?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
ia64/kvm: compilation fix. export account_system_vtime.
[pv_ops_mirror.git] / arch / ia64 / kernel / perfmon.c
blob71d05133f556000530d930537c9dc14c543037c9
1 /*
2 * This file implements the perfmon-2 subsystem which is used
3 * to program the IA-64 Performance Monitoring Unit (PMU).
4 *
5 * The initial version of perfmon.c was written by
6 * Ganesh Venkitachalam, IBM Corp.
7 *
8 * Then it was modified for perfmon-1.x by Stephane Eranian and
9 * David Mosberger, Hewlett Packard Co.
10 *
11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
12 * by Stephane Eranian, Hewlett Packard Co.
13 *
14 * Copyright (C) 1999-2005 Hewlett Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 * David Mosberger-Tang <davidm@hpl.hp.com>
17 *
18 * More information about perfmon available at:
19 * http://www.hpl.hp.com/research/linux/perfmon
20 */
21
22 #include <linux/module.h>
23 #include <linux/kernel.h>
24 #include <linux/sched.h>
25 #include <linux/interrupt.h>
26 #include <linux/proc_fs.h>
27 #include <linux/seq_file.h>
28 #include <linux/init.h>
29 #include <linux/vmalloc.h>
30 #include <linux/mm.h>
31 #include <linux/sysctl.h>
32 #include <linux/list.h>
33 #include <linux/file.h>
34 #include <linux/poll.h>
35 #include <linux/vfs.h>
36 #include <linux/smp.h>
37 #include <linux/pagemap.h>
38 #include <linux/mount.h>
39 #include <linux/bitops.h>
40 #include <linux/capability.h>
41 #include <linux/rcupdate.h>
42 #include <linux/completion.h>
43
44 #include <asm/errno.h>
45 #include <asm/intrinsics.h>
46 #include <asm/page.h>
47 #include <asm/perfmon.h>
48 #include <asm/processor.h>
49 #include <asm/signal.h>
50 #include <asm/system.h>
51 #include <asm/uaccess.h>
52 #include <asm/delay.h>
53
54 #ifdef CONFIG_PERFMON
55 /*
56 * perfmon context state
57 */
58 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
59 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
60 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
61 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
62
63 #define PFM_INVALID_ACTIVATION (~0UL)
64
65 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
66 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
67
68 /*
69 * depth of message queue
70 */
71 #define PFM_MAX_MSGS 32
72 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
73
74 /*
75 * type of a PMU register (bitmask).
76 * bitmask structure:
77 * bit0 : register implemented
78 * bit1 : end marker
79 * bit2-3 : reserved
80 * bit4 : pmc has pmc.pm
81 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
82 * bit6-7 : register type
83 * bit8-31: reserved
84 */
85 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
86 #define PFM_REG_IMPL 0x1 /* register implemented */
87 #define PFM_REG_END 0x2 /* end marker */
88 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
89 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
90 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
91 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
92 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
93
94 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
95 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
96
97 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
98
99 /* i assumed unsigned */
100 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
101 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
102
103 /* XXX: these assume that register i is implemented */
104 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
105 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
106 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
107 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
108
109 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
110 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
111 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
112 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
113
114 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
115 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
116
117 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
118 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
119 #define PFM_CTX_TASK(h) (h)->ctx_task
120
121 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
122
123 /* XXX: does not support more than 64 PMDs */
124 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
125 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
126
127 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
128
129 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
130 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
131 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
132 #define PFM_CODE_RR 0 /* requesting code range restriction */
133 #define PFM_DATA_RR 1 /* requestion data range restriction */
134
135 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
136 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
137 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
138
139 #define RDEP(x) (1UL<<(x))
140
141 /*
142 * context protection macros
143 * in SMP:
144 * - we need to protect against CPU concurrency (spin_lock)
145 * - we need to protect against PMU overflow interrupts (local_irq_disable)
146 * in UP:
147 * - we need to protect against PMU overflow interrupts (local_irq_disable)
148 *
149 * spin_lock_irqsave()/spin_unlock_irqrestore():
150 * in SMP: local_irq_disable + spin_lock
151 * in UP : local_irq_disable
152 *
153 * spin_lock()/spin_lock():
154 * in UP : removed automatically
155 * in SMP: protect against context accesses from other CPU. interrupts
156 * are not masked. This is useful for the PMU interrupt handler
157 * because we know we will not get PMU concurrency in that code.
158 */
159 #define PROTECT_CTX(c, f) \
160 do { \
161 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
162 spin_lock_irqsave(&(c)->ctx_lock, f); \
163 DPRINT(("spinlocked ctx %p by [%d]\n", c, task_pid_nr(current))); \
164 } while(0)
165
166 #define UNPROTECT_CTX(c, f) \
167 do { \
168 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
169 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
170 } while(0)
171
172 #define PROTECT_CTX_NOPRINT(c, f) \
173 do { \
174 spin_lock_irqsave(&(c)->ctx_lock, f); \
175 } while(0)
176
177
178 #define UNPROTECT_CTX_NOPRINT(c, f) \
179 do { \
180 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
181 } while(0)
182
183
184 #define PROTECT_CTX_NOIRQ(c) \
185 do { \
186 spin_lock(&(c)->ctx_lock); \
187 } while(0)
188
189 #define UNPROTECT_CTX_NOIRQ(c) \
190 do { \
191 spin_unlock(&(c)->ctx_lock); \
192 } while(0)
193
194
195 #ifdef CONFIG_SMP
196
197 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
198 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
199 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
200
201 #else /* !CONFIG_SMP */
202 #define SET_ACTIVATION(t) do {} while(0)
203 #define GET_ACTIVATION(t) do {} while(0)
204 #define INC_ACTIVATION(t) do {} while(0)
205 #endif /* CONFIG_SMP */
206
207 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
208 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
209 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
210
211 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
212 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
213
214 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
215
216 /*
217 * cmp0 must be the value of pmc0
218 */
219 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
220
221 #define PFMFS_MAGIC 0xa0b4d889
222
223 /*
224 * debugging
225 */
226 #define PFM_DEBUGGING 1
227 #ifdef PFM_DEBUGGING
228 #define DPRINT(a) \
229 do { \
230 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
231 } while (0)
232
233 #define DPRINT_ovfl(a) \
234 do { \
235 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
236 } while (0)
237 #endif
238
239 /*
240 * 64-bit software counter structure
241 *
242 * the next_reset_type is applied to the next call to pfm_reset_regs()
243 */
244 typedef struct {
245 unsigned long val; /* virtual 64bit counter value */
246 unsigned long lval; /* last reset value */
247 unsigned long long_reset; /* reset value on sampling overflow */
248 unsigned long short_reset; /* reset value on overflow */
249 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
250 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
251 unsigned long seed; /* seed for random-number generator */
252 unsigned long mask; /* mask for random-number generator */
253 unsigned int flags; /* notify/do not notify */
254 unsigned long eventid; /* overflow event identifier */
255 } pfm_counter_t;
256
257 /*
258 * context flags
259 */
260 typedef struct {
261 unsigned int block:1; /* when 1, task will blocked on user notifications */
262 unsigned int system:1; /* do system wide monitoring */
263 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
264 unsigned int is_sampling:1; /* true if using a custom format */
265 unsigned int excl_idle:1; /* exclude idle task in system wide session */
266 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
267 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
268 unsigned int no_msg:1; /* no message sent on overflow */
269 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
270 unsigned int reserved:22;
271 } pfm_context_flags_t;
272
273 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
274 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
275 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
276
277
278 /*
279 * perfmon context: encapsulates all the state of a monitoring session
280 */
281
282 typedef struct pfm_context {
283 spinlock_t ctx_lock; /* context protection */
284
285 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
286 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
287
288 struct task_struct *ctx_task; /* task to which context is attached */
289
290 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
291
292 struct completion ctx_restart_done; /* use for blocking notification mode */
293
294 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
295 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
296 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
297
298 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
299 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
300 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
301
302 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
303
304 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
305 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
306 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
307 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
308
309 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
310
311 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
312 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
313
314 u64 ctx_saved_psr_up; /* only contains psr.up value */
315
316 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
317 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
318 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
319
320 int ctx_fd; /* file descriptor used my this context */
321 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
322
323 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
324 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
325 unsigned long ctx_smpl_size; /* size of sampling buffer */
326 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
327
328 wait_queue_head_t ctx_msgq_wait;
329 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
330 int ctx_msgq_head;
331 int ctx_msgq_tail;
332 struct fasync_struct *ctx_async_queue;
333
334 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
335 } pfm_context_t;
336
337 /*
338 * magic number used to verify that structure is really
339 * a perfmon context
340 */
341 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
342
343 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
344
345 #ifdef CONFIG_SMP
346 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
347 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
348 #else
349 #define SET_LAST_CPU(ctx, v) do {} while(0)
350 #define GET_LAST_CPU(ctx) do {} while(0)
351 #endif
352
353
354 #define ctx_fl_block ctx_flags.block
355 #define ctx_fl_system ctx_flags.system
356 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
357 #define ctx_fl_is_sampling ctx_flags.is_sampling
358 #define ctx_fl_excl_idle ctx_flags.excl_idle
359 #define ctx_fl_going_zombie ctx_flags.going_zombie
360 #define ctx_fl_trap_reason ctx_flags.trap_reason
361 #define ctx_fl_no_msg ctx_flags.no_msg
362 #define ctx_fl_can_restart ctx_flags.can_restart
363
364 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
365 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
366
367 /*
368 * global information about all sessions
369 * mostly used to synchronize between system wide and per-process
370 */
371 typedef struct {
372 spinlock_t pfs_lock; /* lock the structure */
373
374 unsigned int pfs_task_sessions; /* number of per task sessions */
375 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
376 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
377 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
378 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
379 } pfm_session_t;
380
381 /*
382 * information about a PMC or PMD.
383 * dep_pmd[]: a bitmask of dependent PMD registers
384 * dep_pmc[]: a bitmask of dependent PMC registers
385 */
386 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
387 typedef struct {
388 unsigned int type;
389 int pm_pos;
390 unsigned long default_value; /* power-on default value */
391 unsigned long reserved_mask; /* bitmask of reserved bits */
392 pfm_reg_check_t read_check;
393 pfm_reg_check_t write_check;
394 unsigned long dep_pmd[4];
395 unsigned long dep_pmc[4];
396 } pfm_reg_desc_t;
397
398 /* assume cnum is a valid monitor */
399 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
400
401 /*
402 * This structure is initialized at boot time and contains
403 * a description of the PMU main characteristics.
404 *
405 * If the probe function is defined, detection is based
406 * on its return value:
407 * - 0 means recognized PMU
408 * - anything else means not supported
409 * When the probe function is not defined, then the pmu_family field
410 * is used and it must match the host CPU family such that:
411 * - cpu->family & config->pmu_family != 0
412 */
413 typedef struct {
414 unsigned long ovfl_val; /* overflow value for counters */
415
416 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
417 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
418
419 unsigned int num_pmcs; /* number of PMCS: computed at init time */
420 unsigned int num_pmds; /* number of PMDS: computed at init time */
421 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
422 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
423
424 char *pmu_name; /* PMU family name */
425 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
426 unsigned int flags; /* pmu specific flags */
427 unsigned int num_ibrs; /* number of IBRS: computed at init time */
428 unsigned int num_dbrs; /* number of DBRS: computed at init time */
429 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
430 int (*probe)(void); /* customized probe routine */
431 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
432 } pmu_config_t;
433 /*
434 * PMU specific flags
435 */
436 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
437
438 /*
439 * debug register related type definitions
440 */
441 typedef struct {
442 unsigned long ibr_mask:56;
443 unsigned long ibr_plm:4;
444 unsigned long ibr_ig:3;
445 unsigned long ibr_x:1;
446 } ibr_mask_reg_t;
447
448 typedef struct {
449 unsigned long dbr_mask:56;
450 unsigned long dbr_plm:4;
451 unsigned long dbr_ig:2;
452 unsigned long dbr_w:1;
453 unsigned long dbr_r:1;
454 } dbr_mask_reg_t;
455
456 typedef union {
457 unsigned long val;
458 ibr_mask_reg_t ibr;
459 dbr_mask_reg_t dbr;
460 } dbreg_t;
461
462
463 /*
464 * perfmon command descriptions
465 */
466 typedef struct {
467 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
468 char *cmd_name;
469 int cmd_flags;
470 unsigned int cmd_narg;
471 size_t cmd_argsize;
472 int (*cmd_getsize)(void *arg, size_t *sz);
473 } pfm_cmd_desc_t;
474
475 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
476 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
477 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
478 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
479
480
481 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
482 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
483 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
484 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
485 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
486
487 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
488
489 typedef struct {
490 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
491 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
492 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
493 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
496 unsigned long pfm_smpl_handler_calls;
497 unsigned long pfm_smpl_handler_cycles;
498 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
499 } pfm_stats_t;
500
501 /*
502 * perfmon internal variables
503 */
504 static pfm_stats_t pfm_stats[NR_CPUS];
505 static pfm_session_t pfm_sessions; /* global sessions information */
506
507 static DEFINE_SPINLOCK(pfm_alt_install_check);
508 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
509
510 static struct proc_dir_entry *perfmon_dir;
511 static pfm_uuid_t pfm_null_uuid = {0,};
512
513 static spinlock_t pfm_buffer_fmt_lock;
514 static LIST_HEAD(pfm_buffer_fmt_list);
515
516 static pmu_config_t *pmu_conf;
517
518 /* sysctl() controls */
519 pfm_sysctl_t pfm_sysctl;
520 EXPORT_SYMBOL(pfm_sysctl);
521
522 static ctl_table pfm_ctl_table[]={
523 {
524 .ctl_name = CTL_UNNUMBERED,
525 .procname = "debug",
526 .data = &pfm_sysctl.debug,
527 .maxlen = sizeof(int),
528 .mode = 0666,
529 .proc_handler = &proc_dointvec,
530 },
531 {
532 .ctl_name = CTL_UNNUMBERED,
533 .procname = "debug_ovfl",
534 .data = &pfm_sysctl.debug_ovfl,
535 .maxlen = sizeof(int),
536 .mode = 0666,
537 .proc_handler = &proc_dointvec,
538 },
539 {
540 .ctl_name = CTL_UNNUMBERED,
541 .procname = "fastctxsw",
542 .data = &pfm_sysctl.fastctxsw,
543 .maxlen = sizeof(int),
544 .mode = 0600,
545 .proc_handler = &proc_dointvec,
546 },
547 {
548 .ctl_name = CTL_UNNUMBERED,
549 .procname = "expert_mode",
550 .data = &pfm_sysctl.expert_mode,
551 .maxlen = sizeof(int),
552 .mode = 0600,
553 .proc_handler = &proc_dointvec,
554 },
555 {}
556 };
557 static ctl_table pfm_sysctl_dir[] = {
558 {
559 .ctl_name = CTL_UNNUMBERED,
560 .procname = "perfmon",
561 .mode = 0555,
562 .child = pfm_ctl_table,
563 },
564 {}
565 };
566 static ctl_table pfm_sysctl_root[] = {
567 {
568 .ctl_name = CTL_KERN,
569 .procname = "kernel",
570 .mode = 0555,
571 .child = pfm_sysctl_dir,
572 },
573 {}
574 };
575 static struct ctl_table_header *pfm_sysctl_header;
576
577 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
578
579 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
580 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
581
582 static inline void
583 pfm_put_task(struct task_struct *task)
584 {
585 if (task != current) put_task_struct(task);
586 }
587
588 static inline void
589 pfm_reserve_page(unsigned long a)
590 {
591 SetPageReserved(vmalloc_to_page((void *)a));
592 }
593 static inline void
594 pfm_unreserve_page(unsigned long a)
595 {
596 ClearPageReserved(vmalloc_to_page((void*)a));
597 }
598
599 static inline unsigned long
600 pfm_protect_ctx_ctxsw(pfm_context_t *x)
601 {
602 spin_lock(&(x)->ctx_lock);
603 return 0UL;
604 }
605
606 static inline void
607 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
608 {
609 spin_unlock(&(x)->ctx_lock);
610 }
611
612 static inline unsigned int
613 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
614 {
615 return do_munmap(mm, addr, len);
616 }
617
618 static inline unsigned long
619 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
620 {
621 return get_unmapped_area(file, addr, len, pgoff, flags);
622 }
623
624
625 static int
626 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data,
627 struct vfsmount *mnt)
628 {
629 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
630 }
631
632 static struct file_system_type pfm_fs_type = {
633 .name = "pfmfs",
634 .get_sb = pfmfs_get_sb,
635 .kill_sb = kill_anon_super,
636 };
637
638 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
639 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
640 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
641 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
642 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
643
644
645 /* forward declaration */
646 static const struct file_operations pfm_file_ops;
647
648 /*
649 * forward declarations
650 */
651 #ifndef CONFIG_SMP
652 static void pfm_lazy_save_regs (struct task_struct *ta);
653 #endif
654
655 void dump_pmu_state(const char *);
656 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
657
658 #include "perfmon_itanium.h"
659 #include "perfmon_mckinley.h"
660 #include "perfmon_montecito.h"
661 #include "perfmon_generic.h"
662
663 static pmu_config_t *pmu_confs[]={
664 &pmu_conf_mont,
665 &pmu_conf_mck,
666 &pmu_conf_ita,
667 &pmu_conf_gen, /* must be last */
668 NULL
669 };
670
671
672 static int pfm_end_notify_user(pfm_context_t *ctx);
673
674 static inline void
675 pfm_clear_psr_pp(void)
676 {
677 ia64_rsm(IA64_PSR_PP);
678 ia64_srlz_i();
679 }
680
681 static inline void
682 pfm_set_psr_pp(void)
683 {
684 ia64_ssm(IA64_PSR_PP);
685 ia64_srlz_i();
686 }
687
688 static inline void
689 pfm_clear_psr_up(void)
690 {
691 ia64_rsm(IA64_PSR_UP);
692 ia64_srlz_i();
693 }
694
695 static inline void
696 pfm_set_psr_up(void)
697 {
698 ia64_ssm(IA64_PSR_UP);
699 ia64_srlz_i();
700 }
701
702 static inline unsigned long
703 pfm_get_psr(void)
704 {
705 unsigned long tmp;
706 tmp = ia64_getreg(_IA64_REG_PSR);
707 ia64_srlz_i();
708 return tmp;
709 }
710
711 static inline void
712 pfm_set_psr_l(unsigned long val)
713 {
714 ia64_setreg(_IA64_REG_PSR_L, val);
715 ia64_srlz_i();
716 }
717
718 static inline void
719 pfm_freeze_pmu(void)
720 {
721 ia64_set_pmc(0,1UL);
722 ia64_srlz_d();
723 }
724
725 static inline void
726 pfm_unfreeze_pmu(void)
727 {
728 ia64_set_pmc(0,0UL);
729 ia64_srlz_d();
730 }
731
732 static inline void
733 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
734 {
735 int i;
736
737 for (i=0; i < nibrs; i++) {
738 ia64_set_ibr(i, ibrs[i]);
739 ia64_dv_serialize_instruction();
740 }
741 ia64_srlz_i();
742 }
743
744 static inline void
745 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
746 {
747 int i;
748
749 for (i=0; i < ndbrs; i++) {
750 ia64_set_dbr(i, dbrs[i]);
751 ia64_dv_serialize_data();
752 }
753 ia64_srlz_d();
754 }
755
756 /*
757 * PMD[i] must be a counter. no check is made
758 */
759 static inline unsigned long
760 pfm_read_soft_counter(pfm_context_t *ctx, int i)
761 {
762 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
763 }
764
765 /*
766 * PMD[i] must be a counter. no check is made
767 */
768 static inline void
769 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
770 {
771 unsigned long ovfl_val = pmu_conf->ovfl_val;
772
773 ctx->ctx_pmds[i].val = val & ~ovfl_val;
774 /*
775 * writing to unimplemented part is ignore, so we do not need to
776 * mask off top part
777 */
778 ia64_set_pmd(i, val & ovfl_val);
779 }
780
781 static pfm_msg_t *
782 pfm_get_new_msg(pfm_context_t *ctx)
783 {
784 int idx, next;
785
786 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
787
788 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
789 if (next == ctx->ctx_msgq_head) return NULL;
790
791 idx = ctx->ctx_msgq_tail;
792 ctx->ctx_msgq_tail = next;
793
794 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
795
796 return ctx->ctx_msgq+idx;
797 }
798
799 static pfm_msg_t *
800 pfm_get_next_msg(pfm_context_t *ctx)
801 {
802 pfm_msg_t *msg;
803
804 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
805
806 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
807
808 /*
809 * get oldest message
810 */
811 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
812
813 /*
814 * and move forward
815 */
816 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
817
818 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));
819
820 return msg;
821 }
822
823 static void
824 pfm_reset_msgq(pfm_context_t *ctx)
825 {
826 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
827 DPRINT(("ctx=%p msgq reset\n", ctx));
828 }
829
830 static void *
831 pfm_rvmalloc(unsigned long size)
832 {
833 void *mem;
834 unsigned long addr;
835
836 size = PAGE_ALIGN(size);
837 mem = vmalloc(size);
838 if (mem) {
839 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
840 memset(mem, 0, size);
841 addr = (unsigned long)mem;
842 while (size > 0) {
843 pfm_reserve_page(addr);
844 addr+=PAGE_SIZE;
845 size-=PAGE_SIZE;
846 }
847 }
848 return mem;
849 }
850
851 static void
852 pfm_rvfree(void *mem, unsigned long size)
853 {
854 unsigned long addr;
855
856 if (mem) {
857 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
858 addr = (unsigned long) mem;
859 while ((long) size > 0) {
860 pfm_unreserve_page(addr);
861 addr+=PAGE_SIZE;
862 size-=PAGE_SIZE;
863 }
864 vfree(mem);
865 }
866 return;
867 }
868
869 static pfm_context_t *
870 pfm_context_alloc(int ctx_flags)
871 {
872 pfm_context_t *ctx;
873
874 /*
875 * allocate context descriptor
876 * must be able to free with interrupts disabled
877 */
878 ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
879 if (ctx) {
880 DPRINT(("alloc ctx @%p\n", ctx));
881
882 /*
883 * init context protection lock
884 */
885 spin_lock_init(&ctx->ctx_lock);
886
887 /*
888 * context is unloaded
889 */
890 ctx->ctx_state = PFM_CTX_UNLOADED;
891
892 /*
893 * initialization of context's flags
894 */
895 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
896 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
897 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
898 /*
899 * will move to set properties
900 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
901 */
902
903 /*
904 * init restart semaphore to locked
905 */
906 init_completion(&ctx->ctx_restart_done);
907
908 /*
909 * activation is used in SMP only
910 */
911 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
912 SET_LAST_CPU(ctx, -1);
913
914 /*
915 * initialize notification message queue
916 */
917 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
918 init_waitqueue_head(&ctx->ctx_msgq_wait);
919 init_waitqueue_head(&ctx->ctx_zombieq);
920
921 }
922 return ctx;
923 }
924
925 static void
926 pfm_context_free(pfm_context_t *ctx)
927 {
928 if (ctx) {
929 DPRINT(("free ctx @%p\n", ctx));
930 kfree(ctx);
931 }
932 }
933
934 static void
935 pfm_mask_monitoring(struct task_struct *task)
936 {
937 pfm_context_t *ctx = PFM_GET_CTX(task);
938 unsigned long mask, val, ovfl_mask;
939 int i;
940
941 DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
942
943 ovfl_mask = pmu_conf->ovfl_val;
944 /*
945 * monitoring can only be masked as a result of a valid
946 * counter overflow. In UP, it means that the PMU still
947 * has an owner. Note that the owner can be different
948 * from the current task. However the PMU state belongs
949 * to the owner.
