]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - arch/ia64/kernel/perfmon.c
[IA64] perfmon: convert to unlocked_ioctl
[mirror_ubuntu-bionic-kernel.git] / arch / ia64 / kernel / perfmon.c
1 /*
2 * This file implements the perfmon-2 subsystem which is used
3 * to program the IA-64 Performance Monitoring Unit (PMU).
4 *
5 * The initial version of perfmon.c was written by
6 * Ganesh Venkitachalam, IBM Corp.
7 *
8 * Then it was modified for perfmon-1.x by Stephane Eranian and
9 * David Mosberger, Hewlett Packard Co.
10 *
11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
12 * by Stephane Eranian, Hewlett Packard Co.
13 *
14 * Copyright (C) 1999-2005 Hewlett Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 * David Mosberger-Tang <davidm@hpl.hp.com>
17 *
18 * More information about perfmon available at:
19 * http://www.hpl.hp.com/research/linux/perfmon
20 */
21
22 #include <linux/module.h>
23 #include <linux/kernel.h>
24 #include <linux/sched.h>
25 #include <linux/interrupt.h>
26 #include <linux/proc_fs.h>
27 #include <linux/seq_file.h>
28 #include <linux/init.h>
29 #include <linux/vmalloc.h>
30 #include <linux/mm.h>
31 #include <linux/sysctl.h>
32 #include <linux/list.h>
33 #include <linux/file.h>
34 #include <linux/poll.h>
35 #include <linux/vfs.h>
36 #include <linux/smp.h>
37 #include <linux/pagemap.h>
38 #include <linux/mount.h>
39 #include <linux/bitops.h>
40 #include <linux/capability.h>
41 #include <linux/rcupdate.h>
42 #include <linux/completion.h>
43 #include <linux/tracehook.h>
44 #include <linux/slab.h>
45
46 #include <asm/errno.h>
47 #include <asm/intrinsics.h>
48 #include <asm/page.h>
49 #include <asm/perfmon.h>
50 #include <asm/processor.h>
51 #include <asm/signal.h>
52 #include <asm/system.h>
53 #include <asm/uaccess.h>
54 #include <asm/delay.h>
55
56 #ifdef CONFIG_PERFMON
57 /*
58 * perfmon context state
59 */
60 #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
61 #define PFM_CTX_LOADED 2 /* context is loaded onto a task */
62 #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
63 #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
64
65 #define PFM_INVALID_ACTIVATION (~0UL)
66
67 #define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
68 #define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
69
70 /*
71 * depth of message queue
72 */
73 #define PFM_MAX_MSGS 32
74 #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
75
76 /*
77 * type of a PMU register (bitmask).
78 * bitmask structure:
79 * bit0 : register implemented
80 * bit1 : end marker
81 * bit2-3 : reserved
82 * bit4 : pmc has pmc.pm
83 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
84 * bit6-7 : register type
85 * bit8-31: reserved
86 */
87 #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
88 #define PFM_REG_IMPL 0x1 /* register implemented */
89 #define PFM_REG_END 0x2 /* end marker */
90 #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
91 #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
92 #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
93 #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
94 #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
95
96 #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
97 #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
98
99 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
100
101 /* i assumed unsigned */
102 #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
103 #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
104
105 /* XXX: these assume that register i is implemented */
106 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
107 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
108 #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
109 #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
110
111 #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
112 #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
113 #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
114 #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
115
116 #define PFM_NUM_IBRS IA64_NUM_DBG_REGS
117 #define PFM_NUM_DBRS IA64_NUM_DBG_REGS
118
119 #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
120 #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
121 #define PFM_CTX_TASK(h) (h)->ctx_task
122
123 #define PMU_PMC_OI 5 /* position of pmc.oi bit */
124
125 /* XXX: does not support more than 64 PMDs */
126 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
127 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
128
129 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
130
131 #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
132 #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
133 #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
134 #define PFM_CODE_RR 0 /* requesting code range restriction */
135 #define PFM_DATA_RR 1 /* requestion data range restriction */
136
137 #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
138 #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
139 #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
140
141 #define RDEP(x) (1UL<<(x))
142
143 /*
144 * context protection macros
145 * in SMP:
146 * - we need to protect against CPU concurrency (spin_lock)
147 * - we need to protect against PMU overflow interrupts (local_irq_disable)
148 * in UP:
149 * - we need to protect against PMU overflow interrupts (local_irq_disable)
150 *
151 * spin_lock_irqsave()/spin_unlock_irqrestore():
152 * in SMP: local_irq_disable + spin_lock
153 * in UP : local_irq_disable
154 *
155 * spin_lock()/spin_lock():
156 * in UP : removed automatically
157 * in SMP: protect against context accesses from other CPU. interrupts
158 * are not masked. This is useful for the PMU interrupt handler
159 * because we know we will not get PMU concurrency in that code.
160 */
161 #define PROTECT_CTX(c, f) \
162 do { \
163 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
164 spin_lock_irqsave(&(c)->ctx_lock, f); \
165 DPRINT(("spinlocked ctx %p by [%d]\n", c, task_pid_nr(current))); \
166 } while(0)
167
168 #define UNPROTECT_CTX(c, f) \
169 do { \
170 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
171 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
172 } while(0)
173
174 #define PROTECT_CTX_NOPRINT(c, f) \
175 do { \
176 spin_lock_irqsave(&(c)->ctx_lock, f); \
177 } while(0)
178
179
180 #define UNPROTECT_CTX_NOPRINT(c, f) \
181 do { \
182 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
183 } while(0)
184
185
186 #define PROTECT_CTX_NOIRQ(c) \
187 do { \
188 spin_lock(&(c)->ctx_lock); \
189 } while(0)
190
191 #define UNPROTECT_CTX_NOIRQ(c) \
192 do { \
193 spin_unlock(&(c)->ctx_lock); \
194 } while(0)
195
196
197 #ifdef CONFIG_SMP
198
199 #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
200 #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
201 #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
202
203 #else /* !CONFIG_SMP */
204 #define SET_ACTIVATION(t) do {} while(0)
205 #define GET_ACTIVATION(t) do {} while(0)
206 #define INC_ACTIVATION(t) do {} while(0)
207 #endif /* CONFIG_SMP */
208
209 #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
210 #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
211 #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
212
213 #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
214 #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
215
216 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
217
218 /*
219 * cmp0 must be the value of pmc0
220 */
221 #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
222
223 #define PFMFS_MAGIC 0xa0b4d889
224
225 /*
226 * debugging
227 */
228 #define PFM_DEBUGGING 1
229 #ifdef PFM_DEBUGGING
230 #define DPRINT(a) \
231 do { \
232 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
233 } while (0)
234
235 #define DPRINT_ovfl(a) \
236 do { \
237 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; } \
238 } while (0)
239 #endif
240
241 /*
242 * 64-bit software counter structure
243 *
244 * the next_reset_type is applied to the next call to pfm_reset_regs()
245 */
246 typedef struct {
247 unsigned long val; /* virtual 64bit counter value */
248 unsigned long lval; /* last reset value */
249 unsigned long long_reset; /* reset value on sampling overflow */
250 unsigned long short_reset; /* reset value on overflow */
251 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
252 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
253 unsigned long seed; /* seed for random-number generator */
254 unsigned long mask; /* mask for random-number generator */
255 unsigned int flags; /* notify/do not notify */
256 unsigned long eventid; /* overflow event identifier */
257 } pfm_counter_t;
258
259 /*
260 * context flags
261 */
262 typedef struct {
263 unsigned int block:1; /* when 1, task will blocked on user notifications */
264 unsigned int system:1; /* do system wide monitoring */
265 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
266 unsigned int is_sampling:1; /* true if using a custom format */
267 unsigned int excl_idle:1; /* exclude idle task in system wide session */
268 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
269 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
270 unsigned int no_msg:1; /* no message sent on overflow */
271 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
272 unsigned int reserved:22;
273 } pfm_context_flags_t;
274
275 #define PFM_TRAP_REASON_NONE 0x0 /* default value */
276 #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
277 #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
278
279
280 /*
281 * perfmon context: encapsulates all the state of a monitoring session
282 */
283
284 typedef struct pfm_context {
285 spinlock_t ctx_lock; /* context protection */
286
287 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
288 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
289
290 struct task_struct *ctx_task; /* task to which context is attached */
291
292 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
293
294 struct completion ctx_restart_done; /* use for blocking notification mode */
295
296 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
297 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
298 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
299
300 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
301 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
302 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
303
304 unsigned long ctx_pmcs[PFM_NUM_PMC_REGS]; /* saved copies of PMC values */
305
306 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
307 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
308 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
309 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
310
311 pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
312
313 unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
314 unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
315
316 unsigned long ctx_saved_psr_up; /* only contains psr.up value */
317
318 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
319 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
320 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
321
322 int ctx_fd; /* file descriptor used my this context */
323 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
324
325 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
326 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
327 unsigned long ctx_smpl_size; /* size of sampling buffer */
328 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
329
330 wait_queue_head_t ctx_msgq_wait;
331 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
332 int ctx_msgq_head;
333 int ctx_msgq_tail;
334 struct fasync_struct *ctx_async_queue;
335
336 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
337 } pfm_context_t;
338
339 /*
340 * magic number used to verify that structure is really
341 * a perfmon context
342 */
343 #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
344
345 #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
346
347 #ifdef CONFIG_SMP
348 #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
349 #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
350 #else
351 #define SET_LAST_CPU(ctx, v) do {} while(0)
352 #define GET_LAST_CPU(ctx) do {} while(0)
353 #endif
354
355
356 #define ctx_fl_block ctx_flags.block
357 #define ctx_fl_system ctx_flags.system
358 #define ctx_fl_using_dbreg ctx_flags.using_dbreg
359 #define ctx_fl_is_sampling ctx_flags.is_sampling
360 #define ctx_fl_excl_idle ctx_flags.excl_idle
361 #define ctx_fl_going_zombie ctx_flags.going_zombie
362 #define ctx_fl_trap_reason ctx_flags.trap_reason
363 #define ctx_fl_no_msg ctx_flags.no_msg
364 #define ctx_fl_can_restart ctx_flags.can_restart
365
366 #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
367 #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
368
369 /*
370 * global information about all sessions
371 * mostly used to synchronize between system wide and per-process
372 */
373 typedef struct {
374 spinlock_t pfs_lock; /* lock the structure */
375
376 unsigned int pfs_task_sessions; /* number of per task sessions */
377 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
378 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
379 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
380 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
381 } pfm_session_t;
382
383 /*
384 * information about a PMC or PMD.
385 * dep_pmd[]: a bitmask of dependent PMD registers
386 * dep_pmc[]: a bitmask of dependent PMC registers
387 */
388 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
389 typedef struct {
390 unsigned int type;
391 int pm_pos;
392 unsigned long default_value; /* power-on default value */
393 unsigned long reserved_mask; /* bitmask of reserved bits */
394 pfm_reg_check_t read_check;
395 pfm_reg_check_t write_check;
396 unsigned long dep_pmd[4];
397 unsigned long dep_pmc[4];
398 } pfm_reg_desc_t;
399
400 /* assume cnum is a valid monitor */
401 #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
402
403 /*
404 * This structure is initialized at boot time and contains
405 * a description of the PMU main characteristics.
