projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux-2.6
[deliverable/linux.git] / arch / ia64 / kernel / perfmon.c
1 /*
2 * This file implements the perfmon-2 subsystem which is used
3 * to program the IA-64 Performance Monitoring Unit (PMU).
4 *
5 * The initial version of perfmon.c was written by
6 * Ganesh Venkitachalam, IBM Corp.
7 *
8 * Then it was modified for perfmon-1.x by Stephane Eranian and
9 * David Mosberger, Hewlett Packard Co.
10 *
11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
12 * by Stephane Eranian, Hewlett Packard Co.
13 *
14 * Copyright (C) 1999-2005 Hewlett Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 * David Mosberger-Tang <davidm@hpl.hp.com>
17 *
18 * More information about perfmon available at:
19 * http://www.hpl.hp.com/research/linux/perfmon
20 */
21
22 #include <linux/module.h>
23 #include <linux/kernel.h>
24 #include <linux/sched.h>
25 #include <linux/interrupt.h>
26 #include <linux/proc_fs.h>
27 #include <linux/seq_file.h>
28 #include <linux/init.h>
29 #include <linux/vmalloc.h>
30 #include <linux/mm.h>
31 #include <linux/sysctl.h>
32 #include <linux/list.h>
33 #include <linux/file.h>
34 #include <linux/poll.h>
35 #include <linux/vfs.h>
36 #include <linux/smp.h>
37 #include <linux/pagemap.h>
38 #include <linux/mount.h>
39 #include <linux/bitops.h>
40 #include <linux/capability.h>
41 #include <linux/rcupdate.h>
42 #include <linux/completion.h>
43
44 #include <asm/errno.h>
45 #include <asm/intrinsics.h>
46 #include <asm/page.h>
47 #include <asm/perfmon.h>
48 #include <asm/processor.h>
49 #include <asm/signal.h>
50 #include <asm/system.h>
51 #include <asm/uaccess.h>
52 #include <asm/delay.h>
53
54 #ifdef CONFIG_PERFMON
55 /*
56 * perfmon context state
57 */
58 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
59 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
60 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
61 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
62
63 #define PFM_INVALID_ACTIVATION (~0UL)
64
65 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
66 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
67
68 /*
69 * depth of message queue
70 */
71 #define PFM_MAX_MSGS 32
72 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
73
74 /*
75 * type of a PMU register (bitmask).
76 * bitmask structure:
77 * bit0 : register implemented
78 * bit1 : end marker
79 * bit2-3 : reserved
80 * bit4 : pmc has pmc.pm
81 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
82 * bit6-7 : register type
83 * bit8-31: reserved
84 */
85 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
86 #define PFM_REG_IMPL 0x1 /* register implemented */
87 #define PFM_REG_END 0x2 /* end marker */
88 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
89 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
90 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
91 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
92 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
93
94 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
95 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
96
97 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
98
99 /* i assumed unsigned */
100 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
101 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
102
103 /* XXX: these assume that register i is implemented */
104 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
105 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
106 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
107 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
108
109 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
110 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
111 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
112 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
113
114 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
115 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
116
117 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
118 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
119 #define PFM_CTX_TASK(h) (h)->ctx_task
120
121 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
122
123 /* XXX: does not support more than 64 PMDs */
124 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
125 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
126
127 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
128
129 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
130 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
131 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
132 #define PFM_CODE_RR 0 /* requesting code range restriction */
133 #define PFM_DATA_RR 1 /* requestion data range restriction */
134
135 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
136 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
137 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
138
139 #define RDEP(x) (1UL<<(x))
140
141 /*
142 * context protection macros
143 * in SMP:
144 * - we need to protect against CPU concurrency (spin_lock)
145 * - we need to protect against PMU overflow interrupts (local_irq_disable)
146 * in UP:
147 * - we need to protect against PMU overflow interrupts (local_irq_disable)
148 *
149 * spin_lock_irqsave()/spin_unlock_irqrestore():
150 * in SMP: local_irq_disable + spin_lock
151 * in UP : local_irq_disable
152 *
153 * spin_lock()/spin_lock():
154 * in UP : removed automatically
155 * in SMP: protect against context accesses from other CPU. interrupts
156 * are not masked. This is useful for the PMU interrupt handler
157 * because we know we will not get PMU concurrency in that code.
158 */
159 #define PROTECT_CTX(c, f) \
160 do { \
161 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \
162 spin_lock_irqsave(&(c)->ctx_lock, f); \
163 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
164 } while(0)
165
166 #define UNPROTECT_CTX(c, f) \
167 do { \
168 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
169 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
170 } while(0)
171
172 #define PROTECT_CTX_NOPRINT(c, f) \
173 do { \
174 spin_lock_irqsave(&(c)->ctx_lock, f); \
175 } while(0)
176
177
178 #define UNPROTECT_CTX_NOPRINT(c, f) \
179 do { \
180 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
181 } while(0)
182
183
184 #define PROTECT_CTX_NOIRQ(c) \
185 do { \
186 spin_lock(&(c)->ctx_lock); \
187 } while(0)
188
189 #define UNPROTECT_CTX_NOIRQ(c) \
190 do { \
191 spin_unlock(&(c)->ctx_lock); \
192 } while(0)
193
194
195 #ifdef CONFIG_SMP
196
197 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
198 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
199 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
200
201 #else /* !CONFIG_SMP */
202 #define SET_ACTIVATION(t) do {} while(0)
203 #define GET_ACTIVATION(t) do {} while(0)
204 #define INC_ACTIVATION(t) do {} while(0)
205 #endif /* CONFIG_SMP */
206
207 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
208 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
209 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
210
211 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
212 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
213
214 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
215
216 /*
217 * cmp0 must be the value of pmc0
218 */
219 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
220
221 #define PFMFS_MAGIC 0xa0b4d889
222
223 /*
224 * debugging
225 */
226 #define PFM_DEBUGGING 1
227 #ifdef PFM_DEBUGGING
228 #define DPRINT(a) \
229 do { \
230 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
231 } while (0)
232
233 #define DPRINT_ovfl(a) \
234 do { \
235 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
236 } while (0)
237 #endif
238
239 /*
240 * 64-bit software counter structure
241 *
242 * the next_reset_type is applied to the next call to pfm_reset_regs()
243 */
244 typedef struct {
245 unsigned long val; /* virtual 64bit counter value */
246 unsigned long lval; /* last reset value */
247 unsigned long long_reset; /* reset value on sampling overflow */
248 unsigned long short_reset; /* reset value on overflow */
249 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
250 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
251 unsigned long seed; /* seed for random-number generator */
252 unsigned long mask; /* mask for random-number generator */
253 unsigned int flags; /* notify/do not notify */
254 unsigned long eventid; /* overflow event identifier */
255 } pfm_counter_t;
256
257 /*
258 * context flags
259 */
260 typedef struct {
261 unsigned int block:1; /* when 1, task will blocked on user notifications */
262 unsigned int system:1; /* do system wide monitoring */
263 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
264 unsigned int is_sampling:1; /* true if using a custom format */
265 unsigned int excl_idle:1; /* exclude idle task in system wide session */
266 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
267 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
268 unsigned int no_msg:1; /* no message sent on overflow */
269 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
270 unsigned int reserved:22;
271 } pfm_context_flags_t;
272
273 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
274 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
275 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
276
277
278 /*
279 * perfmon context: encapsulates all the state of a monitoring session
280 */
281
282 typedef struct pfm_context {
283 spinlock_t ctx_lock; /* context protection */
284
285 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
286 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
287
288 struct task_struct *ctx_task; /* task to which context is attached */
289
290 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
291
292 struct completion ctx_restart_done; /* use for blocking notification mode */
293
294 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
295 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
296 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
297
298 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
299 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
300 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
301
302 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
303
304 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
305 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
306 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
307 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
308
309 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
310
311 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
312 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
313
314 u64 ctx_saved_psr_up; /* only contains psr.up value */
315
316 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
317 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
318 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
319
320 int ctx_fd; /* file descriptor used my this context */
321 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
322
323 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
324 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
325 unsigned long ctx_smpl_size; /* size of sampling buffer */
326 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
327
328 wait_queue_head_t ctx_msgq_wait;
329 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
330 int ctx_msgq_head;
331 int ctx_msgq_tail;
332 struct fasync_struct *ctx_async_queue;
333
334 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
335 } pfm_context_t;
336
337 /*
338 * magic number used to verify that structure is really
339 * a perfmon context
340 */
341 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
342
343 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
344
345 #ifdef CONFIG_SMP
346 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
347 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
348 #else
349 #define SET_LAST_CPU(ctx, v) do {} while(0)
350 #define GET_LAST_CPU(ctx) do {} while(0)
351 #endif
352
353
354 #define ctx_fl_block ctx_flags.block
355 #define ctx_fl_system ctx_flags.system
356 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
357 #define ctx_fl_is_sampling ctx_flags.is_sampling
358 #define ctx_fl_excl_idle ctx_flags.excl_idle
359 #define ctx_fl_going_zombie ctx_flags.going_zombie
360 #define ctx_fl_trap_reason ctx_flags.trap_reason
361 #define ctx_fl_no_msg ctx_flags.no_msg
362 #define ctx_fl_can_restart ctx_flags.can_restart
363
364 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
365 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
366
367 /*
368 * global information about all sessions
369 * mostly used to synchronize between system wide and per-process
370 */
371 typedef struct {
372 spinlock_t pfs_lock; /* lock the structure */
373
374 unsigned int pfs_task_sessions; /* number of per task sessions */
375 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
376 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
377 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
378 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
379 } pfm_session_t;
380
381 /*
382 * information about a PMC or PMD.
383 * dep_pmd[]: a bitmask of dependent PMD registers
384 * dep_pmc[]: a bitmask of dependent PMC registers
385 */
386 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
387 typedef struct {
388 unsigned int type;
389 int pm_pos;
390 unsigned long default_value; /* power-on default value */
391 unsigned long reserved_mask; /* bitmask of reserved bits */
392 pfm_reg_check_t read_check;
393 pfm_reg_check_t write_check;
394 unsigned long dep_pmd[4];
395 unsigned long dep_pmc[4];
396 } pfm_reg_desc_t;
397
398 /* assume cnum is a valid monitor */
399 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
400
401 /*
402 * This structure is initialized at boot time and contains
403 * a description of the PMU main characteristics.
404 *
405 * If the probe function is defined, detection is based
406 * on its return value:
407 * - 0 means recognized PMU
408 * - anything else means not supported
409 * When the probe function is not defined, then the pmu_family field
410 * is used and it must match the host CPU family such that:
411 * - cpu->family & config->pmu_family != 0
412 */
413 typedef struct {
414 unsigned long ovfl_val; /* overflow value for counters */
415
416 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
417 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
418
419 unsigned int num_pmcs; /* number of PMCS: computed at init time */
420 unsigned int num_pmds; /* number of PMDS: computed at init time */
421 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
422 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
423
424 char *pmu_name; /* PMU family name */
425 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
426 unsigned int flags; /* pmu specific flags */
427 unsigned int num_ibrs; /* number of IBRS: computed at init time */
428 unsigned int num_dbrs; /* number of DBRS: computed at init time */
429 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
430 int (*probe)(void); /* customized probe routine */
431 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
432 } pmu_config_t;
433 /*
434 * PMU specific flags
435 */
436 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
437
438 /*
439 * debug register related type definitions
440 */
441 typedef struct {
442 unsigned long ibr_mask:56;
443 unsigned long ibr_plm:4;
444 unsigned long ibr_ig:3;
445 unsigned long ibr_x:1;
446 } ibr_mask_reg_t;
447
448 typedef struct {
449 unsigned long dbr_mask:56;
450 unsigned long dbr_plm:4;
451 unsigned long dbr_ig:2;
452 unsigned long dbr_w:1;
453 unsigned long dbr_r:1;
454 } dbr_mask_reg_t;
455
456 typedef union {
457 unsigned long val;
458 ibr_mask_reg_t ibr;
459 dbr_mask_reg_t dbr;
460 } dbreg_t;
461
462
463 /*
464 * perfmon command descriptions
465 */
466 typedef struct {
467 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
468 char *cmd_name;
469 int cmd_flags;
470 unsigned int cmd_narg;
471 size_t cmd_argsize;
472 int (*cmd_getsize)(void *arg, size_t *sz);
473 } pfm_cmd_desc_t;
474
475 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
476 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
477 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
478 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
479
480
481 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
482 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
483 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
484 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
485 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
486
487 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
488
489 typedef struct {
490 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
491 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
492 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
493 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
496 unsigned long pfm_smpl_handler_calls;
497 unsigned long pfm_smpl_handler_cycles;
498 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
499 } pfm_stats_t;
500
501 /*
502 * perfmon internal variables
503 */
504 static pfm_stats_t pfm_stats[NR_CPUS];
505 static pfm_session_t pfm_sessions; /* global sessions information */
506
507 static DEFINE_SPINLOCK(pfm_alt_install_check);
508 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
509
510 static struct proc_dir_entry *perfmon_dir;
511 static pfm_uuid_t pfm_null_uuid = {0,};
512
513 static spinlock_t pfm_buffer_fmt_lock;
514 static LIST_HEAD(pfm_buffer_fmt_list);
515
516 static pmu_config_t *pmu_conf;
517
518 /* sysctl() controls */
519 pfm_sysctl_t pfm_sysctl;
520 EXPORT_SYMBOL(pfm_sysctl);
521
522 static ctl_table pfm_ctl_table[]={
523 {
524 .ctl_name = CTL_UNNUMBERED,
525 .procname = "debug",
526 .data = &pfm_sysctl.debug,
527 .maxlen = sizeof(int),
528 .mode = 0666,
529 .proc_handler = &proc_dointvec,
530 },
531 {
532 .ctl_name = CTL_UNNUMBERED,
533 .procname = "debug_ovfl",
534 .data = &pfm_sysctl.debug_ovfl,
535 .maxlen = sizeof(int),
536 .mode = 0666,
537 .proc_handler = &proc_dointvec,
538 },
539 {
540 .ctl_name = CTL_UNNUMBERED,
541 .procname = "fastctxsw",
542 .data = &pfm_sysctl.fastctxsw,
543 .maxlen = sizeof(int),
544 .mode = 0600,
545 .proc_handler = &proc_dointvec,
546 },
547 {
548 .ctl_name = CTL_UNNUMBERED,
549 .procname = "expert_mode",
550 .data = &pfm_sysctl.expert_mode,
551 .maxlen = sizeof(int),
552 .mode = 0600,
553 .proc_handler = &proc_dointvec,
554 },
555 {}
556 };
557 static ctl_table pfm_sysctl_dir[] = {
558 {
559 .ctl_name = CTL_UNNUMBERED,
560 .procname = "perfmon",
561 .mode = 0755,
562 .child = pfm_ctl_table,
563 },
564 {}
565 };
566 static ctl_table pfm_sysctl_root[] = {
567 {
568 .ctl_name = CTL_KERN,
569 .procname = "kernel",
570 .mode = 0755,
571 .child = pfm_sysctl_dir,
572 },
573 {}
574 };
575 static struct ctl_table_header *pfm_sysctl_header;
576
577 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
578
579 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
580 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
581
582 static inline void
583 pfm_put_task(struct task_struct *task)
584 {
585 if (task != current) put_task_struct(task);
586 }
587
588 static inline void
589 pfm_set_task_notify(struct task_struct *task)
590 {
591 struct thread_info *info;
592
593 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
594 set_bit(TIF_PERFMON_WORK, &info->flags);
595 }
596
597 static inline void
598 pfm_clear_task_notify(void)
599 {
600 clear_thread_flag(TIF_PERFMON_WORK);
601 }
602
603 static inline void
604 pfm_reserve_page(unsigned long a)
605 {
606 SetPageReserved(vmalloc_to_page((void *)a));
607 }
608 static inline void
609 pfm_unreserve_page(unsigned long a)
610 {
611 ClearPageReserved(vmalloc_to_page((void*)a));
612 }
613
614 static inline unsigned long
615 pfm_protect_ctx_ctxsw(pfm_context_t *x)
616 {
617 spin_lock(&(x)->ctx_lock);
618 return 0UL;
619 }
620
621 static inline void
622 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
623 {
624 spin_unlock(&(x)->ctx_lock);
625 }
626
627 static inline unsigned int
628 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
629 {
630 return do_munmap(mm, addr, len);
631 }
632
633 static inline unsigned long
634 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
635 {
636 return get_unmapped_area(file, addr, len, pgoff, flags);
637 }
638
639
640 static int
641 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data,
642 struct vfsmount *mnt)
643 {
644 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
645 }
646
647 static struct file_system_type pfm_fs_type = {
648 .name = "pfmfs",
649 .get_sb = pfmfs_get_sb,
650 .kill_sb = kill_anon_super,
651 };
652
653 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
654 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
655 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
656 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
657 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
658
659
660 /* forward declaration */
661 static const struct file_operations pfm_file_ops;
662
663 /*
664 * forward declarations
665 */
666 #ifndef CONFIG_SMP
667 static void pfm_lazy_save_regs (struct task_struct *ta);
668 #endif
669
670 void dump_pmu_state(const char *);
671 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
672
673 #include "perfmon_itanium.h"
674 #include "perfmon_mckinley.h"
675 #include "perfmon_montecito.h"
676 #include "perfmon_generic.h"
677
678 static pmu_config_t *pmu_confs[]={
679 &pmu_conf_mont,
680 &pmu_conf_mck,
681 &pmu_conf_ita,
682 &pmu_conf_gen, /* must be last */
683 NULL
684 };
685
686
687 static int pfm_end_notify_user(pfm_context_t *ctx);
688
689 static inline void
690 pfm_clear_psr_pp(void)
691 {
692 ia64_rsm(IA64_PSR_PP);
693 ia64_srlz_i();
694 }
695
696 static inline void
697 pfm_set_psr_pp(void)
698 {
699 ia64_ssm(IA64_PSR_PP);
700 ia64_srlz_i();
701 }
702
703 static inline void
704 pfm_clear_psr_up(void)
705 {
706 ia64_rsm(IA64_PSR_UP);
707 ia64_srlz_i();
708 }
709
710 static inline void
711 pfm_set_psr_up(void)
712 {
713 ia64_ssm(IA64_PSR_UP);
714 ia64_srlz_i();
715 }
716
717 static inline unsigned long
718 pfm_get_psr(void)
719 {
720 unsigned long tmp;
721 tmp = ia64_getreg(_IA64_REG_PSR);
722 ia64_srlz_i();
723 return tmp;
724 }
725
726 static inline void
727 pfm_set_psr_l(unsigned long val)
728 {
729 ia64_setreg(_IA64_REG_PSR_L, val);
730 ia64_srlz_i();
731 }
732
733 static inline void
734 pfm_freeze_pmu(void)
735 {
736 ia64_set_pmc(0,1UL);
737 ia64_srlz_d();
738 }
739
740 static inline void
741 pfm_unfreeze_pmu(void)
742 {
743 ia64_set_pmc(0,0UL);
744 ia64_srlz_d();
745 }
746
747 static inline void
748 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
749 {
750 int i;
751
752 for (i=0; i < nibrs; i++) {
753 ia64_set_ibr(i, ibrs[i]);
754 ia64_dv_serialize_instruction();
755 }
756 ia64_srlz_i();
757 }
758
759 static inline void
760 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
761 {
762 int i;
763
764 for (i=0; i < ndbrs; i++) {
765 ia64_set_dbr(i, dbrs[i]);
766 ia64_dv_serialize_data();
767 }
768 ia64_srlz_d();
769 }
770
771 /*
772 * PMD[i] must be a counter. no check is made
773 */
774 static inline unsigned long
775 pfm_read_soft_counter(pfm_context_t *ctx, int i)
776 {
777 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
778 }
779
780 /*
781 * PMD[i] must be a counter. no check is made
782 */
783 static inline void
784 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
785 {
786 unsigned long ovfl_val = pmu_conf->ovfl_val;
787
788 ctx->ctx_pmds[i].val = val & ~ovfl_val;
789 /*
790 * writing to unimplemented part is ignore, so we do not need to
791 * mask off top part
792 */
793 ia64_set_pmd(i, val & ovfl_val);
794 }
795
796 static pfm_msg_t *
797 pfm_get_new_msg(pfm_context_t *ctx)
798 {
799 int idx, next;
800
801 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
802
803 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
804 if (next == ctx->ctx_msgq_head) return NULL;
805
806 idx = ctx->ctx_msgq_tail;
807 ctx->ctx_msgq_tail = next;
808
809 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
810
811 return ctx->ctx_msgq+idx;
812 }
813
814 static pfm_msg_t *
815 pfm_get_next_msg(pfm_context_t *ctx)
816 {
817 pfm_msg_t *msg;
818
819 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
820
821 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
822
823 /*
824 * get oldest message
825 */
826 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
827
828 /*
829 * and move forward
830 */
831 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
832
833 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));
834
835 return msg;
836 }
837
838 static void
839 pfm_reset_msgq(pfm_context_t *ctx)
840 {
841 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
842 DPRINT(("ctx=%p msgq reset\n", ctx));
843 }
844
845 static void *
846 pfm_rvmalloc(unsigned long size)
847 {
848 void *mem;
849 unsigned long addr;
850
851 size = PAGE_ALIGN(size);
852 mem = vmalloc(size);
853 if (mem) {
854 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
855 memset(mem, 0, size);
856 addr = (unsigned long)mem;
857 while (size > 0) {
858 pfm_reserve_page(addr);
859 addr+=PAGE_SIZE;
860 size-=PAGE_SIZE;
861 }
862 }
863 return mem;
864 }
865
866 static void
867 pfm_rvfree(void *mem, unsigned long size)
868 {
869 unsigned long addr;
870
871 if (mem) {
872 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
873 addr = (unsigned long) mem;
874 while ((long) size > 0) {
875 pfm_unreserve_page(addr);
876 addr+=PAGE_SIZE;
877 size-=PAGE_SIZE;
878 }
879 vfree(mem);
880 }
881 return;
882 }
883
884 static pfm_context_t *
885 pfm_context_alloc(void)
886 {
887 pfm_context_t *ctx;
888
889 /*
890 * allocate context descriptor
891 * must be able to free with interrupts disabled
892 */
893 ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
894 if (ctx) {
895 DPRINT(("alloc ctx @%p\n", ctx));
896 }
897 return ctx;
898 }
899
900 static void
901 pfm_context_free(pfm_context_t *ctx)
902 {
903 if (ctx) {
904 DPRINT(("free ctx @%p\n", ctx));
905 kfree(ctx);
906 }
907 }
908
909 static void
910 pfm_mask_monitoring(struct task_struct *task)
911 {
912 pfm_context_t *ctx = PFM_GET_CTX(task);
913 unsigned long mask, val, ovfl_mask;
914 int i;
915
916 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
917
918 ovfl_mask = pmu_conf->ovfl_val;
919 /*
920 * monitoring can only be masked as a result of a valid
921 * counter overflow. In UP, it means that the PMU still
922 * has an owner. Note that the owner can be different
923 * from the current task. However the PMU state belongs
924 * to the owner.
