMD[i] must be a counter. no check is made
737 static inline unsigned long
738 pfm_read_soft_counter(pfm_context_t *ctx, int i)
740 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
744 * PMD[i] must be a counter. no check is made
747 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
749 unsigned long ovfl_val = pmu_conf->ovfl_val;
751 ctx->ctx_pmds[i].val = val & ~ovfl_val;
753 * writing to unimplemented part is ignore, so we do not need to
756 ia64_set_pmd(i, val & ovfl_val);
760 pfm_get_new_msg(pfm_context_t *ctx)
764 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
766 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
767 if (next == ctx->ctx_msgq_head) return NULL;
769 idx = ctx->ctx_msgq_tail;
770 ctx->ctx_msgq_tail = next;
772 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
774 return ctx->ctx_msgq+idx;
778 pfm_get_next_msg(pfm_context_t *ctx)
782 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
784 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
789 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
794 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
796 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));
802 pfm_reset_msgq(pfm_context_t *ctx)
804 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
805 DPRINT(("ctx=%p msgq reset\n", ctx));
809 pfm_rvmalloc(unsigned long size)
814 size = PAGE_ALIGN(size);
817 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
818 memset(mem, 0, size);
819 addr = (unsigned long)mem;
821 pfm_reserve_page(addr);
830 pfm_rvfree(void *mem, unsigned long size)
835 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
836 addr = (unsigned long) mem;
837 while ((long) size > 0) {
838 pfm_unreserve_page(addr);
847 static pfm_context_t *
848 pfm_context_alloc(void)
853 * allocate context descriptor
854 * must be able to free with interrupts disabled
856 ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
858 memset(ctx, 0, sizeof(pfm_context_t));
859 DPRINT(("alloc ctx @%p\n", ctx));
865 pfm_context_free(pfm_context_t *ctx)
868 DPRINT(("free ctx @%p\n", ctx));
874 pfm_mask_monitoring(struct task_struct *task)
876 pfm_context_t *ctx = PFM_GET_CTX(task);
877 unsigned long mask, val, ovfl_mask;
880 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
882 ovfl_mask = pmu_conf->ovfl_val;
884 * monitoring can only be masked as a result of a valid
885 * counter overflow. In UP, it means that the PMU still
886 * has an owner. Note that the owner can be different
887 * from the current task. However the PMU state belongs
889 * In SMP, a valid overflow only happens when task is
890 * current. Therefore if we come here, we know that
891 * the PMU state belongs to the current task, therefore
892 * we can access the live registers.
894 * So in both cases, the live register contains the owner's
895 * state. We can ONLY touch the PMU registers and NOT the PSR.
897 * As a consequence to this call, the ctx->th_pmds[] array
898 * contains stale information which must be ignored
899 * when context is reloaded AND monitoring is active (see
902 mask = ctx->ctx_used_pmds[0];
903 for (i = 0; mask; i++, mask>>=1) {
904 /* skip non used pmds */
905 if ((mask & 0x1) == 0) continue;
906 val = ia64_get_pmd(i);
908 if (PMD_IS_COUNTING(i)) {
910 * we rebuild the full 64 bit value of the counter
912 ctx->ctx_pmds[i].val += (val & ovfl_mask);
914 ctx->ctx_pmds[i].val = val;
916 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
918 ctx->ctx_pmds[i].val,
922 * mask monitoring by setting the privilege level to 0
923 * we cannot use psr.pp/psr.up for this, it is controlled by
926 * if task is current, modify actual registers, otherwise modify
927 * thread save state, i.e., what will be restored in pfm_load_regs()
929 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
930 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
931 if ((mask & 0x1) == 0UL) continue;
932 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
933 ctx->th_pmcs[i] &= ~0xfUL;
934 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
937 * make all of this visible
943 * must always be done with task == current
945 * context must be in MASKED state when calling
948 pfm_restore_monitoring(struct task_struct *task)
950 pfm_context_t *ctx = PFM_GET_CTX(task);
951 unsigned long mask, ovfl_mask;
952 unsigned long psr, val;
955 is_system = ctx->ctx_fl_system;
956 ovfl_mask = pmu_conf->ovfl_val;
958 if (task != current) {
959 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
962 if (ctx->ctx_state != PFM_CTX_MASKED) {
963 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
964 task->pid, current->pid, ctx->ctx_state);
969 * monitoring is masked via the PMC.
970 * As we restore their value, we do not want each counter to
971 * restart right away. We stop monitoring using the PSR,
972 * restore the PMC (and PMD) and then re-establish the psr
973 * as it was. Note that there can be no pending overflow at
974 * this point, because monitoring was MASKED.