406 *
407 * If the probe function is defined, detection is based
408 * on its return value:
409 * - 0 means recognized PMU
410 * - anything else means not supported
411 * When the probe function is not defined, then the pmu_family field
412 * is used and it must match the host CPU family such that:
413 * - cpu->family & config->pmu_family != 0
414 */
415 typedef struct {
416 unsigned long ovfl_val; /* overflow value for counters */
417
418 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
419 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
420
421 unsigned int num_pmcs; /* number of PMCS: computed at init time */
422 unsigned int num_pmds; /* number of PMDS: computed at init time */
423 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
424 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
425
426 char *pmu_name; /* PMU family name */
427 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
428 unsigned int flags; /* pmu specific flags */
429 unsigned int num_ibrs; /* number of IBRS: computed at init time */
430 unsigned int num_dbrs; /* number of DBRS: computed at init time */
431 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
432 int (*probe)(void); /* customized probe routine */
433 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
434 } pmu_config_t;
435 /*
436 * PMU specific flags
437 */
438 #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
439
440 /*
441 * debug register related type definitions
442 */
443 typedef struct {
444 unsigned long ibr_mask:56;
445 unsigned long ibr_plm:4;
446 unsigned long ibr_ig:3;
447 unsigned long ibr_x:1;
448 } ibr_mask_reg_t;
449
450 typedef struct {
451 unsigned long dbr_mask:56;
452 unsigned long dbr_plm:4;
453 unsigned long dbr_ig:2;
454 unsigned long dbr_w:1;
455 unsigned long dbr_r:1;
456 } dbr_mask_reg_t;
457
458 typedef union {
459 unsigned long val;
460 ibr_mask_reg_t ibr;
461 dbr_mask_reg_t dbr;
462 } dbreg_t;
463
464
465 /*
466 * perfmon command descriptions
467 */
468 typedef struct {
469 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
470 char *cmd_name;
471 int cmd_flags;
472 unsigned int cmd_narg;
473 size_t cmd_argsize;
474 int (*cmd_getsize)(void *arg, size_t *sz);
475 } pfm_cmd_desc_t;
476
477 #define PFM_CMD_FD 0x01 /* command requires a file descriptor */
478 #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
479 #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
480 #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
481
482
483 #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
484 #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
485 #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
486 #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
487 #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
488
489 #define PFM_CMD_ARG_MANY -1 /* cannot be zero */
490
491 typedef struct {
492 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
493 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
494 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
496 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
497 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
498 unsigned long pfm_smpl_handler_calls;
499 unsigned long pfm_smpl_handler_cycles;
500 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
501 } pfm_stats_t;
502
503 /*
504 * perfmon internal variables
505 */
506 static pfm_stats_t pfm_stats[NR_CPUS];
507 static pfm_session_t pfm_sessions; /* global sessions information */
508
509 static DEFINE_SPINLOCK(pfm_alt_install_check);
510 static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
511
512 static struct proc_dir_entry *perfmon_dir;
513 static pfm_uuid_t pfm_null_uuid = {0,};
514
515 static spinlock_t pfm_buffer_fmt_lock;
516 static LIST_HEAD(pfm_buffer_fmt_list);
517
518 static pmu_config_t *pmu_conf;
519
520 /* sysctl() controls */
521 pfm_sysctl_t pfm_sysctl;
522 EXPORT_SYMBOL(pfm_sysctl);
523
524 static ctl_table pfm_ctl_table[]={
525 {
526 .procname = "debug",
527 .data = &pfm_sysctl.debug,
528 .maxlen = sizeof(int),
529 .mode = 0666,
530 .proc_handler = proc_dointvec,
531 },
532 {
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 .procname = "fastctxsw",
541 .data = &pfm_sysctl.fastctxsw,
542 .maxlen = sizeof(int),
543 .mode = 0600,
544 .proc_handler = proc_dointvec,
545 },
546 {
547 .procname = "expert_mode",
548 .data = &pfm_sysctl.expert_mode,
549 .maxlen = sizeof(int),
550 .mode = 0600,
551 .proc_handler = proc_dointvec,
552 },
553 {}
554 };
555 static ctl_table pfm_sysctl_dir[] = {
556 {
557 .procname = "perfmon",
558 .mode = 0555,
559 .child = pfm_ctl_table,
560 },
561 {}
562 };
563 static ctl_table pfm_sysctl_root[] = {
564 {
565 .procname = "kernel",
566 .mode = 0555,
567 .child = pfm_sysctl_dir,
568 },
569 {}
570 };
571 static struct ctl_table_header *pfm_sysctl_header;
572
573 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
574
575 #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
576 #define pfm_get_cpu_data(a,b) per_cpu(a, b)
577
578 static inline void
579 pfm_put_task(struct task_struct *task)
580 {
581 if (task != current) put_task_struct(task);
582 }
583
584 static inline void
585 pfm_reserve_page(unsigned long a)
586 {
587 SetPageReserved(vmalloc_to_page((void *)a));
588 }
589 static inline void
590 pfm_unreserve_page(unsigned long a)
591 {
592 ClearPageReserved(vmalloc_to_page((void*)a));
593 }
594
595 static inline unsigned long
596 pfm_protect_ctx_ctxsw(pfm_context_t *x)
597 {
598 spin_lock(&(x)->ctx_lock);
599 return 0UL;
600 }
601
602 static inline void
603 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
604 {
605 spin_unlock(&(x)->ctx_lock);
606 }
607
608 static inline unsigned int
609 pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
610 {
611 return do_munmap(mm, addr, len);
612 }
613
614 static inline unsigned long
615 pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
616 {
617 return get_unmapped_area(file, addr, len, pgoff, flags);
618 }
619
620
621 static int
622 pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data,
623 struct vfsmount *mnt)
624 {
625 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
626 }
627
628 static struct file_system_type pfm_fs_type = {
629 .name = "pfmfs",
630 .get_sb = pfmfs_get_sb,
631 .kill_sb = kill_anon_super,
632 };
633
634 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
635 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
636 DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
637 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
638 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
639
640
641 /* forward declaration */
642 static const struct file_operations pfm_file_ops;
643
644 /*
645 * forward declarations
646 */
647 #ifndef CONFIG_SMP
648 static void pfm_lazy_save_regs (struct task_struct *ta);
649 #endif
650
651 void dump_pmu_state(const char *);
652 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
653
654 #include "perfmon_itanium.h"
655 #include "perfmon_mckinley.h"
656 #include "perfmon_montecito.h"
657 #include "perfmon_generic.h"
658
659 static pmu_config_t *pmu_confs[]={
660 &pmu_conf_mont,
661 &pmu_conf_mck,
662 &pmu_conf_ita,
663 &pmu_conf_gen, /* must be last */
664 NULL
665 };
666
667
668 static int pfm_end_notify_user(pfm_context_t *ctx);
669
670 static inline void
671 pfm_clear_psr_pp(void)
672 {
673 ia64_rsm(IA64_PSR_PP);
674 ia64_srlz_i();
675 }
676
677 static inline void
678 pfm_set_psr_pp(void)
679 {
680 ia64_ssm(IA64_PSR_PP);
681 ia64_srlz_i();
682 }
683
684 static inline void
685 pfm_clear_psr_up(void)
686 {
687 ia64_rsm(IA64_PSR_UP);
688 ia64_srlz_i();
689 }
690
691 static inline void
692 pfm_set_psr_up(void)
693 {
694 ia64_ssm(IA64_PSR_UP);
695 ia64_srlz_i();
696 }
697
698 static inline unsigned long
699 pfm_get_psr(void)
700 {
701 unsigned long tmp;
702 tmp = ia64_getreg(_IA64_REG_PSR);
703 ia64_srlz_i();
704 return tmp;
705 }
706
707 static inline void
708 pfm_set_psr_l(unsigned long val)
709 {
710 ia64_setreg(_IA64_REG_PSR_L, val);
711 ia64_srlz_i();
712 }
713
714 static inline void
715 pfm_freeze_pmu(void)
716 {
717 ia64_set_pmc(0,1UL);
718 ia64_srlz_d();
719 }
720
721 static inline void
722 pfm_unfreeze_pmu(void)
723 {
724 ia64_set_pmc(0,0UL);
725 ia64_srlz_d();
726 }
727
728 static inline void
729 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
730 {
731 int i;
732
733 for (i=0; i < nibrs; i++) {
734 ia64_set_ibr(i, ibrs[i]);
735 ia64_dv_serialize_instruction();
736 }
737 ia64_srlz_i();
738 }
739
740 static inline void
741 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
742 {
743 int i;
744
745 for (i=0; i < ndbrs; i++) {
746 ia64_set_dbr(i, dbrs[i]);
747 ia64_dv_serialize_data();
748 }
749 ia64_srlz_d();
750 }
751
752 /*
753 * PMD[i] must be a counter. no check is made
754 */
755 static inline unsigned long
756 pfm_read_soft_counter(pfm_context_t *ctx, int i)
757 {
758 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
759 }
760
761 /*
762 * PMD[i] must be a counter. no check is made
763 */
764 static inline void
765 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
766 {
767 unsigned long ovfl_val = pmu_conf->ovfl_val;
768
769 ctx->ctx_pmds[i].val = val & ~ovfl_val;
770 /*
771 * writing to unimplemented part is ignore, so we do not need to
772 * mask off top part
773 */
774 ia64_set_pmd(i, val & ovfl_val);
775 }
776
777 static pfm_msg_t *
778 pfm_get_new_msg(pfm_context_t *ctx)
779 {
780 int idx, next;
781
782 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
783
784 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
785 if (next == ctx->ctx_msgq_head) return NULL;
786
787 idx = ctx->ctx_msgq_tail;
788 ctx->ctx_msgq_tail = next;
789
790 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
791
792 return ctx->ctx_msgq+idx;
793 }
794
795 static pfm_msg_t *
796 pfm_get_next_msg(pfm_context_t *ctx)
797 {
798 pfm_msg_t *msg;
799
800 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
801
802 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
803
804 /*
805 * get oldest message
806 */
807 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
808
809 /*
810 * and move forward
811 */
812 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
813
814 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));
815
816 return msg;
817 }
818
819 static void
820 pfm_reset_msgq(pfm_context_t *ctx)
821 {
822 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
823 DPRINT(("ctx=%p msgq reset\n", ctx));
824 }
825
826 static void *
827 pfm_rvmalloc(unsigned long size)
828 {
829 void *mem;
830 unsigned long addr;
831
832 size = PAGE_ALIGN(size);
833 mem = vmalloc(size);
834 if (mem) {
835 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
836 memset(mem, 0, size);
837 addr = (unsigned long)mem;
838 while (size > 0) {
839 pfm_reserve_page(addr);
840 addr+=PAGE_SIZE;
841 size-=PAGE_SIZE;
842 }
843 }
844 return mem;
845 }
846
847 static void
848 pfm_rvfree(void *mem, unsigned long size)
849 {
850 unsigned long addr;
851
852 if (mem) {
853 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
854 addr = (unsigned long) mem;
855 while ((long) size > 0) {
856 pfm_unreserve_page(addr);
857 addr+=PAGE_SIZE;
858 size-=PAGE_SIZE;
859 }
860 vfree(mem);
861 }
862 return;
863 }
864
865 static pfm_context_t *
866 pfm_context_alloc(int ctx_flags)
867 {
868 pfm_context_t *ctx;
869
870 /*
871 * allocate context descriptor
872 * must be able to free with interrupts disabled
873 */
874 ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
875 if (ctx) {
876 DPRINT(("alloc ctx @%p\n", ctx));
877
878 /*
879 * init context protection lock
880 */
881 spin_lock_init(&ctx->ctx_lock);
882
883 /*
884 * context is unloaded
885 */
886 ctx->ctx_state = PFM_CTX_UNLOADED;
887
888 /*
889 * initialization of context's flags
890 */
891 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
892 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
893 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
894 /*
895 * will move to set properties
896 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
897 */
898
899 /*
900 * init restart semaphore to locked
901 */
902 init_completion(&ctx->ctx_restart_done);
903
904 /*
905 * activation is used in SMP only
906 */
907 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
908 SET_LAST_CPU(ctx, -1);
909
910 /*
911 * initialize notification message queue
912 */
913 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
914 init_waitqueue_head(&ctx->ctx_msgq_wait);
915 init_waitqueue_head(&ctx->ctx_zombieq);
916
917 }
918 return ctx;
919 }
920
921 static void
922 pfm_context_free(pfm_context_t *ctx)
923 {
924 if (ctx) {
925 DPRINT(("free ctx @%p\n", ctx));
926 kfree(ctx);
927 }
928 }
929
930 static void
931 pfm_mask_monitoring(struct task_struct *task)
932 {
933 pfm_context_t *ctx = PFM_GET_CTX(task);
934 unsigned long mask, val, ovfl_mask;
935 int i;
936
937 DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
938
939 ovfl_mask = pmu_conf->ovfl_val;
940 /*
941 * monitoring can only be masked as a result of a valid
942 * counter overflow. In UP, it means that the PMU still
943 * has an owner. Note that the owner can be different
944 * from the current task. However the PMU state belongs
945 * to the owner.