925 * In SMP, a valid overflow only happens when task is
926 * current. Therefore if we come here, we know that
927 * the PMU state belongs to the current task, therefore
928 * we can access the live registers.
929 *
930 * So in both cases, the live register contains the owner's
931 * state. We can ONLY touch the PMU registers and NOT the PSR.
932 *
933 * As a consequence to this call, the ctx->th_pmds[] array
934 * contains stale information which must be ignored
935 * when context is reloaded AND monitoring is active (see
936 * pfm_restart).
937 */
938 mask = ctx->ctx_used_pmds[0];
939 for (i = 0; mask; i++, mask>>=1) {
940 /* skip non used pmds */
941 if ((mask & 0x1) == 0) continue;
942 val = ia64_get_pmd(i);
943
944 if (PMD_IS_COUNTING(i)) {
945 /*
946 * we rebuild the full 64 bit value of the counter
947 */
948 ctx->ctx_pmds[i].val += (val & ovfl_mask);
949 } else {
950 ctx->ctx_pmds[i].val = val;
951 }
952 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
953 i,
954 ctx->ctx_pmds[i].val,
955 val & ovfl_mask));
956 }
957 /*
958 * mask monitoring by setting the privilege level to 0
959 * we cannot use psr.pp/psr.up for this, it is controlled by
960 * the user
961 *
962 * if task is current, modify actual registers, otherwise modify
963 * thread save state, i.e., what will be restored in pfm_load_regs()
964 */
965 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
966 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
967 if ((mask & 0x1) == 0UL) continue;
968 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
969 ctx->th_pmcs[i] &= ~0xfUL;
970 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
971 }
972 /*
973 * make all of this visible
974 */
975 ia64_srlz_d();
976 }
977
978 /*
979 * must always be done with task == current
980 *
981 * context must be in MASKED state when calling
982 */
983 static void
984 pfm_restore_monitoring(struct task_struct *task)
985 {
986 pfm_context_t *ctx = PFM_GET_CTX(task);
987 unsigned long mask, ovfl_mask;
988 unsigned long psr, val;
989 int i, is_system;
990
991 is_system = ctx->ctx_fl_system;
992 ovfl_mask = pmu_conf->ovfl_val;
993
994 if (task != current) {
995 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
996 return;
997 }
998 if (ctx->ctx_state != PFM_CTX_MASKED) {
999 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1000 task->pid, current->pid, ctx->ctx_state);
1001 return;
1002 }
1003 psr = pfm_get_psr();
1004 /*
1005 * monitoring is masked via the PMC.
1006 * As we restore their value, we do not want each counter to
1007 * restart right away. We stop monitoring using the PSR,
1008 * restore the PMC (and PMD) and then re-establish the psr
1009 * as it was. Note that there can be no pending overflow at
1010 * this point, because monitoring was MASKED.
1011 *
1012 * system-wide session are pinned and self-monitoring
1013 */
1014 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1015 /* disable dcr pp */
1016 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1017 pfm_clear_psr_pp();
1018 } else {
1019 pfm_clear_psr_up();
1020 }
1021 /*
1022 * first, we restore the PMD
1023 */
1024 mask = ctx->ctx_used_pmds[0];
1025 for (i = 0; mask; i++, mask>>=1) {
1026 /* skip non used pmds */
1027 if ((mask & 0x1) == 0) continue;
1028
1029 if (PMD_IS_COUNTING(i)) {
1030 /*
1031 * we split the 64bit value according to
1032 * counter width
1033 */
1034 val = ctx->ctx_pmds[i].val & ovfl_mask;
1035 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1036 } else {
1037 val = ctx->ctx_pmds[i].val;
1038 }
1039 ia64_set_pmd(i, val);
1040
1041 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1042 i,
1043 ctx->ctx_pmds[i].val,
1044 val));
1045 }
1046 /*
1047 * restore the PMCs
1048 */
1049 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1050 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1051 if ((mask & 0x1) == 0UL) continue;
1052 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1053 ia64_set_pmc(i, ctx->th_pmcs[i]);
1054 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, ctx->th_pmcs[i]));
1055 }
1056 ia64_srlz_d();
1057
1058 /*
1059 * must restore DBR/IBR because could be modified while masked
1060 * XXX: need to optimize
1061 */
1062 if (ctx->ctx_fl_using_dbreg) {
1063 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1064 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1065 }
1066
1067 /*
1068 * now restore PSR
1069 */
1070 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1071 /* enable dcr pp */
1072 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1073 ia64_srlz_i();
1074 }
1075 pfm_set_psr_l(psr);
1076 }
1077
1078 static inline void
1079 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1080 {
1081 int i;
1082
1083 ia64_srlz_d();
1084
1085 for (i=0; mask; i++, mask>>=1) {
1086 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1087 }
1088 }
1089
1090 /*
1091 * reload from thread state (used for ctxw only)
1092 */
1093 static inline void
1094 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1095 {
1096 int i;
1097 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1098
1099 for (i=0; mask; i++, mask>>=1) {
1100 if ((mask & 0x1) == 0) continue;
1101 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1102 ia64_set_pmd(i, val);
1103 }
1104 ia64_srlz_d();
1105 }
1106
1107 /*
1108 * propagate PMD from context to thread-state
1109 */
1110 static inline void
1111 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1112 {
1113 unsigned long ovfl_val = pmu_conf->ovfl_val;
1114 unsigned long mask = ctx->ctx_all_pmds[0];
1115 unsigned long val;
1116 int i;
1117
1118 DPRINT(("mask=0x%lx\n", mask));
1119
1120 for (i=0; mask; i++, mask>>=1) {
1121
1122 val = ctx->ctx_pmds[i].val;
1123
1124 /*
1125 * We break up the 64 bit value into 2 pieces
1126 * the lower bits go to the machine state in the
1127 * thread (will be reloaded on ctxsw in).
1128 * The upper part stays in the soft-counter.
1129 */
1130 if (PMD_IS_COUNTING(i)) {
1131 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1132 val &= ovfl_val;
1133 }
1134 ctx->th_pmds[i] = val;
1135
1136 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1137 i,
1138 ctx->th_pmds[i],
1139 ctx->ctx_pmds[i].val));
1140 }
1141 }
1142
1143 /*
1144 * propagate PMC from context to thread-state
1145 */
1146 static inline void
1147 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1148 {
1149 unsigned long mask = ctx->ctx_all_pmcs[0];
1150 int i;
1151
1152 DPRINT(("mask=0x%lx\n", mask));
1153
1154 for (i=0; mask; i++, mask>>=1) {
1155 /* masking 0 with ovfl_val yields 0 */
1156 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1157 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1158 }
1159 }
1160
1161
1162
1163 static inline void
1164 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1165 {
1166 int i;
1167
1168 for (i=0; mask; i++, mask>>=1) {
1169 if ((mask & 0x1) == 0) continue;
1170 ia64_set_pmc(i, pmcs[i]);
1171 }
1172 ia64_srlz_d();
1173 }
1174
1175 static inline int
1176 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1177 {
1178 return memcmp(a, b, sizeof(pfm_uuid_t));
1179 }
1180
1181 static inline int
1182 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1183 {
1184 int ret = 0;
1185 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1186 return ret;
1187 }
1188
1189 static inline int
1190 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1191 {
1192 int ret = 0;
1193 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1194 return ret;
1195 }
1196
1197
1198 static inline int
1199 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1200 int cpu, void *arg)
1201 {
1202 int ret = 0;
1203 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1204 return ret;
1205 }
1206
1207 static inline int
1208 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1209 int cpu, void *arg)
1210 {
1211 int ret = 0;
1212 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1213 return ret;
1214 }
1215
1216 static inline int
1217 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1218 {
1219 int ret = 0;
1220 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1221 return ret;
1222 }
1223
1224 static inline int
1225 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)
1226 {
1227 int ret = 0;
1228 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1229 return ret;
1230 }
1231
1232 static pfm_buffer_fmt_t *
1233 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1234 {
1235 struct list_head * pos;
1236 pfm_buffer_fmt_t * entry;
1237
1238 list_for_each(pos, &pfm_buffer_fmt_list) {
1239 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1240 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1241 return entry;
1242 }
1243 return NULL;
1244 }
1245
1246 /*
1247 * find a buffer format based on its uuid
1248 */
1249 static pfm_buffer_fmt_t *
1250 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1251 {
1252 pfm_buffer_fmt_t * fmt;
1253 spin_lock(&pfm_buffer_fmt_lock);
1254 fmt = __pfm_find_buffer_fmt(uuid);
1255 spin_unlock(&pfm_buffer_fmt_lock);
1256 return fmt;
1257 }
1258
1259 int
1260 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1261 {
1262 int ret = 0;
1263
1264 /* some sanity checks */
1265 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1266
1267 /* we need at least a handler */
1268 if (fmt->fmt_handler == NULL) return -EINVAL;
1269
1270 /*
1271 * XXX: need check validity of fmt_arg_size
1272 */
1273
1274 spin_lock(&pfm_buffer_fmt_lock);
1275
1276 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1277 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1278 ret = -EBUSY;
1279 goto out;
1280 }
1281 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1282 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1283
1284 out:
1285 spin_unlock(&pfm_buffer_fmt_lock);
1286 return ret;
1287 }
1288 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1289
1290 int
1291 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1292 {
1293 pfm_buffer_fmt_t *fmt;
1294 int ret = 0;
1295
1296 spin_lock(&pfm_buffer_fmt_lock);
1297
1298 fmt = __pfm_find_buffer_fmt(uuid);
1299 if (!fmt) {
1300 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1301 ret = -EINVAL;
1302 goto out;
1303 }
1304 list_del_init(&fmt->fmt_list);
1305 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1306
1307 out:
1308 spin_unlock(&pfm_buffer_fmt_lock);
1309 return ret;
1310
1311 }
1312 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1313
1314 extern void update_pal_halt_status(int);
1315
1316 static int
1317 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1318 {
1319 unsigned long flags;
1320 /*
1321 * validity checks on cpu_mask have been done upstream
1322 */
1323 LOCK_PFS(flags);
1324
1325 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1326 pfm_sessions.pfs_sys_sessions,
1327 pfm_sessions.pfs_task_sessions,
1328 pfm_sessions.pfs_sys_use_dbregs,
1329 is_syswide,
1330 cpu));
1331
1332 if (is_syswide) {
1333 /*
1334 * cannot mix system wide and per-task sessions
1335 */
1336 if (pfm_sessions.pfs_task_sessions > 0UL) {
1337 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1338 pfm_sessions.pfs_task_sessions));
1339 goto abort;
1340 }
1341
1342 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1343
1344 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1345
1346 pfm_sessions.pfs_sys_session[cpu] = task;
1347
1348 pfm_sessions.pfs_sys_sessions++ ;
1349
1350 } else {
1351 if (pfm_sessions.pfs_sys_sessions) goto abort;
1352 pfm_sessions.pfs_task_sessions++;
1353 }
1354
1355 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1356 pfm_sessions.pfs_sys_sessions,
1357 pfm_sessions.pfs_task_sessions,
1358 pfm_sessions.pfs_sys_use_dbregs,
1359 is_syswide,
1360 cpu));
1361
1362 /*
1363 * disable default_idle() to go to PAL_HALT
1364 */
1365 update_pal_halt_status(0);
1366
1367 UNLOCK_PFS(flags);
1368
1369 return 0;
1370
1371 error_conflict:
1372 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1373 pfm_sessions.pfs_sys_session[cpu]->pid,
1374 cpu));
1375 abort:
1376 UNLOCK_PFS(flags);
1377
1378 return -EBUSY;
1379
1380 }
1381
1382 static int
1383 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1384 {
1385 unsigned long flags;
1386 /*
1387 * validity checks on cpu_mask have been done upstream
1388 */
1389 LOCK_PFS(flags);
1390
1391 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1392 pfm_sessions.pfs_sys_sessions,
1393 pfm_sessions.pfs_task_sessions,
1394 pfm_sessions.pfs_sys_use_dbregs,
1395 is_syswide,
1396 cpu));
1397
1398
1399 if (is_syswide) {
1400 pfm_sessions.pfs_sys_session[cpu] = NULL;
1401 /*
1402 * would not work with perfmon+more than one bit in cpu_mask
1403 */
1404 if (ctx && ctx->ctx_fl_using_dbreg) {
1405 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1406 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1407 } else {
1408 pfm_sessions.pfs_sys_use_dbregs--;
1409 }
1410 }
1411 pfm_sessions.pfs_sys_sessions--;
1412 } else {
1413 pfm_sessions.pfs_task_sessions--;
1414 }
1415 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1416 pfm_sessions.pfs_sys_sessions,
1417 pfm_sessions.pfs_task_sessions,
1418 pfm_sessions.pfs_sys_use_dbregs,
1419 is_syswide,
1420 cpu));
1421
1422 /*
1423 * if possible, enable default_idle() to go into PAL_HALT
1424 */
1425 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1426 update_pal_halt_status(1);
1427
1428 UNLOCK_PFS(flags);
1429
1430 return 0;
1431 }
1432
1433 /*
1434 * removes virtual mapping of the sampling buffer.
1435 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1436 * a PROTECT_CTX() section.
1437 */
1438 static int
1439 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1440 {
1441 int r;
1442
1443 /* sanity checks */
1444 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1445 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1446 return -EINVAL;
1447 }
1448
1449 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1450
1451 /*
1452 * does the actual unmapping
1453 */
1454 down_write(&task->mm->mmap_sem);
1455
1456 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1457
1458 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1459
1460 up_write(&task->mm->mmap_sem);
1461 if (r !=0) {
1462 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1463 }
1464
1465 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1466
1467 return 0;
1468 }
1469
1470 /*
1471 * free actual physical storage used by sampling buffer
1472 */
1473 #if 0
1474 static int
1475 pfm_free_smpl_buffer(pfm_context_t *ctx)
1476 {
1477 pfm_buffer_fmt_t *fmt;
1478
1479 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1480
1481 /*
1482 * we won't use the buffer format anymore
1483 */
1484 fmt = ctx->ctx_buf_fmt;
1485
1486 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1487 ctx->ctx_smpl_hdr,
1488 ctx->ctx_smpl_size,
1489 ctx->ctx_smpl_vaddr));
1490
1491 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1492
1493 /*
1494 * free the buffer
1495 */
1496 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1497
1498 ctx->ctx_smpl_hdr = NULL;
1499 ctx->ctx_smpl_size = 0UL;
1500
1501 return 0;
1502
1503 invalid_free:
1504 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1505 return -EINVAL;
1506 }
1507 #endif
1508
1509 static inline void
1510 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1511 {
1512 if (fmt == NULL) return;
1513
1514 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1515
1516 }
1517
1518 /*
1519 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1520 * no real gain from having the whole whorehouse mounted. So we don't need
1521 * any operations on the root directory. However, we need a non-trivial
1522 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1523 */
1524 static struct vfsmount *pfmfs_mnt;
1525
1526 static int __init
1527 init_pfm_fs(void)
1528 {
1529 int err = register_filesystem(&pfm_fs_type);
1530 if (!err) {
1531 pfmfs_mnt = kern_mount(&pfm_fs_type);
1532 err = PTR_ERR(pfmfs_mnt);
1533 if (IS_ERR(pfmfs_mnt))
1534 unregister_filesystem(&pfm_fs_type);
1535 else
1536 err = 0;
1537 }
1538 return err;
1539 }
1540
1541 static ssize_t
1542 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1543 {
1544 pfm_context_t *ctx;
1545 pfm_msg_t *msg;
1546 ssize_t ret;
1547 unsigned long flags;
1548 DECLARE_WAITQUEUE(wait, current);
1549 if (PFM_IS_FILE(filp) == 0) {
1550 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1551 return -EINVAL;
1552 }
1553
1554 ctx = (pfm_context_t *)filp->private_data;
1555 if (ctx == NULL) {
1556 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1557 return -EINVAL;
1558 }
1559
1560 /*
1561 * check even when there is no message
1562 */
1563 if (size < sizeof(pfm_msg_t)) {
1564 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1565 return -EINVAL;
1566 }
1567
1568 PROTECT_CTX(ctx, flags);
1569
1570 /*
1571 * put ourselves on the wait queue
1572 */
1573 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1574
1575
1576 for(;;) {
1577 /*
1578 * check wait queue
1579 */
1580
1581 set_current_state(TASK_INTERRUPTIBLE);
1582
1583 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1584
1585 ret = 0;
1586 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1587
1588 UNPROTECT_CTX(ctx, flags);
1589
1590 /*
1591 * check non-blocking read
1592 */
1593 ret = -EAGAIN;
1594 if(filp->f_flags & O_NONBLOCK) break;
1595
1596 /*
1597 * check pending signals
1598 */
1599 if(signal_pending(current)) {
1600 ret = -EINTR;
1601 break;
1602 }
1603 /*
1604 * no message, so wait
1605 */
1606 schedule();
1607
1608 PROTECT_CTX(ctx, flags);
1609 }
1610 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1611 set_current_state(TASK_RUNNING);
1612 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1613
1614 if (ret < 0) goto abort;
1615
1616 ret = -EINVAL;
1617 msg = pfm_get_next_msg(ctx);
1618 if (msg == NULL) {
1619 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1620 goto abort_locked;
1621 }
1622
1623 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1624
1625 ret = -EFAULT;
1626 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1627
1628 abort_locked:
1629 UNPROTECT_CTX(ctx, flags);
1630 abort:
1631 return ret;
1632 }
1633
1634 static ssize_t
1635 pfm_write(struct file *file, const char __user *ubuf,
1636 size_t size, loff_t *ppos)
1637 {
1638 DPRINT(("pfm_write called\n"));
1639 return -EINVAL;
1640 }
1641
1642 static unsigned int
1643 pfm_poll(struct file *filp, poll_table * wait)
1644 {
1645 pfm_context_t *ctx;
1646 unsigned long flags;
1647 unsigned int mask = 0;
1648
1649 if (PFM_IS_FILE(filp) == 0) {
1650 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1651 return 0;
1652 }
1653
1654 ctx = (pfm_context_t *)filp->private_data;
1655 if (ctx == NULL) {
1656 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1657 return 0;
1658 }
1659
1660
1661 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1662
1663 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1664
1665 PROTECT_CTX(ctx, flags);
1666
1667 if (PFM_CTXQ_EMPTY(ctx) == 0)
1668 mask = POLLIN | POLLRDNORM;
1669
1670 UNPROTECT_CTX(ctx, flags);
1671
1672 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1673
1674 return mask;
1675 }
1676
1677 static int
1678 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1679 {
1680 DPRINT(("pfm_ioctl called\n"));
1681 return -EINVAL;
1682 }
1683
1684 /*
1685 * interrupt cannot be masked when coming here
1686 */
1687 static inline int
1688 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1689 {
1690 int ret;
1691
1692 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1693
1694 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1695 current->pid,
1696 fd,
1697 on,
1698 ctx->ctx_async_queue, ret));
1699
1700 return ret;
1701 }
1702
1703 static int
1704 pfm_fasync(int fd, struct file *filp, int on)
1705 {
1706 pfm_context_t *ctx;
1707 int ret;
1708
1709 if (PFM_IS_FILE(filp) == 0) {
1710 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1711 return -EBADF;
1712 }
1713
1714 ctx = (pfm_context_t *)filp->private_data;
1715 if (ctx == NULL) {
1716 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1717 return -EBADF;
1718 }
1719 /*
1720 * we cannot mask interrupts during this call because this may
1721 * may go to sleep if memory is not readily avalaible.
1722 *
1723 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1724 * done in caller. Serialization of this function is ensured by caller.
1725 */
1726 ret = pfm_do_fasync(fd, filp, ctx, on);
1727
1728
1729 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1730 fd,
1731 on,
1732 ctx->ctx_async_queue, ret));
1733
1734 return ret;
1735 }
1736
1737 #ifdef CONFIG_SMP
1738 /*
1739 * this function is exclusively called from pfm_close().
1740 * The context is not protected at that time, nor are interrupts
1741 * on the remote CPU. That's necessary to avoid deadlocks.
1742 */
1743 static void
1744 pfm_syswide_force_stop(void *info)
1745 {
1746 pfm_context_t *ctx = (pfm_context_t *)info;
1747 struct pt_regs *regs = task_pt_regs(current);
1748 struct task_struct *owner;
1749 unsigned long flags;
1750 int ret;
1751
1752 if (ctx->ctx_cpu != smp_processor_id()) {
1753 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1754 ctx->ctx_cpu,
1755 smp_processor_id());
1756 return;
1757 }
1758 owner = GET_PMU_OWNER();
1759 if (owner != ctx->ctx_task) {
1760 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1761 smp_processor_id(),
1762 owner->pid, ctx->ctx_task->pid);
1763 return;
1764 }
1765 if (GET_PMU_CTX() != ctx) {
1766 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1767 smp_processor_id(),
1768 GET_PMU_CTX(), ctx);
1769 return;
1770 }
1771
1772 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1773 /*
1774 * the context is already protected in pfm_close(), we simply
1775 * need to mask interrupts to avoid a PMU interrupt race on
1776 * this CPU
1777 */
1778 local_irq_save(flags);
1779
1780 ret = pfm_context_unload(ctx, NULL, 0, regs);
1781 if (ret) {
1782 DPRINT(("context_unload returned %d\n", ret));
1783 }
1784
1785 /*
1786 * unmask interrupts, PMU interrupts are now spurious here
1787 */
1788 local_irq_restore(flags);
1789 }
1790
1791 static void
1792 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1793 {
1794 int ret;
1795
1796 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1797 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1798 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1799 }
1800 #endif /* CONFIG_SMP */
1801
1802 /*
1803 * called for each close(). Partially free resources.
1804 * When caller is self-monitoring, the context is unloaded.