976 * system-wide session are pinned and self-monitoring
978 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
980 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
986 * first, we restore the PMD
988 mask = ctx->ctx_used_pmds[0];
989 for (i = 0; mask; i++, mask>>=1) {
990 /* skip non used pmds */
991 if ((mask & 0x1) == 0) continue;
993 if (PMD_IS_COUNTING(i)) {
995 * we split the 64bit value according to
998 val = ctx->ctx_pmds[i].val & ovfl_mask;
999 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1001 val = ctx->ctx_pmds[i].val;
1003 ia64_set_pmd(i, val);
1005 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1007 ctx->ctx_pmds[i].val,
1013 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1014 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1015 if ((mask & 0x1) == 0UL) continue;
1016 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1017 ia64_set_pmc(i, ctx->th_pmcs[i]);
1018 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, ctx->th_pmcs[i]));
1023 * must restore DBR/IBR because could be modified while masked
1024 * XXX: need to optimize
1026 if (ctx->ctx_fl_using_dbreg) {
1027 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1028 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1034 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1036 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1043 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1049 for (i=0; mask; i++, mask>>=1) {
1050 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1055 * reload from thread state (used for ctxw only)
1058 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1061 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1063 for (i=0; mask; i++, mask>>=1) {
1064 if ((mask & 0x1) == 0) continue;
1065 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1066 ia64_set_pmd(i, val);
1072 * propagate PMD from context to thread-state
1075 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1077 unsigned long ovfl_val = pmu_conf->ovfl_val;
1078 unsigned long mask = ctx->ctx_all_pmds[0];
1082 DPRINT(("mask=0x%lx\n", mask));
1084 for (i=0; mask; i++, mask>>=1) {
1086 val = ctx->ctx_pmds[i].val;
1089 * We break up the 64 bit value into 2 pieces
1090 * the lower bits go to the machine state in the
1091 * thread (will be reloaded on ctxsw in).
1092 * The upper part stays in the soft-counter.
1094 if (PMD_IS_COUNTING(i)) {
1095 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1098 ctx->th_pmds[i] = val;
1100 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1103 ctx->ctx_pmds[i].val));
1108 * propagate PMC from context to thread-state
1111 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
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 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1121 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_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, fl_owner_t id)
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 mm->total_vm += 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 task_struct *task;
2866 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2867 unsigned long value, pmc_pm;
2868 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2869 unsigned int cnum, reg_flags, flags, pmc_type;
2870 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2871 int is_monitor, is_counting, state;
2873 pfm_reg_check_t wr_func;
2874 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2876 state = ctx->ctx_state;
2877 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2878 is_system = ctx->ctx_fl_system;
2879 task = ctx->ctx_task;
2880 impl_pmds = pmu_conf->impl_pmds[0];
2882 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2886 * In system wide and when the context is loaded, access can only happen
2887 * when the caller is running on the CPU being monitored by the session.
2888 * It does not have to be the owner (ctx_task) of the context per se.
2890 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2891 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2894 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2896 expert_mode = pfm_sysctl.expert_mode;
2898 for (i = 0; i < count; i++, req++) {
2900 cnum = req->reg_num;
2901 reg_flags = req->reg_flags;
2902 value = req->reg_value;
2903 smpl_pmds = req->reg_smpl_pmds[0];
2904 reset_pmds = req->reg_reset_pmds[0];
2908 if (cnum >= PMU_MAX_PMCS) {
2909 DPRINT(("pmc%u is invalid\n", cnum));
2913 pmc_type = pmu_conf->pmc_desc[cnum].type;
2914 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2915 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2916 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2919 * we reject all non implemented PMC as well
2920 * as attempts to modify PMC[0-3] which are used
2921 * as status registers by the PMU
2923 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2924 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2927 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2929 * If the PMC is a monitor, then if the value is not the default:
2930 * - system-wide session: PMCx.pm=1 (privileged monitor)
2931 * - per-task : PMCx.pm=0 (user monitor)
2933 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2934 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2943 * enforce generation of overflow interrupt. Necessary on all
2946 value |= 1 << PMU_PMC_OI;
2948 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2949 flags |= PFM_REGFL_OVFL_NOTIFY;
2952 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2954 /* verify validity of smpl_pmds */
2955 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2956 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2960 /* verify validity of reset_pmds */
2961 if ((reset_pmds & impl_pmds) != reset_pmds) {
2962 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2966 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2967 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2970 /* eventid on non-counting monitors are ignored */
2974 * execute write checker, if any
2976 if (likely(expert_mode == 0 && wr_func)) {
2977 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2978 if (ret) goto error;
2983 * no error on this register
2985 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2988 * Now we commit the changes to the software state
2992 * update overflow information
2996 * full flag update each time a register is programmed
2998 ctx->ctx_pmds[cnum].flags = flags;
3000 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
3001 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
3002 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
3005 * Mark all PMDS to be accessed as used.