946 * In SMP, a valid overflow only happens when task is
947 * current. Therefore if we come here, we know that
948 * the PMU state belongs to the current task, therefore
949 * we can access the live registers.
950 *
951 * So in both cases, the live register contains the owner's
952 * state. We can ONLY touch the PMU registers and NOT the PSR.
953 *
954 * As a consequence to this call, the ctx->th_pmds[] array
955 * contains stale information which must be ignored
956 * when context is reloaded AND monitoring is active (see
957 * pfm_restart).
958 */
959 mask = ctx->ctx_used_pmds[0];
960 for (i = 0; mask; i++, mask>>=1) {
961 /* skip non used pmds */
962 if ((mask & 0x1) == 0) continue;
963 val = ia64_get_pmd(i);
964
965 if (PMD_IS_COUNTING(i)) {
966 /*
967 * we rebuild the full 64 bit value of the counter
968 */
969 ctx->ctx_pmds[i].val += (val & ovfl_mask);
970 } else {
971 ctx->ctx_pmds[i].val = val;
972 }
973 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
974 i,
975 ctx->ctx_pmds[i].val,
976 val & ovfl_mask));
977 }
978 /*
979 * mask monitoring by setting the privilege level to 0
980 * we cannot use psr.pp/psr.up for this, it is controlled by
981 * the user
982 *
983 * if task is current, modify actual registers, otherwise modify
984 * thread save state, i.e., what will be restored in pfm_load_regs()
985 */
986 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
987 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
988 if ((mask & 0x1) == 0UL) continue;
989 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
990 ctx->th_pmcs[i] &= ~0xfUL;
991 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
992 }
993 /*
994 * make all of this visible
995 */
996 ia64_srlz_d();
997 }
998
999 /*
1000 * must always be done with task == current
1001 *
1002 * context must be in MASKED state when calling
1003 */
1004 static void
1005 pfm_restore_monitoring(struct task_struct *task)
1006 {
1007 pfm_context_t *ctx = PFM_GET_CTX(task);
1008 unsigned long mask, ovfl_mask;
1009 unsigned long psr, val;
1010 int i, is_system;
1011
1012 is_system = ctx->ctx_fl_system;
1013 ovfl_mask = pmu_conf->ovfl_val;
1014
1015 if (task != current) {
1016 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1017 return;
1018 }
1019 if (ctx->ctx_state != PFM_CTX_MASKED) {
1020 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1021 task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1022 return;
1023 }
1024 psr = pfm_get_psr();
1025 /*
1026 * monitoring is masked via the PMC.
1027 * As we restore their value, we do not want each counter to
1028 * restart right away. We stop monitoring using the PSR,
1029 * restore the PMC (and PMD) and then re-establish the psr
1030 * as it was. Note that there can be no pending overflow at
1031 * this point, because monitoring was MASKED.
1032 *
1033 * system-wide session are pinned and self-monitoring
1034 */
1035 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1036 /* disable dcr pp */
1037 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1038 pfm_clear_psr_pp();
1039 } else {
1040 pfm_clear_psr_up();
1041 }
1042 /*
1043 * first, we restore the PMD
1044 */
1045 mask = ctx->ctx_used_pmds[0];
1046 for (i = 0; mask; i++, mask>>=1) {
1047 /* skip non used pmds */
1048 if ((mask & 0x1) == 0) continue;
1049
1050 if (PMD_IS_COUNTING(i)) {
1051 /*
1052 * we split the 64bit value according to
1053 * counter width
1054 */
1055 val = ctx->ctx_pmds[i].val & ovfl_mask;
1056 ctx->ctx_pmds[i].val &= ~ovfl_mask;
1057 } else {
1058 val = ctx->ctx_pmds[i].val;
1059 }
1060 ia64_set_pmd(i, val);
1061
1062 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1063 i,
1064 ctx->ctx_pmds[i].val,
1065 val));
1066 }
1067 /*
1068 * restore the PMCs
1069 */
1070 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1071 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1072 if ((mask & 0x1) == 0UL) continue;
1073 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1074 ia64_set_pmc(i, ctx->th_pmcs[i]);
1075 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1076 task_pid_nr(task), i, ctx->th_pmcs[i]));
1077 }
1078 ia64_srlz_d();
1079
1080 /*
1081 * must restore DBR/IBR because could be modified while masked
1082 * XXX: need to optimize
1083 */
1084 if (ctx->ctx_fl_using_dbreg) {
1085 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1086 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1087 }
1088
1089 /*
1090 * now restore PSR
1091 */
1092 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1093 /* enable dcr pp */
1094 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1095 ia64_srlz_i();
1096 }
1097 pfm_set_psr_l(psr);
1098 }
1099
1100 static inline void
1101 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1102 {
1103 int i;
1104
1105 ia64_srlz_d();
1106
1107 for (i=0; mask; i++, mask>>=1) {
1108 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1109 }
1110 }
1111
1112 /*
1113 * reload from thread state (used for ctxw only)
1114 */
1115 static inline void
1116 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1117 {
1118 int i;
1119 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1120
1121 for (i=0; mask; i++, mask>>=1) {
1122 if ((mask & 0x1) == 0) continue;
1123 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1124 ia64_set_pmd(i, val);
1125 }
1126 ia64_srlz_d();
1127 }
1128
1129 /*
1130 * propagate PMD from context to thread-state
1131 */
1132 static inline void
1133 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1134 {
1135 unsigned long ovfl_val = pmu_conf->ovfl_val;
1136 unsigned long mask = ctx->ctx_all_pmds[0];
1137 unsigned long val;
1138 int i;
1139
1140 DPRINT(("mask=0x%lx\n", mask));
1141
1142 for (i=0; mask; i++, mask>>=1) {
1143
1144 val = ctx->ctx_pmds[i].val;
1145
1146 /*
1147 * We break up the 64 bit value into 2 pieces
1148 * the lower bits go to the machine state in the
1149 * thread (will be reloaded on ctxsw in).
1150 * The upper part stays in the soft-counter.
1151 */
1152 if (PMD_IS_COUNTING(i)) {
1153 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1154 val &= ovfl_val;
1155 }
1156 ctx->th_pmds[i] = val;
1157
1158 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1159 i,
1160 ctx->th_pmds[i],
1161 ctx->ctx_pmds[i].val));
1162 }
1163 }
1164
1165 /*
1166 * propagate PMC from context to thread-state
1167 */
1168 static inline void
1169 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1170 {
1171 unsigned long mask = ctx->ctx_all_pmcs[0];
1172 int i;
1173
1174 DPRINT(("mask=0x%lx\n", mask));
1175
1176 for (i=0; mask; i++, mask>>=1) {
1177 /* masking 0 with ovfl_val yields 0 */
1178 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1179 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1180 }
1181 }
1182
1183
1184
1185 static inline void
1186 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1187 {
1188 int i;
1189
1190 for (i=0; mask; i++, mask>>=1) {
1191 if ((mask & 0x1) == 0) continue;
1192 ia64_set_pmc(i, pmcs[i]);
1193 }
1194 ia64_srlz_d();
1195 }
1196
1197 static inline int
1198 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1199 {
1200 return memcmp(a, b, sizeof(pfm_uuid_t));
1201 }
1202
1203 static inline int
1204 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1205 {
1206 int ret = 0;
1207 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1208 return ret;
1209 }
1210
1211 static inline int
1212 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1213 {
1214 int ret = 0;
1215 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1216 return ret;
1217 }
1218
1219
1220 static inline int
1221 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1222 int cpu, void *arg)
1223 {
1224 int ret = 0;
1225 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1226 return ret;
1227 }
1228
1229 static inline int
1230 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1231 int cpu, void *arg)
1232 {
1233 int ret = 0;
1234 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1235 return ret;
1236 }
1237
1238 static inline int
1239 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1240 {
1241 int ret = 0;
1242 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1243 return ret;
1244 }
1245
1246 static inline int
1247 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)
1248 {
1249 int ret = 0;
1250 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1251 return ret;
1252 }
1253
1254 static pfm_buffer_fmt_t *
1255 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1256 {
1257 struct list_head * pos;
1258 pfm_buffer_fmt_t * entry;
1259
1260 list_for_each(pos, &pfm_buffer_fmt_list) {
1261 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1262 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1263 return entry;
1264 }
1265 return NULL;
1266 }
1267
1268 /*
1269 * find a buffer format based on its uuid
1270 */
1271 static pfm_buffer_fmt_t *
1272 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1273 {
1274 pfm_buffer_fmt_t * fmt;
1275 spin_lock(&pfm_buffer_fmt_lock);
1276 fmt = __pfm_find_buffer_fmt(uuid);
1277 spin_unlock(&pfm_buffer_fmt_lock);
1278 return fmt;
1279 }
1280
1281 int
1282 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1283 {
1284 int ret = 0;
1285
1286 /* some sanity checks */
1287 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1288
1289 /* we need at least a handler */
1290 if (fmt->fmt_handler == NULL) return -EINVAL;
1291
1292 /*
1293 * XXX: need check validity of fmt_arg_size
1294 */
1295
1296 spin_lock(&pfm_buffer_fmt_lock);
1297
1298 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1299 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1300 ret = -EBUSY;
1301 goto out;
1302 }
1303 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1304 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1305
1306 out:
1307 spin_unlock(&pfm_buffer_fmt_lock);
1308 return ret;
1309 }
1310 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1311
1312 int
1313 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1314 {
1315 pfm_buffer_fmt_t *fmt;
1316 int ret = 0;
1317
1318 spin_lock(&pfm_buffer_fmt_lock);
1319
1320 fmt = __pfm_find_buffer_fmt(uuid);
1321 if (!fmt) {
1322 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1323 ret = -EINVAL;
1324 goto out;
1325 }
1326 list_del_init(&fmt->fmt_list);
1327 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1328
1329 out:
1330 spin_unlock(&pfm_buffer_fmt_lock);
1331 return ret;
1332
1333 }
1334 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1335
1336 extern void update_pal_halt_status(int);
1337
1338 static int
1339 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1340 {
1341 unsigned long flags;
1342 /*
1343 * validity checks on cpu_mask have been done upstream
1344 */
1345 LOCK_PFS(flags);
1346
1347 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1348 pfm_sessions.pfs_sys_sessions,
1349 pfm_sessions.pfs_task_sessions,
1350 pfm_sessions.pfs_sys_use_dbregs,
1351 is_syswide,
1352 cpu));
1353
1354 if (is_syswide) {
1355 /*
1356 * cannot mix system wide and per-task sessions
1357 */
1358 if (pfm_sessions.pfs_task_sessions > 0UL) {
1359 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1360 pfm_sessions.pfs_task_sessions));
1361 goto abort;
1362 }
1363
1364 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1365
1366 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1367
1368 pfm_sessions.pfs_sys_session[cpu] = task;
1369
1370 pfm_sessions.pfs_sys_sessions++ ;
1371
1372 } else {
1373 if (pfm_sessions.pfs_sys_sessions) goto abort;
1374 pfm_sessions.pfs_task_sessions++;
1375 }
1376
1377 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1378 pfm_sessions.pfs_sys_sessions,
1379 pfm_sessions.pfs_task_sessions,
1380 pfm_sessions.pfs_sys_use_dbregs,
1381 is_syswide,
1382 cpu));
1383
1384 /*
1385 * disable default_idle() to go to PAL_HALT
1386 */
1387 update_pal_halt_status(0);
1388
1389 UNLOCK_PFS(flags);
1390
1391 return 0;
1392
1393 error_conflict:
1394 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1395 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1396 cpu));
1397 abort:
1398 UNLOCK_PFS(flags);
1399
1400 return -EBUSY;
1401
1402 }
1403
1404 static int
1405 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1406 {
1407 unsigned long flags;
1408 /*
1409 * validity checks on cpu_mask have been done upstream
1410 */
1411 LOCK_PFS(flags);
1412
1413 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1414 pfm_sessions.pfs_sys_sessions,
1415 pfm_sessions.pfs_task_sessions,
1416 pfm_sessions.pfs_sys_use_dbregs,
1417 is_syswide,
1418 cpu));
1419
1420
1421 if (is_syswide) {
1422 pfm_sessions.pfs_sys_session[cpu] = NULL;
1423 /*
1424 * would not work with perfmon+more than one bit in cpu_mask
1425 */
1426 if (ctx && ctx->ctx_fl_using_dbreg) {
1427 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1428 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1429 } else {
1430 pfm_sessions.pfs_sys_use_dbregs--;
1431 }
1432 }
1433 pfm_sessions.pfs_sys_sessions--;
1434 } else {
1435 pfm_sessions.pfs_task_sessions--;
1436 }
1437 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1438 pfm_sessions.pfs_sys_sessions,
1439 pfm_sessions.pfs_task_sessions,
1440 pfm_sessions.pfs_sys_use_dbregs,
1441 is_syswide,
1442 cpu));
1443
1444 /*
1445 * if possible, enable default_idle() to go into PAL_HALT
1446 */
1447 if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1448 update_pal_halt_status(1);
1449
1450 UNLOCK_PFS(flags);
1451
1452 return 0;
1453 }
1454
1455 /*
1456 * removes virtual mapping of the sampling buffer.