950 * In SMP, a valid overflow only happens when task is
951 * current. Therefore if we come here, we know that
952 * the PMU state belongs to the current task, therefore
953 * we can access the live registers.
954 *
955 * So in both cases, the live register contains the owner's
956 * state. We can ONLY touch the PMU registers and NOT the PSR.
957 *
958 * As a consequence to this call, the ctx->th_pmds[] array
959 * contains stale information which must be ignored
960 * when context is reloaded AND monitoring is active (see
961 * pfm_restart).
962 */
963 mask = ctx->ctx_used_pmds[0];
964 for (i = 0; mask; i++, mask>>=1) {
965 /* skip non used pmds */
966 if ((mask & 0x1) == 0) continue;
967 val = ia64_get_pmd(i);
968
969 if (PMD_IS_COUNTING(i)) {
970 /*
971 * we rebuild the full 64 bit value of the counter
972 */
973 ctx->ctx_pmds[i].val += (val & ovfl_mask);
974 } else {
975 ctx->ctx_pmds[i].val = val;
976 }
977 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
978 i,
979 ctx->ctx_pmds[i].val,
980 val & ovfl_mask));
981 }
982 /*
983 * mask monitoring by setting the privilege level to 0
984 * we cannot use psr.pp/psr.up for this, it is controlled by
985 * the user
986 *
987 * if task is current, modify actual registers, otherwise modify
988 * thread save state, i.e., what will be restored in pfm_load_regs()
989 */
990 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
991 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
992 if ((mask & 0x1) == 0UL) continue;
993 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
994 ctx->th_pmcs[i] &= ~0xfUL;
995 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
996 }
997 /*
998 * make all of this visible
999 */
1000 ia64_srlz_d();
1001 }
1002
1003 /*
1004 * must always be done with task == current
1005 *
1006 * context must be in MASKED state when calling
1007 */
1008 static void
1009 pfm_restore_monitoring(struct task_struct *task)
1010 {
1011 pfm_context_t *ctx = PFM_GET_CTX(task);
1012 unsigned long mask, ovfl_mask;
1013 unsigned long psr, val;
1014 int i, is_system;
1015
1016 is_system = ctx->ctx_fl_system;
1017 ovfl_mask = pmu_conf->ovfl_val;
1018
1019 if (task != current) {
1020 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1021 return;
1022 }
1023 if (ctx->ctx_state != PFM_CTX_MASKED) {
1024 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1025 task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1026 return;
1027 }
1028 psr = pfm_get_psr();
1029 /*
1030 * monitoring is masked via the PMC.
1031 * As we restore their value, we do not want each counter to
1032 * restart right away. We stop monitoring using the PSR,
1033 * restore the PMC (and PMD) and then re-establish the psr
1034 * as it was. Note that there can be no pending overflow at
1035 * this point, because monitoring was MASKED.
1036 *
1037 * system-wide session are pinned and self-monitoring
1038 */
1039 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1040 /* disable dcr pp */
1041 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1042 pfm_clear_psr_pp();
1043 } else {
1044 pfm_clear_psr_up();
1045 }
1046 /*
1047 * first, we restore the PMD
1048 */
1049 mask = ctx->ctx_used_pmds[0];
1050 for (i = 0; mask; i++, mask>>=1) {
1051 /* skip non used pmds */
1052 if ((mask & 0x1) == 0) continue;
1053
1054 if (PMD_IS_COUNTING(i)) {
1055 /*
1056 * we split the 64bit value according to
1057 * counter width
1058 */
1059 val = ctx->ctx_pmds[i].val & ovfl_mask;
1060 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1061 } else {
1062 val = ctx->ctx_pmds[i].val;
1063 }
1064 ia64_set_pmd(i, val);
1065
1066 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1067 i,
1068 ctx->ctx_pmds[i].val,
1069 val));
1070 }
1071 /*
1072 * restore the PMCs
1073 */
1074 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1075 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1076 if ((mask & 0x1) == 0UL) continue;
1077 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1078 ia64_set_pmc(i, ctx->th_pmcs[i]);
1079 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1080 task_pid_nr(task), i, ctx->th_pmcs[i]));
1081 }
1082 ia64_srlz_d();
1083
1084 /*
1085 * must restore DBR/IBR because could be modified while masked
1086 * XXX: need to optimize
1087 */
1088 if (ctx->ctx_fl_using_dbreg) {
1089 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1090 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1091 }
1092
1093 /*
1094 * now restore PSR
1095 */
1096 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1097 /* enable dcr pp */
1098 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1099 ia64_srlz_i();
1100 }
1101 pfm_set_psr_l(psr);
1102 }
1103
1104 static inline void
1105 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1106 {
1107 int i;
1108
1109 ia64_srlz_d();
1110
1111 for (i=0; mask; i++, mask>>=1) {
1112 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1113 }
1114 }
1115
1116 /*
1117 * reload from thread state (used for ctxw only)
1118 */
1119 static inline void
1120 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1121 {
1122 int i;
1123 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1124
1125 for (i=0; mask; i++, mask>>=1) {
1126 if ((mask & 0x1) == 0) continue;
1127 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1128 ia64_set_pmd(i, val);
1129 }
1130 ia64_srlz_d();
1131 }
1132
1133 /*
1134 * propagate PMD from context to thread-state
1135 */
1136 static inline void
1137 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1138 {
1139 unsigned long ovfl_val = pmu_conf->ovfl_val;
1140 unsigned long mask = ctx->ctx_all_pmds[0];
1141 unsigned long val;
1142 int i;
1143
1144 DPRINT(("mask=0x%lx\n", mask));
1145
1146 for (i=0; mask; i++, mask>>=1) {
1147
1148 val = ctx->ctx_pmds[i].val;
1149
1150 /*
1151 * We break up the 64 bit value into 2 pieces
1152 * the lower bits go to the machine state in the
1153 * thread (will be reloaded on ctxsw in).
1154 * The upper part stays in the soft-counter.
1155 */
1156 if (PMD_IS_COUNTING(i)) {
1157 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1158 val &= ovfl_val;
1159 }
1160 ctx->th_pmds[i] = val;
1161
1162 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1163 i,
1164 ctx->th_pmds[i],
1165 ctx->ctx_pmds[i].val));
1166 }
1167 }
1168
1169 /*
1170 * propagate PMC from context to thread-state
1171 */
1172 static inline void
1173 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1174 {
1175 unsigned long mask = ctx->ctx_all_pmcs[0];
1176 int i;
1177
1178 DPRINT(("mask=0x%lx\n", mask));
1179
1180 for (i=0; mask; i++, mask>>=1) {
1181 /* masking 0 with ovfl_val yields 0 */
1182 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1183 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1184 }
1185 }
1186
1187
1188
1189 static inline void
1190 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1191 {
1192 int i;
1193
1194 for (i=0; mask; i++, mask>>=1) {
1195 if ((mask & 0x1) == 0) continue;
1196 ia64_set_pmc(i, pmcs[i]);
1197 }
1198 ia64_srlz_d();
1199 }
1200
1201 static inline int
1202 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1203 {
1204 return memcmp(a, b, sizeof(pfm_uuid_t));
1205 }
1206
1207 static inline int
1208 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1209 {
1210 int ret = 0;
1211 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1212 return ret;
1213 }
1214
1215 static inline int
1216 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1217 {
1218 int ret = 0;
1219 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1220 return ret;
1221 }
1222
1223
1224 static inline int
1225 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1226 int cpu, void *arg)
1227 {
1228 int ret = 0;
1229 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1230 return ret;
1231 }
1232
1233 static inline int
1234 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1235 int cpu, void *arg)
1236 {
1237 int ret = 0;
1238 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1239 return ret;
1240 }
1241
1242 static inline int
1243 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1244 {
1245 int ret = 0;
1246 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1247 return ret;
1248 }
1249
1250 static inline int
1251 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)
1252 {
1253 int ret = 0;
1254 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1255 return ret;
1256 }
1257
1258 static pfm_buffer_fmt_t *
1259 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1260 {
1261 struct list_head * pos;
1262 pfm_buffer_fmt_t * entry;
1263
1264 list_for_each(pos, &pfm_buffer_fmt_list) {
1265 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1266 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1267 return entry;
1268 }
1269 return NULL;
1270 }
1271
1272 /*
1273 * find a buffer format based on its uuid
1274 */
1275 static pfm_buffer_fmt_t *
1276 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1277 {
1278 pfm_buffer_fmt_t * fmt;
1279 spin_lock(&pfm_buffer_fmt_lock);
1280 fmt = __pfm_find_buffer_fmt(uuid);
1281 spin_unlock(&pfm_buffer_fmt_lock);
1282 return fmt;
1283 }
1284
1285 int
1286 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1287 {
1288 int ret = 0;
1289
1290 /* some sanity checks */
1291 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1292
1293 /* we need at least a handler */
1294 if (fmt->fmt_handler == NULL) return -EINVAL;
1295
1296 /*
1297 * XXX: need check validity of fmt_arg_size
1298 */
1299
1300 spin_lock(&pfm_buffer_fmt_lock);
1301
1302 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1303 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1304 ret = -EBUSY;
1305 goto out;
1306 }
1307 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1308 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1309
1310 out:
1311 spin_unlock(&pfm_buffer_fmt_lock);
1312 return ret;
1313 }
1314 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1315
1316 int
1317 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1318 {
1319 pfm_buffer_fmt_t *fmt;
1320 int ret = 0;
1321
1322 spin_lock(&pfm_buffer_fmt_lock);
1323
1324 fmt = __pfm_find_buffer_fmt(uuid);
1325 if (!fmt) {
1326 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1327 ret = -EINVAL;
1328 goto out;
1329 }
1330 list_del_init(&fmt->fmt_list);
1331 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1332
1333 out:
1334 spin_unlock(&pfm_buffer_fmt_lock);
1335 return ret;
1336
1337 }
1338 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1339
1340 extern void update_pal_halt_status(int);
1341
1342 static int
1343 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1344 {
1345 unsigned long flags;
1346 /*
1347 * validity checks on cpu_mask have been done upstream
1348 */
1349 LOCK_PFS(flags);
1350
1351 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1352 pfm_sessions.pfs_sys_sessions,
1353 pfm_sessions.pfs_task_sessions,
1354 pfm_sessions.pfs_sys_use_dbregs,
1355 is_syswide,
1356 cpu));
1357
1358 if (is_syswide) {
1359 /*
1360 * cannot mix system wide and per-task sessions
1361 */
1362 if (pfm_sessions.pfs_task_sessions > 0UL) {
1363 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1364 pfm_sessions.pfs_task_sessions));
1365 goto abort;
1366 }
1367
1368 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1369
1370 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1371
1372 pfm_sessions.pfs_sys_session[cpu] = task;
1373
1374 pfm_sessions.pfs_sys_sessions++ ;
1375
1376 } else {
1377 if (pfm_sessions.pfs_sys_sessions) goto abort;
1378 pfm_sessions.pfs_task_sessions++;
1379 }
1380
1381 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1382 pfm_sessions.pfs_sys_sessions,
1383 pfm_sessions.pfs_task_sessions,
1384 pfm_sessions.pfs_sys_use_dbregs,
1385 is_syswide,
1386 cpu));
1387
1388 /*
1389 * disable default_idle() to go to PAL_HALT
1390 */
1391 update_pal_halt_status(0);
1392
1393 UNLOCK_PFS(flags);
1394
1395 return 0;
1396
1397 error_conflict:
1398 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1399 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1400 cpu));
1401 abort:
1402 UNLOCK_PFS(flags);
1403
1404 return -EBUSY;
1405
1406 }
1407
1408 static int
1409 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1410 {
1411 unsigned long flags;
1412 /*
1413 * validity checks on cpu_mask have been done upstream
1414 */
1415 LOCK_PFS(flags);
1416
1417 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1418 pfm_sessions.pfs_sys_sessions,
1419 pfm_sessions.pfs_task_sessions,
1420 pfm_sessions.pfs_sys_use_dbregs,
1421 is_syswide,
1422 cpu));
1423
1424
1425 if (is_syswide) {
1426 pfm_sessions.pfs_sys_session[cpu] = NULL;
1427 /*
1428 * would not work with perfmon+more than one bit in cpu_mask
1429 */
1430 if (ctx && ctx->ctx_fl_using_dbreg) {
1431 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1432 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1433 } else {
1434 pfm_sessions.pfs_sys_use_dbregs--;
1435 }
1436 }
1437 pfm_sessions.pfs_sys_sessions--;
1438 } else {
1439 pfm_sessions.pfs_task_sessions--;
1440 }
1441 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1442 pfm_sessions.pfs_sys_sessions,
1443 pfm_sessions.pfs_task_sessions,
1444 pfm_sessions.pfs_sys_use_dbregs,
1445 is_syswide,
1446 cpu));
1447
1448 /*
1449 * if possible, enable default_idle() to go into PAL_HALT
1450 */
1451 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1452 update_pal_halt_status(1);
1453
1454 UNLOCK_PFS(flags);
1455
1456 return 0;
1457 }
1458
1459 /*
1460 * removes virtual mapping of the sampling buffer.
1461 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1462 * a PROTECT_CTX() section.
1463 */
1464 static int
1465 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1466 {
1467 int r;
1468
1469 /* sanity checks */
1470 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1471 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1472 return -EINVAL;
1473 }
1474
1475 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1476
1477 /*
1478 * does the actual unmapping
1479 */
1480 down_write(&task->mm->mmap_sem);
1481
1482 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1483
1484 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1485
1486 up_write(&task->mm->mmap_sem);
1487 if (r !=0) {
1488 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1489 }
1490
1491 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1492
1493 return 0;
1494 }
1495
1496 /*
1497 * free actual physical storage used by sampling buffer
1498 */
1499 #if 0
1500 static int
1501 pfm_free_smpl_buffer(pfm_context_t *ctx)
1502 {
1503 pfm_buffer_fmt_t *fmt;
1504
1505 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1506
1507 /*
1508 * we won't use the buffer format anymore
1509 */
1510 fmt = ctx->ctx_buf_fmt;
1511
1512 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1513 ctx->ctx_smpl_hdr,
1514 ctx->ctx_smpl_size,
1515 ctx->ctx_smpl_vaddr));
1516
1517 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1518
1519 /*
1520 * free the buffer
1521 */
1522 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1523
1524 ctx->ctx_smpl_hdr = NULL;
1525 ctx->ctx_smpl_size = 0UL;
1526
1527 return 0;
1528
1529 invalid_free:
1530 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1531 return -EINVAL;
1532 }
1533 #endif
1534
1535 static inline void
1536 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1537 {
1538 if (fmt == NULL) return;
1539
1540 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1541
1542 }
1543
1544 /*
1545 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1546 * no real gain from having the whole whorehouse mounted. So we don't need
1547 * any operations on the root directory. However, we need a non-trivial
1548 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1549 */
1550 static struct vfsmount *pfmfs_mnt;
1551
1552 static int __init
1553 init_pfm_fs(void)
1554 {
1555 int err = register_filesystem(&pfm_fs_type);
1556 if (!err) {
1557 pfmfs_mnt = kern_mount(&pfm_fs_type);
1558 err = PTR_ERR(pfmfs_mnt);
1559 if (IS_ERR(pfmfs_mnt))
1560 unregister_filesystem(&pfm_fs_type);
1561 else
1562 err = 0;
1563 }
1564 return err;
1565 }
1566
1567 static ssize_t
1568 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1569 {
1570 pfm_context_t *ctx;
1571 pfm_msg_t *msg;
1572 ssize_t ret;
1573 unsigned long flags;
1574 DECLARE_WAITQUEUE(wait, current);
1575 if (PFM_IS_FILE(filp) == 0) {
1576 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1577 return -EINVAL;
1578 }
1579
1580 ctx = (pfm_context_t *)filp->private_data;
1581 if (ctx == NULL) {
1582 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1583 return -EINVAL;
1584 }
1585
1586 /*
1587 * check even when there is no message
1588 */
1589 if (size < sizeof(pfm_msg_t)) {
1590 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1591 return -EINVAL;
1592 }
1593
1594 PROTECT_CTX(ctx, flags);
1595
1596 /*
1597 * put ourselves on the wait queue
1598 */
1599 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1600
1601
1602 for(;;) {
1603 /*
1604 * check wait queue
1605 */
1606
1607 set_current_state(TASK_INTERRUPTIBLE);
1608
1609 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1610
1611 ret = 0;
1612 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1613
1614 UNPROTECT_CTX(ctx, flags);
1615
1616 /*
1617 * check non-blocking read
1618 */
1619 ret = -EAGAIN;
1620 if(filp->f_flags & O_NONBLOCK) break;
1621
1622 /*
1623 * check pending signals
1624 */
1625 if(signal_pending(current)) {
1626 ret = -EINTR;
1627 break;
1628 }
1629 /*
1630 * no message, so wait
1631 */
1632 schedule();
1633
1634 PROTECT_CTX(ctx, flags);
1635 }
1636 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1637 set_current_state(TASK_RUNNING);
1638 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1639
1640 if (ret < 0) goto abort;
1641
1642 ret = -EINVAL;
1643 msg = pfm_get_next_msg(ctx);
1644 if (msg == NULL) {
1645 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1646 goto abort_locked;
1647 }
1648
1649 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1650
1651 ret = -EFAULT;
1652 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1653
1654 abort_locked:
1655 UNPROTECT_CTX(ctx, flags);
1656 abort:
1657 return ret;
1658 }
1659
1660 static ssize_t
1661 pfm_write(struct file *file, const char __user *ubuf,
1662 size_t size, loff_t *ppos)
1663 {
1664 DPRINT(("pfm_write called\n"));
1665 return -EINVAL;
1666 }
1667
1668 static unsigned int
1669 pfm_poll(struct file *filp, poll_table * wait)
1670 {
1671 pfm_context_t *ctx;
1672 unsigned long flags;
1673 unsigned int mask = 0;
1674
1675 if (PFM_IS_FILE(filp) == 0) {
1676 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1677 return 0;
1678 }
1679
1680 ctx = (pfm_context_t *)filp->private_data;
1681 if (ctx == NULL) {
1682 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1683 return 0;
1684 }
1685
1686
1687 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1688
1689 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1690
1691 PROTECT_CTX(ctx, flags);
1692
1693 if (PFM_CTXQ_EMPTY(ctx) == 0)
1694 mask = POLLIN | POLLRDNORM;
1695
1696 UNPROTECT_CTX(ctx, flags);
1697
1698 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1699
1700 return mask;
1701 }
1702
1703 static int
1704 pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1705 {
1706 DPRINT(("pfm_ioctl called\n"));
1707 return -EINVAL;
1708 }
1709
1710 /*
1711 * interrupt cannot be masked when coming here
1712 */
1713 static inline int
1714 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1715 {
1716 int ret;
1717
1718 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1719
1720 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1721 task_pid_nr(current),
1722 fd,
1723 on,
1724 ctx->ctx_async_queue, ret));
1725
1726 return ret;
1727 }
1728
1729 static int
1730 pfm_fasync(int fd, struct file *filp, int on)
1731 {
1732 pfm_context_t *ctx;
1733 int ret;
1734
1735 if (PFM_IS_FILE(filp) == 0) {
1736 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1737 return -EBADF;
1738 }
1739
1740 ctx = (pfm_context_t *)filp->private_data;
1741 if (ctx == NULL) {
1742 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1743 return -EBADF;
1744 }
1745 /*
1746 * we cannot mask interrupts during this call because this may
1747 * may go to sleep if memory is not readily avalaible.
1748 *
1749 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1750 * done in caller. Serialization of this function is ensured by caller.
1751 */
1752 ret = pfm_do_fasync(fd, filp, ctx, on);
1753
1754
1755 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1756 fd,
1757 on,
1758 ctx->ctx_async_queue, ret));
1759
1760 return ret;
1761 }
1762
1763 #ifdef CONFIG_SMP
1764 /*
1765 * this function is exclusively called from pfm_close().
1766 * The context is not protected at that time, nor are interrupts
1767 * on the remote CPU. That's necessary to avoid deadlocks.
1768 */
1769 static void
1770 pfm_syswide_force_stop(void *info)
1771 {
1772 pfm_context_t *ctx = (pfm_context_t *)info;
1773 struct pt_regs *regs = task_pt_regs(current);
1774 struct task_struct *owner;
1775 unsigned long flags;
1776 int ret;
1777
1778 if (ctx->ctx_cpu != smp_processor_id()) {
1779 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1780 ctx->ctx_cpu,
1781 smp_processor_id());
1782 return;
1783 }
1784 owner = GET_PMU_OWNER();
1785 if (owner != ctx->ctx_task) {
1786 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1787 smp_processor_id(),
1788 task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1789 return;
1790 }
1791 if (GET_PMU_CTX() != ctx) {
1792 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1793 smp_processor_id(),
1794 GET_PMU_CTX(), ctx);
1795 return;
1796 }
1797
1798 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1799 /*
1800 * the context is already protected in pfm_close(), we simply
1801 * need to mask interrupts to avoid a PMU interrupt race on
1802 * this CPU
1803 */
1804 local_irq_save(flags);
1805
1806 ret = pfm_context_unload(ctx, NULL, 0, regs);
1807 if (ret) {
1808 DPRINT(("context_unload returned %d\n", ret));
1809 }
1810
1811 /*
1812 * unmask interrupts, PMU interrupts are now spurious here
1813 */
1814 local_irq_restore(flags);
1815 }
1816
1817 static void
1818 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1819 {
1820 int ret;
1821
1822 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1823 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1824 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1825 }
1826 #endif /* CONFIG_SMP */
1827
1828 /*
1829 * called for each close(). Partially free resources.