1805 */
1806 static int
1807 pfm_flush(struct file *filp, fl_owner_t id)
1808 {
1809 pfm_context_t *ctx;
1810 struct task_struct *task;
1811 struct pt_regs *regs;
1812 unsigned long flags;
1813 unsigned long smpl_buf_size = 0UL;
1814 void *smpl_buf_vaddr = NULL;
1815 int state, is_system;
1816
1817 if (PFM_IS_FILE(filp) == 0) {
1818 DPRINT(("bad magic for\n"));
1819 return -EBADF;
1820 }
1821
1822 ctx = (pfm_context_t *)filp->private_data;
1823 if (ctx == NULL) {
1824 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1825 return -EBADF;
1826 }
1827
1828 /*
1829 * remove our file from the async queue, if we use this mode.
1830 * This can be done without the context being protected. We come
1831 * here when the context has become unreachable by other tasks.
1832 *
1833 * We may still have active monitoring at this point and we may
1834 * end up in pfm_overflow_handler(). However, fasync_helper()
1835 * operates with interrupts disabled and it cleans up the
1836 * queue. If the PMU handler is called prior to entering
1837 * fasync_helper() then it will send a signal. If it is
1838 * invoked after, it will find an empty queue and no
1839 * signal will be sent. In both case, we are safe
1840 */
1841 if (filp->f_flags & FASYNC) {
1842 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1843 pfm_do_fasync (-1, filp, ctx, 0);
1844 }
1845
1846 PROTECT_CTX(ctx, flags);
1847
1848 state = ctx->ctx_state;
1849 is_system = ctx->ctx_fl_system;
1850
1851 task = PFM_CTX_TASK(ctx);
1852 regs = task_pt_regs(task);
1853
1854 DPRINT(("ctx_state=%d is_current=%d\n",
1855 state,
1856 task == current ? 1 : 0));
1857
1858 /*
1859 * if state == UNLOADED, then task is NULL
1860 */
1861
1862 /*
1863 * we must stop and unload because we are losing access to the context.
1864 */
1865 if (task == current) {
1866 #ifdef CONFIG_SMP
1867 /*
1868 * the task IS the owner but it migrated to another CPU: that's bad
1869 * but we must handle this cleanly. Unfortunately, the kernel does
1870 * not provide a mechanism to block migration (while the context is loaded).
1871 *
1872 * We need to release the resource on the ORIGINAL cpu.
1873 */
1874 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1875
1876 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1877 /*
1878 * keep context protected but unmask interrupt for IPI
1879 */
1880 local_irq_restore(flags);
1881
1882 pfm_syswide_cleanup_other_cpu(ctx);
1883
1884 /*
1885 * restore interrupt masking
1886 */
1887 local_irq_save(flags);
1888
1889 /*
1890 * context is unloaded at this point
1891 */
1892 } else
1893 #endif /* CONFIG_SMP */
1894 {
1895
1896 DPRINT(("forcing unload\n"));
1897 /*
1898 * stop and unload, returning with state UNLOADED
1899 * and session unreserved.
1900 */
1901 pfm_context_unload(ctx, NULL, 0, regs);
1902
1903 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1904 }
1905 }
1906
1907 /*
1908 * remove virtual mapping, if any, for the calling task.
1909 * cannot reset ctx field until last user is calling close().
1910 *
1911 * ctx_smpl_vaddr must never be cleared because it is needed
1912 * by every task with access to the context
1913 *
1914 * When called from do_exit(), the mm context is gone already, therefore
1915 * mm is NULL, i.e., the VMA is already gone and we do not have to
1916 * do anything here
1917 */
1918 if (ctx->ctx_smpl_vaddr && current->mm) {
1919 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1920 smpl_buf_size = ctx->ctx_smpl_size;
1921 }
1922
1923 UNPROTECT_CTX(ctx, flags);
1924
1925 /*
1926 * if there was a mapping, then we systematically remove it
1927 * at this point. Cannot be done inside critical section
1928 * because some VM function reenables interrupts.
1929 *
1930 */
1931 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1932
1933 return 0;
1934 }
1935 /*
1936 * called either on explicit close() or from exit_files().
1937 * Only the LAST user of the file gets to this point, i.e., it is
1938 * called only ONCE.
1939 *
1940 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1941 * (fput()),i.e, last task to access the file. Nobody else can access the
1942 * file at this point.
1943 *
1944 * When called from exit_files(), the VMA has been freed because exit_mm()
1945 * is executed before exit_files().
1946 *
1947 * When called from exit_files(), the current task is not yet ZOMBIE but we
1948 * flush the PMU state to the context.
1949 */
1950 static int
1951 pfm_close(struct inode *inode, struct file *filp)
1952 {
1953 pfm_context_t *ctx;
1954 struct task_struct *task;
1955 struct pt_regs *regs;
1956 DECLARE_WAITQUEUE(wait, current);
1957 unsigned long flags;
1958 unsigned long smpl_buf_size = 0UL;
1959 void *smpl_buf_addr = NULL;
1960 int free_possible = 1;
1961 int state, is_system;
1962
1963 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1964
1965 if (PFM_IS_FILE(filp) == 0) {
1966 DPRINT(("bad magic\n"));
1967 return -EBADF;
1968 }
1969
1970 ctx = (pfm_context_t *)filp->private_data;
1971 if (ctx == NULL) {
1972 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1973 return -EBADF;
1974 }
1975
1976 PROTECT_CTX(ctx, flags);
1977
1978 state = ctx->ctx_state;
1979 is_system = ctx->ctx_fl_system;
1980
1981 task = PFM_CTX_TASK(ctx);
1982 regs = task_pt_regs(task);
1983
1984 DPRINT(("ctx_state=%d is_current=%d\n",
1985 state,
1986 task == current ? 1 : 0));
1987
1988 /*
1989 * if task == current, then pfm_flush() unloaded the context
1990 */
1991 if (state == PFM_CTX_UNLOADED) goto doit;
1992
1993 /*
1994 * context is loaded/masked and task != current, we need to
1995 * either force an unload or go zombie
1996 */
1997
1998 /*
1999 * The task is currently blocked or will block after an overflow.
2000 * we must force it to wakeup to get out of the
2001 * MASKED state and transition to the unloaded state by itself.
2002 *
2003 * This situation is only possible for per-task mode
2004 */
2005 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2006
2007 /*
2008 * set a "partial" zombie state to be checked
2009 * upon return from down() in pfm_handle_work().
2010 *
2011 * We cannot use the ZOMBIE state, because it is checked
2012 * by pfm_load_regs() which is called upon wakeup from down().
2013 * In such case, it would free the context and then we would
2014 * return to pfm_handle_work() which would access the
2015 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2016 * but visible to pfm_handle_work().
2017 *
2018 * For some window of time, we have a zombie context with
2019 * ctx_state = MASKED and not ZOMBIE
2020 */
2021 ctx->ctx_fl_going_zombie = 1;
2022
2023 /*
2024 * force task to wake up from MASKED state
2025 */
2026 complete(&ctx->ctx_restart_done);
2027
2028 DPRINT(("waking up ctx_state=%d\n", state));
2029
2030 /*
2031 * put ourself to sleep waiting for the other
2032 * task to report completion
2033 *
2034 * the context is protected by mutex, therefore there
2035 * is no risk of being notified of completion before
2036 * begin actually on the waitq.
2037 */
2038 set_current_state(TASK_INTERRUPTIBLE);
2039 add_wait_queue(&ctx->ctx_zombieq, &wait);
2040
2041 UNPROTECT_CTX(ctx, flags);
2042
2043 /*
2044 * XXX: check for signals :
2045 * - ok for explicit close
2046 * - not ok when coming from exit_files()
2047 */
2048 schedule();
2049
2050
2051 PROTECT_CTX(ctx, flags);
2052
2053
2054 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2055 set_current_state(TASK_RUNNING);
2056
2057 /*
2058 * context is unloaded at this point
2059 */
2060 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2061 }
2062 else if (task != current) {
2063 #ifdef CONFIG_SMP
2064 /*
2065 * switch context to zombie state
2066 */
2067 ctx->ctx_state = PFM_CTX_ZOMBIE;
2068
2069 DPRINT(("zombie ctx for [%d]\n", task->pid));
2070 /*
2071 * cannot free the context on the spot. deferred until
2072 * the task notices the ZOMBIE state
2073 */
2074 free_possible = 0;
2075 #else
2076 pfm_context_unload(ctx, NULL, 0, regs);
2077 #endif
2078 }
2079
2080 doit:
2081 /* reload state, may have changed during opening of critical section */
2082 state = ctx->ctx_state;
2083
2084 /*
2085 * the context is still attached to a task (possibly current)
2086 * we cannot destroy it right now
2087 */
2088
2089 /*
2090 * we must free the sampling buffer right here because
2091 * we cannot rely on it being cleaned up later by the
2092 * monitored task. It is not possible to free vmalloc'ed
2093 * memory in pfm_load_regs(). Instead, we remove the buffer
2094 * now. should there be subsequent PMU overflow originally
2095 * meant for sampling, the will be converted to spurious
2096 * and that's fine because the monitoring tools is gone anyway.
2097 */
2098 if (ctx->ctx_smpl_hdr) {
2099 smpl_buf_addr = ctx->ctx_smpl_hdr;
2100 smpl_buf_size = ctx->ctx_smpl_size;
2101 /* no more sampling */
2102 ctx->ctx_smpl_hdr = NULL;
2103 ctx->ctx_fl_is_sampling = 0;
2104 }
2105
2106 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2107 state,
2108 free_possible,
2109 smpl_buf_addr,
2110 smpl_buf_size));
2111
2112 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2113
2114 /*
2115 * UNLOADED that the session has already been unreserved.
2116 */
2117 if (state == PFM_CTX_ZOMBIE) {
2118 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2119 }
2120
2121 /*
2122 * disconnect file descriptor from context must be done
2123 * before we unlock.
2124 */
2125 filp->private_data = NULL;
2126
2127 /*
2128 * if we free on the spot, the context is now completely unreachable
2129 * from the callers side. The monitored task side is also cut, so we
2130 * can freely cut.
2131 *
2132 * If we have a deferred free, only the caller side is disconnected.
2133 */
2134 UNPROTECT_CTX(ctx, flags);
2135
2136 /*
2137 * All memory free operations (especially for vmalloc'ed memory)
2138 * MUST be done with interrupts ENABLED.
2139 */
2140 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2141
2142 /*
2143 * return the memory used by the context
2144 */
2145 if (free_possible) pfm_context_free(ctx);
2146
2147 return 0;
2148 }
2149
2150 static int
2151 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2152 {
2153 DPRINT(("pfm_no_open called\n"));
2154 return -ENXIO;
2155 }
2156
2157
2158
2159 static const struct file_operations pfm_file_ops = {
2160 .llseek = no_llseek,
2161 .read = pfm_read,
2162 .write = pfm_write,
2163 .poll = pfm_poll,
2164 .ioctl = pfm_ioctl,
2165 .open = pfm_no_open, /* special open code to disallow open via /proc */
2166 .fasync = pfm_fasync,
2167 .release = pfm_close,
2168 .flush = pfm_flush
2169 };
2170
2171 static int
2172 pfmfs_delete_dentry(struct dentry *dentry)
2173 {
2174 return 1;
2175 }
2176
2177 static struct dentry_operations pfmfs_dentry_operations = {
2178 .d_delete = pfmfs_delete_dentry,
2179 };
2180
2181
2182 static int
2183 pfm_alloc_fd(struct file **cfile)
2184 {
2185 int fd, ret = 0;
2186 struct file *file = NULL;
2187 struct inode * inode;
2188 char name[32];
2189 struct qstr this;
2190
2191 fd = get_unused_fd();
2192 if (fd < 0) return -ENFILE;
2193
2194 ret = -ENFILE;
2195
2196 file = get_empty_filp();
2197 if (!file) goto out;
2198
2199 /*
2200 * allocate a new inode
2201 */
2202 inode = new_inode(pfmfs_mnt->mnt_sb);
2203 if (!inode) goto out;
2204
2205 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2206
2207 inode->i_mode = S_IFCHR|S_IRUGO;
2208 inode->i_uid = current->fsuid;
2209 inode->i_gid = current->fsgid;
2210
2211 sprintf(name, "[%lu]", inode->i_ino);
2212 this.name = name;
2213 this.len = strlen(name);
2214 this.hash = inode->i_ino;
2215
2216 ret = -ENOMEM;
2217
2218 /*
2219 * allocate a new dcache entry
2220 */
2221 file->f_path.dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2222 if (!file->f_path.dentry) goto out;
2223
2224 file->f_path.dentry->d_op = &pfmfs_dentry_operations;
2225
2226 d_add(file->f_path.dentry, inode);
2227 file->f_path.mnt = mntget(pfmfs_mnt);
2228 file->f_mapping = inode->i_mapping;
2229
2230 file->f_op = &pfm_file_ops;
2231 file->f_mode = FMODE_READ;
2232 file->f_flags = O_RDONLY;
2233 file->f_pos = 0;
2234
2235 /*
2236 * may have to delay until context is attached?
2237 */
2238 fd_install(fd, file);
2239
2240 /*
2241 * the file structure we will use
2242 */
2243 *cfile = file;
2244
2245 return fd;
2246 out:
2247 if (file) put_filp(file);
2248 put_unused_fd(fd);
2249 return ret;
2250 }
2251
2252 static void
2253 pfm_free_fd(int fd, struct file *file)
2254 {
2255 struct files_struct *files = current->files;
2256 struct fdtable *fdt;
2257
2258 /*
2259 * there ie no fd_uninstall(), so we do it here
2260 */
2261 spin_lock(&files->file_lock);
2262 fdt = files_fdtable(files);
2263 rcu_assign_pointer(fdt->fd[fd], NULL);
2264 spin_unlock(&files->file_lock);
2265
2266 if (file)
2267 put_filp(file);
2268 put_unused_fd(fd);
2269 }
2270
2271 static int
2272 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2273 {
2274 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2275
2276 while (size > 0) {
2277 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2278
2279
2280 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2281 return -ENOMEM;
2282
2283 addr += PAGE_SIZE;
2284 buf += PAGE_SIZE;
2285 size -= PAGE_SIZE;
2286 }
2287 return 0;
2288 }
2289
2290 /*
2291 * allocate a sampling buffer and remaps it into the user address space of the task
2292 */
2293 static int
2294 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2295 {
2296 struct mm_struct *mm = task->mm;
2297 struct vm_area_struct *vma = NULL;
2298 unsigned long size;
2299 void *smpl_buf;
2300
2301
2302 /*
2303 * the fixed header + requested size and align to page boundary
2304 */
2305 size = PAGE_ALIGN(rsize);
2306
2307 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2308
2309 /*
2310 * check requested size to avoid Denial-of-service attacks
2311 * XXX: may have to refine this test
2312 * Check against address space limit.
2313 *
2314 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2315 * return -ENOMEM;
2316 */
2317 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2318 return -ENOMEM;
2319
2320 /*
2321 * We do the easy to undo allocations first.
2322 *
2323 * pfm_rvmalloc(), clears the buffer, so there is no leak
2324 */
2325 smpl_buf = pfm_rvmalloc(size);
2326 if (smpl_buf == NULL) {
2327 DPRINT(("Can't allocate sampling buffer\n"));
2328 return -ENOMEM;
2329 }
2330
2331 DPRINT(("smpl_buf @%p\n", smpl_buf));
2332
2333 /* allocate vma */
2334 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2335 if (!vma) {
2336 DPRINT(("Cannot allocate vma\n"));
2337 goto error_kmem;
2338 }
2339
2340 /*
2341 * partially initialize the vma for the sampling buffer
2342 */
2343 vma->vm_mm = mm;
2344 vma->vm_file = filp;
2345 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2346 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2347
2348 /*
2349 * Now we have everything we need and we can initialize
2350 * and connect all the data structures
2351 */
2352
2353 ctx->ctx_smpl_hdr = smpl_buf;
2354 ctx->ctx_smpl_size = size; /* aligned size */
2355
2356 /*
2357 * Let's do the difficult operations next.
2358 *
2359 * now we atomically find some area in the address space and
2360 * remap the buffer in it.
2361 */
2362 down_write(&task->mm->mmap_sem);
2363
2364 /* find some free area in address space, must have mmap sem held */
2365 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2366 if (vma->vm_start == 0UL) {
2367 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2368 up_write(&task->mm->mmap_sem);
2369 goto error;
2370 }
2371 vma->vm_end = vma->vm_start + size;
2372 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2373
2374 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2375
2376 /* can only be applied to current task, need to have the mm semaphore held when called */
2377 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2378 DPRINT(("Can't remap buffer\n"));
2379 up_write(&task->mm->mmap_sem);
2380 goto error;
2381 }
2382
2383 get_file(filp);
2384
2385 /*
2386 * now insert the vma in the vm list for the process, must be
2387 * done with mmap lock held
2388 */
2389 insert_vm_struct(mm, vma);
2390
2391 mm->total_vm += size >> PAGE_SHIFT;
2392 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2393 vma_pages(vma));
2394 up_write(&task->mm->mmap_sem);
2395
2396 /*
2397 * keep track of user level virtual address
2398 */
2399 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2400 *(unsigned long *)user_vaddr = vma->vm_start;
2401
2402 return 0;
2403
2404 error:
2405 kmem_cache_free(vm_area_cachep, vma);
2406 error_kmem:
2407 pfm_rvfree(smpl_buf, size);
2408
2409 return -ENOMEM;
2410 }
2411
2412 /*
2413 * XXX: do something better here
2414 */
2415 static int
2416 pfm_bad_permissions(struct task_struct *task)
2417 {
2418 /* inspired by ptrace_attach() */
2419 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2420 current->uid,
2421 current->gid,
2422 task->euid,
2423 task->suid,
2424 task->uid,
2425 task->egid,
2426 task->sgid));
2427
2428 return ((current->uid != task->euid)
2429 || (current->uid != task->suid)
2430 || (current->uid != task->uid)
2431 || (current->gid != task->egid)
2432 || (current->gid != task->sgid)
2433 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2434 }
2435
2436 static int
2437 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2438 {
2439 int ctx_flags;
2440
2441 /* valid signal */
2442
2443 ctx_flags = pfx->ctx_flags;
2444
2445 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2446
2447 /*
2448 * cannot block in this mode
2449 */
2450 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2451 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2452 return -EINVAL;
2453 }
2454 } else {
2455 }
2456 /* probably more to add here */
2457
2458 return 0;
2459 }
2460
2461 static int
2462 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2463 unsigned int cpu, pfarg_context_t *arg)
2464 {
2465 pfm_buffer_fmt_t *fmt = NULL;
2466 unsigned long size = 0UL;
2467 void *uaddr = NULL;
2468 void *fmt_arg = NULL;
2469 int ret = 0;
2470 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2471
2472 /* invoke and lock buffer format, if found */
2473 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2474 if (fmt == NULL) {
2475 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2476 return -EINVAL;
2477 }
2478
2479 /*
2480 * buffer argument MUST be contiguous to pfarg_context_t
2481 */
2482 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2483
2484 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2485
2486 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2487
2488 if (ret) goto error;
2489
2490 /* link buffer format and context */
2491 ctx->ctx_buf_fmt = fmt;
2492
2493 /*
2494 * check if buffer format wants to use perfmon buffer allocation/mapping service
2495 */
2496 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2497 if (ret) goto error;
2498
2499 if (size) {
2500 /*
2501 * buffer is always remapped into the caller's address space
2502 */
2503 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2504 if (ret) goto error;
2505
2506 /* keep track of user address of buffer */
2507 arg->ctx_smpl_vaddr = uaddr;
2508 }
2509 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2510
2511 error:
2512 return ret;
2513 }
2514
2515 static void
2516 pfm_reset_pmu_state(pfm_context_t *ctx)
2517 {
2518 int i;
2519
2520 /*
2521 * install reset values for PMC.
2522 */
2523 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2524 if (PMC_IS_IMPL(i) == 0) continue;
2525 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2526 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2527 }
2528 /*
2529 * PMD registers are set to 0UL when the context in memset()
2530 */
2531
2532 /*
2533 * On context switched restore, we must restore ALL pmc and ALL pmd even
2534 * when they are not actively used by the task. In UP, the incoming process
2535 * may otherwise pick up left over PMC, PMD state from the previous process.
2536 * As opposed to PMD, stale PMC can cause harm to the incoming
2537 * process because they may change what is being measured.
2538 * Therefore, we must systematically reinstall the entire
2539 * PMC state. In SMP, the same thing is possible on the
2540 * same CPU but also on between 2 CPUs.
2541 *
2542 * The problem with PMD is information leaking especially
2543 * to user level when psr.sp=0
2544 *
2545 * There is unfortunately no easy way to avoid this problem
2546 * on either UP or SMP. This definitively slows down the
2547 * pfm_load_regs() function.
2548 */
2549
2550 /*
2551 * bitmask of all PMCs accessible to this context
2552 *
2553 * PMC0 is treated differently.