1457 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1458 * a PROTECT_CTX() section.
1459 */
1460 static int
1461 pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1462 {
1463 int r;
1464
1465 /* sanity checks */
1466 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1467 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1468 return -EINVAL;
1469 }
1470
1471 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1472
1473 /*
1474 * does the actual unmapping
1475 */
1476 down_write(&task->mm->mmap_sem);
1477
1478 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1479
1480 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1481
1482 up_write(&task->mm->mmap_sem);
1483 if (r !=0) {
1484 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1485 }
1486
1487 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1488
1489 return 0;
1490 }
1491
1492 /*
1493 * free actual physical storage used by sampling buffer
1494 */
1495 #if 0
1496 static int
1497 pfm_free_smpl_buffer(pfm_context_t *ctx)
1498 {
1499 pfm_buffer_fmt_t *fmt;
1500
1501 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1502
1503 /*
1504 * we won't use the buffer format anymore
1505 */
1506 fmt = ctx->ctx_buf_fmt;
1507
1508 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1509 ctx->ctx_smpl_hdr,
1510 ctx->ctx_smpl_size,
1511 ctx->ctx_smpl_vaddr));
1512
1513 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1514
1515 /*
1516 * free the buffer
1517 */
1518 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1519
1520 ctx->ctx_smpl_hdr = NULL;
1521 ctx->ctx_smpl_size = 0UL;
1522
1523 return 0;
1524
1525 invalid_free:
1526 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1527 return -EINVAL;
1528 }
1529 #endif
1530
1531 static inline void
1532 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1533 {
1534 if (fmt == NULL) return;
1535
1536 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1537
1538 }
1539
1540 /*
1541 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1542 * no real gain from having the whole whorehouse mounted. So we don't need
1543 * any operations on the root directory. However, we need a non-trivial
1544 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1545 */
1546 static struct vfsmount *pfmfs_mnt;
1547
1548 static int __init
1549 init_pfm_fs(void)
1550 {
1551 int err = register_filesystem(&pfm_fs_type);
1552 if (!err) {
1553 pfmfs_mnt = kern_mount(&pfm_fs_type);
1554 err = PTR_ERR(pfmfs_mnt);
1555 if (IS_ERR(pfmfs_mnt))
1556 unregister_filesystem(&pfm_fs_type);
1557 else
1558 err = 0;
1559 }
1560 return err;
1561 }
1562
1563 static ssize_t
1564 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1565 {
1566 pfm_context_t *ctx;
1567 pfm_msg_t *msg;
1568 ssize_t ret;
1569 unsigned long flags;
1570 DECLARE_WAITQUEUE(wait, current);
1571 if (PFM_IS_FILE(filp) == 0) {
1572 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1573 return -EINVAL;
1574 }
1575
1576 ctx = (pfm_context_t *)filp->private_data;
1577 if (ctx == NULL) {
1578 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1579 return -EINVAL;
1580 }
1581
1582 /*
1583 * check even when there is no message
1584 */
1585 if (size < sizeof(pfm_msg_t)) {
1586 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1587 return -EINVAL;
1588 }
1589
1590 PROTECT_CTX(ctx, flags);
1591
1592 /*
1593 * put ourselves on the wait queue
1594 */
1595 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1596
1597
1598 for(;;) {
1599 /*
1600 * check wait queue
1601 */
1602
1603 set_current_state(TASK_INTERRUPTIBLE);
1604
1605 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1606
1607 ret = 0;
1608 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1609
1610 UNPROTECT_CTX(ctx, flags);
1611
1612 /*
1613 * check non-blocking read
1614 */
1615 ret = -EAGAIN;
1616 if(filp->f_flags & O_NONBLOCK) break;
1617
1618 /*
1619 * check pending signals
1620 */
1621 if(signal_pending(current)) {
1622 ret = -EINTR;
1623 break;
1624 }
1625 /*
1626 * no message, so wait
1627 */
1628 schedule();
1629
1630 PROTECT_CTX(ctx, flags);
1631 }
1632 DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1633 set_current_state(TASK_RUNNING);
1634 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1635
1636 if (ret < 0) goto abort;
1637
1638 ret = -EINVAL;
1639 msg = pfm_get_next_msg(ctx);
1640 if (msg == NULL) {
1641 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1642 goto abort_locked;
1643 }
1644
1645 DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1646
1647 ret = -EFAULT;
1648 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1649
1650 abort_locked:
1651 UNPROTECT_CTX(ctx, flags);
1652 abort:
1653 return ret;
1654 }
1655
1656 static ssize_t
1657 pfm_write(struct file *file, const char __user *ubuf,
1658 size_t size, loff_t *ppos)
1659 {
1660 DPRINT(("pfm_write called\n"));
1661 return -EINVAL;
1662 }
1663
1664 static unsigned int
1665 pfm_poll(struct file *filp, poll_table * wait)
1666 {
1667 pfm_context_t *ctx;
1668 unsigned long flags;
1669 unsigned int mask = 0;
1670
1671 if (PFM_IS_FILE(filp) == 0) {
1672 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1673 return 0;
1674 }
1675
1676 ctx = (pfm_context_t *)filp->private_data;
1677 if (ctx == NULL) {
1678 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1679 return 0;
1680 }
1681
1682
1683 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1684
1685 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1686
1687 PROTECT_CTX(ctx, flags);
1688
1689 if (PFM_CTXQ_EMPTY(ctx) == 0)
1690 mask = POLLIN | POLLRDNORM;
1691
1692 UNPROTECT_CTX(ctx, flags);
1693
1694 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1695
1696 return mask;
1697 }
1698
1699 static long
1700 pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1701 {
1702 DPRINT(("pfm_ioctl called\n"));
1703 return -EINVAL;
1704 }
1705
1706 /*
1707 * interrupt cannot be masked when coming here
1708 */
1709 static inline int
1710 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1711 {
1712 int ret;
1713
1714 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1715
1716 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1717 task_pid_nr(current),
1718 fd,
1719 on,
1720 ctx->ctx_async_queue, ret));
1721
1722 return ret;
1723 }
1724
1725 static int
1726 pfm_fasync(int fd, struct file *filp, int on)
1727 {
1728 pfm_context_t *ctx;
1729 int ret;
1730
1731 if (PFM_IS_FILE(filp) == 0) {
1732 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1733 return -EBADF;
1734 }
1735
1736 ctx = (pfm_context_t *)filp->private_data;
1737 if (ctx == NULL) {
1738 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1739 return -EBADF;
1740 }
1741 /*
1742 * we cannot mask interrupts during this call because this may
1743 * may go to sleep if memory is not readily avalaible.
1744 *
1745 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1746 * done in caller. Serialization of this function is ensured by caller.
1747 */
1748 ret = pfm_do_fasync(fd, filp, ctx, on);
1749
1750
1751 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1752 fd,
1753 on,
1754 ctx->ctx_async_queue, ret));
1755
1756 return ret;
1757 }
1758
1759 #ifdef CONFIG_SMP
1760 /*
1761 * this function is exclusively called from pfm_close().
1762 * The context is not protected at that time, nor are interrupts
1763 * on the remote CPU. That's necessary to avoid deadlocks.
1764 */
1765 static void
1766 pfm_syswide_force_stop(void *info)
1767 {
1768 pfm_context_t *ctx = (pfm_context_t *)info;
1769 struct pt_regs *regs = task_pt_regs(current);
1770 struct task_struct *owner;
1771 unsigned long flags;
1772 int ret;
1773
1774 if (ctx->ctx_cpu != smp_processor_id()) {
1775 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1776 ctx->ctx_cpu,
1777 smp_processor_id());
1778 return;
1779 }
1780 owner = GET_PMU_OWNER();
1781 if (owner != ctx->ctx_task) {
1782 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1783 smp_processor_id(),
1784 task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1785 return;
1786 }
1787 if (GET_PMU_CTX() != ctx) {
1788 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1789 smp_processor_id(),
1790 GET_PMU_CTX(), ctx);
1791 return;
1792 }
1793
1794 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1795 /*
1796 * the context is already protected in pfm_close(), we simply
1797 * need to mask interrupts to avoid a PMU interrupt race on
1798 * this CPU
1799 */
1800 local_irq_save(flags);
1801
1802 ret = pfm_context_unload(ctx, NULL, 0, regs);
1803 if (ret) {
1804 DPRINT(("context_unload returned %d\n", ret));
1805 }
1806
1807 /*
1808 * unmask interrupts, PMU interrupts are now spurious here
1809 */
1810 local_irq_restore(flags);
1811 }
1812
1813 static void
1814 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1815 {
1816 int ret;
1817
1818 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1819 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1820 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1821 }
1822 #endif /* CONFIG_SMP */
1823
1824 /*
1825 * called for each close(). Partially free resources.
