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/sched/task.h>
26 #include <linux/sched/task_stack.h>
27 #include <linux/interrupt.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/smp.h>
39 #include <linux/pagemap.h>
40 #include <linux/mount.h>
41 #include <linux/bitops.h>
42 #include <linux/capability.h>
43 #include <linux/rcupdate.h>
44 #include <linux/completion.h>
45 #include <linux/tracehook.h>
46 #include <linux/slab.h>
47 #include <linux/cpu.h>
49 #include <asm/errno.h>
50 #include <asm/intrinsics.h>
52 #include <asm/perfmon.h>
53 #include <asm/processor.h>
54 #include <asm/signal.h>
55 #include <linux/uaccess.h>
56 #include <asm/delay.h>
60 * perfmon context state
62 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
63 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
64 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
65 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
67 #define PFM_INVALID_ACTIVATION (~0UL)
69 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
70 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
73 * depth of message queue
75 #define PFM_MAX_MSGS 32
76 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
79 * type of a PMU register (bitmask).
81 * bit0 : register implemented
84 * bit4 : pmc has pmc.pm
85 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
86 * bit6-7 : register type
89 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
90 #define PFM_REG_IMPL 0x1 /* register implemented */
91 #define PFM_REG_END 0x2 /* end marker */
92 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
93 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
94 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
95 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
96 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
98 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
99 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
101 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
103 /* i assumed unsigned */
104 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
105 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
107 /* XXX: these assume that register i is implemented */
108 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
109 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
110 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
111 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
113 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
114 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
115 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
116 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
118 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
119 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
121 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
122 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
123 #define PFM_CTX_TASK(h) (h)->ctx_task
125 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
127 /* XXX: does not support more than 64 PMDs */
128 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
129 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
131 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
133 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
134 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
135 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
136 #define PFM_CODE_RR 0 /* requesting code range restriction */
137 #define PFM_DATA_RR 1 /* requestion data range restriction */
139 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
140 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
141 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
143 #define RDEP(x) (1UL<<(x))
146 * context protection macros
148 * - we need to protect against CPU concurrency (spin_lock)
149 * - we need to protect against PMU overflow interrupts (local_irq_disable)
151 * - we need to protect against PMU overflow interrupts (local_irq_disable)
153 * spin_lock_irqsave()/spin_unlock_irqrestore():
154 * in SMP: local_irq_disable + spin_lock
155 * in UP : local_irq_disable
157 * spin_lock()/spin_lock():
158 * in UP : removed automatically
159 * in SMP: protect against context accesses from other CPU. interrupts
160 * are not masked. This is useful for the PMU interrupt handler
161 * because we know we will not get PMU concurrency in that code.
163 #define PROTECT_CTX(c, f) \
165 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
166 spin_lock_irqsave(&(c)->ctx_lock, f); \
167 DPRINT(("spinlocked ctx %p by [%d]\n", c, task_pid_nr(current))); \
170 #define UNPROTECT_CTX(c, f) \
172 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
173 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
176 #define PROTECT_CTX_NOPRINT(c, f) \
178 spin_lock_irqsave(&(c)->ctx_lock, f); \
182 #define UNPROTECT_CTX_NOPRINT(c, f) \
184 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
188 #define PROTECT_CTX_NOIRQ(c) \
190 spin_lock(&(c)->ctx_lock); \
193 #define UNPROTECT_CTX_NOIRQ(c) \
195 spin_unlock(&(c)->ctx_lock); \
201 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
202 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
203 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
205 #else /* !CONFIG_SMP */
206 #define SET_ACTIVATION(t) do {} while(0)
207 #define GET_ACTIVATION(t) do {} while(0)
208 #define INC_ACTIVATION(t) do {} while(0)
209 #endif /* CONFIG_SMP */
211 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
212 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
213 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
215 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
216 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
218 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
221 * cmp0 must be the value of pmc0
223 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
225 #define PFMFS_MAGIC 0xa0b4d889
230 #define PFM_DEBUGGING 1
234 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
237 #define DPRINT_ovfl(a) \
239 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
244 * 64-bit software counter structure
246 * the next_reset_type is applied to the next call to pfm_reset_regs()
249 unsigned long val; /* virtual 64bit counter value */
250 unsigned long lval; /* last reset value */
251 unsigned long long_reset; /* reset value on sampling overflow */
252 unsigned long short_reset; /* reset value on overflow */
253 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
254 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
255 unsigned long seed; /* seed for random-number generator */
256 unsigned long mask; /* mask for random-number generator */
257 unsigned int flags; /* notify/do not notify */
258 unsigned long eventid; /* overflow event identifier */
265 unsigned int block:1; /* when 1, task will blocked on user notifications */
266 unsigned int system:1; /* do system wide monitoring */
267 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
268 unsigned int is_sampling:1; /* true if using a custom format */
269 unsigned int excl_idle:1; /* exclude idle task in system wide session */
270 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
271 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
272 unsigned int no_msg:1; /* no message sent on overflow */
273 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
274 unsigned int reserved:22;
275 } pfm_context_flags_t;
277 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
278 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
279 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
283 * perfmon context: encapsulates all the state of a monitoring session
286 typedef struct pfm_context {
287 spinlock_t ctx_lock; /* context protection */
289 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
290 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
292 struct task_struct *ctx_task; /* task to which context is attached */
294 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
296 struct completion ctx_restart_done; /* use for blocking notification mode */
298 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
299 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
300 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
302 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
303 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
304 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
306 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
308 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
309 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
310 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
311 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
313 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
315 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
316 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
318 unsigned long ctx_saved_psr_up; /* only contains psr.up value */
320 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
321 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
322 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
324 int ctx_fd; /* file descriptor used my this context */
325 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
327 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
328 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
329 unsigned long ctx_smpl_size; /* size of sampling buffer */
330 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
332 wait_queue_head_t ctx_msgq_wait;
333 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
336 struct fasync_struct *ctx_async_queue;
338 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
342 * magic number used to verify that structure is really
345 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
347 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
350 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
351 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
353 #define SET_LAST_CPU(ctx, v) do {} while(0)
354 #define GET_LAST_CPU(ctx) do {} while(0)
358 #define ctx_fl_block ctx_flags.block
359 #define ctx_fl_system ctx_flags.system
360 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
361 #define ctx_fl_is_sampling ctx_flags.is_sampling
362 #define ctx_fl_excl_idle ctx_flags.excl_idle
363 #define ctx_fl_going_zombie ctx_flags.going_zombie
364 #define ctx_fl_trap_reason ctx_flags.trap_reason
365 #define ctx_fl_no_msg ctx_flags.no_msg
366 #define ctx_fl_can_restart ctx_flags.can_restart
368 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
369 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
372 * global information about all sessions
373 * mostly used to synchronize between system wide and per-process
376 spinlock_t pfs_lock; /* lock the structure */
378 unsigned int pfs_task_sessions; /* number of per task sessions */
379 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
380 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
381 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
382 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
386 * information about a PMC or PMD.
387 * dep_pmd[]: a bitmask of dependent PMD registers
388 * dep_pmc[]: a bitmask of dependent PMC registers
390 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
394 unsigned long default_value; /* power-on default value */
395 unsigned long reserved_mask; /* bitmask of reserved bits */
396 pfm_reg_check_t read_check;
397 pfm_reg_check_t write_check;
398 unsigned long dep_pmd[4];
399 unsigned long dep_pmc[4];
402 /* assume cnum is a valid monitor */
403 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
406 * This structure is initialized at boot time and contains
407 * a description of the PMU main characteristics.
409 * If the probe function is defined, detection is based
410 * on its return value:
411 * - 0 means recognized PMU
412 * - anything else means not supported
413 * When the probe function is not defined, then the pmu_family field
414 * is used and it must match the host CPU family such that:
415 * - cpu->family & config->pmu_family != 0
418 unsigned long ovfl_val; /* overflow value for counters */
420 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
421 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
423 unsigned int num_pmcs; /* number of PMCS: computed at init time */
424 unsigned int num_pmds; /* number of PMDS: computed at init time */
425 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
426 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
428 char *pmu_name; /* PMU family name */
429 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
430 unsigned int flags; /* pmu specific flags */
431 unsigned int num_ibrs; /* number of IBRS: computed at init time */
432 unsigned int num_dbrs; /* number of DBRS: computed at init time */
433 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
434 int (*probe)(void); /* customized probe routine */
435 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
440 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
443 * debug register related type definitions
446 unsigned long ibr_mask:56;
447 unsigned long ibr_plm:4;
448 unsigned long ibr_ig:3;
449 unsigned long ibr_x:1;
453 unsigned long dbr_mask:56;
454 unsigned long dbr_plm:4;
455 unsigned long dbr_ig:2;
456 unsigned long dbr_w:1;
457 unsigned long dbr_r:1;
468 * perfmon command descriptions
471 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
474 unsigned int cmd_narg;
476 int (*cmd_getsize)(void *arg, size_t *sz);
479 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
480 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
481 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
482 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
485 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
486 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
487 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
488 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
489 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
491 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
494 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
495 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
496 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
497 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
498 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
499 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
500 unsigned long pfm_smpl_handler_calls;
501 unsigned long pfm_smpl_handler_cycles;
502 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
506 * perfmon internal variables
508 static pfm_stats_t pfm_stats[NR_CPUS];
509 static pfm_session_t pfm_sessions; /* global sessions information */
511 static DEFINE_SPINLOCK(pfm_alt_install_check);
512 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
514 static struct proc_dir_entry *perfmon_dir;
515 static pfm_uuid_t pfm_null_uuid = {0,};
517 static spinlock_t pfm_buffer_fmt_lock;
518 static LIST_HEAD(pfm_buffer_fmt_list);
520 static pmu_config_t *pmu_conf;
522 /* sysctl() controls */
523 pfm_sysctl_t pfm_sysctl;
524 EXPORT_SYMBOL(pfm_sysctl);
526 static struct ctl_table pfm_ctl_table[] = {
529 .data = &pfm_sysctl.debug,
530 .maxlen = sizeof(int),
532 .proc_handler = proc_dointvec,
535 .procname = "debug_ovfl",
536 .data = &pfm_sysctl.debug_ovfl,
537 .maxlen = sizeof(int),
539 .proc_handler = proc_dointvec,
542 .procname = "fastctxsw",
543 .data = &pfm_sysctl.fastctxsw,
544 .maxlen = sizeof(int),
546 .proc_handler = proc_dointvec,
549 .procname = "expert_mode",
550 .data = &pfm_sysctl.expert_mode,
551 .maxlen = sizeof(int),
553 .proc_handler = proc_dointvec,
557 static struct ctl_table pfm_sysctl_dir[] = {
559 .procname = "perfmon",
561 .child = pfm_ctl_table,
565 static struct ctl_table pfm_sysctl_root[] = {
567 .procname = "kernel",
569 .child = pfm_sysctl_dir,
573 static struct ctl_table_header *pfm_sysctl_header;
575 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
577 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
578 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
581 pfm_put_task(struct task_struct *task)
583 if (task != current) put_task_struct(task);
586 static inline unsigned long
587 pfm_protect_ctx_ctxsw(pfm_context_t *x)
589 spin_lock(&(x)->ctx_lock);
594 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
596 spin_unlock(&(x)->ctx_lock);
599 /* forward declaration */
600 static const struct dentry_operations pfmfs_dentry_operations;
602 static struct dentry *
603 pfmfs_mount(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
605 return mount_pseudo(fs_type, "pfm:", NULL, &pfmfs_dentry_operations,
609 static struct file_system_type pfm_fs_type = {
611 .mount = pfmfs_mount,
612 .kill_sb = kill_anon_super,
614 MODULE_ALIAS_FS("pfmfs");
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 const 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));
808 static pfm_context_t *
809 pfm_context_alloc(int ctx_flags)
814 * allocate context descriptor
815 * must be able to free with interrupts disabled
817 ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
819 DPRINT(("alloc ctx @%p\n", ctx));
822 * init context protection lock
824 spin_lock_init(&ctx->ctx_lock);
827 * context is unloaded
829 ctx->ctx_state = PFM_CTX_UNLOADED;
832 * initialization of context's flags
834 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
835 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
836 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
838 * will move to set properties
839 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
843 * init restart semaphore to locked
845 init_completion(&ctx->ctx_restart_done);
848 * activation is used in SMP only
850 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
851 SET_LAST_CPU(ctx, -1);
854 * initialize notification message queue
856 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
857 init_waitqueue_head(&ctx->ctx_msgq_wait);
858 init_waitqueue_head(&ctx->ctx_zombieq);
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_nr(task)));
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_nr(task), task_pid_nr(current));
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_nr(task), task_pid_nr(current), 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",
1019 task_pid_nr(task), i, ctx->th_pmcs[i]));
1024 * must restore DBR/IBR because could be modified while masked
1025 * XXX: need to optimize
1027 if (ctx->ctx_fl_using_dbreg) {
1028 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1029 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1035 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1037 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1044 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1050 for (i=0; mask; i++, mask>>=1) {
1051 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1056 * reload from thread state (used for ctxw only)
1059 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1062 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1064 for (i=0; mask; i++, mask>>=1) {
1065 if ((mask & 0x1) == 0) continue;
1066 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1067 ia64_set_pmd(i, val);
1073 * propagate PMD from context to thread-state
1076 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1078 unsigned long ovfl_val = pmu_conf->ovfl_val;
1079 unsigned long mask = ctx->ctx_all_pmds[0];
1083 DPRINT(("mask=0x%lx\n", mask));
1085 for (i=0; mask; i++, mask>>=1) {
1087 val = ctx->ctx_pmds[i].val;
1090 * We break up the 64 bit value into 2 pieces
1091 * the lower bits go to the machine state in the
1092 * thread (will be reloaded on ctxsw in).
1093 * The upper part stays in the soft-counter.