1830 * When caller is self-monitoring, the context is unloaded.
1831 */
1832 static int
1833 pfm_flush(struct file *filp, fl_owner_t id)
1834 {
1835 pfm_context_t *ctx;
1836 struct task_struct *task;
1837 struct pt_regs *regs;
1838 unsigned long flags;
1839 unsigned long smpl_buf_size = 0UL;
1840 void *smpl_buf_vaddr = NULL;
1841 int state, is_system;
1842
1843 if (PFM_IS_FILE(filp) == 0) {
1844 DPRINT(("bad magic for\n"));
1845 return -EBADF;
1846 }
1847
1848 ctx = (pfm_context_t *)filp->private_data;
1849 if (ctx == NULL) {
1850 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1851 return -EBADF;
1852 }
1853
1854 /*
1855 * remove our file from the async queue, if we use this mode.
1856 * This can be done without the context being protected. We come
1857 * here when the context has become unreachable by other tasks.
1858 *
1859 * We may still have active monitoring at this point and we may
1860 * end up in pfm_overflow_handler(). However, fasync_helper()
1861 * operates with interrupts disabled and it cleans up the
1862 * queue. If the PMU handler is called prior to entering
1863 * fasync_helper() then it will send a signal. If it is
1864 * invoked after, it will find an empty queue and no
1865 * signal will be sent. In both case, we are safe
1866 */
1867 if (filp->f_flags & FASYNC) {
1868 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1869 pfm_do_fasync (-1, filp, ctx, 0);
1870 }
1871
1872 PROTECT_CTX(ctx, flags);
1873
1874 state = ctx->ctx_state;
1875 is_system = ctx->ctx_fl_system;
1876
1877 task = PFM_CTX_TASK(ctx);
1878 regs = task_pt_regs(task);
1879
1880 DPRINT(("ctx_state=%d is_current=%d\n",
1881 state,
1882 task == current ? 1 : 0));
1883
1884 /*
1885 * if state == UNLOADED, then task is NULL
1886 */
1887
1888 /*
1889 * we must stop and unload because we are losing access to the context.
1890 */
1891 if (task == current) {
1892 #ifdef CONFIG_SMP
1893 /*
1894 * the task IS the owner but it migrated to another CPU: that's bad
1895 * but we must handle this cleanly. Unfortunately, the kernel does
1896 * not provide a mechanism to block migration (while the context is loaded).
1897 *
1898 * We need to release the resource on the ORIGINAL cpu.
1899 */
1900 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1901
1902 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1903 /*
1904 * keep context protected but unmask interrupt for IPI
1905 */
1906 local_irq_restore(flags);
1907
1908 pfm_syswide_cleanup_other_cpu(ctx);
1909
1910 /*
1911 * restore interrupt masking
1912 */
1913 local_irq_save(flags);
1914
1915 /*
1916 * context is unloaded at this point
1917 */
1918 } else
1919 #endif /* CONFIG_SMP */
1920 {
1921
1922 DPRINT(("forcing unload\n"));
1923 /*
1924 * stop and unload, returning with state UNLOADED
1925 * and session unreserved.
1926 */
1927 pfm_context_unload(ctx, NULL, 0, regs);
1928
1929 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1930 }
1931 }
1932
1933 /*
1934 * remove virtual mapping, if any, for the calling task.
1935 * cannot reset ctx field until last user is calling close().
1936 *
1937 * ctx_smpl_vaddr must never be cleared because it is needed
1938 * by every task with access to the context
1939 *
1940 * When called from do_exit(), the mm context is gone already, therefore
1941 * mm is NULL, i.e., the VMA is already gone and we do not have to
1942 * do anything here
1943 */
1944 if (ctx->ctx_smpl_vaddr && current->mm) {
1945 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1946 smpl_buf_size = ctx->ctx_smpl_size;
1947 }
1948
1949 UNPROTECT_CTX(ctx, flags);
1950
1951 /*
1952 * if there was a mapping, then we systematically remove it
1953 * at this point. Cannot be done inside critical section
1954 * because some VM function reenables interrupts.
1955 *
1956 */
1957 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1958
1959 return 0;
1960 }
1961 /*
1962 * called either on explicit close() or from exit_files().
1963 * Only the LAST user of the file gets to this point, i.e., it is
1964 * called only ONCE.
1965 *
1966 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1967 * (fput()),i.e, last task to access the file. Nobody else can access the
1968 * file at this point.
1969 *
1970 * When called from exit_files(), the VMA has been freed because exit_mm()
1971 * is executed before exit_files().
1972 *
1973 * When called from exit_files(), the current task is not yet ZOMBIE but we
1974 * flush the PMU state to the context.
1975 */
1976 static int
1977 pfm_close(struct inode *inode, struct file *filp)
1978 {
1979 pfm_context_t *ctx;
1980 struct task_struct *task;
1981 struct pt_regs *regs;
1982 DECLARE_WAITQUEUE(wait, current);
1983 unsigned long flags;
1984 unsigned long smpl_buf_size = 0UL;
1985 void *smpl_buf_addr = NULL;
1986 int free_possible = 1;
1987 int state, is_system;
1988
1989 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1990
1991 if (PFM_IS_FILE(filp) == 0) {
1992 DPRINT(("bad magic\n"));
1993 return -EBADF;
1994 }
1995
1996 ctx = (pfm_context_t *)filp->private_data;
1997 if (ctx == NULL) {
1998 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1999 return -EBADF;
2000 }
2001
2002 PROTECT_CTX(ctx, flags);
2003
2004 state = ctx->ctx_state;
2005 is_system = ctx->ctx_fl_system;
2006
2007 task = PFM_CTX_TASK(ctx);
2008 regs = task_pt_regs(task);
2009
2010 DPRINT(("ctx_state=%d is_current=%d\n",
2011 state,
2012 task == current ? 1 : 0));
2013
2014 /*
2015 * if task == current, then pfm_flush() unloaded the context
2016 */
2017 if (state == PFM_CTX_UNLOADED) goto doit;
2018
2019 /*
2020 * context is loaded/masked and task != current, we need to
2021 * either force an unload or go zombie
2022 */
2023
2024 /*
2025 * The task is currently blocked or will block after an overflow.
2026 * we must force it to wakeup to get out of the
2027 * MASKED state and transition to the unloaded state by itself.
2028 *
2029 * This situation is only possible for per-task mode
2030 */
2031 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2032
2033 /*
2034 * set a "partial" zombie state to be checked
2035 * upon return from down() in pfm_handle_work().
2036 *
2037 * We cannot use the ZOMBIE state, because it is checked
2038 * by pfm_load_regs() which is called upon wakeup from down().
2039 * In such case, it would free the context and then we would
2040 * return to pfm_handle_work() which would access the
2041 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2042 * but visible to pfm_handle_work().
2043 *
2044 * For some window of time, we have a zombie context with
2045 * ctx_state = MASKED and not ZOMBIE
2046 */
2047 ctx->ctx_fl_going_zombie = 1;
2048
2049 /*
2050 * force task to wake up from MASKED state
2051 */
2052 complete(&ctx->ctx_restart_done);
2053
2054 DPRINT(("waking up ctx_state=%d\n", state));
2055
2056 /*
2057 * put ourself to sleep waiting for the other
2058 * task to report completion
2059 *
2060 * the context is protected by mutex, therefore there
2061 * is no risk of being notified of completion before
2062 * begin actually on the waitq.
2063 */
2064 set_current_state(TASK_INTERRUPTIBLE);
2065 add_wait_queue(&ctx->ctx_zombieq, &wait);
2066
2067 UNPROTECT_CTX(ctx, flags);
2068
2069 /*
2070 * XXX: check for signals :
2071 * - ok for explicit close
2072 * - not ok when coming from exit_files()
2073 */
2074 schedule();
2075
2076
2077 PROTECT_CTX(ctx, flags);
2078
2079
2080 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2081 set_current_state(TASK_RUNNING);
2082
2083 /*
2084 * context is unloaded at this point
2085 */
2086 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2087 }
2088 else if (task != current) {
2089 #ifdef CONFIG_SMP
2090 /*
2091 * switch context to zombie state
2092 */
2093 ctx->ctx_state = PFM_CTX_ZOMBIE;
2094
2095 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2096 /*
2097 * cannot free the context on the spot. deferred until
2098 * the task notices the ZOMBIE state
2099 */
2100 free_possible = 0;
2101 #else
2102 pfm_context_unload(ctx, NULL, 0, regs);
2103 #endif
2104 }
2105
2106 doit:
2107 /* reload state, may have changed during opening of critical section */
2108 state = ctx->ctx_state;
2109
2110 /*
2111 * the context is still attached to a task (possibly current)
2112 * we cannot destroy it right now
2113 */
2114
2115 /*
2116 * we must free the sampling buffer right here because
2117 * we cannot rely on it being cleaned up later by the
2118 * monitored task. It is not possible to free vmalloc'ed
2119 * memory in pfm_load_regs(). Instead, we remove the buffer
2120 * now. should there be subsequent PMU overflow originally
2121 * meant for sampling, the will be converted to spurious
2122 * and that's fine because the monitoring tools is gone anyway.
2123 */
2124 if (ctx->ctx_smpl_hdr) {
2125 smpl_buf_addr = ctx->ctx_smpl_hdr;
2126 smpl_buf_size = ctx->ctx_smpl_size;
2127 /* no more sampling */
2128 ctx->ctx_smpl_hdr = NULL;
2129 ctx->ctx_fl_is_sampling = 0;
2130 }
2131
2132 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2133 state,
2134 free_possible,
2135 smpl_buf_addr,
2136 smpl_buf_size));
2137
2138 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2139
2140 /*
2141 * UNLOADED that the session has already been unreserved.
2142 */
2143 if (state == PFM_CTX_ZOMBIE) {
2144 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2145 }
2146
2147 /*
2148 * disconnect file descriptor from context must be done
2149 * before we unlock.
2150 */
2151 filp->private_data = NULL;
2152
2153 /*
2154 * if we free on the spot, the context is now completely unreachable
2155 * from the callers side. The monitored task side is also cut, so we
2156 * can freely cut.
2157 *
2158 * If we have a deferred free, only the caller side is disconnected.
2159 */
2160 UNPROTECT_CTX(ctx, flags);
2161
2162 /*
2163 * All memory free operations (especially for vmalloc'ed memory)
2164 * MUST be done with interrupts ENABLED.
2165 */
2166 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2167
2168 /*
2169 * return the memory used by the context
2170 */
2171 if (free_possible) pfm_context_free(ctx);
2172
2173 return 0;
2174 }
2175
2176 static int
2177 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2178 {
2179 DPRINT(("pfm_no_open called\n"));
2180 return -ENXIO;
2181 }
2182
2183
2184
2185 static const struct file_operations pfm_file_ops = {
2186 .llseek = no_llseek,
2187 .read = pfm_read,
2188 .write = pfm_write,
2189 .poll = pfm_poll,
2190 .ioctl = pfm_ioctl,
2191 .open = pfm_no_open, /* special open code to disallow open via /proc */
2192 .fasync = pfm_fasync,
2193 .release = pfm_close,
2194 .flush = pfm_flush
2195 };
2196
2197 static int
2198 pfmfs_delete_dentry(struct dentry *dentry)
2199 {
2200 return 1;
2201 }
2202
2203 static struct dentry_operations pfmfs_dentry_operations = {
2204 .d_delete = pfmfs_delete_dentry,
2205 };
2206
2207
2208 static struct file *
2209 pfm_alloc_file(pfm_context_t *ctx)
2210 {
2211 struct file *file;
2212 struct inode *inode;
2213 struct dentry *dentry;
2214 char name[32];
2215 struct qstr this;
2216
2217 /*
2218 * allocate a new inode
2219 */
2220 inode = new_inode(pfmfs_mnt->mnt_sb);
2221 if (!inode)
2222 return ERR_PTR(-ENOMEM);
2223
2224 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2225
2226 inode->i_mode = S_IFCHR|S_IRUGO;
2227 inode->i_uid = current->fsuid;
2228 inode->i_gid = current->fsgid;
2229
2230 sprintf(name, "[%lu]", inode->i_ino);
2231 this.name = name;
2232 this.len = strlen(name);
2233 this.hash = inode->i_ino;
2234
2235 /*
2236 * allocate a new dcache entry
2237 */
2238 dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2239 if (!dentry) {
2240 iput(inode);
2241 return ERR_PTR(-ENOMEM);
2242 }
2243
2244 dentry->d_op = &pfmfs_dentry_operations;
2245 d_add(dentry, inode);
2246
2247 file = alloc_file(pfmfs_mnt, dentry, FMODE_READ, &pfm_file_ops);
2248 if (!file) {
2249 dput(dentry);
2250 return ERR_PTR(-ENFILE);
2251 }
2252
2253 file->f_flags = O_RDONLY;
2254 file->private_data = ctx;
2255
2256 return file;
2257 }
2258
2259 static int
2260 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2261 {
2262 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2263
2264 while (size > 0) {
2265 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2266
2267
2268 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2269 return -ENOMEM;
2270
2271 addr += PAGE_SIZE;
2272 buf += PAGE_SIZE;
2273 size -= PAGE_SIZE;
2274 }
2275 return 0;
2276 }
2277
2278 /*
2279 * allocate a sampling buffer and remaps it into the user address space of the task
2280 */
2281 static int
2282 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2283 {
2284 struct mm_struct *mm = task->mm;
2285 struct vm_area_struct *vma = NULL;
2286 unsigned long size;
2287 void *smpl_buf;
2288
2289
2290 /*
2291 * the fixed header + requested size and align to page boundary
2292 */
2293 size = PAGE_ALIGN(rsize);
2294
2295 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2296
2297 /*
2298 * check requested size to avoid Denial-of-service attacks
2299 * XXX: may have to refine this test
2300 * Check against address space limit.
2301 *
2302 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2303 * return -ENOMEM;
2304 */
2305 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2306 return -ENOMEM;
2307
2308 /*
2309 * We do the easy to undo allocations first.
2310 *
2311 * pfm_rvmalloc(), clears the buffer, so there is no leak
2312 */
2313 smpl_buf = pfm_rvmalloc(size);
2314 if (smpl_buf == NULL) {
2315 DPRINT(("Can't allocate sampling buffer\n"));
2316 return -ENOMEM;
2317 }
2318
2319 DPRINT(("smpl_buf @%p\n", smpl_buf));
2320
2321 /* allocate vma */
2322 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2323 if (!vma) {
2324 DPRINT(("Cannot allocate vma\n"));
2325 goto error_kmem;
2326 }
2327
2328 /*
2329 * partially initialize the vma for the sampling buffer
2330 */
2331 vma->vm_mm = mm;
2332 vma->vm_file = filp;
2333 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2334 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2335
2336 /*
2337 * Now we have everything we need and we can initialize
2338 * and connect all the data structures
2339 */
2340
2341 ctx->ctx_smpl_hdr = smpl_buf;
2342 ctx->ctx_smpl_size = size; /* aligned size */
2343
2344 /*
2345 * Let's do the difficult operations next.
2346 *
2347 * now we atomically find some area in the address space and
2348 * remap the buffer in it.
2349 */
2350 down_write(&task->mm->mmap_sem);
2351
2352 /* find some free area in address space, must have mmap sem held */
2353 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2354 if (vma->vm_start == 0UL) {
2355 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2356 up_write(&task->mm->mmap_sem);
2357 goto error;
2358 }
2359 vma->vm_end = vma->vm_start + size;
2360 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2361
2362 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2363
2364 /* can only be applied to current task, need to have the mm semaphore held when called */
2365 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2366 DPRINT(("Can't remap buffer\n"));
2367 up_write(&task->mm->mmap_sem);
2368 goto error;
2369 }
2370
2371 get_file(filp);
2372
2373 /*
2374 * now insert the vma in the vm list for the process, must be
2375 * done with mmap lock held
2376 */
2377 insert_vm_struct(mm, vma);
2378
2379 mm->total_vm += size >> PAGE_SHIFT;
2380 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2381 vma_pages(vma));
2382 up_write(&task->mm->mmap_sem);
2383
2384 /*
2385 * keep track of user level virtual address
2386 */
2387 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2388 *(unsigned long *)user_vaddr = vma->vm_start;
2389
2390 return 0;
2391
2392 error:
2393 kmem_cache_free(vm_area_cachep, vma);
2394 error_kmem:
2395 pfm_rvfree(smpl_buf, size);
2396
2397 return -ENOMEM;
2398 }
2399
2400 /*
2401 * XXX: do something better here
2402 */
2403 static int
2404 pfm_bad_permissions(struct task_struct *task)
2405 {
2406 /* inspired by ptrace_attach() */
2407 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2408 current->uid,
2409 current->gid,
2410 task->euid,
2411 task->suid,
2412 task->uid,
2413 task->egid,
2414 task->sgid));
2415
2416 return ((current->uid != task->euid)
2417 || (current->uid != task->suid)
2418 || (current->uid != task->uid)
2419 || (current->gid != task->egid)
2420 || (current->gid != task->sgid)
2421 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2422 }
2423
2424 static int
2425 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2426 {
2427 int ctx_flags;
2428
2429 /* valid signal */
2430
2431 ctx_flags = pfx->ctx_flags;
2432
2433 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2434
2435 /*
2436 * cannot block in this mode
2437 */
2438 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2439 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2440 return -EINVAL;
2441 }
2442 } else {
2443 }
2444 /* probably more to add here */
2445
2446 return 0;
2447 }
2448
2449 static int
2450 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2451 unsigned int cpu, pfarg_context_t *arg)
2452 {
2453 pfm_buffer_fmt_t *fmt = NULL;
2454 unsigned long size = 0UL;
2455 void *uaddr = NULL;
2456 void *fmt_arg = NULL;
2457 int ret = 0;
2458 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2459
2460 /* invoke and lock buffer format, if found */
2461 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2462 if (fmt == NULL) {
2463 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2464 return -EINVAL;
2465 }
2466
2467 /*
2468 * buffer argument MUST be contiguous to pfarg_context_t
2469 */
2470 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2471
2472 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2473
2474 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2475
2476 if (ret) goto error;
2477
2478 /* link buffer format and context */
2479 ctx->ctx_buf_fmt = fmt;
2480 ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2481
2482 /*
2483 * check if buffer format wants to use perfmon buffer allocation/mapping service
2484 */
2485 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2486 if (ret) goto error;
2487
2488 if (size) {
2489 /*
2490 * buffer is always remapped into the caller's address space
2491 */
2492 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2493 if (ret) goto error;
2494
2495 /* keep track of user address of buffer */
2496 arg->ctx_smpl_vaddr = uaddr;
2497 }
2498 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2499
2500 error:
2501 return ret;
2502 }
2503
2504 static void
2505 pfm_reset_pmu_state(pfm_context_t *ctx)
2506 {
2507 int i;
2508
2509 /*
2510 * install reset values for PMC.
2511 */
2512 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2513 if (PMC_IS_IMPL(i) == 0) continue;
2514 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2515 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2516 }
2517 /*
2518 * PMD registers are set to 0UL when the context in memset()
2519 */
2520
2521 /*
2522 * On context switched restore, we must restore ALL pmc and ALL pmd even
2523 * when they are not actively used by the task. In UP, the incoming process
2524 * may otherwise pick up left over PMC, PMD state from the previous process.
2525 * As opposed to PMD, stale PMC can cause harm to the incoming
2526 * process because they may change what is being measured.
2527 * Therefore, we must systematically reinstall the entire
2528 * PMC state. In SMP, the same thing is possible on the
2529 * same CPU but also on between 2 CPUs.
2530 *
2531 * The problem with PMD is information leaking especially
2532 * to user level when psr.sp=0
2533 *
2534 * There is unfortunately no easy way to avoid this problem
2535 * on either UP or SMP. This definitively slows down the
2536 * pfm_load_regs() function.
2537 */
2538
2539 /*
2540 * bitmask of all PMCs accessible to this context
2541 *
2542 * PMC0 is treated differently.