1826 * When caller is self-monitoring, the context is unloaded.
1827 */
1828 static int
1829 pfm_flush(struct file *filp, fl_owner_t id)
1830 {
1831 pfm_context_t *ctx;
1832 struct task_struct *task;
1833 struct pt_regs *regs;
1834 unsigned long flags;
1835 unsigned long smpl_buf_size = 0UL;
1836 void *smpl_buf_vaddr = NULL;
1837 int state, is_system;
1838
1839 if (PFM_IS_FILE(filp) == 0) {
1840 DPRINT(("bad magic for\n"));
1841 return -EBADF;
1842 }
1843
1844 ctx = (pfm_context_t *)filp->private_data;
1845 if (ctx == NULL) {
1846 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1847 return -EBADF;
1848 }
1849
1850 /*
1851 * remove our file from the async queue, if we use this mode.
1852 * This can be done without the context being protected. We come
1853 * here when the context has become unreachable by other tasks.
1854 *
1855 * We may still have active monitoring at this point and we may
1856 * end up in pfm_overflow_handler(). However, fasync_helper()
1857 * operates with interrupts disabled and it cleans up the
1858 * queue. If the PMU handler is called prior to entering
1859 * fasync_helper() then it will send a signal. If it is
1860 * invoked after, it will find an empty queue and no
1861 * signal will be sent. In both case, we are safe
1862 */
1863 PROTECT_CTX(ctx, flags);
1864
1865 state = ctx->ctx_state;
1866 is_system = ctx->ctx_fl_system;
1867
1868 task = PFM_CTX_TASK(ctx);
1869 regs = task_pt_regs(task);
1870
1871 DPRINT(("ctx_state=%d is_current=%d\n",
1872 state,
1873 task == current ? 1 : 0));
1874
1875 /*
1876 * if state == UNLOADED, then task is NULL
1877 */
1878
1879 /*
1880 * we must stop and unload because we are losing access to the context.
1881 */
1882 if (task == current) {
1883 #ifdef CONFIG_SMP
1884 /*
1885 * the task IS the owner but it migrated to another CPU: that's bad
1886 * but we must handle this cleanly. Unfortunately, the kernel does
1887 * not provide a mechanism to block migration (while the context is loaded).
1888 *
1889 * We need to release the resource on the ORIGINAL cpu.
1890 */
1891 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1892
1893 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1894 /*
1895 * keep context protected but unmask interrupt for IPI
1896 */
1897 local_irq_restore(flags);
1898
1899 pfm_syswide_cleanup_other_cpu(ctx);
1900
1901 /*
1902 * restore interrupt masking
1903 */
1904 local_irq_save(flags);
1905
1906 /*
1907 * context is unloaded at this point
1908 */
1909 } else
1910 #endif /* CONFIG_SMP */
1911 {
1912
1913 DPRINT(("forcing unload\n"));
1914 /*
1915 * stop and unload, returning with state UNLOADED
1916 * and session unreserved.
1917 */
1918 pfm_context_unload(ctx, NULL, 0, regs);
1919
1920 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1921 }
1922 }
1923
1924 /*
1925 * remove virtual mapping, if any, for the calling task.
1926 * cannot reset ctx field until last user is calling close().
1927 *
1928 * ctx_smpl_vaddr must never be cleared because it is needed
1929 * by every task with access to the context
1930 *
1931 * When called from do_exit(), the mm context is gone already, therefore
1932 * mm is NULL, i.e., the VMA is already gone and we do not have to
1933 * do anything here
1934 */
1935 if (ctx->ctx_smpl_vaddr && current->mm) {
1936 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1937 smpl_buf_size = ctx->ctx_smpl_size;
1938 }
1939
1940 UNPROTECT_CTX(ctx, flags);
1941
1942 /*
1943 * if there was a mapping, then we systematically remove it
1944 * at this point. Cannot be done inside critical section
1945 * because some VM function reenables interrupts.
1946 *
1947 */
1948 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1949
1950 return 0;
1951 }
1952 /*
1953 * called either on explicit close() or from exit_files().
1954 * Only the LAST user of the file gets to this point, i.e., it is
1955 * called only ONCE.
1956 *
1957 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1958 * (fput()),i.e, last task to access the file. Nobody else can access the
1959 * file at this point.
1960 *
1961 * When called from exit_files(), the VMA has been freed because exit_mm()
1962 * is executed before exit_files().
1963 *
1964 * When called from exit_files(), the current task is not yet ZOMBIE but we
1965 * flush the PMU state to the context.
1966 */
1967 static int
1968 pfm_close(struct inode *inode, struct file *filp)
1969 {
1970 pfm_context_t *ctx;
1971 struct task_struct *task;
1972 struct pt_regs *regs;
1973 DECLARE_WAITQUEUE(wait, current);
1974 unsigned long flags;
1975 unsigned long smpl_buf_size = 0UL;
1976 void *smpl_buf_addr = NULL;
1977 int free_possible = 1;
1978 int state, is_system;
1979
1980 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1981
1982 if (PFM_IS_FILE(filp) == 0) {
1983 DPRINT(("bad magic\n"));
1984 return -EBADF;
1985 }
1986
1987 ctx = (pfm_context_t *)filp->private_data;
1988 if (ctx == NULL) {
1989 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1990 return -EBADF;
1991 }
1992
1993 PROTECT_CTX(ctx, flags);
1994
1995 state = ctx->ctx_state;
1996 is_system = ctx->ctx_fl_system;
1997
1998 task = PFM_CTX_TASK(ctx);
1999 regs = task_pt_regs(task);
2000
2001 DPRINT(("ctx_state=%d is_current=%d\n",
2002 state,
2003 task == current ? 1 : 0));
2004
2005 /*
2006 * if task == current, then pfm_flush() unloaded the context
2007 */
2008 if (state == PFM_CTX_UNLOADED) goto doit;
2009
2010 /*
2011 * context is loaded/masked and task != current, we need to
2012 * either force an unload or go zombie
2013 */
2014
2015 /*
2016 * The task is currently blocked or will block after an overflow.
2017 * we must force it to wakeup to get out of the
2018 * MASKED state and transition to the unloaded state by itself.
2019 *
2020 * This situation is only possible for per-task mode
2021 */
2022 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2023
2024 /*
2025 * set a "partial" zombie state to be checked
2026 * upon return from down() in pfm_handle_work().
2027 *
2028 * We cannot use the ZOMBIE state, because it is checked
2029 * by pfm_load_regs() which is called upon wakeup from down().
2030 * In such case, it would free the context and then we would
2031 * return to pfm_handle_work() which would access the
2032 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2033 * but visible to pfm_handle_work().
2034 *
2035 * For some window of time, we have a zombie context with
2036 * ctx_state = MASKED and not ZOMBIE
2037 */
2038 ctx->ctx_fl_going_zombie = 1;
2039
2040 /*
2041 * force task to wake up from MASKED state
2042 */
2043 complete(&ctx->ctx_restart_done);
2044
2045 DPRINT(("waking up ctx_state=%d\n", state));
2046
2047 /*
2048 * put ourself to sleep waiting for the other
2049 * task to report completion
2050 *
2051 * the context is protected by mutex, therefore there
2052 * is no risk of being notified of completion before
2053 * begin actually on the waitq.
2054 */
2055 set_current_state(TASK_INTERRUPTIBLE);
2056 add_wait_queue(&ctx->ctx_zombieq, &wait);
2057
2058 UNPROTECT_CTX(ctx, flags);
2059
2060 /*
2061 * XXX: check for signals :
2062 * - ok for explicit close
2063 * - not ok when coming from exit_files()
2064 */
2065 schedule();
2066
2067
2068 PROTECT_CTX(ctx, flags);
2069
2070
2071 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2072 set_current_state(TASK_RUNNING);
2073
2074 /*
2075 * context is unloaded at this point
2076 */
2077 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2078 }
2079 else if (task != current) {
2080 #ifdef CONFIG_SMP
2081 /*
2082 * switch context to zombie state
2083 */
2084 ctx->ctx_state = PFM_CTX_ZOMBIE;
2085
2086 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2087 /*
2088 * cannot free the context on the spot. deferred until
2089 * the task notices the ZOMBIE state
2090 */
2091 free_possible = 0;
2092 #else
2093 pfm_context_unload(ctx, NULL, 0, regs);
2094 #endif
2095 }
2096
2097 doit:
2098 /* reload state, may have changed during opening of critical section */
2099 state = ctx->ctx_state;
2100
2101 /*
2102 * the context is still attached to a task (possibly current)
2103 * we cannot destroy it right now
2104 */
2105
2106 /*
2107 * we must free the sampling buffer right here because
2108 * we cannot rely on it being cleaned up later by the
2109 * monitored task. It is not possible to free vmalloc'ed
2110 * memory in pfm_load_regs(). Instead, we remove the buffer
2111 * now. should there be subsequent PMU overflow originally
2112 * meant for sampling, the will be converted to spurious
2113 * and that's fine because the monitoring tools is gone anyway.
2114 */
2115 if (ctx->ctx_smpl_hdr) {
2116 smpl_buf_addr = ctx->ctx_smpl_hdr;
2117 smpl_buf_size = ctx->ctx_smpl_size;
2118 /* no more sampling */
2119 ctx->ctx_smpl_hdr = NULL;
2120 ctx->ctx_fl_is_sampling = 0;
2121 }
2122
2123 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2124 state,
2125 free_possible,
2126 smpl_buf_addr,
2127 smpl_buf_size));
2128
2129 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2130
2131 /*
2132 * UNLOADED that the session has already been unreserved.
2133 */
2134 if (state == PFM_CTX_ZOMBIE) {
2135 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2136 }
2137
2138 /*
2139 * disconnect file descriptor from context must be done
2140 * before we unlock.
2141 */
2142 filp->private_data = NULL;
2143
2144 /*
2145 * if we free on the spot, the context is now completely unreachable
2146 * from the callers side. The monitored task side is also cut, so we
2147 * can freely cut.
2148 *
2149 * If we have a deferred free, only the caller side is disconnected.
2150 */
2151 UNPROTECT_CTX(ctx, flags);
2152
2153 /*
2154 * All memory free operations (especially for vmalloc'ed memory)
2155 * MUST be done with interrupts ENABLED.