2554 */
2555 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2556
2557 /*
2558 * bitmask of all PMDs that are accessible to this context
2559 */
2560 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2561
2562 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2563
2564 /*
2565 * useful in case of re-enable after disable
2566 */
2567 ctx->ctx_used_ibrs[0] = 0UL;
2568 ctx->ctx_used_dbrs[0] = 0UL;
2569 }
2570
2571 static int
2572 pfm_ctx_getsize(void *arg, size_t *sz)
2573 {
2574 pfarg_context_t *req = (pfarg_context_t *)arg;
2575 pfm_buffer_fmt_t *fmt;
2576
2577 *sz = 0;
2578
2579 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2580
2581 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2582 if (fmt == NULL) {
2583 DPRINT(("cannot find buffer format\n"));
2584 return -EINVAL;
2585 }
2586 /* get just enough to copy in user parameters */
2587 *sz = fmt->fmt_arg_size;
2588 DPRINT(("arg_size=%lu\n", *sz));
2589
2590 return 0;
2591 }
2592
2593
2594
2595 /*
2596 * cannot attach if :
2597 * - kernel task
2598 * - task not owned by caller
2599 * - task incompatible with context mode
2600 */
2601 static int
2602 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2603 {
2604 /*
2605 * no kernel task or task not owner by caller
2606 */
2607 if (task->mm == NULL) {
2608 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2609 return -EPERM;
2610 }
2611 if (pfm_bad_permissions(task)) {
2612 DPRINT(("no permission to attach to [%d]\n", task->pid));
2613 return -EPERM;
2614 }
2615 /*
2616 * cannot block in self-monitoring mode
2617 */
2618 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2619 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2620 return -EINVAL;
2621 }
2622
2623 if (task->exit_state == EXIT_ZOMBIE) {
2624 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2625 return -EBUSY;
2626 }
2627
2628 /*
2629 * always ok for self
2630 */
2631 if (task == current) return 0;
2632
2633 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
2634 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2635 return -EBUSY;
2636 }
2637 /*
2638 * make sure the task is off any CPU
2639 */
2640 wait_task_inactive(task);
2641
2642 /* more to come... */
2643
2644 return 0;
2645 }
2646
2647 static int
2648 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2649 {
2650 struct task_struct *p = current;
2651 int ret;
2652
2653 /* XXX: need to add more checks here */
2654 if (pid < 2) return -EPERM;
2655
2656 if (pid != current->pid) {
2657
2658 read_lock(&tasklist_lock);
2659
2660 p = find_task_by_pid(pid);
2661
2662 /* make sure task cannot go away while we operate on it */
2663 if (p) get_task_struct(p);
2664
2665 read_unlock(&tasklist_lock);
2666
2667 if (p == NULL) return -ESRCH;
2668 }
2669
2670 ret = pfm_task_incompatible(ctx, p);
2671 if (ret == 0) {
2672 *task = p;
2673 } else if (p != current) {
2674 pfm_put_task(p);
2675 }
2676 return ret;
2677 }
2678
2679
2680
2681 static int
2682 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2683 {
2684 pfarg_context_t *req = (pfarg_context_t *)arg;
2685 struct file *filp;
2686 int ctx_flags;
2687 int ret;
2688
2689 /* let's check the arguments first */
2690 ret = pfarg_is_sane(current, req);
2691 if (ret < 0) return ret;
2692
2693 ctx_flags = req->ctx_flags;
2694
2695 ret = -ENOMEM;
2696
2697 ctx = pfm_context_alloc();
2698 if (!ctx) goto error;
2699
2700 ret = pfm_alloc_fd(&filp);
2701 if (ret < 0) goto error_file;
2702
2703 req->ctx_fd = ctx->ctx_fd = ret;
2704
2705 /*
2706 * attach context to file
2707 */
2708 filp->private_data = ctx;
2709
2710 /*
2711 * does the user want to sample?
2712 */
2713 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2714 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2715 if (ret) goto buffer_error;
2716 }
2717
2718 /*
2719 * init context protection lock
2720 */
2721 spin_lock_init(&ctx->ctx_lock);
2722
2723 /*
2724 * context is unloaded
2725 */
2726 ctx->ctx_state = PFM_CTX_UNLOADED;
2727
2728 /*
2729 * initialization of context's flags
2730 */
2731 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2732 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2733 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2734 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2735 /*
2736 * will move to set properties
2737 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2738 */
2739
2740 /*
2741 * init restart semaphore to locked
2742 */
2743 init_completion(&ctx->ctx_restart_done);
2744
2745 /*
2746 * activation is used in SMP only
2747 */
2748 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2749 SET_LAST_CPU(ctx, -1);
2750
2751 /*
2752 * initialize notification message queue
2753 */
2754 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2755 init_waitqueue_head(&ctx->ctx_msgq_wait);
2756 init_waitqueue_head(&ctx->ctx_zombieq);
2757
2758 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2759 ctx,
2760 ctx_flags,
2761 ctx->ctx_fl_system,
2762 ctx->ctx_fl_block,
2763 ctx->ctx_fl_excl_idle,
2764 ctx->ctx_fl_no_msg,
2765 ctx->ctx_fd));
2766
2767 /*
2768 * initialize soft PMU state
2769 */
2770 pfm_reset_pmu_state(ctx);
2771
2772 return 0;
2773
2774 buffer_error:
2775 pfm_free_fd(ctx->ctx_fd, filp);
2776
2777 if (ctx->ctx_buf_fmt) {
2778 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2779 }
2780 error_file:
2781 pfm_context_free(ctx);
2782
2783 error:
2784 return ret;
2785 }
2786
2787 static inline unsigned long
2788 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2789 {
2790 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2791 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2792 extern unsigned long carta_random32 (unsigned long seed);
2793
2794 if (reg->flags & PFM_REGFL_RANDOM) {
2795 new_seed = carta_random32(old_seed);
2796 val -= (old_seed & mask); /* counter values are negative numbers! */
2797 if ((mask >> 32) != 0)
2798 /* construct a full 64-bit random value: */
2799 new_seed |= carta_random32(old_seed >> 32) << 32;
2800 reg->seed = new_seed;
2801 }
2802 reg->lval = val;
2803 return val;
2804 }
2805
2806 static void
2807 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2808 {
2809 unsigned long mask = ovfl_regs[0];
2810 unsigned long reset_others = 0UL;
2811 unsigned long val;
2812 int i;
2813
2814 /*
2815 * now restore reset value on sampling overflowed counters
2816 */
2817 mask >>= PMU_FIRST_COUNTER;
2818 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2819
2820 if ((mask & 0x1UL) == 0UL) continue;
2821
2822 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2823 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2824
2825 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2826 }
2827
2828 /*
2829 * Now take care of resetting the other registers
2830 */
2831 for(i = 0; reset_others; i++, reset_others >>= 1) {
2832
2833 if ((reset_others & 0x1) == 0) continue;
2834
2835 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2836
2837 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2838 is_long_reset ? "long" : "short", i, val));
2839 }
2840 }
2841
2842 static void
2843 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2844 {
2845 unsigned long mask = ovfl_regs[0];
2846 unsigned long reset_others = 0UL;
2847 unsigned long val;
2848 int i;
2849
2850 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2851
2852 if (ctx->ctx_state == PFM_CTX_MASKED) {
2853 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2854 return;
2855 }
2856
2857 /*
2858 * now restore reset value on sampling overflowed counters
2859 */
2860 mask >>= PMU_FIRST_COUNTER;
2861 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2862
2863 if ((mask & 0x1UL) == 0UL) continue;
2864
2865 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2866 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2867
2868 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2869
2870 pfm_write_soft_counter(ctx, i, val);
2871 }
2872
2873 /*
2874 * Now take care of resetting the other registers
2875 */
2876 for(i = 0; reset_others; i++, reset_others >>= 1) {
2877
2878 if ((reset_others & 0x1) == 0) continue;
2879
2880 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2881
2882 if (PMD_IS_COUNTING(i)) {
2883 pfm_write_soft_counter(ctx, i, val);
2884 } else {
2885 ia64_set_pmd(i, val);
2886 }
2887 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2888 is_long_reset ? "long" : "short", i, val));
2889 }
2890 ia64_srlz_d();
2891 }
2892
2893 static int
2894 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2895 {
2896 struct task_struct *task;
2897 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2898 unsigned long value, pmc_pm;
2899 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2900 unsigned int cnum, reg_flags, flags, pmc_type;
2901 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2902 int is_monitor, is_counting, state;
2903 int ret = -EINVAL;
2904 pfm_reg_check_t wr_func;
2905 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2906
2907 state = ctx->ctx_state;
2908 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2909 is_system = ctx->ctx_fl_system;
2910 task = ctx->ctx_task;
2911 impl_pmds = pmu_conf->impl_pmds[0];
2912
2913 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2914
2915 if (is_loaded) {
2916 /*
2917 * In system wide and when the context is loaded, access can only happen
2918 * when the caller is running on the CPU being monitored by the session.
2919 * It does not have to be the owner (ctx_task) of the context per se.
2920 */
2921 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2922 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2923 return -EBUSY;
2924 }
2925 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2926 }
2927 expert_mode = pfm_sysctl.expert_mode;
2928
2929 for (i = 0; i < count; i++, req++) {
2930
2931 cnum = req->reg_num;
2932 reg_flags = req->reg_flags;
2933 value = req->reg_value;
2934 smpl_pmds = req->reg_smpl_pmds[0];
2935 reset_pmds = req->reg_reset_pmds[0];
2936 flags = 0;
2937
2938
2939 if (cnum >= PMU_MAX_PMCS) {
2940 DPRINT(("pmc%u is invalid\n", cnum));
2941 goto error;
2942 }
2943
2944 pmc_type = pmu_conf->pmc_desc[cnum].type;
2945 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2946 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2947 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2948
2949 /*
2950 * we reject all non implemented PMC as well
2951 * as attempts to modify PMC[0-3] which are used
2952 * as status registers by the PMU
2953 */
2954 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2955 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2956 goto error;
2957 }
2958 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2959 /*
2960 * If the PMC is a monitor, then if the value is not the default:
2961 * - system-wide session: PMCx.pm=1 (privileged monitor)
2962 * - per-task : PMCx.pm=0 (user monitor)
2963 */
2964 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2965 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2966 cnum,
2967 pmc_pm,
2968 is_system));
2969 goto error;
2970 }
2971
2972 if (is_counting) {
2973 /*
2974 * enforce generation of overflow interrupt. Necessary on all
2975 * CPUs.
2976 */
2977 value |= 1 << PMU_PMC_OI;
2978
2979 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2980 flags |= PFM_REGFL_OVFL_NOTIFY;
2981 }
2982
2983 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2984
2985 /* verify validity of smpl_pmds */
2986 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2987 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2988 goto error;
2989 }
2990
2991 /* verify validity of reset_pmds */
2992 if ((reset_pmds & impl_pmds) != reset_pmds) {
2993 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2994 goto error;
2995 }
2996 } else {
2997 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2998 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2999 goto error;
3000 }
3001 /* eventid on non-counting monitors are ignored */
3002 }
3003
3004 /*
3005 * execute write checker, if any
3006 */
3007 if (likely(expert_mode == 0 && wr_func)) {
3008 ret = (*wr_func)(task, ctx, cnum, &value, regs);
3009 if (ret) goto error;
3010 ret = -EINVAL;
3011 }
3012
3013 /*
3014 * no error on this register
3015 */
3016 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3017
3018 /*
3019 * Now we commit the changes to the software state
3020 */
3021
3022 /*
3023 * update overflow information
3024 */
3025 if (is_counting) {
3026 /*
3027 * full flag update each time a register is programmed
3028 */
3029 ctx->ctx_pmds[cnum].flags = flags;
3030
3031 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
3032 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
3033 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
3034
3035 /*
3036 * Mark all PMDS to be accessed as used.
3037 *
3038 * We do not keep track of PMC because we have to
3039 * systematically restore ALL of them.
3040 *
3041 * We do not update the used_monitors mask, because
3042 * if we have not programmed them, then will be in
3043 * a quiescent state, therefore we will not need to
3044 * mask/restore then when context is MASKED.
3045 */
3046 CTX_USED_PMD(ctx, reset_pmds);
3047 CTX_USED_PMD(ctx, smpl_pmds);
3048 /*
3049 * make sure we do not try to reset on
3050 * restart because we have established new values
3051 */
3052 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3053 }
3054 /*
3055 * Needed in case the user does not initialize the equivalent
3056 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3057 * possible leak here.
3058 */
3059 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3060
3061 /*
3062 * keep track of the monitor PMC that we are using.
3063 * we save the value of the pmc in ctx_pmcs[] and if
3064 * the monitoring is not stopped for the context we also
3065 * place it in the saved state area so that it will be
3066 * picked up later by the context switch code.
3067 *
3068 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3069 *
3070 * The value in th_pmcs[] may be modified on overflow, i.e., when
3071 * monitoring needs to be stopped.
3072 */
3073 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3074
3075 /*
3076 * update context state
3077 */
3078 ctx->ctx_pmcs[cnum] = value;
3079
3080 if (is_loaded) {
3081 /*
3082 * write thread state
3083 */
3084 if (is_system == 0) ctx->th_pmcs[cnum] = value;
3085
3086 /*
3087 * write hardware register if we can
3088 */
3089 if (can_access_pmu) {
3090 ia64_set_pmc(cnum, value);
3091 }
3092 #ifdef CONFIG_SMP
3093 else {
3094 /*
3095 * per-task SMP only here
3096 *
3097 * we are guaranteed that the task is not running on the other CPU,
3098 * we indicate that this PMD will need to be reloaded if the task
3099 * is rescheduled on the CPU it ran last on.
3100 */
3101 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3102 }
3103 #endif
3104 }
3105
3106 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",
3107 cnum,
3108 value,
3109 is_loaded,
3110 can_access_pmu,
3111 flags,
3112 ctx->ctx_all_pmcs[0],
3113 ctx->ctx_used_pmds[0],
3114 ctx->ctx_pmds[cnum].eventid,
3115 smpl_pmds,
3116 reset_pmds,
3117 ctx->ctx_reload_pmcs[0],
3118 ctx->ctx_used_monitors[0],
3119 ctx->ctx_ovfl_regs[0]));
3120 }
3121
3122 /*
3123 * make sure the changes are visible
3124 */
3125 if (can_access_pmu) ia64_srlz_d();
3126
3127 return 0;
3128 error:
3129 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3130 return ret;
3131 }
3132
3133 static int
3134 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3135 {
3136 struct task_struct *task;
3137 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3138 unsigned long value, hw_value, ovfl_mask;
3139 unsigned int cnum;
3140 int i, can_access_pmu = 0, state;
3141 int is_counting, is_loaded, is_system, expert_mode;
3142 int ret = -EINVAL;
3143 pfm_reg_check_t wr_func;
3144
3145
3146 state = ctx->ctx_state;
3147 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3148 is_system = ctx->ctx_fl_system;
3149 ovfl_mask = pmu_conf->ovfl_val;
3150 task = ctx->ctx_task;
3151
3152 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3153
3154 /*
3155 * on both UP and SMP, we can only write to the PMC when the task is
3156 * the owner of the local PMU.
3157 */
3158 if (likely(is_loaded)) {
3159 /*
3160 * In system wide and when the context is loaded, access can only happen
3161 * when the caller is running on the CPU being monitored by the session.
3162 * It does not have to be the owner (ctx_task) of the context per se.
3163 */
3164 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3165 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3166 return -EBUSY;
3167 }
3168 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3169 }
3170 expert_mode = pfm_sysctl.expert_mode;
3171
3172 for (i = 0; i < count; i++, req++) {
3173
3174 cnum = req->reg_num;
3175 value = req->reg_value;
3176
3177 if (!PMD_IS_IMPL(cnum)) {
3178 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3179 goto abort_mission;
3180 }
3181 is_counting = PMD_IS_COUNTING(cnum);
3182 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3183
3184 /*
3185 * execute write checker, if any
3186 */
3187 if (unlikely(expert_mode == 0 && wr_func)) {
3188 unsigned long v = value;
3189
3190 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3191 if (ret) goto abort_mission;
3192
3193 value = v;
3194 ret = -EINVAL;
3195 }
3196
3197 /*
3198 * no error on this register
3199 */
3200 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3201
3202 /*
3203 * now commit changes to software state
3204 */
3205 hw_value = value;
3206
3207 /*
3208 * update virtualized (64bits) counter
3209 */
3210 if (is_counting) {
3211 /*
3212 * write context state
3213 */
3214 ctx->ctx_pmds[cnum].lval = value;
3215
3216 /*
3217 * when context is load we use the split value
3218 */
3219 if (is_loaded) {
3220 hw_value = value & ovfl_mask;
3221 value = value & ~ovfl_mask;
3222 }
3223 }
3224 /*
3225 * update reset values (not just for counters)
3226 */
3227 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3228 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3229
3230 /*
3231 * update randomization parameters (not just for counters)
3232 */
3233 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3234 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3235
3236 /*
3237 * update context value
3238 */
3239 ctx->ctx_pmds[cnum].val = value;
3240
3241 /*
3242 * Keep track of what we use
3243 *
3244 * We do not keep track of PMC because we have to
3245 * systematically restore ALL of them.
3246 */
3247 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3248
3249 /*
3250 * mark this PMD register used as well
3251 */
3252 CTX_USED_PMD(ctx, RDEP(cnum));
3253
3254 /*
3255 * make sure we do not try to reset on
3256 * restart because we have established new values
3257 */
3258 if (is_counting && state == PFM_CTX_MASKED) {
3259 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3260 }
3261
3262 if (is_loaded) {
3263 /*
3264 * write thread state
3265 */
3266 if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3267
3268 /*
3269 * write hardware register if we can
3270 */
3271 if (can_access_pmu) {
3272 ia64_set_pmd(cnum, hw_value);
3273 } else {
3274 #ifdef CONFIG_SMP
3275 /*
3276 * we are guaranteed that the task is not running on the other CPU,
3277 * we indicate that this PMD will need to be reloaded if the task
3278 * is rescheduled on the CPU it ran last on.
3279 */
3280 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3281 #endif
3282 }
3283 }
3284
3285 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3286 "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",
3287 cnum,
3288 value,
3289 is_loaded,
3290 can_access_pmu,
3291 hw_value,
3292 ctx->ctx_pmds[cnum].val,
3293 ctx->ctx_pmds[cnum].short_reset,
3294 ctx->ctx_pmds[cnum].long_reset,
3295 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3296 ctx->ctx_pmds[cnum].seed,
3297 ctx->ctx_pmds[cnum].mask,
3298 ctx->ctx_used_pmds[0],
3299 ctx->ctx_pmds[cnum].reset_pmds[0],
3300 ctx->ctx_reload_pmds[0],
3301 ctx->ctx_all_pmds[0],
3302 ctx->ctx_ovfl_regs[0]));
3303 }
3304
3305 /*
3306 * make changes visible
3307 */
3308 if (can_access_pmu) ia64_srlz_d();
3309
3310 return 0;
3311
3312 abort_mission:
3313 /*
3314 * for now, we have only one possibility for error
3315 */
3316 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3317 return ret;
3318 }
3319
3320 /*
3321 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3322 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3323 * interrupt is delivered during the call, it will be kept pending until we leave, making
3324 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3325 * guaranteed to return consistent data to the user, it may simply be old. It is not
3326 * trivial to treat the overflow while inside the call because you may end up in
3327 * some module sampling buffer code causing deadlocks.
3328 */
3329 static int
3330 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3331 {
3332 struct task_struct *task;
3333 unsigned long val = 0UL, lval, ovfl_mask, sval;
3334 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3335 unsigned int cnum, reg_flags = 0;
3336 int i, can_access_pmu = 0, state;
3337 int is_loaded, is_system, is_counting, expert_mode;
3338 int ret = -EINVAL;
3339 pfm_reg_check_t rd_func;
3340
3341 /*
3342 * access is possible when loaded only for
3343 * self-monitoring tasks or in UP mode
3344 */
3345
3346 state = ctx->ctx_state;
3347 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3348 is_system = ctx->ctx_fl_system;
3349 ovfl_mask = pmu_conf->ovfl_val;
3350 task = ctx->ctx_task;
3351
3352 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3353
3354 if (likely(is_loaded)) {
3355 /*
3356 * In system wide and when the context is loaded, access can only happen
3357 * when the caller is running on the CPU being monitored by the session.
3358 * It does not have to be the owner (ctx_task) of the context per se.
3359 */
3360 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3361 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3362 return -EBUSY;
3363 }
3364 /*
3365 * this can be true when not self-monitoring only in UP
3366 */
3367 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3368
3369 if (can_access_pmu) ia64_srlz_d();
3370 }
3371 expert_mode = pfm_sysctl.expert_mode;
3372
3373 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3374 is_loaded,
3375 can_access_pmu,
3376 state));
3377
3378 /*
3379 * on both UP and SMP, we can only read the PMD from the hardware register when
3380 * the task is the owner of the local PMU.
3381 */
3382
3383 for (i = 0; i < count; i++, req++) {
3384
3385 cnum = req->reg_num;
3386 reg_flags = req->reg_flags;
3387
3388 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3389 /*
3390 * we can only read the register that we use. That includes
3391 * the one we explicitly initialize AND the one we want included
3392 * in the sampling buffer (smpl_regs).
3393 *
3394 * Having this restriction allows optimization in the ctxsw routine
3395 * without compromising security (leaks)
3396 */
3397 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3398
3399 sval = ctx->ctx_pmds[cnum].val;
3400 lval = ctx->ctx_pmds[cnum].lval;
3401 is_counting = PMD_IS_COUNTING(cnum);
3402
3403 /*
3404 * If the task is not the current one, then we check if the
3405 * PMU state is still in the local live register due to lazy ctxsw.
3406 * If true, then we read directly from the registers.
3407 */
3408 if (can_access_pmu){
3409 val = ia64_get_pmd(cnum);
3410 } else {
3411 /*
3412 * context has been saved
3413 * if context is zombie, then task does not exist anymore.
3414 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3415 */
3416 val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3417 }
3418 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3419
3420 if (is_counting) {
3421 /*
3422 * XXX: need to check for overflow when loaded
3423 */
3424 val &= ovfl_mask;
3425 val += sval;
3426 }
3427
3428 /*
3429 * execute read checker, if any
3430 */
3431 if (unlikely(expert_mode == 0 && rd_func)) {
3432 unsigned long v = val;
3433 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3434 if (ret) goto error;
3435 val = v;
3436 ret = -EINVAL;
3437 }
3438
3439 PFM_REG_RETFLAG_SET(reg_flags, 0);
3440
3441 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3442
3443 /*
3444 * update register return value, abort all if problem during copy.
3445 * we only modify the reg_flags field. no check mode is fine because
3446 * access has been verified upfront in sys_perfmonctl().
3447 */
3448 req->reg_value = val;
3449 req->reg_flags = reg_flags;
3450 req->reg_last_reset_val = lval;
3451 }
3452
3453 return 0;
3454
3455 error:
3456 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3457 return ret;
3458 }
3459
3460 int
3461 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3462 {
3463 pfm_context_t *ctx;
3464
3465 if (req == NULL) return -EINVAL;
3466
3467 ctx = GET_PMU_CTX();
3468
3469 if (ctx == NULL) return -EINVAL;
3470
3471 /*
3472 * for now limit to current task, which is enough when calling
3473 * from overflow handler
3474 */
3475 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3476
3477 return pfm_write_pmcs(ctx, req, nreq, regs);
3478 }
3479 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3480
3481 int
3482 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3483 {
3484 pfm_context_t *ctx;
3485
3486 if (req == NULL) return -EINVAL;
3487
3488 ctx = GET_PMU_CTX();
3489
3490 if (ctx == NULL) return -EINVAL;
3491
3492 /*
3493 * for now limit to current task, which is enough when calling
3494 * from overflow handler
3495 */
3496 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3497
3498 return pfm_read_pmds(ctx, req, nreq, regs);
3499 }
3500 EXPORT_SYMBOL(pfm_mod_read_pmds);
3501
3502 /*
3503 * Only call this function when a process it trying to
3504 * write the debug registers (reading is always allowed)
3505 */
3506 int
3507 pfm_use_debug_registers(struct task_struct *task)
3508 {
3509 pfm_context_t *ctx = task->thread.pfm_context;
3510 unsigned long flags;
3511 int ret = 0;
3512
3513 if (pmu_conf->use_rr_dbregs == 0) return 0;
3514
3515 DPRINT(("called for [%d]\n", task->pid));
3516
3517 /*
3518 * do it only once
3519 */
3520 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3521
3522 /*
3523 * Even on SMP, we do not need to use an atomic here because
3524 * the only way in is via ptrace() and this is possible only when the
3525 * process is stopped. Even in the case where the ctxsw out is not totally
3526 * completed by the time we come here, there is no way the 'stopped' process
3527 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3528 * So this is always safe.