2543 */
2544 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2545
2546 /*
2547 * bitmask of all PMDs that are accessible to this context
2548 */
2549 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2550
2551 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2552
2553 /*
2554 * useful in case of re-enable after disable
2555 */
2556 ctx->ctx_used_ibrs[0] = 0UL;
2557 ctx->ctx_used_dbrs[0] = 0UL;
2558 }
2559
2560 static int
2561 pfm_ctx_getsize(void *arg, size_t *sz)
2562 {
2563 pfarg_context_t *req = (pfarg_context_t *)arg;
2564 pfm_buffer_fmt_t *fmt;
2565
2566 *sz = 0;
2567
2568 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2569
2570 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2571 if (fmt == NULL) {
2572 DPRINT(("cannot find buffer format\n"));
2573 return -EINVAL;
2574 }
2575 /* get just enough to copy in user parameters */
2576 *sz = fmt->fmt_arg_size;
2577 DPRINT(("arg_size=%lu\n", *sz));
2578
2579 return 0;
2580 }
2581
2582
2583
2584 /*
2585 * cannot attach if :
2586 * - kernel task
2587 * - task not owned by caller
2588 * - task incompatible with context mode
2589 */
2590 static int
2591 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2592 {
2593 /*
2594 * no kernel task or task not owner by caller
2595 */
2596 if (task->mm == NULL) {
2597 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2598 return -EPERM;
2599 }
2600 if (pfm_bad_permissions(task)) {
2601 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task)));
2602 return -EPERM;
2603 }
2604 /*
2605 * cannot block in self-monitoring mode
2606 */
2607 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2608 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2609 return -EINVAL;
2610 }
2611
2612 if (task->exit_state == EXIT_ZOMBIE) {
2613 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task)));
2614 return -EBUSY;
2615 }
2616
2617 /*
2618 * always ok for self
2619 */
2620 if (task == current) return 0;
2621
2622 if (!task_is_stopped_or_traced(task)) {
2623 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2624 return -EBUSY;
2625 }
2626 /*
2627 * make sure the task is off any CPU
2628 */
2629 wait_task_inactive(task);
2630
2631 /* more to come... */
2632
2633 return 0;
2634 }
2635
2636 static int
2637 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2638 {
2639 struct task_struct *p = current;
2640 int ret;
2641
2642 /* XXX: need to add more checks here */
2643 if (pid < 2) return -EPERM;
2644
2645 if (pid != task_pid_vnr(current)) {
2646
2647 read_lock(&tasklist_lock);
2648
2649 p = find_task_by_vpid(pid);
2650
2651 /* make sure task cannot go away while we operate on it */
2652 if (p) get_task_struct(p);
2653
2654 read_unlock(&tasklist_lock);
2655
2656 if (p == NULL) return -ESRCH;
2657 }
2658
2659 ret = pfm_task_incompatible(ctx, p);
2660 if (ret == 0) {
2661 *task = p;
2662 } else if (p != current) {
2663 pfm_put_task(p);
2664 }
2665 return ret;
2666 }
2667
2668
2669
2670 static int
2671 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2672 {
2673 pfarg_context_t *req = (pfarg_context_t *)arg;
2674 struct file *filp;
2675 struct path path;
2676 int ctx_flags;
2677 int fd;
2678 int ret;
2679
2680 /* let's check the arguments first */
2681 ret = pfarg_is_sane(current, req);
2682 if (ret < 0)
2683 return ret;
2684
2685 ctx_flags = req->ctx_flags;
2686
2687 ret = -ENOMEM;
2688
2689 fd = get_unused_fd();
2690 if (fd < 0)
2691 return fd;
2692
2693 ctx = pfm_context_alloc(ctx_flags);
2694 if (!ctx)
2695 goto error;
2696
2697 filp = pfm_alloc_file(ctx);
2698 if (IS_ERR(filp)) {
2699 ret = PTR_ERR(filp);
2700 goto error_file;
2701 }
2702
2703 req->ctx_fd = ctx->ctx_fd = fd;
2704
2705 /*
2706 * does the user want to sample?
2707 */
2708 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2709 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2710 if (ret)
2711 goto buffer_error;
2712 }
2713
2714 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2715 ctx,
2716 ctx_flags,
2717 ctx->ctx_fl_system,
2718 ctx->ctx_fl_block,
2719 ctx->ctx_fl_excl_idle,
2720 ctx->ctx_fl_no_msg,
2721 ctx->ctx_fd));
2722
2723 /*
2724 * initialize soft PMU state
2725 */
2726 pfm_reset_pmu_state(ctx);
2727
2728 fd_install(fd, filp);
2729
2730 return 0;
2731
2732 buffer_error:
2733 path = filp->f_path;
2734 put_filp(filp);
2735 path_put(&path);
2736
2737 if (ctx->ctx_buf_fmt) {
2738 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2739 }
2740 error_file:
2741 pfm_context_free(ctx);
2742
2743 error:
2744 put_unused_fd(fd);
2745 return ret;
2746 }
2747
2748 static inline unsigned long
2749 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2750 {
2751 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2752 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2753 extern unsigned long carta_random32 (unsigned long seed);
2754
2755 if (reg->flags & PFM_REGFL_RANDOM) {
2756 new_seed = carta_random32(old_seed);
2757 val -= (old_seed & mask); /* counter values are negative numbers! */
2758 if ((mask >> 32) != 0)
2759 /* construct a full 64-bit random value: */
2760 new_seed |= carta_random32(old_seed >> 32) << 32;
2761 reg->seed = new_seed;
2762 }
2763 reg->lval = val;
2764 return val;
2765 }
2766
2767 static void
2768 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2769 {
2770 unsigned long mask = ovfl_regs[0];
2771 unsigned long reset_others = 0UL;
2772 unsigned long val;
2773 int i;
2774
2775 /*
2776 * now restore reset value on sampling overflowed counters
2777 */
2778 mask >>= PMU_FIRST_COUNTER;
2779 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2780
2781 if ((mask & 0x1UL) == 0UL) continue;
2782
2783 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2784 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2785
2786 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2787 }
2788
2789 /*
2790 * Now take care of resetting the other registers
2791 */
2792 for(i = 0; reset_others; i++, reset_others >>= 1) {
2793
2794 if ((reset_others & 0x1) == 0) continue;
2795
2796 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2797
2798 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2799 is_long_reset ? "long" : "short", i, val));
2800 }
2801 }
2802
2803 static void
2804 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2805 {
2806 unsigned long mask = ovfl_regs[0];
2807 unsigned long reset_others = 0UL;
2808 unsigned long val;
2809 int i;
2810
2811 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2812
2813 if (ctx->ctx_state == PFM_CTX_MASKED) {
2814 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2815 return;
2816 }
2817
2818 /*
2819 * now restore reset value on sampling overflowed counters
2820 */
2821 mask >>= PMU_FIRST_COUNTER;
2822 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2823
2824 if ((mask & 0x1UL) == 0UL) continue;
2825
2826 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2827 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2828
2829 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2830
2831 pfm_write_soft_counter(ctx, i, val);
2832 }
2833
2834 /*
2835 * Now take care of resetting the other registers
2836 */
2837 for(i = 0; reset_others; i++, reset_others >>= 1) {
2838
2839 if ((reset_others & 0x1) == 0) continue;
2840
2841 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2842
2843 if (PMD_IS_COUNTING(i)) {
2844 pfm_write_soft_counter(ctx, i, val);
2845 } else {
2846 ia64_set_pmd(i, val);
2847 }
2848 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2849 is_long_reset ? "long" : "short", i, val));
2850 }
2851 ia64_srlz_d();
2852 }
2853
2854 static int
2855 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2856 {
2857 struct task_struct *task;
2858 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2859 unsigned long value, pmc_pm;
2860 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2861 unsigned int cnum, reg_flags, flags, pmc_type;
2862 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2863 int is_monitor, is_counting, state;
2864 int ret = -EINVAL;
2865 pfm_reg_check_t wr_func;
2866 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2867
2868 state = ctx->ctx_state;
2869 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2870 is_system = ctx->ctx_fl_system;
2871 task = ctx->ctx_task;
2872 impl_pmds = pmu_conf->impl_pmds[0];
2873
2874 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2875
2876 if (is_loaded) {
2877 /*
2878 * In system wide and when the context is loaded, access can only happen
2879 * when the caller is running on the CPU being monitored by the session.
2880 * It does not have to be the owner (ctx_task) of the context per se.
2881 */
2882 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2883 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2884 return -EBUSY;
2885 }
2886 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2887 }
2888 expert_mode = pfm_sysctl.expert_mode;
2889
2890 for (i = 0; i < count; i++, req++) {
2891
2892 cnum = req->reg_num;
2893 reg_flags = req->reg_flags;
2894 value = req->reg_value;
2895 smpl_pmds = req->reg_smpl_pmds[0];
2896 reset_pmds = req->reg_reset_pmds[0];
2897 flags = 0;
2898
2899
2900 if (cnum >= PMU_MAX_PMCS) {
2901 DPRINT(("pmc%u is invalid\n", cnum));
2902 goto error;
2903 }
2904
2905 pmc_type = pmu_conf->pmc_desc[cnum].type;
2906 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2907 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2908 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2909
2910 /*
2911 * we reject all non implemented PMC as well
2912 * as attempts to modify PMC[0-3] which are used
2913 * as status registers by the PMU
2914 */
2915 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2916 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2917 goto error;
2918 }
2919 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2920 /*
2921 * If the PMC is a monitor, then if the value is not the default:
2922 * - system-wide session: PMCx.pm=1 (privileged monitor)
2923 * - per-task : PMCx.pm=0 (user monitor)
2924 */
2925 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2926 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2927 cnum,
2928 pmc_pm,
2929 is_system));
2930 goto error;
2931 }
2932
2933 if (is_counting) {
2934 /*
2935 * enforce generation of overflow interrupt. Necessary on all
2936 * CPUs.
2937 */
2938 value |= 1 << PMU_PMC_OI;
2939
2940 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2941 flags |= PFM_REGFL_OVFL_NOTIFY;
2942 }
2943
2944 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2945
2946 /* verify validity of smpl_pmds */
2947 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2948 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2949 goto error;
2950 }
2951
2952 /* verify validity of reset_pmds */
2953 if ((reset_pmds & impl_pmds) != reset_pmds) {
2954 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2955 goto error;
2956 }
2957 } else {
2958 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2959 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2960 goto error;
2961 }
2962 /* eventid on non-counting monitors are ignored */
2963 }
2964
2965 /*
2966 * execute write checker, if any
2967 */
2968 if (likely(expert_mode == 0 && wr_func)) {
2969 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2970 if (ret) goto error;
2971 ret = -EINVAL;
2972 }
2973
2974 /*
2975 * no error on this register
2976 */
2977 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2978
2979 /*
2980 * Now we commit the changes to the software state
2981 */
2982
2983 /*
2984 * update overflow information
2985 */
2986 if (is_counting) {
2987 /*
2988 * full flag update each time a register is programmed
2989 */
2990 ctx->ctx_pmds[cnum].flags = flags;
2991
2992 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2993 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2994 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2995
2996 /*
2997 * Mark all PMDS to be accessed as used.
2998 *
2999 * We do not keep track of PMC because we have to
3000 * systematically restore ALL of them.
3001 *
3002 * We do not update the used_monitors mask, because
3003 * if we have not programmed them, then will be in
3004 * a quiescent state, therefore we will not need to
3005 * mask/restore then when context is MASKED.
3006 */
3007 CTX_USED_PMD(ctx, reset_pmds);
3008 CTX_USED_PMD(ctx, smpl_pmds);
3009 /*
3010 * make sure we do not try to reset on
3011 * restart because we have established new values
3012 */
3013 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3014 }
3015 /*
3016 * Needed in case the user does not initialize the equivalent
3017 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3018 * possible leak here.
3019 */
3020 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3021
3022 /*
3023 * keep track of the monitor PMC that we are using.
3024 * we save the value of the pmc in ctx_pmcs[] and if
3025 * the monitoring is not stopped for the context we also
3026 * place it in the saved state area so that it will be
3027 * picked up later by the context switch code.
3028 *
3029 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3030 *
3031 * The value in th_pmcs[] may be modified on overflow, i.e., when
3032 * monitoring needs to be stopped.
3033 */
3034 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3035
3036 /*
3037 * update context state
3038 */
3039 ctx->ctx_pmcs[cnum] = value;
3040
3041 if (is_loaded) {
3042 /*
3043 * write thread state
3044 */
3045 if (is_system == 0) ctx->th_pmcs[cnum] = value;
3046
3047 /*
3048 * write hardware register if we can
3049 */
3050 if (can_access_pmu) {
3051 ia64_set_pmc(cnum, value);
3052 }
3053 #ifdef CONFIG_SMP
3054 else {
3055 /*
3056 * per-task SMP only here
3057 *
3058 * we are guaranteed that the task is not running on the other CPU,
3059 * we indicate that this PMD will need to be reloaded if the task
3060 * is rescheduled on the CPU it ran last on.
3061 */
3062 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3063 }
3064 #endif
3065 }
3066
3067 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",
3068 cnum,
3069 value,
3070 is_loaded,
3071 can_access_pmu,
3072 flags,
3073 ctx->ctx_all_pmcs[0],
3074 ctx->ctx_used_pmds[0],
3075 ctx->ctx_pmds[cnum].eventid,
3076 smpl_pmds,
3077 reset_pmds,
3078 ctx->ctx_reload_pmcs[0],
3079 ctx->ctx_used_monitors[0],
3080 ctx->ctx_ovfl_regs[0]));
3081 }
3082
3083 /*
3084 * make sure the changes are visible
3085 */
3086 if (can_access_pmu) ia64_srlz_d();
3087
3088 return 0;
3089 error:
3090 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3091 return ret;
3092 }
3093
3094 static int
3095 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3096 {
3097 struct task_struct *task;
3098 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3099 unsigned long value, hw_value, ovfl_mask;
3100 unsigned int cnum;
3101 int i, can_access_pmu = 0, state;
3102 int is_counting, is_loaded, is_system, expert_mode;
3103 int ret = -EINVAL;
3104 pfm_reg_check_t wr_func;
3105
3106
3107 state = ctx->ctx_state;
3108 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3109 is_system = ctx->ctx_fl_system;
3110 ovfl_mask = pmu_conf->ovfl_val;
3111 task = ctx->ctx_task;
3112
3113 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3114
3115 /*
3116 * on both UP and SMP, we can only write to the PMC when the task is
3117 * the owner of the local PMU.
3118 */
3119 if (likely(is_loaded)) {
3120 /*
3121 * In system wide and when the context is loaded, access can only happen
3122 * when the caller is running on the CPU being monitored by the session.
3123 * It does not have to be the owner (ctx_task) of the context per se.
3124 */
3125 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3126 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3127 return -EBUSY;
3128 }
3129 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3130 }
3131 expert_mode = pfm_sysctl.expert_mode;
3132
3133 for (i = 0; i < count; i++, req++) {
3134
3135 cnum = req->reg_num;
3136 value = req->reg_value;
3137
3138 if (!PMD_IS_IMPL(cnum)) {
3139 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3140 goto abort_mission;
3141 }
3142 is_counting = PMD_IS_COUNTING(cnum);
3143 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3144
3145 /*
3146 * execute write checker, if any
3147 */
3148 if (unlikely(expert_mode == 0 && wr_func)) {
3149 unsigned long v = value;
3150
3151 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3152 if (ret) goto abort_mission;
3153
3154 value = v;
3155 ret = -EINVAL;
3156 }
3157
3158 /*
3159 * no error on this register
3160 */
3161 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3162
3163 /*
3164 * now commit changes to software state
3165 */
3166 hw_value = value;
3167
3168 /*
3169 * update virtualized (64bits) counter
3170 */
3171 if (is_counting) {
3172 /*
3173 * write context state
3174 */
3175 ctx->ctx_pmds[cnum].lval = value;
3176
3177 /*
3178 * when context is load we use the split value
3179 */
3180 if (is_loaded) {
3181 hw_value = value & ovfl_mask;
3182 value = value & ~ovfl_mask;
3183 }
3184 }
3185 /*
3186 * update reset values (not just for counters)
3187 */
3188 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3189 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3190
3191 /*
3192 * update randomization parameters (not just for counters)
3193 */
3194 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3195 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3196
3197 /*
3198 * update context value
3199 */
3200 ctx->ctx_pmds[cnum].val = value;
3201
3202 /*
3203 * Keep track of what we use
3204 *
3205 * We do not keep track of PMC because we have to
3206 * systematically restore ALL of them.
3207 */
3208 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3209
3210 /*
3211 * mark this PMD register used as well
3212 */
3213 CTX_USED_PMD(ctx, RDEP(cnum));
3214
3215 /*
3216 * make sure we do not try to reset on
3217 * restart because we have established new values
3218 */
3219 if (is_counting && state == PFM_CTX_MASKED) {
3220 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3221 }
3222
3223 if (is_loaded) {
3224 /*
3225 * write thread state
3226 */
3227 if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3228
3229 /*
3230 * write hardware register if we can
3231 */
3232 if (can_access_pmu) {
3233 ia64_set_pmd(cnum, hw_value);
3234 } else {
3235 #ifdef CONFIG_SMP
3236 /*
3237 * we are guaranteed that the task is not running on the other CPU,
3238 * we indicate that this PMD will need to be reloaded if the task
3239 * is rescheduled on the CPU it ran last on.
3240 */
3241 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3242 #endif
3243 }
3244 }
3245
3246 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3247 "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",
3248 cnum,
3249 value,
3250 is_loaded,
3251 can_access_pmu,
3252 hw_value,
3253 ctx->ctx_pmds[cnum].val,
3254 ctx->ctx_pmds[cnum].short_reset,
3255 ctx->ctx_pmds[cnum].long_reset,
3256 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3257 ctx->ctx_pmds[cnum].seed,
3258 ctx->ctx_pmds[cnum].mask,
3259 ctx->ctx_used_pmds[0],
3260 ctx->ctx_pmds[cnum].reset_pmds[0],
3261 ctx->ctx_reload_pmds[0],
3262 ctx->ctx_all_pmds[0],
3263 ctx->ctx_ovfl_regs[0]));
3264 }
3265
3266 /*
3267 * make changes visible
3268 */
3269 if (can_access_pmu) ia64_srlz_d();
3270
3271 return 0;
3272
3273 abort_mission:
3274 /*
3275 * for now, we have only one possibility for error
3276 */
3277 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3278 return ret;
3279 }
3280
3281 /*
3282 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3283 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3284 * interrupt is delivered during the call, it will be kept pending until we leave, making
3285 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3286 * guaranteed to return consistent data to the user, it may simply be old. It is not
3287 * trivial to treat the overflow while inside the call because you may end up in
3288 * some module sampling buffer code causing deadlocks.
3289 */
3290 static int
3291 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3292 {
3293 struct task_struct *task;
3294 unsigned long val = 0UL, lval, ovfl_mask, sval;
3295 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3296 unsigned int cnum, reg_flags = 0;
3297 int i, can_access_pmu = 0, state;
3298 int is_loaded, is_system, is_counting, expert_mode;
3299 int ret = -EINVAL;
3300 pfm_reg_check_t rd_func;
3301
3302 /*
3303 * access is possible when loaded only for
3304 * self-monitoring tasks or in UP mode
3305 */
3306
3307 state = ctx->ctx_state;
3308 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3309 is_system = ctx->ctx_fl_system;
3310 ovfl_mask = pmu_conf->ovfl_val;
3311 task = ctx->ctx_task;
3312
3313 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3314
3315 if (likely(is_loaded)) {
3316 /*
3317 * In system wide and when the context is loaded, access can only happen
3318 * when the caller is running on the CPU being monitored by the session.
3319 * It does not have to be the owner (ctx_task) of the context per se.
3320 */
3321 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3322 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3323 return -EBUSY;
3324 }
3325 /*
3326 * this can be true when not self-monitoring only in UP
3327 */
3328 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3329
3330 if (can_access_pmu) ia64_srlz_d();
3331 }
3332 expert_mode = pfm_sysctl.expert_mode;
3333
3334 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3335 is_loaded,
3336 can_access_pmu,
3337 state));
3338
3339 /*
3340 * on both UP and SMP, we can only read the PMD from the hardware register when
3341 * the task is the owner of the local PMU.
3342 */
3343
3344 for (i = 0; i < count; i++, req++) {
3345
3346 cnum = req->reg_num;
3347 reg_flags = req->reg_flags;
3348
3349 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3350 /*
3351 * we can only read the register that we use. That includes
3352 * the one we explicitly initialize AND the one we want included
3353 * in the sampling buffer (smpl_regs).
3354 *
3355 * Having this restriction allows optimization in the ctxsw routine
3356 * without compromising security (leaks)
3357 */
3358 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3359
3360 sval = ctx->ctx_pmds[cnum].val;
3361 lval = ctx->ctx_pmds[cnum].lval;
3362 is_counting = PMD_IS_COUNTING(cnum);
3363
3364 /*
3365 * If the task is not the current one, then we check if the
3366 * PMU state is still in the local live register due to lazy ctxsw.
3367 * If true, then we read directly from the registers.
3368 */
3369 if (can_access_pmu){
3370 val = ia64_get_pmd(cnum);
3371 } else {
3372 /*
3373 * context has been saved
3374 * if context is zombie, then task does not exist anymore.
3375 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3376 */
3377 val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3378 }
3379 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3380
3381 if (is_counting) {
3382 /*
3383 * XXX: need to check for overflow when loaded
3384 */
3385 val &= ovfl_mask;
3386 val += sval;
3387 }
3388
3389 /*
3390 * execute read checker, if any
3391 */
3392 if (unlikely(expert_mode == 0 && rd_func)) {
3393 unsigned long v = val;
3394 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3395 if (ret) goto error;
3396 val = v;
3397 ret = -EINVAL;
3398 }
3399
3400 PFM_REG_RETFLAG_SET(reg_flags, 0);
3401
3402 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3403
3404 /*
3405 * update register return value, abort all if problem during copy.
3406 * we only modify the reg_flags field. no check mode is fine because
3407 * access has been verified upfront in sys_perfmonctl().
3408 */
3409 req->reg_value = val;
3410 req->reg_flags = reg_flags;
3411 req->reg_last_reset_val = lval;
3412 }
3413
3414 return 0;
3415
3416 error:
3417 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3418 return ret;
3419 }
3420
3421 int
3422 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3423 {
3424 pfm_context_t *ctx;
3425
3426 if (req == NULL) return -EINVAL;
3427
3428 ctx = GET_PMU_CTX();
3429
3430 if (ctx == NULL) return -EINVAL;
3431
3432 /*
3433 * for now limit to current task, which is enough when calling
3434 * from overflow handler
3435 */
3436 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3437
3438 return pfm_write_pmcs(ctx, req, nreq, regs);
3439 }
3440 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3441
3442 int
3443 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3444 {
3445 pfm_context_t *ctx;
3446
3447 if (req == NULL) return -EINVAL;
3448
3449 ctx = GET_PMU_CTX();
3450
3451 if (ctx == NULL) return -EINVAL;
3452
3453 /*
3454 * for now limit to current task, which is enough when calling
3455 * from overflow handler
3456 */
3457 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3458
3459 return pfm_read_pmds(ctx, req, nreq, regs);
3460 }
3461 EXPORT_SYMBOL(pfm_mod_read_pmds);
3462
3463 /*
3464 * Only call this function when a process it trying to
3465 * write the debug registers (reading is always allowed)
3466 */
3467 int
3468 pfm_use_debug_registers(struct task_struct *task)
3469 {
3470 pfm_context_t *ctx = task->thread.pfm_context;
3471 unsigned long flags;
3472 int ret = 0;
3473
3474 if (pmu_conf->use_rr_dbregs == 0) return 0;
3475
3476 DPRINT(("called for [%d]\n", task_pid_nr(task)));
3477
3478 /*
3479 * do it only once
3480 */
3481 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3482
3483 /*
3484 * Even on SMP, we do not need to use an atomic here because
3485 * the only way in is via ptrace() and this is possible only when the
3486 * process is stopped. Even in the case where the ctxsw out is not totally
3487 * completed by the time we come here, there is no way the 'stopped' process
3488 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3489 * So this is always safe.