2156 */
2157 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2158
2159 /*
2160 * return the memory used by the context
2161 */
2162 if (free_possible) pfm_context_free(ctx);
2163
2164 return 0;
2165 }
2166
2167 static int
2168 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2169 {
2170 DPRINT(("pfm_no_open called\n"));
2171 return -ENXIO;
2172 }
2173
2174
2175
2176 static const struct file_operations pfm_file_ops = {
2177 .llseek = no_llseek,
2178 .read = pfm_read,
2179 .write = pfm_write,
2180 .poll = pfm_poll,
2181 .unlocked_ioctl = pfm_ioctl,
2182 .open = pfm_no_open, /* special open code to disallow open via /proc */
2183 .fasync = pfm_fasync,
2184 .release = pfm_close,
2185 .flush = pfm_flush
2186 };
2187
2188 static int
2189 pfmfs_delete_dentry(struct dentry *dentry)
2190 {
2191 return 1;
2192 }
2193
2194 static const struct dentry_operations pfmfs_dentry_operations = {
2195 .d_delete = pfmfs_delete_dentry,
2196 };
2197
2198
2199 static struct file *
2200 pfm_alloc_file(pfm_context_t *ctx)
2201 {
2202 struct file *file;
2203 struct inode *inode;
2204 struct path path;
2205 char name[32];
2206 struct qstr this;
2207
2208 /*
2209 * allocate a new inode
2210 */
2211 inode = new_inode(pfmfs_mnt->mnt_sb);
2212 if (!inode)
2213 return ERR_PTR(-ENOMEM);
2214
2215 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2216
2217 inode->i_mode = S_IFCHR|S_IRUGO;
2218 inode->i_uid = current_fsuid();
2219 inode->i_gid = current_fsgid();
2220
2221 sprintf(name, "[%lu]", inode->i_ino);
2222 this.name = name;
2223 this.len = strlen(name);
2224 this.hash = inode->i_ino;
2225
2226 /*
2227 * allocate a new dcache entry
2228 */
2229 path.dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2230 if (!path.dentry) {
2231 iput(inode);
2232 return ERR_PTR(-ENOMEM);
2233 }
2234 path.mnt = mntget(pfmfs_mnt);
2235
2236 path.dentry->d_op = &pfmfs_dentry_operations;
2237 d_add(path.dentry, inode);
2238
2239 file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2240 if (!file) {
2241 path_put(&path);
2242 return ERR_PTR(-ENFILE);
2243 }
2244
2245 file->f_flags = O_RDONLY;
2246 file->private_data = ctx;
2247
2248 return file;
2249 }
2250
2251 static int
2252 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2253 {
2254 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2255
2256 while (size > 0) {
2257 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2258
2259
2260 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2261 return -ENOMEM;
2262
2263 addr += PAGE_SIZE;
2264 buf += PAGE_SIZE;
2265 size -= PAGE_SIZE;
2266 }
2267 return 0;
2268 }
2269
2270 /*
2271 * allocate a sampling buffer and remaps it into the user address space of the task
2272 */
2273 static int
2274 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2275 {
2276 struct mm_struct *mm = task->mm;
2277 struct vm_area_struct *vma = NULL;
2278 unsigned long size;
2279 void *smpl_buf;
2280
2281
2282 /*
2283 * the fixed header + requested size and align to page boundary
2284 */
2285 size = PAGE_ALIGN(rsize);
2286
2287 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2288
2289 /*
2290 * check requested size to avoid Denial-of-service attacks
2291 * XXX: may have to refine this test
2292 * Check against address space limit.
2293 *
2294 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2295 * return -ENOMEM;
2296 */
2297 if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2298 return -ENOMEM;
2299
2300 /*
2301 * We do the easy to undo allocations first.
2302 *
2303 * pfm_rvmalloc(), clears the buffer, so there is no leak
2304 */
2305 smpl_buf = pfm_rvmalloc(size);
2306 if (smpl_buf == NULL) {
2307 DPRINT(("Can't allocate sampling buffer\n"));
2308 return -ENOMEM;
2309 }
2310
2311 DPRINT(("smpl_buf @%p\n", smpl_buf));
2312
2313 /* allocate vma */
2314 vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2315 if (!vma) {
2316 DPRINT(("Cannot allocate vma\n"));
2317 goto error_kmem;
2318 }
2319 INIT_LIST_HEAD(&vma->anon_vma_chain);
2320
2321 /*
2322 * partially initialize the vma for the sampling buffer
2323 */
2324 vma->vm_mm = mm;
2325 vma->vm_file = filp;
2326 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2327 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2328
2329 /*
2330 * Now we have everything we need and we can initialize
2331 * and connect all the data structures
2332 */
2333
2334 ctx->ctx_smpl_hdr = smpl_buf;
2335 ctx->ctx_smpl_size = size; /* aligned size */
2336
2337 /*
2338 * Let's do the difficult operations next.
2339 *
2340 * now we atomically find some area in the address space and
2341 * remap the buffer in it.
2342 */
2343 down_write(&task->mm->mmap_sem);
2344
2345 /* find some free area in address space, must have mmap sem held */
2346 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2347 if (vma->vm_start == 0UL) {
2348 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2349 up_write(&task->mm->mmap_sem);
2350 goto error;
2351 }
2352 vma->vm_end = vma->vm_start + size;
2353 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2354
2355 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2356
2357 /* can only be applied to current task, need to have the mm semaphore held when called */
2358 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2359 DPRINT(("Can't remap buffer\n"));
2360 up_write(&task->mm->mmap_sem);
2361 goto error;
2362 }
2363
2364 get_file(filp);
2365
2366 /*
2367 * now insert the vma in the vm list for the process, must be
2368 * done with mmap lock held
2369 */
2370 insert_vm_struct(mm, vma);
2371
2372 mm->total_vm += size >> PAGE_SHIFT;
2373 vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2374 vma_pages(vma));
2375 up_write(&task->mm->mmap_sem);
2376
2377 /*
2378 * keep track of user level virtual address
2379 */
2380 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2381 *(unsigned long *)user_vaddr = vma->vm_start;
2382
2383 return 0;
2384
2385 error:
2386 kmem_cache_free(vm_area_cachep, vma);
2387 error_kmem:
2388 pfm_rvfree(smpl_buf, size);
2389
2390 return -ENOMEM;
2391 }
2392
2393 /*
2394 * XXX: do something better here
2395 */
2396 static int
2397 pfm_bad_permissions(struct task_struct *task)
2398 {
2399 const struct cred *tcred;
2400 uid_t uid = current_uid();
2401 gid_t gid = current_gid();
2402 int ret;
2403
2404 rcu_read_lock();
2405 tcred = __task_cred(task);
2406
2407 /* inspired by ptrace_attach() */
2408 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2409 uid,
2410 gid,
2411 tcred->euid,
2412 tcred->suid,
2413 tcred->uid,
2414 tcred->egid,
2415 tcred->sgid));
2416
2417 ret = ((uid != tcred->euid)
2418 || (uid != tcred->suid)
2419 || (uid != tcred->uid)
2420 || (gid != tcred->egid)
2421 || (gid != tcred->sgid)
2422 || (gid != tcred->gid)) && !capable(CAP_SYS_PTRACE);
2423
2424 rcu_read_unlock();
2425 return ret;
2426 }
2427
2428 static int
2429 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2430 {
2431 int ctx_flags;
2432
2433 /* valid signal */
2434
2435 ctx_flags = pfx->ctx_flags;
2436
2437 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2438
2439 /*
2440 * cannot block in this mode
2441 */
2442 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2443 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2444 return -EINVAL;
2445 }
2446 } else {
2447 }
2448 /* probably more to add here */
2449
2450 return 0;
2451 }
2452
2453 static int
2454 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2455 unsigned int cpu, pfarg_context_t *arg)
2456 {
2457 pfm_buffer_fmt_t *fmt = NULL;
2458 unsigned long size = 0UL;
2459 void *uaddr = NULL;
2460 void *fmt_arg = NULL;
2461 int ret = 0;
2462 #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2463
2464 /* invoke and lock buffer format, if found */
2465 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2466 if (fmt == NULL) {
2467 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2468 return -EINVAL;
2469 }
2470
2471 /*
2472 * buffer argument MUST be contiguous to pfarg_context_t
2473 */
2474 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2475
2476 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2477
2478 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2479
2480 if (ret) goto error;
2481
2482 /* link buffer format and context */
2483 ctx->ctx_buf_fmt = fmt;
2484 ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2485
2486 /*
2487 * check if buffer format wants to use perfmon buffer allocation/mapping service
2488 */
2489 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2490 if (ret) goto error;
2491
2492 if (size) {
2493 /*
2494 * buffer is always remapped into the caller's address space
2495 */
2496 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2497 if (ret) goto error;
2498
2499 /* keep track of user address of buffer */
2500 arg->ctx_smpl_vaddr = uaddr;
2501 }
2502 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2503
2504 error:
2505 return ret;
2506 }
2507
2508 static void
2509 pfm_reset_pmu_state(pfm_context_t *ctx)
2510 {
2511 int i;
2512
2513 /*
2514 * install reset values for PMC.
2515 */
2516 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2517 if (PMC_IS_IMPL(i) == 0) continue;
2518 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2519 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2520 }
2521 /*
2522 * PMD registers are set to 0UL when the context in memset()
2523 */
2524
2525 /*
2526 * On context switched restore, we must restore ALL pmc and ALL pmd even
2527 * when they are not actively used by the task. In UP, the incoming process
2528 * may otherwise pick up left over PMC, PMD state from the previous process.
2529 * As opposed to PMD, stale PMC can cause harm to the incoming
2530 * process because they may change what is being measured.
2531 * Therefore, we must systematically reinstall the entire
2532 * PMC state. In SMP, the same thing is possible on the
2533 * same CPU but also on between 2 CPUs.
2534 *
2535 * The problem with PMD is information leaking especially
2536 * to user level when psr.sp=0
2537 *
2538 * There is unfortunately no easy way to avoid this problem
2539 * on either UP or SMP. This definitively slows down the
2540 * pfm_load_regs() function.
2541 */
2542
2543 /*
2544 * bitmask of all PMCs accessible to this context
2545 *
2546 * PMC0 is treated differently.