3529 */
3530 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3531
3532 LOCK_PFS(flags);
3533
3534 /*
3535 * We cannot allow setting breakpoints when system wide monitoring
3536 * sessions are using the debug registers.
3537 */
3538 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3539 ret = -1;
3540 else
3541 pfm_sessions.pfs_ptrace_use_dbregs++;
3542
3543 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3544 pfm_sessions.pfs_ptrace_use_dbregs,
3545 pfm_sessions.pfs_sys_use_dbregs,
3546 task->pid, ret));
3547
3548 UNLOCK_PFS(flags);
3549
3550 return ret;
3551 }
3552
3553 /*
3554 * This function is called for every task that exits with the
3555 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3556 * able to use the debug registers for debugging purposes via
3557 * ptrace(). Therefore we know it was not using them for
3558 * perfmormance monitoring, so we only decrement the number
3559 * of "ptraced" debug register users to keep the count up to date
3560 */
3561 int
3562 pfm_release_debug_registers(struct task_struct *task)
3563 {
3564 unsigned long flags;
3565 int ret;
3566
3567 if (pmu_conf->use_rr_dbregs == 0) return 0;
3568
3569 LOCK_PFS(flags);
3570 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3571 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3572 ret = -1;
3573 } else {
3574 pfm_sessions.pfs_ptrace_use_dbregs--;
3575 ret = 0;
3576 }
3577 UNLOCK_PFS(flags);
3578
3579 return ret;
3580 }
3581
3582 static int
3583 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3584 {
3585 struct task_struct *task;
3586 pfm_buffer_fmt_t *fmt;
3587 pfm_ovfl_ctrl_t rst_ctrl;
3588 int state, is_system;
3589 int ret = 0;
3590
3591 state = ctx->ctx_state;
3592 fmt = ctx->ctx_buf_fmt;
3593 is_system = ctx->ctx_fl_system;
3594 task = PFM_CTX_TASK(ctx);
3595
3596 switch(state) {
3597 case PFM_CTX_MASKED:
3598 break;
3599 case PFM_CTX_LOADED:
3600 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3601 /* fall through */
3602 case PFM_CTX_UNLOADED:
3603 case PFM_CTX_ZOMBIE:
3604 DPRINT(("invalid state=%d\n", state));
3605 return -EBUSY;
3606 default:
3607 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3608 return -EINVAL;
3609 }
3610
3611 /*
3612 * In system wide and when the context is loaded, access can only happen
3613 * when the caller is running on the CPU being monitored by the session.
3614 * It does not have to be the owner (ctx_task) of the context per se.
3615 */
3616 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3617 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3618 return -EBUSY;
3619 }
3620
3621 /* sanity check */
3622 if (unlikely(task == NULL)) {
3623 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3624 return -EINVAL;
3625 }
3626
3627 if (task == current || is_system) {
3628
3629 fmt = ctx->ctx_buf_fmt;
3630
3631 DPRINT(("restarting self %d ovfl=0x%lx\n",
3632 task->pid,
3633 ctx->ctx_ovfl_regs[0]));
3634
3635 if (CTX_HAS_SMPL(ctx)) {
3636
3637 prefetch(ctx->ctx_smpl_hdr);
3638
3639 rst_ctrl.bits.mask_monitoring = 0;
3640 rst_ctrl.bits.reset_ovfl_pmds = 0;
3641
3642 if (state == PFM_CTX_LOADED)
3643 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3644 else
3645 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3646 } else {
3647 rst_ctrl.bits.mask_monitoring = 0;
3648 rst_ctrl.bits.reset_ovfl_pmds = 1;
3649 }
3650
3651 if (ret == 0) {
3652 if (rst_ctrl.bits.reset_ovfl_pmds)
3653 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3654
3655 if (rst_ctrl.bits.mask_monitoring == 0) {
3656 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3657
3658 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3659 } else {
3660 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3661
3662 // cannot use pfm_stop_monitoring(task, regs);
3663 }
3664 }
3665 /*
3666 * clear overflowed PMD mask to remove any stale information
3667 */
3668 ctx->ctx_ovfl_regs[0] = 0UL;
3669
3670 /*
3671 * back to LOADED state
3672 */
3673 ctx->ctx_state = PFM_CTX_LOADED;
3674
3675 /*
3676 * XXX: not really useful for self monitoring
3677 */
3678 ctx->ctx_fl_can_restart = 0;
3679
3680 return 0;
3681 }
3682
3683 /*
3684 * restart another task
3685 */
3686
3687 /*
3688 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3689 * one is seen by the task.
3690 */
3691 if (state == PFM_CTX_MASKED) {
3692 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3693 /*
3694 * will prevent subsequent restart before this one is
3695 * seen by other task
3696 */
3697 ctx->ctx_fl_can_restart = 0;
3698 }
3699
3700 /*
3701 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3702 * the task is blocked or on its way to block. That's the normal
3703 * restart path. If the monitoring is not masked, then the task
3704 * can be actively monitoring and we cannot directly intervene.
3705 * Therefore we use the trap mechanism to catch the task and
3706 * force it to reset the buffer/reset PMDs.
3707 *
3708 * if non-blocking, then we ensure that the task will go into
3709 * pfm_handle_work() before returning to user mode.
3710 *
3711 * We cannot explicitly reset another task, it MUST always
3712 * be done by the task itself. This works for system wide because
3713 * the tool that is controlling the session is logically doing
3714 * "self-monitoring".
3715 */
3716 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3717 DPRINT(("unblocking [%d] \n", task->pid));
3718 complete(&ctx->ctx_restart_done);
3719 } else {
3720 DPRINT(("[%d] armed exit trap\n", task->pid));
3721
3722 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3723
3724 PFM_SET_WORK_PENDING(task, 1);
3725
3726 pfm_set_task_notify(task);
3727
3728 /*
3729 * XXX: send reschedule if task runs on another CPU
3730 */
3731 }
3732 return 0;
3733 }
3734
3735 static int
3736 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3737 {
3738 unsigned int m = *(unsigned int *)arg;
3739
3740 pfm_sysctl.debug = m == 0 ? 0 : 1;
3741
3742 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3743
3744 if (m == 0) {
3745 memset(pfm_stats, 0, sizeof(pfm_stats));
3746 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3747 }
3748 return 0;
3749 }
3750
3751 /*
3752 * arg can be NULL and count can be zero for this function
3753 */
3754 static int
3755 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3756 {
3757 struct thread_struct *thread = NULL;
3758 struct task_struct *task;
3759 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3760 unsigned long flags;
3761 dbreg_t dbreg;
3762 unsigned int rnum;
3763 int first_time;
3764 int ret = 0, state;
3765 int i, can_access_pmu = 0;
3766 int is_system, is_loaded;
3767
3768 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3769
3770 state = ctx->ctx_state;
3771 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3772 is_system = ctx->ctx_fl_system;
3773 task = ctx->ctx_task;
3774
3775 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3776
3777 /*
3778 * on both UP and SMP, we can only write to the PMC when the task is
3779 * the owner of the local PMU.
3780 */
3781 if (is_loaded) {
3782 thread = &task->thread;
3783 /*
3784 * In system wide and when the context is loaded, access can only happen
3785 * when the caller is running on the CPU being monitored by the session.
3786 * It does not have to be the owner (ctx_task) of the context per se.
3787 */
3788 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3789 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3790 return -EBUSY;
3791 }
3792 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3793 }
3794
3795 /*
3796 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3797 * ensuring that no real breakpoint can be installed via this call.
3798 *
3799 * IMPORTANT: regs can be NULL in this function
3800 */
3801
3802 first_time = ctx->ctx_fl_using_dbreg == 0;
3803
3804 /*
3805 * don't bother if we are loaded and task is being debugged
3806 */
3807 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3808 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3809 return -EBUSY;
3810 }
3811
3812 /*
3813 * check for debug registers in system wide mode
3814 *
3815 * If though a check is done in pfm_context_load(),
3816 * we must repeat it here, in case the registers are
3817 * written after the context is loaded
3818 */
3819 if (is_loaded) {
3820 LOCK_PFS(flags);
3821
3822 if (first_time && is_system) {
3823 if (pfm_sessions.pfs_ptrace_use_dbregs)
3824 ret = -EBUSY;
3825 else
3826 pfm_sessions.pfs_sys_use_dbregs++;
3827 }
3828 UNLOCK_PFS(flags);
3829 }
3830
3831 if (ret != 0) return ret;
3832
3833 /*
3834 * mark ourself as user of the debug registers for
3835 * perfmon purposes.
3836 */
3837 ctx->ctx_fl_using_dbreg = 1;
3838
3839 /*
3840 * clear hardware registers to make sure we don't
3841 * pick up stale state.
3842 *
3843 * for a system wide session, we do not use
3844 * thread.dbr, thread.ibr because this process
3845 * never leaves the current CPU and the state
3846 * is shared by all processes running on it
3847 */
3848 if (first_time && can_access_pmu) {
3849 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3850 for (i=0; i < pmu_conf->num_ibrs; i++) {
3851 ia64_set_ibr(i, 0UL);
3852 ia64_dv_serialize_instruction();
3853 }
3854 ia64_srlz_i();
3855 for (i=0; i < pmu_conf->num_dbrs; i++) {
3856 ia64_set_dbr(i, 0UL);
3857 ia64_dv_serialize_data();
3858 }
3859 ia64_srlz_d();
3860 }
3861
3862 /*
3863 * Now install the values into the registers
3864 */
3865 for (i = 0; i < count; i++, req++) {
3866
3867 rnum = req->dbreg_num;
3868 dbreg.val = req->dbreg_value;
3869
3870 ret = -EINVAL;
3871
3872 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3873 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3874 rnum, dbreg.val, mode, i, count));
3875
3876 goto abort_mission;
3877 }
3878
3879 /*
3880 * make sure we do not install enabled breakpoint
3881 */
3882 if (rnum & 0x1) {
3883 if (mode == PFM_CODE_RR)
3884 dbreg.ibr.ibr_x = 0;
3885 else
3886 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3887 }
3888
3889 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3890
3891 /*
3892 * Debug registers, just like PMC, can only be modified
3893 * by a kernel call. Moreover, perfmon() access to those
3894 * registers are centralized in this routine. The hardware
3895 * does not modify the value of these registers, therefore,
3896 * if we save them as they are written, we can avoid having
3897 * to save them on context switch out. This is made possible
3898 * by the fact that when perfmon uses debug registers, ptrace()
3899 * won't be able to modify them concurrently.
3900 */
3901 if (mode == PFM_CODE_RR) {
3902 CTX_USED_IBR(ctx, rnum);
3903
3904 if (can_access_pmu) {
3905 ia64_set_ibr(rnum, dbreg.val);
3906 ia64_dv_serialize_instruction();
3907 }
3908
3909 ctx->ctx_ibrs[rnum] = dbreg.val;
3910
3911 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3912 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3913 } else {
3914 CTX_USED_DBR(ctx, rnum);
3915
3916 if (can_access_pmu) {
3917 ia64_set_dbr(rnum, dbreg.val);
3918 ia64_dv_serialize_data();
3919 }
3920 ctx->ctx_dbrs[rnum] = dbreg.val;
3921
3922 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3923 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3924 }
3925 }
3926
3927 return 0;
3928
3929 abort_mission:
3930 /*
3931 * in case it was our first attempt, we undo the global modifications
3932 */
3933 if (first_time) {
3934 LOCK_PFS(flags);
3935 if (ctx->ctx_fl_system) {
3936 pfm_sessions.pfs_sys_use_dbregs--;
3937 }
3938 UNLOCK_PFS(flags);
3939 ctx->ctx_fl_using_dbreg = 0;
3940 }
3941 /*
3942 * install error return flag
3943 */
3944 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3945
3946 return ret;
3947 }
3948
3949 static int
3950 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3951 {
3952 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3953 }
3954
3955 static int
3956 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3957 {
3958 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3959 }
3960
3961 int
3962 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3963 {
3964 pfm_context_t *ctx;
3965
3966 if (req == NULL) return -EINVAL;
3967
3968 ctx = GET_PMU_CTX();
3969
3970 if (ctx == NULL) return -EINVAL;
3971
3972 /*
3973 * for now limit to current task, which is enough when calling
3974 * from overflow handler
3975 */
3976 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3977
3978 return pfm_write_ibrs(ctx, req, nreq, regs);
3979 }
3980 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3981
3982 int
3983 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3984 {
3985 pfm_context_t *ctx;
3986
3987 if (req == NULL) return -EINVAL;
3988
3989 ctx = GET_PMU_CTX();
3990
3991 if (ctx == NULL) return -EINVAL;
3992
3993 /*
3994 * for now limit to current task, which is enough when calling
3995 * from overflow handler
3996 */
3997 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3998
3999 return pfm_write_dbrs(ctx, req, nreq, regs);
4000 }
4001 EXPORT_SYMBOL(pfm_mod_write_dbrs);
4002
4003
4004 static int
4005 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4006 {
4007 pfarg_features_t *req = (pfarg_features_t *)arg;
4008
4009 req->ft_version = PFM_VERSION;
4010 return 0;
4011 }
4012
4013 static int
4014 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4015 {
4016 struct pt_regs *tregs;
4017 struct task_struct *task = PFM_CTX_TASK(ctx);
4018 int state, is_system;
4019
4020 state = ctx->ctx_state;
4021 is_system = ctx->ctx_fl_system;
4022
4023 /*
4024 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
4025 */
4026 if (state == PFM_CTX_UNLOADED) return -EINVAL;
4027
4028 /*
4029 * In system wide and when the context is loaded, access can only happen
4030 * when the caller is running on the CPU being monitored by the session.
4031 * It does not have to be the owner (ctx_task) of the context per se.
4032 */
4033 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4034 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4035 return -EBUSY;
4036 }
4037 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4038 PFM_CTX_TASK(ctx)->pid,
4039 state,
4040 is_system));
4041 /*
4042 * in system mode, we need to update the PMU directly
4043 * and the user level state of the caller, which may not
4044 * necessarily be the creator of the context.
4045 */
4046 if (is_system) {
4047 /*
4048 * Update local PMU first
4049 *
4050 * disable dcr pp
4051 */
4052 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4053 ia64_srlz_i();
4054
4055 /*
4056 * update local cpuinfo
4057 */
4058 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4059
4060 /*
4061 * stop monitoring, does srlz.i
4062 */
4063 pfm_clear_psr_pp();
4064
4065 /*
4066 * stop monitoring in the caller
4067 */
4068 ia64_psr(regs)->pp = 0;
4069
4070 return 0;
4071 }
4072 /*
4073 * per-task mode
4074 */
4075
4076 if (task == current) {
4077 /* stop monitoring at kernel level */
4078 pfm_clear_psr_up();
4079
4080 /*
4081 * stop monitoring at the user level
4082 */
4083 ia64_psr(regs)->up = 0;
4084 } else {
4085 tregs = task_pt_regs(task);
4086
4087 /*
4088 * stop monitoring at the user level
4089 */
4090 ia64_psr(tregs)->up = 0;
4091
4092 /*
4093 * monitoring disabled in kernel at next reschedule
4094 */
4095 ctx->ctx_saved_psr_up = 0;
4096 DPRINT(("task=[%d]\n", task->pid));
4097 }
4098 return 0;
4099 }
4100
4101
4102 static int
4103 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4104 {
4105 struct pt_regs *tregs;
4106 int state, is_system;
4107
4108 state = ctx->ctx_state;
4109 is_system = ctx->ctx_fl_system;
4110
4111 if (state != PFM_CTX_LOADED) return -EINVAL;
4112
4113 /*
4114 * In system wide and when the context is loaded, access can only happen
4115 * when the caller is running on the CPU being monitored by the session.
4116 * It does not have to be the owner (ctx_task) of the context per se.
4117 */
4118 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4119 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4120 return -EBUSY;
4121 }
4122
4123 /*
4124 * in system mode, we need to update the PMU directly
4125 * and the user level state of the caller, which may not
4126 * necessarily be the creator of the context.
4127 */
4128 if (is_system) {
4129
4130 /*
4131 * set user level psr.pp for the caller
4132 */
4133 ia64_psr(regs)->pp = 1;
4134
4135 /*
4136 * now update the local PMU and cpuinfo
4137 */
4138 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4139
4140 /*
4141 * start monitoring at kernel level
4142 */
4143 pfm_set_psr_pp();
4144
4145 /* enable dcr pp */
4146 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4147 ia64_srlz_i();
4148
4149 return 0;
4150 }
4151
4152 /*
4153 * per-process mode
4154 */
4155
4156 if (ctx->ctx_task == current) {
4157
4158 /* start monitoring at kernel level */
4159 pfm_set_psr_up();
4160
4161 /*
4162 * activate monitoring at user level
4163 */
4164 ia64_psr(regs)->up = 1;
4165
4166 } else {
4167 tregs = task_pt_regs(ctx->ctx_task);
4168
4169 /*
4170 * start monitoring at the kernel level the next
4171 * time the task is scheduled
4172 */
4173 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4174
4175 /*
4176 * activate monitoring at user level
4177 */
4178 ia64_psr(tregs)->up = 1;
4179 }
4180 return 0;
4181 }
4182
4183 static int
4184 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4185 {
4186 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4187 unsigned int cnum;
4188 int i;
4189 int ret = -EINVAL;
4190
4191 for (i = 0; i < count; i++, req++) {
4192
4193 cnum = req->reg_num;
4194
4195 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4196
4197 req->reg_value = PMC_DFL_VAL(cnum);
4198
4199 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4200
4201 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4202 }
4203 return 0;
4204
4205 abort_mission:
4206 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4207 return ret;
4208 }
4209
4210 static int
4211 pfm_check_task_exist(pfm_context_t *ctx)
4212 {
4213 struct task_struct *g, *t;
4214 int ret = -ESRCH;
4215
4216 read_lock(&tasklist_lock);
4217
4218 do_each_thread (g, t) {
4219 if (t->thread.pfm_context == ctx) {
4220 ret = 0;
4221 break;
4222 }
4223 } while_each_thread (g, t);
4224
4225 read_unlock(&tasklist_lock);
4226
4227 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4228
4229 return ret;
4230 }
4231
4232 static int
4233 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4234 {
4235 struct task_struct *task;
4236 struct thread_struct *thread;
4237 struct pfm_context_t *old;
4238 unsigned long flags;
4239 #ifndef CONFIG_SMP
4240 struct task_struct *owner_task = NULL;
4241 #endif
4242 pfarg_load_t *req = (pfarg_load_t *)arg;
4243 unsigned long *pmcs_source, *pmds_source;
4244 int the_cpu;
4245 int ret = 0;
4246 int state, is_system, set_dbregs = 0;
4247
4248 state = ctx->ctx_state;
4249 is_system = ctx->ctx_fl_system;
4250 /*
4251 * can only load from unloaded or terminated state
4252 */
4253 if (state != PFM_CTX_UNLOADED) {
4254 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4255 req->load_pid,
4256 ctx->ctx_state));
4257 return -EBUSY;
4258 }
4259
4260 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4261
4262 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4263 DPRINT(("cannot use blocking mode on self\n"));
4264 return -EINVAL;
4265 }
4266
4267 ret = pfm_get_task(ctx, req->load_pid, &task);
4268 if (ret) {
4269 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4270 return ret;
4271 }
4272
4273 ret = -EINVAL;
4274
4275 /*
4276 * system wide is self monitoring only
4277 */
4278 if (is_system && task != current) {
4279 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4280 req->load_pid));
4281 goto error;
4282 }
4283
4284 thread = &task->thread;
4285
4286 ret = 0;
4287 /*
4288 * cannot load a context which is using range restrictions,
4289 * into a task that is being debugged.
4290 */
4291 if (ctx->ctx_fl_using_dbreg) {
4292 if (thread->flags & IA64_THREAD_DBG_VALID) {
4293 ret = -EBUSY;
4294 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4295 goto error;
4296 }
4297 LOCK_PFS(flags);
4298
4299 if (is_system) {
4300 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4301 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4302 ret = -EBUSY;
4303 } else {
4304 pfm_sessions.pfs_sys_use_dbregs++;
4305 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4306 set_dbregs = 1;
4307 }
4308 }
4309
4310 UNLOCK_PFS(flags);
4311
4312 if (ret) goto error;
4313 }
4314
4315 /*
4316 * SMP system-wide monitoring implies self-monitoring.
4317 *
4318 * The programming model expects the task to
4319 * be pinned on a CPU throughout the session.
4320 * Here we take note of the current CPU at the
4321 * time the context is loaded. No call from
4322 * another CPU will be allowed.
4323 *
4324 * The pinning via shed_setaffinity()
4325 * must be done by the calling task prior
4326 * to this call.
4327 *
4328 * systemwide: keep track of CPU this session is supposed to run on
4329 */
4330 the_cpu = ctx->ctx_cpu = smp_processor_id();
4331
4332 ret = -EBUSY;
4333 /*
4334 * now reserve the session
4335 */
4336 ret = pfm_reserve_session(current, is_system, the_cpu);
4337 if (ret) goto error;
4338
4339 /*
4340 * task is necessarily stopped at this point.
4341 *
4342 * If the previous context was zombie, then it got removed in
4343 * pfm_save_regs(). Therefore we should not see it here.