1095 if (PMD_IS_COUNTING(i)) {
1096 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1099 ctx->th_pmds[i] = val;
1101 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1104 ctx->ctx_pmds[i].val));
1109 * propagate PMC from context to thread-state
1112 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1114 unsigned long mask = ctx->ctx_all_pmcs[0];
1117 DPRINT(("mask=0x%lx\n", mask));
1119 for (i=0; mask; i++, mask>>=1) {
1120 /* masking 0 with ovfl_val yields 0 */
1121 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1122 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1129 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1133 for (i=0; mask; i++, mask>>=1) {
1134 if ((mask & 0x1) == 0) continue;
1135 ia64_set_pmc(i, pmcs[i]);
1141 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1143 return memcmp(a, b, sizeof(pfm_uuid_t));
1147 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1150 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1155 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1158 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1164 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1168 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1173 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1177 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1182 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1185 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1190 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)
1193 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1197 static pfm_buffer_fmt_t *
1198 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1200 struct list_head * pos;
1201 pfm_buffer_fmt_t * entry;
1203 list_for_each(pos, &pfm_buffer_fmt_list) {
1204 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1205 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1212 * find a buffer format based on its uuid
1214 static pfm_buffer_fmt_t *
1215 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1217 pfm_buffer_fmt_t * fmt;
1218 spin_lock(&pfm_buffer_fmt_lock);
1219 fmt = __pfm_find_buffer_fmt(uuid);
1220 spin_unlock(&pfm_buffer_fmt_lock);
1225 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1229 /* some sanity checks */
1230 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1232 /* we need at least a handler */
1233 if (fmt->fmt_handler == NULL) return -EINVAL;
1236 * XXX: need check validity of fmt_arg_size
1239 spin_lock(&pfm_buffer_fmt_lock);
1241 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1242 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1246 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1247 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1250 spin_unlock(&pfm_buffer_fmt_lock);
1253 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1256 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1258 pfm_buffer_fmt_t *fmt;
1261 spin_lock(&pfm_buffer_fmt_lock);
1263 fmt = __pfm_find_buffer_fmt(uuid);
1265 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1269 list_del_init(&fmt->fmt_list);
1270 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1273 spin_unlock(&pfm_buffer_fmt_lock);
1277 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1280 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1282 unsigned long flags;
1284 * validity checks on cpu_mask have been done upstream
1288 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1289 pfm_sessions.pfs_sys_sessions,
1290 pfm_sessions.pfs_task_sessions,
1291 pfm_sessions.pfs_sys_use_dbregs,
1297 * cannot mix system wide and per-task sessions
1299 if (pfm_sessions.pfs_task_sessions > 0UL) {
1300 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1301 pfm_sessions.pfs_task_sessions));
1305 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1307 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1309 pfm_sessions.pfs_sys_session[cpu] = task;
1311 pfm_sessions.pfs_sys_sessions++ ;
1314 if (pfm_sessions.pfs_sys_sessions) goto abort;
1315 pfm_sessions.pfs_task_sessions++;
1318 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1319 pfm_sessions.pfs_sys_sessions,
1320 pfm_sessions.pfs_task_sessions,
1321 pfm_sessions.pfs_sys_use_dbregs,
1326 * Force idle() into poll mode
1328 cpu_idle_poll_ctrl(true);
1335 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1336 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1346 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1348 unsigned long flags;
1350 * validity checks on cpu_mask have been done upstream
1354 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1355 pfm_sessions.pfs_sys_sessions,
1356 pfm_sessions.pfs_task_sessions,
1357 pfm_sessions.pfs_sys_use_dbregs,
1363 pfm_sessions.pfs_sys_session[cpu] = NULL;
1365 * would not work with perfmon+more than one bit in cpu_mask
1367 if (ctx && ctx->ctx_fl_using_dbreg) {
1368 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1369 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1371 pfm_sessions.pfs_sys_use_dbregs--;
1374 pfm_sessions.pfs_sys_sessions--;
1376 pfm_sessions.pfs_task_sessions--;
1378 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1379 pfm_sessions.pfs_sys_sessions,
1380 pfm_sessions.pfs_task_sessions,
1381 pfm_sessions.pfs_sys_use_dbregs,
1385 /* Undo forced polling. Last session reenables pal_halt */
1386 cpu_idle_poll_ctrl(false);
1394 * removes virtual mapping of the sampling buffer.
1395 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1396 * a PROTECT_CTX() section.
1399 pfm_remove_smpl_mapping(void *vaddr, unsigned long size)
1401 struct task_struct *task = current;
1405 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1406 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1410 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1413 * does the actual unmapping
1415 r = vm_munmap((unsigned long)vaddr, size);
1418 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1421 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1427 * free actual physical storage used by sampling buffer
1431 pfm_free_smpl_buffer(pfm_context_t *ctx)
1433 pfm_buffer_fmt_t *fmt;
1435 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1438 * we won't use the buffer format anymore
1440 fmt = ctx->ctx_buf_fmt;
1442 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1445 ctx->ctx_smpl_vaddr));
1447 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1452 vfree(ctx->ctx_smpl_hdr);
1454 ctx->ctx_smpl_hdr = NULL;
1455 ctx->ctx_smpl_size = 0UL;
1460 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1466 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1468 if (fmt == NULL) return;
1470 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1475 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1476 * no real gain from having the whole whorehouse mounted. So we don't need
1477 * any operations on the root directory. However, we need a non-trivial
1478 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1480 static struct vfsmount *pfmfs_mnt __read_mostly;
1485 int err = register_filesystem(&pfm_fs_type);
1487 pfmfs_mnt = kern_mount(&pfm_fs_type);
1488 err = PTR_ERR(pfmfs_mnt);
1489 if (IS_ERR(pfmfs_mnt))
1490 unregister_filesystem(&pfm_fs_type);
1498 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1503 unsigned long flags;
1504 DECLARE_WAITQUEUE(wait, current);
1505 if (PFM_IS_FILE(filp) == 0) {
1506 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1510 ctx = filp->private_data;
1512 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1517 * check even when there is no message
1519 if (size < sizeof(pfm_msg_t)) {
1520 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1524 PROTECT_CTX(ctx, flags);
1527 * put ourselves on the wait queue
1529 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1537 set_current_state(TASK_INTERRUPTIBLE);
1539 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1542 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1544 UNPROTECT_CTX(ctx, flags);
1547 * check non-blocking read
1550 if(filp->f_flags & O_NONBLOCK) break;
1553 * check pending signals
1555 if(signal_pending(current)) {
1560 * no message, so wait
1564 PROTECT_CTX(ctx, flags);
1566 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1567 set_current_state(TASK_RUNNING);
1568 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1570 if (ret < 0) goto abort;
1573 msg = pfm_get_next_msg(ctx);
1575 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1579 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1582 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1585 UNPROTECT_CTX(ctx, flags);
1591 pfm_write(struct file *file, const char __user *ubuf,
1592 size_t size, loff_t *ppos)
1594 DPRINT(("pfm_write called\n"));
1599 pfm_poll(struct file *filp, poll_table * wait)
1602 unsigned long flags;
1605 if (PFM_IS_FILE(filp) == 0) {
1606 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1610 ctx = filp->private_data;
1612 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1617 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1619 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1621 PROTECT_CTX(ctx, flags);
1623 if (PFM_CTXQ_EMPTY(ctx) == 0)
1624 mask = EPOLLIN | EPOLLRDNORM;
1626 UNPROTECT_CTX(ctx, flags);
1628 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1634 pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1636 DPRINT(("pfm_ioctl called\n"));
1641 * interrupt cannot be masked when coming here
1644 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1648 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1650 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1651 task_pid_nr(current),
1654 ctx->ctx_async_queue, ret));
1660 pfm_fasync(int fd, struct file *filp, int on)
1665 if (PFM_IS_FILE(filp) == 0) {
1666 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1670 ctx = filp->private_data;
1672 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1676 * we cannot mask interrupts during this call because this may
1677 * may go to sleep if memory is not readily avalaible.
1679 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1680 * done in caller. Serialization of this function is ensured by caller.
1682 ret = pfm_do_fasync(fd, filp, ctx, on);
1685 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1688 ctx->ctx_async_queue, ret));
1695 * this function is exclusively called from pfm_close().
1696 * The context is not protected at that time, nor are interrupts
1697 * on the remote CPU. That's necessary to avoid deadlocks.
1700 pfm_syswide_force_stop(void *info)
1702 pfm_context_t *ctx = (pfm_context_t *)info;
1703 struct pt_regs *regs = task_pt_regs(current);
1704 struct task_struct *owner;
1705 unsigned long flags;
1708 if (ctx->ctx_cpu != smp_processor_id()) {
1709 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1711 smp_processor_id());
1714 owner = GET_PMU_OWNER();
1715 if (owner != ctx->ctx_task) {
1716 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1718 task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1721 if (GET_PMU_CTX() != ctx) {
1722 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1724 GET_PMU_CTX(), ctx);
1728 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1730 * the context is already protected in pfm_close(), we simply
1731 * need to mask interrupts to avoid a PMU interrupt race on
1734 local_irq_save(flags);
1736 ret = pfm_context_unload(ctx, NULL, 0, regs);
1738 DPRINT(("context_unload returned %d\n", ret));
1742 * unmask interrupts, PMU interrupts are now spurious here
1744 local_irq_restore(flags);
1748 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1752 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1753 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1754 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1756 #endif /* CONFIG_SMP */
1759 * called for each close(). Partially free resources.
1760 * When caller is self-monitoring, the context is unloaded.
1763 pfm_flush(struct file *filp, fl_owner_t id)
1766 struct task_struct *task;
1767 struct pt_regs *regs;
1768 unsigned long flags;
1769 unsigned long smpl_buf_size = 0UL;
1770 void *smpl_buf_vaddr = NULL;
1771 int state, is_system;
1773 if (PFM_IS_FILE(filp) == 0) {
1774 DPRINT(("bad magic for\n"));
1778 ctx = filp->private_data;
1780 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1785 * remove our file from the async queue, if we use this mode.
1786 * This can be done without the context being protected. We come
1787 * here when the context has become unreachable by other tasks.
1789 * We may still have active monitoring at this point and we may
1790 * end up in pfm_overflow_handler(). However, fasync_helper()
1791 * operates with interrupts disabled and it cleans up the
1792 * queue. If the PMU handler is called prior to entering
1793 * fasync_helper() then it will send a signal. If it is
1794 * invoked after, it will find an empty queue and no
1795 * signal will be sent. In both case, we are safe
1797 PROTECT_CTX(ctx, flags);
1799 state = ctx->ctx_state;
1800 is_system = ctx->ctx_fl_system;
1802 task = PFM_CTX_TASK(ctx);
1803 regs = task_pt_regs(task);
1805 DPRINT(("ctx_state=%d is_current=%d\n",
1807 task == current ? 1 : 0));
1810 * if state == UNLOADED, then task is NULL
1814 * we must stop and unload because we are losing access to the context.
1816 if (task == current) {
1819 * the task IS the owner but it migrated to another CPU: that's bad
1820 * but we must handle this cleanly. Unfortunately, the kernel does
1821 * not provide a mechanism to block migration (while the context is loaded).
1823 * We need to release the resource on the ORIGINAL cpu.
1825 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1827 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1829 * keep context protected but unmask interrupt for IPI
1831 local_irq_restore(flags);
1833 pfm_syswide_cleanup_other_cpu(ctx);
1836 * restore interrupt masking
1838 local_irq_save(flags);
1841 * context is unloaded at this point
1844 #endif /* CONFIG_SMP */
1847 DPRINT(("forcing unload\n"));
1849 * stop and unload, returning with state UNLOADED
1850 * and session unreserved.
1852 pfm_context_unload(ctx, NULL, 0, regs);
1854 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1859 * remove virtual mapping, if any, for the calling task.
1860 * cannot reset ctx field until last user is calling close().
1862 * ctx_smpl_vaddr must never be cleared because it is needed
1863 * by every task with access to the context
1865 * When called from do_exit(), the mm context is gone already, therefore
1866 * mm is NULL, i.e., the VMA is already gone and we do not have to
1869 if (ctx->ctx_smpl_vaddr && current->mm) {
1870 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1871 smpl_buf_size = ctx->ctx_smpl_size;
1874 UNPROTECT_CTX(ctx, flags);
1877 * if there was a mapping, then we systematically remove it
1878 * at this point. Cannot be done inside critical section
1879 * because some VM function reenables interrupts.
1882 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(smpl_buf_vaddr, smpl_buf_size);
1887 * called either on explicit close() or from exit_files().
1888 * Only the LAST user of the file gets to this point, i.e., it is
1891 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1892 * (fput()),i.e, last task to access the file. Nobody else can access the
1893 * file at this point.
1895 * When called from exit_files(), the VMA has been freed because exit_mm()
1896 * is executed before exit_files().
1898 * When called from exit_files(), the current task is not yet ZOMBIE but we
1899 * flush the PMU state to the context.
1902 pfm_close(struct inode *inode, struct file *filp)
1905 struct task_struct *task;
1906 struct pt_regs *regs;
1907 DECLARE_WAITQUEUE(wait, current);
1908 unsigned long flags;
1909 unsigned long smpl_buf_size = 0UL;
1910 void *smpl_buf_addr = NULL;
1911 int free_possible = 1;
1912 int state, is_system;
1914 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1916 if (PFM_IS_FILE(filp) == 0) {
1917 DPRINT(("bad magic\n"));
1921 ctx = filp->private_data;
1923 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1927 PROTECT_CTX(ctx, flags);
1929 state = ctx->ctx_state;
1930 is_system = ctx->ctx_fl_system;
1932 task = PFM_CTX_TASK(ctx);
1933 regs = task_pt_regs(task);
1935 DPRINT(("ctx_state=%d is_current=%d\n",
1937 task == current ? 1 : 0));
1940 * if task == current, then pfm_flush() unloaded the context
1942 if (state == PFM_CTX_UNLOADED) goto doit;
1945 * context is loaded/masked and task != current, we need to
1946 * either force an unload or go zombie
1950 * The task is currently blocked or will block after an overflow.
1951 * we must force it to wakeup to get out of the
1952 * MASKED state and transition to the unloaded state by itself.
1954 * This situation is only possible for per-task mode
1956 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1959 * set a "partial" zombie state to be checked
1960 * upon return from down() in pfm_handle_work().
1962 * We cannot use the ZOMBIE state, because it is checked
1963 * by pfm_load_regs() which is called upon wakeup from down().
1964 * In such case, it would free the context and then we would
1965 * return to pfm_handle_work() which would access the
1966 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1967 * but visible to pfm_handle_work().
1969 * For some window of time, we have a zombie context with
1970 * ctx_state = MASKED and not ZOMBIE
1972 ctx->ctx_fl_going_zombie = 1;
1975 * force task to wake up from MASKED state
1977 complete(&ctx->ctx_restart_done);
1979 DPRINT(("waking up ctx_state=%d\n", state));
1982 * put ourself to sleep waiting for the other
1983 * task to report completion
1985 * the context is protected by mutex, therefore there
1986 * is no risk of being notified of completion before
1987 * begin actually on the waitq.