3490 */
3491 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3492
3493 LOCK_PFS(flags);
3494
3495 /*
3496 * We cannot allow setting breakpoints when system wide monitoring
3497 * sessions are using the debug registers.
3498 */
3499 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3500 ret = -1;
3501 else
3502 pfm_sessions.pfs_ptrace_use_dbregs++;
3503
3504 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3505 pfm_sessions.pfs_ptrace_use_dbregs,
3506 pfm_sessions.pfs_sys_use_dbregs,
3507 task_pid_nr(task), ret));
3508
3509 UNLOCK_PFS(flags);
3510
3511 return ret;
3512 }
3513
3514 /*
3515 * This function is called for every task that exits with the
3516 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3517 * able to use the debug registers for debugging purposes via
3518 * ptrace(). Therefore we know it was not using them for
3519 * perfmormance monitoring, so we only decrement the number
3520 * of "ptraced" debug register users to keep the count up to date
3521 */
3522 int
3523 pfm_release_debug_registers(struct task_struct *task)
3524 {
3525 unsigned long flags;
3526 int ret;
3527
3528 if (pmu_conf->use_rr_dbregs == 0) return 0;
3529
3530 LOCK_PFS(flags);
3531 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3532 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3533 ret = -1;
3534 } else {
3535 pfm_sessions.pfs_ptrace_use_dbregs--;
3536 ret = 0;
3537 }
3538 UNLOCK_PFS(flags);
3539
3540 return ret;
3541 }
3542
3543 static int
3544 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3545 {
3546 struct task_struct *task;
3547 pfm_buffer_fmt_t *fmt;
3548 pfm_ovfl_ctrl_t rst_ctrl;
3549 int state, is_system;
3550 int ret = 0;
3551
3552 state = ctx->ctx_state;
3553 fmt = ctx->ctx_buf_fmt;
3554 is_system = ctx->ctx_fl_system;
3555 task = PFM_CTX_TASK(ctx);
3556
3557 switch(state) {
3558 case PFM_CTX_MASKED:
3559 break;
3560 case PFM_CTX_LOADED:
3561 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3562 /* fall through */
3563 case PFM_CTX_UNLOADED:
3564 case PFM_CTX_ZOMBIE:
3565 DPRINT(("invalid state=%d\n", state));
3566 return -EBUSY;
3567 default:
3568 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3569 return -EINVAL;
3570 }
3571
3572 /*
3573 * In system wide and when the context is loaded, access can only happen
3574 * when the caller is running on the CPU being monitored by the session.
3575 * It does not have to be the owner (ctx_task) of the context per se.
3576 */
3577 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3578 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3579 return -EBUSY;
3580 }
3581
3582 /* sanity check */
3583 if (unlikely(task == NULL)) {
3584 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3585 return -EINVAL;
3586 }
3587
3588 if (task == current || is_system) {
3589
3590 fmt = ctx->ctx_buf_fmt;
3591
3592 DPRINT(("restarting self %d ovfl=0x%lx\n",
3593 task_pid_nr(task),
3594 ctx->ctx_ovfl_regs[0]));
3595
3596 if (CTX_HAS_SMPL(ctx)) {
3597
3598 prefetch(ctx->ctx_smpl_hdr);
3599
3600 rst_ctrl.bits.mask_monitoring = 0;
3601 rst_ctrl.bits.reset_ovfl_pmds = 0;
3602
3603 if (state == PFM_CTX_LOADED)
3604 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3605 else
3606 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3607 } else {
3608 rst_ctrl.bits.mask_monitoring = 0;
3609 rst_ctrl.bits.reset_ovfl_pmds = 1;
3610 }
3611
3612 if (ret == 0) {
3613 if (rst_ctrl.bits.reset_ovfl_pmds)
3614 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3615
3616 if (rst_ctrl.bits.mask_monitoring == 0) {
3617 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3618
3619 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3620 } else {
3621 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3622
3623 // cannot use pfm_stop_monitoring(task, regs);
3624 }
3625 }
3626 /*
3627 * clear overflowed PMD mask to remove any stale information
3628 */
3629 ctx->ctx_ovfl_regs[0] = 0UL;
3630
3631 /*
3632 * back to LOADED state
3633 */
3634 ctx->ctx_state = PFM_CTX_LOADED;
3635
3636 /*
3637 * XXX: not really useful for self monitoring
3638 */
3639 ctx->ctx_fl_can_restart = 0;
3640
3641 return 0;
3642 }
3643
3644 /*
3645 * restart another task
3646 */
3647
3648 /*
3649 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3650 * one is seen by the task.
3651 */
3652 if (state == PFM_CTX_MASKED) {
3653 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3654 /*
3655 * will prevent subsequent restart before this one is
3656 * seen by other task
3657 */
3658 ctx->ctx_fl_can_restart = 0;
3659 }
3660
3661 /*
3662 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3663 * the task is blocked or on its way to block. That's the normal
3664 * restart path. If the monitoring is not masked, then the task
3665 * can be actively monitoring and we cannot directly intervene.
3666 * Therefore we use the trap mechanism to catch the task and
3667 * force it to reset the buffer/reset PMDs.
3668 *
3669 * if non-blocking, then we ensure that the task will go into
3670 * pfm_handle_work() before returning to user mode.
3671 *
3672 * We cannot explicitly reset another task, it MUST always
3673 * be done by the task itself. This works for system wide because
3674 * the tool that is controlling the session is logically doing
3675 * "self-monitoring".
3676 */
3677 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3678 DPRINT(("unblocking [%d] \n", task_pid_nr(task)));
3679 complete(&ctx->ctx_restart_done);
3680 } else {
3681 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3682
3683 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3684
3685 PFM_SET_WORK_PENDING(task, 1);
3686
3687 tsk_set_notify_resume(task);
3688
3689 /*
3690 * XXX: send reschedule if task runs on another CPU
3691 */
3692 }
3693 return 0;
3694 }
3695
3696 static int
3697 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3698 {
3699 unsigned int m = *(unsigned int *)arg;
3700
3701 pfm_sysctl.debug = m == 0 ? 0 : 1;
3702
3703 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3704
3705 if (m == 0) {
3706 memset(pfm_stats, 0, sizeof(pfm_stats));
3707 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3708 }
3709 return 0;
3710 }
3711
3712 /*
3713 * arg can be NULL and count can be zero for this function
3714 */
3715 static int
3716 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3717 {
3718 struct thread_struct *thread = NULL;
3719 struct task_struct *task;
3720 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3721 unsigned long flags;
3722 dbreg_t dbreg;
3723 unsigned int rnum;
3724 int first_time;
3725 int ret = 0, state;
3726 int i, can_access_pmu = 0;
3727 int is_system, is_loaded;
3728
3729 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3730
3731 state = ctx->ctx_state;
3732 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3733 is_system = ctx->ctx_fl_system;
3734 task = ctx->ctx_task;
3735
3736 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3737
3738 /*
3739 * on both UP and SMP, we can only write to the PMC when the task is
3740 * the owner of the local PMU.
3741 */
3742 if (is_loaded) {
3743 thread = &task->thread;
3744 /*
3745 * In system wide and when the context is loaded, access can only happen
3746 * when the caller is running on the CPU being monitored by the session.
3747 * It does not have to be the owner (ctx_task) of the context per se.
3748 */
3749 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3750 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3751 return -EBUSY;
3752 }
3753 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3754 }
3755
3756 /*
3757 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3758 * ensuring that no real breakpoint can be installed via this call.
3759 *
3760 * IMPORTANT: regs can be NULL in this function
3761 */
3762
3763 first_time = ctx->ctx_fl_using_dbreg == 0;
3764
3765 /*
3766 * don't bother if we are loaded and task is being debugged
3767 */
3768 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3769 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3770 return -EBUSY;
3771 }
3772
3773 /*
3774 * check for debug registers in system wide mode
3775 *
3776 * If though a check is done in pfm_context_load(),
3777 * we must repeat it here, in case the registers are
3778 * written after the context is loaded
3779 */
3780 if (is_loaded) {
3781 LOCK_PFS(flags);
3782
3783 if (first_time && is_system) {
3784 if (pfm_sessions.pfs_ptrace_use_dbregs)
3785 ret = -EBUSY;
3786 else
3787 pfm_sessions.pfs_sys_use_dbregs++;
3788 }
3789 UNLOCK_PFS(flags);
3790 }
3791
3792 if (ret != 0) return ret;
3793
3794 /*
3795 * mark ourself as user of the debug registers for
3796 * perfmon purposes.
3797 */
3798 ctx->ctx_fl_using_dbreg = 1;
3799
3800 /*
3801 * clear hardware registers to make sure we don't
3802 * pick up stale state.
3803 *
3804 * for a system wide session, we do not use
3805 * thread.dbr, thread.ibr because this process
3806 * never leaves the current CPU and the state
3807 * is shared by all processes running on it
3808 */
3809 if (first_time && can_access_pmu) {
3810 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3811 for (i=0; i < pmu_conf->num_ibrs; i++) {
3812 ia64_set_ibr(i, 0UL);
3813 ia64_dv_serialize_instruction();
3814 }
3815 ia64_srlz_i();
3816 for (i=0; i < pmu_conf->num_dbrs; i++) {
3817 ia64_set_dbr(i, 0UL);
3818 ia64_dv_serialize_data();
3819 }
3820 ia64_srlz_d();
3821 }
3822
3823 /*
3824 * Now install the values into the registers
3825 */
3826 for (i = 0; i < count; i++, req++) {
3827
3828 rnum = req->dbreg_num;
3829 dbreg.val = req->dbreg_value;
3830
3831 ret = -EINVAL;
3832
3833 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3834 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3835 rnum, dbreg.val, mode, i, count));
3836
3837 goto abort_mission;
3838 }
3839
3840 /*
3841 * make sure we do not install enabled breakpoint
3842 */
3843 if (rnum & 0x1) {
3844 if (mode == PFM_CODE_RR)
3845 dbreg.ibr.ibr_x = 0;
3846 else
3847 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3848 }
3849
3850 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3851
3852 /*
3853 * Debug registers, just like PMC, can only be modified
3854 * by a kernel call. Moreover, perfmon() access to those
3855 * registers are centralized in this routine. The hardware
3856 * does not modify the value of these registers, therefore,
3857 * if we save them as they are written, we can avoid having
3858 * to save them on context switch out. This is made possible
3859 * by the fact that when perfmon uses debug registers, ptrace()
3860 * won't be able to modify them concurrently.
3861 */
3862 if (mode == PFM_CODE_RR) {
3863 CTX_USED_IBR(ctx, rnum);
3864
3865 if (can_access_pmu) {
3866 ia64_set_ibr(rnum, dbreg.val);
3867 ia64_dv_serialize_instruction();
3868 }
3869
3870 ctx->ctx_ibrs[rnum] = dbreg.val;
3871
3872 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3873 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3874 } else {
3875 CTX_USED_DBR(ctx, rnum);
3876
3877 if (can_access_pmu) {
3878 ia64_set_dbr(rnum, dbreg.val);
3879 ia64_dv_serialize_data();
3880 }
3881 ctx->ctx_dbrs[rnum] = dbreg.val;
3882
3883 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3884 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3885 }
3886 }
3887
3888 return 0;
3889
3890 abort_mission:
3891 /*
3892 * in case it was our first attempt, we undo the global modifications
3893 */
3894 if (first_time) {
3895 LOCK_PFS(flags);
3896 if (ctx->ctx_fl_system) {
3897 pfm_sessions.pfs_sys_use_dbregs--;
3898 }
3899 UNLOCK_PFS(flags);
3900 ctx->ctx_fl_using_dbreg = 0;
3901 }
3902 /*
3903 * install error return flag
3904 */
3905 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3906
3907 return ret;
3908 }
3909
3910 static int
3911 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3912 {
3913 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3914 }
3915
3916 static int
3917 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3918 {
3919 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3920 }
3921
3922 int
3923 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3924 {
3925 pfm_context_t *ctx;
3926
3927 if (req == NULL) return -EINVAL;
3928
3929 ctx = GET_PMU_CTX();
3930
3931 if (ctx == NULL) return -EINVAL;
3932
3933 /*
3934 * for now limit to current task, which is enough when calling
3935 * from overflow handler
3936 */
3937 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3938
3939 return pfm_write_ibrs(ctx, req, nreq, regs);
3940 }
3941 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3942
3943 int
3944 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3945 {
3946 pfm_context_t *ctx;
3947
3948 if (req == NULL) return -EINVAL;
3949
3950 ctx = GET_PMU_CTX();
3951
3952 if (ctx == NULL) return -EINVAL;
3953
3954 /*
3955 * for now limit to current task, which is enough when calling
3956 * from overflow handler
3957 */
3958 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3959
3960 return pfm_write_dbrs(ctx, req, nreq, regs);
3961 }
3962 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3963
3964
3965 static int
3966 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3967 {
3968 pfarg_features_t *req = (pfarg_features_t *)arg;
3969
3970 req->ft_version = PFM_VERSION;
3971 return 0;
3972 }
3973
3974 static int
3975 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3976 {
3977 struct pt_regs *tregs;
3978 struct task_struct *task = PFM_CTX_TASK(ctx);
3979 int state, is_system;
3980
3981 state = ctx->ctx_state;
3982 is_system = ctx->ctx_fl_system;
3983
3984 /*
3985 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3986 */
3987 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3988
3989 /*
3990 * In system wide and when the context is loaded, access can only happen
3991 * when the caller is running on the CPU being monitored by the session.
3992 * It does not have to be the owner (ctx_task) of the context per se.
3993 */
3994 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3995 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3996 return -EBUSY;
3997 }
3998 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3999 task_pid_nr(PFM_CTX_TASK(ctx)),
4000 state,
4001 is_system));
4002 /*
4003 * in system mode, we need to update the PMU directly
4004 * and the user level state of the caller, which may not
4005 * necessarily be the creator of the context.
4006 */
4007 if (is_system) {
4008 /*
4009 * Update local PMU first
4010 *
4011 * disable dcr pp
4012 */
4013 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4014 ia64_srlz_i();
4015
4016 /*
4017 * update local cpuinfo
4018 */
4019 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4020
4021 /*
4022 * stop monitoring, does srlz.i
4023 */
4024 pfm_clear_psr_pp();
4025
4026 /*
4027 * stop monitoring in the caller
4028 */
4029 ia64_psr(regs)->pp = 0;
4030
4031 return 0;
4032 }
4033 /*
4034 * per-task mode
4035 */
4036
4037 if (task == current) {
4038 /* stop monitoring at kernel level */
4039 pfm_clear_psr_up();
4040
4041 /*
4042 * stop monitoring at the user level
4043 */
4044 ia64_psr(regs)->up = 0;
4045 } else {
4046 tregs = task_pt_regs(task);
4047
4048 /*
4049 * stop monitoring at the user level
4050 */
4051 ia64_psr(tregs)->up = 0;
4052
4053 /*
4054 * monitoring disabled in kernel at next reschedule
4055 */
4056 ctx->ctx_saved_psr_up = 0;
4057 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4058 }
4059 return 0;
4060 }
4061
4062
4063 static int
4064 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4065 {
4066 struct pt_regs *tregs;
4067 int state, is_system;
4068
4069 state = ctx->ctx_state;
4070 is_system = ctx->ctx_fl_system;
4071
4072 if (state != PFM_CTX_LOADED) return -EINVAL;
4073
4074 /*
4075 * In system wide and when the context is loaded, access can only happen
4076 * when the caller is running on the CPU being monitored by the session.
4077 * It does not have to be the owner (ctx_task) of the context per se.
4078 */
4079 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4080 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4081 return -EBUSY;
4082 }
4083
4084 /*
4085 * in system mode, we need to update the PMU directly
4086 * and the user level state of the caller, which may not
4087 * necessarily be the creator of the context.
4088 */
4089 if (is_system) {
4090
4091 /*
4092 * set user level psr.pp for the caller
4093 */
4094 ia64_psr(regs)->pp = 1;
4095
4096 /*
4097 * now update the local PMU and cpuinfo
4098 */
4099 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4100
4101 /*
4102 * start monitoring at kernel level
4103 */
4104 pfm_set_psr_pp();
4105
4106 /* enable dcr pp */
4107 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4108 ia64_srlz_i();
4109
4110 return 0;
4111 }
4112
4113 /*
4114 * per-process mode
4115 */
4116
4117 if (ctx->ctx_task == current) {
4118
4119 /* start monitoring at kernel level */
4120 pfm_set_psr_up();
4121
4122 /*
4123 * activate monitoring at user level
4124 */
4125 ia64_psr(regs)->up = 1;
4126
4127 } else {
4128 tregs = task_pt_regs(ctx->ctx_task);
4129
4130 /*
4131 * start monitoring at the kernel level the next
4132 * time the task is scheduled
4133 */
4134 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4135
4136 /*
4137 * activate monitoring at user level
4138 */
4139 ia64_psr(tregs)->up = 1;
4140 }
4141 return 0;
4142 }
4143
4144 static int
4145 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4146 {
4147 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4148 unsigned int cnum;
4149 int i;
4150 int ret = -EINVAL;
4151
4152 for (i = 0; i < count; i++, req++) {
4153
4154 cnum = req->reg_num;
4155
4156 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4157
4158 req->reg_value = PMC_DFL_VAL(cnum);
4159
4160 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4161
4162 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4163 }
4164 return 0;
4165
4166 abort_mission:
4167 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4168 return ret;
4169 }
4170
4171 static int
4172 pfm_check_task_exist(pfm_context_t *ctx)
4173 {
4174 struct task_struct *g, *t;
4175 int ret = -ESRCH;
4176
4177 read_lock(&tasklist_lock);
4178
4179 do_each_thread (g, t) {
4180 if (t->thread.pfm_context == ctx) {
4181 ret = 0;
4182 goto out;
4183 }
4184 } while_each_thread (g, t);
4185 out:
4186 read_unlock(&tasklist_lock);
4187
4188 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4189
4190 return ret;
4191 }
4192
4193 static int
4194 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4195 {
4196 struct task_struct *task;
4197 struct thread_struct *thread;
4198 struct pfm_context_t *old;
4199 unsigned long flags;
4200 #ifndef CONFIG_SMP
4201 struct task_struct *owner_task = NULL;
4202 #endif
4203 pfarg_load_t *req = (pfarg_load_t *)arg;
4204 unsigned long *pmcs_source, *pmds_source;
4205 int the_cpu;
4206 int ret = 0;
4207 int state, is_system, set_dbregs = 0;
4208
4209 state = ctx->ctx_state;
4210 is_system = ctx->ctx_fl_system;
4211 /*
4212 * can only load from unloaded or terminated state
4213 */
4214 if (state != PFM_CTX_UNLOADED) {
4215 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4216 req->load_pid,
4217 ctx->ctx_state));
4218 return -EBUSY;
4219 }
4220
4221 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4222
4223 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4224 DPRINT(("cannot use blocking mode on self\n"));
4225 return -EINVAL;
4226 }
4227
4228 ret = pfm_get_task(ctx, req->load_pid, &task);
4229 if (ret) {
4230 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4231 return ret;
4232 }
4233
4234 ret = -EINVAL;
4235
4236 /*
4237 * system wide is self monitoring only
4238 */
4239 if (is_system && task != current) {
4240 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4241 req->load_pid));
4242 goto error;
4243 }
4244
4245 thread = &task->thread;
4246
4247 ret = 0;
4248 /*
4249 * cannot load a context which is using range restrictions,
4250 * into a task that is being debugged.
4251 */
4252 if (ctx->ctx_fl_using_dbreg) {
4253 if (thread->flags & IA64_THREAD_DBG_VALID) {
4254 ret = -EBUSY;
4255 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4256 goto error;
4257 }
4258 LOCK_PFS(flags);
4259
4260 if (is_system) {
4261 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4262 DPRINT(("cannot load [%d] dbregs in use\n",
4263 task_pid_nr(task)));
4264 ret = -EBUSY;
4265 } else {
4266 pfm_sessions.pfs_sys_use_dbregs++;
4267 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4268 set_dbregs = 1;
4269 }
4270 }
4271
4272 UNLOCK_PFS(flags);
4273
4274 if (ret) goto error;
4275 }
4276
4277 /*
4278 * SMP system-wide monitoring implies self-monitoring.
4279 *
4280 * The programming model expects the task to
4281 * be pinned on a CPU throughout the session.
4282 * Here we take note of the current CPU at the
4283 * time the context is loaded. No call from
4284 * another CPU will be allowed.
4285 *
4286 * The pinning via shed_setaffinity()
4287 * must be done by the calling task prior
4288 * to this call.
4289 *
4290 * systemwide: keep track of CPU this session is supposed to run on
4291 */
4292 the_cpu = ctx->ctx_cpu = smp_processor_id();
4293
4294 ret = -EBUSY;
4295 /*
4296 * now reserve the session
4297 */
4298 ret = pfm_reserve_session(current, is_system, the_cpu);
4299 if (ret) goto error;
4300
4301 /*
4302 * task is necessarily stopped at this point.
4303 *
4304 * If the previous context was zombie, then it got removed in
4305 * pfm_save_regs(). Therefore we should not see it here.
