?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/trivial
[linux-2.6/linux-2.6-openrd.git] / arch / ia64 / kernel / perfmon.c
blob599b233bef7514500a3daf8782a0f339dacce974
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 #include <linux/tracehook.h>
44
45 #include <asm/errno.h>
46 #include <asm/intrinsics.h>
47 #include <asm/page.h>
48 #include <asm/perfmon.h>
49 #include <asm/processor.h>
50 #include <asm/signal.h>
51 #include <asm/system.h>
52 #include <asm/uaccess.h>
53 #include <asm/delay.h>
54
55 #ifdef CONFIG_PERFMON
56 /*
57 * perfmon context state
58 */
59 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
60 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
61 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
62 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
63
64 #define PFM_INVALID_ACTIVATION (~0UL)
65
66 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
67 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
68
69 /*
70 * depth of message queue
71 */
72 #define PFM_MAX_MSGS 32
73 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
74
75 /*
76 * type of a PMU register (bitmask).
77 * bitmask structure:
78 * bit0 : register implemented
79 * bit1 : end marker
80 * bit2-3 : reserved
81 * bit4 : pmc has pmc.pm
82 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
83 * bit6-7 : register type
84 * bit8-31: reserved
85 */
86 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
87 #define PFM_REG_IMPL 0x1 /* register implemented */
88 #define PFM_REG_END 0x2 /* end marker */
89 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
90 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
91 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
92 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
93 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
94
95 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
96 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
97
98 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
99
100 /* i assumed unsigned */
101 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
102 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
103
104 /* XXX: these assume that register i is implemented */
105 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
106 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
107 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
108 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
109
110 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
111 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
112 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
113 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
114
115 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
116 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
117
118 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
119 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
120 #define PFM_CTX_TASK(h) (h)->ctx_task
121
122 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
123
124 /* XXX: does not support more than 64 PMDs */
125 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
126 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
127
128 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
129
130 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
131 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
132 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
133 #define PFM_CODE_RR 0 /* requesting code range restriction */
134 #define PFM_DATA_RR 1 /* requestion data range restriction */
135
136 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
137 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
138 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
139
140 #define RDEP(x) (1UL<<(x))
141
142 /*
143 * context protection macros
144 * in SMP:
145 * - we need to protect against CPU concurrency (spin_lock)
146 * - we need to protect against PMU overflow interrupts (local_irq_disable)
147 * in UP:
148 * - we need to protect against PMU overflow interrupts (local_irq_disable)
149 *
150 * spin_lock_irqsave()/spin_unlock_irqrestore():
151 * in SMP: local_irq_disable + spin_lock
152 * in UP : local_irq_disable
153 *
154 * spin_lock()/spin_lock():
155 * in UP : removed automatically
156 * in SMP: protect against context accesses from other CPU. interrupts
157 * are not masked. This is useful for the PMU interrupt handler
158 * because we know we will not get PMU concurrency in that code.
159 */
160 #define PROTECT_CTX(c, f) \
161 do { \
162 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
163 spin_lock_irqsave(&(c)->ctx_lock, f); \
164 DPRINT(("spinlocked ctx %p by [%d]\n", c, task_pid_nr(current))); \
165 } while(0)
166
167 #define UNPROTECT_CTX(c, f) \
168 do { \
169 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
170 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
171 } while(0)
172
173 #define PROTECT_CTX_NOPRINT(c, f) \
174 do { \
175 spin_lock_irqsave(&(c)->ctx_lock, f); \
176 } while(0)
177
178
179 #define UNPROTECT_CTX_NOPRINT(c, f) \
180 do { \
181 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
182 } while(0)
183
184
185 #define PROTECT_CTX_NOIRQ(c) \
186 do { \
187 spin_lock(&(c)->ctx_lock); \
188 } while(0)
189
190 #define UNPROTECT_CTX_NOIRQ(c) \
191 do { \
192 spin_unlock(&(c)->ctx_lock); \
193 } while(0)
194
195
196 #ifdef CONFIG_SMP
197
198 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
199 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
200 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
201
202 #else /* !CONFIG_SMP */
203 #define SET_ACTIVATION(t) do {} while(0)
204 #define GET_ACTIVATION(t) do {} while(0)
205 #define INC_ACTIVATION(t) do {} while(0)
206 #endif /* CONFIG_SMP */
207
208 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
209 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
210 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
211
212 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
213 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
214
215 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
216
217 /*
218 * cmp0 must be the value of pmc0
219 */
220 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
221
222 #define PFMFS_MAGIC 0xa0b4d889
223
224 /*
225 * debugging
226 */
227 #define PFM_DEBUGGING 1
228 #ifdef PFM_DEBUGGING
229 #define DPRINT(a) \
230 do { \
231 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
232 } while (0)
233
234 #define DPRINT_ovfl(a) \
235 do { \
236 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
237 } while (0)
238 #endif
239
240 /*
241 * 64-bit software counter structure
242 *
243 * the next_reset_type is applied to the next call to pfm_reset_regs()
244 */
245 typedef struct {
246 unsigned long val; /* virtual 64bit counter value */
247 unsigned long lval; /* last reset value */
248 unsigned long long_reset; /* reset value on sampling overflow */
249 unsigned long short_reset; /* reset value on overflow */
250 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
251 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
252 unsigned long seed; /* seed for random-number generator */
253 unsigned long mask; /* mask for random-number generator */
254 unsigned int flags; /* notify/do not notify */
255 unsigned long eventid; /* overflow event identifier */
256 } pfm_counter_t;
257
258 /*
259 * context flags
260 */
261 typedef struct {
262 unsigned int block:1; /* when 1, task will blocked on user notifications */
263 unsigned int system:1; /* do system wide monitoring */
264 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
265 unsigned int is_sampling:1; /* true if using a custom format */
266 unsigned int excl_idle:1; /* exclude idle task in system wide session */
267 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
268 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
269 unsigned int no_msg:1; /* no message sent on overflow */
270 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
271 unsigned int reserved:22;
272 } pfm_context_flags_t;
273
274 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
275 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
276 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
277
278
279 /*
280 * perfmon context: encapsulates all the state of a monitoring session
281 */
282
283 typedef struct pfm_context {
284 spinlock_t ctx_lock; /* context protection */
285
286 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
287 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
288
289 struct task_struct *ctx_task; /* task to which context is attached */
290
291 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
292
293 struct completion ctx_restart_done; /* use for blocking notification mode */
294
295 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
296 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
297 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
298
299 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
300 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
301 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
302
303 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
304
305 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
306 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
307 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
308 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
309
310 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
311
312 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
313 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
314
315 unsigned long ctx_saved_psr_up; /* only contains psr.up value */
316
317 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
318 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
319 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
320
321 int ctx_fd; /* file descriptor used my this context */
322 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
323
324 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
325 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
326 unsigned long ctx_smpl_size; /* size of sampling buffer */
327 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
328
329 wait_queue_head_t ctx_msgq_wait;
330 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
331 int ctx_msgq_head;
332 int ctx_msgq_tail;
333 struct fasync_struct *ctx_async_queue;
334
335 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
336 } pfm_context_t;
337
338 /*
339 * magic number used to verify that structure is really
340 * a perfmon context
341 */
342 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
343
344 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
345
346 #ifdef CONFIG_SMP
347 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
348 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
349 #else
350 #define SET_LAST_CPU(ctx, v) do {} while(0)
351 #define GET_LAST_CPU(ctx) do {} while(0)
352 #endif
353
354
355 #define ctx_fl_block ctx_flags.block
356 #define ctx_fl_system ctx_flags.system
357 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
358 #define ctx_fl_is_sampling ctx_flags.is_sampling
359 #define ctx_fl_excl_idle ctx_flags.excl_idle
360 #define ctx_fl_going_zombie ctx_flags.going_zombie
361 #define ctx_fl_trap_reason ctx_flags.trap_reason
362 #define ctx_fl_no_msg ctx_flags.no_msg
363 #define ctx_fl_can_restart ctx_flags.can_restart
364
365 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
366 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
367
368 /*
369 * global information about all sessions
370 * mostly used to synchronize between system wide and per-process
371 */
372 typedef struct {
373 spinlock_t pfs_lock; /* lock the structure */
374
375 unsigned int pfs_task_sessions; /* number of per task sessions */
376 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
377 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
378 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
379 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
380 } pfm_session_t;
381
382 /*
383 * information about a PMC or PMD.
384 * dep_pmd[]: a bitmask of dependent PMD registers
385 * dep_pmc[]: a bitmask of dependent PMC registers
386 */
387 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
388 typedef struct {
389 unsigned int type;
390 int pm_pos;
391 unsigned long default_value; /* power-on default value */
392 unsigned long reserved_mask; /* bitmask of reserved bits */
393 pfm_reg_check_t read_check;
394 pfm_reg_check_t write_check;
395 unsigned long dep_pmd[4];
396 unsigned long dep_pmc[4];
397 } pfm_reg_desc_t;
398
399 /* assume cnum is a valid monitor */
400 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
401
402 /*
403 * This structure is initialized at boot time and contains
404 * a description of the PMU main characteristics.
405 *
406 * If the probe function is defined, detection is based
407 * on its return value:
408 * - 0 means recognized PMU
409 * - anything else means not supported
410 * When the probe function is not defined, then the pmu_family field
411 * is used and it must match the host CPU family such that:
412 * - cpu->family & config->pmu_family != 0
413 */
414 typedef struct {
415 unsigned long ovfl_val; /* overflow value for counters */
416
417 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
418 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
419
420 unsigned int num_pmcs; /* number of PMCS: computed at init time */
421 unsigned int num_pmds; /* number of PMDS: computed at init time */
422 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
423 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
424
425 char *pmu_name; /* PMU family name */
426 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
427 unsigned int flags; /* pmu specific flags */
428 unsigned int num_ibrs; /* number of IBRS: computed at init time */
429 unsigned int num_dbrs; /* number of DBRS: computed at init time */
430 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
431 int (*probe)(void); /* customized probe routine */
432 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
433 } pmu_config_t;
434 /*
435 * PMU specific flags
436 */
437 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
438
439 /*
440 * debug register related type definitions
441 */
442 typedef struct {
443 unsigned long ibr_mask:56;
444 unsigned long ibr_plm:4;
445 unsigned long ibr_ig:3;
446 unsigned long ibr_x:1;
447 } ibr_mask_reg_t;
448
449 typedef struct {
450 unsigned long dbr_mask:56;
451 unsigned long dbr_plm:4;
452 unsigned long dbr_ig:2;
453 unsigned long dbr_w:1;
454 unsigned long dbr_r:1;
455 } dbr_mask_reg_t;
456
457 typedef union {
458 unsigned long val;
459 ibr_mask_reg_t ibr;
460 dbr_mask_reg_t dbr;
461 } dbreg_t;
462
463
464 /*
465 * perfmon command descriptions
466 */
467 typedef struct {
468 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
469 char *cmd_name;
470 int cmd_flags;
471 unsigned int cmd_narg;
472 size_t cmd_argsize;
473 int (*cmd_getsize)(void *arg, size_t *sz);
474 } pfm_cmd_desc_t;
475
476 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
477 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
478 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
479 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
480
481
482 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
483 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
484 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
485 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
486 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
487
488 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
489
490 typedef struct {
491 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
492 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
493 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
496 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
497 unsigned long pfm_smpl_handler_calls;
498 unsigned long pfm_smpl_handler_cycles;
499 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
500 } pfm_stats_t;
501
502 /*
503 * perfmon internal variables
504 */
505 static pfm_stats_t pfm_stats[NR_CPUS];
506 static pfm_session_t pfm_sessions; /* global sessions information */
507
508 static DEFINE_SPINLOCK(pfm_alt_install_check);
509 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
510
511 static struct proc_dir_entry *perfmon_dir;
512 static pfm_uuid_t pfm_null_uuid = {0,};
513
514 static spinlock_t pfm_buffer_fmt_lock;
515 static LIST_HEAD(pfm_buffer_fmt_list);
516
517 static pmu_config_t *pmu_conf;
518
519 /* sysctl() controls */
520 pfm_sysctl_t pfm_sysctl;
521 EXPORT_SYMBOL(pfm_sysctl);
522
523 static ctl_table pfm_ctl_table[]={
524 {
525 .procname = "debug",
526 .data = &pfm_sysctl.debug,
527 .maxlen = sizeof(int),
528 .mode = 0666,
529 .proc_handler = proc_dointvec,
530 },
531 {
532 .procname = "debug_ovfl",
533 .data = &pfm_sysctl.debug_ovfl,
534 .maxlen = sizeof(int),
535 .mode = 0666,
536 .proc_handler = proc_dointvec,
537 },
538 {
539 .procname = "fastctxsw",
540 .data = &pfm_sysctl.fastctxsw,
541 .maxlen = sizeof(int),
542 .mode = 0600,
543 .proc_handler = proc_dointvec,
544 },
545 {
546 .procname = "expert_mode",
547 .data = &pfm_sysctl.expert_mode,
548 .maxlen = sizeof(int),
549 .mode = 0600,
550 .proc_handler = proc_dointvec,
551 },
552 {}
553 };
554 static ctl_table pfm_sysctl_dir[] = {
555 {
556 .procname = "perfmon",
557 .mode = 0555,
558 .child = pfm_ctl_table,
559 },
560 {}
561 };
562 static ctl_table pfm_sysctl_root[] = {
563 {
564 .procname = "kernel",
565 .mode = 0555,
566 .child = pfm_sysctl_dir,
567 },
568 {}
569 };
570 static struct ctl_table_header *pfm_sysctl_header;
571
572 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
573
574 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
575 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
576
577 static inline void
578 pfm_put_task(struct task_struct *task)
579 {
580 if (task != current) put_task_struct(task);
581 }
582
583 static inline void
584 pfm_reserve_page(unsigned long a)
585 {
586 SetPageReserved(vmalloc_to_page((void *)a));
587 }
588 static inline void
589 pfm_unreserve_page(unsigned long a)
590 {
591 ClearPageReserved(vmalloc_to_page((void*)a));
592 }
593
594 static inline unsigned long
595 pfm_protect_ctx_ctxsw(pfm_context_t *x)
596 {
597 spin_lock(&(x)->ctx_lock);
598 return 0UL;
599 }
600
601 static inline void
602 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
603 {
604 spin_unlock(&(x)->ctx_lock);
605 }
606
607 static inline unsigned int
608 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
609 {
610 return do_munmap(mm, addr, len);
611 }
612
613 static inline unsigned long
614 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
615 {
616 return get_unmapped_area(file, addr, len, pgoff, flags);
617 }
618
619
620 static int
621 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data,
622 struct vfsmount *mnt)
623 {
624 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
625 }
626
627 static struct file_system_type pfm_fs_type = {
628 .name = "pfmfs",
629 .get_sb = pfmfs_get_sb,
630 .kill_sb = kill_anon_super,
631 };
632
633 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
634 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
635 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
636 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
637 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
638
639
640 /* forward declaration */
641 static const struct file_operations pfm_file_ops;
642
643 /*
644 * forward declarations
645 */
646 #ifndef CONFIG_SMP
647 static void pfm_lazy_save_regs (struct task_struct *ta);
648 #endif
649
650 void dump_pmu_state(const char *);
651 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
652
653 #include "perfmon_itanium.h"
654 #include "perfmon_mckinley.h"
655 #include "perfmon_montecito.h"
656 #include "perfmon_generic.h"
657
658 static pmu_config_t *pmu_confs[]={
659 &pmu_conf_mont,
660 &pmu_conf_mck,
661 &pmu_conf_ita,
662 &pmu_conf_gen, /* must be last */
663 NULL
664 };
665
666
667 static int pfm_end_notify_user(pfm_context_t *ctx);
668
669 static inline void
670 pfm_clear_psr_pp(void)
671 {
672 ia64_rsm(IA64_PSR_PP);
673 ia64_srlz_i();
674 }
675
676 static inline void
677 pfm_set_psr_pp(void)
678 {
679 ia64_ssm(IA64_PSR_PP);
680 ia64_srlz_i();
681 }
682
683 static inline void
684 pfm_clear_psr_up(void)
685 {
686 ia64_rsm(IA64_PSR_UP);
687 ia64_srlz_i();
688 }
689
690 static inline void
691 pfm_set_psr_up(void)
692 {
693 ia64_ssm(IA64_PSR_UP);
694 ia64_srlz_i();
695 }
696
697 static inline unsigned long
698 pfm_get_psr(void)
699 {
700 unsigned long tmp;
701 tmp = ia64_getreg(_IA64_REG_PSR);
702 ia64_srlz_i();
703 return tmp;
704 }
705
706 static inline void
707 pfm_set_psr_l(unsigned long val)
708 {
709 ia64_setreg(_IA64_REG_PSR_L, val);
710 ia64_srlz_i();
711 }
712
713 static inline void
714 pfm_freeze_pmu(void)
715 {
716 ia64_set_pmc(0,1UL);
717 ia64_srlz_d();
718 }
719
720 static inline void
721 pfm_unfreeze_pmu(void)
722 {
723 ia64_set_pmc(0,0UL);
724 ia64_srlz_d();
725 }
726
727 static inline void
728 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
729 {
730 int i;
731
732 for (i=0; i < nibrs; i++) {
733 ia64_set_ibr(i, ibrs[i]);
734 ia64_dv_serialize_instruction();
735 }
736 ia64_srlz_i();
737 }
738
739 static inline void
740 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
741 {
742 int i;
743
744 for (i=0; i < ndbrs; i++) {
745 ia64_set_dbr(i, dbrs[i]);
746 ia64_dv_serialize_data();
747 }
748 ia64_srlz_d();
749 }
750
751 /*
752 * PMD[i] must be a counter. no check is made
753 */
754 static inline unsigned long
755 pfm_read_soft_counter(pfm_context_t *ctx, int i)
756 {
757 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
758 }
759
760 /*
761 * PMD[i] must be a counter. no check is made
762 */
763 static inline void
764 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
765 {
766 unsigned long ovfl_val = pmu_conf->ovfl_val;
767
768 ctx->ctx_pmds[i].val = val & ~ovfl_val;
769 /*
770 * writing to unimplemented part is ignore, so we do not need to
771 * mask off top part
772 */
773 ia64_set_pmd(i, val & ovfl_val);
774 }
775
776 static pfm_msg_t *
777 pfm_get_new_msg(pfm_context_t *ctx)
778 {
779 int idx, next;
780
781 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
782
783 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
784 if (next == ctx->ctx_msgq_head) return NULL;
785
786 idx = ctx->ctx_msgq_tail;
787 ctx->ctx_msgq_tail = next;
788
789 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
790
791 return ctx->ctx_msgq+idx;
792 }
793
794 static pfm_msg_t *
795 pfm_get_next_msg(pfm_context_t *ctx)
796 {
797 pfm_msg_t *msg;
798
799 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
800
801 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
802
803 /*
804 * get oldest message
805 */
806 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
807
808 /*
809 * and move forward
810 */
811 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
812
813 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));
814
815 return msg;
816 }
817
818 static void
819 pfm_reset_msgq(pfm_context_t *ctx)
820 {
821 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
822 DPRINT(("ctx=%p msgq reset\n", ctx));
823 }
824
825 static void *
826 pfm_rvmalloc(unsigned long size)
827 {
828 void *mem;
829 unsigned long addr;
830
831 size = PAGE_ALIGN(size);
832 mem = vmalloc(size);
833 if (mem) {
834 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
835 memset(mem, 0, size);
836 addr = (unsigned long)mem;
837 while (size > 0) {
838 pfm_reserve_page(addr);
839 addr+=PAGE_SIZE;
840 size-=PAGE_SIZE;
841 }
842 }
843 return mem;
844 }
845
846 static void
847 pfm_rvfree(void *mem, unsigned long size)
848 {
849 unsigned long addr;
850
851 if (mem) {
852 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
853 addr = (unsigned long) mem;
854 while ((long) size > 0) {
855 pfm_unreserve_page(addr);
856 addr+=PAGE_SIZE;
857 size-=PAGE_SIZE;
858 }
859 vfree(mem);
860 }
861 return;
862 }
863
864 static pfm_context_t *
865 pfm_context_alloc(int ctx_flags)
866 {
867 pfm_context_t *ctx;
868
869 /*
870 * allocate context descriptor
871 * must be able to free with interrupts disabled
872 */
873 ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
874 if (ctx) {
875 DPRINT(("alloc ctx @%p\n", ctx));
876
877 /*
878 * init context protection lock
879 */
880 spin_lock_init(&ctx->ctx_lock);
881
882 /*
883 * context is unloaded
884 */
885 ctx->ctx_state = PFM_CTX_UNLOADED;
886
887 /*
888 * initialization of context's flags
889 */
890 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
891 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
892 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
893 /*
894 * will move to set properties
895 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
896 */
897
898 /*
899 * init restart semaphore to locked
900 */
901 init_completion(&ctx->ctx_restart_done);
902
903 /*
904 * activation is used in SMP only
905 */
906 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
907 SET_LAST_CPU(ctx, -1);
908
909 /*
910 * initialize notification message queue
911 */
912 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
913 init_waitqueue_head(&ctx->ctx_msgq_wait);
914 init_waitqueue_head(&ctx->ctx_zombieq);
915
916 }
917 return ctx;
918 }
919
920 static void
921 pfm_context_free(pfm_context_t *ctx)
922 {
923 if (ctx) {
924 DPRINT(("free ctx @%p\n", ctx));
925 kfree(ctx);
926 }
927 }
928
929 static void
930 pfm_mask_monitoring(struct task_struct *task)
931 {
932 pfm_context_t *ctx = PFM_GET_CTX(task);
933 unsigned long mask, val, ovfl_mask;
934 int i;
935
936 DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
937
938 ovfl_mask = pmu_conf->ovfl_val;
939 /*
940 * monitoring can only be masked as a result of a valid
941 * counter overflow. In UP, it means that the PMU still
942 * has an owner. Note that the owner can be different
943 * from the current task. However the PMU state belongs
944 * to the owner.
945 * In SMP, a valid overflow only happens when task is
946 * current. Therefore if we come here, we know that
947 * the PMU state belongs to the current task, therefore
948 * we can access the live registers.