1989 set_current_state(TASK_INTERRUPTIBLE);
1990 add_wait_queue(&ctx->ctx_zombieq, &wait);
1992 UNPROTECT_CTX(ctx, flags);
1995 * XXX: check for signals :
1996 * - ok for explicit close
1997 * - not ok when coming from exit_files()
2002 PROTECT_CTX(ctx, flags);
2005 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2006 set_current_state(TASK_RUNNING);
2009 * context is unloaded at this point
2011 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2013 else if (task != current) {
2016 * switch context to zombie state
2018 ctx->ctx_state = PFM_CTX_ZOMBIE;
2020 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2022 * cannot free the context on the spot. deferred until
2023 * the task notices the ZOMBIE state
2027 pfm_context_unload(ctx, NULL, 0, regs);
2032 /* reload state, may have changed during opening of critical section */
2033 state = ctx->ctx_state;
2036 * the context is still attached to a task (possibly current)
2037 * we cannot destroy it right now
2041 * we must free the sampling buffer right here because
2042 * we cannot rely on it being cleaned up later by the
2043 * monitored task. It is not possible to free vmalloc'ed
2044 * memory in pfm_load_regs(). Instead, we remove the buffer
2045 * now. should there be subsequent PMU overflow originally
2046 * meant for sampling, the will be converted to spurious
2047 * and that's fine because the monitoring tools is gone anyway.
2049 if (ctx->ctx_smpl_hdr) {
2050 smpl_buf_addr = ctx->ctx_smpl_hdr;
2051 smpl_buf_size = ctx->ctx_smpl_size;
2052 /* no more sampling */
2053 ctx->ctx_smpl_hdr = NULL;
2054 ctx->ctx_fl_is_sampling = 0;
2057 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2063 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2066 * UNLOADED that the session has already been unreserved.
2068 if (state == PFM_CTX_ZOMBIE) {
2069 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2073 * disconnect file descriptor from context must be done
2076 filp->private_data = NULL;
2079 * if we free on the spot, the context is now completely unreachable
2080 * from the callers side. The monitored task side is also cut, so we
2083 * If we have a deferred free, only the caller side is disconnected.
2085 UNPROTECT_CTX(ctx, flags);
2088 * All memory free operations (especially for vmalloc'ed memory)
2089 * MUST be done with interrupts ENABLED.
2091 vfree(smpl_buf_addr);
2094 * return the memory used by the context
2096 if (free_possible) pfm_context_free(ctx);
2101 static const struct file_operations pfm_file_ops = {
2102 .llseek = no_llseek,
2106 .unlocked_ioctl = pfm_ioctl,
2107 .fasync = pfm_fasync,
2108 .release = pfm_close,
2112 static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2114 return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2115 d_inode(dentry)->i_ino);
2118 static const struct dentry_operations pfmfs_dentry_operations = {
2119 .d_delete = always_delete_dentry,
2120 .d_dname = pfmfs_dname,
2124 static struct file *
2125 pfm_alloc_file(pfm_context_t *ctx)
2128 struct inode *inode;
2130 struct qstr this = { .name = "" };
2133 * allocate a new inode
2135 inode = new_inode(pfmfs_mnt->mnt_sb);
2137 return ERR_PTR(-ENOMEM);
2139 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2141 inode->i_mode = S_IFCHR|S_IRUGO;
2142 inode->i_uid = current_fsuid();
2143 inode->i_gid = current_fsgid();
2146 * allocate a new dcache entry
2148 path.dentry = d_alloc(pfmfs_mnt->mnt_root, &this);
2151 return ERR_PTR(-ENOMEM);
2153 path.mnt = mntget(pfmfs_mnt);
2155 d_add(path.dentry, inode);
2157 file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2163 file->f_flags = O_RDONLY;
2164 file->private_data = ctx;
2170 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2172 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2175 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2178 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2189 * allocate a sampling buffer and remaps it into the user address space of the task
2192 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2194 struct mm_struct *mm = task->mm;
2195 struct vm_area_struct *vma = NULL;
2201 * the fixed header + requested size and align to page boundary
2203 size = PAGE_ALIGN(rsize);
2205 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2208 * check requested size to avoid Denial-of-service attacks
2209 * XXX: may have to refine this test
2210 * Check against address space limit.
2212 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2215 if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2219 * We do the easy to undo allocations first.
2221 smpl_buf = vzalloc(size);
2222 if (smpl_buf == NULL) {
2223 DPRINT(("Can't allocate sampling buffer\n"));
2227 DPRINT(("smpl_buf @%p\n", smpl_buf));
2230 vma = vm_area_alloc(mm);
2232 DPRINT(("Cannot allocate vma\n"));
2237 * partially initialize the vma for the sampling buffer
2239 vma->vm_file = get_file(filp);
2240 vma->vm_flags = VM_READ|VM_MAYREAD|VM_DONTEXPAND|VM_DONTDUMP;
2241 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2244 * Now we have everything we need and we can initialize
2245 * and connect all the data structures
2248 ctx->ctx_smpl_hdr = smpl_buf;
2249 ctx->ctx_smpl_size = size; /* aligned size */
2252 * Let's do the difficult operations next.
2254 * now we atomically find some area in the address space and
2255 * remap the buffer in it.
2257 down_write(&task->mm->mmap_sem);
2259 /* find some free area in address space, must have mmap sem held */
2260 vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2261 if (IS_ERR_VALUE(vma->vm_start)) {
2262 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2263 up_write(&task->mm->mmap_sem);
2266 vma->vm_end = vma->vm_start + size;
2267 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2269 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2271 /* can only be applied to current task, need to have the mm semaphore held when called */
2272 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2273 DPRINT(("Can't remap buffer\n"));
2274 up_write(&task->mm->mmap_sem);
2279 * now insert the vma in the vm list for the process, must be
2280 * done with mmap lock held
2282 insert_vm_struct(mm, vma);
2284 vm_stat_account(vma->vm_mm, vma->vm_flags, vma_pages(vma));
2285 up_write(&task->mm->mmap_sem);
2288 * keep track of user level virtual address
2290 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2291 *(unsigned long *)user_vaddr = vma->vm_start;
2304 * XXX: do something better here
2307 pfm_bad_permissions(struct task_struct *task)
2309 const struct cred *tcred;
2310 kuid_t uid = current_uid();
2311 kgid_t gid = current_gid();
2315 tcred = __task_cred(task);
2317 /* inspired by ptrace_attach() */
2318 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2319 from_kuid(&init_user_ns, uid),
2320 from_kgid(&init_user_ns, gid),
2321 from_kuid(&init_user_ns, tcred->euid),
2322 from_kuid(&init_user_ns, tcred->suid),
2323 from_kuid(&init_user_ns, tcred->uid),
2324 from_kgid(&init_user_ns, tcred->egid),
2325 from_kgid(&init_user_ns, tcred->sgid)));
2327 ret = ((!uid_eq(uid, tcred->euid))
2328 || (!uid_eq(uid, tcred->suid))
2329 || (!uid_eq(uid, tcred->uid))
2330 || (!gid_eq(gid, tcred->egid))
2331 || (!gid_eq(gid, tcred->sgid))
2332 || (!gid_eq(gid, tcred->gid))) && !capable(CAP_SYS_PTRACE);
2339 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2345 ctx_flags = pfx->ctx_flags;
2347 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2350 * cannot block in this mode
2352 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2353 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2358 /* probably more to add here */
2364 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2365 unsigned int cpu, pfarg_context_t *arg)
2367 pfm_buffer_fmt_t *fmt = NULL;
2368 unsigned long size = 0UL;
2370 void *fmt_arg = NULL;
2372 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2374 /* invoke and lock buffer format, if found */
2375 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2377 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2382 * buffer argument MUST be contiguous to pfarg_context_t
2384 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2386 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2388 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2390 if (ret) goto error;
2392 /* link buffer format and context */
2393 ctx->ctx_buf_fmt = fmt;
2394 ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2397 * check if buffer format wants to use perfmon buffer allocation/mapping service
2399 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2400 if (ret) goto error;
2404 * buffer is always remapped into the caller's address space
2406 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2407 if (ret) goto error;
2409 /* keep track of user address of buffer */
2410 arg->ctx_smpl_vaddr = uaddr;
2412 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2419 pfm_reset_pmu_state(pfm_context_t *ctx)
2424 * install reset values for PMC.
2426 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2427 if (PMC_IS_IMPL(i) == 0) continue;
2428 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2429 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2432 * PMD registers are set to 0UL when the context in memset()
2436 * On context switched restore, we must restore ALL pmc and ALL pmd even
2437 * when they are not actively used by the task. In UP, the incoming process
2438 * may otherwise pick up left over PMC, PMD state from the previous process.
2439 * As opposed to PMD, stale PMC can cause harm to the incoming
2440 * process because they may change what is being measured.
2441 * Therefore, we must systematically reinstall the entire
2442 * PMC state. In SMP, the same thing is possible on the
2443 * same CPU but also on between 2 CPUs.
2445 * The problem with PMD is information leaking especially
2446 * to user level when psr.sp=0
2448 * There is unfortunately no easy way to avoid this problem
2449 * on either UP or SMP. This definitively slows down the
2450 * pfm_load_regs() function.
2454 * bitmask of all PMCs accessible to this context
2456 * PMC0 is treated differently.
2458 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2461 * bitmask of all PMDs that are accessible to this context
2463 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2465 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2468 * useful in case of re-enable after disable
2470 ctx->ctx_used_ibrs[0] = 0UL;
2471 ctx->ctx_used_dbrs[0] = 0UL;
2475 pfm_ctx_getsize(void *arg, size_t *sz)
2477 pfarg_context_t *req = (pfarg_context_t *)arg;
2478 pfm_buffer_fmt_t *fmt;
2482 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2484 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2486 DPRINT(("cannot find buffer format\n"));
2489 /* get just enough to copy in user parameters */
2490 *sz = fmt->fmt_arg_size;
2491 DPRINT(("arg_size=%lu\n", *sz));
2499 * cannot attach if :
2501 * - task not owned by caller
2502 * - task incompatible with context mode
2505 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2508 * no kernel task or task not owner by caller
2510 if (task->mm == NULL) {
2511 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2514 if (pfm_bad_permissions(task)) {
2515 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task)));
2519 * cannot block in self-monitoring mode
2521 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2522 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2526 if (task->exit_state == EXIT_ZOMBIE) {
2527 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task)));
2532 * always ok for self
2534 if (task == current) return 0;
2536 if (!task_is_stopped_or_traced(task)) {
2537 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2541 * make sure the task is off any CPU
2543 wait_task_inactive(task, 0);
2545 /* more to come... */
2551 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2553 struct task_struct *p = current;
2556 /* XXX: need to add more checks here */
2557 if (pid < 2) return -EPERM;
2559 if (pid != task_pid_vnr(current)) {
2560 /* make sure task cannot go away while we operate on it */
2561 p = find_get_task_by_vpid(pid);
2566 ret = pfm_task_incompatible(ctx, p);
2569 } else if (p != current) {
2578 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2580 pfarg_context_t *req = (pfarg_context_t *)arg;
2587 /* let's check the arguments first */
2588 ret = pfarg_is_sane(current, req);
2592 ctx_flags = req->ctx_flags;
2596 fd = get_unused_fd_flags(0);
2600 ctx = pfm_context_alloc(ctx_flags);
2604 filp = pfm_alloc_file(ctx);
2606 ret = PTR_ERR(filp);
2610 req->ctx_fd = ctx->ctx_fd = fd;
2613 * does the user want to sample?
2615 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2616 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2621 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2626 ctx->ctx_fl_excl_idle,
2631 * initialize soft PMU state
2633 pfm_reset_pmu_state(ctx);
2635 fd_install(fd, filp);
2640 path = filp->f_path;
2644 if (ctx->ctx_buf_fmt) {
2645 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2648 pfm_context_free(ctx);
2655 static inline unsigned long
2656 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2658 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2659 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2660 extern unsigned long carta_random32 (unsigned long seed);
2662 if (reg->flags & PFM_REGFL_RANDOM) {
2663 new_seed = carta_random32(old_seed);
2664 val -= (old_seed & mask); /* counter values are negative numbers! */
2665 if ((mask >> 32) != 0)
2666 /* construct a full 64-bit random value: */
2667 new_seed |= carta_random32(old_seed >> 32) << 32;
2668 reg->seed = new_seed;
2675 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2677 unsigned long mask = ovfl_regs[0];
2678 unsigned long reset_others = 0UL;
2683 * now restore reset value on sampling overflowed counters
2685 mask >>= PMU_FIRST_COUNTER;
2686 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2688 if ((mask & 0x1UL) == 0UL) continue;
2690 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2691 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2693 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2697 * Now take care of resetting the other registers
2699 for(i = 0; reset_others; i++, reset_others >>= 1) {
2701 if ((reset_others & 0x1) == 0) continue;
2703 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2705 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2706 is_long_reset ? "long" : "short", i, val));
2711 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2713 unsigned long mask = ovfl_regs[0];
2714 unsigned long reset_others = 0UL;
2718 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2720 if (ctx->ctx_state == PFM_CTX_MASKED) {
2721 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2726 * now restore reset value on sampling overflowed counters
2728 mask >>= PMU_FIRST_COUNTER;
2729 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2731 if ((mask & 0x1UL) == 0UL) continue;
2733 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2734 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2736 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2738 pfm_write_soft_counter(ctx, i, val);
2742 * Now take care of resetting the other registers
2744 for(i = 0; reset_others; i++, reset_others >>= 1) {
2746 if ((reset_others & 0x1) == 0) continue;
2748 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2750 if (PMD_IS_COUNTING(i)) {
2751 pfm_write_soft_counter(ctx, i, val);
2753 ia64_set_pmd(i, val);
2755 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2756 is_long_reset ? "long" : "short", i, val));
2762 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2764 struct task_struct *task;
2765 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2766 unsigned long value, pmc_pm;
2767 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2768 unsigned int cnum, reg_flags, flags, pmc_type;
2769 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2770 int is_monitor, is_counting, state;
2772 pfm_reg_check_t wr_func;
2773 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2775 state = ctx->ctx_state;
2776 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2777 is_system = ctx->ctx_fl_system;
2778 task = ctx->ctx_task;
2779 impl_pmds = pmu_conf->impl_pmds[0];
2781 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2785 * In system wide and when the context is loaded, access can only happen
2786 * when the caller is running on the CPU being monitored by the session.
2787 * It does not have to be the owner (ctx_task) of the context per se.