4306 * If we see a context, then this is an active context
4307 *
4308 * XXX: needs to be atomic
4309 */
4310 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4311 thread->pfm_context, ctx));
4312
4313 ret = -EBUSY;
4314 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4315 if (old != NULL) {
4316 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4317 goto error_unres;
4318 }
4319
4320 pfm_reset_msgq(ctx);
4321
4322 ctx->ctx_state = PFM_CTX_LOADED;
4323
4324 /*
4325 * link context to task
4326 */
4327 ctx->ctx_task = task;
4328
4329 if (is_system) {
4330 /*
4331 * we load as stopped
4332 */
4333 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4334 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4335
4336 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4337 } else {
4338 thread->flags |= IA64_THREAD_PM_VALID;
4339 }
4340
4341 /*
4342 * propagate into thread-state
4343 */
4344 pfm_copy_pmds(task, ctx);
4345 pfm_copy_pmcs(task, ctx);
4346
4347 pmcs_source = ctx->th_pmcs;
4348 pmds_source = ctx->th_pmds;
4349
4350 /*
4351 * always the case for system-wide
4352 */
4353 if (task == current) {
4354
4355 if (is_system == 0) {
4356
4357 /* allow user level control */
4358 ia64_psr(regs)->sp = 0;
4359 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4360
4361 SET_LAST_CPU(ctx, smp_processor_id());
4362 INC_ACTIVATION();
4363 SET_ACTIVATION(ctx);
4364 #ifndef CONFIG_SMP
4365 /*
4366 * push the other task out, if any
4367 */
4368 owner_task = GET_PMU_OWNER();
4369 if (owner_task) pfm_lazy_save_regs(owner_task);
4370 #endif
4371 }
4372 /*
4373 * load all PMD from ctx to PMU (as opposed to thread state)
4374 * restore all PMC from ctx to PMU
4375 */
4376 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4377 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4378
4379 ctx->ctx_reload_pmcs[0] = 0UL;
4380 ctx->ctx_reload_pmds[0] = 0UL;
4381
4382 /*
4383 * guaranteed safe by earlier check against DBG_VALID
4384 */
4385 if (ctx->ctx_fl_using_dbreg) {
4386 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4387 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4388 }
4389 /*
4390 * set new ownership
4391 */
4392 SET_PMU_OWNER(task, ctx);
4393
4394 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4395 } else {
4396 /*
4397 * when not current, task MUST be stopped, so this is safe
4398 */
4399 regs = task_pt_regs(task);
4400
4401 /* force a full reload */
4402 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4403 SET_LAST_CPU(ctx, -1);
4404
4405 /* initial saved psr (stopped) */
4406 ctx->ctx_saved_psr_up = 0UL;
4407 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4408 }
4409
4410 ret = 0;
4411
4412 error_unres:
4413 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4414 error:
4415 /*
4416 * we must undo the dbregs setting (for system-wide)
4417 */
4418 if (ret && set_dbregs) {
4419 LOCK_PFS(flags);
4420 pfm_sessions.pfs_sys_use_dbregs--;
4421 UNLOCK_PFS(flags);
4422 }
4423 /*
4424 * release task, there is now a link with the context
4425 */
4426 if (is_system == 0 && task != current) {
4427 pfm_put_task(task);
4428
4429 if (ret == 0) {
4430 ret = pfm_check_task_exist(ctx);
4431 if (ret) {
4432 ctx->ctx_state = PFM_CTX_UNLOADED;
4433 ctx->ctx_task = NULL;
4434 }
4435 }
4436 }
4437 return ret;
4438 }
4439
4440 /*
4441 * in this function, we do not need to increase the use count
4442 * for the task via get_task_struct(), because we hold the
4443 * context lock. If the task were to disappear while having
4444 * a context attached, it would go through pfm_exit_thread()
4445 * which also grabs the context lock and would therefore be blocked
4446 * until we are here.
4447 */
4448 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4449
4450 static int
4451 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4452 {
4453 struct task_struct *task = PFM_CTX_TASK(ctx);
4454 struct pt_regs *tregs;
4455 int prev_state, is_system;
4456 int ret;
4457
4458 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4459
4460 prev_state = ctx->ctx_state;
4461 is_system = ctx->ctx_fl_system;
4462
4463 /*
4464 * unload only when necessary
4465 */
4466 if (prev_state == PFM_CTX_UNLOADED) {
4467 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4468 return 0;
4469 }
4470
4471 /*
4472 * clear psr and dcr bits
4473 */
4474 ret = pfm_stop(ctx, NULL, 0, regs);
4475 if (ret) return ret;
4476
4477 ctx->ctx_state = PFM_CTX_UNLOADED;
4478
4479 /*
4480 * in system mode, we need to update the PMU directly
4481 * and the user level state of the caller, which may not
4482 * necessarily be the creator of the context.
4483 */
4484 if (is_system) {
4485
4486 /*
4487 * Update cpuinfo
4488 *
4489 * local PMU is taken care of in pfm_stop()
4490 */
4491 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4492 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4493
4494 /*
4495 * save PMDs in context
4496 * release ownership
4497 */
4498 pfm_flush_pmds(current, ctx);
4499
4500 /*
4501 * at this point we are done with the PMU
4502 * so we can unreserve the resource.
4503 */
4504 if (prev_state != PFM_CTX_ZOMBIE)
4505 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4506
4507 /*
4508 * disconnect context from task
4509 */
4510 task->thread.pfm_context = NULL;
4511 /*
4512 * disconnect task from context
4513 */
4514 ctx->ctx_task = NULL;
4515
4516 /*
4517 * There is nothing more to cleanup here.
4518 */
4519 return 0;
4520 }
4521
4522 /*
4523 * per-task mode
4524 */
4525 tregs = task == current ? regs : task_pt_regs(task);
4526
4527 if (task == current) {
4528 /*
4529 * cancel user level control
4530 */
4531 ia64_psr(regs)->sp = 1;
4532
4533 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4534 }
4535 /*
4536 * save PMDs to context
4537 * release ownership
4538 */
4539 pfm_flush_pmds(task, ctx);
4540
4541 /*
4542 * at this point we are done with the PMU
4543 * so we can unreserve the resource.
4544 *
4545 * when state was ZOMBIE, we have already unreserved.
4546 */
4547 if (prev_state != PFM_CTX_ZOMBIE)
4548 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4549
4550 /*
4551 * reset activation counter and psr
4552 */
4553 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4554 SET_LAST_CPU(ctx, -1);
4555
4556 /*
4557 * PMU state will not be restored
4558 */
4559 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4560
4561 /*
4562 * break links between context and task
4563 */
4564 task->thread.pfm_context = NULL;
4565 ctx->ctx_task = NULL;
4566
4567 PFM_SET_WORK_PENDING(task, 0);
4568
4569 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4570 ctx->ctx_fl_can_restart = 0;
4571 ctx->ctx_fl_going_zombie = 0;
4572
4573 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4574
4575 return 0;
4576 }
4577
4578
4579 /*
4580 * called only from exit_thread(): task == current
4581 * we come here only if current has a context attached (loaded or masked)
4582 */
4583 void
4584 pfm_exit_thread(struct task_struct *task)
4585 {
4586 pfm_context_t *ctx;
4587 unsigned long flags;
4588 struct pt_regs *regs = task_pt_regs(task);
4589 int ret, state;
4590 int free_ok = 0;
4591
4592 ctx = PFM_GET_CTX(task);
4593
4594 PROTECT_CTX(ctx, flags);
4595
4596 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4597
4598 state = ctx->ctx_state;
4599 switch(state) {
4600 case PFM_CTX_UNLOADED:
4601 /*
4602 * only comes to this function if pfm_context is not NULL, i.e., cannot
4603 * be in unloaded state
4604 */
4605 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4606 break;
4607 case PFM_CTX_LOADED:
4608 case PFM_CTX_MASKED:
4609 ret = pfm_context_unload(ctx, NULL, 0, regs);
4610 if (ret) {
4611 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4612 }
4613 DPRINT(("ctx unloaded for current state was %d\n", state));
4614
4615 pfm_end_notify_user(ctx);
4616 break;
4617 case PFM_CTX_ZOMBIE:
4618 ret = pfm_context_unload(ctx, NULL, 0, regs);
4619 if (ret) {
4620 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4621 }
4622 free_ok = 1;
4623 break;
4624 default:
4625 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4626 break;
4627 }
4628 UNPROTECT_CTX(ctx, flags);
4629
4630 { u64 psr = pfm_get_psr();
4631 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4632 BUG_ON(GET_PMU_OWNER());
4633 BUG_ON(ia64_psr(regs)->up);
4634 BUG_ON(ia64_psr(regs)->pp);
4635 }
4636
4637 /*
4638 * All memory free operations (especially for vmalloc'ed memory)
4639 * MUST be done with interrupts ENABLED.
4640 */
4641 if (free_ok) pfm_context_free(ctx);
4642 }
4643
4644 /*
4645 * functions MUST be listed in the increasing order of their index (see permfon.h)
4646 */
4647 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4648 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4649 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4650 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4651 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4652
4653 static pfm_cmd_desc_t pfm_cmd_tab[]={
4654 /* 0 */PFM_CMD_NONE,
4655 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4656 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4657 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4658 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4659 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4660 /* 6 */PFM_CMD_NONE,
4661 /* 7 */PFM_CMD_NONE,
4662 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4663 /* 9 */PFM_CMD_NONE,
4664 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4665 /* 11 */PFM_CMD_NONE,
4666 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4667 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4668 /* 14 */PFM_CMD_NONE,
4669 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4670 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4671 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4672 /* 18 */PFM_CMD_NONE,
4673 /* 19 */PFM_CMD_NONE,
4674 /* 20 */PFM_CMD_NONE,
4675 /* 21 */PFM_CMD_NONE,
4676 /* 22 */PFM_CMD_NONE,
4677 /* 23 */PFM_CMD_NONE,
4678 /* 24 */PFM_CMD_NONE,
4679 /* 25 */PFM_CMD_NONE,
4680 /* 26 */PFM_CMD_NONE,
4681 /* 27 */PFM_CMD_NONE,
4682 /* 28 */PFM_CMD_NONE,
4683 /* 29 */PFM_CMD_NONE,
4684 /* 30 */PFM_CMD_NONE,
4685 /* 31 */PFM_CMD_NONE,
4686 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4687 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4688 };
4689 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4690
4691 static int
4692 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4693 {
4694 struct task_struct *task;
4695 int state, old_state;
4696
4697 recheck:
4698 state = ctx->ctx_state;
4699 task = ctx->ctx_task;
4700
4701 if (task == NULL) {
4702 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4703 return 0;
4704 }
4705
4706 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4707 ctx->ctx_fd,
4708 state,
4709 task_pid_nr(task),
4710 task->state, PFM_CMD_STOPPED(cmd)));
4711
4712 /*
4713 * self-monitoring always ok.
4714 *
4715 * for system-wide the caller can either be the creator of the
4716 * context (to one to which the context is attached to) OR
4717 * a task running on the same CPU as the session.
4718 */
4719 if (task == current || ctx->ctx_fl_system) return 0;
4720
4721 /*
4722 * we are monitoring another thread
4723 */
4724 switch(state) {
4725 case PFM_CTX_UNLOADED:
4726 /*
4727 * if context is UNLOADED we are safe to go
4728 */
4729 return 0;
4730 case PFM_CTX_ZOMBIE:
4731 /*
4732 * no command can operate on a zombie context
4733 */
4734 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4735 return -EINVAL;
4736 case PFM_CTX_MASKED:
4737 /*
4738 * PMU state has been saved to software even though
4739 * the thread may still be running.
4740 */
4741 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4742 }
4743
4744 /*
4745 * context is LOADED or MASKED. Some commands may need to have
4746 * the task stopped.
4747 *
4748 * We could lift this restriction for UP but it would mean that
4749 * the user has no guarantee the task would not run between
4750 * two successive calls to perfmonctl(). That's probably OK.
4751 * If this user wants to ensure the task does not run, then
4752 * the task must be stopped.
4753 */
4754 if (PFM_CMD_STOPPED(cmd)) {
4755 if (!task_is_stopped_or_traced(task)) {
4756 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4757 return -EBUSY;
4758 }
4759 /*
4760 * task is now stopped, wait for ctxsw out
4761 *
4762 * This is an interesting point in the code.
4763 * We need to unprotect the context because
4764 * the pfm_save_regs() routines needs to grab
4765 * the same lock. There are danger in doing
4766 * this because it leaves a window open for
4767 * another task to get access to the context
4768 * and possibly change its state. The one thing
4769 * that is not possible is for the context to disappear
4770 * because we are protected by the VFS layer, i.e.,
4771 * get_fd()/put_fd().
4772 */
4773 old_state = state;
4774
4775 UNPROTECT_CTX(ctx, flags);
4776
4777 wait_task_inactive(task);
4778
4779 PROTECT_CTX(ctx, flags);
4780
4781 /*
4782 * we must recheck to verify if state has changed
4783 */
4784 if (ctx->ctx_state != old_state) {
4785 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4786 goto recheck;
4787 }
4788 }
4789 return 0;
4790 }
4791
4792 /*
4793 * system-call entry point (must return long)
4794 */
4795 asmlinkage long
4796 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4797 {
4798 struct file *file = NULL;
4799 pfm_context_t *ctx = NULL;
4800 unsigned long flags = 0UL;
4801 void *args_k = NULL;
4802 long ret; /* will expand int return types */
4803 size_t base_sz, sz, xtra_sz = 0;
4804 int narg, completed_args = 0, call_made = 0, cmd_flags;
4805 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4806 int (*getsize)(void *arg, size_t *sz);
4807 #define PFM_MAX_ARGSIZE 4096
4808
4809 /*
4810 * reject any call if perfmon was disabled at initialization
4811 */
4812 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4813
4814 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4815 DPRINT(("invalid cmd=%d\n", cmd));
4816 return -EINVAL;
4817 }
4818
4819 func = pfm_cmd_tab[cmd].cmd_func;
4820 narg = pfm_cmd_tab[cmd].cmd_narg;
4821 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4822 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4823 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4824
4825 if (unlikely(func == NULL)) {
4826 DPRINT(("invalid cmd=%d\n", cmd));
4827 return -EINVAL;
4828 }
4829
4830 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4831 PFM_CMD_NAME(cmd),
4832 cmd,
4833 narg,
4834 base_sz,
4835 count));
4836
4837 /*
4838 * check if number of arguments matches what the command expects
4839 */
4840 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4841 return -EINVAL;
4842
4843 restart_args:
4844 sz = xtra_sz + base_sz*count;
4845 /*
4846 * limit abuse to min page size
4847 */
4848 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4849 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4850 return -E2BIG;
4851 }
4852
4853 /*
4854 * allocate default-sized argument buffer
4855 */
4856 if (likely(count && args_k == NULL)) {
4857 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4858 if (args_k == NULL) return -ENOMEM;
4859 }
4860
4861 ret = -EFAULT;
4862
4863 /*
4864 * copy arguments
4865 *
4866 * assume sz = 0 for command without parameters
4867 */
4868 if (sz && copy_from_user(args_k, arg, sz)) {
4869 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4870 goto error_args;
4871 }
4872
4873 /*
4874 * check if command supports extra parameters
4875 */
4876 if (completed_args == 0 && getsize) {
4877 /*
4878 * get extra parameters size (based on main argument)
4879 */
4880 ret = (*getsize)(args_k, &xtra_sz);
4881 if (ret) goto error_args;
4882
4883 completed_args = 1;
4884
4885 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4886
4887 /* retry if necessary */
4888 if (likely(xtra_sz)) goto restart_args;
4889 }
4890
4891 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4892
4893 ret = -EBADF;
4894
4895 file = fget(fd);
4896 if (unlikely(file == NULL)) {
4897 DPRINT(("invalid fd %d\n", fd));
4898 goto error_args;
4899 }
4900 if (unlikely(PFM_IS_FILE(file) == 0)) {
4901 DPRINT(("fd %d not related to perfmon\n", fd));
4902 goto error_args;
4903 }
4904
4905 ctx = (pfm_context_t *)file->private_data;
4906 if (unlikely(ctx == NULL)) {
4907 DPRINT(("no context for fd %d\n", fd));
4908 goto error_args;
4909 }
4910 prefetch(&ctx->ctx_state);
4911
4912 PROTECT_CTX(ctx, flags);
4913
4914 /*
4915 * check task is stopped
4916 */
4917 ret = pfm_check_task_state(ctx, cmd, flags);
4918 if (unlikely(ret)) goto abort_locked;
4919
4920 skip_fd:
4921 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4922
4923 call_made = 1;
4924
4925 abort_locked:
4926 if (likely(ctx)) {
4927 DPRINT(("context unlocked\n"));
4928 UNPROTECT_CTX(ctx, flags);
4929 }
4930
4931 /* copy argument back to user, if needed */
4932 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4933
4934 error_args:
4935 if (file)
4936 fput(file);
4937
4938 kfree(args_k);
4939
4940 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4941
4942 return ret;
4943 }
4944
4945 static void
4946 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4947 {
4948 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4949 pfm_ovfl_ctrl_t rst_ctrl;
4950 int state;
4951 int ret = 0;
4952
4953 state = ctx->ctx_state;
4954 /*
4955 * Unlock sampling buffer and reset index atomically
4956 * XXX: not really needed when blocking
4957 */
4958 if (CTX_HAS_SMPL(ctx)) {
4959
4960 rst_ctrl.bits.mask_monitoring = 0;
4961 rst_ctrl.bits.reset_ovfl_pmds = 0;
4962
4963 if (state == PFM_CTX_LOADED)
4964 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4965 else
4966 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4967 } else {
4968 rst_ctrl.bits.mask_monitoring = 0;
4969 rst_ctrl.bits.reset_ovfl_pmds = 1;
4970 }
4971
4972 if (ret == 0) {
4973 if (rst_ctrl.bits.reset_ovfl_pmds) {
4974 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4975 }
4976 if (rst_ctrl.bits.mask_monitoring == 0) {
4977 DPRINT(("resuming monitoring\n"));
4978 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4979 } else {
4980 DPRINT(("stopping monitoring\n"));
4981 //pfm_stop_monitoring(current, regs);
4982 }
4983 ctx->ctx_state = PFM_CTX_LOADED;
4984 }
4985 }
4986
4987 /*
4988 * context MUST BE LOCKED when calling
4989 * can only be called for current
4990 */
4991 static void
4992 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4993 {
4994 int ret;
4995
4996 DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4997
4998 ret = pfm_context_unload(ctx, NULL, 0, regs);
4999 if (ret) {
5000 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
5001 }
5002
5003 /*
5004 * and wakeup controlling task, indicating we are now disconnected
5005 */
5006 wake_up_interruptible(&ctx->ctx_zombieq);
5007
5008 /*
5009 * given that context is still locked, the controlling
5010 * task will only get access when we return from
5011 * pfm_handle_work().
5012 */
5013 }
5014
5015 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5016
5017 /*
5018 * pfm_handle_work() can be called with interrupts enabled
5019 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5020 * call may sleep, therefore we must re-enable interrupts
5021 * to avoid deadlocks. It is safe to do so because this function
5022 * is called ONLY when returning to user level (pUStk=1), in which case
5023 * there is no risk of kernel stack overflow due to deep
5024 * interrupt nesting.
5025 */
5026 void
5027 pfm_handle_work(void)
5028 {
5029 pfm_context_t *ctx;
5030 struct pt_regs *regs;
5031 unsigned long flags, dummy_flags;
5032 unsigned long ovfl_regs;
5033 unsigned int reason;
5034 int ret;
5035
5036 ctx = PFM_GET_CTX(current);
5037 if (ctx == NULL) {
5038 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5039 task_pid_nr(current));
5040 return;
5041 }
5042
5043 PROTECT_CTX(ctx, flags);
5044
5045 PFM_SET_WORK_PENDING(current, 0);
5046
5047 tsk_clear_notify_resume(current);
5048
5049 regs = task_pt_regs(current);
5050
5051 /*
5052 * extract reason for being here and clear
5053 */
5054 reason = ctx->ctx_fl_trap_reason;
5055 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5056 ovfl_regs = ctx->ctx_ovfl_regs[0];
5057
5058 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5059
5060 /*
5061 * must be done before we check for simple-reset mode
5062 */
5063 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5064 goto do_zombie;
5065
5066 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5067 if (reason == PFM_TRAP_REASON_RESET)
5068 goto skip_blocking;
5069
5070 /*
5071 * restore interrupt mask to what it was on entry.
5072 * Could be enabled/diasbled.
5073 */
5074 UNPROTECT_CTX(ctx, flags);
5075
5076 /*
5077 * force interrupt enable because of down_interruptible()
5078 */
5079 local_irq_enable();
5080
5081 DPRINT(("before block sleeping\n"));
5082
5083 /*
5084 * may go through without blocking on SMP systems
5085 * if restart has been received already by the time we call down()
5086 */
5087 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5088
5089 DPRINT(("after block sleeping ret=%d\n", ret));
5090
5091 /*
5092 * lock context and mask interrupts again
5093 * We save flags into a dummy because we may have
5094 * altered interrupts mask compared to entry in this
5095 * function.
5096 */
5097 PROTECT_CTX(ctx, dummy_flags);
5098
5099 /*
5100 * we need to read the ovfl_regs only after wake-up
5101 * because we may have had pfm_write_pmds() in between
5102 * and that can changed PMD values and therefore
5103 * ovfl_regs is reset for these new PMD values.
5104 */
5105 ovfl_regs = ctx->ctx_ovfl_regs[0];
5106
5107 if (ctx->ctx_fl_going_zombie) {
5108 do_zombie:
5109 DPRINT(("context is zombie, bailing out\n"));
5110 pfm_context_force_terminate(ctx, regs);
5111 goto nothing_to_do;
5112 }
5113 /*
5114 * in case of interruption of down() we don't restart anything
5115 */
5116 if (ret < 0)
5117 goto nothing_to_do;
5118
5119 skip_blocking:
5120 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5121 ctx->ctx_ovfl_regs[0] = 0UL;
5122
5123 nothing_to_do:
5124 /*
5125 * restore flags as they were upon entry
5126 */
5127 UNPROTECT_CTX(ctx, flags);
5128 }
5129
5130 static int
5131 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5132 {
5133 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5134 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5135 return 0;
5136 }
5137
5138 DPRINT(("waking up somebody\n"));
5139
5140 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5141
5142 /*
5143 * safe, we are not in intr handler, nor in ctxsw when
5144 * we come here
5145 */
5146 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5147
5148 return 0;
5149 }
5150
5151 static int
5152 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5153 {
5154 pfm_msg_t *msg = NULL;
5155
5156 if (ctx->ctx_fl_no_msg == 0) {
5157 msg = pfm_get_new_msg(ctx);
5158 if (msg == NULL) {
5159 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5160 return -1;
5161 }
5162
5163 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5164 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5165 msg->pfm_ovfl_msg.msg_active_set = 0;
5166 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5167 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5168 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5169 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5170 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5171 }
5172
5173 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5174 msg,
5175 ctx->ctx_fl_no_msg,
5176 ctx->ctx_fd,
5177 ovfl_pmds));
5178
5179 return pfm_notify_user(ctx, msg);
5180 }
5181
5182 static int
5183 pfm_end_notify_user(pfm_context_t *ctx)
5184 {
5185 pfm_msg_t *msg;
5186
5187 msg = pfm_get_new_msg(ctx);
5188 if (msg == NULL) {
5189 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5190 return -1;
5191 }
5192 /* no leak */
5193 memset(msg, 0, sizeof(*msg));
5194
5195 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5196 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5197 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5198
5199 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5200 msg,
5201 ctx->ctx_fl_no_msg,
5202 ctx->ctx_fd));
5203
5204 return pfm_notify_user(ctx, msg);
5205 }
5206
5207 /*
5208 * main overflow processing routine.