3007 * We do not keep track of PMC because we have to
3008 * systematically restore ALL of them.
3010 * We do not update the used_monitors mask, because
3011 * if we have not programmed them, then will be in
3012 * a quiescent state, therefore we will not need to
3013 * mask/restore then when context is MASKED.
3015 CTX_USED_PMD(ctx, reset_pmds);
3016 CTX_USED_PMD(ctx, smpl_pmds);
3018 * make sure we do not try to reset on
3019 * restart because we have established new values
3021 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3024 * Needed in case the user does not initialize the equivalent
3025 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3026 * possible leak here.
3028 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3031 * keep track of the monitor PMC that we are using.
3032 * we save the value of the pmc in ctx_pmcs[] and if
3033 * the monitoring is not stopped for the context we also
3034 * place it in the saved state area so that it will be
3035 * picked up later by the context switch code.
3037 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3039 * The value in th_pmcs[] may be modified on overflow, i.e., when
3040 * monitoring needs to be stopped.
3042 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3045 * update context state
3047 ctx->ctx_pmcs[cnum] = value;
3051 * write thread state
3053 if (is_system == 0) ctx->th_pmcs[cnum] = value;
3056 * write hardware register if we can
3058 if (can_access_pmu) {
3059 ia64_set_pmc(cnum, value);
3064 * per-task SMP only here
3066 * we are guaranteed that the task is not running on the other CPU,
3067 * we indicate that this PMD will need to be reloaded if the task
3068 * is rescheduled on the CPU it ran last on.
3070 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3075 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",
3081 ctx->ctx_all_pmcs[0],
3082 ctx->ctx_used_pmds[0],
3083 ctx->ctx_pmds[cnum].eventid,
3086 ctx->ctx_reload_pmcs[0],
3087 ctx->ctx_used_monitors[0],
3088 ctx->ctx_ovfl_regs[0]));
3092 * make sure the changes are visible
3094 if (can_access_pmu) ia64_srlz_d();
3098 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3103 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3105 struct task_struct *task;
3106 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3107 unsigned long value, hw_value, ovfl_mask;
3109 int i, can_access_pmu = 0, state;
3110 int is_counting, is_loaded, is_system, expert_mode;
3112 pfm_reg_check_t wr_func;
3115 state = ctx->ctx_state;
3116 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3117 is_system = ctx->ctx_fl_system;
3118 ovfl_mask = pmu_conf->ovfl_val;
3119 task = ctx->ctx_task;
3121 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3124 * on both UP and SMP, we can only write to the PMC when the task is
3125 * the owner of the local PMU.
3127 if (likely(is_loaded)) {
3129 * In system wide and when the context is loaded, access can only happen
3130 * when the caller is running on the CPU being monitored by the session.
3131 * It does not have to be the owner (ctx_task) of the context per se.
3133 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3134 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3137 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3139 expert_mode = pfm_sysctl.expert_mode;
3141 for (i = 0; i < count; i++, req++) {
3143 cnum = req->reg_num;
3144 value = req->reg_value;
3146 if (!PMD_IS_IMPL(cnum)) {
3147 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3150 is_counting = PMD_IS_COUNTING(cnum);
3151 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3154 * execute write checker, if any
3156 if (unlikely(expert_mode == 0 && wr_func)) {
3157 unsigned long v = value;
3159 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3160 if (ret) goto abort_mission;
3167 * no error on this register
3169 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3172 * now commit changes to software state
3177 * update virtualized (64bits) counter
3181 * write context state
3183 ctx->ctx_pmds[cnum].lval = value;
3186 * when context is load we use the split value
3189 hw_value = value & ovfl_mask;
3190 value = value & ~ovfl_mask;
3194 * update reset values (not just for counters)
3196 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3197 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3200 * update randomization parameters (not just for counters)
3202 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3203 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3206 * update context value
3208 ctx->ctx_pmds[cnum].val = value;
3211 * Keep track of what we use
3213 * We do not keep track of PMC because we have to
3214 * systematically restore ALL of them.
3216 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3219 * mark this PMD register used as well
3221 CTX_USED_PMD(ctx, RDEP(cnum));