4344 * If we see a context, then this is an active context
4345 *
4346 * XXX: needs to be atomic
4347 */
4348 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4349 thread->pfm_context, ctx));
4350
4351 ret = -EBUSY;
4352 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4353 if (old != NULL) {
4354 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4355 goto error_unres;
4356 }
4357
4358 pfm_reset_msgq(ctx);
4359
4360 ctx->ctx_state = PFM_CTX_LOADED;
4361
4362 /*
4363 * link context to task
4364 */
4365 ctx->ctx_task = task;
4366
4367 if (is_system) {
4368 /*
4369 * we load as stopped
4370 */
4371 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4372 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4373
4374 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4375 } else {
4376 thread->flags |= IA64_THREAD_PM_VALID;
4377 }
4378
4379 /*
4380 * propagate into thread-state
4381 */
4382 pfm_copy_pmds(task, ctx);
4383 pfm_copy_pmcs(task, ctx);
4384
4385 pmcs_source = ctx->th_pmcs;
4386 pmds_source = ctx->th_pmds;
4387
4388 /*
4389 * always the case for system-wide
4390 */
4391 if (task == current) {
4392
4393 if (is_system == 0) {
4394
4395 /* allow user level control */
4396 ia64_psr(regs)->sp = 0;
4397 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4398
4399 SET_LAST_CPU(ctx, smp_processor_id());
4400 INC_ACTIVATION();
4401 SET_ACTIVATION(ctx);
4402 #ifndef CONFIG_SMP
4403 /*
4404 * push the other task out, if any
4405 */
4406 owner_task = GET_PMU_OWNER();
4407 if (owner_task) pfm_lazy_save_regs(owner_task);
4408 #endif
4409 }
4410 /*
4411 * load all PMD from ctx to PMU (as opposed to thread state)
4412 * restore all PMC from ctx to PMU
4413 */
4414 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4415 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4416
4417 ctx->ctx_reload_pmcs[0] = 0UL;
4418 ctx->ctx_reload_pmds[0] = 0UL;
4419
4420 /*
4421 * guaranteed safe by earlier check against DBG_VALID
4422 */
4423 if (ctx->ctx_fl_using_dbreg) {
4424 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4425 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4426 }
4427 /*
4428 * set new ownership
4429 */
4430 SET_PMU_OWNER(task, ctx);
4431
4432 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4433 } else {
4434 /*
4435 * when not current, task MUST be stopped, so this is safe
4436 */
4437 regs = task_pt_regs(task);
4438
4439 /* force a full reload */
4440 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4441 SET_LAST_CPU(ctx, -1);
4442
4443 /* initial saved psr (stopped) */
4444 ctx->ctx_saved_psr_up = 0UL;
4445 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4446 }
4447
4448 ret = 0;
4449
4450 error_unres:
4451 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4452 error:
4453 /*
4454 * we must undo the dbregs setting (for system-wide)
4455 */
4456 if (ret && set_dbregs) {
4457 LOCK_PFS(flags);
4458 pfm_sessions.pfs_sys_use_dbregs--;
4459 UNLOCK_PFS(flags);
4460 }
4461 /*
4462 * release task, there is now a link with the context
4463 */
4464 if (is_system == 0 && task != current) {
4465 pfm_put_task(task);
4466
4467 if (ret == 0) {
4468 ret = pfm_check_task_exist(ctx);
4469 if (ret) {
4470 ctx->ctx_state = PFM_CTX_UNLOADED;
4471 ctx->ctx_task = NULL;
4472 }
4473 }
4474 }
4475 return ret;
4476 }
4477
4478 /*
4479 * in this function, we do not need to increase the use count
4480 * for the task via get_task_struct(), because we hold the
4481 * context lock. If the task were to disappear while having
4482 * a context attached, it would go through pfm_exit_thread()
4483 * which also grabs the context lock and would therefore be blocked
4484 * until we are here.
4485 */
4486 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4487
4488 static int
4489 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4490 {
4491 struct task_struct *task = PFM_CTX_TASK(ctx);
4492 struct pt_regs *tregs;
4493 int prev_state, is_system;
4494 int ret;
4495
4496 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4497
4498 prev_state = ctx->ctx_state;
4499 is_system = ctx->ctx_fl_system;
4500
4501 /*
4502 * unload only when necessary
4503 */
4504 if (prev_state == PFM_CTX_UNLOADED) {
4505 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4506 return 0;
4507 }
4508
4509 /*
4510 * clear psr and dcr bits
4511 */
4512 ret = pfm_stop(ctx, NULL, 0, regs);
4513 if (ret) return ret;
4514
4515 ctx->ctx_state = PFM_CTX_UNLOADED;
4516
4517 /*
4518 * in system mode, we need to update the PMU directly
4519 * and the user level state of the caller, which may not
4520 * necessarily be the creator of the context.
4521 */
4522 if (is_system) {
4523
4524 /*
4525 * Update cpuinfo
4526 *
4527 * local PMU is taken care of in pfm_stop()
4528 */
4529 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4530 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4531
4532 /*
4533 * save PMDs in context
4534 * release ownership
4535 */
4536 pfm_flush_pmds(current, ctx);
4537
4538 /*
4539 * at this point we are done with the PMU
4540 * so we can unreserve the resource.
4541 */
4542 if (prev_state != PFM_CTX_ZOMBIE)
4543 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4544
4545 /*
4546 * disconnect context from task
4547 */
4548 task->thread.pfm_context = NULL;
4549 /*
4550 * disconnect task from context
4551 */
4552 ctx->ctx_task = NULL;
4553
4554 /*
4555 * There is nothing more to cleanup here.
4556 */
4557 return 0;
4558 }
4559
4560 /*
4561 * per-task mode
4562 */
4563 tregs = task == current ? regs : task_pt_regs(task);
4564
4565 if (task == current) {
4566 /*
4567 * cancel user level control
4568 */
4569 ia64_psr(regs)->sp = 1;
4570
4571 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4572 }
4573 /*
4574 * save PMDs to context
4575 * release ownership
4576 */
4577 pfm_flush_pmds(task, ctx);
4578
4579 /*
4580 * at this point we are done with the PMU
4581 * so we can unreserve the resource.
4582 *
4583 * when state was ZOMBIE, we have already unreserved.
4584 */
4585 if (prev_state != PFM_CTX_ZOMBIE)
4586 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4587
4588 /*
4589 * reset activation counter and psr
4590 */
4591 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4592 SET_LAST_CPU(ctx, -1);
4593
4594 /*
4595 * PMU state will not be restored
4596 */
4597 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4598
4599 /*
4600 * break links between context and task
4601 */
4602 task->thread.pfm_context = NULL;
4603 ctx->ctx_task = NULL;
4604
4605 PFM_SET_WORK_PENDING(task, 0);
4606
4607 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4608 ctx->ctx_fl_can_restart = 0;
4609 ctx->ctx_fl_going_zombie = 0;
4610
4611 DPRINT(("disconnected [%d] from context\n", task->pid));
4612
4613 return 0;
4614 }
4615
4616
4617 /*
4618 * called only from exit_thread(): task == current
4619 * we come here only if current has a context attached (loaded or masked)
4620 */
4621 void
4622 pfm_exit_thread(struct task_struct *task)
4623 {
4624 pfm_context_t *ctx;
4625 unsigned long flags;
4626 struct pt_regs *regs = task_pt_regs(task);
4627 int ret, state;
4628 int free_ok = 0;
4629
4630 ctx = PFM_GET_CTX(task);
4631
4632 PROTECT_CTX(ctx, flags);
4633
4634 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4635
4636 state = ctx->ctx_state;
4637 switch(state) {
4638 case PFM_CTX_UNLOADED:
4639 /*
4640 * only comes to this function if pfm_context is not NULL, i.e., cannot
4641 * be in unloaded state
4642 */
4643 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4644 break;
4645 case PFM_CTX_LOADED:
4646 case PFM_CTX_MASKED:
4647 ret = pfm_context_unload(ctx, NULL, 0, regs);
4648 if (ret) {
4649 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4650 }
4651 DPRINT(("ctx unloaded for current state was %d\n", state));
4652
4653 pfm_end_notify_user(ctx);
4654 break;
4655 case PFM_CTX_ZOMBIE:
4656 ret = pfm_context_unload(ctx, NULL, 0, regs);
4657 if (ret) {
4658 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4659 }
4660 free_ok = 1;
4661 break;
4662 default:
4663 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4664 break;
4665 }
4666 UNPROTECT_CTX(ctx, flags);
4667
4668 { u64 psr = pfm_get_psr();
4669 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4670 BUG_ON(GET_PMU_OWNER());
4671 BUG_ON(ia64_psr(regs)->up);
4672 BUG_ON(ia64_psr(regs)->pp);
4673 }
4674
4675 /*
4676 * All memory free operations (especially for vmalloc'ed memory)
4677 * MUST be done with interrupts ENABLED.
4678 */
4679 if (free_ok) pfm_context_free(ctx);
4680 }
4681
4682 /*
4683 * functions MUST be listed in the increasing order of their index (see permfon.h)
4684 */
4685 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4686 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4687 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4688 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4689 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4690
4691 static pfm_cmd_desc_t pfm_cmd_tab[]={
4692 /* 0 */PFM_CMD_NONE,
4693 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4694 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4695 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4696 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4697 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4698 /* 6 */PFM_CMD_NONE,
4699 /* 7 */PFM_CMD_NONE,
4700 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4701 /* 9 */PFM_CMD_NONE,
4702 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4703 /* 11 */PFM_CMD_NONE,
4704 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4705 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4706 /* 14 */PFM_CMD_NONE,
4707 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4708 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4709 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4710 /* 18 */PFM_CMD_NONE,
4711 /* 19 */PFM_CMD_NONE,
4712 /* 20 */PFM_CMD_NONE,
4713 /* 21 */PFM_CMD_NONE,
4714 /* 22 */PFM_CMD_NONE,
4715 /* 23 */PFM_CMD_NONE,
4716 /* 24 */PFM_CMD_NONE,
4717 /* 25 */PFM_CMD_NONE,
4718 /* 26 */PFM_CMD_NONE,
4719 /* 27 */PFM_CMD_NONE,
4720 /* 28 */PFM_CMD_NONE,
4721 /* 29 */PFM_CMD_NONE,
4722 /* 30 */PFM_CMD_NONE,
4723 /* 31 */PFM_CMD_NONE,
4724 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4725 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4726 };
4727 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4728
4729 static int
4730 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4731 {
4732 struct task_struct *task;
4733 int state, old_state;
4734
4735 recheck:
4736 state = ctx->ctx_state;
4737 task = ctx->ctx_task;
4738
4739 if (task == NULL) {
4740 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4741 return 0;
4742 }
4743
4744 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4745 ctx->ctx_fd,
4746 state,
4747 task->pid,
4748 task->state, PFM_CMD_STOPPED(cmd)));
4749
4750 /*
4751 * self-monitoring always ok.
4752 *
4753 * for system-wide the caller can either be the creator of the
4754 * context (to one to which the context is attached to) OR
4755 * a task running on the same CPU as the session.
4756 */
4757 if (task == current || ctx->ctx_fl_system) return 0;
4758
4759 /*
4760 * we are monitoring another thread
4761 */
4762 switch(state) {
4763 case PFM_CTX_UNLOADED:
4764 /*
4765 * if context is UNLOADED we are safe to go
4766 */
4767 return 0;
4768 case PFM_CTX_ZOMBIE:
4769 /*
4770 * no command can operate on a zombie context
4771 */
4772 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4773 return -EINVAL;
4774 case PFM_CTX_MASKED:
4775 /*
4776 * PMU state has been saved to software even though
4777 * the thread may still be running.
4778 */
4779 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4780 }
4781
4782 /*
4783 * context is LOADED or MASKED. Some commands may need to have
4784 * the task stopped.
4785 *
4786 * We could lift this restriction for UP but it would mean that
4787 * the user has no guarantee the task would not run between
4788 * two successive calls to perfmonctl(). That's probably OK.
4789 * If this user wants to ensure the task does not run, then
4790 * the task must be stopped.
4791 */
4792 if (PFM_CMD_STOPPED(cmd)) {
4793 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
4794 DPRINT(("[%d] task not in stopped state\n", task->pid));
4795 return -EBUSY;
4796 }
4797 /*
4798 * task is now stopped, wait for ctxsw out
4799 *
4800 * This is an interesting point in the code.
4801 * We need to unprotect the context because
4802 * the pfm_save_regs() routines needs to grab
4803 * the same lock. There are danger in doing
4804 * this because it leaves a window open for
4805 * another task to get access to the context
4806 * and possibly change its state. The one thing
4807 * that is not possible is for the context to disappear
4808 * because we are protected by the VFS layer, i.e.,
4809 * get_fd()/put_fd().
4810 */
4811 old_state = state;
4812
4813 UNPROTECT_CTX(ctx, flags);
4814
4815 wait_task_inactive(task);
4816
4817 PROTECT_CTX(ctx, flags);
4818
4819 /*
4820 * we must recheck to verify if state has changed
4821 */
4822 if (ctx->ctx_state != old_state) {
4823 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4824 goto recheck;
4825 }
4826 }
4827 return 0;
4828 }
4829
4830 /*
4831 * system-call entry point (must return long)
4832 */
4833 asmlinkage long
4834 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4835 {
4836 struct file *file = NULL;
4837 pfm_context_t *ctx = NULL;
4838 unsigned long flags = 0UL;
4839 void *args_k = NULL;
4840 long ret; /* will expand int return types */
4841 size_t base_sz, sz, xtra_sz = 0;
4842 int narg, completed_args = 0, call_made = 0, cmd_flags;
4843 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4844 int (*getsize)(void *arg, size_t *sz);
4845 #define PFM_MAX_ARGSIZE 4096
4846
4847 /*
4848 * reject any call if perfmon was disabled at initialization
4849 */
4850 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4851
4852 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4853 DPRINT(("invalid cmd=%d\n", cmd));
4854 return -EINVAL;
4855 }
4856
4857 func = pfm_cmd_tab[cmd].cmd_func;
4858 narg = pfm_cmd_tab[cmd].cmd_narg;
4859 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4860 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4861 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4862
4863 if (unlikely(func == NULL)) {
4864 DPRINT(("invalid cmd=%d\n", cmd));
4865 return -EINVAL;
4866 }
4867
4868 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4869 PFM_CMD_NAME(cmd),
4870 cmd,
4871 narg,
4872 base_sz,
4873 count));
4874
4875 /*
4876 * check if number of arguments matches what the command expects
4877 */
4878 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4879 return -EINVAL;
4880
4881 restart_args:
4882 sz = xtra_sz + base_sz*count;
4883 /*
4884 * limit abuse to min page size
4885 */
4886 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4887 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4888 return -E2BIG;
4889 }
4890
4891 /*
4892 * allocate default-sized argument buffer
4893 */
4894 if (likely(count && args_k == NULL)) {
4895 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4896 if (args_k == NULL) return -ENOMEM;
4897 }
4898
4899 ret = -EFAULT;
4900
4901 /*
4902 * copy arguments
4903 *
4904 * assume sz = 0 for command without parameters
4905 */
4906 if (sz && copy_from_user(args_k, arg, sz)) {
4907 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4908 goto error_args;
4909 }
4910
4911 /*
4912 * check if command supports extra parameters
4913 */
4914 if (completed_args == 0 && getsize) {
4915 /*
4916 * get extra parameters size (based on main argument)
4917 */
4918 ret = (*getsize)(args_k, &xtra_sz);
4919 if (ret) goto error_args;
4920
4921 completed_args = 1;
4922
4923 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4924
4925 /* retry if necessary */
4926 if (likely(xtra_sz)) goto restart_args;
4927 }
4928
4929 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4930
4931 ret = -EBADF;
4932
4933 file = fget(fd);
4934 if (unlikely(file == NULL)) {
4935 DPRINT(("invalid fd %d\n", fd));
4936 goto error_args;
4937 }
4938 if (unlikely(PFM_IS_FILE(file) == 0)) {
4939 DPRINT(("fd %d not related to perfmon\n", fd));
4940 goto error_args;
4941 }
4942
4943 ctx = (pfm_context_t *)file->private_data;
4944 if (unlikely(ctx == NULL)) {
4945 DPRINT(("no context for fd %d\n", fd));
4946 goto error_args;
4947 }
4948 prefetch(&ctx->ctx_state);
4949
4950 PROTECT_CTX(ctx, flags);
4951
4952 /*
4953 * check task is stopped
4954 */
4955 ret = pfm_check_task_state(ctx, cmd, flags);
4956 if (unlikely(ret)) goto abort_locked;
4957
4958 skip_fd:
4959 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4960
4961 call_made = 1;
4962
4963 abort_locked:
4964 if (likely(ctx)) {
4965 DPRINT(("context unlocked\n"));
4966 UNPROTECT_CTX(ctx, flags);
4967 }
4968
4969 /* copy argument back to user, if needed */
4970 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4971
4972 error_args:
4973 if (file)
4974 fput(file);
4975
4976 kfree(args_k);
4977
4978 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4979
4980 return ret;
4981 }
4982
4983 static void
4984 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4985 {
4986 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4987 pfm_ovfl_ctrl_t rst_ctrl;
4988 int state;
4989 int ret = 0;
4990
4991 state = ctx->ctx_state;
4992 /*
4993 * Unlock sampling buffer and reset index atomically
4994 * XXX: not really needed when blocking
4995 */
4996 if (CTX_HAS_SMPL(ctx)) {
4997
4998 rst_ctrl.bits.mask_monitoring = 0;
4999 rst_ctrl.bits.reset_ovfl_pmds = 0;
5000
5001 if (state == PFM_CTX_LOADED)
5002 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
5003 else
5004 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
5005 } else {
5006 rst_ctrl.bits.mask_monitoring = 0;
5007 rst_ctrl.bits.reset_ovfl_pmds = 1;
5008 }
5009
5010 if (ret == 0) {
5011 if (rst_ctrl.bits.reset_ovfl_pmds) {
5012 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
5013 }
5014 if (rst_ctrl.bits.mask_monitoring == 0) {
5015 DPRINT(("resuming monitoring\n"));
5016 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
5017 } else {
5018 DPRINT(("stopping monitoring\n"));
5019 //pfm_stop_monitoring(current, regs);
5020 }
5021 ctx->ctx_state = PFM_CTX_LOADED;
5022 }
5023 }
5024
5025 /*
5026 * context MUST BE LOCKED when calling
5027 * can only be called for current
5028 */
5029 static void
5030 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
5031 {
5032 int ret;
5033
5034 DPRINT(("entering for [%d]\n", current->pid));
5035
5036 ret = pfm_context_unload(ctx, NULL, 0, regs);
5037 if (ret) {
5038 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", current->pid, ret);
5039 }
5040
5041 /*
5042 * and wakeup controlling task, indicating we are now disconnected
5043 */
5044 wake_up_interruptible(&ctx->ctx_zombieq);
5045
5046 /*
5047 * given that context is still locked, the controlling
5048 * task will only get access when we return from
5049 * pfm_handle_work().
5050 */
5051 }
5052
5053 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5054 /*
5055 * pfm_handle_work() can be called with interrupts enabled
5056 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5057 * call may sleep, therefore we must re-enable interrupts
5058 * to avoid deadlocks. It is safe to do so because this function
5059 * is called ONLY when returning to user level (PUStk=1), in which case
5060 * there is no risk of kernel stack overflow due to deep
5061 * interrupt nesting.
5062 */
5063 void
5064 pfm_handle_work(void)
5065 {
5066 pfm_context_t *ctx;
5067 struct pt_regs *regs;
5068 unsigned long flags, dummy_flags;
5069 unsigned long ovfl_regs;
5070 unsigned int reason;
5071 int ret;
5072
5073 ctx = PFM_GET_CTX(current);
5074 if (ctx == NULL) {
5075 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5076 return;
5077 }
5078
5079 PROTECT_CTX(ctx, flags);
5080
5081 PFM_SET_WORK_PENDING(current, 0);
5082
5083 pfm_clear_task_notify();
5084
5085 regs = task_pt_regs(current);
5086
5087 /*
5088 * extract reason for being here and clear
5089 */
5090 reason = ctx->ctx_fl_trap_reason;
5091 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5092 ovfl_regs = ctx->ctx_ovfl_regs[0];
5093
5094 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5095
5096 /*
5097 * must be done before we check for simple-reset mode
5098 */
5099 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5100
5101
5102 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5103 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5104
5105 /*
5106 * restore interrupt mask to what it was on entry.
5107 * Could be enabled/diasbled.
5108 */
5109 UNPROTECT_CTX(ctx, flags);
5110
5111 /*
5112 * force interrupt enable because of down_interruptible()
5113 */
5114 local_irq_enable();
5115
5116 DPRINT(("before block sleeping\n"));
5117
5118 /*
5119 * may go through without blocking on SMP systems
5120 * if restart has been received already by the time we call down()
5121 */
5122 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5123
5124 DPRINT(("after block sleeping ret=%d\n", ret));
5125
5126 /*
5127 * lock context and mask interrupts again
5128 * We save flags into a dummy because we may have
5129 * altered interrupts mask compared to entry in this
5130 * function.
5131 */
5132 PROTECT_CTX(ctx, dummy_flags);
5133
5134 /*
5135 * we need to read the ovfl_regs only after wake-up
5136 * because we may have had pfm_write_pmds() in between
5137 * and that can changed PMD values and therefore
5138 * ovfl_regs is reset for these new PMD values.
5139 */
5140 ovfl_regs = ctx->ctx_ovfl_regs[0];
5141
5142 if (ctx->ctx_fl_going_zombie) {
5143 do_zombie:
5144 DPRINT(("context is zombie, bailing out\n"));
5145 pfm_context_force_terminate(ctx, regs);
5146 goto nothing_to_do;
5147 }
5148 /*
5149 * in case of interruption of down() we don't restart anything
5150 */
5151 if (ret < 0) goto nothing_to_do;
5152
5153 skip_blocking:
5154 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5155 ctx->ctx_ovfl_regs[0] = 0UL;
5156
5157 nothing_to_do:
5158 /*
5159 * restore flags as they were upon entry
5160 */
5161 UNPROTECT_CTX(ctx, flags);
5162 }
5163
5164 static int
5165 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5166 {
5167 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5168 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5169 return 0;
5170 }
5171
5172 DPRINT(("waking up somebody\n"));
5173
5174 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5175
5176 /*
5177 * safe, we are not in intr handler, nor in ctxsw when
5178 * we come here
5179 */
5180 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5181
5182 return 0;
5183 }
5184
5185 static int
5186 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5187 {
5188 pfm_msg_t *msg = NULL;
5189
5190 if (ctx->ctx_fl_no_msg == 0) {
5191 msg = pfm_get_new_msg(ctx);
5192 if (msg == NULL) {
5193 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5194 return -1;
5195 }
5196
5197 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5198 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5199 msg->pfm_ovfl_msg.msg_active_set = 0;
5200 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5201 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5202 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5203 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5204 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5205 }
5206
5207 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5208 msg,
5209 ctx->ctx_fl_no_msg,
5210 ctx->ctx_fd,
5211 ovfl_pmds));
5212
5213 return pfm_notify_user(ctx, msg);
5214 }
5215
5216 static int
5217 pfm_end_notify_user(pfm_context_t *ctx)
5218 {
5219 pfm_msg_t *msg;
5220
5221 msg = pfm_get_new_msg(ctx);
5222 if (msg == NULL) {
5223 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5224 return -1;
5225 }
5226 /* no leak */
5227 memset(msg, 0, sizeof(*msg));
5228
5229 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5230 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5231 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5232
5233 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5234 msg,
5235 ctx->ctx_fl_no_msg,
5236 ctx->ctx_fd));
5237
5238 return pfm_notify_user(ctx, msg);
5239 }
5240
5241 /*
5242 * main overflow processing routine.