949 *
950 * So in both cases, the live register contains the owner's
951 * state. We can ONLY touch the PMU registers and NOT the PSR.
952 *
953 * As a consequence to this call, the ctx->th_pmds[] array
954 * contains stale information which must be ignored
955 * when context is reloaded AND monitoring is active (see
956 * pfm_restart).
957 */
958 mask = ctx->ctx_used_pmds[0];
959 for (i = 0; mask; i++, mask>>=1) {
960 /* skip non used pmds */
961 if ((mask & 0x1) == 0) continue;
962 val = ia64_get_pmd(i);
963
964 if (PMD_IS_COUNTING(i)) {
965 /*
966 * we rebuild the full 64 bit value of the counter
967 */
968 ctx->ctx_pmds[i].val += (val & ovfl_mask);
969 } else {
970 ctx->ctx_pmds[i].val = val;
971 }
972 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
973 i,
974 ctx->ctx_pmds[i].val,
975 val & ovfl_mask));
976 }
977 /*
978 * mask monitoring by setting the privilege level to 0
979 * we cannot use psr.pp/psr.up for this, it is controlled by
980 * the user
981 *
982 * if task is current, modify actual registers, otherwise modify
983 * thread save state, i.e., what will be restored in pfm_load_regs()
984 */
985 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
986 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
987 if ((mask & 0x1) == 0UL) continue;
988 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
989 ctx->th_pmcs[i] &= ~0xfUL;
990 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
991 }
992 /*
993 * make all of this visible
994 */
995 ia64_srlz_d();
996 }
997
998 /*
999 * must always be done with task == current
1000 *
1001 * context must be in MASKED state when calling
1002 */
1003 static void
1004 pfm_restore_monitoring(struct task_struct *task)
1005 {
1006 pfm_context_t *ctx = PFM_GET_CTX(task);
1007 unsigned long mask, ovfl_mask;
1008 unsigned long psr, val;
1009 int i, is_system;
1010
1011 is_system = ctx->ctx_fl_system;
1012 ovfl_mask = pmu_conf->ovfl_val;
1013
1014 if (task != current) {
1015 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1016 return;
1017 }
1018 if (ctx->ctx_state != PFM_CTX_MASKED) {
1019 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1020 task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1021 return;
1022 }
1023 psr = pfm_get_psr();
1024 /*
1025 * monitoring is masked via the PMC.
1026 * As we restore their value, we do not want each counter to
1027 * restart right away. We stop monitoring using the PSR,
1028 * restore the PMC (and PMD) and then re-establish the psr
1029 * as it was. Note that there can be no pending overflow at
1030 * this point, because monitoring was MASKED.
1031 *
1032 * system-wide session are pinned and self-monitoring
1033 */
1034 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1035 /* disable dcr pp */
1036 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1037 pfm_clear_psr_pp();
1038 } else {
1039 pfm_clear_psr_up();
1040 }
1041 /*
1042 * first, we restore the PMD
1043 */
1044 mask = ctx->ctx_used_pmds[0];
1045 for (i = 0; mask; i++, mask>>=1) {
1046 /* skip non used pmds */
1047 if ((mask & 0x1) == 0) continue;
1048
1049 if (PMD_IS_COUNTING(i)) {
1050 /*
1051 * we split the 64bit value according to
1052 * counter width
1053 */
1054 val = ctx->ctx_pmds[i].val & ovfl_mask;
1055 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1056 } else {
1057 val = ctx->ctx_pmds[i].val;
1058 }
1059 ia64_set_pmd(i, val);
1060
1061 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1062 i,
1063 ctx->ctx_pmds[i].val,
1064 val));
1065 }
1066 /*
1067 * restore the PMCs
1068 */
1069 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1070 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1071 if ((mask & 0x1) == 0UL) continue;
1072 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1073 ia64_set_pmc(i, ctx->th_pmcs[i]);
1074 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1075 task_pid_nr(task), i, ctx->th_pmcs[i]));
1076 }
1077 ia64_srlz_d();
1078
1079 /*
1080 * must restore DBR/IBR because could be modified while masked
1081 * XXX: need to optimize
1082 */
1083 if (ctx->ctx_fl_using_dbreg) {
1084 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1085 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1086 }
1087
1088 /*
1089 * now restore PSR
1090 */
1091 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1092 /* enable dcr pp */
1093 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1094 ia64_srlz_i();
1095 }
1096 pfm_set_psr_l(psr);
1097 }
1098
1099 static inline void
1100 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1101 {
1102 int i;
1103
1104 ia64_srlz_d();
1105
1106 for (i=0; mask; i++, mask>>=1) {
1107 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1108 }
1109 }
1110
1111 /*
1112 * reload from thread state (used for ctxw only)
1113 */
1114 static inline void
1115 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1116 {
1117 int i;
1118 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1119
1120 for (i=0; mask; i++, mask>>=1) {
1121 if ((mask & 0x1) == 0) continue;
1122 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1123 ia64_set_pmd(i, val);
1124 }
1125 ia64_srlz_d();
1126 }
1127
1128 /*
1129 * propagate PMD from context to thread-state
1130 */
1131 static inline void
1132 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1133 {
1134 unsigned long ovfl_val = pmu_conf->ovfl_val;
1135 unsigned long mask = ctx->ctx_all_pmds[0];
1136 unsigned long val;
1137 int i;
1138
1139 DPRINT(("mask=0x%lx\n", mask));
1140
1141 for (i=0; mask; i++, mask>>=1) {
1142
1143 val = ctx->ctx_pmds[i].val;
1144
1145 /*
1146 * We break up the 64 bit value into 2 pieces
1147 * the lower bits go to the machine state in the
1148 * thread (will be reloaded on ctxsw in).
1149 * The upper part stays in the soft-counter.
1150 */
1151 if (PMD_IS_COUNTING(i)) {
1152 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1153 val &= ovfl_val;
1154 }
1155 ctx->th_pmds[i] = val;
1156
1157 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1158 i,
1159 ctx->th_pmds[i],
1160 ctx->ctx_pmds[i].val));
1161 }
1162 }
1163
1164 /*
1165 * propagate PMC from context to thread-state
1166 */
1167 static inline void
1168 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1169 {
1170 unsigned long mask = ctx->ctx_all_pmcs[0];
1171 int i;
1172
1173 DPRINT(("mask=0x%lx\n", mask));
1174
1175 for (i=0; mask; i++, mask>>=1) {
1176 /* masking 0 with ovfl_val yields 0 */
1177 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1178 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1179 }
1180 }
1181
1182
1183
1184 static inline void
1185 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1186 {
1187 int i;
1188
1189 for (i=0; mask; i++, mask>>=1) {
1190 if ((mask & 0x1) == 0) continue;
1191 ia64_set_pmc(i, pmcs[i]);
1192 }
1193 ia64_srlz_d();
1194 }
1195
1196 static inline int
1197 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1198 {
1199 return memcmp(a, b, sizeof(pfm_uuid_t));
1200 }
1201
1202 static inline int
1203 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1204 {
1205 int ret = 0;
1206 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1207 return ret;
1208 }
1209
1210 static inline int
1211 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1212 {
1213 int ret = 0;
1214 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1215 return ret;
1216 }
1217
1218
1219 static inline int
1220 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1221 int cpu, void *arg)
1222 {
1223 int ret = 0;
1224 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1225 return ret;
1226 }
1227
1228 static inline int
1229 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1230 int cpu, void *arg)
1231 {
1232 int ret = 0;
1233 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1234 return ret;
1235 }
1236
1237 static inline int
1238 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1239 {
1240 int ret = 0;
1241 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1242 return ret;
1243 }
1244
1245 static inline int
1246 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)
1247 {
1248 int ret = 0;
1249 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1250 return ret;
1251 }
1252
1253 static pfm_buffer_fmt_t *
1254 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1255 {
1256 struct list_head * pos;
1257 pfm_buffer_fmt_t * entry;
1258
1259 list_for_each(pos, &pfm_buffer_fmt_list) {
1260 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1261 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1262 return entry;
1263 }
1264 return NULL;
1265 }
1266
1267 /*
1268 * find a buffer format based on its uuid
1269 */
1270 static pfm_buffer_fmt_t *
1271 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1272 {
1273 pfm_buffer_fmt_t * fmt;
1274 spin_lock(&pfm_buffer_fmt_lock);
1275 fmt = __pfm_find_buffer_fmt(uuid);
1276 spin_unlock(&pfm_buffer_fmt_lock);
1277 return fmt;
1278 }
1279
1280 int
1281 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1282 {
1283 int ret = 0;
1284
1285 /* some sanity checks */
1286 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1287
1288 /* we need at least a handler */
1289 if (fmt->fmt_handler == NULL) return -EINVAL;
1290
1291 /*
1292 * XXX: need check validity of fmt_arg_size
1293 */
1294
1295 spin_lock(&pfm_buffer_fmt_lock);
1296
1297 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1298 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1299 ret = -EBUSY;
1300 goto out;
1301 }
1302 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1303 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1304
1305 out:
1306 spin_unlock(&pfm_buffer_fmt_lock);
1307 return ret;
1308 }
1309 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1310
1311 int
1312 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1313 {
1314 pfm_buffer_fmt_t *fmt;
1315 int ret = 0;
1316
1317 spin_lock(&pfm_buffer_fmt_lock);
1318
1319 fmt = __pfm_find_buffer_fmt(uuid);
1320 if (!fmt) {
1321 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1322 ret = -EINVAL;
1323 goto out;
1324 }
1325 list_del_init(&fmt->fmt_list);
1326 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1327
1328 out:
1329 spin_unlock(&pfm_buffer_fmt_lock);
1330 return ret;
1331
1332 }
1333 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1334
1335 extern void update_pal_halt_status(int);
1336
1337 static int
1338 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1339 {
1340 unsigned long flags;
1341 /*
1342 * validity checks on cpu_mask have been done upstream
1343 */
1344 LOCK_PFS(flags);
1345
1346 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1347 pfm_sessions.pfs_sys_sessions,
1348 pfm_sessions.pfs_task_sessions,
1349 pfm_sessions.pfs_sys_use_dbregs,
1350 is_syswide,
1351 cpu));
1352
1353 if (is_syswide) {
1354 /*
1355 * cannot mix system wide and per-task sessions
1356 */
1357 if (pfm_sessions.pfs_task_sessions > 0UL) {
1358 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1359 pfm_sessions.pfs_task_sessions));
1360 goto abort;
1361 }
1362
1363 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1364
1365 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1366
1367 pfm_sessions.pfs_sys_session[cpu] = task;
1368
1369 pfm_sessions.pfs_sys_sessions++ ;
1370
1371 } else {
1372 if (pfm_sessions.pfs_sys_sessions) goto abort;
1373 pfm_sessions.pfs_task_sessions++;
1374 }
1375
1376 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1377 pfm_sessions.pfs_sys_sessions,
1378 pfm_sessions.pfs_task_sessions,
1379 pfm_sessions.pfs_sys_use_dbregs,
1380 is_syswide,
1381 cpu));
1382
1383 /*
1384 * disable default_idle() to go to PAL_HALT
1385 */
1386 update_pal_halt_status(0);
1387
1388 UNLOCK_PFS(flags);
1389
1390 return 0;
1391
1392 error_conflict:
1393 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1394 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1395 cpu));
1396 abort:
1397 UNLOCK_PFS(flags);
1398
1399 return -EBUSY;
1400
1401 }
1402
1403 static int
1404 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1405 {
1406 unsigned long flags;
1407 /*
1408 * validity checks on cpu_mask have been done upstream
1409 */
1410 LOCK_PFS(flags);
1411
1412 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1413 pfm_sessions.pfs_sys_sessions,
1414 pfm_sessions.pfs_task_sessions,
1415 pfm_sessions.pfs_sys_use_dbregs,
1416 is_syswide,
1417 cpu));
1418
1419
1420 if (is_syswide) {
1421 pfm_sessions.pfs_sys_session[cpu] = NULL;
1422 /*
1423 * would not work with perfmon+more than one bit in cpu_mask
1424 */
1425 if (ctx && ctx->ctx_fl_using_dbreg) {
1426 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1427 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1428 } else {
1429 pfm_sessions.pfs_sys_use_dbregs--;
1430 }
1431 }
1432 pfm_sessions.pfs_sys_sessions--;
1433 } else {
1434 pfm_sessions.pfs_task_sessions--;
1435 }
1436 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1437 pfm_sessions.pfs_sys_sessions,
1438 pfm_sessions.pfs_task_sessions,
1439 pfm_sessions.pfs_sys_use_dbregs,
1440 is_syswide,
1441 cpu));
1442
1443 /*
1444 * if possible, enable default_idle() to go into PAL_HALT
1445 */
1446 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1447 update_pal_halt_status(1);
1448
1449 UNLOCK_PFS(flags);
1450
1451 return 0;
1452 }
1453
1454 /*
1455 * removes virtual mapping of the sampling buffer.
1456 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1457 * a PROTECT_CTX() section.
1458 */
1459 static int
1460 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1461 {
1462 int r;
1463
1464 /* sanity checks */
1465 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1466 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1467 return -EINVAL;
1468 }
1469
1470 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1471
1472 /*
1473 * does the actual unmapping
1474 */
1475 down_write(&task->mm->mmap_sem);
1476
1477 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1478
1479 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1480
1481 up_write(&task->mm->mmap_sem);
1482 if (r !=0) {
1483 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1484 }
1485
1486 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1487
1488 return 0;
1489 }
1490
1491 /*
1492 * free actual physical storage used by sampling buffer
1493 */
1494 #if 0
1495 static int
1496 pfm_free_smpl_buffer(pfm_context_t *ctx)
1497 {
1498 pfm_buffer_fmt_t *fmt;
1499
1500 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1501
1502 /*
1503 * we won't use the buffer format anymore
1504 */
1505 fmt = ctx->ctx_buf_fmt;
1506
1507 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1508 ctx->ctx_smpl_hdr,
1509 ctx->ctx_smpl_size,
1510 ctx->ctx_smpl_vaddr));
1511
1512 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1513
1514 /*
1515 * free the buffer
1516 */
1517 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1518
1519 ctx->ctx_smpl_hdr = NULL;
1520 ctx->ctx_smpl_size = 0UL;
1521
1522 return 0;
1523
1524 invalid_free:
1525 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1526 return -EINVAL;
1527 }
1528 #endif
1529
1530 static inline void
1531 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1532 {
1533 if (fmt == NULL) return;
1534
1535 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1536
1537 }
1538
1539 /*
1540 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1541 * no real gain from having the whole whorehouse mounted. So we don't need
1542 * any operations on the root directory. However, we need a non-trivial
1543 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1544 */
1545 static struct vfsmount *pfmfs_mnt;
1546
1547 static int __init
1548 init_pfm_fs(void)
1549 {
1550 int err = register_filesystem(&pfm_fs_type);
1551 if (!err) {
1552 pfmfs_mnt = kern_mount(&pfm_fs_type);
1553 err = PTR_ERR(pfmfs_mnt);
1554 if (IS_ERR(pfmfs_mnt))
1555 unregister_filesystem(&pfm_fs_type);
1556 else
1557 err = 0;
1558 }
1559 return err;
1560 }
1561
1562 static ssize_t
1563 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1564 {
1565 pfm_context_t *ctx;
1566 pfm_msg_t *msg;
1567 ssize_t ret;
1568 unsigned long flags;
1569 DECLARE_WAITQUEUE(wait, current);
1570 if (PFM_IS_FILE(filp) == 0) {
1571 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1572 return -EINVAL;
1573 }
1574
1575 ctx = (pfm_context_t *)filp->private_data;
1576 if (ctx == NULL) {
1577 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1578 return -EINVAL;
1579 }
1580
1581 /*
1582 * check even when there is no message
1583 */
1584 if (size < sizeof(pfm_msg_t)) {
1585 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1586 return -EINVAL;
1587 }
1588
1589 PROTECT_CTX(ctx, flags);
1590
1591 /*
1592 * put ourselves on the wait queue
1593 */
1594 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1595
1596
1597 for(;;) {
1598 /*
1599 * check wait queue
1600 */
1601
1602 set_current_state(TASK_INTERRUPTIBLE);
1603
1604 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1605
1606 ret = 0;
1607 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1608
1609 UNPROTECT_CTX(ctx, flags);
1610
1611 /*
1612 * check non-blocking read
1613 */
1614 ret = -EAGAIN;
1615 if(filp->f_flags & O_NONBLOCK) break;
1616
1617 /*
1618 * check pending signals
1619 */
1620 if(signal_pending(current)) {
1621 ret = -EINTR;
1622 break;
1623 }
1624 /*
1625 * no message, so wait
1626 */
1627 schedule();
1628
1629 PROTECT_CTX(ctx, flags);
1630 }
1631 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1632 set_current_state(TASK_RUNNING);
1633 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1634
1635 if (ret < 0) goto abort;
1636
1637 ret = -EINVAL;
1638 msg = pfm_get_next_msg(ctx);
1639 if (msg == NULL) {
1640 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1641 goto abort_locked;
1642 }
1643
1644 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1645
1646 ret = -EFAULT;
1647 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1648
1649 abort_locked:
1650 UNPROTECT_CTX(ctx, flags);
1651 abort:
1652 return ret;
1653 }
1654
1655 static ssize_t
1656 pfm_write(struct file *file, const char __user *ubuf,
1657 size_t size, loff_t *ppos)
1658 {
1659 DPRINT(("pfm_write called\n"));
1660 return -EINVAL;
1661 }
1662
1663 static unsigned int
1664 pfm_poll(struct file *filp, poll_table * wait)
1665 {
1666 pfm_context_t *ctx;
1667 unsigned long flags;
1668 unsigned int mask = 0;
1669
1670 if (PFM_IS_FILE(filp) == 0) {
1671 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1672 return 0;
1673 }
1674
1675 ctx = (pfm_context_t *)filp->private_data;
1676 if (ctx == NULL) {
1677 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1678 return 0;
1679 }
1680
1681
1682 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1683
1684 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1685
1686 PROTECT_CTX(ctx, flags);
1687
1688 if (PFM_CTXQ_EMPTY(ctx) == 0)
1689 mask = POLLIN | POLLRDNORM;
1690
1691 UNPROTECT_CTX(ctx, flags);
1692
1693 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1694
1695 return mask;
1696 }
1697
1698 static int
1699 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1700 {
1701 DPRINT(("pfm_ioctl called\n"));
1702 return -EINVAL;
1703 }
1704
1705 /*
1706 * interrupt cannot be masked when coming here
1707 */
1708 static inline int
1709 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1710 {
1711 int ret;
1712
1713 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1714
1715 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1716 task_pid_nr(current),
1717 fd,
1718 on,
1719 ctx->ctx_async_queue, ret));
1720
1721 return ret;
1722 }
1723
1724 static int
1725 pfm_fasync(int fd, struct file *filp, int on)
1726 {
1727 pfm_context_t *ctx;
1728 int ret;
1729
1730 if (PFM_IS_FILE(filp) == 0) {
1731 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1732 return -EBADF;
1733 }
1734
1735 ctx = (pfm_context_t *)filp->private_data;
1736 if (ctx == NULL) {
1737 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1738 return -EBADF;
1739 }
1740 /*
1741 * we cannot mask interrupts during this call because this may
1742 * may go to sleep if memory is not readily avalaible.
1743 *
1744 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1745 * done in caller. Serialization of this function is ensured by caller.
1746 */
1747 ret = pfm_do_fasync(fd, filp, ctx, on);
1748
1749
1750 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1751 fd,
1752 on,
1753 ctx->ctx_async_queue, ret));
1754
1755 return ret;
1756 }
1757
1758 #ifdef CONFIG_SMP
1759 /*
1760 * this function is exclusively called from pfm_close().
1761 * The context is not protected at that time, nor are interrupts
1762 * on the remote CPU. That's necessary to avoid deadlocks.
1763 */
1764 static void
1765 pfm_syswide_force_stop(void *info)
1766 {
1767 pfm_context_t *ctx = (pfm_context_t *)info;
1768 struct pt_regs *regs = task_pt_regs(current);
1769 struct task_struct *owner;
1770 unsigned long flags;
1771 int ret;
1772
1773 if (ctx->ctx_cpu != smp_processor_id()) {
1774 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1775 ctx->ctx_cpu,
1776 smp_processor_id());
1777 return;
1778 }
1779 owner = GET_PMU_OWNER();
1780 if (owner != ctx->ctx_task) {
1781 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1782 smp_processor_id(),
1783 task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1784 return;
1785 }
1786 if (GET_PMU_CTX() != ctx) {
1787 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1788 smp_processor_id(),
1789 GET_PMU_CTX(), ctx);
1790 return;
1791 }
1792
1793 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1794 /*
1795 * the context is already protected in pfm_close(), we simply
1796 * need to mask interrupts to avoid a PMU interrupt race on
1797 * this CPU
1798 */
1799 local_irq_save(flags);
1800
1801 ret = pfm_context_unload(ctx, NULL, 0, regs);
1802 if (ret) {
1803 DPRINT(("context_unload returned %d\n", ret));
1804 }
1805
1806 /*
1807 * unmask interrupts, PMU interrupts are now spurious here
1808 */
1809 local_irq_restore(flags);
1810 }
1811
1812 static void
1813 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1814 {
1815 int ret;
1816
1817 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1818 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1819 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1820 }
1821 #endif /* CONFIG_SMP */
1822
1823 /*
1824 * called for each close(). Partially free resources.
1825 * When caller is self-monitoring, the context is unloaded.
1826 */
1827 static int
1828 pfm_flush(struct file *filp, fl_owner_t id)
1829 {
1830 pfm_context_t *ctx;
1831 struct task_struct *task;
1832 struct pt_regs *regs;
1833 unsigned long flags;
1834 unsigned long smpl_buf_size = 0UL;
1835 void *smpl_buf_vaddr = NULL;
1836 int state, is_system;
1837
1838 if (PFM_IS_FILE(filp) == 0) {
1839 DPRINT(("bad magic for\n"));
1840 return -EBADF;
1841 }
1842
1843 ctx = (pfm_context_t *)filp->private_data;
1844 if (ctx == NULL) {
1845 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1846 return -EBADF;
1847 }
1848
1849 /*
1850 * remove our file from the async queue, if we use this mode.