5209 * it can be called from the interrupt path or explicitly during the context switch code
5210 */
5211 static void
5212 pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5213 {
5214 pfm_ovfl_arg_t *ovfl_arg;
5215 unsigned long mask;
5216 unsigned long old_val, ovfl_val, new_val;
5217 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5218 unsigned long tstamp;
5219 pfm_ovfl_ctrl_t ovfl_ctrl;
5220 unsigned int i, has_smpl;
5221 int must_notify = 0;
5222
5223 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5224
5225 /*
5226 * sanity test. Should never happen
5227 */
5228 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5229
5230 tstamp = ia64_get_itc();
5231 mask = pmc0 >> PMU_FIRST_COUNTER;
5232 ovfl_val = pmu_conf->ovfl_val;
5233 has_smpl = CTX_HAS_SMPL(ctx);
5234
5235 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5236 "used_pmds=0x%lx\n",
5237 pmc0,
5238 task ? task_pid_nr(task): -1,
5239 (regs ? regs->cr_iip : 0),
5240 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5241 ctx->ctx_used_pmds[0]));
5242
5243
5244 /*
5245 * first we update the virtual counters
5246 * assume there was a prior ia64_srlz_d() issued
5247 */
5248 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5249
5250 /* skip pmd which did not overflow */
5251 if ((mask & 0x1) == 0) continue;
5252
5253 /*
5254 * Note that the pmd is not necessarily 0 at this point as qualified events
5255 * may have happened before the PMU was frozen. The residual count is not
5256 * taken into consideration here but will be with any read of the pmd via
5257 * pfm_read_pmds().
5258 */
5259 old_val = new_val = ctx->ctx_pmds[i].val;
5260 new_val += 1 + ovfl_val;
5261 ctx->ctx_pmds[i].val = new_val;
5262
5263 /*
5264 * check for overflow condition
5265 */
5266 if (likely(old_val > new_val)) {
5267 ovfl_pmds |= 1UL << i;
5268 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5269 }
5270
5271 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5272 i,
5273 new_val,
5274 old_val,
5275 ia64_get_pmd(i) & ovfl_val,
5276 ovfl_pmds,
5277 ovfl_notify));
5278 }
5279
5280 /*
5281 * there was no 64-bit overflow, nothing else to do
5282 */
5283 if (ovfl_pmds == 0UL) return;
5284
5285 /*
5286 * reset all control bits
5287 */
5288 ovfl_ctrl.val = 0;
5289 reset_pmds = 0UL;
5290
5291 /*
5292 * if a sampling format module exists, then we "cache" the overflow by
5293 * calling the module's handler() routine.
5294 */
5295 if (has_smpl) {
5296 unsigned long start_cycles, end_cycles;
5297 unsigned long pmd_mask;
5298 int j, k, ret = 0;
5299 int this_cpu = smp_processor_id();
5300
5301 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5302 ovfl_arg = &ctx->ctx_ovfl_arg;
5303
5304 prefetch(ctx->ctx_smpl_hdr);
5305
5306 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5307
5308 mask = 1UL << i;
5309
5310 if ((pmd_mask & 0x1) == 0) continue;
5311
5312 ovfl_arg->ovfl_pmd = (unsigned char )i;
5313 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5314 ovfl_arg->active_set = 0;
5315 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5316 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5317
5318 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5319 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5320 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5321
5322 /*
5323 * copy values of pmds of interest. Sampling format may copy them
5324 * into sampling buffer.
5325 */
5326 if (smpl_pmds) {
5327 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5328 if ((smpl_pmds & 0x1) == 0) continue;
5329 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5330 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5331 }
5332 }
5333
5334 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5335
5336 start_cycles = ia64_get_itc();
5337
5338 /*
5339 * call custom buffer format record (handler) routine
5340 */
5341 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5342
5343 end_cycles = ia64_get_itc();
5344
5345 /*
5346 * For those controls, we take the union because they have
5347 * an all or nothing behavior.
5348 */
5349 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5350 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5351 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5352 /*
5353 * build the bitmask of pmds to reset now
5354 */
5355 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5356
5357 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5358 }
5359 /*
5360 * when the module cannot handle the rest of the overflows, we abort right here
5361 */
5362 if (ret && pmd_mask) {
5363 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5364 pmd_mask<<PMU_FIRST_COUNTER));
5365 }
5366 /*
5367 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5368 */
5369 ovfl_pmds &= ~reset_pmds;
5370 } else {
5371 /*
5372 * when no sampling module is used, then the default
5373 * is to notify on overflow if requested by user
5374 */
5375 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5376 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5377 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5378 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5379 /*
5380 * if needed, we reset all overflowed pmds
5381 */
5382 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5383 }
5384
5385 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5386
5387 /*
5388 * reset the requested PMD registers using the short reset values
5389 */
5390 if (reset_pmds) {
5391 unsigned long bm = reset_pmds;
5392 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5393 }
5394
5395 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5396 /*
5397 * keep track of what to reset when unblocking
5398 */
5399 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5400
5401 /*
5402 * check for blocking context
5403 */
5404 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5405
5406 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5407
5408 /*
5409 * set the perfmon specific checking pending work for the task
5410 */
5411 PFM_SET_WORK_PENDING(task, 1);
5412
5413 /*
5414 * when coming from ctxsw, current still points to the
5415 * previous task, therefore we must work with task and not current.
5416 */
5417 tsk_set_notify_resume(task);
5418 }
5419 /*
5420 * defer until state is changed (shorten spin window). the context is locked
5421 * anyway, so the signal receiver would come spin for nothing.
5422 */
5423 must_notify = 1;
5424 }
5425
5426 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5427 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5428 PFM_GET_WORK_PENDING(task),
5429 ctx->ctx_fl_trap_reason,
5430 ovfl_pmds,
5431 ovfl_notify,
5432 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5433 /*
5434 * in case monitoring must be stopped, we toggle the psr bits
5435 */
5436 if (ovfl_ctrl.bits.mask_monitoring) {
5437 pfm_mask_monitoring(task);
5438 ctx->ctx_state = PFM_CTX_MASKED;
5439 ctx->ctx_fl_can_restart = 1;
5440 }
5441
5442 /*
5443 * send notification now
5444 */
5445 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5446
5447 return;
5448
5449 sanity_check:
5450 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5451 smp_processor_id(),
5452 task ? task_pid_nr(task) : -1,
5453 pmc0);
5454 return;
5455
5456 stop_monitoring:
5457 /*
5458 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5459 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5460 * come here as zombie only if the task is the current task. In which case, we
5461 * can access the PMU hardware directly.
5462 *
5463 * Note that zombies do have PM_VALID set. So here we do the minimal.
5464 *
5465 * In case the context was zombified it could not be reclaimed at the time
5466 * the monitoring program exited. At this point, the PMU reservation has been
5467 * returned, the sampiing buffer has been freed. We must convert this call
5468 * into a spurious interrupt. However, we must also avoid infinite overflows
5469 * by stopping monitoring for this task. We can only come here for a per-task
5470 * context. All we need to do is to stop monitoring using the psr bits which
5471 * are always task private. By re-enabling secure montioring, we ensure that
5472 * the monitored task will not be able to re-activate monitoring.
5473 * The task will eventually be context switched out, at which point the context
5474 * will be reclaimed (that includes releasing ownership of the PMU).
5475 *
5476 * So there might be a window of time where the number of per-task session is zero
5477 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5478 * context. This is safe because if a per-task session comes in, it will push this one
5479 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5480 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5481 * also push our zombie context out.
5482 *
5483 * Overall pretty hairy stuff....
5484 */
5485 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5486 pfm_clear_psr_up();
5487 ia64_psr(regs)->up = 0;
5488 ia64_psr(regs)->sp = 1;
5489 return;
5490 }
5491
5492 static int
5493 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5494 {
5495 struct task_struct *task;
5496 pfm_context_t *ctx;
5497 unsigned long flags;
5498 u64 pmc0;
5499 int this_cpu = smp_processor_id();
5500 int retval = 0;
5501
5502 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5503
5504 /*
5505 * srlz.d done before arriving here
5506 */
5507 pmc0 = ia64_get_pmc(0);
5508
5509 task = GET_PMU_OWNER();
5510 ctx = GET_PMU_CTX();
5511
5512 /*
5513 * if we have some pending bits set
5514 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5515 */
5516 if (PMC0_HAS_OVFL(pmc0) && task) {
5517 /*
5518 * we assume that pmc0.fr is always set here
5519 */
5520
5521 /* sanity check */
5522 if (!ctx) goto report_spurious1;
5523
5524 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5525 goto report_spurious2;
5526
5527 PROTECT_CTX_NOPRINT(ctx, flags);
5528
5529 pfm_overflow_handler(task, ctx, pmc0, regs);
5530
5531 UNPROTECT_CTX_NOPRINT(ctx, flags);
5532
5533 } else {
5534 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5535 retval = -1;
5536 }
5537 /*
5538 * keep it unfrozen at all times
5539 */
5540 pfm_unfreeze_pmu();
5541
5542 return retval;
5543
5544 report_spurious1:
5545 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5546 this_cpu, task_pid_nr(task));
5547 pfm_unfreeze_pmu();
5548 return -1;
5549 report_spurious2:
5550 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5551 this_cpu,
5552 task_pid_nr(task));
5553 pfm_unfreeze_pmu();
5554 return -1;
5555 }
5556
5557 static irqreturn_t
5558 pfm_interrupt_handler(int irq, void *arg)
5559 {
5560 unsigned long start_cycles, total_cycles;
5561 unsigned long min, max;
5562 int this_cpu;
5563 int ret;
5564 struct pt_regs *regs = get_irq_regs();
5565
5566 this_cpu = get_cpu();
5567 if (likely(!pfm_alt_intr_handler)) {
5568 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5569 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5570
5571 start_cycles = ia64_get_itc();
5572
5573 ret = pfm_do_interrupt_handler(arg, regs);
5574
5575 total_cycles = ia64_get_itc();
5576
5577 /*
5578 * don't measure spurious interrupts
5579 */
5580 if (likely(ret == 0)) {
5581 total_cycles -= start_cycles;
5582
5583 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5584 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5585
5586 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5587 }
5588 }
5589 else {
5590 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5591 }
5592
5593 put_cpu_no_resched();
5594 return IRQ_HANDLED;
5595 }
5596
5597 /*
5598 * /proc/perfmon interface, for debug only
5599 */
5600
5601 #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5602
5603 static void *
5604 pfm_proc_start(struct seq_file *m, loff_t *pos)
5605 {
5606 if (*pos == 0) {
5607 return PFM_PROC_SHOW_HEADER;
5608 }
5609
5610 while (*pos <= NR_CPUS) {
5611 if (cpu_online(*pos - 1)) {
5612 return (void *)*pos;
5613 }
5614 ++*pos;
5615 }
5616 return NULL;
5617 }
5618
5619 static void *
5620 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5621 {
5622 ++*pos;
5623 return pfm_proc_start(m, pos);
5624 }
5625
5626 static void
5627 pfm_proc_stop(struct seq_file *m, void *v)
5628 {
5629 }
5630
5631 static void
5632 pfm_proc_show_header(struct seq_file *m)
5633 {
5634 struct list_head * pos;
5635 pfm_buffer_fmt_t * entry;
5636 unsigned long flags;
5637
5638 seq_printf(m,
5639 "perfmon version : %u.%u\n"
5640 "model : %s\n"
5641 "fastctxsw : %s\n"
5642 "expert mode : %s\n"
5643 "ovfl_mask : 0x%lx\n"
5644 "PMU flags : 0x%x\n",
5645 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5646 pmu_conf->pmu_name,
5647 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5648 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5649 pmu_conf->ovfl_val,
5650 pmu_conf->flags);
5651
5652 LOCK_PFS(flags);
5653
5654 seq_printf(m,
5655 "proc_sessions : %u\n"
5656 "sys_sessions : %u\n"
5657 "sys_use_dbregs : %u\n"
5658 "ptrace_use_dbregs : %u\n",
5659 pfm_sessions.pfs_task_sessions,
5660 pfm_sessions.pfs_sys_sessions,
5661 pfm_sessions.pfs_sys_use_dbregs,
5662 pfm_sessions.pfs_ptrace_use_dbregs);
5663
5664 UNLOCK_PFS(flags);
5665
5666 spin_lock(&pfm_buffer_fmt_lock);
5667
5668 list_for_each(pos, &pfm_buffer_fmt_list) {
5669 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5670 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5671 entry->fmt_uuid[0],
5672 entry->fmt_uuid[1],
5673 entry->fmt_uuid[2],
5674 entry->fmt_uuid[3],
5675 entry->fmt_uuid[4],
5676 entry->fmt_uuid[5],
5677 entry->fmt_uuid[6],
5678 entry->fmt_uuid[7],
5679 entry->fmt_uuid[8],
5680 entry->fmt_uuid[9],
5681 entry->fmt_uuid[10],
5682 entry->fmt_uuid[11],
5683 entry->fmt_uuid[12],
5684 entry->fmt_uuid[13],
5685 entry->fmt_uuid[14],
5686 entry->fmt_uuid[15],
5687 entry->fmt_name);
5688 }
5689 spin_unlock(&pfm_buffer_fmt_lock);
5690
5691 }
5692
5693 static int
5694 pfm_proc_show(struct seq_file *m, void *v)
5695 {
5696 unsigned long psr;
5697 unsigned int i;
5698 int cpu;
5699
5700 if (v == PFM_PROC_SHOW_HEADER) {
5701 pfm_proc_show_header(m);
5702 return 0;
5703 }
5704
5705 /* show info for CPU (v - 1) */
5706
5707 cpu = (long)v - 1;
5708 seq_printf(m,
5709 "CPU%-2d overflow intrs : %lu\n"
5710 "CPU%-2d overflow cycles : %lu\n"
5711 "CPU%-2d overflow min : %lu\n"
5712 "CPU%-2d overflow max : %lu\n"
5713 "CPU%-2d smpl handler calls : %lu\n"
5714 "CPU%-2d smpl handler cycles : %lu\n"
5715 "CPU%-2d spurious intrs : %lu\n"
5716 "CPU%-2d replay intrs : %lu\n"
5717 "CPU%-2d syst_wide : %d\n"
5718 "CPU%-2d dcr_pp : %d\n"
5719 "CPU%-2d exclude idle : %d\n"
5720 "CPU%-2d owner : %d\n"
5721 "CPU%-2d context : %p\n"
5722 "CPU%-2d activations : %lu\n",
5723 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5724 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5725 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5726 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5727 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5728 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5729 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5730 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5731 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5732 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5733 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5734 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5735 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5736 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5737
5738 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5739
5740 psr = pfm_get_psr();
5741
5742 ia64_srlz_d();
5743
5744 seq_printf(m,
5745 "CPU%-2d psr : 0x%lx\n"
5746 "CPU%-2d pmc0 : 0x%lx\n",
5747 cpu, psr,
5748 cpu, ia64_get_pmc(0));
5749
5750 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5751 if (PMC_IS_COUNTING(i) == 0) continue;
5752 seq_printf(m,
5753 "CPU%-2d pmc%u : 0x%lx\n"
5754 "CPU%-2d pmd%u : 0x%lx\n",
5755 cpu, i, ia64_get_pmc(i),
5756 cpu, i, ia64_get_pmd(i));
5757 }
5758 }
5759 return 0;
5760 }
5761
5762 const struct seq_operations pfm_seq_ops = {
5763 .start = pfm_proc_start,
5764 .next = pfm_proc_next,
5765 .stop = pfm_proc_stop,
5766 .show = pfm_proc_show
5767 };
5768
5769 static int
5770 pfm_proc_open(struct inode *inode, struct file *file)
5771 {
5772 return seq_open(file, &pfm_seq_ops);
5773 }
5774
5775
5776 /*
5777 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5778 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5779 * is active or inactive based on mode. We must rely on the value in
5780 * local_cpu_data->pfm_syst_info
5781 */
5782 void
5783 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5784 {
5785 struct pt_regs *regs;
5786 unsigned long dcr;
5787 unsigned long dcr_pp;
5788
5789 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5790
5791 /*
5792 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5793 * on every CPU, so we can rely on the pid to identify the idle task.
5794 */
5795 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5796 regs = task_pt_regs(task);
5797 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5798 return;
5799 }
5800 /*
5801 * if monitoring has started
5802 */
5803 if (dcr_pp) {
5804 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5805 /*
5806 * context switching in?
5807 */
5808 if (is_ctxswin) {
5809 /* mask monitoring for the idle task */
5810 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5811 pfm_clear_psr_pp();
5812 ia64_srlz_i();
5813 return;
5814 }
5815 /*
5816 * context switching out
5817 * restore monitoring for next task
5818 *
5819 * Due to inlining this odd if-then-else construction generates
5820 * better code.
5821 */
5822 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5823 pfm_set_psr_pp();
5824 ia64_srlz_i();
5825 }
5826 }
5827
5828 #ifdef CONFIG_SMP
5829
5830 static void
5831 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5832 {
5833 struct task_struct *task = ctx->ctx_task;
5834
5835 ia64_psr(regs)->up = 0;
5836 ia64_psr(regs)->sp = 1;
5837
5838 if (GET_PMU_OWNER() == task) {
5839 DPRINT(("cleared ownership for [%d]\n",
5840 task_pid_nr(ctx->ctx_task)));
5841 SET_PMU_OWNER(NULL, NULL);
5842 }
5843
5844 /*
5845 * disconnect the task from the context and vice-versa
5846 */
5847 PFM_SET_WORK_PENDING(task, 0);
5848
5849 task->thread.pfm_context = NULL;
5850 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5851
5852 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task)));
5853 }
5854
5855
5856 /*
5857 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5858 */
5859 void
5860 pfm_save_regs(struct task_struct *task)
5861 {
5862 pfm_context_t *ctx;
5863 unsigned long flags;
5864 u64 psr;
5865
5866
5867 ctx = PFM_GET_CTX(task);
5868 if (ctx == NULL) return;
5869
5870 /*
5871 * we always come here with interrupts ALREADY disabled by
5872 * the scheduler. So we simply need to protect against concurrent
5873 * access, not CPU concurrency.
5874 */
5875 flags = pfm_protect_ctx_ctxsw(ctx);
5876
5877 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5878 struct pt_regs *regs = task_pt_regs(task);
5879
5880 pfm_clear_psr_up();
5881
5882 pfm_force_cleanup(ctx, regs);
5883
5884 BUG_ON(ctx->ctx_smpl_hdr);
5885
5886 pfm_unprotect_ctx_ctxsw(ctx, flags);
5887
5888 pfm_context_free(ctx);
5889 return;
5890 }
5891
5892 /*
5893 * save current PSR: needed because we modify it
5894 */
5895 ia64_srlz_d();
5896 psr = pfm_get_psr();
5897
5898 BUG_ON(psr & (IA64_PSR_I));
5899
5900 /*
5901 * stop monitoring:
5902 * This is the last instruction which may generate an overflow
5903 *
5904 * We do not need to set psr.sp because, it is irrelevant in kernel.
5905 * It will be restored from ipsr when going back to user level
5906 */
5907 pfm_clear_psr_up();
5908
5909 /*
5910 * keep a copy of psr.up (for reload)
5911 */
5912 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5913
5914 /*
5915 * release ownership of this PMU.
5916 * PM interrupts are masked, so nothing
5917 * can happen.
5918 */
5919 SET_PMU_OWNER(NULL, NULL);
5920
5921 /*
5922 * we systematically save the PMD as we have no
5923 * guarantee we will be schedule at that same
5924 * CPU again.
5925 */
5926 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5927
5928 /*
5929 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5930 * we will need it on the restore path to check
5931 * for pending overflow.
5932 */
5933 ctx->th_pmcs[0] = ia64_get_pmc(0);
5934
5935 /*
5936 * unfreeze PMU if had pending overflows
5937 */
5938 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5939
5940 /*
5941 * finally, allow context access.
5942 * interrupts will still be masked after this call.
5943 */
5944 pfm_unprotect_ctx_ctxsw(ctx, flags);
5945 }
5946
5947 #else /* !CONFIG_SMP */
5948 void
5949 pfm_save_regs(struct task_struct *task)
5950 {
5951 pfm_context_t *ctx;
5952 u64 psr;
5953
5954 ctx = PFM_GET_CTX(task);
5955 if (ctx == NULL) return;
5956
5957 /*
5958 * save current PSR: needed because we modify it
5959 */
5960 psr = pfm_get_psr();
5961
5962 BUG_ON(psr & (IA64_PSR_I));
5963
5964 /*
5965 * stop monitoring:
5966 * This is the last instruction which may generate an overflow
5967 *
5968 * We do not need to set psr.sp because, it is irrelevant in kernel.
5969 * It will be restored from ipsr when going back to user level
5970 */
5971 pfm_clear_psr_up();
5972
5973 /*
5974 * keep a copy of psr.up (for reload)
5975 */
5976 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5977 }
5978
5979 static void
5980 pfm_lazy_save_regs (struct task_struct *task)
5981 {
5982 pfm_context_t *ctx;
5983 unsigned long flags;
5984
5985 { u64 psr = pfm_get_psr();
5986 BUG_ON(psr & IA64_PSR_UP);
5987 }
5988
5989 ctx = PFM_GET_CTX(task);
5990
5991 /*
5992 * we need to mask PMU overflow here to
5993 * make sure that we maintain pmc0 until
5994 * we save it. overflow interrupts are
5995 * treated as spurious if there is no
5996 * owner.
5997 *
5998 * XXX: I don't think this is necessary
5999 */
6000 PROTECT_CTX(ctx,flags);
6001
6002 /*
6003 * release ownership of this PMU.
6004 * must be done before we save the registers.