2789 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2790 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2793 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2795 expert_mode = pfm_sysctl.expert_mode;
2797 for (i = 0; i < count; i++, req++) {
2799 cnum = req->reg_num;
2800 reg_flags = req->reg_flags;
2801 value = req->reg_value;
2802 smpl_pmds = req->reg_smpl_pmds[0];
2803 reset_pmds = req->reg_reset_pmds[0];
2807 if (cnum >= PMU_MAX_PMCS) {
2808 DPRINT(("pmc%u is invalid\n", cnum));
2812 pmc_type = pmu_conf->pmc_desc[cnum].type;
2813 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2814 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2815 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2818 * we reject all non implemented PMC as well
2819 * as attempts to modify PMC[0-3] which are used
2820 * as status registers by the PMU
2822 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2823 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2826 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2828 * If the PMC is a monitor, then if the value is not the default:
2829 * - system-wide session: PMCx.pm=1 (privileged monitor)
2830 * - per-task : PMCx.pm=0 (user monitor)
2832 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2833 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2842 * enforce generation of overflow interrupt. Necessary on all
2845 value |= 1 << PMU_PMC_OI;
2847 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2848 flags |= PFM_REGFL_OVFL_NOTIFY;
2851 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2853 /* verify validity of smpl_pmds */
2854 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2855 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2859 /* verify validity of reset_pmds */
2860 if ((reset_pmds & impl_pmds) != reset_pmds) {
2861 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2865 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2866 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2869 /* eventid on non-counting monitors are ignored */
2873 * execute write checker, if any
2875 if (likely(expert_mode == 0 && wr_func)) {
2876 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2877 if (ret) goto error;
2882 * no error on this register
2884 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2887 * Now we commit the changes to the software state
2891 * update overflow information
2895 * full flag update each time a register is programmed
2897 ctx->ctx_pmds[cnum].flags = flags;
2899 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2900 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2901 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2904 * Mark all PMDS to be accessed as used.
2906 * We do not keep track of PMC because we have to
2907 * systematically restore ALL of them.
2909 * We do not update the used_monitors mask, because
2910 * if we have not programmed them, then will be in
2911 * a quiescent state, therefore we will not need to
2912 * mask/restore then when context is MASKED.
2914 CTX_USED_PMD(ctx, reset_pmds);
2915 CTX_USED_PMD(ctx, smpl_pmds);
2917 * make sure we do not try to reset on
2918 * restart because we have established new values
2920 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
2923 * Needed in case the user does not initialize the equivalent
2924 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2925 * possible leak here.
2927 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
2930 * keep track of the monitor PMC that we are using.
2931 * we save the value of the pmc in ctx_pmcs[] and if
2932 * the monitoring is not stopped for the context we also
2933 * place it in the saved state area so that it will be
2934 * picked up later by the context switch code.
2936 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
2938 * The value in th_pmcs[] may be modified on overflow, i.e., when
2939 * monitoring needs to be stopped.
2941 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
2944 * update context state
2946 ctx->ctx_pmcs[cnum] = value;
2950 * write thread state
2952 if (is_system == 0) ctx->th_pmcs[cnum] = value;
2955 * write hardware register if we can
2957 if (can_access_pmu) {
2958 ia64_set_pmc(cnum, value);
2963 * per-task SMP only here
2965 * we are guaranteed that the task is not running on the other CPU,
2966 * we indicate that this PMD will need to be reloaded if the task
2967 * is rescheduled on the CPU it ran last on.
2969 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
2974 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",
2980 ctx->ctx_all_pmcs[0],
2981 ctx->ctx_used_pmds[0],
2982 ctx->ctx_pmds[cnum].eventid,
2985 ctx->ctx_reload_pmcs[0],
2986 ctx->ctx_used_monitors[0],
2987 ctx->ctx_ovfl_regs[0]));
2991 * make sure the changes are visible
2993 if (can_access_pmu) ia64_srlz_d();
2997 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3002 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3004 struct task_struct *task;
3005 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3006 unsigned long value, hw_value, ovfl_mask;
3008 int i, can_access_pmu = 0, state;
3009 int is_counting, is_loaded, is_system, expert_mode;
3011 pfm_reg_check_t wr_func;
3014 state = ctx->ctx_state;
3015 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3016 is_system = ctx->ctx_fl_system;
3017 ovfl_mask = pmu_conf->ovfl_val;
3018 task = ctx->ctx_task;
3020 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3023 * on both UP and SMP, we can only write to the PMC when the task is
3024 * the owner of the local PMU.
3026 if (likely(is_loaded)) {
3028 * In system wide and when the context is loaded, access can only happen
3029 * when the caller is running on the CPU being monitored by the session.
3030 * It does not have to be the owner (ctx_task) of the context per se.
3032 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3033 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3036 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3038 expert_mode = pfm_sysctl.expert_mode;
3040 for (i = 0; i < count; i++, req++) {
3042 cnum = req->reg_num;
3043 value = req->reg_value;
3045 if (!PMD_IS_IMPL(cnum)) {
3046 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3049 is_counting = PMD_IS_COUNTING(cnum);
3050 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3053 * execute write checker, if any
3055 if (unlikely(expert_mode == 0 && wr_func)) {
3056 unsigned long v = value;
3058 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3059 if (ret) goto abort_mission;
3066 * no error on this register
3068 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3071 * now commit changes to software state
3076 * update virtualized (64bits) counter
3080 * write context state
3082 ctx->ctx_pmds[cnum].lval = value;
3085 * when context is load we use the split value
3088 hw_value = value & ovfl_mask;
3089 value = value & ~ovfl_mask;
3093 * update reset values (not just for counters)
3095 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3096 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3099 * update randomization parameters (not just for counters)
3101 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3102 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3105 * update context value
3107 ctx->ctx_pmds[cnum].val = value;
3110 * Keep track of what we use
3112 * We do not keep track of PMC because we have to
3113 * systematically restore ALL of them.
3115 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3118 * mark this PMD register used as well
3120 CTX_USED_PMD(ctx, RDEP(cnum));
3123 * make sure we do not try to reset on
3124 * restart because we have established new values
3126 if (is_counting && state == PFM_CTX_MASKED) {
3127 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3132 * write thread state
3134 if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3137 * write hardware register if we can
3139 if (can_access_pmu) {
3140 ia64_set_pmd(cnum, hw_value);
3144 * we are guaranteed that the task is not running on the other CPU,
3145 * we indicate that this PMD will need to be reloaded if the task
3146 * is rescheduled on the CPU it ran last on.
3148 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3153 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3154 "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",
3160 ctx->ctx_pmds[cnum].val,
3161 ctx->ctx_pmds[cnum].short_reset,
3162 ctx->ctx_pmds[cnum].long_reset,
3163 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3164 ctx->ctx_pmds[cnum].seed,
3165 ctx->ctx_pmds[cnum].mask,
3166 ctx->ctx_used_pmds[0],
3167 ctx->ctx_pmds[cnum].reset_pmds[0],
3168 ctx->ctx_reload_pmds[0],
3169 ctx->ctx_all_pmds[0],
3170 ctx->ctx_ovfl_regs[0]));
3174 * make changes visible
3176 if (can_access_pmu) ia64_srlz_d();
3182 * for now, we have only one possibility for error
3184 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3189 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3190 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3191 * interrupt is delivered during the call, it will be kept pending until we leave, making
3192 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3193 * guaranteed to return consistent data to the user, it may simply be old. It is not
3194 * trivial to treat the overflow while inside the call because you may end up in
3195 * some module sampling buffer code causing deadlocks.
3198 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3200 struct task_struct *task;
3201 unsigned long val = 0UL, lval, ovfl_mask, sval;
3202 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3203 unsigned int cnum, reg_flags = 0;
3204 int i, can_access_pmu = 0, state;
3205 int is_loaded, is_system, is_counting, expert_mode;
3207 pfm_reg_check_t rd_func;
3210 * access is possible when loaded only for
3211 * self-monitoring tasks or in UP mode
3214 state = ctx->ctx_state;
3215 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3216 is_system = ctx->ctx_fl_system;
3217 ovfl_mask = pmu_conf->ovfl_val;
3218 task = ctx->ctx_task;
3220 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3222 if (likely(is_loaded)) {
3224 * In system wide and when the context is loaded, access can only happen
3225 * when the caller is running on the CPU being monitored by the session.
3226 * It does not have to be the owner (ctx_task) of the context per se.
3228 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3229 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3233 * this can be true when not self-monitoring only in UP
3235 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3237 if (can_access_pmu) ia64_srlz_d();
3239 expert_mode = pfm_sysctl.expert_mode;
3241 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3247 * on both UP and SMP, we can only read the PMD from the hardware register when
3248 * the task is the owner of the local PMU.
3251 for (i = 0; i < count; i++, req++) {
3253 cnum = req->reg_num;
3254 reg_flags = req->reg_flags;
3256 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3258 * we can only read the register that we use. That includes
3259 * the one we explicitly initialize AND the one we want included
3260 * in the sampling buffer (smpl_regs).
3262 * Having this restriction allows optimization in the ctxsw routine
3263 * without compromising security (leaks)
3265 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3267 sval = ctx->ctx_pmds[cnum].val;
3268 lval = ctx->ctx_pmds[cnum].lval;
3269 is_counting = PMD_IS_COUNTING(cnum);
3272 * If the task is not the current one, then we check if the
3273 * PMU state is still in the local live register due to lazy ctxsw.
3274 * If true, then we read directly from the registers.
3276 if (can_access_pmu){
3277 val = ia64_get_pmd(cnum);
3280 * context has been saved
3281 * if context is zombie, then task does not exist anymore.
3282 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3284 val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3286 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3290 * XXX: need to check for overflow when loaded
3297 * execute read checker, if any
3299 if (unlikely(expert_mode == 0 && rd_func)) {
3300 unsigned long v = val;
3301 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3302 if (ret) goto error;
3307 PFM_REG_RETFLAG_SET(reg_flags, 0);
3309 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3312 * update register return value, abort all if problem during copy.
3313 * we only modify the reg_flags field. no check mode is fine because
3314 * access has been verified upfront in sys_perfmonctl().
3316 req->reg_value = val;
3317 req->reg_flags = reg_flags;
3318 req->reg_last_reset_val = lval;
3324 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3329 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3333 if (req == NULL) return -EINVAL;
3335 ctx = GET_PMU_CTX();
3337 if (ctx == NULL) return -EINVAL;
3340 * for now limit to current task, which is enough when calling
3341 * from overflow handler
3343 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3345 return pfm_write_pmcs(ctx, req, nreq, regs);
3347 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3350 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3354 if (req == NULL) return -EINVAL;
3356 ctx = GET_PMU_CTX();
3358 if (ctx == NULL) return -EINVAL;
3361 * for now limit to current task, which is enough when calling
3362 * from overflow handler
3364 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3366 return pfm_read_pmds(ctx, req, nreq, regs);
3368 EXPORT_SYMBOL(pfm_mod_read_pmds);
3371 * Only call this function when a process it trying to
3372 * write the debug registers (reading is always allowed)
3375 pfm_use_debug_registers(struct task_struct *task)
3377 pfm_context_t *ctx = task->thread.pfm_context;
3378 unsigned long flags;
3381 if (pmu_conf->use_rr_dbregs == 0) return 0;
3383 DPRINT(("called for [%d]\n", task_pid_nr(task)));
3388 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3391 * Even on SMP, we do not need to use an atomic here because
3392 * the only way in is via ptrace() and this is possible only when the
3393 * process is stopped. Even in the case where the ctxsw out is not totally
3394 * completed by the time we come here, there is no way the 'stopped' process
3395 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3396 * So this is always safe.
3398 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3403 * We cannot allow setting breakpoints when system wide monitoring
3404 * sessions are using the debug registers.
3406 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3409 pfm_sessions.pfs_ptrace_use_dbregs++;
3411 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3412 pfm_sessions.pfs_ptrace_use_dbregs,
3413 pfm_sessions.pfs_sys_use_dbregs,
3414 task_pid_nr(task), ret));
3422 * This function is called for every task that exits with the
3423 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3424 * able to use the debug registers for debugging purposes via
3425 * ptrace(). Therefore we know it was not using them for
3426 * performance monitoring, so we only decrement the number
3427 * of "ptraced" debug register users to keep the count up to date
3430 pfm_release_debug_registers(struct task_struct *task)
3432 unsigned long flags;
3435 if (pmu_conf->use_rr_dbregs == 0) return 0;
3438 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3439 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3442 pfm_sessions.pfs_ptrace_use_dbregs--;
3451 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3453 struct task_struct *task;
3454 pfm_buffer_fmt_t *fmt;
3455 pfm_ovfl_ctrl_t rst_ctrl;
3456 int state, is_system;
3459 state = ctx->ctx_state;
3460 fmt = ctx->ctx_buf_fmt;
3461 is_system = ctx->ctx_fl_system;
3462 task = PFM_CTX_TASK(ctx);
3465 case PFM_CTX_MASKED:
3467 case PFM_CTX_LOADED:
3468 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3470 case PFM_CTX_UNLOADED:
3471 case PFM_CTX_ZOMBIE:
3472 DPRINT(("invalid state=%d\n", state));
3475 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3480 * In system wide and when the context is loaded, access can only happen
3481 * when the caller is running on the CPU being monitored by the session.
3482 * It does not have to be the owner (ctx_task) of the context per se.
3484 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3485 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3490 if (unlikely(task == NULL)) {
3491 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3495 if (task == current || is_system) {
3497 fmt = ctx->ctx_buf_fmt;
3499 DPRINT(("restarting self %d ovfl=0x%lx\n",
3501 ctx->ctx_ovfl_regs[0]));
3503 if (CTX_HAS_SMPL(ctx)) {
3505 prefetch(ctx->ctx_smpl_hdr);
3507 rst_ctrl.bits.mask_monitoring = 0;
3508 rst_ctrl.bits.reset_ovfl_pmds = 0;
3510 if (state == PFM_CTX_LOADED)
3511 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3513 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3515 rst_ctrl.bits.mask_monitoring = 0;
3516 rst_ctrl.bits.reset_ovfl_pmds = 1;
3520 if (rst_ctrl.bits.reset_ovfl_pmds)
3521 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3523 if (rst_ctrl.bits.mask_monitoring == 0) {
3524 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3526 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3528 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3530 // cannot use pfm_stop_monitoring(task, regs);
3534 * clear overflowed PMD mask to remove any stale information
3536 ctx->ctx_ovfl_regs[0] = 0UL;
3539 * back to LOADED state
3541 ctx->ctx_state = PFM_CTX_LOADED;
3544 * XXX: not really useful for self monitoring
3546 ctx->ctx_fl_can_restart = 0;
3552 * restart another task
3556 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3557 * one is seen by the task.