5243 * it can be called from the interrupt path or explicitly during the context switch code
5244 */
5245 static void
5246 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5247 {
5248 pfm_ovfl_arg_t *ovfl_arg;
5249 unsigned long mask;
5250 unsigned long old_val, ovfl_val, new_val;
5251 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5252 unsigned long tstamp;
5253 pfm_ovfl_ctrl_t ovfl_ctrl;
5254 unsigned int i, has_smpl;
5255 int must_notify = 0;
5256
5257 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5258
5259 /*
5260 * sanity test. Should never happen
5261 */
5262 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5263
5264 tstamp = ia64_get_itc();
5265 mask = pmc0 >> PMU_FIRST_COUNTER;
5266 ovfl_val = pmu_conf->ovfl_val;
5267 has_smpl = CTX_HAS_SMPL(ctx);
5268
5269 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5270 "used_pmds=0x%lx\n",
5271 pmc0,
5272 task ? task->pid: -1,
5273 (regs ? regs->cr_iip : 0),
5274 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5275 ctx->ctx_used_pmds[0]));
5276
5277
5278 /*
5279 * first we update the virtual counters
5280 * assume there was a prior ia64_srlz_d() issued
5281 */
5282 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5283
5284 /* skip pmd which did not overflow */
5285 if ((mask & 0x1) == 0) continue;
5286
5287 /*
5288 * Note that the pmd is not necessarily 0 at this point as qualified events
5289 * may have happened before the PMU was frozen. The residual count is not
5290 * taken into consideration here but will be with any read of the pmd via
5291 * pfm_read_pmds().
5292 */
5293 old_val = new_val = ctx->ctx_pmds[i].val;
5294 new_val += 1 + ovfl_val;
5295 ctx->ctx_pmds[i].val = new_val;
5296
5297 /*
5298 * check for overflow condition
5299 */
5300 if (likely(old_val > new_val)) {
5301 ovfl_pmds |= 1UL << i;
5302 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5303 }
5304
5305 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5306 i,
5307 new_val,
5308 old_val,
5309 ia64_get_pmd(i) & ovfl_val,
5310 ovfl_pmds,
5311 ovfl_notify));
5312 }
5313
5314 /*
5315 * there was no 64-bit overflow, nothing else to do
5316 */
5317 if (ovfl_pmds == 0UL) return;
5318
5319 /*
5320 * reset all control bits
5321 */
5322 ovfl_ctrl.val = 0;
5323 reset_pmds = 0UL;
5324
5325 /*
5326 * if a sampling format module exists, then we "cache" the overflow by
5327 * calling the module's handler() routine.
5328 */
5329 if (has_smpl) {
5330 unsigned long start_cycles, end_cycles;
5331 unsigned long pmd_mask;
5332 int j, k, ret = 0;
5333 int this_cpu = smp_processor_id();
5334
5335 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5336 ovfl_arg = &ctx->ctx_ovfl_arg;
5337
5338 prefetch(ctx->ctx_smpl_hdr);
5339
5340 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5341
5342 mask = 1UL << i;
5343
5344 if ((pmd_mask & 0x1) == 0) continue;
5345
5346 ovfl_arg->ovfl_pmd = (unsigned char )i;
5347 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5348 ovfl_arg->active_set = 0;
5349 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5350 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5351
5352 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5353 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5354 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5355
5356 /*
5357 * copy values of pmds of interest. Sampling format may copy them
5358 * into sampling buffer.
5359 */
5360 if (smpl_pmds) {
5361 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5362 if ((smpl_pmds & 0x1) == 0) continue;
5363 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5364 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5365 }
5366 }
5367
5368 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5369
5370 start_cycles = ia64_get_itc();
5371
5372 /*
5373 * call custom buffer format record (handler) routine
5374 */
5375 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5376
5377 end_cycles = ia64_get_itc();
5378
5379 /*
5380 * For those controls, we take the union because they have
5381 * an all or nothing behavior.
5382 */
5383 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5384 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5385 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5386 /*
5387 * build the bitmask of pmds to reset now
5388 */
5389 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5390
5391 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5392 }
5393 /*
5394 * when the module cannot handle the rest of the overflows, we abort right here
5395 */
5396 if (ret && pmd_mask) {
5397 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5398 pmd_mask<<PMU_FIRST_COUNTER));
5399 }
5400 /*
5401 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5402 */
5403 ovfl_pmds &= ~reset_pmds;
5404 } else {
5405 /*
5406 * when no sampling module is used, then the default
5407 * is to notify on overflow if requested by user
5408 */
5409 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5410 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5411 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5412 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5413 /*
5414 * if needed, we reset all overflowed pmds
5415 */
5416 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5417 }
5418
5419 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5420
5421 /*
5422 * reset the requested PMD registers using the short reset values
5423 */
5424 if (reset_pmds) {
5425 unsigned long bm = reset_pmds;
5426 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5427 }
5428
5429 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5430 /*
5431 * keep track of what to reset when unblocking
5432 */
5433 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5434
5435 /*
5436 * check for blocking context
5437 */
5438 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5439
5440 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5441
5442 /*
5443 * set the perfmon specific checking pending work for the task
5444 */
5445 PFM_SET_WORK_PENDING(task, 1);
5446
5447 /*
5448 * when coming from ctxsw, current still points to the
5449 * previous task, therefore we must work with task and not current.
5450 */
5451 pfm_set_task_notify(task);
5452 }
5453 /*
5454 * defer until state is changed (shorten spin window). the context is locked
5455 * anyway, so the signal receiver would come spin for nothing.
5456 */
5457 must_notify = 1;
5458 }
5459
5460 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5461 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5462 PFM_GET_WORK_PENDING(task),
5463 ctx->ctx_fl_trap_reason,
5464 ovfl_pmds,
5465 ovfl_notify,
5466 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5467 /*
5468 * in case monitoring must be stopped, we toggle the psr bits
5469 */
5470 if (ovfl_ctrl.bits.mask_monitoring) {
5471 pfm_mask_monitoring(task);
5472 ctx->ctx_state = PFM_CTX_MASKED;
5473 ctx->ctx_fl_can_restart = 1;
5474 }
5475
5476 /*
5477 * send notification now
5478 */
5479 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5480
5481 return;
5482
5483 sanity_check:
5484 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5485 smp_processor_id(),
5486 task ? task->pid : -1,
5487 pmc0);
5488 return;
5489
5490 stop_monitoring:
5491 /*
5492 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5493 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5494 * come here as zombie only if the task is the current task. In which case, we
5495 * can access the PMU hardware directly.
5496 *
5497 * Note that zombies do have PM_VALID set. So here we do the minimal.
5498 *
5499 * In case the context was zombified it could not be reclaimed at the time
5500 * the monitoring program exited. At this point, the PMU reservation has been
5501 * returned, the sampiing buffer has been freed. We must convert this call
5502 * into a spurious interrupt. However, we must also avoid infinite overflows
5503 * by stopping monitoring for this task. We can only come here for a per-task
5504 * context. All we need to do is to stop monitoring using the psr bits which
5505 * are always task private. By re-enabling secure montioring, we ensure that
5506 * the monitored task will not be able to re-activate monitoring.
5507 * The task will eventually be context switched out, at which point the context
5508 * will be reclaimed (that includes releasing ownership of the PMU).
5509 *
5510 * So there might be a window of time where the number of per-task session is zero
5511 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5512 * context. This is safe because if a per-task session comes in, it will push this one
5513 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5514 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5515 * also push our zombie context out.
5516 *
5517 * Overall pretty hairy stuff....
5518 */
5519 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5520 pfm_clear_psr_up();
5521 ia64_psr(regs)->up = 0;
5522 ia64_psr(regs)->sp = 1;
5523 return;
5524 }
5525
5526 static int
5527 pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5528 {
5529 struct task_struct *task;
5530 pfm_context_t *ctx;
5531 unsigned long flags;
5532 u64 pmc0;
5533 int this_cpu = smp_processor_id();
5534 int retval = 0;
5535
5536 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5537
5538 /*
5539 * srlz.d done before arriving here
5540 */
5541 pmc0 = ia64_get_pmc(0);
5542
5543 task = GET_PMU_OWNER();
5544 ctx = GET_PMU_CTX();
5545
5546 /*
5547 * if we have some pending bits set
5548 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5549 */
5550 if (PMC0_HAS_OVFL(pmc0) && task) {
5551 /*
5552 * we assume that pmc0.fr is always set here
5553 */
5554
5555 /* sanity check */
5556 if (!ctx) goto report_spurious1;
5557
5558 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5559 goto report_spurious2;
5560
5561 PROTECT_CTX_NOPRINT(ctx, flags);
5562
5563 pfm_overflow_handler(task, ctx, pmc0, regs);
5564
5565 UNPROTECT_CTX_NOPRINT(ctx, flags);
5566
5567 } else {
5568 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5569 retval = -1;
5570 }
5571 /*
5572 * keep it unfrozen at all times
5573 */
5574 pfm_unfreeze_pmu();
5575
5576 return retval;
5577
5578 report_spurious1:
5579 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5580 this_cpu, task->pid);
5581 pfm_unfreeze_pmu();
5582 return -1;
5583 report_spurious2:
5584 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5585 this_cpu,
5586 task->pid);
5587 pfm_unfreeze_pmu();
5588 return -1;
5589 }
5590
5591 static irqreturn_t
5592 pfm_interrupt_handler(int irq, void *arg)
5593 {
5594 unsigned long start_cycles, total_cycles;
5595 unsigned long min, max;
5596 int this_cpu;
5597 int ret;
5598 struct pt_regs *regs = get_irq_regs();
5599
5600 this_cpu = get_cpu();
5601 if (likely(!pfm_alt_intr_handler)) {
5602 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5603 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5604
5605 start_cycles = ia64_get_itc();
5606
5607 ret = pfm_do_interrupt_handler(irq, arg, regs);
5608
5609 total_cycles = ia64_get_itc();
5610
5611 /*
5612 * don't measure spurious interrupts
5613 */
5614 if (likely(ret == 0)) {
5615 total_cycles -= start_cycles;
5616
5617 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5618 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5619
5620 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5621 }
5622 }
5623 else {
5624 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5625 }
5626
5627 put_cpu_no_resched();
5628 return IRQ_HANDLED;
5629 }
5630
5631 /*
5632 * /proc/perfmon interface, for debug only
5633 */
5634
5635 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5636
5637 static void *
5638 pfm_proc_start(struct seq_file *m, loff_t *pos)
5639 {
5640 if (*pos == 0) {
5641 return PFM_PROC_SHOW_HEADER;
5642 }
5643
5644 while (*pos <= NR_CPUS) {
5645 if (cpu_online(*pos - 1)) {
5646 return (void *)*pos;
5647 }
5648 ++*pos;
5649 }
5650 return NULL;
5651 }
5652
5653 static void *
5654 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5655 {
5656 ++*pos;
5657 return pfm_proc_start(m, pos);
5658 }
5659
5660 static void
5661 pfm_proc_stop(struct seq_file *m, void *v)
5662 {
5663 }
5664
5665 static void
5666 pfm_proc_show_header(struct seq_file *m)
5667 {
5668 struct list_head * pos;
5669 pfm_buffer_fmt_t * entry;
5670 unsigned long flags;
5671
5672 seq_printf(m,
5673 "perfmon version : %u.%u\n"
5674 "model : %s\n"
5675 "fastctxsw : %s\n"
5676 "expert mode : %s\n"
5677 "ovfl_mask : 0x%lx\n"
5678 "PMU flags : 0x%x\n",
5679 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5680 pmu_conf->pmu_name,
5681 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5682 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5683 pmu_conf->ovfl_val,
5684 pmu_conf->flags);
5685
5686 LOCK_PFS(flags);
5687
5688 seq_printf(m,
5689 "proc_sessions : %u\n"
5690 "sys_sessions : %u\n"
5691 "sys_use_dbregs : %u\n"
5692 "ptrace_use_dbregs : %u\n",
5693 pfm_sessions.pfs_task_sessions,
5694 pfm_sessions.pfs_sys_sessions,
5695 pfm_sessions.pfs_sys_use_dbregs,
5696 pfm_sessions.pfs_ptrace_use_dbregs);
5697
5698 UNLOCK_PFS(flags);
5699
5700 spin_lock(&pfm_buffer_fmt_lock);
5701
5702 list_for_each(pos, &pfm_buffer_fmt_list) {
5703 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5704 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5705 entry->fmt_uuid[0],
5706 entry->fmt_uuid[1],
5707 entry->fmt_uuid[2],
5708 entry->fmt_uuid[3],
5709 entry->fmt_uuid[4],
5710 entry->fmt_uuid[5],
5711 entry->fmt_uuid[6],
5712 entry->fmt_uuid[7],
5713 entry->fmt_uuid[8],
5714 entry->fmt_uuid[9],
5715 entry->fmt_uuid[10],
5716 entry->fmt_uuid[11],
5717 entry->fmt_uuid[12],
5718 entry->fmt_uuid[13],
5719 entry->fmt_uuid[14],
5720 entry->fmt_uuid[15],
5721 entry->fmt_name);
5722 }
5723 spin_unlock(&pfm_buffer_fmt_lock);
5724
5725 }
5726
5727 static int
5728 pfm_proc_show(struct seq_file *m, void *v)
5729 {
5730 unsigned long psr;
5731 unsigned int i;
5732 int cpu;
5733
5734 if (v == PFM_PROC_SHOW_HEADER) {
5735 pfm_proc_show_header(m);
5736 return 0;
5737 }
5738
5739 /* show info for CPU (v - 1) */
5740
5741 cpu = (long)v - 1;
5742 seq_printf(m,
5743 "CPU%-2d overflow intrs : %lu\n"
5744 "CPU%-2d overflow cycles : %lu\n"
5745 "CPU%-2d overflow min : %lu\n"
5746 "CPU%-2d overflow max : %lu\n"
5747 "CPU%-2d smpl handler calls : %lu\n"
5748 "CPU%-2d smpl handler cycles : %lu\n"
5749 "CPU%-2d spurious intrs : %lu\n"
5750 "CPU%-2d replay intrs : %lu\n"
5751 "CPU%-2d syst_wide : %d\n"
5752 "CPU%-2d dcr_pp : %d\n"
5753 "CPU%-2d exclude idle : %d\n"
5754 "CPU%-2d owner : %d\n"
5755 "CPU%-2d context : %p\n"
5756 "CPU%-2d activations : %lu\n",
5757 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5758 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5759 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5760 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5761 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5762 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5763 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5764 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5765 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5766 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5767 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5768 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5769 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5770 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5771
5772 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5773
5774 psr = pfm_get_psr();
5775
5776 ia64_srlz_d();
5777
5778 seq_printf(m,
5779 "CPU%-2d psr : 0x%lx\n"
5780 "CPU%-2d pmc0 : 0x%lx\n",
5781 cpu, psr,
5782 cpu, ia64_get_pmc(0));
5783
5784 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5785 if (PMC_IS_COUNTING(i) == 0) continue;
5786 seq_printf(m,
5787 "CPU%-2d pmc%u : 0x%lx\n"
5788 "CPU%-2d pmd%u : 0x%lx\n",
5789 cpu, i, ia64_get_pmc(i),
5790 cpu, i, ia64_get_pmd(i));
5791 }
5792 }
5793 return 0;
5794 }
5795
5796 struct seq_operations pfm_seq_ops = {
5797 .start = pfm_proc_start,
5798 .next = pfm_proc_next,
5799 .stop = pfm_proc_stop,
5800 .show = pfm_proc_show
5801 };
5802
5803 static int
5804 pfm_proc_open(struct inode *inode, struct file *file)
5805 {
5806 return seq_open(file, &pfm_seq_ops);
5807 }
5808
5809
5810 /*
5811 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5812 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5813 * is active or inactive based on mode. We must rely on the value in
5814 * local_cpu_data->pfm_syst_info
5815 */
5816 void
5817 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5818 {
5819 struct pt_regs *regs;
5820 unsigned long dcr;
5821 unsigned long dcr_pp;
5822
5823 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5824
5825 /*
5826 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5827 * on every CPU, so we can rely on the pid to identify the idle task.
5828 */
5829 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5830 regs = task_pt_regs(task);
5831 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5832 return;
5833 }
5834 /*
5835 * if monitoring has started
5836 */
5837 if (dcr_pp) {
5838 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5839 /*
5840 * context switching in?
5841 */
5842 if (is_ctxswin) {
5843 /* mask monitoring for the idle task */
5844 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5845 pfm_clear_psr_pp();
5846 ia64_srlz_i();
5847 return;
5848 }
5849 /*
5850 * context switching out
5851 * restore monitoring for next task
5852 *
5853 * Due to inlining this odd if-then-else construction generates
5854 * better code.
5855 */
5856 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5857 pfm_set_psr_pp();
5858 ia64_srlz_i();
5859 }
5860 }
5861
5862 #ifdef CONFIG_SMP
5863
5864 static void
5865 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5866 {
5867 struct task_struct *task = ctx->ctx_task;
5868
5869 ia64_psr(regs)->up = 0;
5870 ia64_psr(regs)->sp = 1;
5871
5872 if (GET_PMU_OWNER() == task) {
5873 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5874 SET_PMU_OWNER(NULL, NULL);
5875 }
5876
5877 /*
5878 * disconnect the task from the context and vice-versa
5879 */
5880 PFM_SET_WORK_PENDING(task, 0);
5881
5882 task->thread.pfm_context = NULL;
5883 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5884
5885 DPRINT(("force cleanup for [%d]\n", task->pid));
5886 }
5887
5888
5889 /*
5890 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5891 */
5892 void
5893 pfm_save_regs(struct task_struct *task)
5894 {
5895 pfm_context_t *ctx;
5896 unsigned long flags;
5897 u64 psr;
5898
5899
5900 ctx = PFM_GET_CTX(task);
5901 if (ctx == NULL) return;
5902
5903 /*
5904 * we always come here with interrupts ALREADY disabled by
5905 * the scheduler. So we simply need to protect against concurrent
5906 * access, not CPU concurrency.
5907 */
5908 flags = pfm_protect_ctx_ctxsw(ctx);
5909
5910 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5911 struct pt_regs *regs = task_pt_regs(task);
5912
5913 pfm_clear_psr_up();
5914
5915 pfm_force_cleanup(ctx, regs);
5916
5917 BUG_ON(ctx->ctx_smpl_hdr);
5918
5919 pfm_unprotect_ctx_ctxsw(ctx, flags);
5920
5921 pfm_context_free(ctx);
5922 return;
5923 }
5924
5925 /*
5926 * save current PSR: needed because we modify it
5927 */
5928 ia64_srlz_d();
5929 psr = pfm_get_psr();
5930
5931 BUG_ON(psr & (IA64_PSR_I));
5932
5933 /*
5934 * stop monitoring:
5935 * This is the last instruction which may generate an overflow
5936 *
5937 * We do not need to set psr.sp because, it is irrelevant in kernel.
5938 * It will be restored from ipsr when going back to user level
5939 */
5940 pfm_clear_psr_up();
5941
5942 /*
5943 * keep a copy of psr.up (for reload)
5944 */
5945 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5946
5947 /*
5948 * release ownership of this PMU.
5949 * PM interrupts are masked, so nothing
5950 * can happen.
5951 */
5952 SET_PMU_OWNER(NULL, NULL);
5953
5954 /*
5955 * we systematically save the PMD as we have no
5956 * guarantee we will be schedule at that same
5957 * CPU again.
5958 */
5959 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5960
5961 /*
5962 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5963 * we will need it on the restore path to check
5964 * for pending overflow.
5965 */
5966 ctx->th_pmcs[0] = ia64_get_pmc(0);
5967
5968 /*
5969 * unfreeze PMU if had pending overflows
5970 */
5971 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5972
5973 /*
5974 * finally, allow context access.
5975 * interrupts will still be masked after this call.
5976 */
5977 pfm_unprotect_ctx_ctxsw(ctx, flags);
5978 }
5979
5980 #else /* !CONFIG_SMP */
5981 void
5982 pfm_save_regs(struct task_struct *task)
5983 {
5984 pfm_context_t *ctx;
5985 u64 psr;
5986
5987 ctx = PFM_GET_CTX(task);
5988 if (ctx == NULL) return;
5989
5990 /*
5991 * save current PSR: needed because we modify it
5992 */
5993 psr = pfm_get_psr();
5994
5995 BUG_ON(psr & (IA64_PSR_I));
5996
5997 /*
5998 * stop monitoring:
5999 * This is the last instruction which may generate an overflow
6000 *
6001 * We do not need to set psr.sp because, it is irrelevant in kernel.
6002 * It will be restored from ipsr when going back to user level
6003 */
6004 pfm_clear_psr_up();
6005
6006 /*
6007 * keep a copy of psr.up (for reload)
6008 */
6009 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
6010 }
6011
6012 static void
6013 pfm_lazy_save_regs (struct task_struct *task)
6014 {
6015 pfm_context_t *ctx;
6016 unsigned long flags;
6017
6018 { u64 psr = pfm_get_psr();
6019 BUG_ON(psr & IA64_PSR_UP);
6020 }
6021
6022 ctx = PFM_GET_CTX(task);
6023
6024 /*
6025 * we need to mask PMU overflow here to
6026 * make sure that we maintain pmc0 until
6027 * we save it. overflow interrupts are
6028 * treated as spurious if there is no
6029 * owner.
6030 *
6031 * XXX: I don't think this is necessary
6032 */
6033 PROTECT_CTX(ctx,flags);
6034
6035 /*
6036 * release ownership of this PMU.
6037 * must be done before we save the registers.
6038 *
6039 * after this call any PMU interrupt is treated
6040 * as spurious.