1851 * This can be done without the context being protected. We come
1852 * here when the context has become unreachable by other tasks.
1853 *
1854 * We may still have active monitoring at this point and we may
1855 * end up in pfm_overflow_handler(). However, fasync_helper()
1856 * operates with interrupts disabled and it cleans up the
1857 * queue. If the PMU handler is called prior to entering
1858 * fasync_helper() then it will send a signal. If it is
1859 * invoked after, it will find an empty queue and no
1860 * signal will be sent. In both case, we are safe
1861 */
1862 PROTECT_CTX(ctx, flags);
1863
1864 state = ctx->ctx_state;
1865 is_system = ctx->ctx_fl_system;
1866
1867 task = PFM_CTX_TASK(ctx);
1868 regs = task_pt_regs(task);
1869
1870 DPRINT(("ctx_state=%d is_current=%d\n",
1871 state,
1872 task == current ? 1 : 0));
1873
1874 /*
1875 * if state == UNLOADED, then task is NULL
1876 */
1877
1878 /*
1879 * we must stop and unload because we are losing access to the context.
1880 */
1881 if (task == current) {
1882 #ifdef CONFIG_SMP
1883 /*
1884 * the task IS the owner but it migrated to another CPU: that's bad
1885 * but we must handle this cleanly. Unfortunately, the kernel does
1886 * not provide a mechanism to block migration (while the context is loaded).
1887 *
1888 * We need to release the resource on the ORIGINAL cpu.
1889 */
1890 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1891
1892 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1893 /*
1894 * keep context protected but unmask interrupt for IPI
1895 */
1896 local_irq_restore(flags);
1897
1898 pfm_syswide_cleanup_other_cpu(ctx);
1899
1900 /*
1901 * restore interrupt masking
1902 */
1903 local_irq_save(flags);
1904
1905 /*
1906 * context is unloaded at this point
1907 */
1908 } else
1909 #endif /* CONFIG_SMP */
1910 {
1911
1912 DPRINT(("forcing unload\n"));
1913 /*
1914 * stop and unload, returning with state UNLOADED
1915 * and session unreserved.
1916 */
1917 pfm_context_unload(ctx, NULL, 0, regs);
1918
1919 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1920 }
1921 }
1922
1923 /*
1924 * remove virtual mapping, if any, for the calling task.
1925 * cannot reset ctx field until last user is calling close().
1926 *
1927 * ctx_smpl_vaddr must never be cleared because it is needed
1928 * by every task with access to the context
1929 *
1930 * When called from do_exit(), the mm context is gone already, therefore
1931 * mm is NULL, i.e., the VMA is already gone and we do not have to
1932 * do anything here
1933 */
1934 if (ctx->ctx_smpl_vaddr && current->mm) {
1935 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1936 smpl_buf_size = ctx->ctx_smpl_size;
1937 }
1938
1939 UNPROTECT_CTX(ctx, flags);
1940
1941 /*
1942 * if there was a mapping, then we systematically remove it
1943 * at this point. Cannot be done inside critical section
1944 * because some VM function reenables interrupts.
1945 *
1946 */
1947 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1948
1949 return 0;
1950 }
1951 /*
1952 * called either on explicit close() or from exit_files().
1953 * Only the LAST user of the file gets to this point, i.e., it is
1954 * called only ONCE.
1955 *
1956 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1957 * (fput()),i.e, last task to access the file. Nobody else can access the
1958 * file at this point.
1959 *
1960 * When called from exit_files(), the VMA has been freed because exit_mm()
1961 * is executed before exit_files().
1962 *
1963 * When called from exit_files(), the current task is not yet ZOMBIE but we
1964 * flush the PMU state to the context.
1965 */
1966 static int
1967 pfm_close(struct inode *inode, struct file *filp)
1968 {
1969 pfm_context_t *ctx;
1970 struct task_struct *task;
1971 struct pt_regs *regs;
1972 DECLARE_WAITQUEUE(wait, current);
1973 unsigned long flags;
1974 unsigned long smpl_buf_size = 0UL;
1975 void *smpl_buf_addr = NULL;
1976 int free_possible = 1;
1977 int state, is_system;
1978
1979 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1980
1981 if (PFM_IS_FILE(filp) == 0) {
1982 DPRINT(("bad magic\n"));
1983 return -EBADF;
1984 }
1985
1986 ctx = (pfm_context_t *)filp->private_data;
1987 if (ctx == NULL) {
1988 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1989 return -EBADF;
1990 }
1991
1992 PROTECT_CTX(ctx, flags);
1993
1994 state = ctx->ctx_state;
1995 is_system = ctx->ctx_fl_system;
1996
1997 task = PFM_CTX_TASK(ctx);
1998 regs = task_pt_regs(task);
1999
2000 DPRINT(("ctx_state=%d is_current=%d\n",
2001 state,
2002 task == current ? 1 : 0));
2003
2004 /*
2005 * if task == current, then pfm_flush() unloaded the context
2006 */
2007 if (state == PFM_CTX_UNLOADED) goto doit;
2008
2009 /*
2010 * context is loaded/masked and task != current, we need to
2011 * either force an unload or go zombie
2012 */
2013
2014 /*
2015 * The task is currently blocked or will block after an overflow.
2016 * we must force it to wakeup to get out of the
2017 * MASKED state and transition to the unloaded state by itself.
2018 *
2019 * This situation is only possible for per-task mode
2020 */
2021 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2022
2023 /*
2024 * set a "partial" zombie state to be checked
2025 * upon return from down() in pfm_handle_work().
2026 *
2027 * We cannot use the ZOMBIE state, because it is checked
2028 * by pfm_load_regs() which is called upon wakeup from down().
2029 * In such case, it would free the context and then we would
2030 * return to pfm_handle_work() which would access the
2031 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2032 * but visible to pfm_handle_work().
2033 *
2034 * For some window of time, we have a zombie context with
2035 * ctx_state = MASKED and not ZOMBIE
2036 */
2037 ctx->ctx_fl_going_zombie = 1;
2038
2039 /*
2040 * force task to wake up from MASKED state
2041 */
2042 complete(&ctx->ctx_restart_done);
2043
2044 DPRINT(("waking up ctx_state=%d\n", state));
2045
2046 /*
2047 * put ourself to sleep waiting for the other
2048 * task to report completion
2049 *
2050 * the context is protected by mutex, therefore there
2051 * is no risk of being notified of completion before
2052 * begin actually on the waitq.
2053 */
2054 set_current_state(TASK_INTERRUPTIBLE);
2055 add_wait_queue(&ctx->ctx_zombieq, &wait);
2056
2057 UNPROTECT_CTX(ctx, flags);
2058
2059 /*
2060 * XXX: check for signals :
2061 * - ok for explicit close
2062 * - not ok when coming from exit_files()
2063 */
2064 schedule();
2065
2066
2067 PROTECT_CTX(ctx, flags);
2068
2069
2070 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2071 set_current_state(TASK_RUNNING);
2072
2073 /*
2074 * context is unloaded at this point
2075 */
2076 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2077 }
2078 else if (task != current) {
2079 #ifdef CONFIG_SMP
2080 /*
2081 * switch context to zombie state
2082 */
2083 ctx->ctx_state = PFM_CTX_ZOMBIE;
2084
2085 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2086 /*
2087 * cannot free the context on the spot. deferred until
2088 * the task notices the ZOMBIE state
2089 */
2090 free_possible = 0;
2091 #else
2092 pfm_context_unload(ctx, NULL, 0, regs);
2093 #endif
2094 }
2095
2096 doit:
2097 /* reload state, may have changed during opening of critical section */
2098 state = ctx->ctx_state;
2099
2100 /*
2101 * the context is still attached to a task (possibly current)
2102 * we cannot destroy it right now
2103 */
2104
2105 /*
2106 * we must free the sampling buffer right here because
2107 * we cannot rely on it being cleaned up later by the
2108 * monitored task. It is not possible to free vmalloc'ed
2109 * memory in pfm_load_regs(). Instead, we remove the buffer
2110 * now. should there be subsequent PMU overflow originally
2111 * meant for sampling, the will be converted to spurious
2112 * and that's fine because the monitoring tools is gone anyway.
2113 */
2114 if (ctx->ctx_smpl_hdr) {
2115 smpl_buf_addr = ctx->ctx_smpl_hdr;
2116 smpl_buf_size = ctx->ctx_smpl_size;
2117 /* no more sampling */
2118 ctx->ctx_smpl_hdr = NULL;
2119 ctx->ctx_fl_is_sampling = 0;
2120 }
2121
2122 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2123 state,
2124 free_possible,
2125 smpl_buf_addr,
2126 smpl_buf_size));
2127
2128 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2129
2130 /*
2131 * UNLOADED that the session has already been unreserved.
2132 */
2133 if (state == PFM_CTX_ZOMBIE) {
2134 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2135 }
2136
2137 /*
2138 * disconnect file descriptor from context must be done
2139 * before we unlock.
2140 */
2141 filp->private_data = NULL;
2142
2143 /*
2144 * if we free on the spot, the context is now completely unreachable
2145 * from the callers side. The monitored task side is also cut, so we
2146 * can freely cut.
2147 *
2148 * If we have a deferred free, only the caller side is disconnected.
2149 */
2150 UNPROTECT_CTX(ctx, flags);
2151
2152 /*
2153 * All memory free operations (especially for vmalloc'ed memory)
2154 * MUST be done with interrupts ENABLED.
2155 */
2156 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2157
2158 /*
2159 * return the memory used by the context
2160 */
2161 if (free_possible) pfm_context_free(ctx);
2162
2163 return 0;
2164 }
2165
2166 static int
2167 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2168 {
2169 DPRINT(("pfm_no_open called\n"));
2170 return -ENXIO;
2171 }
2172
2173
2174
2175 static const struct file_operations pfm_file_ops = {
2176 .llseek = no_llseek,
2177 .read = pfm_read,
2178 .write = pfm_write,
2179 .poll = pfm_poll,
2180 .ioctl = pfm_ioctl,
2181 .open = pfm_no_open, /* special open code to disallow open via /proc */
2182 .fasync = pfm_fasync,
2183 .release = pfm_close,
2184 .flush = pfm_flush
2185 };
2186
2187 static int
2188 pfmfs_delete_dentry(struct dentry *dentry)
2189 {
2190 return 1;
2191 }
2192
2193 static const struct dentry_operations pfmfs_dentry_operations = {
2194 .d_delete = pfmfs_delete_dentry,
2195 };
2196
2197
2198 static struct file *
2199 pfm_alloc_file(pfm_context_t *ctx)
2200 {
2201 struct file *file;
2202 struct inode *inode;
2203 struct dentry *dentry;
2204 char name[32];
2205 struct qstr this;
2206
2207 /*
2208 * allocate a new inode
2209 */
2210 inode = new_inode(pfmfs_mnt->mnt_sb);
2211 if (!inode)
2212 return ERR_PTR(-ENOMEM);
2213
2214 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2215
2216 inode->i_mode = S_IFCHR|S_IRUGO;
2217 inode->i_uid = current_fsuid();
2218 inode->i_gid = current_fsgid();
2219
2220 sprintf(name, "[%lu]", inode->i_ino);
2221 this.name = name;
2222 this.len = strlen(name);
2223 this.hash = inode->i_ino;
2224
2225 /*
2226 * allocate a new dcache entry
2227 */
2228 dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2229 if (!dentry) {
2230 iput(inode);
2231 return ERR_PTR(-ENOMEM);
2232 }
2233
2234 dentry->d_op = &pfmfs_dentry_operations;
2235 d_add(dentry, inode);
2236
2237 file = alloc_file(pfmfs_mnt, dentry, FMODE_READ, &pfm_file_ops);
2238 if (!file) {
2239 dput(dentry);
2240 return ERR_PTR(-ENFILE);
2241 }
2242
2243 file->f_flags = O_RDONLY;
2244 file->private_data = ctx;
2245
2246 return file;
2247 }
2248
2249 static int
2250 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2251 {
2252 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2253
2254 while (size > 0) {
2255 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2256
2257
2258 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2259 return -ENOMEM;
2260
2261 addr += PAGE_SIZE;
2262 buf += PAGE_SIZE;
2263 size -= PAGE_SIZE;
2264 }
2265 return 0;
2266 }
2267
2268 /*
2269 * allocate a sampling buffer and remaps it into the user address space of the task
2270 */
2271 static int
2272 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2273 {
2274 struct mm_struct *mm = task->mm;
2275 struct vm_area_struct *vma = NULL;
2276 unsigned long size;
2277 void *smpl_buf;
2278
2279
2280 /*
2281 * the fixed header + requested size and align to page boundary
2282 */
2283 size = PAGE_ALIGN(rsize);
2284
2285 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2286
2287 /*
2288 * check requested size to avoid Denial-of-service attacks
2289 * XXX: may have to refine this test
2290 * Check against address space limit.
2291 *
2292 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2293 * return -ENOMEM;
2294 */
2295 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2296 return -ENOMEM;
2297
2298 /*
2299 * We do the easy to undo allocations first.
2300 *
2301 * pfm_rvmalloc(), clears the buffer, so there is no leak
2302 */
2303 smpl_buf = pfm_rvmalloc(size);
2304 if (smpl_buf == NULL) {
2305 DPRINT(("Can't allocate sampling buffer\n"));
2306 return -ENOMEM;
2307 }
2308
2309 DPRINT(("smpl_buf @%p\n", smpl_buf));
2310
2311 /* allocate vma */
2312 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2313 if (!vma) {
2314 DPRINT(("Cannot allocate vma\n"));
2315 goto error_kmem;
2316 }
2317
2318 /*
2319 * partially initialize the vma for the sampling buffer
2320 */
2321 vma->vm_mm = mm;
2322 vma->vm_file = filp;
2323 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2324 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2325
2326 /*
2327 * Now we have everything we need and we can initialize
2328 * and connect all the data structures
2329 */
2330
2331 ctx->ctx_smpl_hdr = smpl_buf;
2332 ctx->ctx_smpl_size = size; /* aligned size */
2333
2334 /*
2335 * Let's do the difficult operations next.
2336 *
2337 * now we atomically find some area in the address space and
2338 * remap the buffer in it.
2339 */
2340 down_write(&task->mm->mmap_sem);
2341
2342 /* find some free area in address space, must have mmap sem held */
2343 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2344 if (vma->vm_start == 0UL) {
2345 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2346 up_write(&task->mm->mmap_sem);
2347 goto error;
2348 }
2349 vma->vm_end = vma->vm_start + size;
2350 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2351
2352 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2353
2354 /* can only be applied to current task, need to have the mm semaphore held when called */
2355 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2356 DPRINT(("Can't remap buffer\n"));
2357 up_write(&task->mm->mmap_sem);
2358 goto error;
2359 }
2360
2361 get_file(filp);
2362
2363 /*
2364 * now insert the vma in the vm list for the process, must be
2365 * done with mmap lock held
2366 */
2367 insert_vm_struct(mm, vma);
2368
2369 mm->total_vm += size >> PAGE_SHIFT;
2370 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2371 vma_pages(vma));
2372 up_write(&task->mm->mmap_sem);
2373
2374 /*
2375 * keep track of user level virtual address
2376 */
2377 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2378 *(unsigned long *)user_vaddr = vma->vm_start;
2379
2380 return 0;
2381
2382 error:
2383 kmem_cache_free(vm_area_cachep, vma);
2384 error_kmem:
2385 pfm_rvfree(smpl_buf, size);
2386
2387 return -ENOMEM;
2388 }
2389
2390 /*
2391 * XXX: do something better here
2392 */
2393 static int
2394 pfm_bad_permissions(struct task_struct *task)
2395 {
2396 const struct cred *tcred;
2397 uid_t uid = current_uid();
2398 gid_t gid = current_gid();
2399 int ret;
2400
2401 rcu_read_lock();
2402 tcred = __task_cred(task);
2403
2404 /* inspired by ptrace_attach() */
2405 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2406 uid,
2407 gid,
2408 tcred->euid,
2409 tcred->suid,
2410 tcred->uid,
2411 tcred->egid,
2412 tcred->sgid));
2413
2414 ret = ((uid != tcred->euid)
2415 || (uid != tcred->suid)
2416 || (uid != tcred->uid)
2417 || (gid != tcred->egid)
2418 || (gid != tcred->sgid)
2419 || (gid != tcred->gid)) && !capable(CAP_SYS_PTRACE);
2420
2421 rcu_read_unlock();
2422 return ret;
2423 }
2424
2425 static int
2426 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2427 {
2428 int ctx_flags;
2429
2430 /* valid signal */
2431
2432 ctx_flags = pfx->ctx_flags;
2433
2434 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2435
2436 /*
2437 * cannot block in this mode
2438 */
2439 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2440 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2441 return -EINVAL;
2442 }
2443 } else {
2444 }
2445 /* probably more to add here */
2446
2447 return 0;
2448 }
2449
2450 static int
2451 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2452 unsigned int cpu, pfarg_context_t *arg)
2453 {
2454 pfm_buffer_fmt_t *fmt = NULL;
2455 unsigned long size = 0UL;
2456 void *uaddr = NULL;
2457 void *fmt_arg = NULL;
2458 int ret = 0;
2459 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2460
2461 /* invoke and lock buffer format, if found */
2462 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2463 if (fmt == NULL) {
2464 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2465 return -EINVAL;
2466 }
2467
2468 /*
2469 * buffer argument MUST be contiguous to pfarg_context_t
2470 */
2471 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2472
2473 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2474
2475 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2476
2477 if (ret) goto error;
2478
2479 /* link buffer format and context */
2480 ctx->ctx_buf_fmt = fmt;
2481 ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2482
2483 /*
2484 * check if buffer format wants to use perfmon buffer allocation/mapping service
2485 */
2486 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2487 if (ret) goto error;
2488
2489 if (size) {
2490 /*
2491 * buffer is always remapped into the caller's address space
2492 */
2493 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2494 if (ret) goto error;
2495
2496 /* keep track of user address of buffer */
2497 arg->ctx_smpl_vaddr = uaddr;
2498 }
2499 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2500
2501 error:
2502 return ret;
2503 }
2504
2505 static void
2506 pfm_reset_pmu_state(pfm_context_t *ctx)
2507 {
2508 int i;
2509
2510 /*
2511 * install reset values for PMC.
2512 */
2513 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2514 if (PMC_IS_IMPL(i) == 0) continue;
2515 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2516 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2517 }
2518 /*
2519 * PMD registers are set to 0UL when the context in memset()
2520 */
2521
2522 /*
2523 * On context switched restore, we must restore ALL pmc and ALL pmd even
2524 * when they are not actively used by the task. In UP, the incoming process
2525 * may otherwise pick up left over PMC, PMD state from the previous process.
2526 * As opposed to PMD, stale PMC can cause harm to the incoming
2527 * process because they may change what is being measured.
2528 * Therefore, we must systematically reinstall the entire
2529 * PMC state. In SMP, the same thing is possible on the
2530 * same CPU but also on between 2 CPUs.
2531 *
2532 * The problem with PMD is information leaking especially
2533 * to user level when psr.sp=0
2534 *
2535 * There is unfortunately no easy way to avoid this problem
2536 * on either UP or SMP. This definitively slows down the
2537 * pfm_load_regs() function.
2538 */
2539
2540 /*
2541 * bitmask of all PMCs accessible to this context
2542 *
2543 * PMC0 is treated differently.
2544 */
2545 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2546
2547 /*
2548 * bitmask of all PMDs that are accessible to this context
2549 */
2550 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2551
2552 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2553
2554 /*
2555 * useful in case of re-enable after disable
2556 */
2557 ctx->ctx_used_ibrs[0] = 0UL;
2558 ctx->ctx_used_dbrs[0] = 0UL;
2559 }
2560
2561 static int
2562 pfm_ctx_getsize(void *arg, size_t *sz)
2563 {
2564 pfarg_context_t *req = (pfarg_context_t *)arg;
2565 pfm_buffer_fmt_t *fmt;
2566
2567 *sz = 0;
2568
2569 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2570
2571 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2572 if (fmt == NULL) {
2573 DPRINT(("cannot find buffer format\n"));
2574 return -EINVAL;
2575 }
2576 /* get just enough to copy in user parameters */
2577 *sz = fmt->fmt_arg_size;
2578 DPRINT(("arg_size=%lu\n", *sz));
2579
2580 return 0;
2581 }
2582
2583
2584
2585 /*
2586 * cannot attach if :
2587 * - kernel task
2588 * - task not owned by caller
2589 * - task incompatible with context mode
2590 */
2591 static int
2592 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2593 {
2594 /*
2595 * no kernel task or task not owner by caller
2596 */
2597 if (task->mm == NULL) {
2598 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2599 return -EPERM;
2600 }
2601 if (pfm_bad_permissions(task)) {
2602 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task)));
2603 return -EPERM;
2604 }
2605 /*
2606 * cannot block in self-monitoring mode
2607 */
2608 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2609 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2610 return -EINVAL;
2611 }
2612
2613 if (task->exit_state == EXIT_ZOMBIE) {
2614 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task)));
2615 return -EBUSY;
2616 }
2617
2618 /*
2619 * always ok for self
2620 */
2621 if (task == current) return 0;
2622
2623 if (!task_is_stopped_or_traced(task)) {
2624 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2625 return -EBUSY;
2626 }
2627 /*
2628 * make sure the task is off any CPU
2629 */
2630 wait_task_inactive(task, 0);
2631
2632 /* more to come... */
2633
2634 return 0;
2635 }
2636
2637 static int
2638 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2639 {
2640 struct task_struct *p = current;
2641 int ret;
2642
2643 /* XXX: need to add more checks here */
2644 if (pid < 2) return -EPERM;
2645
2646 if (pid != task_pid_vnr(current)) {
2647
2648 read_lock(&tasklist_lock);
2649
2650 p = find_task_by_vpid(pid);
2651
2652 /* make sure task cannot go away while we operate on it */
2653 if (p) get_task_struct(p);
2654
2655 read_unlock(&tasklist_lock);
2656
2657 if (p == NULL) return -ESRCH;
2658 }
2659
2660 ret = pfm_task_incompatible(ctx, p);
2661 if (ret == 0) {
2662 *task = p;
2663 } else if (p != current) {
2664 pfm_put_task(p);
2665 }
2666 return ret;
2667 }
2668
2669
2670
2671 static int
2672 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2673 {
2674 pfarg_context_t *req = (pfarg_context_t *)arg;
2675 struct file *filp;
2676 struct path path;
2677 int ctx_flags;
2678 int fd;
2679 int ret;
2680
2681 /* let's check the arguments first */
2682 ret = pfarg_is_sane(current, req);
2683 if (ret < 0)
2684 return ret;
2685
2686 ctx_flags = req->ctx_flags;
2687
2688 ret = -ENOMEM;
2689
2690 fd = get_unused_fd();
2691 if (fd < 0)
2692 return fd;
2693
2694 ctx = pfm_context_alloc(ctx_flags);
2695 if (!ctx)
2696 goto error;
2697
2698 filp = pfm_alloc_file(ctx);
2699 if (IS_ERR(filp)) {
2700 ret = PTR_ERR(filp);
2701 goto error_file;
2702 }
2703
2704 req->ctx_fd = ctx->ctx_fd = fd;
2705
2706 /*
2707 * does the user want to sample?