6005 *
6006 * after this call any PMU interrupt is treated
6007 * as spurious.
6008 */
6009 SET_PMU_OWNER(NULL, NULL);
6010
6011 /*
6012 * save all the pmds we use
6013 */
6014 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6015
6016 /*
6017 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6018 * it is needed to check for pended overflow
6019 * on the restore path
6020 */
6021 ctx->th_pmcs[0] = ia64_get_pmc(0);
6022
6023 /*
6024 * unfreeze PMU if had pending overflows
6025 */
6026 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6027
6028 /*
6029 * now get can unmask PMU interrupts, they will
6030 * be treated as purely spurious and we will not
6031 * lose any information
6032 */
6033 UNPROTECT_CTX(ctx,flags);
6034 }
6035 #endif /* CONFIG_SMP */
6036
6037 #ifdef CONFIG_SMP
6038 /*
6039 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6040 */
6041 void
6042 pfm_load_regs (struct task_struct *task)
6043 {
6044 pfm_context_t *ctx;
6045 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6046 unsigned long flags;
6047 u64 psr, psr_up;
6048 int need_irq_resend;
6049
6050 ctx = PFM_GET_CTX(task);
6051 if (unlikely(ctx == NULL)) return;
6052
6053 BUG_ON(GET_PMU_OWNER());
6054
6055 /*
6056 * possible on unload
6057 */
6058 if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6059
6060 /*
6061 * we always come here with interrupts ALREADY disabled by
6062 * the scheduler. So we simply need to protect against concurrent
6063 * access, not CPU concurrency.
6064 */
6065 flags = pfm_protect_ctx_ctxsw(ctx);
6066 psr = pfm_get_psr();
6067
6068 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6069
6070 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6071 BUG_ON(psr & IA64_PSR_I);
6072
6073 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6074 struct pt_regs *regs = task_pt_regs(task);
6075
6076 BUG_ON(ctx->ctx_smpl_hdr);
6077
6078 pfm_force_cleanup(ctx, regs);
6079
6080 pfm_unprotect_ctx_ctxsw(ctx, flags);
6081
6082 /*
6083 * this one (kmalloc'ed) is fine with interrupts disabled
6084 */
6085 pfm_context_free(ctx);
6086
6087 return;
6088 }
6089
6090 /*
6091 * we restore ALL the debug registers to avoid picking up
6092 * stale state.
6093 */
6094 if (ctx->ctx_fl_using_dbreg) {
6095 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6096 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6097 }
6098 /*
6099 * retrieve saved psr.up
6100 */
6101 psr_up = ctx->ctx_saved_psr_up;
6102
6103 /*
6104 * if we were the last user of the PMU on that CPU,
6105 * then nothing to do except restore psr
6106 */
6107 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6108
6109 /*
6110 * retrieve partial reload masks (due to user modifications)
6111 */
6112 pmc_mask = ctx->ctx_reload_pmcs[0];
6113 pmd_mask = ctx->ctx_reload_pmds[0];
6114
6115 } else {
6116 /*
6117 * To avoid leaking information to the user level when psr.sp=0,
6118 * we must reload ALL implemented pmds (even the ones we don't use).
6119 * In the kernel we only allow PFM_READ_PMDS on registers which
6120 * we initialized or requested (sampling) so there is no risk there.
6121 */
6122 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6123
6124 /*
6125 * ALL accessible PMCs are systematically reloaded, unused registers
6126 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6127 * up stale configuration.
6128 *
6129 * PMC0 is never in the mask. It is always restored separately.
6130 */
6131 pmc_mask = ctx->ctx_all_pmcs[0];
6132 }
6133 /*
6134 * when context is MASKED, we will restore PMC with plm=0
6135 * and PMD with stale information, but that's ok, nothing
6136 * will be captured.
6137 *
6138 * XXX: optimize here
6139 */
6140 if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6141 if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6142
6143 /*
6144 * check for pending overflow at the time the state
6145 * was saved.
6146 */
6147 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6148 /*
6149 * reload pmc0 with the overflow information
6150 * On McKinley PMU, this will trigger a PMU interrupt
6151 */
6152 ia64_set_pmc(0, ctx->th_pmcs[0]);
6153 ia64_srlz_d();
6154 ctx->th_pmcs[0] = 0UL;
6155
6156 /*
6157 * will replay the PMU interrupt
6158 */
6159 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6160
6161 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6162 }
6163
6164 /*
6165 * we just did a reload, so we reset the partial reload fields
6166 */
6167 ctx->ctx_reload_pmcs[0] = 0UL;
6168 ctx->ctx_reload_pmds[0] = 0UL;
6169
6170 SET_LAST_CPU(ctx, smp_processor_id());
6171
6172 /*
6173 * dump activation value for this PMU
6174 */
6175 INC_ACTIVATION();
6176 /*
6177 * record current activation for this context
6178 */
6179 SET_ACTIVATION(ctx);
6180
6181 /*
6182 * establish new ownership.
6183 */
6184 SET_PMU_OWNER(task, ctx);
6185
6186 /*
6187 * restore the psr.up bit. measurement
6188 * is active again.
6189 * no PMU interrupt can happen at this point
6190 * because we still have interrupts disabled.
6191 */
6192 if (likely(psr_up)) pfm_set_psr_up();
6193
6194 /*
6195 * allow concurrent access to context
6196 */
6197 pfm_unprotect_ctx_ctxsw(ctx, flags);
6198 }
6199 #else /* !CONFIG_SMP */
6200 /*
6201 * reload PMU state for UP kernels
6202 * in 2.5 we come here with interrupts disabled
6203 */
6204 void
6205 pfm_load_regs (struct task_struct *task)
6206 {
6207 pfm_context_t *ctx;
6208 struct task_struct *owner;
6209 unsigned long pmd_mask, pmc_mask;
6210 u64 psr, psr_up;
6211 int need_irq_resend;
6212
6213 owner = GET_PMU_OWNER();
6214 ctx = PFM_GET_CTX(task);
6215 psr = pfm_get_psr();
6216
6217 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6218 BUG_ON(psr & IA64_PSR_I);
6219
6220 /*
6221 * we restore ALL the debug registers to avoid picking up
6222 * stale state.
6223 *
6224 * This must be done even when the task is still the owner
6225 * as the registers may have been modified via ptrace()
6226 * (not perfmon) by the previous task.
6227 */
6228 if (ctx->ctx_fl_using_dbreg) {
6229 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6230 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6231 }
6232
6233 /*
6234 * retrieved saved psr.up
6235 */
6236 psr_up = ctx->ctx_saved_psr_up;
6237 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6238
6239 /*
6240 * short path, our state is still there, just
6241 * need to restore psr and we go
6242 *
6243 * we do not touch either PMC nor PMD. the psr is not touched
6244 * by the overflow_handler. So we are safe w.r.t. to interrupt
6245 * concurrency even without interrupt masking.
6246 */
6247 if (likely(owner == task)) {
6248 if (likely(psr_up)) pfm_set_psr_up();
6249 return;
6250 }
6251
6252 /*
6253 * someone else is still using the PMU, first push it out and
6254 * then we'll be able to install our stuff !
6255 *
6256 * Upon return, there will be no owner for the current PMU
6257 */
6258 if (owner) pfm_lazy_save_regs(owner);
6259
6260 /*
6261 * To avoid leaking information to the user level when psr.sp=0,
6262 * we must reload ALL implemented pmds (even the ones we don't use).
6263 * In the kernel we only allow PFM_READ_PMDS on registers which
6264 * we initialized or requested (sampling) so there is no risk there.
6265 */
6266 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6267
6268 /*
6269 * ALL accessible PMCs are systematically reloaded, unused registers
6270 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6271 * up stale configuration.
6272 *
6273 * PMC0 is never in the mask. It is always restored separately
6274 */
6275 pmc_mask = ctx->ctx_all_pmcs[0];
6276
6277 pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6278 pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6279
6280 /*
6281 * check for pending overflow at the time the state
6282 * was saved.
6283 */
6284 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6285 /*
6286 * reload pmc0 with the overflow information
6287 * On McKinley PMU, this will trigger a PMU interrupt
6288 */
6289 ia64_set_pmc(0, ctx->th_pmcs[0]);
6290 ia64_srlz_d();
6291
6292 ctx->th_pmcs[0] = 0UL;
6293
6294 /*
6295 * will replay the PMU interrupt
6296 */
6297 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6298
6299 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6300 }
6301
6302 /*
6303 * establish new ownership.
6304 */
6305 SET_PMU_OWNER(task, ctx);
6306
6307 /*
6308 * restore the psr.up bit. measurement
6309 * is active again.
6310 * no PMU interrupt can happen at this point
6311 * because we still have interrupts disabled.
6312 */
6313 if (likely(psr_up)) pfm_set_psr_up();
6314 }
6315 #endif /* CONFIG_SMP */
6316
6317 /*
6318 * this function assumes monitoring is stopped
6319 */
6320 static void
6321 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6322 {
6323 u64 pmc0;
6324 unsigned long mask2, val, pmd_val, ovfl_val;
6325 int i, can_access_pmu = 0;
6326 int is_self;
6327
6328 /*
6329 * is the caller the task being monitored (or which initiated the
6330 * session for system wide measurements)
6331 */
6332 is_self = ctx->ctx_task == task ? 1 : 0;
6333
6334 /*
6335 * can access PMU is task is the owner of the PMU state on the current CPU
6336 * or if we are running on the CPU bound to the context in system-wide mode
6337 * (that is not necessarily the task the context is attached to in this mode).
6338 * In system-wide we always have can_access_pmu true because a task running on an
6339 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6340 */
6341 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6342 if (can_access_pmu) {
6343 /*
6344 * Mark the PMU as not owned
6345 * This will cause the interrupt handler to do nothing in case an overflow
6346 * interrupt was in-flight
6347 * This also guarantees that pmc0 will contain the final state
6348 * It virtually gives us full control on overflow processing from that point
6349 * on.
6350 */
6351 SET_PMU_OWNER(NULL, NULL);
6352 DPRINT(("releasing ownership\n"));
6353
6354 /*
6355 * read current overflow status:
6356 *
6357 * we are guaranteed to read the final stable state
6358 */
6359 ia64_srlz_d();
6360 pmc0 = ia64_get_pmc(0); /* slow */
6361
6362 /*
6363 * reset freeze bit, overflow status information destroyed
6364 */
6365 pfm_unfreeze_pmu();
6366 } else {
6367 pmc0 = ctx->th_pmcs[0];
6368 /*
6369 * clear whatever overflow status bits there were
6370 */
6371 ctx->th_pmcs[0] = 0;
6372 }
6373 ovfl_val = pmu_conf->ovfl_val;
6374 /*
6375 * we save all the used pmds
6376 * we take care of overflows for counting PMDs
6377 *
6378 * XXX: sampling situation is not taken into account here
6379 */
6380 mask2 = ctx->ctx_used_pmds[0];
6381
6382 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6383
6384 for (i = 0; mask2; i++, mask2>>=1) {
6385
6386 /* skip non used pmds */
6387 if ((mask2 & 0x1) == 0) continue;
6388
6389 /*
6390 * can access PMU always true in system wide mode
6391 */
6392 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6393
6394 if (PMD_IS_COUNTING(i)) {
6395 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6396 task_pid_nr(task),
6397 i,
6398 ctx->ctx_pmds[i].val,
6399 val & ovfl_val));
6400
6401 /*
6402 * we rebuild the full 64 bit value of the counter
6403 */
6404 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6405
6406 /*
6407 * now everything is in ctx_pmds[] and we need
6408 * to clear the saved context from save_regs() such that
6409 * pfm_read_pmds() gets the correct value
6410 */
6411 pmd_val = 0UL;
6412
6413 /*
6414 * take care of overflow inline
6415 */
6416 if (pmc0 & (1UL << i)) {
6417 val += 1 + ovfl_val;
6418 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6419 }
6420 }
6421
6422 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6423
6424 if (is_self) ctx->th_pmds[i] = pmd_val;
6425
6426 ctx->ctx_pmds[i].val = val;
6427 }
6428 }
6429
6430 static struct irqaction perfmon_irqaction = {
6431 .handler = pfm_interrupt_handler,
6432 .flags = IRQF_DISABLED,
6433 .name = "perfmon"
6434 };
6435
6436 static void
6437 pfm_alt_save_pmu_state(void *data)
6438 {
6439 struct pt_regs *regs;
6440
6441 regs = task_pt_regs(current);
6442
6443 DPRINT(("called\n"));
6444
6445 /*
6446 * should not be necessary but
6447 * let's take not risk
6448 */
6449 pfm_clear_psr_up();
6450 pfm_clear_psr_pp();
6451 ia64_psr(regs)->pp = 0;
6452
6453 /*
6454 * This call is required
6455 * May cause a spurious interrupt on some processors
6456 */
6457 pfm_freeze_pmu();
6458
6459 ia64_srlz_d();
6460 }
6461
6462 void
6463 pfm_alt_restore_pmu_state(void *data)
6464 {
6465 struct pt_regs *regs;
6466
6467 regs = task_pt_regs(current);
6468
6469 DPRINT(("called\n"));
6470
6471 /*
6472 * put PMU back in state expected
6473 * by perfmon
6474 */
6475 pfm_clear_psr_up();
6476 pfm_clear_psr_pp();
6477 ia64_psr(regs)->pp = 0;
6478
6479 /*
6480 * perfmon runs with PMU unfrozen at all times
6481 */
6482 pfm_unfreeze_pmu();
6483
6484 ia64_srlz_d();
6485 }
6486
6487 int
6488 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6489 {
6490 int ret, i;
6491 int reserve_cpu;
6492
6493 /* some sanity checks */
6494 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6495
6496 /* do the easy test first */
6497 if (pfm_alt_intr_handler) return -EBUSY;
6498
6499 /* one at a time in the install or remove, just fail the others */
6500 if (!spin_trylock(&pfm_alt_install_check)) {
6501 return -EBUSY;
6502 }
6503
6504 /* reserve our session */
6505 for_each_online_cpu(reserve_cpu) {
6506 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6507 if (ret) goto cleanup_reserve;
6508 }
6509
6510 /* save the current system wide pmu states */
6511 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 0, 1);
6512 if (ret) {
6513 DPRINT(("on_each_cpu() failed: %d\n", ret));
6514 goto cleanup_reserve;
6515 }
6516
6517 /* officially change to the alternate interrupt handler */
6518 pfm_alt_intr_handler = hdl;
6519
6520 spin_unlock(&pfm_alt_install_check);
6521
6522 return 0;
6523
6524 cleanup_reserve:
6525 for_each_online_cpu(i) {
6526 /* don't unreserve more than we reserved */
6527 if (i >= reserve_cpu) break;
6528
6529 pfm_unreserve_session(NULL, 1, i);
6530 }
6531
6532 spin_unlock(&pfm_alt_install_check);
6533
6534 return ret;
6535 }
6536 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6537
6538 int
6539 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6540 {
6541 int i;
6542 int ret;
6543
6544 if (hdl == NULL) return -EINVAL;
6545
6546 /* cannot remove someone else's handler! */
6547 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6548
6549 /* one at a time in the install or remove, just fail the others */
6550 if (!spin_trylock(&pfm_alt_install_check)) {
6551 return -EBUSY;
6552 }
6553
6554 pfm_alt_intr_handler = NULL;
6555
6556 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 0, 1);
6557 if (ret) {
6558 DPRINT(("on_each_cpu() failed: %d\n", ret));
6559 }
6560
6561 for_each_online_cpu(i) {
6562 pfm_unreserve_session(NULL, 1, i);
6563 }
6564
6565 spin_unlock(&pfm_alt_install_check);
6566
6567 return 0;
6568 }
6569 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6570
6571 /*
6572 * perfmon initialization routine, called from the initcall() table
6573 */
6574 static int init_pfm_fs(void);
6575
6576 static int __init
6577 pfm_probe_pmu(void)
6578 {
6579 pmu_config_t **p;
6580 int family;
6581
6582 family = local_cpu_data->family;
6583 p = pmu_confs;
6584
6585 while(*p) {
6586 if ((*p)->probe) {
6587 if ((*p)->probe() == 0) goto found;
6588 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6589 goto found;
6590 }
6591 p++;
6592 }
6593 return -1;
6594 found:
6595 pmu_conf = *p;
6596 return 0;
6597 }
6598
6599 static const struct file_operations pfm_proc_fops = {
6600 .open = pfm_proc_open,
6601 .read = seq_read,
6602 .llseek = seq_lseek,
6603 .release = seq_release,
6604 };
6605
6606 int __init
6607 pfm_init(void)
6608 {
6609 unsigned int n, n_counters, i;
6610
6611 printk("perfmon: version %u.%u IRQ %u\n",
6612 PFM_VERSION_MAJ,
6613 PFM_VERSION_MIN,
6614 IA64_PERFMON_VECTOR);
6615
6616 if (pfm_probe_pmu()) {
6617 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6618 local_cpu_data->family);
6619 return -ENODEV;
6620 }
6621
6622 /*
6623 * compute the number of implemented PMD/PMC from the
6624 * description tables
6625 */
6626 n = 0;
6627 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6628 if (PMC_IS_IMPL(i) == 0) continue;
6629 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6630 n++;
6631 }
6632 pmu_conf->num_pmcs = n;
6633
6634 n = 0; n_counters = 0;
6635 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6636 if (PMD_IS_IMPL(i) == 0) continue;
6637 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6638 n++;
6639 if (PMD_IS_COUNTING(i)) n_counters++;
6640 }
6641 pmu_conf->num_pmds = n;
6642 pmu_conf->num_counters = n_counters;
6643
6644 /*
6645 * sanity checks on the number of debug registers
6646 */
6647 if (pmu_conf->use_rr_dbregs) {
6648 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6649 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6650 pmu_conf = NULL;
6651 return -1;
6652 }
6653 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6654 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6655 pmu_conf = NULL;
6656 return -1;
6657 }
6658 }
6659
6660 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6661 pmu_conf->pmu_name,
6662 pmu_conf->num_pmcs,
6663 pmu_conf->num_pmds,
6664 pmu_conf->num_counters,
6665 ffz(pmu_conf->ovfl_val));
6666
6667 /* sanity check */
6668 if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6669 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6670 pmu_conf = NULL;
6671 return -1;
6672 }
6673
6674 /*
6675 * create /proc/perfmon (mostly for debugging purposes)
6676 */
6677 perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6678 if (perfmon_dir == NULL) {
6679 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6680 pmu_conf = NULL;
6681 return -1;
6682 }
6683
6684 /*
6685 * create /proc/sys/kernel/perfmon (for debugging purposes)
6686 */
6687 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6688
6689 /*
6690 * initialize all our spinlocks
6691 */
6692 spin_lock_init(&pfm_sessions.pfs_lock);
6693 spin_lock_init(&pfm_buffer_fmt_lock);
6694
6695 init_pfm_fs();
6696
6697 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6698
6699 return 0;
6700 }
6701
6702 __initcall(pfm_init);
6703
6704 /*
6705 * this function is called before pfm_init()
6706 */
6707 void
6708 pfm_init_percpu (void)
6709 {
6710 static int first_time=1;
6711 /*
6712 * make sure no measurement is active
6713 * (may inherit programmed PMCs from EFI).
6714 */
6715 pfm_clear_psr_pp();
6716 pfm_clear_psr_up();
6717
6718 /*
6719 * we run with the PMU not frozen at all times
6720 */
6721 pfm_unfreeze_pmu();
6722
6723 if (first_time) {
6724 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6725 first_time=0;
6726 }
6727
6728 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6729 ia64_srlz_d();
6730 }
6731
6732 /*
6733 * used for debug purposes only
6734 */
6735 void
6736 dump_pmu_state(const char *from)
6737 {
6738 struct task_struct *task;
6739 struct pt_regs *regs;
6740 pfm_context_t *ctx;
6741 unsigned long psr, dcr, info, flags;
6742 int i, this_cpu;
6743
6744 local_irq_save(flags);
6745
6746 this_cpu = smp_processor_id();
6747 regs = task_pt_regs(current);
6748 info = PFM_CPUINFO_GET();
6749 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6750
6751 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6752 local_irq_restore(flags);
6753 return;
6754 }
6755
6756 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6757 this_cpu,
6758 from,
6759 task_pid_nr(current),
6760 regs->cr_iip,
6761 current->comm);
6762
6763 task = GET_PMU_OWNER();
6764 ctx = GET_PMU_CTX();
6765
6766 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6767
6768 psr = pfm_get_psr();
6769
6770 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",
6771 this_cpu,
6772 ia64_get_pmc(0),
6773 psr & IA64_PSR_PP ? 1 : 0,
6774 psr & IA64_PSR_UP ? 1 : 0,
6775 dcr & IA64_DCR_PP ? 1 : 0,
6776 info,
6777 ia64_psr(regs)->up,
6778 ia64_psr(regs)->pp);
6779
6780 ia64_psr(regs)->up = 0;
6781 ia64_psr(regs)->pp = 0;
6782
6783 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6784 if (PMC_IS_IMPL(i) == 0) continue;
6785 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]);
6786 }
6787
6788 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6789 if (PMD_IS_IMPL(i) == 0) continue;
6790 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]);
6791 }
6792
6793 if (ctx) {
6794 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6795 this_cpu,
6796 ctx->ctx_state,
6797 ctx->ctx_smpl_vaddr,
6798 ctx->ctx_smpl_hdr,
6799 ctx->ctx_msgq_head,
6800 ctx->ctx_msgq_tail,
6801 ctx->ctx_saved_psr_up);
6802 }
6803 local_irq_restore(flags);
6804 }
6805
6806 /*
6807 * called from process.c:copy_thread(). task is new child.
6808 */
6809 void
6810 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6811 {
6812 struct thread_struct *thread;
6813
6814 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6815
6816 thread = &task->thread;
6817
6818 /*
6819 * cut links inherited from parent (current)
6820 */
6821 thread->pfm_context = NULL;
6822
6823 PFM_SET_WORK_PENDING(task, 0);
6824
6825 /*
6826 * the psr bits are already set properly in copy_threads()
6827 */
6828 }
6829 #else /* !CONFIG_PERFMON */
6830 asmlinkage long
6831 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6832 {
6833 return -ENOSYS;
6834 }
6835 #endif /* CONFIG_PERFMON */
Mirror for ia64_pv_ops.git