2547 */
2548 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2549
2550 /*
2551 * bitmask of all PMDs that are accessible to this context
2552 */
2553 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2554
2555 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2556
2557 /*
2558 * useful in case of re-enable after disable
2559 */
2560 ctx->ctx_used_ibrs[0] = 0UL;
2561 ctx->ctx_used_dbrs[0] = 0UL;
2562 }
2563
2564 static int
2565 pfm_ctx_getsize(void *arg, size_t *sz)
2566 {
2567 pfarg_context_t *req = (pfarg_context_t *)arg;
2568 pfm_buffer_fmt_t *fmt;
2569
2570 *sz = 0;
2571
2572 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2573
2574 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2575 if (fmt == NULL) {
2576 DPRINT(("cannot find buffer format\n"));
2577 return -EINVAL;
2578 }
2579 /* get just enough to copy in user parameters */
2580 *sz = fmt->fmt_arg_size;
2581 DPRINT(("arg_size=%lu\n", *sz));
2582
2583 return 0;
2584 }
2585
2586
2587
2588 /*
2589 * cannot attach if :
2590 * - kernel task
2591 * - task not owned by caller
2592 * - task incompatible with context mode
2593 */
2594 static int
2595 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2596 {
2597 /*
2598 * no kernel task or task not owner by caller
2599 */
2600 if (task->mm == NULL) {
2601 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2602 return -EPERM;
2603 }
2604 if (pfm_bad_permissions(task)) {
2605 DPRINT(("no permission to attach to [%d]\n", task_pid_nr(task)));
2606 return -EPERM;
2607 }
2608 /*
2609 * cannot block in self-monitoring mode
2610 */
2611 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2612 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2613 return -EINVAL;
2614 }
2615
2616 if (task->exit_state == EXIT_ZOMBIE) {
2617 DPRINT(("cannot attach to zombie task [%d]\n", task_pid_nr(task)));
2618 return -EBUSY;
2619 }
2620
2621 /*
2622 * always ok for self
2623 */
2624 if (task == current) return 0;
2625
2626 if (!task_is_stopped_or_traced(task)) {
2627 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2628 return -EBUSY;
2629 }
2630 /*
2631 * make sure the task is off any CPU
2632 */
2633 wait_task_inactive(task, 0);
2634
2635 /* more to come... */
2636
2637 return 0;
2638 }
2639
2640 static int
2641 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2642 {
2643 struct task_struct *p = current;
2644 int ret;
2645
2646 /* XXX: need to add more checks here */
2647 if (pid < 2) return -EPERM;
2648
2649 if (pid != task_pid_vnr(current)) {
2650
2651 read_lock(&tasklist_lock);
2652
2653 p = find_task_by_vpid(pid);
2654
2655 /* make sure task cannot go away while we operate on it */
2656 if (p) get_task_struct(p);
2657
2658 read_unlock(&tasklist_lock);
2659
2660 if (p == NULL) return -ESRCH;
2661 }
2662
2663 ret = pfm_task_incompatible(ctx, p);
2664 if (ret == 0) {
2665 *task = p;
2666 } else if (p != current) {
2667 pfm_put_task(p);
2668 }
2669 return ret;
2670 }
2671
2672
2673
2674 static int
2675 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2676 {
2677 pfarg_context_t *req = (pfarg_context_t *)arg;
2678 struct file *filp;
2679 struct path path;
2680 int ctx_flags;
2681 int fd;
2682 int ret;
2683
2684 /* let's check the arguments first */
2685 ret = pfarg_is_sane(current, req);
2686 if (ret < 0)
2687 return ret;
2688
2689 ctx_flags = req->ctx_flags;
2690
2691 ret = -ENOMEM;
2692
2693 fd = get_unused_fd();
2694 if (fd < 0)
2695 return fd;
2696
2697 ctx = pfm_context_alloc(ctx_flags);
2698 if (!ctx)
2699 goto error;
2700
2701 filp = pfm_alloc_file(ctx);
2702 if (IS_ERR(filp)) {
2703 ret = PTR_ERR(filp);
2704 goto error_file;
2705 }
2706
2707 req->ctx_fd = ctx->ctx_fd = fd;
2708
2709 /*
2710 * does the user want to sample?
2711 */
2712 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2713 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2714 if (ret)
2715 goto buffer_error;
2716 }
2717
2718 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2719 ctx,
2720 ctx_flags,
2721 ctx->ctx_fl_system,
2722 ctx->ctx_fl_block,
2723 ctx->ctx_fl_excl_idle,
2724 ctx->ctx_fl_no_msg,
2725 ctx->ctx_fd));
2726
2727 /*
2728 * initialize soft PMU state
2729 */
2730 pfm_reset_pmu_state(ctx);
2731
2732 fd_install(fd, filp);
2733
2734 return 0;
2735
2736 buffer_error:
2737 path = filp->f_path;
2738 put_filp(filp);
2739 path_put(&path);
2740
2741 if (ctx->ctx_buf_fmt) {
2742 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2743 }
2744 error_file:
2745 pfm_context_free(ctx);
2746
2747 error:
2748 put_unused_fd(fd);
2749 return ret;
2750 }
2751
2752 static inline unsigned long
2753 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2754 {
2755 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2756 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2757 extern unsigned long carta_random32 (unsigned long seed);
2758
2759 if (reg->flags & PFM_REGFL_RANDOM) {
2760 new_seed = carta_random32(old_seed);
2761 val -= (old_seed & mask); /* counter values are negative numbers! */
2762 if ((mask >> 32) != 0)
2763 /* construct a full 64-bit random value: */
2764 new_seed |= carta_random32(old_seed >> 32) << 32;
2765 reg->seed = new_seed;
2766 }
2767 reg->lval = val;
2768 return val;
2769 }
2770
2771 static void
2772 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2773 {
2774 unsigned long mask = ovfl_regs[0];
2775 unsigned long reset_others = 0UL;
2776 unsigned long val;
2777 int i;
2778
2779 /*
2780 * now restore reset value on sampling overflowed counters
2781 */
2782 mask >>= PMU_FIRST_COUNTER;
2783 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2784
2785 if ((mask & 0x1UL) == 0UL) continue;
2786
2787 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2788 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2789
2790 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2791 }
2792
2793 /*
2794 * Now take care of resetting the other registers
2795 */
2796 for(i = 0; reset_others; i++, reset_others >>= 1) {
2797
2798 if ((reset_others & 0x1) == 0) continue;
2799
2800 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2801
2802 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2803 is_long_reset ? "long" : "short", i, val));
2804 }
2805 }
2806
2807 static void
2808 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2809 {
2810 unsigned long mask = ovfl_regs[0];
2811 unsigned long reset_others = 0UL;
2812 unsigned long val;
2813 int i;
2814
2815 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2816
2817 if (ctx->ctx_state == PFM_CTX_MASKED) {
2818 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2819 return;
2820 }
2821
2822 /*
2823 * now restore reset value on sampling overflowed counters
2824 */
2825 mask >>= PMU_FIRST_COUNTER;
2826 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2827
2828 if ((mask & 0x1UL) == 0UL) continue;
2829
2830 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2831 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2832
2833 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2834
2835 pfm_write_soft_counter(ctx, i, val);
2836 }
2837
2838 /*
2839 * Now take care of resetting the other registers
2840 */
2841 for(i = 0; reset_others; i++, reset_others >>= 1) {
2842
2843 if ((reset_others & 0x1) == 0) continue;
2844
2845 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2846
2847 if (PMD_IS_COUNTING(i)) {
2848 pfm_write_soft_counter(ctx, i, val);
2849 } else {
2850 ia64_set_pmd(i, val);
2851 }
2852 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2853 is_long_reset ? "long" : "short", i, val));
2854 }
2855 ia64_srlz_d();
2856 }
2857
2858 static int
2859 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2860 {
2861 struct task_struct *task;
2862 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2863 unsigned long value, pmc_pm;
2864 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2865 unsigned int cnum, reg_flags, flags, pmc_type;
2866 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2867 int is_monitor, is_counting, state;
2868 int ret = -EINVAL;
2869 pfm_reg_check_t wr_func;
2870 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2871
2872 state = ctx->ctx_state;
2873 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2874 is_system = ctx->ctx_fl_system;
2875 task = ctx->ctx_task;
2876 impl_pmds = pmu_conf->impl_pmds[0];
2877
2878 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2879
2880 if (is_loaded) {
2881 /*
2882 * In system wide and when the context is loaded, access can only happen
2883 * when the caller is running on the CPU being monitored by the session.
2884 * It does not have to be the owner (ctx_task) of the context per se.
2885 */
2886 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2887 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2888 return -EBUSY;
2889 }
2890 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2891 }
2892 expert_mode = pfm_sysctl.expert_mode;
2893
2894 for (i = 0; i < count; i++, req++) {
2895
2896 cnum = req->reg_num;
2897 reg_flags = req->reg_flags;
2898 value = req->reg_value;
2899 smpl_pmds = req->reg_smpl_pmds[0];
2900 reset_pmds = req->reg_reset_pmds[0];
2901 flags = 0;
2902
2903
2904 if (cnum >= PMU_MAX_PMCS) {
2905 DPRINT(("pmc%u is invalid\n", cnum));
2906 goto error;
2907 }
2908
2909 pmc_type = pmu_conf->pmc_desc[cnum].type;
2910 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2911 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2912 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2913
2914 /*
2915 * we reject all non implemented PMC as well
2916 * as attempts to modify PMC[0-3] which are used
2917 * as status registers by the PMU
2918 */
2919 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2920 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2921 goto error;
2922 }
2923 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2924 /*
2925 * If the PMC is a monitor, then if the value is not the default:
2926 * - system-wide session: PMCx.pm=1 (privileged monitor)
2927 * - per-task : PMCx.pm=0 (user monitor)
2928 */
2929 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2930 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2931 cnum,
2932 pmc_pm,
2933 is_system));
2934 goto error;
2935 }
2936
2937 if (is_counting) {
2938 /*
2939 * enforce generation of overflow interrupt. Necessary on all
2940 * CPUs.
2941 */
2942 value |= 1 << PMU_PMC_OI;
2943
2944 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2945 flags |= PFM_REGFL_OVFL_NOTIFY;
2946 }
2947
2948 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2949
2950 /* verify validity of smpl_pmds */
2951 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2952 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2953 goto error;
2954 }
2955
2956 /* verify validity of reset_pmds */
2957 if ((reset_pmds & impl_pmds) != reset_pmds) {
2958 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2959 goto error;
2960 }
2961 } else {
2962 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2963 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2964 goto error;
2965 }
2966 /* eventid on non-counting monitors are ignored */
2967 }
2968
2969 /*
2970 * execute write checker, if any
2971 */
2972 if (likely(expert_mode == 0 && wr_func)) {
2973 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2974 if (ret) goto error;
2975 ret = -EINVAL;
2976 }
2977
2978 /*
2979 * no error on this register
2980 */
2981 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2982
2983 /*
2984 * Now we commit the changes to the software state
2985 */
2986
2987 /*
2988 * update overflow information
2989 */
2990 if (is_counting) {
2991 /*
2992 * full flag update each time a register is programmed
2993 */
2994 ctx->ctx_pmds[cnum].flags = flags;
2995
2996 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2997 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2998 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2999
3000 /*
3001 * Mark all PMDS to be accessed as used.
3002 *
3003 * We do not keep track of PMC because we have to
3004 * systematically restore ALL of them.
3005 *
3006 * We do not update the used_monitors mask, because
3007 * if we have not programmed them, then will be in
3008 * a quiescent state, therefore we will not need to
3009 * mask/restore then when context is MASKED.
3010 */
3011 CTX_USED_PMD(ctx, reset_pmds);
3012 CTX_USED_PMD(ctx, smpl_pmds);
3013 /*
3014 * make sure we do not try to reset on
3015 * restart because we have established new values
3016 */
3017 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3018 }
3019 /*
3020 * Needed in case the user does not initialize the equivalent
3021 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3022 * possible leak here.
3023 */
3024 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3025
3026 /*
3027 * keep track of the monitor PMC that we are using.
3028 * we save the value of the pmc in ctx_pmcs[] and if
3029 * the monitoring is not stopped for the context we also
3030 * place it in the saved state area so that it will be
3031 * picked up later by the context switch code.
3032 *
3033 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3034 *
3035 * The value in th_pmcs[] may be modified on overflow, i.e., when
3036 * monitoring needs to be stopped.
3037 */
3038 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3039
3040 /*
3041 * update context state
3042 */
3043 ctx->ctx_pmcs[cnum] = value;
3044
3045 if (is_loaded) {
3046 /*
3047 * write thread state
3048 */
3049 if (is_system == 0) ctx->th_pmcs[cnum] = value;
3050
3051 /*
3052 * write hardware register if we can
3053 */
3054 if (can_access_pmu) {
3055 ia64_set_pmc(cnum, value);
3056 }
3057 #ifdef CONFIG_SMP
3058 else {
3059 /*
3060 * per-task SMP only here
3061 *
3062 * we are guaranteed that the task is not running on the other CPU,
3063 * we indicate that this PMD will need to be reloaded if the task
3064 * is rescheduled on the CPU it ran last on.