2708 */
2709 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2710 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2711 if (ret)
2712 goto buffer_error;
2713 }
2714
2715 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2716 ctx,
2717 ctx_flags,
2718 ctx->ctx_fl_system,
2719 ctx->ctx_fl_block,
2720 ctx->ctx_fl_excl_idle,
2721 ctx->ctx_fl_no_msg,
2722 ctx->ctx_fd));
2723
2724 /*
2725 * initialize soft PMU state
2726 */
2727 pfm_reset_pmu_state(ctx);
2728
2729 fd_install(fd, filp);
2730
2731 return 0;
2732
2733 buffer_error:
2734 path = filp->f_path;
2735 put_filp(filp);
2736 path_put(&path);
2737
2738 if (ctx->ctx_buf_fmt) {
2739 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2740 }
2741 error_file:
2742 pfm_context_free(ctx);
2743
2744 error:
2745 put_unused_fd(fd);
2746 return ret;
2747 }
2748
2749 static inline unsigned long
2750 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2751 {
2752 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2753 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2754 extern unsigned long carta_random32 (unsigned long seed);
2755
2756 if (reg->flags & PFM_REGFL_RANDOM) {
2757 new_seed = carta_random32(old_seed);
2758 val -= (old_seed & mask); /* counter values are negative numbers! */
2759 if ((mask >> 32) != 0)
2760 /* construct a full 64-bit random value: */
2761 new_seed |= carta_random32(old_seed >> 32) << 32;
2762 reg->seed = new_seed;
2763 }
2764 reg->lval = val;
2765 return val;
2766 }
2767
2768 static void
2769 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2770 {
2771 unsigned long mask = ovfl_regs[0];
2772 unsigned long reset_others = 0UL;
2773 unsigned long val;
2774 int i;
2775
2776 /*
2777 * now restore reset value on sampling overflowed counters
2778 */
2779 mask >>= PMU_FIRST_COUNTER;
2780 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2781
2782 if ((mask & 0x1UL) == 0UL) continue;
2783
2784 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2785 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2786
2787 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2788 }
2789
2790 /*
2791 * Now take care of resetting the other registers
2792 */
2793 for(i = 0; reset_others; i++, reset_others >>= 1) {
2794
2795 if ((reset_others & 0x1) == 0) continue;
2796
2797 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2798
2799 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2800 is_long_reset ? "long" : "short", i, val));
2801 }
2802 }
2803
2804 static void
2805 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2806 {
2807 unsigned long mask = ovfl_regs[0];
2808 unsigned long reset_others = 0UL;
2809 unsigned long val;
2810 int i;
2811
2812 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2813
2814 if (ctx->ctx_state == PFM_CTX_MASKED) {
2815 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2816 return;
2817 }
2818
2819 /*
2820 * now restore reset value on sampling overflowed counters
2821 */
2822 mask >>= PMU_FIRST_COUNTER;
2823 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2824
2825 if ((mask & 0x1UL) == 0UL) continue;
2826
2827 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2828 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2829
2830 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2831
2832 pfm_write_soft_counter(ctx, i, val);
2833 }
2834
2835 /*
2836 * Now take care of resetting the other registers
2837 */
2838 for(i = 0; reset_others; i++, reset_others >>= 1) {
2839
2840 if ((reset_others & 0x1) == 0) continue;
2841
2842 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2843
2844 if (PMD_IS_COUNTING(i)) {
2845 pfm_write_soft_counter(ctx, i, val);
2846 } else {
2847 ia64_set_pmd(i, val);
2848 }
2849 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2850 is_long_reset ? "long" : "short", i, val));
2851 }
2852 ia64_srlz_d();
2853 }
2854
2855 static int
2856 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2857 {
2858 struct task_struct *task;
2859 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2860 unsigned long value, pmc_pm;
2861 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2862 unsigned int cnum, reg_flags, flags, pmc_type;
2863 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2864 int is_monitor, is_counting, state;
2865 int ret = -EINVAL;
2866 pfm_reg_check_t wr_func;
2867 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2868
2869 state = ctx->ctx_state;
2870 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2871 is_system = ctx->ctx_fl_system;
2872 task = ctx->ctx_task;
2873 impl_pmds = pmu_conf->impl_pmds[0];
2874
2875 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2876
2877 if (is_loaded) {
2878 /*
2879 * In system wide and when the context is loaded, access can only happen
2880 * when the caller is running on the CPU being monitored by the session.
2881 * It does not have to be the owner (ctx_task) of the context per se.
2882 */
2883 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2884 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2885 return -EBUSY;
2886 }
2887 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2888 }
2889 expert_mode = pfm_sysctl.expert_mode;
2890
2891 for (i = 0; i < count; i++, req++) {
2892
2893 cnum = req->reg_num;
2894 reg_flags = req->reg_flags;
2895 value = req->reg_value;
2896 smpl_pmds = req->reg_smpl_pmds[0];
2897 reset_pmds = req->reg_reset_pmds[0];
2898 flags = 0;
2899
2900
2901 if (cnum >= PMU_MAX_PMCS) {
2902 DPRINT(("pmc%u is invalid\n", cnum));
2903 goto error;
2904 }
2905
2906 pmc_type = pmu_conf->pmc_desc[cnum].type;
2907 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2908 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2909 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2910
2911 /*
2912 * we reject all non implemented PMC as well
2913 * as attempts to modify PMC[0-3] which are used
2914 * as status registers by the PMU
2915 */
2916 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2917 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2918 goto error;
2919 }
2920 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2921 /*
2922 * If the PMC is a monitor, then if the value is not the default:
2923 * - system-wide session: PMCx.pm=1 (privileged monitor)
2924 * - per-task : PMCx.pm=0 (user monitor)
2925 */
2926 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2927 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2928 cnum,
2929 pmc_pm,
2930 is_system));
2931 goto error;
2932 }
2933
2934 if (is_counting) {
2935 /*
2936 * enforce generation of overflow interrupt. Necessary on all
2937 * CPUs.
2938 */
2939 value |= 1 << PMU_PMC_OI;
2940
2941 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2942 flags |= PFM_REGFL_OVFL_NOTIFY;
2943 }
2944
2945 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2946
2947 /* verify validity of smpl_pmds */
2948 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2949 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2950 goto error;
2951 }
2952
2953 /* verify validity of reset_pmds */
2954 if ((reset_pmds & impl_pmds) != reset_pmds) {
2955 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2956 goto error;
2957 }
2958 } else {
2959 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2960 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2961 goto error;
2962 }
2963 /* eventid on non-counting monitors are ignored */
2964 }
2965
2966 /*
2967 * execute write checker, if any
2968 */
2969 if (likely(expert_mode == 0 && wr_func)) {
2970 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2971 if (ret) goto error;
2972 ret = -EINVAL;
2973 }
2974
2975 /*
2976 * no error on this register
2977 */
2978 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2979
2980 /*
2981 * Now we commit the changes to the software state
2982 */
2983
2984 /*
2985 * update overflow information
2986 */
2987 if (is_counting) {
2988 /*
2989 * full flag update each time a register is programmed
2990 */
2991 ctx->ctx_pmds[cnum].flags = flags;
2992
2993 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2994 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2995 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2996
2997 /*
2998 * Mark all PMDS to be accessed as used.
2999 *
3000 * We do not keep track of PMC because we have to
3001 * systematically restore ALL of them.
3002 *
3003 * We do not update the used_monitors mask, because
3004 * if we have not programmed them, then will be in
3005 * a quiescent state, therefore we will not need to
3006 * mask/restore then when context is MASKED.
3007 */
3008 CTX_USED_PMD(ctx, reset_pmds);
3009 CTX_USED_PMD(ctx, smpl_pmds);
3010 /*
3011 * make sure we do not try to reset on
3012 * restart because we have established new values
3013 */
3014 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3015 }
3016 /*
3017 * Needed in case the user does not initialize the equivalent
3018 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3019 * possible leak here.
3020 */
3021 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3022
3023 /*
3024 * keep track of the monitor PMC that we are using.
3025 * we save the value of the pmc in ctx_pmcs[] and if
3026 * the monitoring is not stopped for the context we also
3027 * place it in the saved state area so that it will be
3028 * picked up later by the context switch code.
3029 *
3030 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3031 *
3032 * The value in th_pmcs[] may be modified on overflow, i.e., when
3033 * monitoring needs to be stopped.
3034 */
3035 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3036
3037 /*
3038 * update context state
3039 */
3040 ctx->ctx_pmcs[cnum] = value;
3041
3042 if (is_loaded) {
3043 /*
3044 * write thread state
3045 */
3046 if (is_system == 0) ctx->th_pmcs[cnum] = value;
3047
3048 /*
3049 * write hardware register if we can
3050 */
3051 if (can_access_pmu) {
3052 ia64_set_pmc(cnum, value);
3053 }
3054 #ifdef CONFIG_SMP
3055 else {
3056 /*
3057 * per-task SMP only here
3058 *
3059 * we are guaranteed that the task is not running on the other CPU,
3060 * we indicate that this PMD will need to be reloaded if the task
3061 * is rescheduled on the CPU it ran last on.
3062 */
3063 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3064 }
3065 #endif
3066 }
3067
3068 DPRINT(("pmc[%u]=0x%lx ld=%d apmu=%d flags=0x%x all_pmcs=0x%lx used_pmds=0x%lx eventid=%ld smpl_pmds=0x%lx reset_pmds=0x%lx reloads_pmcs=0x%lx used_monitors=0x%lx ovfl_regs=0x%lx\n",
3069 cnum,
3070 value,
3071 is_loaded,
3072 can_access_pmu,
3073 flags,
3074 ctx->ctx_all_pmcs[0],
3075 ctx->ctx_used_pmds[0],
3076 ctx->ctx_pmds[cnum].eventid,
3077 smpl_pmds,
3078 reset_pmds,
3079 ctx->ctx_reload_pmcs[0],
3080 ctx->ctx_used_monitors[0],
3081 ctx->ctx_ovfl_regs[0]));
3082 }
3083
3084 /*
3085 * make sure the changes are visible
3086 */
3087 if (can_access_pmu) ia64_srlz_d();
3088
3089 return 0;
3090 error:
3091 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3092 return ret;
3093 }
3094
3095 static int
3096 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3097 {
3098 struct task_struct *task;
3099 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3100 unsigned long value, hw_value, ovfl_mask;
3101 unsigned int cnum;
3102 int i, can_access_pmu = 0, state;
3103 int is_counting, is_loaded, is_system, expert_mode;
3104 int ret = -EINVAL;
3105 pfm_reg_check_t wr_func;
3106
3107
3108 state = ctx->ctx_state;
3109 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3110 is_system = ctx->ctx_fl_system;
3111 ovfl_mask = pmu_conf->ovfl_val;
3112 task = ctx->ctx_task;
3113
3114 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3115
3116 /*
3117 * on both UP and SMP, we can only write to the PMC when the task is
3118 * the owner of the local PMU.
3119 */
3120 if (likely(is_loaded)) {
3121 /*
3122 * In system wide and when the context is loaded, access can only happen
3123 * when the caller is running on the CPU being monitored by the session.
3124 * It does not have to be the owner (ctx_task) of the context per se.
3125 */
3126 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3127 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3128 return -EBUSY;
3129 }
3130 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3131 }
3132 expert_mode = pfm_sysctl.expert_mode;
3133
3134 for (i = 0; i < count; i++, req++) {
3135
3136 cnum = req->reg_num;
3137 value = req->reg_value;
3138
3139 if (!PMD_IS_IMPL(cnum)) {
3140 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3141 goto abort_mission;
3142 }
3143 is_counting = PMD_IS_COUNTING(cnum);
3144 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3145
3146 /*
3147 * execute write checker, if any
3148 */
3149 if (unlikely(expert_mode == 0 && wr_func)) {
3150 unsigned long v = value;
3151
3152 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3153 if (ret) goto abort_mission;
3154
3155 value = v;
3156 ret = -EINVAL;
3157 }
3158
3159 /*
3160 * no error on this register
3161 */
3162 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3163
3164 /*
3165 * now commit changes to software state
3166 */
3167 hw_value = value;
3168
3169 /*
3170 * update virtualized (64bits) counter
3171 */
3172 if (is_counting) {
3173 /*
3174 * write context state
3175 */
3176 ctx->ctx_pmds[cnum].lval = value;
3177
3178 /*
3179 * when context is load we use the split value
3180 */
3181 if (is_loaded) {
3182 hw_value = value & ovfl_mask;
3183 value = value & ~ovfl_mask;
3184 }
3185 }
3186 /*
3187 * update reset values (not just for counters)
3188 */
3189 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3190 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3191
3192 /*
3193 * update randomization parameters (not just for counters)
3194 */
3195 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3196 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3197
3198 /*
3199 * update context value
3200 */
3201 ctx->ctx_pmds[cnum].val = value;
3202
3203 /*
3204 * Keep track of what we use
3205 *
3206 * We do not keep track of PMC because we have to
3207 * systematically restore ALL of them.
3208 */
3209 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3210
3211 /*
3212 * mark this PMD register used as well
3213 */
3214 CTX_USED_PMD(ctx, RDEP(cnum));
3215
3216 /*
3217 * make sure we do not try to reset on
3218 * restart because we have established new values
3219 */
3220 if (is_counting && state == PFM_CTX_MASKED) {
3221 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3222 }
3223
3224 if (is_loaded) {
3225 /*
3226 * write thread state
3227 */
3228 if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3229
3230 /*
3231 * write hardware register if we can
3232 */
3233 if (can_access_pmu) {
3234 ia64_set_pmd(cnum, hw_value);
3235 } else {
3236 #ifdef CONFIG_SMP
3237 /*
3238 * we are guaranteed that the task is not running on the other CPU,
3239 * we indicate that this PMD will need to be reloaded if the task
3240 * is rescheduled on the CPU it ran last on.
3241 */
3242 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3243 #endif
3244 }
3245 }
3246
3247 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3248 "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",
3249 cnum,
3250 value,
3251 is_loaded,
3252 can_access_pmu,
3253 hw_value,
3254 ctx->ctx_pmds[cnum].val,
3255 ctx->ctx_pmds[cnum].short_reset,
3256 ctx->ctx_pmds[cnum].long_reset,
3257 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3258 ctx->ctx_pmds[cnum].seed,
3259 ctx->ctx_pmds[cnum].mask,
3260 ctx->ctx_used_pmds[0],
3261 ctx->ctx_pmds[cnum].reset_pmds[0],
3262 ctx->ctx_reload_pmds[0],
3263 ctx->ctx_all_pmds[0],
3264 ctx->ctx_ovfl_regs[0]));
3265 }
3266
3267 /*
3268 * make changes visible
3269 */
3270 if (can_access_pmu) ia64_srlz_d();
3271
3272 return 0;
3273
3274 abort_mission:
3275 /*
3276 * for now, we have only one possibility for error
3277 */
3278 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3279 return ret;
3280 }
3281
3282 /*
3283 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3284 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3285 * interrupt is delivered during the call, it will be kept pending until we leave, making
3286 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3287 * guaranteed to return consistent data to the user, it may simply be old. It is not
3288 * trivial to treat the overflow while inside the call because you may end up in
3289 * some module sampling buffer code causing deadlocks.
3290 */
3291 static int
3292 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3293 {
3294 struct task_struct *task;
3295 unsigned long val = 0UL, lval, ovfl_mask, sval;
3296 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3297 unsigned int cnum, reg_flags = 0;
3298 int i, can_access_pmu = 0, state;
3299 int is_loaded, is_system, is_counting, expert_mode;
3300 int ret = -EINVAL;
3301 pfm_reg_check_t rd_func;
3302
3303 /*
3304 * access is possible when loaded only for
3305 * self-monitoring tasks or in UP mode
3306 */
3307
3308 state = ctx->ctx_state;
3309 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3310 is_system = ctx->ctx_fl_system;
3311 ovfl_mask = pmu_conf->ovfl_val;
3312 task = ctx->ctx_task;
3313
3314 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3315
3316 if (likely(is_loaded)) {
3317 /*
3318 * In system wide and when the context is loaded, access can only happen
3319 * when the caller is running on the CPU being monitored by the session.
3320 * It does not have to be the owner (ctx_task) of the context per se.
3321 */
3322 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3323 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3324 return -EBUSY;
3325 }
3326 /*
3327 * this can be true when not self-monitoring only in UP
3328 */
3329 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3330
3331 if (can_access_pmu) ia64_srlz_d();
3332 }
3333 expert_mode = pfm_sysctl.expert_mode;
3334
3335 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3336 is_loaded,
3337 can_access_pmu,
3338 state));
3339
3340 /*
3341 * on both UP and SMP, we can only read the PMD from the hardware register when
3342 * the task is the owner of the local PMU.
3343 */
3344
3345 for (i = 0; i < count; i++, req++) {
3346
3347 cnum = req->reg_num;
3348 reg_flags = req->reg_flags;
3349
3350 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3351 /*
3352 * we can only read the register that we use. That includes
3353 * the one we explicitly initialize AND the one we want included
3354 * in the sampling buffer (smpl_regs).
3355 *
3356 * Having this restriction allows optimization in the ctxsw routine
3357 * without compromising security (leaks)
3358 */
3359 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3360
3361 sval = ctx->ctx_pmds[cnum].val;
3362 lval = ctx->ctx_pmds[cnum].lval;
3363 is_counting = PMD_IS_COUNTING(cnum);
3364
3365 /*
3366 * If the task is not the current one, then we check if the
3367 * PMU state is still in the local live register due to lazy ctxsw.
3368 * If true, then we read directly from the registers.
3369 */
3370 if (can_access_pmu){
3371 val = ia64_get_pmd(cnum);
3372 } else {
3373 /*
3374 * context has been saved
3375 * if context is zombie, then task does not exist anymore.
3376 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3377 */
3378 val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3379 }
3380 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3381
3382 if (is_counting) {
3383 /*
3384 * XXX: need to check for overflow when loaded
3385 */
3386 val &= ovfl_mask;
3387 val += sval;
3388 }
3389
3390 /*
3391 * execute read checker, if any
3392 */
3393 if (unlikely(expert_mode == 0 && rd_func)) {
3394 unsigned long v = val;
3395 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3396 if (ret) goto error;
3397 val = v;
3398 ret = -EINVAL;
3399 }
3400
3401 PFM_REG_RETFLAG_SET(reg_flags, 0);
3402
3403 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3404
3405 /*
3406 * update register return value, abort all if problem during copy.
3407 * we only modify the reg_flags field. no check mode is fine because
3408 * access has been verified upfront in sys_perfmonctl().
3409 */
3410 req->reg_value = val;
3411 req->reg_flags = reg_flags;
3412 req->reg_last_reset_val = lval;
3413 }
3414
3415 return 0;
3416
3417 error:
3418 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3419 return ret;
3420 }
3421
3422 int
3423 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3424 {
3425 pfm_context_t *ctx;
3426
3427 if (req == NULL) return -EINVAL;
3428
3429 ctx = GET_PMU_CTX();
3430
3431 if (ctx == NULL) return -EINVAL;
3432
3433 /*
3434 * for now limit to current task, which is enough when calling
3435 * from overflow handler
3436 */
3437 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3438
3439 return pfm_write_pmcs(ctx, req, nreq, regs);
3440 }
3441 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3442
3443 int
3444 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3445 {
3446 pfm_context_t *ctx;
3447
3448 if (req == NULL) return -EINVAL;
3449
3450 ctx = GET_PMU_CTX();
3451
3452 if (ctx == NULL) return -EINVAL;
3453
3454 /*
3455 * for now limit to current task, which is enough when calling
3456 * from overflow handler
3457 */
3458 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3459
3460 return pfm_read_pmds(ctx, req, nreq, regs);
3461 }
3462 EXPORT_SYMBOL(pfm_mod_read_pmds);
3463
3464 /*
3465 * Only call this function when a process it trying to
3466 * write the debug registers (reading is always allowed)
3467 */
3468 int
3469 pfm_use_debug_registers(struct task_struct *task)
3470 {
3471 pfm_context_t *ctx = task->thread.pfm_context;
3472 unsigned long flags;
3473 int ret = 0;
3474
3475 if (pmu_conf->use_rr_dbregs == 0) return 0;
3476
3477 DPRINT(("called for [%d]\n", task_pid_nr(task)));
3478
3479 /*
3480 * do it only once
3481 */
3482 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3483
3484 /*
3485 * Even on SMP, we do not need to use an atomic here because
3486 * the only way in is via ptrace() and this is possible only when the
3487 * process is stopped. Even in the case where the ctxsw out is not totally
3488 * completed by the time we come here, there is no way the 'stopped' process
3489 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3490 * So this is always safe.
3491 */
3492 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3493
3494 LOCK_PFS(flags);
3495
3496 /*
3497 * We cannot allow setting breakpoints when system wide monitoring
3498 * sessions are using the debug registers.