6041 */
6042 SET_PMU_OWNER(NULL, NULL);
6043
6044 /*
6045 * save all the pmds we use
6046 */
6047 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6048
6049 /*
6050 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6051 * it is needed to check for pended overflow
6052 * on the restore path
6053 */
6054 ctx->th_pmcs[0] = ia64_get_pmc(0);
6055
6056 /*
6057 * unfreeze PMU if had pending overflows
6058 */
6059 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6060
6061 /*
6062 * now get can unmask PMU interrupts, they will
6063 * be treated as purely spurious and we will not
6064 * lose any information
6065 */
6066 UNPROTECT_CTX(ctx,flags);
6067 }
6068 #endif /* CONFIG_SMP */
6069
6070 #ifdef CONFIG_SMP
6071 /*
6072 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6073 */
6074 void
6075 pfm_load_regs (struct task_struct *task)
6076 {
6077 pfm_context_t *ctx;
6078 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6079 unsigned long flags;
6080 u64 psr, psr_up;
6081 int need_irq_resend;
6082
6083 ctx = PFM_GET_CTX(task);
6084 if (unlikely(ctx == NULL)) return;
6085
6086 BUG_ON(GET_PMU_OWNER());
6087
6088 /*
6089 * possible on unload
6090 */
6091 if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6092
6093 /*
6094 * we always come here with interrupts ALREADY disabled by
6095 * the scheduler. So we simply need to protect against concurrent
6096 * access, not CPU concurrency.
6097 */
6098 flags = pfm_protect_ctx_ctxsw(ctx);
6099 psr = pfm_get_psr();
6100
6101 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6102
6103 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6104 BUG_ON(psr & IA64_PSR_I);
6105
6106 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6107 struct pt_regs *regs = task_pt_regs(task);
6108
6109 BUG_ON(ctx->ctx_smpl_hdr);
6110
6111 pfm_force_cleanup(ctx, regs);
6112
6113 pfm_unprotect_ctx_ctxsw(ctx, flags);
6114
6115 /*
6116 * this one (kmalloc'ed) is fine with interrupts disabled
6117 */
6118 pfm_context_free(ctx);
6119
6120 return;
6121 }
6122
6123 /*
6124 * we restore ALL the debug registers to avoid picking up
6125 * stale state.
6126 */
6127 if (ctx->ctx_fl_using_dbreg) {
6128 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6129 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6130 }
6131 /*
6132 * retrieve saved psr.up
6133 */
6134 psr_up = ctx->ctx_saved_psr_up;
6135
6136 /*
6137 * if we were the last user of the PMU on that CPU,
6138 * then nothing to do except restore psr
6139 */
6140 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6141
6142 /*
6143 * retrieve partial reload masks (due to user modifications)
6144 */
6145 pmc_mask = ctx->ctx_reload_pmcs[0];
6146 pmd_mask = ctx->ctx_reload_pmds[0];
6147
6148 } else {
6149 /*
6150 * To avoid leaking information to the user level when psr.sp=0,
6151 * we must reload ALL implemented pmds (even the ones we don't use).
6152 * In the kernel we only allow PFM_READ_PMDS on registers which
6153 * we initialized or requested (sampling) so there is no risk there.
6154 */
6155 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6156
6157 /*
6158 * ALL accessible PMCs are systematically reloaded, unused registers
6159 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6160 * up stale configuration.
6161 *
6162 * PMC0 is never in the mask. It is always restored separately.
6163 */
6164 pmc_mask = ctx->ctx_all_pmcs[0];
6165 }
6166 /*
6167 * when context is MASKED, we will restore PMC with plm=0
6168 * and PMD with stale information, but that's ok, nothing
6169 * will be captured.
6170 *
6171 * XXX: optimize here
6172 */
6173 if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6174 if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6175
6176 /*
6177 * check for pending overflow at the time the state
6178 * was saved.
6179 */
6180 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6181 /*
6182 * reload pmc0 with the overflow information
6183 * On McKinley PMU, this will trigger a PMU interrupt
6184 */
6185 ia64_set_pmc(0, ctx->th_pmcs[0]);
6186 ia64_srlz_d();
6187 ctx->th_pmcs[0] = 0UL;
6188
6189 /*
6190 * will replay the PMU interrupt
6191 */
6192 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6193
6194 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6195 }
6196
6197 /*
6198 * we just did a reload, so we reset the partial reload fields
6199 */
6200 ctx->ctx_reload_pmcs[0] = 0UL;
6201 ctx->ctx_reload_pmds[0] = 0UL;
6202
6203 SET_LAST_CPU(ctx, smp_processor_id());
6204
6205 /*
6206 * dump activation value for this PMU
6207 */
6208 INC_ACTIVATION();
6209 /*
6210 * record current activation for this context
6211 */
6212 SET_ACTIVATION(ctx);
6213
6214 /*
6215 * establish new ownership.
6216 */
6217 SET_PMU_OWNER(task, ctx);
6218
6219 /*
6220 * restore the psr.up bit. measurement
6221 * is active again.
6222 * no PMU interrupt can happen at this point
6223 * because we still have interrupts disabled.
6224 */
6225 if (likely(psr_up)) pfm_set_psr_up();
6226
6227 /*
6228 * allow concurrent access to context
6229 */
6230 pfm_unprotect_ctx_ctxsw(ctx, flags);
6231 }
6232 #else /* !CONFIG_SMP */
6233 /*
6234 * reload PMU state for UP kernels
6235 * in 2.5 we come here with interrupts disabled
6236 */
6237 void
6238 pfm_load_regs (struct task_struct *task)
6239 {
6240 pfm_context_t *ctx;
6241 struct task_struct *owner;
6242 unsigned long pmd_mask, pmc_mask;
6243 u64 psr, psr_up;
6244 int need_irq_resend;
6245
6246 owner = GET_PMU_OWNER();
6247 ctx = PFM_GET_CTX(task);
6248 psr = pfm_get_psr();
6249
6250 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6251 BUG_ON(psr & IA64_PSR_I);
6252
6253 /*
6254 * we restore ALL the debug registers to avoid picking up
6255 * stale state.
6256 *
6257 * This must be done even when the task is still the owner
6258 * as the registers may have been modified via ptrace()
6259 * (not perfmon) by the previous task.
6260 */
6261 if (ctx->ctx_fl_using_dbreg) {
6262 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6263 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6264 }
6265
6266 /*
6267 * retrieved saved psr.up
6268 */
6269 psr_up = ctx->ctx_saved_psr_up;
6270 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6271
6272 /*
6273 * short path, our state is still there, just
6274 * need to restore psr and we go
6275 *
6276 * we do not touch either PMC nor PMD. the psr is not touched
6277 * by the overflow_handler. So we are safe w.r.t. to interrupt
6278 * concurrency even without interrupt masking.
6279 */
6280 if (likely(owner == task)) {
6281 if (likely(psr_up)) pfm_set_psr_up();
6282 return;
6283 }
6284
6285 /*
6286 * someone else is still using the PMU, first push it out and
6287 * then we'll be able to install our stuff !
6288 *
6289 * Upon return, there will be no owner for the current PMU
6290 */
6291 if (owner) pfm_lazy_save_regs(owner);
6292
6293 /*
6294 * To avoid leaking information to the user level when psr.sp=0,
6295 * we must reload ALL implemented pmds (even the ones we don't use).
6296 * In the kernel we only allow PFM_READ_PMDS on registers which
6297 * we initialized or requested (sampling) so there is no risk there.
6298 */
6299 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6300
6301 /*
6302 * ALL accessible PMCs are systematically reloaded, unused registers
6303 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6304 * up stale configuration.
6305 *
6306 * PMC0 is never in the mask. It is always restored separately
6307 */
6308 pmc_mask = ctx->ctx_all_pmcs[0];
6309
6310 pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6311 pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6312
6313 /*
6314 * check for pending overflow at the time the state
6315 * was saved.
6316 */
6317 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6318 /*
6319 * reload pmc0 with the overflow information
6320 * On McKinley PMU, this will trigger a PMU interrupt
6321 */
6322 ia64_set_pmc(0, ctx->th_pmcs[0]);
6323 ia64_srlz_d();
6324
6325 ctx->th_pmcs[0] = 0UL;
6326
6327 /*
6328 * will replay the PMU interrupt
6329 */
6330 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6331
6332 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6333 }
6334
6335 /*
6336 * establish new ownership.
6337 */
6338 SET_PMU_OWNER(task, ctx);
6339
6340 /*
6341 * restore the psr.up bit. measurement
6342 * is active again.
6343 * no PMU interrupt can happen at this point
6344 * because we still have interrupts disabled.
6345 */
6346 if (likely(psr_up)) pfm_set_psr_up();
6347 }
6348 #endif /* CONFIG_SMP */
6349
6350 /*
6351 * this function assumes monitoring is stopped
6352 */
6353 static void
6354 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6355 {
6356 u64 pmc0;
6357 unsigned long mask2, val, pmd_val, ovfl_val;
6358 int i, can_access_pmu = 0;
6359 int is_self;
6360
6361 /*
6362 * is the caller the task being monitored (or which initiated the
6363 * session for system wide measurements)
6364 */
6365 is_self = ctx->ctx_task == task ? 1 : 0;
6366
6367 /*
6368 * can access PMU is task is the owner of the PMU state on the current CPU
6369 * or if we are running on the CPU bound to the context in system-wide mode
6370 * (that is not necessarily the task the context is attached to in this mode).
6371 * In system-wide we always have can_access_pmu true because a task running on an
6372 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6373 */
6374 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6375 if (can_access_pmu) {
6376 /*
6377 * Mark the PMU as not owned
6378 * This will cause the interrupt handler to do nothing in case an overflow
6379 * interrupt was in-flight
6380 * This also guarantees that pmc0 will contain the final state
6381 * It virtually gives us full control on overflow processing from that point
6382 * on.
6383 */
6384 SET_PMU_OWNER(NULL, NULL);
6385 DPRINT(("releasing ownership\n"));
6386
6387 /*
6388 * read current overflow status:
6389 *
6390 * we are guaranteed to read the final stable state
6391 */
6392 ia64_srlz_d();
6393 pmc0 = ia64_get_pmc(0); /* slow */
6394
6395 /*
6396 * reset freeze bit, overflow status information destroyed
6397 */
6398 pfm_unfreeze_pmu();
6399 } else {
6400 pmc0 = ctx->th_pmcs[0];
6401 /*
6402 * clear whatever overflow status bits there were
6403 */
6404 ctx->th_pmcs[0] = 0;
6405 }
6406 ovfl_val = pmu_conf->ovfl_val;
6407 /*
6408 * we save all the used pmds
6409 * we take care of overflows for counting PMDs
6410 *
6411 * XXX: sampling situation is not taken into account here
6412 */
6413 mask2 = ctx->ctx_used_pmds[0];
6414
6415 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6416
6417 for (i = 0; mask2; i++, mask2>>=1) {
6418
6419 /* skip non used pmds */
6420 if ((mask2 & 0x1) == 0) continue;
6421
6422 /*
6423 * can access PMU always true in system wide mode
6424 */
6425 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6426
6427 if (PMD_IS_COUNTING(i)) {
6428 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6429 task->pid,
6430 i,
6431 ctx->ctx_pmds[i].val,
6432 val & ovfl_val));
6433
6434 /*
6435 * we rebuild the full 64 bit value of the counter
6436 */
6437 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6438
6439 /*
6440 * now everything is in ctx_pmds[] and we need
6441 * to clear the saved context from save_regs() such that
6442 * pfm_read_pmds() gets the correct value
6443 */
6444 pmd_val = 0UL;
6445
6446 /*
6447 * take care of overflow inline
6448 */
6449 if (pmc0 & (1UL << i)) {
6450 val += 1 + ovfl_val;
6451 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6452 }
6453 }
6454
6455 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, i, val, pmd_val));
6456
6457 if (is_self) ctx->th_pmds[i] = pmd_val;
6458
6459 ctx->ctx_pmds[i].val = val;
6460 }
6461 }
6462
6463 static struct irqaction perfmon_irqaction = {
6464 .handler = pfm_interrupt_handler,
6465 .flags = IRQF_DISABLED,
6466 .name = "perfmon"
6467 };
6468
6469 static void
6470 pfm_alt_save_pmu_state(void *data)
6471 {
6472 struct pt_regs *regs;
6473
6474 regs = task_pt_regs(current);
6475
6476 DPRINT(("called\n"));
6477
6478 /*
6479 * should not be necessary but
6480 * let's take not risk
6481 */
6482 pfm_clear_psr_up();
6483 pfm_clear_psr_pp();
6484 ia64_psr(regs)->pp = 0;
6485
6486 /*
6487 * This call is required
6488 * May cause a spurious interrupt on some processors
6489 */
6490 pfm_freeze_pmu();
6491
6492 ia64_srlz_d();
6493 }
6494
6495 void
6496 pfm_alt_restore_pmu_state(void *data)
6497 {
6498 struct pt_regs *regs;
6499
6500 regs = task_pt_regs(current);
6501
6502 DPRINT(("called\n"));
6503
6504 /*
6505 * put PMU back in state expected
6506 * by perfmon
6507 */
6508 pfm_clear_psr_up();
6509 pfm_clear_psr_pp();
6510 ia64_psr(regs)->pp = 0;
6511
6512 /*
6513 * perfmon runs with PMU unfrozen at all times
6514 */
6515 pfm_unfreeze_pmu();
6516
6517 ia64_srlz_d();
6518 }
6519
6520 int
6521 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6522 {
6523 int ret, i;
6524 int reserve_cpu;
6525
6526 /* some sanity checks */
6527 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6528
6529 /* do the easy test first */
6530 if (pfm_alt_intr_handler) return -EBUSY;
6531
6532 /* one at a time in the install or remove, just fail the others */
6533 if (!spin_trylock(&pfm_alt_install_check)) {
6534 return -EBUSY;
6535 }
6536
6537 /* reserve our session */
6538 for_each_online_cpu(reserve_cpu) {
6539 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6540 if (ret) goto cleanup_reserve;
6541 }
6542
6543 /* save the current system wide pmu states */
6544 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 0, 1);
6545 if (ret) {
6546 DPRINT(("on_each_cpu() failed: %d\n", ret));
6547 goto cleanup_reserve;
6548 }
6549
6550 /* officially change to the alternate interrupt handler */
6551 pfm_alt_intr_handler = hdl;
6552
6553 spin_unlock(&pfm_alt_install_check);
6554
6555 return 0;
6556
6557 cleanup_reserve:
6558 for_each_online_cpu(i) {
6559 /* don't unreserve more than we reserved */
6560 if (i >= reserve_cpu) break;
6561
6562 pfm_unreserve_session(NULL, 1, i);
6563 }
6564
6565 spin_unlock(&pfm_alt_install_check);
6566
6567 return ret;
6568 }
6569 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6570
6571 int
6572 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6573 {
6574 int i;
6575 int ret;
6576
6577 if (hdl == NULL) return -EINVAL;
6578
6579 /* cannot remove someone else's handler! */
6580 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6581
6582 /* one at a time in the install or remove, just fail the others */
6583 if (!spin_trylock(&pfm_alt_install_check)) {
6584 return -EBUSY;
6585 }
6586
6587 pfm_alt_intr_handler = NULL;
6588
6589 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 0, 1);
6590 if (ret) {
6591 DPRINT(("on_each_cpu() failed: %d\n", ret));
6592 }
6593
6594 for_each_online_cpu(i) {
6595 pfm_unreserve_session(NULL, 1, i);
6596 }
6597
6598 spin_unlock(&pfm_alt_install_check);
6599
6600 return 0;
6601 }
6602 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6603
6604 /*
6605 * perfmon initialization routine, called from the initcall() table
6606 */
6607 static int init_pfm_fs(void);
6608
6609 static int __init
6610 pfm_probe_pmu(void)
6611 {
6612 pmu_config_t **p;
6613 int family;
6614
6615 family = local_cpu_data->family;
6616 p = pmu_confs;
6617
6618 while(*p) {
6619 if ((*p)->probe) {
6620 if ((*p)->probe() == 0) goto found;
6621 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6622 goto found;
6623 }
6624 p++;
6625 }
6626 return -1;
6627 found:
6628 pmu_conf = *p;
6629 return 0;
6630 }
6631
6632 static const struct file_operations pfm_proc_fops = {
6633 .open = pfm_proc_open,
6634 .read = seq_read,
6635 .llseek = seq_lseek,
6636 .release = seq_release,
6637 };
6638
6639 int __init
6640 pfm_init(void)
6641 {
6642 unsigned int n, n_counters, i;
6643
6644 printk("perfmon: version %u.%u IRQ %u\n",
6645 PFM_VERSION_MAJ,
6646 PFM_VERSION_MIN,
6647 IA64_PERFMON_VECTOR);
6648
6649 if (pfm_probe_pmu()) {
6650 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6651 local_cpu_data->family);
6652 return -ENODEV;
6653 }
6654
6655 /*
6656 * compute the number of implemented PMD/PMC from the
6657 * description tables
6658 */
6659 n = 0;
6660 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6661 if (PMC_IS_IMPL(i) == 0) continue;
6662 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6663 n++;
6664 }
6665 pmu_conf->num_pmcs = n;
6666
6667 n = 0; n_counters = 0;
6668 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6669 if (PMD_IS_IMPL(i) == 0) continue;
6670 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6671 n++;
6672 if (PMD_IS_COUNTING(i)) n_counters++;
6673 }
6674 pmu_conf->num_pmds = n;
6675 pmu_conf->num_counters = n_counters;
6676
6677 /*
6678 * sanity checks on the number of debug registers
6679 */
6680 if (pmu_conf->use_rr_dbregs) {
6681 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6682 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6683 pmu_conf = NULL;
6684 return -1;
6685 }
6686 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6687 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6688 pmu_conf = NULL;
6689 return -1;
6690 }
6691 }
6692
6693 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6694 pmu_conf->pmu_name,
6695 pmu_conf->num_pmcs,
6696 pmu_conf->num_pmds,
6697 pmu_conf->num_counters,
6698 ffz(pmu_conf->ovfl_val));
6699
6700 /* sanity check */
6701 if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6702 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6703 pmu_conf = NULL;
6704 return -1;
6705 }
6706
6707 /*
6708 * create /proc/perfmon (mostly for debugging purposes)
6709 */
6710 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6711 if (perfmon_dir == NULL) {
6712 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6713 pmu_conf = NULL;
6714 return -1;
6715 }
6716 /*
6717 * install customized file operations for /proc/perfmon entry
6718 */
6719 perfmon_dir->proc_fops = &pfm_proc_fops;
6720
6721 /*
6722 * create /proc/sys/kernel/perfmon (for debugging purposes)
6723 */
6724 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6725
6726 /*
6727 * initialize all our spinlocks
6728 */
6729 spin_lock_init(&pfm_sessions.pfs_lock);
6730 spin_lock_init(&pfm_buffer_fmt_lock);
6731
6732 init_pfm_fs();
6733
6734 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6735
6736 return 0;
6737 }
6738
6739 __initcall(pfm_init);
6740
6741 /*
6742 * this function is called before pfm_init()
6743 */
6744 void
6745 pfm_init_percpu (void)
6746 {
6747 static int first_time=1;
6748 /*
6749 * make sure no measurement is active
6750 * (may inherit programmed PMCs from EFI).
6751 */
6752 pfm_clear_psr_pp();
6753 pfm_clear_psr_up();
6754
6755 /*
6756 * we run with the PMU not frozen at all times
6757 */
6758 pfm_unfreeze_pmu();
6759
6760 if (first_time) {
6761 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6762 first_time=0;
6763 }
6764
6765 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6766 ia64_srlz_d();
6767 }
6768
6769 /*
6770 * used for debug purposes only
6771 */
6772 void
6773 dump_pmu_state(const char *from)
6774 {
6775 struct task_struct *task;
6776 struct pt_regs *regs;
6777 pfm_context_t *ctx;
6778 unsigned long psr, dcr, info, flags;
6779 int i, this_cpu;
6780
6781 local_irq_save(flags);
6782
6783 this_cpu = smp_processor_id();
6784 regs = task_pt_regs(current);
6785 info = PFM_CPUINFO_GET();
6786 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6787
6788 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6789 local_irq_restore(flags);
6790 return;
6791 }
6792
6793 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6794 this_cpu,
6795 from,
6796 current->pid,
6797 regs->cr_iip,
6798 current->comm);
6799
6800 task = GET_PMU_OWNER();
6801 ctx = GET_PMU_CTX();
6802
6803 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6804
6805 psr = pfm_get_psr();
6806
6807 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",
6808 this_cpu,
6809 ia64_get_pmc(0),
6810 psr & IA64_PSR_PP ? 1 : 0,
6811 psr & IA64_PSR_UP ? 1 : 0,
6812 dcr & IA64_DCR_PP ? 1 : 0,
6813 info,
6814 ia64_psr(regs)->up,
6815 ia64_psr(regs)->pp);
6816
6817 ia64_psr(regs)->up = 0;
6818 ia64_psr(regs)->pp = 0;
6819
6820 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6821 if (PMC_IS_IMPL(i) == 0) continue;
6822 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]);
6823 }
6824
6825 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6826 if (PMD_IS_IMPL(i) == 0) continue;
6827 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]);
6828 }
6829
6830 if (ctx) {
6831 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6832 this_cpu,
6833 ctx->ctx_state,
6834 ctx->ctx_smpl_vaddr,
6835 ctx->ctx_smpl_hdr,
6836 ctx->ctx_msgq_head,
6837 ctx->ctx_msgq_tail,
6838 ctx->ctx_saved_psr_up);
6839 }
6840 local_irq_restore(flags);
6841 }
6842
6843 /*
6844 * called from process.c:copy_thread(). task is new child.
6845 */
6846 void
6847 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6848 {
6849 struct thread_struct *thread;
6850
6851 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6852
6853 thread = &task->thread;
6854
6855 /*
6856 * cut links inherited from parent (current)
6857 */
6858 thread->pfm_context = NULL;
6859
6860 PFM_SET_WORK_PENDING(task, 0);
6861
6862 /*
6863 * the psr bits are already set properly in copy_threads()
6864 */
6865 }
6866 #else /* !CONFIG_PERFMON */
6867 asmlinkage long
6868 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6869 {
6870 return -ENOSYS;
6871 }
6872 #endif /* CONFIG_PERFMON */
Linux kernel - Deliverables
This page took 0.320144 seconds and 6 git commands to generate.