3559 if (state == PFM_CTX_MASKED) {
3560 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3562 * will prevent subsequent restart before this one is
3563 * seen by other task
3565 ctx->ctx_fl_can_restart = 0;
3569 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3570 * the task is blocked or on its way to block. That's the normal
3571 * restart path. If the monitoring is not masked, then the task
3572 * can be actively monitoring and we cannot directly intervene.
3573 * Therefore we use the trap mechanism to catch the task and
3574 * force it to reset the buffer/reset PMDs.
3576 * if non-blocking, then we ensure that the task will go into
3577 * pfm_handle_work() before returning to user mode.
3579 * We cannot explicitly reset another task, it MUST always
3580 * be done by the task itself. This works for system wide because
3581 * the tool that is controlling the session is logically doing
3582 * "self-monitoring".
3584 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3585 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3586 complete(&ctx->ctx_restart_done);
3588 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3590 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3592 PFM_SET_WORK_PENDING(task, 1);
3594 set_notify_resume(task);
3597 * XXX: send reschedule if task runs on another CPU
3604 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3606 unsigned int m = *(unsigned int *)arg;
3608 pfm_sysctl.debug = m == 0 ? 0 : 1;
3610 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3613 memset(pfm_stats, 0, sizeof(pfm_stats));
3614 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3620 * arg can be NULL and count can be zero for this function
3623 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3625 struct thread_struct *thread = NULL;
3626 struct task_struct *task;
3627 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3628 unsigned long flags;
3633 int i, can_access_pmu = 0;
3634 int is_system, is_loaded;
3636 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3638 state = ctx->ctx_state;
3639 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3640 is_system = ctx->ctx_fl_system;
3641 task = ctx->ctx_task;
3643 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3646 * on both UP and SMP, we can only write to the PMC when the task is
3647 * the owner of the local PMU.
3650 thread = &task->thread;
3652 * In system wide and when the context is loaded, access can only happen
3653 * when the caller is running on the CPU being monitored by the session.
3654 * It does not have to be the owner (ctx_task) of the context per se.
3656 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3657 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3660 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3664 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3665 * ensuring that no real breakpoint can be installed via this call.
3667 * IMPORTANT: regs can be NULL in this function
3670 first_time = ctx->ctx_fl_using_dbreg == 0;
3673 * don't bother if we are loaded and task is being debugged
3675 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3676 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3681 * check for debug registers in system wide mode
3683 * If though a check is done in pfm_context_load(),
3684 * we must repeat it here, in case the registers are
3685 * written after the context is loaded
3690 if (first_time && is_system) {
3691 if (pfm_sessions.pfs_ptrace_use_dbregs)
3694 pfm_sessions.pfs_sys_use_dbregs++;
3699 if (ret != 0) return ret;
3702 * mark ourself as user of the debug registers for
3705 ctx->ctx_fl_using_dbreg = 1;
3708 * clear hardware registers to make sure we don't
3709 * pick up stale state.
3711 * for a system wide session, we do not use
3712 * thread.dbr, thread.ibr because this process
3713 * never leaves the current CPU and the state
3714 * is shared by all processes running on it
3716 if (first_time && can_access_pmu) {
3717 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3718 for (i=0; i < pmu_conf->num_ibrs; i++) {
3719 ia64_set_ibr(i, 0UL);
3720 ia64_dv_serialize_instruction();
3723 for (i=0; i < pmu_conf->num_dbrs; i++) {
3724 ia64_set_dbr(i, 0UL);
3725 ia64_dv_serialize_data();
3731 * Now install the values into the registers
3733 for (i = 0; i < count; i++, req++) {
3735 rnum = req->dbreg_num;
3736 dbreg.val = req->dbreg_value;
3740 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3741 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3742 rnum, dbreg.val, mode, i, count));
3748 * make sure we do not install enabled breakpoint
3751 if (mode == PFM_CODE_RR)
3752 dbreg.ibr.ibr_x = 0;
3754 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3757 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3760 * Debug registers, just like PMC, can only be modified
3761 * by a kernel call. Moreover, perfmon() access to those
3762 * registers are centralized in this routine. The hardware
3763 * does not modify the value of these registers, therefore,
3764 * if we save them as they are written, we can avoid having
3765 * to save them on context switch out. This is made possible
3766 * by the fact that when perfmon uses debug registers, ptrace()
3767 * won't be able to modify them concurrently.
3769 if (mode == PFM_CODE_RR) {
3770 CTX_USED_IBR(ctx, rnum);
3772 if (can_access_pmu) {
3773 ia64_set_ibr(rnum, dbreg.val);
3774 ia64_dv_serialize_instruction();
3777 ctx->ctx_ibrs[rnum] = dbreg.val;
3779 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3780 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3782 CTX_USED_DBR(ctx, rnum);
3784 if (can_access_pmu) {
3785 ia64_set_dbr(rnum, dbreg.val);
3786 ia64_dv_serialize_data();
3788 ctx->ctx_dbrs[rnum] = dbreg.val;
3790 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3791 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3799 * in case it was our first attempt, we undo the global modifications
3803 if (ctx->ctx_fl_system) {
3804 pfm_sessions.pfs_sys_use_dbregs--;
3807 ctx->ctx_fl_using_dbreg = 0;
3810 * install error return flag
3812 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3818 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3820 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3824 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3826 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3830 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3834 if (req == NULL) return -EINVAL;
3836 ctx = GET_PMU_CTX();
3838 if (ctx == NULL) return -EINVAL;
3841 * for now limit to current task, which is enough when calling
3842 * from overflow handler
3844 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3846 return pfm_write_ibrs(ctx, req, nreq, regs);
3848 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3851 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3855 if (req == NULL) return -EINVAL;
3857 ctx = GET_PMU_CTX();
3859 if (ctx == NULL) return -EINVAL;
3862 * for now limit to current task, which is enough when calling
3863 * from overflow handler
3865 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3867 return pfm_write_dbrs(ctx, req, nreq, regs);
3869 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3873 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3875 pfarg_features_t *req = (pfarg_features_t *)arg;
3877 req->ft_version = PFM_VERSION;
3882 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3884 struct pt_regs *tregs;
3885 struct task_struct *task = PFM_CTX_TASK(ctx);
3886 int state, is_system;
3888 state = ctx->ctx_state;
3889 is_system = ctx->ctx_fl_system;
3892 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3894 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3897 * In system wide and when the context is loaded, access can only happen
3898 * when the caller is running on the CPU being monitored by the session.
3899 * It does not have to be the owner (ctx_task) of the context per se.
3901 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3902 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3905 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3906 task_pid_nr(PFM_CTX_TASK(ctx)),
3910 * in system mode, we need to update the PMU directly
3911 * and the user level state of the caller, which may not
3912 * necessarily be the creator of the context.
3916 * Update local PMU first
3920 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
3924 * update local cpuinfo
3926 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
3929 * stop monitoring, does srlz.i
3934 * stop monitoring in the caller
3936 ia64_psr(regs)->pp = 0;
3944 if (task == current) {
3945 /* stop monitoring at kernel level */
3949 * stop monitoring at the user level
3951 ia64_psr(regs)->up = 0;
3953 tregs = task_pt_regs(task);
3956 * stop monitoring at the user level
3958 ia64_psr(tregs)->up = 0;
3961 * monitoring disabled in kernel at next reschedule
3963 ctx->ctx_saved_psr_up = 0;
3964 DPRINT(("task=[%d]\n", task_pid_nr(task)));
3971 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3973 struct pt_regs *tregs;
3974 int state, is_system;
3976 state = ctx->ctx_state;
3977 is_system = ctx->ctx_fl_system;
3979 if (state != PFM_CTX_LOADED) return -EINVAL;
3982 * In system wide and when the context is loaded, access can only happen
3983 * when the caller is running on the CPU being monitored by the session.
3984 * It does not have to be the owner (ctx_task) of the context per se.
3986 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3987 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3992 * in system mode, we need to update the PMU directly
3993 * and the user level state of the caller, which may not
3994 * necessarily be the creator of the context.
3999 * set user level psr.pp for the caller
4001 ia64_psr(regs)->pp = 1;
4004 * now update the local PMU and cpuinfo
4006 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4009 * start monitoring at kernel level
4014 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4024 if (ctx->ctx_task == current) {
4026 /* start monitoring at kernel level */
4030 * activate monitoring at user level
4032 ia64_psr(regs)->up = 1;
4035 tregs = task_pt_regs(ctx->ctx_task);
4038 * start monitoring at the kernel level the next
4039 * time the task is scheduled
4041 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4044 * activate monitoring at user level
4046 ia64_psr(tregs)->up = 1;
4052 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4054 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4059 for (i = 0; i < count; i++, req++) {
4061 cnum = req->reg_num;
4063 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4065 req->reg_value = PMC_DFL_VAL(cnum);
4067 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4069 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4074 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4079 pfm_check_task_exist(pfm_context_t *ctx)
4081 struct task_struct *g, *t;
4084 read_lock(&tasklist_lock);
4086 do_each_thread (g, t) {
4087 if (t->thread.pfm_context == ctx) {
4091 } while_each_thread (g, t);
4093 read_unlock(&tasklist_lock);
4095 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4101 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4103 struct task_struct *task;
4104 struct thread_struct *thread;
4105 struct pfm_context_t *old;
4106 unsigned long flags;
4108 struct task_struct *owner_task = NULL;
4110 pfarg_load_t *req = (pfarg_load_t *)arg;
4111 unsigned long *pmcs_source, *pmds_source;
4114 int state, is_system, set_dbregs = 0;
4116 state = ctx->ctx_state;
4117 is_system = ctx->ctx_fl_system;
4119 * can only load from unloaded or terminated state
4121 if (state != PFM_CTX_UNLOADED) {
4122 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4128 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4130 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4131 DPRINT(("cannot use blocking mode on self\n"));
4135 ret = pfm_get_task(ctx, req->load_pid, &task);
4137 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4144 * system wide is self monitoring only
4146 if (is_system && task != current) {
4147 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4152 thread = &task->thread;
4156 * cannot load a context which is using range restrictions,
4157 * into a task that is being debugged.
4159 if (ctx->ctx_fl_using_dbreg) {
4160 if (thread->flags & IA64_THREAD_DBG_VALID) {
4162 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4168 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4169 DPRINT(("cannot load [%d] dbregs in use\n",
4170 task_pid_nr(task)));
4173 pfm_sessions.pfs_sys_use_dbregs++;
4174 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4181 if (ret) goto error;
4185 * SMP system-wide monitoring implies self-monitoring.
4187 * The programming model expects the task to
4188 * be pinned on a CPU throughout the session.
4189 * Here we take note of the current CPU at the
4190 * time the context is loaded. No call from
4191 * another CPU will be allowed.
4193 * The pinning via shed_setaffinity()
4194 * must be done by the calling task prior
4197 * systemwide: keep track of CPU this session is supposed to run on
4199 the_cpu = ctx->ctx_cpu = smp_processor_id();
4203 * now reserve the session
4205 ret = pfm_reserve_session(current, is_system, the_cpu);
4206 if (ret) goto error;
4209 * task is necessarily stopped at this point.
4211 * If the previous context was zombie, then it got removed in
4212 * pfm_save_regs(). Therefore we should not see it here.
4213 * If we see a context, then this is an active context
4215 * XXX: needs to be atomic
4217 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4218 thread->pfm_context, ctx));
4221 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4223 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4227 pfm_reset_msgq(ctx);
4229 ctx->ctx_state = PFM_CTX_LOADED;
4232 * link context to task
4234 ctx->ctx_task = task;
4238 * we load as stopped
4240 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4241 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4243 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4245 thread->flags |= IA64_THREAD_PM_VALID;
4249 * propagate into thread-state
4251 pfm_copy_pmds(task, ctx);
4252 pfm_copy_pmcs(task, ctx);
4254 pmcs_source = ctx->th_pmcs;
4255 pmds_source = ctx->th_pmds;
4258 * always the case for system-wide
4260 if (task == current) {
4262 if (is_system == 0) {
4264 /* allow user level control */
4265 ia64_psr(regs)->sp = 0;
4266 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4268 SET_LAST_CPU(ctx, smp_processor_id());
4270 SET_ACTIVATION(ctx);
4273 * push the other task out, if any
4275 owner_task = GET_PMU_OWNER();
4276 if (owner_task) pfm_lazy_save_regs(owner_task);
4280 * load all PMD from ctx to PMU (as opposed to thread state)
4281 * restore all PMC from ctx to PMU
4283 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4284 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4286 ctx->ctx_reload_pmcs[0] = 0UL;
4287 ctx->ctx_reload_pmds[0] = 0UL;
4290 * guaranteed safe by earlier check against DBG_VALID
4292 if (ctx->ctx_fl_using_dbreg) {
4293 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4294 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4299 SET_PMU_OWNER(task, ctx);
4301 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4304 * when not current, task MUST be stopped, so this is safe
4306 regs = task_pt_regs(task);
4308 /* force a full reload */
4309 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4310 SET_LAST_CPU(ctx, -1);
4312 /* initial saved psr (stopped) */
4313 ctx->ctx_saved_psr_up = 0UL;
4314 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4320 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4323 * we must undo the dbregs setting (for system-wide)
4325 if (ret && set_dbregs) {
4327 pfm_sessions.pfs_sys_use_dbregs--;
4331 * release task, there is now a link with the context
4333 if (is_system == 0 && task != current) {
4337 ret = pfm_check_task_exist(ctx);
4339 ctx->ctx_state = PFM_CTX_UNLOADED;
4340 ctx->ctx_task = NULL;
4348 * in this function, we do not need to increase the use count
4349 * for the task via get_task_struct(), because we hold the
4350 * context lock. If the task were to disappear while having
4351 * a context attached, it would go through pfm_exit_thread()
4352 * which also grabs the context lock and would therefore be blocked
4353 * until we are here.
4355 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4358 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4360 struct task_struct *task = PFM_CTX_TASK(ctx);
4361 struct pt_regs *tregs;
4362 int prev_state, is_system;
4365 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4367 prev_state = ctx->ctx_state;
4368 is_system = ctx->ctx_fl_system;
4371 * unload only when necessary
4373 if (prev_state == PFM_CTX_UNLOADED) {
4374 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4379 * clear psr and dcr bits
4381 ret = pfm_stop(ctx, NULL, 0, regs);
4382 if (ret) return ret;
4384 ctx->ctx_state = PFM_CTX_UNLOADED;
4387 * in system mode, we need to update the PMU directly
4388 * and the user level state of the caller, which may not
4389 * necessarily be the creator of the context.
4396 * local PMU is taken care of in pfm_stop()
4398 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4399 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4402 * save PMDs in context
4405 pfm_flush_pmds(current, ctx);
4408 * at this point we are done with the PMU
4409 * so we can unreserve the resource.