3499 */
3500 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3501 ret = -1;
3502 else
3503 pfm_sessions.pfs_ptrace_use_dbregs++;
3504
3505 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3506 pfm_sessions.pfs_ptrace_use_dbregs,
3507 pfm_sessions.pfs_sys_use_dbregs,
3508 task_pid_nr(task), ret));
3509
3510 UNLOCK_PFS(flags);
3511
3512 return ret;
3513 }
3514
3515 /*
3516 * This function is called for every task that exits with the
3517 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3518 * able to use the debug registers for debugging purposes via
3519 * ptrace(). Therefore we know it was not using them for
3520 * performance monitoring, so we only decrement the number
3521 * of "ptraced" debug register users to keep the count up to date
3522 */
3523 int
3524 pfm_release_debug_registers(struct task_struct *task)
3525 {
3526 unsigned long flags;
3527 int ret;
3528
3529 if (pmu_conf->use_rr_dbregs == 0) return 0;
3530
3531 LOCK_PFS(flags);
3532 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3533 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3534 ret = -1;
3535 } else {
3536 pfm_sessions.pfs_ptrace_use_dbregs--;
3537 ret = 0;
3538 }
3539 UNLOCK_PFS(flags);
3540
3541 return ret;
3542 }
3543
3544 static int
3545 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3546 {
3547 struct task_struct *task;
3548 pfm_buffer_fmt_t *fmt;
3549 pfm_ovfl_ctrl_t rst_ctrl;
3550 int state, is_system;
3551 int ret = 0;
3552
3553 state = ctx->ctx_state;
3554 fmt = ctx->ctx_buf_fmt;
3555 is_system = ctx->ctx_fl_system;
3556 task = PFM_CTX_TASK(ctx);
3557
3558 switch(state) {
3559 case PFM_CTX_MASKED:
3560 break;
3561 case PFM_CTX_LOADED:
3562 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3563 /* fall through */
3564 case PFM_CTX_UNLOADED:
3565 case PFM_CTX_ZOMBIE:
3566 DPRINT(("invalid state=%d\n", state));
3567 return -EBUSY;
3568 default:
3569 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3570 return -EINVAL;
3571 }
3572
3573 /*
3574 * In system wide and when the context is loaded, access can only happen
3575 * when the caller is running on the CPU being monitored by the session.
3576 * It does not have to be the owner (ctx_task) of the context per se.
3577 */
3578 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3579 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3580 return -EBUSY;
3581 }
3582
3583 /* sanity check */
3584 if (unlikely(task == NULL)) {
3585 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3586 return -EINVAL;
3587 }
3588
3589 if (task == current || is_system) {
3590
3591 fmt = ctx->ctx_buf_fmt;
3592
3593 DPRINT(("restarting self %d ovfl=0x%lx\n",
3594 task_pid_nr(task),
3595 ctx->ctx_ovfl_regs[0]));
3596
3597 if (CTX_HAS_SMPL(ctx)) {
3598
3599 prefetch(ctx->ctx_smpl_hdr);
3600
3601 rst_ctrl.bits.mask_monitoring = 0;
3602 rst_ctrl.bits.reset_ovfl_pmds = 0;
3603
3604 if (state == PFM_CTX_LOADED)
3605 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3606 else
3607 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3608 } else {
3609 rst_ctrl.bits.mask_monitoring = 0;
3610 rst_ctrl.bits.reset_ovfl_pmds = 1;
3611 }
3612
3613 if (ret == 0) {
3614 if (rst_ctrl.bits.reset_ovfl_pmds)
3615 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3616
3617 if (rst_ctrl.bits.mask_monitoring == 0) {
3618 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3619
3620 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3621 } else {
3622 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3623
3624 // cannot use pfm_stop_monitoring(task, regs);
3625 }
3626 }
3627 /*
3628 * clear overflowed PMD mask to remove any stale information
3629 */
3630 ctx->ctx_ovfl_regs[0] = 0UL;
3631
3632 /*
3633 * back to LOADED state
3634 */
3635 ctx->ctx_state = PFM_CTX_LOADED;
3636
3637 /*
3638 * XXX: not really useful for self monitoring
3639 */
3640 ctx->ctx_fl_can_restart = 0;
3641
3642 return 0;
3643 }
3644
3645 /*
3646 * restart another task
3647 */
3648
3649 /*
3650 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3651 * one is seen by the task.
3652 */
3653 if (state == PFM_CTX_MASKED) {
3654 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3655 /*
3656 * will prevent subsequent restart before this one is
3657 * seen by other task
3658 */
3659 ctx->ctx_fl_can_restart = 0;
3660 }
3661
3662 /*
3663 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3664 * the task is blocked or on its way to block. That's the normal
3665 * restart path. If the monitoring is not masked, then the task
3666 * can be actively monitoring and we cannot directly intervene.
3667 * Therefore we use the trap mechanism to catch the task and
3668 * force it to reset the buffer/reset PMDs.
3669 *
3670 * if non-blocking, then we ensure that the task will go into
3671 * pfm_handle_work() before returning to user mode.
3672 *
3673 * We cannot explicitly reset another task, it MUST always
3674 * be done by the task itself. This works for system wide because
3675 * the tool that is controlling the session is logically doing
3676 * "self-monitoring".
3677 */
3678 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3679 DPRINT(("unblocking [%d] \n", task_pid_nr(task)));
3680 complete(&ctx->ctx_restart_done);
3681 } else {
3682 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3683
3684 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3685
3686 PFM_SET_WORK_PENDING(task, 1);
3687
3688 set_notify_resume(task);
3689
3690 /*
3691 * XXX: send reschedule if task runs on another CPU
3692 */
3693 }
3694 return 0;
3695 }
3696
3697 static int
3698 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3699 {
3700 unsigned int m = *(unsigned int *)arg;
3701
3702 pfm_sysctl.debug = m == 0 ? 0 : 1;
3703
3704 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3705
3706 if (m == 0) {
3707 memset(pfm_stats, 0, sizeof(pfm_stats));
3708 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3709 }
3710 return 0;
3711 }
3712
3713 /*
3714 * arg can be NULL and count can be zero for this function
3715 */
3716 static int
3717 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3718 {
3719 struct thread_struct *thread = NULL;
3720 struct task_struct *task;
3721 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3722 unsigned long flags;
3723 dbreg_t dbreg;
3724 unsigned int rnum;
3725 int first_time;
3726 int ret = 0, state;
3727 int i, can_access_pmu = 0;
3728 int is_system, is_loaded;
3729
3730 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3731
3732 state = ctx->ctx_state;
3733 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3734 is_system = ctx->ctx_fl_system;
3735 task = ctx->ctx_task;
3736
3737 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3738
3739 /*
3740 * on both UP and SMP, we can only write to the PMC when the task is
3741 * the owner of the local PMU.
3742 */
3743 if (is_loaded) {
3744 thread = &task->thread;
3745 /*
3746 * In system wide and when the context is loaded, access can only happen
3747 * when the caller is running on the CPU being monitored by the session.
3748 * It does not have to be the owner (ctx_task) of the context per se.
3749 */
3750 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3751 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3752 return -EBUSY;
3753 }
3754 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3755 }
3756
3757 /*
3758 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3759 * ensuring that no real breakpoint can be installed via this call.
3760 *
3761 * IMPORTANT: regs can be NULL in this function
3762 */
3763
3764 first_time = ctx->ctx_fl_using_dbreg == 0;
3765
3766 /*
3767 * don't bother if we are loaded and task is being debugged
3768 */
3769 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3770 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3771 return -EBUSY;
3772 }
3773
3774 /*
3775 * check for debug registers in system wide mode
3776 *
3777 * If though a check is done in pfm_context_load(),
3778 * we must repeat it here, in case the registers are
3779 * written after the context is loaded
3780 */
3781 if (is_loaded) {
3782 LOCK_PFS(flags);
3783
3784 if (first_time && is_system) {
3785 if (pfm_sessions.pfs_ptrace_use_dbregs)
3786 ret = -EBUSY;
3787 else
3788 pfm_sessions.pfs_sys_use_dbregs++;
3789 }
3790 UNLOCK_PFS(flags);
3791 }
3792
3793 if (ret != 0) return ret;
3794
3795 /*
3796 * mark ourself as user of the debug registers for
3797 * perfmon purposes.
3798 */
3799 ctx->ctx_fl_using_dbreg = 1;
3800
3801 /*
3802 * clear hardware registers to make sure we don't
3803 * pick up stale state.
3804 *
3805 * for a system wide session, we do not use
3806 * thread.dbr, thread.ibr because this process
3807 * never leaves the current CPU and the state
3808 * is shared by all processes running on it
3809 */
3810 if (first_time && can_access_pmu) {
3811 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3812 for (i=0; i < pmu_conf->num_ibrs; i++) {
3813 ia64_set_ibr(i, 0UL);
3814 ia64_dv_serialize_instruction();
3815 }
3816 ia64_srlz_i();
3817 for (i=0; i < pmu_conf->num_dbrs; i++) {
3818 ia64_set_dbr(i, 0UL);
3819 ia64_dv_serialize_data();
3820 }
3821 ia64_srlz_d();
3822 }
3823
3824 /*
3825 * Now install the values into the registers
3826 */
3827 for (i = 0; i < count; i++, req++) {
3828
3829 rnum = req->dbreg_num;
3830 dbreg.val = req->dbreg_value;
3831
3832 ret = -EINVAL;
3833
3834 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3835 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3836 rnum, dbreg.val, mode, i, count));
3837
3838 goto abort_mission;
3839 }
3840
3841 /*
3842 * make sure we do not install enabled breakpoint
3843 */
3844 if (rnum & 0x1) {
3845 if (mode == PFM_CODE_RR)
3846 dbreg.ibr.ibr_x = 0;
3847 else
3848 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3849 }
3850
3851 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3852
3853 /*
3854 * Debug registers, just like PMC, can only be modified
3855 * by a kernel call. Moreover, perfmon() access to those
3856 * registers are centralized in this routine. The hardware
3857 * does not modify the value of these registers, therefore,
3858 * if we save them as they are written, we can avoid having
3859 * to save them on context switch out. This is made possible
3860 * by the fact that when perfmon uses debug registers, ptrace()
3861 * won't be able to modify them concurrently.
3862 */
3863 if (mode == PFM_CODE_RR) {
3864 CTX_USED_IBR(ctx, rnum);
3865
3866 if (can_access_pmu) {
3867 ia64_set_ibr(rnum, dbreg.val);
3868 ia64_dv_serialize_instruction();
3869 }
3870
3871 ctx->ctx_ibrs[rnum] = dbreg.val;
3872
3873 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3874 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3875 } else {
3876 CTX_USED_DBR(ctx, rnum);
3877
3878 if (can_access_pmu) {
3879 ia64_set_dbr(rnum, dbreg.val);
3880 ia64_dv_serialize_data();
3881 }
3882 ctx->ctx_dbrs[rnum] = dbreg.val;
3883
3884 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3885 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3886 }
3887 }
3888
3889 return 0;
3890
3891 abort_mission:
3892 /*
3893 * in case it was our first attempt, we undo the global modifications
3894 */
3895 if (first_time) {
3896 LOCK_PFS(flags);
3897 if (ctx->ctx_fl_system) {
3898 pfm_sessions.pfs_sys_use_dbregs--;
3899 }
3900 UNLOCK_PFS(flags);
3901 ctx->ctx_fl_using_dbreg = 0;
3902 }
3903 /*
3904 * install error return flag
3905 */
3906 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3907
3908 return ret;
3909 }
3910
3911 static int
3912 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3913 {
3914 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3915 }
3916
3917 static int
3918 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3919 {
3920 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3921 }
3922
3923 int
3924 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3925 {
3926 pfm_context_t *ctx;
3927
3928 if (req == NULL) return -EINVAL;
3929
3930 ctx = GET_PMU_CTX();
3931
3932 if (ctx == NULL) return -EINVAL;
3933
3934 /*
3935 * for now limit to current task, which is enough when calling
3936 * from overflow handler
3937 */
3938 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3939
3940 return pfm_write_ibrs(ctx, req, nreq, regs);
3941 }
3942 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3943
3944 int
3945 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3946 {
3947 pfm_context_t *ctx;
3948
3949 if (req == NULL) return -EINVAL;
3950
3951 ctx = GET_PMU_CTX();
3952
3953 if (ctx == NULL) return -EINVAL;
3954
3955 /*
3956 * for now limit to current task, which is enough when calling
3957 * from overflow handler
3958 */
3959 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3960
3961 return pfm_write_dbrs(ctx, req, nreq, regs);
3962 }
3963 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3964
3965
3966 static int
3967 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3968 {
3969 pfarg_features_t *req = (pfarg_features_t *)arg;
3970
3971 req->ft_version = PFM_VERSION;
3972 return 0;
3973 }
3974
3975 static int
3976 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3977 {
3978 struct pt_regs *tregs;
3979 struct task_struct *task = PFM_CTX_TASK(ctx);
3980 int state, is_system;
3981
3982 state = ctx->ctx_state;
3983 is_system = ctx->ctx_fl_system;
3984
3985 /*
3986 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3987 */
3988 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3989
3990 /*
3991 * In system wide and when the context is loaded, access can only happen
3992 * when the caller is running on the CPU being monitored by the session.
3993 * It does not have to be the owner (ctx_task) of the context per se.
3994 */
3995 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3996 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3997 return -EBUSY;
3998 }
3999 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4000 task_pid_nr(PFM_CTX_TASK(ctx)),
4001 state,
4002 is_system));
4003 /*
4004 * in system mode, we need to update the PMU directly
4005 * and the user level state of the caller, which may not
4006 * necessarily be the creator of the context.
4007 */
4008 if (is_system) {
4009 /*
4010 * Update local PMU first
4011 *
4012 * disable dcr pp
4013 */
4014 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4015 ia64_srlz_i();
4016
4017 /*
4018 * update local cpuinfo
4019 */
4020 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4021
4022 /*
4023 * stop monitoring, does srlz.i
4024 */
4025 pfm_clear_psr_pp();
4026
4027 /*
4028 * stop monitoring in the caller
4029 */
4030 ia64_psr(regs)->pp = 0;
4031
4032 return 0;
4033 }
4034 /*
4035 * per-task mode
4036 */
4037
4038 if (task == current) {
4039 /* stop monitoring at kernel level */
4040 pfm_clear_psr_up();
4041
4042 /*
4043 * stop monitoring at the user level
4044 */
4045 ia64_psr(regs)->up = 0;
4046 } else {
4047 tregs = task_pt_regs(task);
4048
4049 /*
4050 * stop monitoring at the user level
4051 */
4052 ia64_psr(tregs)->up = 0;
4053
4054 /*
4055 * monitoring disabled in kernel at next reschedule
4056 */
4057 ctx->ctx_saved_psr_up = 0;
4058 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4059 }
4060 return 0;
4061 }
4062
4063
4064 static int
4065 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4066 {
4067 struct pt_regs *tregs;
4068 int state, is_system;
4069
4070 state = ctx->ctx_state;
4071 is_system = ctx->ctx_fl_system;
4072
4073 if (state != PFM_CTX_LOADED) return -EINVAL;
4074
4075 /*
4076 * In system wide and when the context is loaded, access can only happen
4077 * when the caller is running on the CPU being monitored by the session.
4078 * It does not have to be the owner (ctx_task) of the context per se.
4079 */
4080 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4081 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4082 return -EBUSY;
4083 }
4084
4085 /*
4086 * in system mode, we need to update the PMU directly
4087 * and the user level state of the caller, which may not
4088 * necessarily be the creator of the context.
4089 */
4090 if (is_system) {
4091
4092 /*
4093 * set user level psr.pp for the caller
4094 */
4095 ia64_psr(regs)->pp = 1;
4096
4097 /*
4098 * now update the local PMU and cpuinfo
4099 */
4100 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4101
4102 /*
4103 * start monitoring at kernel level
4104 */
4105 pfm_set_psr_pp();
4106
4107 /* enable dcr pp */
4108 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4109 ia64_srlz_i();
4110
4111 return 0;
4112 }
4113
4114 /*
4115 * per-process mode
4116 */
4117
4118 if (ctx->ctx_task == current) {
4119
4120 /* start monitoring at kernel level */
4121 pfm_set_psr_up();
4122
4123 /*
4124 * activate monitoring at user level
4125 */
4126 ia64_psr(regs)->up = 1;
4127
4128 } else {
4129 tregs = task_pt_regs(ctx->ctx_task);
4130
4131 /*
4132 * start monitoring at the kernel level the next
4133 * time the task is scheduled
4134 */
4135 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4136
4137 /*
4138 * activate monitoring at user level
4139 */
4140 ia64_psr(tregs)->up = 1;
4141 }
4142 return 0;
4143 }
4144
4145 static int
4146 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4147 {
4148 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4149 unsigned int cnum;
4150 int i;
4151 int ret = -EINVAL;
4152
4153 for (i = 0; i < count; i++, req++) {
4154
4155 cnum = req->reg_num;
4156
4157 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4158
4159 req->reg_value = PMC_DFL_VAL(cnum);
4160
4161 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4162
4163 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4164 }
4165 return 0;
4166
4167 abort_mission:
4168 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4169 return ret;
4170 }
4171
4172 static int
4173 pfm_check_task_exist(pfm_context_t *ctx)
4174 {
4175 struct task_struct *g, *t;
4176 int ret = -ESRCH;
4177
4178 read_lock(&tasklist_lock);
4179
4180 do_each_thread (g, t) {
4181 if (t->thread.pfm_context == ctx) {
4182 ret = 0;
4183 goto out;
4184 }
4185 } while_each_thread (g, t);
4186 out:
4187 read_unlock(&tasklist_lock);
4188
4189 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4190
4191 return ret;
4192 }
4193
4194 static int
4195 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4196 {
4197 struct task_struct *task;
4198 struct thread_struct *thread;
4199 struct pfm_context_t *old;
4200 unsigned long flags;
4201 #ifndef CONFIG_SMP
4202 struct task_struct *owner_task = NULL;
4203 #endif
4204 pfarg_load_t *req = (pfarg_load_t *)arg;
4205 unsigned long *pmcs_source, *pmds_source;
4206 int the_cpu;
4207 int ret = 0;
4208 int state, is_system, set_dbregs = 0;
4209
4210 state = ctx->ctx_state;
4211 is_system = ctx->ctx_fl_system;
4212 /*
4213 * can only load from unloaded or terminated state
4214 */
4215 if (state != PFM_CTX_UNLOADED) {
4216 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4217 req->load_pid,
4218 ctx->ctx_state));
4219 return -EBUSY;
4220 }
4221
4222 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4223
4224 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4225 DPRINT(("cannot use blocking mode on self\n"));
4226 return -EINVAL;
4227 }
4228
4229 ret = pfm_get_task(ctx, req->load_pid, &task);
4230 if (ret) {
4231 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4232 return ret;
4233 }
4234
4235 ret = -EINVAL;
4236
4237 /*
4238 * system wide is self monitoring only
4239 */
4240 if (is_system && task != current) {
4241 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4242 req->load_pid));
4243 goto error;
4244 }
4245
4246 thread = &task->thread;
4247
4248 ret = 0;
4249 /*
4250 * cannot load a context which is using range restrictions,
4251 * into a task that is being debugged.
4252 */
4253 if (ctx->ctx_fl_using_dbreg) {
4254 if (thread->flags & IA64_THREAD_DBG_VALID) {
4255 ret = -EBUSY;
4256 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4257 goto error;
4258 }
4259 LOCK_PFS(flags);
4260
4261 if (is_system) {
4262 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4263 DPRINT(("cannot load [%d] dbregs in use\n",
4264 task_pid_nr(task)));
4265 ret = -EBUSY;
4266 } else {
4267 pfm_sessions.pfs_sys_use_dbregs++;
4268 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4269 set_dbregs = 1;
4270 }
4271 }
4272
4273 UNLOCK_PFS(flags);
4274
4275 if (ret) goto error;
4276 }
4277
4278 /*
4279 * SMP system-wide monitoring implies self-monitoring.
4280 *
4281 * The programming model expects the task to
4282 * be pinned on a CPU throughout the session.
4283 * Here we take note of the current CPU at the
4284 * time the context is loaded. No call from
4285 * another CPU will be allowed.
4286 *
4287 * The pinning via shed_setaffinity()
4288 * must be done by the calling task prior
4289 * to this call.
4290 *
4291 * systemwide: keep track of CPU this session is supposed to run on
4292 */
4293 the_cpu = ctx->ctx_cpu = smp_processor_id();
4294
4295 ret = -EBUSY;
4296 /*
4297 * now reserve the session
4298 */
4299 ret = pfm_reserve_session(current, is_system, the_cpu);
4300 if (ret) goto error;
4301
4302 /*
4303 * task is necessarily stopped at this point.
4304 *
4305 * If the previous context was zombie, then it got removed in
4306 * pfm_save_regs(). Therefore we should not see it here.
4307 * If we see a context, then this is an active context
4308 *
4309 * XXX: needs to be atomic
4310 */
4311 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4312 thread->pfm_context, ctx));
4313
4314 ret = -EBUSY;
4315 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4316 if (old != NULL) {
4317 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4318 goto error_unres;
4319 }
4320
4321 pfm_reset_msgq(ctx);
4322
4323 ctx->ctx_state = PFM_CTX_LOADED;
4324
4325 /*
4326 * link context to task
4327 */
4328 ctx->ctx_task = task;
4329
4330 if (is_system) {
4331 /*
4332 * we load as stopped
4333 */
4334 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4335 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4336
4337 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4338 } else {
4339 thread->flags |= IA64_THREAD_PM_VALID;
4340 }
4341
4342 /*
4343 * propagate into thread-state
4344 */
4345 pfm_copy_pmds(task, ctx);
4346 pfm_copy_pmcs(task, ctx);
4347
4348 pmcs_source = ctx->th_pmcs;
4349 pmds_source = ctx->th_pmds;
4350
4351 /*
4352 * always the case for system-wide
4353 */
4354 if (task == current) {
4355
4356 if (is_system == 0) {
4357
4358 /* allow user level control */
4359 ia64_psr(regs)->sp = 0;
4360 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4361
4362 SET_LAST_CPU(ctx, smp_processor_id());
4363 INC_ACTIVATION();
4364 SET_ACTIVATION(ctx);
4365 #ifndef CONFIG_SMP
4366 /*
4367 * push the other task out, if any
4368 */
4369 owner_task = GET_PMU_OWNER();
4370 if (owner_task) pfm_lazy_save_regs(owner_task);
4371 #endif
4372 }
4373 /*
4374 * load all PMD from ctx to PMU (as opposed to thread state)
4375 * restore all PMC from ctx to PMU
4376 */
4377 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4378 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4379
4380 ctx->ctx_reload_pmcs[0] = 0UL;
4381 ctx->ctx_reload_pmds[0] = 0UL;
4382
4383 /*
4384 * guaranteed safe by earlier check against DBG_VALID
4385 */
4386 if (ctx->ctx_fl_using_dbreg) {
4387 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4388 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4389 }
4390 /*
4391 * set new ownership
4392 */
4393 SET_PMU_OWNER(task, ctx);
4394
4395 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4396 } else {
4397 /*
4398 * when not current, task MUST be stopped, so this is safe
4399 */
4400 regs = task_pt_regs(task);
4401
4402 /* force a full reload */
4403 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4404 SET_LAST_CPU(ctx, -1);
4405
4406 /* initial saved psr (stopped) */
4407 ctx->ctx_saved_psr_up = 0UL;
4408 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4409 }
4410
4411 ret = 0;
4412
4413 error_unres:
4414 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4415 error:
4416 /*
4417 * we must undo the dbregs setting (for system-wide)
4418 */
4419 if (ret && set_dbregs) {
4420 LOCK_PFS(flags);
4421 pfm_sessions.pfs_sys_use_dbregs--;
4422 UNLOCK_PFS(flags);
4423 }
4424 /*
4425 * release task, there is now a link with the context
4426 */
4427 if (is_system == 0 && task != current) {
4428 pfm_put_task(task);
4429
4430 if (ret == 0) {
4431 ret = pfm_check_task_exist(ctx);
4432 if (ret) {
4433 ctx->ctx_state = PFM_CTX_UNLOADED;
4434 ctx->ctx_task = NULL;
4435 }
4436 }
4437 }
4438 return ret;
4439 }
4440
4441 /*
4442 * in this function, we do not need to increase the use count
4443 * for the task via get_task_struct(), because we hold the
4444 * context lock. If the task were to disappear while having
4445 * a context attached, it would go through pfm_exit_thread()
4446 * which also grabs the context lock and would therefore be blocked
4447 * until we are here.
4448 */
4449 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4450
4451 static int
4452 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4453 {
4454 struct task_struct *task = PFM_CTX_TASK(ctx);
4455 struct pt_regs *tregs;
4456 int prev_state, is_system;
4457 int ret;
4458
4459 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4460
4461 prev_state = ctx->ctx_state;
4462 is_system = ctx->ctx_fl_system;
4463
4464 /*
4465 * unload only when necessary
4466 */
4467 if (prev_state == PFM_CTX_UNLOADED) {
4468 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4469 return 0;
4470 }
4471
4472 /*
4473 * clear psr and dcr bits
4474 */
4475 ret = pfm_stop(ctx, NULL, 0, regs);
4476 if (ret) return ret;
4477
4478 ctx->ctx_state = PFM_CTX_UNLOADED;
4479
4480 /*
4481 * in system mode, we need to update the PMU directly
4482 * and the user level state of the caller, which may not
4483 * necessarily be the creator of the context.
4484 */
4485 if (is_system) {
4486
4487 /*
4488 * Update cpuinfo
4489 *
4490 * local PMU is taken care of in pfm_stop()
4491 */
4492 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4493 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4494
4495 /*
4496 * save PMDs in context
4497 * release ownership
4498 */
4499 pfm_flush_pmds(current, ctx);
4500
4501 /*
4502 * at this point we are done with the PMU
4503 * so we can unreserve the resource.
4504 */
4505 if (prev_state != PFM_CTX_ZOMBIE)
4506 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4507
4508 /*
4509 * disconnect context from task
4510 */
4511 task->thread.pfm_context = NULL;
4512 /*
4513 * disconnect task from context
4514 */
4515 ctx->ctx_task = NULL;
4516
4517 /*
4518 * There is nothing more to cleanup here.
4519 */
4520 return 0;
4521 }
4522
4523 /*
4524 * per-task mode
4525 */
4526 tregs = task == current ? regs : task_pt_regs(task);
4527
4528 if (task == current) {
4529 /*
4530 * cancel user level control
4531 */
4532 ia64_psr(regs)->sp = 1;
4533
4534 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4535 }
4536 /*
4537 * save PMDs to context
4538 * release ownership
4539 */
4540 pfm_flush_pmds(task, ctx);
4541
4542 /*
4543 * at this point we are done with the PMU
4544 * so we can unreserve the resource.
4545 *
4546 * when state was ZOMBIE, we have already unreserved.
4547 */
4548 if (prev_state != PFM_CTX_ZOMBIE)
4549 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4550
4551 /*
4552 * reset activation counter and psr
4553 */
4554 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4555 SET_LAST_CPU(ctx, -1);
4556
4557 /*
4558 * PMU state will not be restored
4559 */
4560 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4561
4562 /*
4563 * break links between context and task
4564 */
4565 task->thread.pfm_context = NULL;
4566 ctx->ctx_task = NULL;
4567
4568 PFM_SET_WORK_PENDING(task, 0);
4569
4570 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4571 ctx->ctx_fl_can_restart = 0;
4572 ctx->ctx_fl_going_zombie = 0;
4573
4574 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4575
4576 return 0;
4577 }
4578
4579
4580 /*
4581 * called only from exit_thread(): task == current
4582 * we come here only if current has a context attached (loaded or masked)
4583 */
4584 void
4585 pfm_exit_thread(struct task_struct *task)
4586 {
4587 pfm_context_t *ctx;
4588 unsigned long flags;
4589 struct pt_regs *regs = task_pt_regs(task);
4590 int ret, state;
4591 int free_ok = 0;
4592
4593 ctx = PFM_GET_CTX(task);
4594
4595 PROTECT_CTX(ctx, flags);
4596
4597 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4598
4599 state = ctx->ctx_state;
4600 switch(state) {
4601 case PFM_CTX_UNLOADED:
4602 /*
4603 * only comes to this function if pfm_context is not NULL, i.e., cannot
4604 * be in unloaded state
4605 */
4606 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4607 break;
4608 case PFM_CTX_LOADED:
4609 case PFM_CTX_MASKED:
4610 ret = pfm_context_unload(ctx, NULL, 0, regs);
4611 if (ret) {
4612 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4613 }
4614 DPRINT(("ctx unloaded for current state was %d\n", state));
4615
4616 pfm_end_notify_user(ctx);
4617 break;
4618 case PFM_CTX_ZOMBIE:
4619 ret = pfm_context_unload(ctx, NULL, 0, regs);
4620 if (ret) {
4621 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4622 }
4623 free_ok = 1;
4624 break;
4625 default:
4626 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4627 break;
4628 }
4629 UNPROTECT_CTX(ctx, flags);
4630
4631 { u64 psr = pfm_get_psr();
4632 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4633 BUG_ON(GET_PMU_OWNER());
4634 BUG_ON(ia64_psr(regs)->up);
4635 BUG_ON(ia64_psr(regs)->pp);
4636 }
4637
4638 /*
4639 * All memory free operations (especially for vmalloc'ed memory)
4640 * MUST be done with interrupts ENABLED.
4641 */
4642 if (free_ok) pfm_context_free(ctx);
4643 }
4644
4645 /*
4646 * functions MUST be listed in the increasing order of their index (see permfon.h)
4647 */
4648 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4649 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4650 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4651 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4652 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4653
4654 static pfm_cmd_desc_t pfm_cmd_tab[]={
4655 /* 0 */PFM_CMD_NONE,
4656 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4657 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4658 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4659 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4660 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4661 /* 6 */PFM_CMD_NONE,
4662 /* 7 */PFM_CMD_NONE,
4663 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4664 /* 9 */PFM_CMD_NONE,
4665 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4666 /* 11 */PFM_CMD_NONE,
4667 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4668 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4669 /* 14 */PFM_CMD_NONE,
4670 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4671 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4672 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4673 /* 18 */PFM_CMD_NONE,
4674 /* 19 */PFM_CMD_NONE,
4675 /* 20 */PFM_CMD_NONE,
4676 /* 21 */PFM_CMD_NONE,
4677 /* 22 */PFM_CMD_NONE,
4678 /* 23 */PFM_CMD_NONE,
4679 /* 24 */PFM_CMD_NONE,
4680 /* 25 */PFM_CMD_NONE,
4681 /* 26 */PFM_CMD_NONE,
4682 /* 27 */PFM_CMD_NONE,
4683 /* 28 */PFM_CMD_NONE,
4684 /* 29 */PFM_CMD_NONE,
4685 /* 30 */PFM_CMD_NONE,
4686 /* 31 */PFM_CMD_NONE,
4687 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4688 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4689 };
4690 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4691
4692 static int
4693 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4694 {
4695 struct task_struct *task;
4696 int state, old_state;
4697
4698 recheck:
4699 state = ctx->ctx_state;
4700 task = ctx->ctx_task;
4701
4702 if (task == NULL) {
4703 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4704 return 0;
4705 }
4706
4707 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4708 ctx->ctx_fd,
4709 state,
4710 task_pid_nr(task),
4711 task->state, PFM_CMD_STOPPED(cmd)));
4712
4713 /*
4714 * self-monitoring always ok.
4715 *
4716 * for system-wide the caller can either be the creator of the
4717 * context (to one to which the context is attached to) OR
4718 * a task running on the same CPU as the session.
4719 */
4720 if (task == current || ctx->ctx_fl_system) return 0;
4721
4722 /*
4723 * we are monitoring another thread
4724 */
4725 switch(state) {
4726 case PFM_CTX_UNLOADED:
4727 /*
4728 * if context is UNLOADED we are safe to go
4729 */
4730 return 0;
4731 case PFM_CTX_ZOMBIE:
4732 /*
4733 * no command can operate on a zombie context
4734 */
4735 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4736 return -EINVAL;
4737 case PFM_CTX_MASKED:
4738 /*
4739 * PMU state has been saved to software even though
4740 * the thread may still be running.
4741 */
4742 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4743 }
4744
4745 /*
4746 * context is LOADED or MASKED. Some commands may need to have
4747 * the task stopped.
4748 *
4749 * We could lift this restriction for UP but it would mean that
4750 * the user has no guarantee the task would not run between
4751 * two successive calls to perfmonctl(). That's probably OK.
4752 * If this user wants to ensure the task does not run, then
4753 * the task must be stopped.
4754 */
4755 if (PFM_CMD_STOPPED(cmd)) {
4756 if (!task_is_stopped_or_traced(task)) {
4757 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4758 return -EBUSY;
4759 }
4760 /*
4761 * task is now stopped, wait for ctxsw out
4762 *
4763 * This is an interesting point in the code.
4764 * We need to unprotect the context because
4765 * the pfm_save_regs() routines needs to grab
4766 * the same lock. There are danger in doing
4767 * this because it leaves a window open for
4768 * another task to get access to the context
4769 * and possibly change its state. The one thing
4770 * that is not possible is for the context to disappear
4771 * because we are protected by the VFS layer, i.e.,
4772 * get_fd()/put_fd().
4773 */
4774 old_state = state;
4775
4776 UNPROTECT_CTX(ctx, flags);
4777
4778 wait_task_inactive(task, 0);
4779
4780 PROTECT_CTX(ctx, flags);
4781
4782 /*
4783 * we must recheck to verify if state has changed
4784 */
4785 if (ctx->ctx_state != old_state) {
4786 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4787 goto recheck;
4788 }
4789 }
4790 return 0;
4791 }
4792
4793 /*
4794 * system-call entry point (must return long)
4795 */
4796 asmlinkage long
4797 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4798 {
4799 struct file *file = NULL;
4800 pfm_context_t *ctx = NULL;
4801 unsigned long flags = 0UL;
4802 void *args_k = NULL;
4803 long ret; /* will expand int return types */
4804 size_t base_sz, sz, xtra_sz = 0;
4805 int narg, completed_args = 0, call_made = 0, cmd_flags;
4806 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4807 int (*getsize)(void *arg, size_t *sz);
4808 #define PFM_MAX_ARGSIZE 4096
4809
4810 /*
4811 * reject any call if perfmon was disabled at initialization
4812 */
4813 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4814
4815 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4816 DPRINT(("invalid cmd=%d\n", cmd));
4817 return -EINVAL;
4818 }
4819
4820 func = pfm_cmd_tab[cmd].cmd_func;
4821 narg = pfm_cmd_tab[cmd].cmd_narg;
4822 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4823 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4824 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4825
4826 if (unlikely(func == NULL)) {
4827 DPRINT(("invalid cmd=%d\n", cmd));
4828 return -EINVAL;
4829 }
4830
4831 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4832 PFM_CMD_NAME(cmd),
4833 cmd,
4834 narg,
4835 base_sz,
4836 count));
4837
4838 /*
4839 * check if number of arguments matches what the command expects
4840 */
4841 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4842 return -EINVAL;
4843
4844 restart_args:
4845 sz = xtra_sz + base_sz*count;
4846 /*
4847 * limit abuse to min page size
4848 */
4849 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4850 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4851 return -E2BIG;
4852 }
4853
4854 /*
4855 * allocate default-sized argument buffer
4856 */
4857 if (likely(count && args_k == NULL)) {
4858 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4859 if (args_k == NULL) return -ENOMEM;
4860 }
4861
4862 ret = -EFAULT;
4863
4864 /*
4865 * copy arguments
4866 *
4867 * assume sz = 0 for command without parameters
4868 */
4869 if (sz && copy_from_user(args_k, arg, sz)) {
4870 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4871 goto error_args;
4872 }
4873
4874 /*
4875 * check if command supports extra parameters
4876 */
4877 if (completed_args == 0 && getsize) {
4878 /*
4879 * get extra parameters size (based on main argument)
4880 */
4881 ret = (*getsize)(args_k, &xtra_sz);
4882 if (ret) goto error_args;
4883
4884 completed_args = 1;
4885
4886 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4887
4888 /* retry if necessary */
4889 if (likely(xtra_sz)) goto restart_args;
4890 }
4891
4892 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4893
4894 ret = -EBADF;
4895
4896 file = fget(fd);
4897 if (unlikely(file == NULL)) {
4898 DPRINT(("invalid fd %d\n", fd));
4899 goto error_args;
4900 }
4901 if (unlikely(PFM_IS_FILE(file) == 0)) {
4902 DPRINT(("fd %d not related to perfmon\n", fd));
4903 goto error_args;
4904 }
4905
4906 ctx = (pfm_context_t *)file->private_data;
4907 if (unlikely(ctx == NULL)) {
4908 DPRINT(("no context for fd %d\n", fd));
4909 goto error_args;
4910 }
4911 prefetch(&ctx->ctx_state);
4912
4913 PROTECT_CTX(ctx, flags);
4914
4915 /*
4916 * check task is stopped
4917 */
4918 ret = pfm_check_task_state(ctx, cmd, flags);
4919 if (unlikely(ret)) goto abort_locked;
4920
4921 skip_fd:
4922 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4923
4924 call_made = 1;
4925
4926 abort_locked:
4927 if (likely(ctx)) {
4928 DPRINT(("context unlocked\n"));
4929 UNPROTECT_CTX(ctx, flags);
4930 }
4931
4932 /* copy argument back to user, if needed */
4933 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4934
4935 error_args:
4936 if (file)
4937 fput(file);
4938
4939 kfree(args_k);
4940
4941 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4942
4943 return ret;
4944 }
4945
4946 static void
4947 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4948 {
4949 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4950 pfm_ovfl_ctrl_t rst_ctrl;
4951 int state;
4952 int ret = 0;
4953
4954 state = ctx->ctx_state;
4955 /*
4956 * Unlock sampling buffer and reset index atomically
4957 * XXX: not really needed when blocking
4958 */
4959 if (CTX_HAS_SMPL(ctx)) {
4960
4961 rst_ctrl.bits.mask_monitoring = 0;
4962 rst_ctrl.bits.reset_ovfl_pmds = 0;
4963
4964 if (state == PFM_CTX_LOADED)
4965 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4966 else
4967 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4968 } else {
4969 rst_ctrl.bits.mask_monitoring = 0;
4970 rst_ctrl.bits.reset_ovfl_pmds = 1;
4971 }
4972
4973 if (ret == 0) {
4974 if (rst_ctrl.bits.reset_ovfl_pmds) {
4975 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4976 }
4977 if (rst_ctrl.bits.mask_monitoring == 0) {
4978 DPRINT(("resuming monitoring\n"));
4979 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4980 } else {
4981 DPRINT(("stopping monitoring\n"));
4982 //pfm_stop_monitoring(current, regs);
4983 }
4984 ctx->ctx_state = PFM_CTX_LOADED;
4985 }
4986 }
4987
4988 /*
4989 * context MUST BE LOCKED when calling
4990 * can only be called for current
4991 */
4992 static void
4993 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4994 {
4995 int ret;
4996
4997 DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4998
4999 ret = pfm_context_unload(ctx, NULL, 0, regs);
5000 if (ret) {
5001 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
5002 }
5003
5004 /*
5005 * and wakeup controlling task, indicating we are now disconnected
5006 */
5007 wake_up_interruptible(&ctx->ctx_zombieq);
5008
5009 /*
5010 * given that context is still locked, the controlling
5011 * task will only get access when we return from
5012 * pfm_handle_work().
5013 */
5014 }
5015
5016 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5017
5018 /*
5019 * pfm_handle_work() can be called with interrupts enabled
5020 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5021 * call may sleep, therefore we must re-enable interrupts
5022 * to avoid deadlocks. It is safe to do so because this function
5023 * is called ONLY when returning to user level (pUStk=1), in which case
5024 * there is no risk of kernel stack overflow due to deep
5025 * interrupt nesting.
5026 */
5027 void
5028 pfm_handle_work(void)
5029 {
5030 pfm_context_t *ctx;
5031 struct pt_regs *regs;
5032 unsigned long flags, dummy_flags;
5033 unsigned long ovfl_regs;
5034 unsigned int reason;
5035 int ret;
5036
5037 ctx = PFM_GET_CTX(current);
5038 if (ctx == NULL) {
5039 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5040 task_pid_nr(current));
5041 return;
5042 }
5043
5044 PROTECT_CTX(ctx, flags);
5045
5046 PFM_SET_WORK_PENDING(current, 0);
5047
5048 regs = task_pt_regs(current);
5049
5050 /*
5051 * extract reason for being here and clear
5052 */
5053 reason = ctx->ctx_fl_trap_reason;
5054 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5055 ovfl_regs = ctx->ctx_ovfl_regs[0];
5056
5057 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5058
5059 /*
5060 * must be done before we check for simple-reset mode
5061 */
5062 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5063 goto do_zombie;
5064
5065 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5066 if (reason == PFM_TRAP_REASON_RESET)
5067 goto skip_blocking;
5068
5069 /*
5070 * restore interrupt mask to what it was on entry.
5071 * Could be enabled/diasbled.
5072 */
5073 UNPROTECT_CTX(ctx, flags);
5074
5075 /*
5076 * force interrupt enable because of down_interruptible()
5077 */
5078 local_irq_enable();
5079
5080 DPRINT(("before block sleeping\n"));
5081
5082 /*
5083 * may go through without blocking on SMP systems
5084 * if restart has been received already by the time we call down()
5085 */
5086 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5087
5088 DPRINT(("after block sleeping ret=%d\n", ret));
5089
5090 /*
5091 * lock context and mask interrupts again
5092 * We save flags into a dummy because we may have
5093 * altered interrupts mask compared to entry in this
5094 * function.
5095 */
5096 PROTECT_CTX(ctx, dummy_flags);
5097
5098 /*
5099 * we need to read the ovfl_regs only after wake-up
5100 * because we may have had pfm_write_pmds() in between
5101 * and that can changed PMD values and therefore
5102 * ovfl_regs is reset for these new PMD values.
5103 */
5104 ovfl_regs = ctx->ctx_ovfl_regs[0];
5105
5106 if (ctx->ctx_fl_going_zombie) {
5107 do_zombie:
5108 DPRINT(("context is zombie, bailing out\n"));
5109 pfm_context_force_terminate(ctx, regs);
5110 goto nothing_to_do;
5111 }
5112 /*
5113 * in case of interruption of down() we don't restart anything
5114 */
5115 if (ret < 0)
5116 goto nothing_to_do;
5117
5118 skip_blocking:
5119 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5120 ctx->ctx_ovfl_regs[0] = 0UL;
5121
5122 nothing_to_do:
5123 /*
5124 * restore flags as they were upon entry
5125 */
5126 UNPROTECT_CTX(ctx, flags);
5127 }
5128
5129 static int
5130 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5131 {
5132 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5133 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5134 return 0;
5135 }
5136
5137 DPRINT(("waking up somebody\n"));
5138
5139 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5140
5141 /*
5142 * safe, we are not in intr handler, nor in ctxsw when
5143 * we come here
5144 */
5145 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5146
5147 return 0;
5148 }
5149
5150 static int
5151 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5152 {
5153 pfm_msg_t *msg = NULL;
5154
5155 if (ctx->ctx_fl_no_msg == 0) {
5156 msg = pfm_get_new_msg(ctx);
5157 if (msg == NULL) {
5158 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5159 return -1;
5160 }
5161
5162 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5163 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5164 msg->pfm_ovfl_msg.msg_active_set = 0;
5165 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5166 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5167 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5168 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5169 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5170 }
5171
5172 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5173 msg,
5174 ctx->ctx_fl_no_msg,
5175 ctx->ctx_fd,
5176 ovfl_pmds));
5177
5178 return pfm_notify_user(ctx, msg);
5179 }
5180
5181 static int
5182 pfm_end_notify_user(pfm_context_t *ctx)
5183 {
5184 pfm_msg_t *msg;
5185
5186 msg = pfm_get_new_msg(ctx);
5187 if (msg == NULL) {
5188 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5189 return -1;
5190 }
5191 /* no leak */
5192 memset(msg, 0, sizeof(*msg));
5193
5194 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5195 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5196 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5197
5198 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5199 msg,
5200 ctx->ctx_fl_no_msg,
5201 ctx->ctx_fd));
5202
5203 return pfm_notify_user(ctx, msg);
5204 }
5205
5206 /*
5207 * main overflow processing routine.