4411 if (prev_state != PFM_CTX_ZOMBIE)
4412 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4415 * disconnect context from task
4417 task->thread.pfm_context = NULL;
4419 * disconnect task from context
4421 ctx->ctx_task = NULL;
4424 * There is nothing more to cleanup here.
4432 tregs = task == current ? regs : task_pt_regs(task);
4434 if (task == current) {
4436 * cancel user level control
4438 ia64_psr(regs)->sp = 1;
4440 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4443 * save PMDs to context
4446 pfm_flush_pmds(task, ctx);
4449 * at this point we are done with the PMU
4450 * so we can unreserve the resource.
4452 * when state was ZOMBIE, we have already unreserved.
4454 if (prev_state != PFM_CTX_ZOMBIE)
4455 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4458 * reset activation counter and psr
4460 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4461 SET_LAST_CPU(ctx, -1);
4464 * PMU state will not be restored
4466 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4469 * break links between context and task
4471 task->thread.pfm_context = NULL;
4472 ctx->ctx_task = NULL;
4474 PFM_SET_WORK_PENDING(task, 0);
4476 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4477 ctx->ctx_fl_can_restart = 0;
4478 ctx->ctx_fl_going_zombie = 0;
4480 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4487 * called only from exit_thread()
4488 * we come here only if the task has a context attached (loaded or masked)
4491 pfm_exit_thread(struct task_struct *task)
4494 unsigned long flags;
4495 struct pt_regs *regs = task_pt_regs(task);
4499 ctx = PFM_GET_CTX(task);
4501 PROTECT_CTX(ctx, flags);
4503 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4505 state = ctx->ctx_state;
4507 case PFM_CTX_UNLOADED:
4509 * only comes to this function if pfm_context is not NULL, i.e., cannot
4510 * be in unloaded state
4512 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4514 case PFM_CTX_LOADED:
4515 case PFM_CTX_MASKED:
4516 ret = pfm_context_unload(ctx, NULL, 0, regs);
4518 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4520 DPRINT(("ctx unloaded for current state was %d\n", state));
4522 pfm_end_notify_user(ctx);
4524 case PFM_CTX_ZOMBIE:
4525 ret = pfm_context_unload(ctx, NULL, 0, regs);
4527 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4532 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4535 UNPROTECT_CTX(ctx, flags);
4537 { u64 psr = pfm_get_psr();
4538 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4539 BUG_ON(GET_PMU_OWNER());
4540 BUG_ON(ia64_psr(regs)->up);
4541 BUG_ON(ia64_psr(regs)->pp);
4545 * All memory free operations (especially for vmalloc'ed memory)
4546 * MUST be done with interrupts ENABLED.
4548 if (free_ok) pfm_context_free(ctx);
4552 * functions MUST be listed in the increasing order of their index (see permfon.h)
4554 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4555 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4556 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4557 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4558 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4560 static pfm_cmd_desc_t pfm_cmd_tab[]={
4561 /* 0 */PFM_CMD_NONE,
4562 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4563 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4564 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4565 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4566 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4567 /* 6 */PFM_CMD_NONE,
4568 /* 7 */PFM_CMD_NONE,
4569 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4570 /* 9 */PFM_CMD_NONE,
4571 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4572 /* 11 */PFM_CMD_NONE,
4573 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4574 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4575 /* 14 */PFM_CMD_NONE,
4576 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4577 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4578 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4579 /* 18 */PFM_CMD_NONE,
4580 /* 19 */PFM_CMD_NONE,
4581 /* 20 */PFM_CMD_NONE,
4582 /* 21 */PFM_CMD_NONE,
4583 /* 22 */PFM_CMD_NONE,
4584 /* 23 */PFM_CMD_NONE,
4585 /* 24 */PFM_CMD_NONE,
4586 /* 25 */PFM_CMD_NONE,
4587 /* 26 */PFM_CMD_NONE,
4588 /* 27 */PFM_CMD_NONE,
4589 /* 28 */PFM_CMD_NONE,
4590 /* 29 */PFM_CMD_NONE,
4591 /* 30 */PFM_CMD_NONE,
4592 /* 31 */PFM_CMD_NONE,
4593 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4594 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4596 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4599 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4601 struct task_struct *task;
4602 int state, old_state;
4605 state = ctx->ctx_state;
4606 task = ctx->ctx_task;
4609 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4613 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4617 task->state, PFM_CMD_STOPPED(cmd)));
4620 * self-monitoring always ok.
4622 * for system-wide the caller can either be the creator of the
4623 * context (to one to which the context is attached to) OR
4624 * a task running on the same CPU as the session.
4626 if (task == current || ctx->ctx_fl_system) return 0;
4629 * we are monitoring another thread
4632 case PFM_CTX_UNLOADED:
4634 * if context is UNLOADED we are safe to go
4637 case PFM_CTX_ZOMBIE:
4639 * no command can operate on a zombie context
4641 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4643 case PFM_CTX_MASKED:
4645 * PMU state has been saved to software even though
4646 * the thread may still be running.
4648 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4652 * context is LOADED or MASKED. Some commands may need to have
4655 * We could lift this restriction for UP but it would mean that
4656 * the user has no guarantee the task would not run between
4657 * two successive calls to perfmonctl(). That's probably OK.
4658 * If this user wants to ensure the task does not run, then
4659 * the task must be stopped.
4661 if (PFM_CMD_STOPPED(cmd)) {
4662 if (!task_is_stopped_or_traced(task)) {
4663 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4667 * task is now stopped, wait for ctxsw out
4669 * This is an interesting point in the code.
4670 * We need to unprotect the context because
4671 * the pfm_save_regs() routines needs to grab
4672 * the same lock. There are danger in doing
4673 * this because it leaves a window open for
4674 * another task to get access to the context
4675 * and possibly change its state. The one thing
4676 * that is not possible is for the context to disappear
4677 * because we are protected by the VFS layer, i.e.,
4678 * get_fd()/put_fd().
4682 UNPROTECT_CTX(ctx, flags);
4684 wait_task_inactive(task, 0);
4686 PROTECT_CTX(ctx, flags);
4689 * we must recheck to verify if state has changed
4691 if (ctx->ctx_state != old_state) {
4692 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4700 * system-call entry point (must return long)
4703 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4705 struct fd f = {NULL, 0};
4706 pfm_context_t *ctx = NULL;
4707 unsigned long flags = 0UL;
4708 void *args_k = NULL;
4709 long ret; /* will expand int return types */
4710 size_t base_sz, sz, xtra_sz = 0;
4711 int narg, completed_args = 0, call_made = 0, cmd_flags;
4712 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4713 int (*getsize)(void *arg, size_t *sz);
4714 #define PFM_MAX_ARGSIZE 4096
4717 * reject any call if perfmon was disabled at initialization
4719 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4721 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4722 DPRINT(("invalid cmd=%d\n", cmd));
4726 func = pfm_cmd_tab[cmd].cmd_func;
4727 narg = pfm_cmd_tab[cmd].cmd_narg;
4728 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4729 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4730 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4732 if (unlikely(func == NULL)) {
4733 DPRINT(("invalid cmd=%d\n", cmd));
4737 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4745 * check if number of arguments matches what the command expects
4747 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4751 sz = xtra_sz + base_sz*count;
4753 * limit abuse to min page size
4755 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4756 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4761 * allocate default-sized argument buffer
4763 if (likely(count && args_k == NULL)) {
4764 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4765 if (args_k == NULL) return -ENOMEM;
4773 * assume sz = 0 for command without parameters
4775 if (sz && copy_from_user(args_k, arg, sz)) {
4776 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4781 * check if command supports extra parameters
4783 if (completed_args == 0 && getsize) {
4785 * get extra parameters size (based on main argument)
4787 ret = (*getsize)(args_k, &xtra_sz);
4788 if (ret) goto error_args;
4792 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4794 /* retry if necessary */
4795 if (likely(xtra_sz)) goto restart_args;
4798 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4803 if (unlikely(f.file == NULL)) {
4804 DPRINT(("invalid fd %d\n", fd));
4807 if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4808 DPRINT(("fd %d not related to perfmon\n", fd));
4812 ctx = f.file->private_data;
4813 if (unlikely(ctx == NULL)) {
4814 DPRINT(("no context for fd %d\n", fd));
4817 prefetch(&ctx->ctx_state);
4819 PROTECT_CTX(ctx, flags);
4822 * check task is stopped
4824 ret = pfm_check_task_state(ctx, cmd, flags);
4825 if (unlikely(ret)) goto abort_locked;
4828 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4834 DPRINT(("context unlocked\n"));
4835 UNPROTECT_CTX(ctx, flags);
4838 /* copy argument back to user, if needed */
4839 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4847 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4853 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4855 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4856 pfm_ovfl_ctrl_t rst_ctrl;
4860 state = ctx->ctx_state;
4862 * Unlock sampling buffer and reset index atomically
4863 * XXX: not really needed when blocking
4865 if (CTX_HAS_SMPL(ctx)) {
4867 rst_ctrl.bits.mask_monitoring = 0;
4868 rst_ctrl.bits.reset_ovfl_pmds = 0;
4870 if (state == PFM_CTX_LOADED)
4871 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4873 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4875 rst_ctrl.bits.mask_monitoring = 0;
4876 rst_ctrl.bits.reset_ovfl_pmds = 1;
4880 if (rst_ctrl.bits.reset_ovfl_pmds) {
4881 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4883 if (rst_ctrl.bits.mask_monitoring == 0) {
4884 DPRINT(("resuming monitoring\n"));
4885 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4887 DPRINT(("stopping monitoring\n"));
4888 //pfm_stop_monitoring(current, regs);
4890 ctx->ctx_state = PFM_CTX_LOADED;
4895 * context MUST BE LOCKED when calling
4896 * can only be called for current
4899 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4903 DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4905 ret = pfm_context_unload(ctx, NULL, 0, regs);
4907 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4911 * and wakeup controlling task, indicating we are now disconnected
4913 wake_up_interruptible(&ctx->ctx_zombieq);
4916 * given that context is still locked, the controlling
4917 * task will only get access when we return from
4918 * pfm_handle_work().
4922 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4925 * pfm_handle_work() can be called with interrupts enabled
4926 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4927 * call may sleep, therefore we must re-enable interrupts
4928 * to avoid deadlocks. It is safe to do so because this function
4929 * is called ONLY when returning to user level (pUStk=1), in which case
4930 * there is no risk of kernel stack overflow due to deep
4931 * interrupt nesting.
4934 pfm_handle_work(void)
4937 struct pt_regs *regs;
4938 unsigned long flags, dummy_flags;
4939 unsigned long ovfl_regs;
4940 unsigned int reason;
4943 ctx = PFM_GET_CTX(current);
4945 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
4946 task_pid_nr(current));
4950 PROTECT_CTX(ctx, flags);
4952 PFM_SET_WORK_PENDING(current, 0);
4954 regs = task_pt_regs(current);
4957 * extract reason for being here and clear
4959 reason = ctx->ctx_fl_trap_reason;
4960 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4961 ovfl_regs = ctx->ctx_ovfl_regs[0];
4963 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
4966 * must be done before we check for simple-reset mode
4968 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
4971 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
4972 if (reason == PFM_TRAP_REASON_RESET)
4976 * restore interrupt mask to what it was on entry.
4977 * Could be enabled/diasbled.
4979 UNPROTECT_CTX(ctx, flags);
4982 * force interrupt enable because of down_interruptible()
4986 DPRINT(("before block sleeping\n"));
4989 * may go through without blocking on SMP systems
4990 * if restart has been received already by the time we call down()
4992 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
4994 DPRINT(("after block sleeping ret=%d\n", ret));
4997 * lock context and mask interrupts again
4998 * We save flags into a dummy because we may have
4999 * altered interrupts mask compared to entry in this
5002 PROTECT_CTX(ctx, dummy_flags);
5005 * we need to read the ovfl_regs only after wake-up
5006 * because we may have had pfm_write_pmds() in between
5007 * and that can changed PMD values and therefore
5008 * ovfl_regs is reset for these new PMD values.
5010 ovfl_regs = ctx->ctx_ovfl_regs[0];
5012 if (ctx->ctx_fl_going_zombie) {
5014 DPRINT(("context is zombie, bailing out\n"));
5015 pfm_context_force_terminate(ctx, regs);
5019 * in case of interruption of down() we don't restart anything
5025 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5026 ctx->ctx_ovfl_regs[0] = 0UL;
5030 * restore flags as they were upon entry
5032 UNPROTECT_CTX(ctx, flags);
5036 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5038 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5039 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5043 DPRINT(("waking up somebody\n"));
5045 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5048 * safe, we are not in intr handler, nor in ctxsw when
5051 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5057 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5059 pfm_msg_t *msg = NULL;
5061 if (ctx->ctx_fl_no_msg == 0) {
5062 msg = pfm_get_new_msg(ctx);
5064 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5068 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5069 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5070 msg->pfm_ovfl_msg.msg_active_set = 0;
5071 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5072 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5073 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5074 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5075 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5078 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5084 return pfm_notify_user(ctx, msg);
5088 pfm_end_notify_user(pfm_context_t *ctx)
5092 msg = pfm_get_new_msg(ctx);
5094 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5098 memset(msg, 0, sizeof(*msg));
5100 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5101 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5102 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5104 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5109 return pfm_notify_user(ctx, msg);
5113 * main overflow processing routine.
5114 * it can be called from the interrupt path or explicitly during the context switch code
5116 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5117 unsigned long pmc0, struct pt_regs *regs)
5119 pfm_ovfl_arg_t *ovfl_arg;
5121 unsigned long old_val, ovfl_val, new_val;
5122 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5123 unsigned long tstamp;
5124 pfm_ovfl_ctrl_t ovfl_ctrl;
5125 unsigned int i, has_smpl;
5126 int must_notify = 0;
5128 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5131 * sanity test. Should never happen
5133 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5135 tstamp = ia64_get_itc();
5136 mask = pmc0 >> PMU_FIRST_COUNTER;
5137 ovfl_val = pmu_conf->ovfl_val;
5138 has_smpl = CTX_HAS_SMPL(ctx);
5140 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5141 "used_pmds=0x%lx\n",
5143 task ? task_pid_nr(task): -1,
5144 (regs ? regs->cr_iip : 0),
5145 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5146 ctx->ctx_used_pmds[0]));
5150 * first we update the virtual counters
5151 * assume there was a prior ia64_srlz_d() issued
5153 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5155 /* skip pmd which did not overflow */
5156 if ((mask & 0x1) == 0) continue;
5159 * Note that the pmd is not necessarily 0 at this point as qualified events
5160 * may have happened before the PMU was frozen. The residual count is not
5161 * taken into consideration here but will be with any read of the pmd via
5164 old_val = new_val = ctx->ctx_pmds[i].val;
5165 new_val += 1 + ovfl_val;
5166 ctx->ctx_pmds[i].val = new_val;
5169 * check for overflow condition
5171 if (likely(old_val > new_val)) {
5172 ovfl_pmds |= 1UL << i;
5173 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5176 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5180 ia64_get_pmd(i) & ovfl_val,
5186 * there was no 64-bit overflow, nothing else to do
5188 if (ovfl_pmds == 0UL) return;
5191 * reset all control bits
5197 * if a sampling format module exists, then we "cache" the overflow by
5198 * calling the module's handler() routine.