5208 * it can be called from the interrupt path or explicitly during the context switch code
5209 */
5210 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5211 unsigned long pmc0, struct pt_regs *regs)
5212 {
5213 pfm_ovfl_arg_t *ovfl_arg;
5214 unsigned long mask;
5215 unsigned long old_val, ovfl_val, new_val;
5216 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5217 unsigned long tstamp;
5218 pfm_ovfl_ctrl_t ovfl_ctrl;
5219 unsigned int i, has_smpl;
5220 int must_notify = 0;
5221
5222 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5223
5224 /*
5225 * sanity test. Should never happen
5226 */
5227 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5228
5229 tstamp = ia64_get_itc();
5230 mask = pmc0 >> PMU_FIRST_COUNTER;
5231 ovfl_val = pmu_conf->ovfl_val;
5232 has_smpl = CTX_HAS_SMPL(ctx);
5233
5234 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5235 "used_pmds=0x%lx\n",
5236 pmc0,
5237 task ? task_pid_nr(task): -1,
5238 (regs ? regs->cr_iip : 0),
5239 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5240 ctx->ctx_used_pmds[0]));
5241
5242
5243 /*
5244 * first we update the virtual counters
5245 * assume there was a prior ia64_srlz_d() issued
5246 */
5247 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5248
5249 /* skip pmd which did not overflow */
5250 if ((mask & 0x1) == 0) continue;
5251
5252 /*
5253 * Note that the pmd is not necessarily 0 at this point as qualified events
5254 * may have happened before the PMU was frozen. The residual count is not
5255 * taken into consideration here but will be with any read of the pmd via
5256 * pfm_read_pmds().
5257 */
5258 old_val = new_val = ctx->ctx_pmds[i].val;
5259 new_val += 1 + ovfl_val;
5260 ctx->ctx_pmds[i].val = new_val;
5261
5262 /*
5263 * check for overflow condition
5264 */
5265 if (likely(old_val > new_val)) {
5266 ovfl_pmds |= 1UL << i;
5267 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5268 }
5269
5270 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5271 i,
5272 new_val,
5273 old_val,
5274 ia64_get_pmd(i) & ovfl_val,
5275 ovfl_pmds,
5276 ovfl_notify));
5277 }
5278
5279 /*
5280 * there was no 64-bit overflow, nothing else to do
5281 */
5282 if (ovfl_pmds == 0UL) return;
5283
5284 /*
5285 * reset all control bits
5286 */
5287 ovfl_ctrl.val = 0;
5288 reset_pmds = 0UL;
5289
5290 /*
5291 * if a sampling format module exists, then we "cache" the overflow by
5292 * calling the module's handler() routine.
5293 */
5294 if (has_smpl) {
5295 unsigned long start_cycles, end_cycles;
5296 unsigned long pmd_mask;
5297 int j, k, ret = 0;
5298 int this_cpu = smp_processor_id();
5299
5300 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5301 ovfl_arg = &ctx->ctx_ovfl_arg;
5302
5303 prefetch(ctx->ctx_smpl_hdr);
5304
5305 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5306
5307 mask = 1UL << i;
5308
5309 if ((pmd_mask & 0x1) == 0) continue;
5310
5311 ovfl_arg->ovfl_pmd = (unsigned char )i;
5312 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5313 ovfl_arg->active_set = 0;
5314 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5315 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5316
5317 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5318 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5319 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5320
5321 /*
5322 * copy values of pmds of interest. Sampling format may copy them
5323 * into sampling buffer.
5324 */
5325 if (smpl_pmds) {
5326 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5327 if ((smpl_pmds & 0x1) == 0) continue;
5328 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5329 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5330 }
5331 }
5332
5333 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5334
5335 start_cycles = ia64_get_itc();
5336
5337 /*
5338 * call custom buffer format record (handler) routine
5339 */
5340 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5341
5342 end_cycles = ia64_get_itc();
5343
5344 /*
5345 * For those controls, we take the union because they have
5346 * an all or nothing behavior.
5347 */
5348 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5349 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5350 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5351 /*
5352 * build the bitmask of pmds to reset now
5353 */
5354 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5355
5356 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5357 }
5358 /*
5359 * when the module cannot handle the rest of the overflows, we abort right here
5360 */
5361 if (ret && pmd_mask) {
5362 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5363 pmd_mask<<PMU_FIRST_COUNTER));
5364 }
5365 /*
5366 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5367 */
5368 ovfl_pmds &= ~reset_pmds;
5369 } else {
5370 /*
5371 * when no sampling module is used, then the default
5372 * is to notify on overflow if requested by user
5373 */
5374 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5375 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5376 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5377 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5378 /*
5379 * if needed, we reset all overflowed pmds
5380 */
5381 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5382 }
5383
5384 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5385
5386 /*
5387 * reset the requested PMD registers using the short reset values
5388 */
5389 if (reset_pmds) {
5390 unsigned long bm = reset_pmds;
5391 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5392 }
5393
5394 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5395 /*
5396 * keep track of what to reset when unblocking
5397 */
5398 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5399
5400 /*
5401 * check for blocking context
5402 */
5403 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5404
5405 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5406
5407 /*
5408 * set the perfmon specific checking pending work for the task
5409 */
5410 PFM_SET_WORK_PENDING(task, 1);
5411
5412 /*
5413 * when coming from ctxsw, current still points to the
5414 * previous task, therefore we must work with task and not current.
5415 */
5416 set_notify_resume(task);
5417 }
5418 /*
5419 * defer until state is changed (shorten spin window). the context is locked
5420 * anyway, so the signal receiver would come spin for nothing.
5421 */
5422 must_notify = 1;
5423 }
5424
5425 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5426 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5427 PFM_GET_WORK_PENDING(task),
5428 ctx->ctx_fl_trap_reason,
5429 ovfl_pmds,
5430 ovfl_notify,
5431 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5432 /*
5433 * in case monitoring must be stopped, we toggle the psr bits
5434 */
5435 if (ovfl_ctrl.bits.mask_monitoring) {
5436 pfm_mask_monitoring(task);
5437 ctx->ctx_state = PFM_CTX_MASKED;
5438 ctx->ctx_fl_can_restart = 1;
5439 }
5440
5441 /*
5442 * send notification now
5443 */
5444 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5445
5446 return;
5447
5448 sanity_check:
5449 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5450 smp_processor_id(),
5451 task ? task_pid_nr(task) : -1,
5452 pmc0);
5453 return;
5454
5455 stop_monitoring:
5456 /*
5457 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5458 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5459 * come here as zombie only if the task is the current task. In which case, we
5460 * can access the PMU hardware directly.
5461 *
5462 * Note that zombies do have PM_VALID set. So here we do the minimal.
5463 *
5464 * In case the context was zombified it could not be reclaimed at the time
5465 * the monitoring program exited. At this point, the PMU reservation has been
5466 * returned, the sampiing buffer has been freed. We must convert this call
5467 * into a spurious interrupt. However, we must also avoid infinite overflows
5468 * by stopping monitoring for this task. We can only come here for a per-task
5469 * context. All we need to do is to stop monitoring using the psr bits which
5470 * are always task private. By re-enabling secure montioring, we ensure that
5471 * the monitored task will not be able to re-activate monitoring.
5472 * The task will eventually be context switched out, at which point the context
5473 * will be reclaimed (that includes releasing ownership of the PMU).
5474 *
5475 * So there might be a window of time where the number of per-task session is zero
5476 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5477 * context. This is safe because if a per-task session comes in, it will push this one
5478 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5479 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5480 * also push our zombie context out.
5481 *
5482 * Overall pretty hairy stuff....
5483 */
5484 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5485 pfm_clear_psr_up();
5486 ia64_psr(regs)->up = 0;
5487 ia64_psr(regs)->sp = 1;
5488 return;
5489 }
5490
5491 static int
5492 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5493 {
5494 struct task_struct *task;
5495 pfm_context_t *ctx;
5496 unsigned long flags;
5497 u64 pmc0;
5498 int this_cpu = smp_processor_id();
5499 int retval = 0;
5500
5501 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5502
5503 /*
5504 * srlz.d done before arriving here
5505 */
5506 pmc0 = ia64_get_pmc(0);
5507
5508 task = GET_PMU_OWNER();
5509 ctx = GET_PMU_CTX();
5510
5511 /*
5512 * if we have some pending bits set
5513 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5514 */
5515 if (PMC0_HAS_OVFL(pmc0) && task) {
5516 /*
5517 * we assume that pmc0.fr is always set here
5518 */
5519
5520 /* sanity check */
5521 if (!ctx) goto report_spurious1;
5522
5523 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5524 goto report_spurious2;
5525
5526 PROTECT_CTX_NOPRINT(ctx, flags);
5527
5528 pfm_overflow_handler(task, ctx, pmc0, regs);
5529
5530 UNPROTECT_CTX_NOPRINT(ctx, flags);
5531
5532 } else {
5533 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5534 retval = -1;
5535 }
5536 /*
5537 * keep it unfrozen at all times
5538 */
5539 pfm_unfreeze_pmu();
5540
5541 return retval;
5542
5543 report_spurious1:
5544 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5545 this_cpu, task_pid_nr(task));
5546 pfm_unfreeze_pmu();
5547 return -1;
5548 report_spurious2:
5549 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5550 this_cpu,
5551 task_pid_nr(task));
5552 pfm_unfreeze_pmu();
5553 return -1;
5554 }
5555
5556 static irqreturn_t
5557 pfm_interrupt_handler(int irq, void *arg)
5558 {
5559 unsigned long start_cycles, total_cycles;
5560 unsigned long min, max;
5561 int this_cpu;
5562 int ret;
5563 struct pt_regs *regs = get_irq_regs();
5564
5565 this_cpu = get_cpu();
5566 if (likely(!pfm_alt_intr_handler)) {
5567 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5568 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5569
5570 start_cycles = ia64_get_itc();
5571
5572 ret = pfm_do_interrupt_handler(arg, regs);
5573
5574 total_cycles = ia64_get_itc();
5575
5576 /*
5577 * don't measure spurious interrupts
5578 */
5579 if (likely(ret == 0)) {
5580 total_cycles -= start_cycles;
5581
5582 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5583 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5584
5585 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5586 }
5587 }
5588 else {
5589 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5590 }
5591
5592 put_cpu();
5593 return IRQ_HANDLED;
5594 }
5595
5596 /*
5597 * /proc/perfmon interface, for debug only
5598 */
5599
5600 #define PFM_PROC_SHOW_HEADER ((void *)(long)nr_cpu_ids+1)
5601
5602 static void *
5603 pfm_proc_start(struct seq_file *m, loff_t *pos)
5604 {
5605 if (*pos == 0) {
5606 return PFM_PROC_SHOW_HEADER;
5607 }
5608
5609 while (*pos <= nr_cpu_ids) {
5610 if (cpu_online(*pos - 1)) {
5611 return (void *)*pos;
5612 }
5613 ++*pos;
5614 }
5615 return NULL;
5616 }
5617
5618 static void *
5619 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5620 {
5621 ++*pos;
5622 return pfm_proc_start(m, pos);
5623 }
5624
5625 static void
5626 pfm_proc_stop(struct seq_file *m, void *v)
5627 {
5628 }
5629
5630 static void
5631 pfm_proc_show_header(struct seq_file *m)
5632 {
5633 struct list_head * pos;
5634 pfm_buffer_fmt_t * entry;
5635 unsigned long flags;
5636
5637 seq_printf(m,
5638 "perfmon version : %u.%u\n"
5639 "model : %s\n"
5640 "fastctxsw : %s\n"
5641 "expert mode : %s\n"
5642 "ovfl_mask : 0x%lx\n"
5643 "PMU flags : 0x%x\n",
5644 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5645 pmu_conf->pmu_name,
5646 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5647 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5648 pmu_conf->ovfl_val,
5649 pmu_conf->flags);
5650
5651 LOCK_PFS(flags);
5652
5653 seq_printf(m,
5654 "proc_sessions : %u\n"
5655 "sys_sessions : %u\n"
5656 "sys_use_dbregs : %u\n"
5657 "ptrace_use_dbregs : %u\n",
5658 pfm_sessions.pfs_task_sessions,
5659 pfm_sessions.pfs_sys_sessions,
5660 pfm_sessions.pfs_sys_use_dbregs,
5661 pfm_sessions.pfs_ptrace_use_dbregs);
5662
5663 UNLOCK_PFS(flags);
5664
5665 spin_lock(&pfm_buffer_fmt_lock);
5666
5667 list_for_each(pos, &pfm_buffer_fmt_list) {
5668 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5669 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5670 entry->fmt_uuid[0],
5671 entry->fmt_uuid[1],
5672 entry->fmt_uuid[2],
5673 entry->fmt_uuid[3],
5674 entry->fmt_uuid[4],
5675 entry->fmt_uuid[5],
5676 entry->fmt_uuid[6],
5677 entry->fmt_uuid[7],
5678 entry->fmt_uuid[8],
5679 entry->fmt_uuid[9],
5680 entry->fmt_uuid[10],
5681 entry->fmt_uuid[11],
5682 entry->fmt_uuid[12],
5683 entry->fmt_uuid[13],
5684 entry->fmt_uuid[14],
5685 entry->fmt_uuid[15],
5686 entry->fmt_name);
5687 }
5688 spin_unlock(&pfm_buffer_fmt_lock);
5689
5690 }
5691
5692 static int
5693 pfm_proc_show(struct seq_file *m, void *v)
5694 {
5695 unsigned long psr;
5696 unsigned int i;
5697 int cpu;
5698
5699 if (v == PFM_PROC_SHOW_HEADER) {
5700 pfm_proc_show_header(m);
5701 return 0;
5702 }
5703
5704 /* show info for CPU (v - 1) */
5705
5706 cpu = (long)v - 1;
5707 seq_printf(m,
5708 "CPU%-2d overflow intrs : %lu\n"
5709 "CPU%-2d overflow cycles : %lu\n"
5710 "CPU%-2d overflow min : %lu\n"
5711 "CPU%-2d overflow max : %lu\n"
5712 "CPU%-2d smpl handler calls : %lu\n"
5713 "CPU%-2d smpl handler cycles : %lu\n"
5714 "CPU%-2d spurious intrs : %lu\n"
5715 "CPU%-2d replay intrs : %lu\n"
5716 "CPU%-2d syst_wide : %d\n"
5717 "CPU%-2d dcr_pp : %d\n"
5718 "CPU%-2d exclude idle : %d\n"
5719 "CPU%-2d owner : %d\n"
5720 "CPU%-2d context : %p\n"
5721 "CPU%-2d activations : %lu\n",
5722 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5723 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5724 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5725 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5726 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5727 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5728 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5729 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5730 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5731 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5732 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5733 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5734 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5735 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5736
5737 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5738
5739 psr = pfm_get_psr();
5740
5741 ia64_srlz_d();
5742
5743 seq_printf(m,
5744 "CPU%-2d psr : 0x%lx\n"
5745 "CPU%-2d pmc0 : 0x%lx\n",
5746 cpu, psr,
5747 cpu, ia64_get_pmc(0));
5748
5749 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5750 if (PMC_IS_COUNTING(i) == 0) continue;
5751 seq_printf(m,
5752 "CPU%-2d pmc%u : 0x%lx\n"
5753 "CPU%-2d pmd%u : 0x%lx\n",
5754 cpu, i, ia64_get_pmc(i),
5755 cpu, i, ia64_get_pmd(i));
5756 }
5757 }
5758 return 0;
5759 }
5760
5761 const struct seq_operations pfm_seq_ops = {
5762 .start = pfm_proc_start,
5763 .next = pfm_proc_next,
5764 .stop = pfm_proc_stop,
5765 .show = pfm_proc_show
5766 };
5767
5768 static int
5769 pfm_proc_open(struct inode *inode, struct file *file)
5770 {
5771 return seq_open(file, &pfm_seq_ops);
5772 }
5773
5774
5775 /*
5776 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5777 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5778 * is active or inactive based on mode. We must rely on the value in
5779 * local_cpu_data->pfm_syst_info
5780 */
5781 void
5782 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5783 {
5784 struct pt_regs *regs;
5785 unsigned long dcr;
5786 unsigned long dcr_pp;
5787
5788 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5789
5790 /*
5791 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5792 * on every CPU, so we can rely on the pid to identify the idle task.
5793 */
5794 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5795 regs = task_pt_regs(task);
5796 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5797 return;
5798 }
5799 /*
5800 * if monitoring has started
5801 */
5802 if (dcr_pp) {
5803 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5804 /*
5805 * context switching in?
5806 */
5807 if (is_ctxswin) {
5808 /* mask monitoring for the idle task */
5809 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5810 pfm_clear_psr_pp();
5811 ia64_srlz_i();
5812 return;
5813 }
5814 /*
5815 * context switching out
5816 * restore monitoring for next task
5817 *
5818 * Due to inlining this odd if-then-else construction generates
5819 * better code.
5820 */
5821 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5822 pfm_set_psr_pp();
5823 ia64_srlz_i();
5824 }
5825 }
5826
5827 #ifdef CONFIG_SMP
5828
5829 static void
5830 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5831 {
5832 struct task_struct *task = ctx->ctx_task;
5833
5834 ia64_psr(regs)->up = 0;
5835 ia64_psr(regs)->sp = 1;
5836
5837 if (GET_PMU_OWNER() == task) {
5838 DPRINT(("cleared ownership for [%d]\n",
5839 task_pid_nr(ctx->ctx_task)));
5840 SET_PMU_OWNER(NULL, NULL);
5841 }
5842
5843 /*
5844 * disconnect the task from the context and vice-versa
5845 */
5846 PFM_SET_WORK_PENDING(task, 0);
5847
5848 task->thread.pfm_context = NULL;
5849 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5850
5851 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task)));
5852 }
5853
5854
5855 /*
5856 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5857 */
5858 void
5859 pfm_save_regs(struct task_struct *task)
5860 {
5861 pfm_context_t *ctx;
5862 unsigned long flags;
5863 u64 psr;
5864
5865
5866 ctx = PFM_GET_CTX(task);
5867 if (ctx == NULL) return;
5868
5869 /*
5870 * we always come here with interrupts ALREADY disabled by
5871 * the scheduler. So we simply need to protect against concurrent
5872 * access, not CPU concurrency.
5873 */
5874 flags = pfm_protect_ctx_ctxsw(ctx);
5875
5876 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5877 struct pt_regs *regs = task_pt_regs(task);
5878
5879 pfm_clear_psr_up();
5880
5881 pfm_force_cleanup(ctx, regs);
5882
5883 BUG_ON(ctx->ctx_smpl_hdr);
5884
5885 pfm_unprotect_ctx_ctxsw(ctx, flags);
5886
5887 pfm_context_free(ctx);
5888 return;
5889 }
5890
5891 /*
5892 * save current PSR: needed because we modify it
5893 */
5894 ia64_srlz_d();
5895 psr = pfm_get_psr();
5896
5897 BUG_ON(psr & (IA64_PSR_I));
5898
5899 /*
5900 * stop monitoring:
5901 * This is the last instruction which may generate an overflow
5902 *
5903 * We do not need to set psr.sp because, it is irrelevant in kernel.
5904 * It will be restored from ipsr when going back to user level
5905 */
5906 pfm_clear_psr_up();
5907
5908 /*
5909 * keep a copy of psr.up (for reload)
5910 */
5911 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5912
5913 /*
5914 * release ownership of this PMU.
5915 * PM interrupts are masked, so nothing
5916 * can happen.
5917 */
5918 SET_PMU_OWNER(NULL, NULL);
5919
5920 /*
5921 * we systematically save the PMD as we have no
5922 * guarantee we will be schedule at that same
5923 * CPU again.
5924 */
5925 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5926
5927 /*
5928 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5929 * we will need it on the restore path to check
5930 * for pending overflow.
5931 */
5932 ctx->th_pmcs[0] = ia64_get_pmc(0);
5933
5934 /*
5935 * unfreeze PMU if had pending overflows
5936 */
5937 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5938
5939 /*
5940 * finally, allow context access.
5941 * interrupts will still be masked after this call.
5942 */
5943 pfm_unprotect_ctx_ctxsw(ctx, flags);
5944 }
5945
5946 #else /* !CONFIG_SMP */
5947 void
5948 pfm_save_regs(struct task_struct *task)
5949 {
5950 pfm_context_t *ctx;
5951 u64 psr;
5952
5953 ctx = PFM_GET_CTX(task);
5954 if (ctx == NULL) return;
5955
5956 /*
5957 * save current PSR: needed because we modify it
5958 */
5959 psr = pfm_get_psr();
5960
5961 BUG_ON(psr & (IA64_PSR_I));
5962
5963 /*
5964 * stop monitoring:
5965 * This is the last instruction which may generate an overflow
5966 *
5967 * We do not need to set psr.sp because, it is irrelevant in kernel.
5968 * It will be restored from ipsr when going back to user level
5969 */
5970 pfm_clear_psr_up();
5971
5972 /*
5973 * keep a copy of psr.up (for reload)
5974 */
5975 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5976 }
5977
5978 static void
5979 pfm_lazy_save_regs (struct task_struct *task)
5980 {
5981 pfm_context_t *ctx;
5982 unsigned long flags;
5983
5984 { u64 psr = pfm_get_psr();
5985 BUG_ON(psr & IA64_PSR_UP);
5986 }
5987
5988 ctx = PFM_GET_CTX(task);
5989
5990 /*
5991 * we need to mask PMU overflow here to
5992 * make sure that we maintain pmc0 until
5993 * we save it. overflow interrupts are
5994 * treated as spurious if there is no
5995 * owner.
5996 *
5997 * XXX: I don't think this is necessary
5998 */
5999 PROTECT_CTX(ctx,flags);
6000
6001 /*
6002 * release ownership of this PMU.
6003 * must be done before we save the registers.
6004 *
6005 * after this call any PMU interrupt is treated
6006 * as spurious.
6007 */
6008 SET_PMU_OWNER(NULL, NULL);
6009
6010 /*
6011 * save all the pmds we use
6012 */
6013 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6014
6015 /*
6016 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6017 * it is needed to check for pended overflow
6018 * on the restore path
6019 */
6020 ctx->th_pmcs[0] = ia64_get_pmc(0);
6021
6022 /*
6023 * unfreeze PMU if had pending overflows
6024 */
6025 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6026
6027 /*
6028 * now get can unmask PMU interrupts, they will
6029 * be treated as purely spurious and we will not
6030 * lose any information
6031 */
6032 UNPROTECT_CTX(ctx,flags);
6033 }
6034 #endif /* CONFIG_SMP */
6035
6036 #ifdef CONFIG_SMP
6037 /*
6038 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6039 */
6040 void
6041 pfm_load_regs (struct task_struct *task)
6042 {
6043 pfm_context_t *ctx;
6044 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6045 unsigned long flags;
6046 u64 psr, psr_up;
6047 int need_irq_resend;
6048
6049 ctx = PFM_GET_CTX(task);
6050 if (unlikely(ctx == NULL)) return;
6051
6052 BUG_ON(GET_PMU_OWNER());
6053
6054 /*
6055 * possible on unload
6056 */
6057 if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6058
6059 /*
6060 * we always come here with interrupts ALREADY disabled by
6061 * the scheduler. So we simply need to protect against concurrent
6062 * access, not CPU concurrency.
6063 */
6064 flags = pfm_protect_ctx_ctxsw(ctx);
6065 psr = pfm_get_psr();
6066
6067 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6068
6069 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6070 BUG_ON(psr & IA64_PSR_I);
6071
6072 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6073 struct pt_regs *regs = task_pt_regs(task);
6074
6075 BUG_ON(ctx->ctx_smpl_hdr);
6076
6077 pfm_force_cleanup(ctx, regs);
6078
6079 pfm_unprotect_ctx_ctxsw(ctx, flags);
6080
6081 /*
6082 * this one (kmalloc'ed) is fine with interrupts disabled
6083 */
6084 pfm_context_free(ctx);
6085
6086 return;
6087 }
6088
6089 /*
6090 * we restore ALL the debug registers to avoid picking up
6091 * stale state.
6092 */
6093 if (ctx->ctx_fl_using_dbreg) {
6094 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6095 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6096 }
6097 /*
6098 * retrieve saved psr.up
6099 */
6100 psr_up = ctx->ctx_saved_psr_up;
6101
6102 /*
6103 * if we were the last user of the PMU on that CPU,
6104 * then nothing to do except restore psr
6105 */
6106 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6107
6108 /*
6109 * retrieve partial reload masks (due to user modifications)
6110 */
6111 pmc_mask = ctx->ctx_reload_pmcs[0];
6112 pmd_mask = ctx->ctx_reload_pmds[0];
6113
6114 } else {
6115 /*
6116 * To avoid leaking information to the user level when psr.sp=0,
6117 * we must reload ALL implemented pmds (even the ones we don't use).
6118 * In the kernel we only allow PFM_READ_PMDS on registers which
6119 * we initialized or requested (sampling) so there is no risk there.
6120 */
6121 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6122
6123 /*
6124 * ALL accessible PMCs are systematically reloaded, unused registers
6125 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6126 * up stale configuration.
6127 *
6128 * PMC0 is never in the mask. It is always restored separately.
6129 */
6130 pmc_mask = ctx->ctx_all_pmcs[0];
6131 }
6132 /*
6133 * when context is MASKED, we will restore PMC with plm=0
6134 * and PMD with stale information, but that's ok, nothing
6135 * will be captured.
6136 *
6137 * XXX: optimize here
6138 */
6139 if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6140 if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6141
6142 /*
6143 * check for pending overflow at the time the state
6144 * was saved.
6145 */
6146 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6147 /*
6148 * reload pmc0 with the overflow information
6149 * On McKinley PMU, this will trigger a PMU interrupt
6150 */
6151 ia64_set_pmc(0, ctx->th_pmcs[0]);
6152 ia64_srlz_d();
6153 ctx->th_pmcs[0] = 0UL;
6154
6155 /*
6156 * will replay the PMU interrupt
6157 */
6158 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6159
6160 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6161 }
6162
6163 /*
6164 * we just did a reload, so we reset the partial reload fields
6165 */
6166 ctx->ctx_reload_pmcs[0] = 0UL;
6167 ctx->ctx_reload_pmds[0] = 0UL;
6168
6169 SET_LAST_CPU(ctx, smp_processor_id());
6170
6171 /*
6172 * dump activation value for this PMU
6173 */
6174 INC_ACTIVATION();
6175 /*
6176 * record current activation for this context
6177 */
6178 SET_ACTIVATION(ctx);
6179
6180 /*
6181 * establish new ownership.
6182 */
6183 SET_PMU_OWNER(task, ctx);
6184
6185 /*
6186 * restore the psr.up bit. measurement
6187 * is active again.
6188 * no PMU interrupt can happen at this point
6189 * because we still have interrupts disabled.
6190 */
6191 if (likely(psr_up)) pfm_set_psr_up();
6192
6193 /*
6194 * allow concurrent access to context
6195 */
6196 pfm_unprotect_ctx_ctxsw(ctx, flags);
6197 }
6198 #else /* !CONFIG_SMP */
6199 /*
6200 * reload PMU state for UP kernels
6201 * in 2.5 we come here with interrupts disabled
6202 */
6203 void
6204 pfm_load_regs (struct task_struct *task)
6205 {
6206 pfm_context_t *ctx;
6207 struct task_struct *owner;
6208 unsigned long pmd_mask, pmc_mask;
6209 u64 psr, psr_up;
6210 int need_irq_resend;
6211
6212 owner = GET_PMU_OWNER();
6213 ctx = PFM_GET_CTX(task);
6214 psr = pfm_get_psr();
6215
6216 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6217 BUG_ON(psr & IA64_PSR_I);
6218
6219 /*
6220 * we restore ALL the debug registers to avoid picking up
6221 * stale state.
6222 *
6223 * This must be done even when the task is still the owner
6224 * as the registers may have been modified via ptrace()
6225 * (not perfmon) by the previous task.
6226 */
6227 if (ctx->ctx_fl_using_dbreg) {
6228 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6229 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6230 }
6231
6232 /*
6233 * retrieved saved psr.up
6234 */
6235 psr_up = ctx->ctx_saved_psr_up;
6236 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6237
6238 /*
6239 * short path, our state is still there, just
6240 * need to restore psr and we go
6241 *
6242 * we do not touch either PMC nor PMD. the psr is not touched
6243 * by the overflow_handler. So we are safe w.r.t. to interrupt
6244 * concurrency even without interrupt masking.
6245 */
6246 if (likely(owner == task)) {
6247 if (likely(psr_up)) pfm_set_psr_up();
6248 return;
6249 }
6250
6251 /*
6252 * someone else is still using the PMU, first push it out and
6253 * then we'll be able to install our stuff !
6254 *
6255 * Upon return, there will be no owner for the current PMU
6256 */
6257 if (owner) pfm_lazy_save_regs(owner);
6258
6259 /*
6260 * To avoid leaking information to the user level when psr.sp=0,
6261 * we must reload ALL implemented pmds (even the ones we don't use).
6262 * In the kernel we only allow PFM_READ_PMDS on registers which
6263 * we initialized or requested (sampling) so there is no risk there.
6264 */
6265 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6266
6267 /*
6268 * ALL accessible PMCs are systematically reloaded, unused registers
6269 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6270 * up stale configuration.
6271 *
6272 * PMC0 is never in the mask. It is always restored separately
6273 */
6274 pmc_mask = ctx->ctx_all_pmcs[0];
6275
6276 pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6277 pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6278
6279 /*
6280 * check for pending overflow at the time the state
6281 * was saved.
6282 */
6283 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6284 /*
6285 * reload pmc0 with the overflow information
6286 * On McKinley PMU, this will trigger a PMU interrupt
6287 */
6288 ia64_set_pmc(0, ctx->th_pmcs[0]);
6289 ia64_srlz_d();
6290
6291 ctx->th_pmcs[0] = 0UL;
6292
6293 /*
6294 * will replay the PMU interrupt
6295 */
6296 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6297
6298 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6299 }
6300
6301 /*
6302 * establish new ownership.
6303 */
6304 SET_PMU_OWNER(task, ctx);
6305
6306 /*
6307 * restore the psr.up bit. measurement
6308 * is active again.
6309 * no PMU interrupt can happen at this point
6310 * because we still have interrupts disabled.
6311 */
6312 if (likely(psr_up)) pfm_set_psr_up();
6313 }
6314 #endif /* CONFIG_SMP */
6315
6316 /*
6317 * this function assumes monitoring is stopped
6318 */
6319 static void
6320 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6321 {
6322 u64 pmc0;
6323 unsigned long mask2, val, pmd_val, ovfl_val;
6324 int i, can_access_pmu = 0;
6325 int is_self;
6326
6327 /*
6328 * is the caller the task being monitored (or which initiated the
6329 * session for system wide measurements)
6330 */
6331 is_self = ctx->ctx_task == task ? 1 : 0;
6332
6333 /*
6334 * can access PMU is task is the owner of the PMU state on the current CPU
6335 * or if we are running on the CPU bound to the context in system-wide mode
6336 * (that is not necessarily the task the context is attached to in this mode).
6337 * In system-wide we always have can_access_pmu true because a task running on an
6338 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6339 */
6340 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6341 if (can_access_pmu) {
6342 /*
6343 * Mark the PMU as not owned
6344 * This will cause the interrupt handler to do nothing in case an overflow
6345 * interrupt was in-flight
6346 * This also guarantees that pmc0 will contain the final state
6347 * It virtually gives us full control on overflow processing from that point
6348 * on.
6349 */
6350 SET_PMU_OWNER(NULL, NULL);
6351 DPRINT(("releasing ownership\n"));
6352
6353 /*
6354 * read current overflow status:
6355 *
6356 * we are guaranteed to read the final stable state
6357 */
6358 ia64_srlz_d();
6359 pmc0 = ia64_get_pmc(0); /* slow */
6360
6361 /*
6362 * reset freeze bit, overflow status information destroyed
6363 */
6364 pfm_unfreeze_pmu();
6365 } else {
6366 pmc0 = ctx->th_pmcs[0];
6367 /*
6368 * clear whatever overflow status bits there were
6369 */
6370 ctx->th_pmcs[0] = 0;
6371 }
6372 ovfl_val = pmu_conf->ovfl_val;
6373 /*
6374 * we save all the used pmds
6375 * we take care of overflows for counting PMDs
6376 *
6377 * XXX: sampling situation is not taken into account here
6378 */
6379 mask2 = ctx->ctx_used_pmds[0];
6380
6381 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6382
6383 for (i = 0; mask2; i++, mask2>>=1) {
6384
6385 /* skip non used pmds */
6386 if ((mask2 & 0x1) == 0) continue;
6387
6388 /*
6389 * can access PMU always true in system wide mode
6390 */
6391 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6392
6393 if (PMD_IS_COUNTING(i)) {
6394 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6395 task_pid_nr(task),
6396 i,
6397 ctx->ctx_pmds[i].val,
6398 val & ovfl_val));
6399
6400 /*
6401 * we rebuild the full 64 bit value of the counter
6402 */
6403 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6404
6405 /*
6406 * now everything is in ctx_pmds[] and we need
6407 * to clear the saved context from save_regs() such that
6408 * pfm_read_pmds() gets the correct value
6409 */
6410 pmd_val = 0UL;
6411
6412 /*
6413 * take care of overflow inline
6414 */
6415 if (pmc0 & (1UL << i)) {
6416 val += 1 + ovfl_val;
6417 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6418 }
6419 }
6420
6421 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6422
6423 if (is_self) ctx->th_pmds[i] = pmd_val;
6424
6425 ctx->ctx_pmds[i].val = val;
6426 }
6427 }
6428
6429 static struct irqaction perfmon_irqaction = {
6430 .handler = pfm_interrupt_handler,
6431 .flags = IRQF_DISABLED,
6432 .name = "perfmon"
6433 };
6434
6435 static void
6436 pfm_alt_save_pmu_state(void *data)
6437 {
6438 struct pt_regs *regs;
6439
6440 regs = task_pt_regs(current);
6441
6442 DPRINT(("called\n"));
6443
6444 /*
6445 * should not be necessary but
6446 * let's take not risk
6447 */
6448 pfm_clear_psr_up();
6449 pfm_clear_psr_pp();
6450 ia64_psr(regs)->pp = 0;
6451
6452 /*
6453 * This call is required
6454 * May cause a spurious interrupt on some processors
6455 */
6456 pfm_freeze_pmu();
6457
6458 ia64_srlz_d();
6459 }
6460
6461 void
6462 pfm_alt_restore_pmu_state(void *data)
6463 {
6464 struct pt_regs *regs;
6465
6466 regs = task_pt_regs(current);
6467
6468 DPRINT(("called\n"));
6469
6470 /*
6471 * put PMU back in state expected
6472 * by perfmon
6473 */
6474 pfm_clear_psr_up();
6475 pfm_clear_psr_pp();
6476 ia64_psr(regs)->pp = 0;
6477
6478 /*
6479 * perfmon runs with PMU unfrozen at all times
6480 */
6481 pfm_unfreeze_pmu();
6482
6483 ia64_srlz_d();
6484 }
6485
6486 int
6487 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6488 {
6489 int ret, i;
6490 int reserve_cpu;
6491
6492 /* some sanity checks */
6493 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6494
6495 /* do the easy test first */
6496 if (pfm_alt_intr_handler) return -EBUSY;
6497
6498 /* one at a time in the install or remove, just fail the others */
6499 if (!spin_trylock(&pfm_alt_install_check)) {
6500 return -EBUSY;
6501 }
6502
6503 /* reserve our session */
6504 for_each_online_cpu(reserve_cpu) {
6505 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6506 if (ret) goto cleanup_reserve;
6507 }
6508
6509 /* save the current system wide pmu states */
6510 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6511 if (ret) {
6512 DPRINT(("on_each_cpu() failed: %d\n", ret));
6513 goto cleanup_reserve;
6514 }
6515
6516 /* officially change to the alternate interrupt handler */
6517 pfm_alt_intr_handler = hdl;
6518
6519 spin_unlock(&pfm_alt_install_check);
6520
6521 return 0;
6522
6523 cleanup_reserve:
6524 for_each_online_cpu(i) {
6525 /* don't unreserve more than we reserved */
6526 if (i >= reserve_cpu) break;
6527
6528 pfm_unreserve_session(NULL, 1, i);
6529 }
6530
6531 spin_unlock(&pfm_alt_install_check);
6532
6533 return ret;
6534 }
6535 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6536
6537 int
6538 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6539 {
6540 int i;
6541 int ret;
6542
6543 if (hdl == NULL) return -EINVAL;
6544
6545 /* cannot remove someone else's handler! */
6546 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6547
6548 /* one at a time in the install or remove, just fail the others */
6549 if (!spin_trylock(&pfm_alt_install_check)) {
6550 return -EBUSY;
6551 }
6552
6553 pfm_alt_intr_handler = NULL;
6554
6555 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6556 if (ret) {
6557 DPRINT(("on_each_cpu() failed: %d\n", ret));
6558 }
6559
6560 for_each_online_cpu(i) {
6561 pfm_unreserve_session(NULL, 1, i);
6562 }
6563
6564 spin_unlock(&pfm_alt_install_check);
6565
6566 return 0;
6567 }
6568 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6569
6570 /*
6571 * perfmon initialization routine, called from the initcall() table
6572 */
6573 static int init_pfm_fs(void);
6574
6575 static int __init
6576 pfm_probe_pmu(void)
6577 {
6578 pmu_config_t **p;
6579 int family;
6580
6581 family = local_cpu_data->family;
6582 p = pmu_confs;
6583
6584 while(*p) {
6585 if ((*p)->probe) {
6586 if ((*p)->probe() == 0) goto found;
6587 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6588 goto found;
6589 }
6590 p++;
6591 }
6592 return -1;
6593 found:
6594 pmu_conf = *p;
6595 return 0;
6596 }
6597
6598 static const struct file_operations pfm_proc_fops = {
6599 .open = pfm_proc_open,
6600 .read = seq_read,
6601 .llseek = seq_lseek,
6602 .release = seq_release,
6603 };
6604
6605 int __init
6606 pfm_init(void)
6607 {
6608 unsigned int n, n_counters, i;
6609
6610 printk("perfmon: version %u.%u IRQ %u\n",
6611 PFM_VERSION_MAJ,
6612 PFM_VERSION_MIN,
6613 IA64_PERFMON_VECTOR);
6614
6615 if (pfm_probe_pmu()) {
6616 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6617 local_cpu_data->family);
6618 return -ENODEV;
6619 }
6620
6621 /*
6622 * compute the number of implemented PMD/PMC from the
6623 * description tables
6624 */
6625 n = 0;
6626 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6627 if (PMC_IS_IMPL(i) == 0) continue;
6628 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6629 n++;
6630 }
6631 pmu_conf->num_pmcs = n;
6632
6633 n = 0; n_counters = 0;
6634 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6635 if (PMD_IS_IMPL(i) == 0) continue;
6636 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6637 n++;
6638 if (PMD_IS_COUNTING(i)) n_counters++;
6639 }
6640 pmu_conf->num_pmds = n;
6641 pmu_conf->num_counters = n_counters;
6642
6643 /*
6644 * sanity checks on the number of debug registers
6645 */
6646 if (pmu_conf->use_rr_dbregs) {
6647 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6648 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6649 pmu_conf = NULL;
6650 return -1;
6651 }
6652 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6653 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6654 pmu_conf = NULL;
6655 return -1;
6656 }
6657 }
6658
6659 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6660 pmu_conf->pmu_name,
6661 pmu_conf->num_pmcs,
6662 pmu_conf->num_pmds,
6663 pmu_conf->num_counters,
6664 ffz(pmu_conf->ovfl_val));
6665
6666 /* sanity check */
6667 if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6668 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6669 pmu_conf = NULL;
6670 return -1;
6671 }
6672
6673 /*
6674 * create /proc/perfmon (mostly for debugging purposes)
6675 */
6676 perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6677 if (perfmon_dir == NULL) {
6678 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6679 pmu_conf = NULL;
6680 return -1;
6681 }
6682
6683 /*
6684 * create /proc/sys/kernel/perfmon (for debugging purposes)
6685 */
6686 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6687
6688 /*
6689 * initialize all our spinlocks
6690 */
6691 spin_lock_init(&pfm_sessions.pfs_lock);
6692 spin_lock_init(&pfm_buffer_fmt_lock);
6693
6694 init_pfm_fs();
6695
6696 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6697
6698 return 0;
6699 }
6700
6701 __initcall(pfm_init);
6702
6703 /*
6704 * this function is called before pfm_init()
6705 */
6706 void
6707 pfm_init_percpu (void)
6708 {
6709 static int first_time=1;
6710 /*
6711 * make sure no measurement is active
6712 * (may inherit programmed PMCs from EFI).
6713 */
6714 pfm_clear_psr_pp();
6715 pfm_clear_psr_up();
6716
6717 /*
6718 * we run with the PMU not frozen at all times
6719 */
6720 pfm_unfreeze_pmu();
6721
6722 if (first_time) {
6723 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6724 first_time=0;
6725 }
6726
6727 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6728 ia64_srlz_d();
6729 }
6730
6731 /*
6732 * used for debug purposes only
6733 */
6734 void
6735 dump_pmu_state(const char *from)
6736 {
6737 struct task_struct *task;
6738 struct pt_regs *regs;
6739 pfm_context_t *ctx;
6740 unsigned long psr, dcr, info, flags;
6741 int i, this_cpu;
6742
6743 local_irq_save(flags);
6744
6745 this_cpu = smp_processor_id();
6746 regs = task_pt_regs(current);
6747 info = PFM_CPUINFO_GET();
6748 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6749
6750 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6751 local_irq_restore(flags);
6752 return;
6753 }
6754
6755 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6756 this_cpu,
6757 from,
6758 task_pid_nr(current),
6759 regs->cr_iip,
6760 current->comm);
6761
6762 task = GET_PMU_OWNER();
6763 ctx = GET_PMU_CTX();
6764
6765 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6766
6767 psr = pfm_get_psr();
6768
6769 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",
6770 this_cpu,
6771 ia64_get_pmc(0),
6772 psr & IA64_PSR_PP ? 1 : 0,
6773 psr & IA64_PSR_UP ? 1 : 0,
6774 dcr & IA64_DCR_PP ? 1 : 0,
6775 info,
6776 ia64_psr(regs)->up,
6777 ia64_psr(regs)->pp);
6778
6779 ia64_psr(regs)->up = 0;
6780 ia64_psr(regs)->pp = 0;
6781
6782 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6783 if (PMC_IS_IMPL(i) == 0) continue;
6784 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]);
6785 }
6786
6787 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6788 if (PMD_IS_IMPL(i) == 0) continue;
6789 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]);
6790 }
6791
6792 if (ctx) {
6793 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6794 this_cpu,
6795 ctx->ctx_state,
6796 ctx->ctx_smpl_vaddr,
6797 ctx->ctx_smpl_hdr,
6798 ctx->ctx_msgq_head,
6799 ctx->ctx_msgq_tail,
6800 ctx->ctx_saved_psr_up);
6801 }
6802 local_irq_restore(flags);
6803 }
6804
6805 /*
6806 * called from process.c:copy_thread(). task is new child.
6807 */
6808 void
6809 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6810 {
6811 struct thread_struct *thread;
6812
6813 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6814
6815 thread = &task->thread;
6816
6817 /*
6818 * cut links inherited from parent (current)
6819 */
6820 thread->pfm_context = NULL;
6821
6822 PFM_SET_WORK_PENDING(task, 0);
6823
6824 /*
6825 * the psr bits are already set properly in copy_threads()
6826 */
6827 }
6828 #else /* !CONFIG_PERFMON */
6829 asmlinkage long
6830 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6831 {
6832 return -ENOSYS;
6833 }
6834 #endif /* CONFIG_PERFMON */
linux 2.6 git repository for openrd