5201 unsigned long start_cycles, end_cycles;
5202 unsigned long pmd_mask;
5204 int this_cpu = smp_processor_id();
5206 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5207 ovfl_arg = &ctx->ctx_ovfl_arg;
5209 prefetch(ctx->ctx_smpl_hdr);
5211 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5215 if ((pmd_mask & 0x1) == 0) continue;
5217 ovfl_arg->ovfl_pmd = (unsigned char )i;
5218 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5219 ovfl_arg->active_set = 0;
5220 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5221 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5223 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5224 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5225 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5228 * copy values of pmds of interest. Sampling format may copy them
5229 * into sampling buffer.
5232 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5233 if ((smpl_pmds & 0x1) == 0) continue;
5234 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5235 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5239 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5241 start_cycles = ia64_get_itc();
5244 * call custom buffer format record (handler) routine
5246 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5248 end_cycles = ia64_get_itc();
5251 * For those controls, we take the union because they have
5252 * an all or nothing behavior.
5254 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5255 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5256 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5258 * build the bitmask of pmds to reset now
5260 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5262 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5265 * when the module cannot handle the rest of the overflows, we abort right here
5267 if (ret && pmd_mask) {
5268 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5269 pmd_mask<<PMU_FIRST_COUNTER));
5272 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5274 ovfl_pmds &= ~reset_pmds;
5277 * when no sampling module is used, then the default
5278 * is to notify on overflow if requested by user
5280 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5281 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5282 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5283 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5285 * if needed, we reset all overflowed pmds
5287 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5290 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5293 * reset the requested PMD registers using the short reset values
5296 unsigned long bm = reset_pmds;
5297 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5300 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5302 * keep track of what to reset when unblocking
5304 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5307 * check for blocking context
5309 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5311 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5314 * set the perfmon specific checking pending work for the task
5316 PFM_SET_WORK_PENDING(task, 1);
5319 * when coming from ctxsw, current still points to the
5320 * previous task, therefore we must work with task and not current.
5322 set_notify_resume(task);
5325 * defer until state is changed (shorten spin window). the context is locked
5326 * anyway, so the signal receiver would come spin for nothing.
5331 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5332 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5333 PFM_GET_WORK_PENDING(task),
5334 ctx->ctx_fl_trap_reason,
5337 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5339 * in case monitoring must be stopped, we toggle the psr bits
5341 if (ovfl_ctrl.bits.mask_monitoring) {
5342 pfm_mask_monitoring(task);
5343 ctx->ctx_state = PFM_CTX_MASKED;
5344 ctx->ctx_fl_can_restart = 1;
5348 * send notification now
5350 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5355 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5357 task ? task_pid_nr(task) : -1,
5363 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5364 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5365 * come here as zombie only if the task is the current task. In which case, we
5366 * can access the PMU hardware directly.
5368 * Note that zombies do have PM_VALID set. So here we do the minimal.
5370 * In case the context was zombified it could not be reclaimed at the time
5371 * the monitoring program exited. At this point, the PMU reservation has been
5372 * returned, the sampiing buffer has been freed. We must convert this call
5373 * into a spurious interrupt. However, we must also avoid infinite overflows
5374 * by stopping monitoring for this task. We can only come here for a per-task
5375 * context. All we need to do is to stop monitoring using the psr bits which
5376 * are always task private. By re-enabling secure montioring, we ensure that
5377 * the monitored task will not be able to re-activate monitoring.
5378 * The task will eventually be context switched out, at which point the context
5379 * will be reclaimed (that includes releasing ownership of the PMU).
5381 * So there might be a window of time where the number of per-task session is zero
5382 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5383 * context. This is safe because if a per-task session comes in, it will push this one
5384 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5385 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5386 * also push our zombie context out.
5388 * Overall pretty hairy stuff....
5390 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5392 ia64_psr(regs)->up = 0;
5393 ia64_psr(regs)->sp = 1;
5398 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5400 struct task_struct *task;
5402 unsigned long flags;
5404 int this_cpu = smp_processor_id();
5407 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5410 * srlz.d done before arriving here
5412 pmc0 = ia64_get_pmc(0);
5414 task = GET_PMU_OWNER();
5415 ctx = GET_PMU_CTX();
5418 * if we have some pending bits set
5419 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5421 if (PMC0_HAS_OVFL(pmc0) && task) {
5423 * we assume that pmc0.fr is always set here
5427 if (!ctx) goto report_spurious1;
5429 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5430 goto report_spurious2;
5432 PROTECT_CTX_NOPRINT(ctx, flags);
5434 pfm_overflow_handler(task, ctx, pmc0, regs);
5436 UNPROTECT_CTX_NOPRINT(ctx, flags);
5439 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5443 * keep it unfrozen at all times
5450 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5451 this_cpu, task_pid_nr(task));
5455 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5463 pfm_interrupt_handler(int irq, void *arg)
5465 unsigned long start_cycles, total_cycles;
5466 unsigned long min, max;
5469 struct pt_regs *regs = get_irq_regs();
5471 this_cpu = get_cpu();
5472 if (likely(!pfm_alt_intr_handler)) {
5473 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5474 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5476 start_cycles = ia64_get_itc();
5478 ret = pfm_do_interrupt_handler(arg, regs);
5480 total_cycles = ia64_get_itc();
5483 * don't measure spurious interrupts
5485 if (likely(ret == 0)) {
5486 total_cycles -= start_cycles;
5488 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5489 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5491 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5495 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5503 * /proc/perfmon interface, for debug only
5506 #define PFM_PROC_SHOW_HEADER ((void *)(long)nr_cpu_ids+1)
5509 pfm_proc_start(struct seq_file *m, loff_t *pos)
5512 return PFM_PROC_SHOW_HEADER;
5515 while (*pos <= nr_cpu_ids) {
5516 if (cpu_online(*pos - 1)) {
5517 return (void *)*pos;
5525 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5528 return pfm_proc_start(m, pos);
5532 pfm_proc_stop(struct seq_file *m, void *v)
5537 pfm_proc_show_header(struct seq_file *m)
5539 struct list_head * pos;
5540 pfm_buffer_fmt_t * entry;
5541 unsigned long flags;
5544 "perfmon version : %u.%u\n"
5547 "expert mode : %s\n"
5548 "ovfl_mask : 0x%lx\n"
5549 "PMU flags : 0x%x\n",
5550 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5552 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5553 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5560 "proc_sessions : %u\n"
5561 "sys_sessions : %u\n"
5562 "sys_use_dbregs : %u\n"
5563 "ptrace_use_dbregs : %u\n",
5564 pfm_sessions.pfs_task_sessions,
5565 pfm_sessions.pfs_sys_sessions,
5566 pfm_sessions.pfs_sys_use_dbregs,
5567 pfm_sessions.pfs_ptrace_use_dbregs);
5571 spin_lock(&pfm_buffer_fmt_lock);
5573 list_for_each(pos, &pfm_buffer_fmt_list) {
5574 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5575 seq_printf(m, "format : %16phD %s\n",
5576 entry->fmt_uuid, entry->fmt_name);
5578 spin_unlock(&pfm_buffer_fmt_lock);
5583 pfm_proc_show(struct seq_file *m, void *v)
5589 if (v == PFM_PROC_SHOW_HEADER) {
5590 pfm_proc_show_header(m);
5594 /* show info for CPU (v - 1) */
5598 "CPU%-2d overflow intrs : %lu\n"
5599 "CPU%-2d overflow cycles : %lu\n"
5600 "CPU%-2d overflow min : %lu\n"
5601 "CPU%-2d overflow max : %lu\n"
5602 "CPU%-2d smpl handler calls : %lu\n"
5603 "CPU%-2d smpl handler cycles : %lu\n"
5604 "CPU%-2d spurious intrs : %lu\n"
5605 "CPU%-2d replay intrs : %lu\n"
5606 "CPU%-2d syst_wide : %d\n"
5607 "CPU%-2d dcr_pp : %d\n"
5608 "CPU%-2d exclude idle : %d\n"
5609 "CPU%-2d owner : %d\n"
5610 "CPU%-2d context : %p\n"
5611 "CPU%-2d activations : %lu\n",
5612 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5613 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5614 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5615 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5616 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5617 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5618 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5619 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5620 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5621 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5622 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5623 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5624 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5625 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5627 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5629 psr = pfm_get_psr();
5634 "CPU%-2d psr : 0x%lx\n"
5635 "CPU%-2d pmc0 : 0x%lx\n",
5637 cpu, ia64_get_pmc(0));
5639 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5640 if (PMC_IS_COUNTING(i) == 0) continue;
5642 "CPU%-2d pmc%u : 0x%lx\n"
5643 "CPU%-2d pmd%u : 0x%lx\n",
5644 cpu, i, ia64_get_pmc(i),
5645 cpu, i, ia64_get_pmd(i));
5651 const struct seq_operations pfm_seq_ops = {
5652 .start = pfm_proc_start,
5653 .next = pfm_proc_next,
5654 .stop = pfm_proc_stop,
5655 .show = pfm_proc_show
5659 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5660 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5661 * is active or inactive based on mode. We must rely on the value in
5662 * local_cpu_data->pfm_syst_info
5665 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5667 struct pt_regs *regs;
5669 unsigned long dcr_pp;
5671 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5674 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5675 * on every CPU, so we can rely on the pid to identify the idle task.
5677 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5678 regs = task_pt_regs(task);
5679 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5683 * if monitoring has started
5686 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5688 * context switching in?
5691 /* mask monitoring for the idle task */
5692 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5698 * context switching out
5699 * restore monitoring for next task
5701 * Due to inlining this odd if-then-else construction generates
5704 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5713 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5715 struct task_struct *task = ctx->ctx_task;
5717 ia64_psr(regs)->up = 0;
5718 ia64_psr(regs)->sp = 1;
5720 if (GET_PMU_OWNER() == task) {
5721 DPRINT(("cleared ownership for [%d]\n",
5722 task_pid_nr(ctx->ctx_task)));
5723 SET_PMU_OWNER(NULL, NULL);
5727 * disconnect the task from the context and vice-versa
5729 PFM_SET_WORK_PENDING(task, 0);
5731 task->thread.pfm_context = NULL;
5732 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5734 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task)));
5739 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5742 pfm_save_regs(struct task_struct *task)
5745 unsigned long flags;
5749 ctx = PFM_GET_CTX(task);
5750 if (ctx == NULL) return;
5753 * we always come here with interrupts ALREADY disabled by
5754 * the scheduler. So we simply need to protect against concurrent
5755 * access, not CPU concurrency.
5757 flags = pfm_protect_ctx_ctxsw(ctx);
5759 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5760 struct pt_regs *regs = task_pt_regs(task);
5764 pfm_force_cleanup(ctx, regs);
5766 BUG_ON(ctx->ctx_smpl_hdr);
5768 pfm_unprotect_ctx_ctxsw(ctx, flags);
5770 pfm_context_free(ctx);
5775 * save current PSR: needed because we modify it
5778 psr = pfm_get_psr();
5780 BUG_ON(psr & (IA64_PSR_I));
5784 * This is the last instruction which may generate an overflow
5786 * We do not need to set psr.sp because, it is irrelevant in kernel.
5787 * It will be restored from ipsr when going back to user level
5792 * keep a copy of psr.up (for reload)
5794 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5797 * release ownership of this PMU.
5798 * PM interrupts are masked, so nothing
5801 SET_PMU_OWNER(NULL, NULL);
5804 * we systematically save the PMD as we have no
5805 * guarantee we will be schedule at that same
5808 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5811 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5812 * we will need it on the restore path to check
5813 * for pending overflow.
5815 ctx->th_pmcs[0] = ia64_get_pmc(0);
5818 * unfreeze PMU if had pending overflows
5820 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5823 * finally, allow context access.
5824 * interrupts will still be masked after this call.
5826 pfm_unprotect_ctx_ctxsw(ctx, flags);
5829 #else /* !CONFIG_SMP */
5831 pfm_save_regs(struct task_struct *task)
5836 ctx = PFM_GET_CTX(task);
5837 if (ctx == NULL) return;
5840 * save current PSR: needed because we modify it
5842 psr = pfm_get_psr();
5844 BUG_ON(psr & (IA64_PSR_I));
5848 * This is the last instruction which may generate an overflow
5850 * We do not need to set psr.sp because, it is irrelevant in kernel.
5851 * It will be restored from ipsr when going back to user level
5856 * keep a copy of psr.up (for reload)
5858 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5862 pfm_lazy_save_regs (struct task_struct *task)
5865 unsigned long flags;
5867 { u64 psr = pfm_get_psr();
5868 BUG_ON(psr & IA64_PSR_UP);
5871 ctx = PFM_GET_CTX(task);
5874 * we need to mask PMU overflow here to
5875 * make sure that we maintain pmc0 until
5876 * we save it. overflow interrupts are
5877 * treated as spurious if there is no
5880 * XXX: I don't think this is necessary
5882 PROTECT_CTX(ctx,flags);
5885 * release ownership of this PMU.
5886 * must be done before we save the registers.
5888 * after this call any PMU interrupt is treated
5891 SET_PMU_OWNER(NULL, NULL);
5894 * save all the pmds we use
5896 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5899 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5900 * it is needed to check for pended overflow
5901 * on the restore path
5903 ctx->th_pmcs[0] = ia64_get_pmc(0);
5906 * unfreeze PMU if had pending overflows
5908 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5911 * now get can unmask PMU interrupts, they will
5912 * be treated as purely spurious and we will not
5913 * lose any information
5915 UNPROTECT_CTX(ctx,flags);
5917 #endif /* CONFIG_SMP */
5921 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5924 pfm_load_regs (struct task_struct *task)
5927 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
5928 unsigned long flags;
5930 int need_irq_resend;
5932 ctx = PFM_GET_CTX(task);
5933 if (unlikely(ctx == NULL)) return;
5935 BUG_ON(GET_PMU_OWNER());
5938 * possible on unload
5940 if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
5943 * we always come here with interrupts ALREADY disabled by
5944 * the scheduler. So we simply need to protect against concurrent
5945 * access, not CPU concurrency.
5947 flags = pfm_protect_ctx_ctxsw(ctx);
5948 psr = pfm_get_psr();
5950 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
5952 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
5953 BUG_ON(psr & IA64_PSR_I);
5955 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
5956 struct pt_regs *regs = task_pt_regs(task);
5958 BUG_ON(ctx->ctx_smpl_hdr);
5960 pfm_force_cleanup(ctx, regs);
5962 pfm_unprotect_ctx_ctxsw(ctx, flags);
5965 * this one (kmalloc'ed) is fine with interrupts disabled
5967 pfm_context_free(ctx);
5973 * we restore ALL the debug registers to avoid picking up
5976 if (ctx->ctx_fl_using_dbreg) {
5977 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
5978 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
5981 * retrieve saved psr.up
5983 psr_up = ctx->ctx_saved_psr_up;
5986 * if we were the last user of the PMU on that CPU,
5987 * then nothing to do except restore psr
5989 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
5992 * retrieve partial reload masks (due to user modifications)
5994 pmc_mask = ctx->ctx_reload_pmcs[0];
5995 pmd_mask = ctx->ctx_reload_pmds[0];
5999 * To avoid leaking information to the user level when psr.sp=0,
6000 * we must reload ALL implemented pmds (even the ones we don't use).
6001 * In the kernel we only allow PFM_READ_PMDS on registers which
6002 * we initialized or requested (sampling) so there is no risk there.
6004 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6007 * ALL accessible PMCs are systematically reloaded, unused registers
6008 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6009 * up stale configuration.
6011 * PMC0 is never in the mask. It is always restored separately.
6013 pmc_mask = ctx->ctx_all_pmcs[0];
6016 * when context is MASKED, we will restore PMC with plm=0
6017 * and PMD with stale information, but that's ok, nothing
6020 * XXX: optimize here
6022 if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6023 if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6026 * check for pending overflow at the time the state
6029 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6031 * reload pmc0 with the overflow information
6032 * On McKinley PMU, this will trigger a PMU interrupt
6034 ia64_set_pmc(0, ctx->th_pmcs[0]);
6036 ctx->th_pmcs[0] = 0UL;
6039 * will replay the PMU interrupt
6041 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6043 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6047 * we just did a reload, so we reset the partial reload fields
6049 ctx->ctx_reload_pmcs[0] = 0UL;
6050 ctx->ctx_reload_pmds[0] = 0UL;
6052 SET_LAST_CPU(ctx, smp_processor_id());
6055 * dump activation value for this PMU
6059 * record current activation for this context
6061 SET_ACTIVATION(ctx);
6064 * establish new ownership.
6066 SET_PMU_OWNER(task, ctx);
6069 * restore the psr.up bit. measurement
6071 * no PMU interrupt can happen at this point
6072 * because we still have interrupts disabled.
6074 if (likely(psr_up)) pfm_set_psr_up();
6077 * allow concurrent access to context
6079 pfm_unprotect_ctx_ctxsw(ctx, flags);
6081 #else /* !CONFIG_SMP */
6083 * reload PMU state for UP kernels
6084 * in 2.5 we come here with interrupts disabled
6087 pfm_load_regs (struct task_struct *task)
6090 struct task_struct *owner;
6091 unsigned long pmd_mask, pmc_mask;
6093 int need_irq_resend;
6095 owner = GET_PMU_OWNER();
6096 ctx = PFM_GET_CTX(task);
6097 psr = pfm_get_psr();
6099 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6100 BUG_ON(psr & IA64_PSR_I);
6103 * we restore ALL the debug registers to avoid picking up
6106 * This must be done even when the task is still the owner
6107 * as the registers may have been modified via ptrace()
6108 * (not perfmon) by the previous task.
6110 if (ctx->ctx_fl_using_dbreg) {
6111 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6112 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6116 * retrieved saved psr.up
6118 psr_up = ctx->ctx_saved_psr_up;
6119 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6122 * short path, our state is still there, just
6123 * need to restore psr and we go
6125 * we do not touch either PMC nor PMD. the psr is not touched
6126 * by the overflow_handler. So we are safe w.r.t. to interrupt
6127 * concurrency even without interrupt masking.
6129 if (likely(owner == task)) {
6130 if (likely(psr_up)) pfm_set_psr_up();
6135 * someone else is still using the PMU, first push it out and
6136 * then we'll be able to install our stuff !
6138 * Upon return, there will be no owner for the current PMU
6140 if (owner) pfm_lazy_save_regs(owner);
6143 * To avoid leaking information to the user level when psr.sp=0,
6144 * we must reload ALL implemented pmds (even the ones we don't use).
6145 * In the kernel we only allow PFM_READ_PMDS on registers which
6146 * we initialized or requested (sampling) so there is no risk there.
6148 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6151 * ALL accessible PMCs are systematically reloaded, unused registers
6152 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6153 * up stale configuration.
6155 * PMC0 is never in the mask. It is always restored separately
6157 pmc_mask = ctx->ctx_all_pmcs[0];
6159 pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6160 pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6163 * check for pending overflow at the time the state
6166 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6168 * reload pmc0 with the overflow information
6169 * On McKinley PMU, this will trigger a PMU interrupt
6171 ia64_set_pmc(0, ctx->th_pmcs[0]);
6174 ctx->th_pmcs[0] = 0UL;
6177 * will replay the PMU interrupt
6179 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6181 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6185 * establish new ownership.
6187 SET_PMU_OWNER(task, ctx);
6190 * restore the psr.up bit. measurement
6192 * no PMU interrupt can happen at this point
6193 * because we still have interrupts disabled.
6195 if (likely(psr_up)) pfm_set_psr_up();
6197 #endif /* CONFIG_SMP */
6200 * this function assumes monitoring is stopped
6203 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6206 unsigned long mask2, val, pmd_val, ovfl_val;
6207 int i, can_access_pmu = 0;
6211 * is the caller the task being monitored (or which initiated the
6212 * session for system wide measurements)
6214 is_self = ctx->ctx_task == task ? 1 : 0;
6217 * can access PMU is task is the owner of the PMU state on the current CPU
6218 * or if we are running on the CPU bound to the context in system-wide mode
6219 * (that is not necessarily the task the context is attached to in this mode).
6220 * In system-wide we always have can_access_pmu true because a task running on an
6221 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6223 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6224 if (can_access_pmu) {
6226 * Mark the PMU as not owned
6227 * This will cause the interrupt handler to do nothing in case an overflow
6228 * interrupt was in-flight
6229 * This also guarantees that pmc0 will contain the final state
6230 * It virtually gives us full control on overflow processing from that point
6233 SET_PMU_OWNER(NULL, NULL);
6234 DPRINT(("releasing ownership\n"));
6237 * read current overflow status:
6239 * we are guaranteed to read the final stable state
6242 pmc0 = ia64_get_pmc(0); /* slow */
6245 * reset freeze bit, overflow status information destroyed
6249 pmc0 = ctx->th_pmcs[0];
6251 * clear whatever overflow status bits there were
6253 ctx->th_pmcs[0] = 0;
6255 ovfl_val = pmu_conf->ovfl_val;
6257 * we save all the used pmds
6258 * we take care of overflows for counting PMDs
6260 * XXX: sampling situation is not taken into account here
6262 mask2 = ctx->ctx_used_pmds[0];
6264 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6266 for (i = 0; mask2; i++, mask2>>=1) {
6268 /* skip non used pmds */
6269 if ((mask2 & 0x1) == 0) continue;
6272 * can access PMU always true in system wide mode
6274 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6276 if (PMD_IS_COUNTING(i)) {
6277 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6280 ctx->ctx_pmds[i].val,
6284 * we rebuild the full 64 bit value of the counter
6286 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6289 * now everything is in ctx_pmds[] and we need
6290 * to clear the saved context from save_regs() such that
6291 * pfm_read_pmds() gets the correct value
6296 * take care of overflow inline
6298 if (pmc0 & (1UL << i)) {
6299 val += 1 + ovfl_val;
6300 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6304 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6306 if (is_self) ctx->th_pmds[i] = pmd_val;
6308 ctx->ctx_pmds[i].val = val;
6312 static struct irqaction perfmon_irqaction = {
6313 .handler = pfm_interrupt_handler,
6318 pfm_alt_save_pmu_state(void *data)
6320 struct pt_regs *regs;
6322 regs = task_pt_regs(current);
6324 DPRINT(("called\n"));
6327 * should not be necessary but
6328 * let's take not risk
6332 ia64_psr(regs)->pp = 0;
6335 * This call is required
6336 * May cause a spurious interrupt on some processors
6344 pfm_alt_restore_pmu_state(void *data)
6346 struct pt_regs *regs;
6348 regs = task_pt_regs(current);
6350 DPRINT(("called\n"));
6353 * put PMU back in state expected
6358 ia64_psr(regs)->pp = 0;
6361 * perfmon runs with PMU unfrozen at all times
6369 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6374 /* some sanity checks */
6375 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6377 /* do the easy test first */
6378 if (pfm_alt_intr_handler) return -EBUSY;
6380 /* one at a time in the install or remove, just fail the others */
6381 if (!spin_trylock(&pfm_alt_install_check)) {
6385 /* reserve our session */
6386 for_each_online_cpu(reserve_cpu) {
6387 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6388 if (ret) goto cleanup_reserve;
6391 /* save the current system wide pmu states */
6392 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6394 DPRINT(("on_each_cpu() failed: %d\n", ret));
6395 goto cleanup_reserve;
6398 /* officially change to the alternate interrupt handler */
6399 pfm_alt_intr_handler = hdl;
6401 spin_unlock(&pfm_alt_install_check);
6406 for_each_online_cpu(i) {
6407 /* don't unreserve more than we reserved */
6408 if (i >= reserve_cpu) break;
6410 pfm_unreserve_session(NULL, 1, i);
6413 spin_unlock(&pfm_alt_install_check);
6417 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6420 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6425 if (hdl == NULL) return -EINVAL;
6427 /* cannot remove someone else's handler! */
6428 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6430 /* one at a time in the install or remove, just fail the others */
6431 if (!spin_trylock(&pfm_alt_install_check)) {
6435 pfm_alt_intr_handler = NULL;
6437 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6439 DPRINT(("on_each_cpu() failed: %d\n", ret));
6442 for_each_online_cpu(i) {
6443 pfm_unreserve_session(NULL, 1, i);
6446 spin_unlock(&pfm_alt_install_check);
6450 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6453 * perfmon initialization routine, called from the initcall() table
6455 static int init_pfm_fs(void);
6463 family = local_cpu_data->family;
6468 if ((*p)->probe() == 0) goto found;
6469 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6483 unsigned int n, n_counters, i;
6485 printk("perfmon: version %u.%u IRQ %u\n",
6488 IA64_PERFMON_VECTOR);
6490 if (pfm_probe_pmu()) {
6491 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6492 local_cpu_data->family);
6497 * compute the number of implemented PMD/PMC from the
6498 * description tables
6501 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6502 if (PMC_IS_IMPL(i) == 0) continue;
6503 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6506 pmu_conf->num_pmcs = n;
6508 n = 0; n_counters = 0;
6509 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6510 if (PMD_IS_IMPL(i) == 0) continue;
6511 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6513 if (PMD_IS_COUNTING(i)) n_counters++;
6515 pmu_conf->num_pmds = n;
6516 pmu_conf->num_counters = n_counters;
6519 * sanity checks on the number of debug registers
6521 if (pmu_conf->use_rr_dbregs) {
6522 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6523 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6527 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6528 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6534 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6538 pmu_conf->num_counters,
6539 ffz(pmu_conf->ovfl_val));
6542 if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6543 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6549 * create /proc/perfmon (mostly for debugging purposes)
6551 perfmon_dir = proc_create_seq("perfmon", S_IRUGO, NULL, &pfm_seq_ops);
6552 if (perfmon_dir == NULL) {
6553 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6559 * create /proc/sys/kernel/perfmon (for debugging purposes)
6561 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6564 * initialize all our spinlocks
6566 spin_lock_init(&pfm_sessions.pfs_lock);
6567 spin_lock_init(&pfm_buffer_fmt_lock);
6571 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6576 __initcall(pfm_init);
6579 * this function is called before pfm_init()
6582 pfm_init_percpu (void)
6584 static int first_time=1;
6586 * make sure no measurement is active
6587 * (may inherit programmed PMCs from EFI).
6593 * we run with the PMU not frozen at all times
6598 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6602 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6607 * used for debug purposes only
6610 dump_pmu_state(const char *from)
6612 struct task_struct *task;
6613 struct pt_regs *regs;
6615 unsigned long psr, dcr, info, flags;
6618 local_irq_save(flags);
6620 this_cpu = smp_processor_id();
6621 regs = task_pt_regs(current);
6622 info = PFM_CPUINFO_GET();
6623 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6625 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6626 local_irq_restore(flags);
6630 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6633 task_pid_nr(current),
6637 task = GET_PMU_OWNER();
6638 ctx = GET_PMU_CTX();
6640 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6642 psr = pfm_get_psr();
6644 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",
6647 psr & IA64_PSR_PP ? 1 : 0,
6648 psr & IA64_PSR_UP ? 1 : 0,
6649 dcr & IA64_DCR_PP ? 1 : 0,
6652 ia64_psr(regs)->pp);
6654 ia64_psr(regs)->up = 0;
6655 ia64_psr(regs)->pp = 0;
6657 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6658 if (PMC_IS_IMPL(i) == 0) continue;
6659 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, ctx->th_pmcs[i]);
6662 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6663 if (PMD_IS_IMPL(i) == 0) continue;
6664 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, ctx->th_pmds[i]);
6668 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6671 ctx->ctx_smpl_vaddr,
6675 ctx->ctx_saved_psr_up);
6677 local_irq_restore(flags);
6681 * called from process.c:copy_thread(). task is new child.
6684 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6686 struct thread_struct *thread;
6688 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6690 thread = &task->thread;
6693 * cut links inherited from parent (current)
6695 thread->pfm_context = NULL;
6697 PFM_SET_WORK_PENDING(task, 0);
6700 * the psr bits are already set properly in copy_threads()
6703 #else /* !CONFIG_PERFMON */
6705 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6709 #endif /* CONFIG_PERFMON */