3065 */
3066 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3067 }
3068 #endif
3069 }
3070
3071 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",
3072 cnum,
3073 value,
3074 is_loaded,
3075 can_access_pmu,
3076 flags,
3077 ctx->ctx_all_pmcs[0],
3078 ctx->ctx_used_pmds[0],
3079 ctx->ctx_pmds[cnum].eventid,
3080 smpl_pmds,
3081 reset_pmds,
3082 ctx->ctx_reload_pmcs[0],
3083 ctx->ctx_used_monitors[0],
3084 ctx->ctx_ovfl_regs[0]));
3085 }
3086
3087 /*
3088 * make sure the changes are visible
3089 */
3090 if (can_access_pmu) ia64_srlz_d();
3091
3092 return 0;
3093 error:
3094 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3095 return ret;
3096 }
3097
3098 static int
3099 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3100 {
3101 struct task_struct *task;
3102 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3103 unsigned long value, hw_value, ovfl_mask;
3104 unsigned int cnum;
3105 int i, can_access_pmu = 0, state;
3106 int is_counting, is_loaded, is_system, expert_mode;
3107 int ret = -EINVAL;
3108 pfm_reg_check_t wr_func;
3109
3110
3111 state = ctx->ctx_state;
3112 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3113 is_system = ctx->ctx_fl_system;
3114 ovfl_mask = pmu_conf->ovfl_val;
3115 task = ctx->ctx_task;
3116
3117 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3118
3119 /*
3120 * on both UP and SMP, we can only write to the PMC when the task is
3121 * the owner of the local PMU.
3122 */
3123 if (likely(is_loaded)) {
3124 /*
3125 * In system wide and when the context is loaded, access can only happen
3126 * when the caller is running on the CPU being monitored by the session.
3127 * It does not have to be the owner (ctx_task) of the context per se.
3128 */
3129 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3130 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3131 return -EBUSY;
3132 }
3133 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3134 }
3135 expert_mode = pfm_sysctl.expert_mode;
3136
3137 for (i = 0; i < count; i++, req++) {
3138
3139 cnum = req->reg_num;
3140 value = req->reg_value;
3141
3142 if (!PMD_IS_IMPL(cnum)) {
3143 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3144 goto abort_mission;
3145 }
3146 is_counting = PMD_IS_COUNTING(cnum);
3147 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3148
3149 /*
3150 * execute write checker, if any
3151 */
3152 if (unlikely(expert_mode == 0 && wr_func)) {
3153 unsigned long v = value;
3154
3155 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3156 if (ret) goto abort_mission;
3157
3158 value = v;
3159 ret = -EINVAL;
3160 }
3161
3162 /*
3163 * no error on this register
3164 */
3165 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3166
3167 /*
3168 * now commit changes to software state
3169 */
3170 hw_value = value;
3171
3172 /*
3173 * update virtualized (64bits) counter
3174 */
3175 if (is_counting) {
3176 /*
3177 * write context state
3178 */
3179 ctx->ctx_pmds[cnum].lval = value;
3180
3181 /*
3182 * when context is load we use the split value
3183 */
3184 if (is_loaded) {
3185 hw_value = value & ovfl_mask;
3186 value = value & ~ovfl_mask;
3187 }
3188 }
3189 /*
3190 * update reset values (not just for counters)
3191 */
3192 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3193 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3194
3195 /*
3196 * update randomization parameters (not just for counters)
3197 */
3198 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3199 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3200
3201 /*
3202 * update context value
3203 */
3204 ctx->ctx_pmds[cnum].val = value;
3205
3206 /*
3207 * Keep track of what we use
3208 *
3209 * We do not keep track of PMC because we have to
3210 * systematically restore ALL of them.
3211 */
3212 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3213
3214 /*
3215 * mark this PMD register used as well
3216 */
3217 CTX_USED_PMD(ctx, RDEP(cnum));
3218
3219 /*
3220 * make sure we do not try to reset on
3221 * restart because we have established new values
3222 */
3223 if (is_counting && state == PFM_CTX_MASKED) {
3224 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3225 }
3226
3227 if (is_loaded) {
3228 /*
3229 * write thread state
3230 */
3231 if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3232
3233 /*
3234 * write hardware register if we can
3235 */
3236 if (can_access_pmu) {
3237 ia64_set_pmd(cnum, hw_value);
3238 } else {
3239 #ifdef CONFIG_SMP
3240 /*
3241 * we are guaranteed that the task is not running on the other CPU,
3242 * we indicate that this PMD will need to be reloaded if the task
3243 * is rescheduled on the CPU it ran last on.
3244 */
3245 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3246 #endif
3247 }
3248 }
3249
3250 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3251 "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",
3252 cnum,
3253 value,
3254 is_loaded,
3255 can_access_pmu,
3256 hw_value,
3257 ctx->ctx_pmds[cnum].val,
3258 ctx->ctx_pmds[cnum].short_reset,
3259 ctx->ctx_pmds[cnum].long_reset,
3260 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3261 ctx->ctx_pmds[cnum].seed,
3262 ctx->ctx_pmds[cnum].mask,
3263 ctx->ctx_used_pmds[0],
3264 ctx->ctx_pmds[cnum].reset_pmds[0],
3265 ctx->ctx_reload_pmds[0],
3266 ctx->ctx_all_pmds[0],
3267 ctx->ctx_ovfl_regs[0]));
3268 }
3269
3270 /*
3271 * make changes visible
3272 */
3273 if (can_access_pmu) ia64_srlz_d();
3274
3275 return 0;
3276
3277 abort_mission:
3278 /*
3279 * for now, we have only one possibility for error
3280 */
3281 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3282 return ret;
3283 }
3284
3285 /*
3286 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3287 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3288 * interrupt is delivered during the call, it will be kept pending until we leave, making
3289 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3290 * guaranteed to return consistent data to the user, it may simply be old. It is not
3291 * trivial to treat the overflow while inside the call because you may end up in
3292 * some module sampling buffer code causing deadlocks.
3293 */
3294 static int
3295 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3296 {
3297 struct task_struct *task;
3298 unsigned long val = 0UL, lval, ovfl_mask, sval;
3299 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3300 unsigned int cnum, reg_flags = 0;
3301 int i, can_access_pmu = 0, state;
3302 int is_loaded, is_system, is_counting, expert_mode;
3303 int ret = -EINVAL;
3304 pfm_reg_check_t rd_func;
3305
3306 /*
3307 * access is possible when loaded only for
3308 * self-monitoring tasks or in UP mode
3309 */
3310
3311 state = ctx->ctx_state;
3312 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3313 is_system = ctx->ctx_fl_system;
3314 ovfl_mask = pmu_conf->ovfl_val;
3315 task = ctx->ctx_task;
3316
3317 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3318
3319 if (likely(is_loaded)) {
3320 /*
3321 * In system wide and when the context is loaded, access can only happen
3322 * when the caller is running on the CPU being monitored by the session.
3323 * It does not have to be the owner (ctx_task) of the context per se.
3324 */
3325 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3326 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3327 return -EBUSY;
3328 }
3329 /*
3330 * this can be true when not self-monitoring only in UP
3331 */
3332 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3333
3334 if (can_access_pmu) ia64_srlz_d();
3335 }
3336 expert_mode = pfm_sysctl.expert_mode;
3337
3338 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3339 is_loaded,
3340 can_access_pmu,
3341 state));
3342
3343 /*
3344 * on both UP and SMP, we can only read the PMD from the hardware register when
3345 * the task is the owner of the local PMU.
3346 */
3347
3348 for (i = 0; i < count; i++, req++) {
3349
3350 cnum = req->reg_num;
3351 reg_flags = req->reg_flags;
3352
3353 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3354 /*
3355 * we can only read the register that we use. That includes
3356 * the one we explicitly initialize AND the one we want included
3357 * in the sampling buffer (smpl_regs).
3358 *
3359 * Having this restriction allows optimization in the ctxsw routine
3360 * without compromising security (leaks)
3361 */
3362 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3363
3364 sval = ctx->ctx_pmds[cnum].val;
3365 lval = ctx->ctx_pmds[cnum].lval;
3366 is_counting = PMD_IS_COUNTING(cnum);
3367
3368 /*
3369 * If the task is not the current one, then we check if the
3370 * PMU state is still in the local live register due to lazy ctxsw.
3371 * If true, then we read directly from the registers.
3372 */
3373 if (can_access_pmu){
3374 val = ia64_get_pmd(cnum);
3375 } else {
3376 /*
3377 * context has been saved
3378 * if context is zombie, then task does not exist anymore.
3379 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3380 */
3381 val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3382 }
3383 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3384
3385 if (is_counting) {
3386 /*
3387 * XXX: need to check for overflow when loaded
3388 */
3389 val &= ovfl_mask;
3390 val += sval;
3391 }
3392
3393 /*
3394 * execute read checker, if any
3395 */
3396 if (unlikely(expert_mode == 0 && rd_func)) {
3397 unsigned long v = val;
3398 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3399 if (ret) goto error;
3400 val = v;
3401 ret = -EINVAL;
3402 }
3403
3404 PFM_REG_RETFLAG_SET(reg_flags, 0);
3405
3406 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3407
3408 /*
3409 * update register return value, abort all if problem during copy.
3410 * we only modify the reg_flags field. no check mode is fine because
3411 * access has been verified upfront in sys_perfmonctl().
3412 */
3413 req->reg_value = val;
3414 req->reg_flags = reg_flags;
3415 req->reg_last_reset_val = lval;
3416 }
3417
3418 return 0;
3419
3420 error:
3421 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3422 return ret;
3423 }
3424
3425 int
3426 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3427 {
3428 pfm_context_t *ctx;
3429
3430 if (req == NULL) return -EINVAL;
3431
3432 ctx = GET_PMU_CTX();
3433
3434 if (ctx == NULL) return -EINVAL;
3435
3436 /*
3437 * for now limit to current task, which is enough when calling
3438 * from overflow handler
3439 */
3440 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3441
3442 return pfm_write_pmcs(ctx, req, nreq, regs);
3443 }
3444 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3445
3446 int
3447 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3448 {
3449 pfm_context_t *ctx;
3450
3451 if (req == NULL) return -EINVAL;
3452
3453 ctx = GET_PMU_CTX();
3454
3455 if (ctx == NULL) return -EINVAL;
3456
3457 /*
3458 * for now limit to current task, which is enough when calling
3459 * from overflow handler
3460 */
3461 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3462
3463 return pfm_read_pmds(ctx, req, nreq, regs);
3464 }
3465 EXPORT_SYMBOL(pfm_mod_read_pmds);
3466
3467 /*
3468 * Only call this function when a process it trying to
3469 * write the debug registers (reading is always allowed)
3470 */
3471 int
3472 pfm_use_debug_registers(struct task_struct *task)
3473 {
3474 pfm_context_t *ctx = task->thread.pfm_context;
3475 unsigned long flags;
3476 int ret = 0;
3477
3478 if (pmu_conf->use_rr_dbregs == 0) return 0;
3479
3480 DPRINT(("called for [%d]\n", task_pid_nr(task)));
3481
3482 /*
3483 * do it only once
3484 */
3485 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3486
3487 /*
3488 * Even on SMP, we do not need to use an atomic here because
3489 * the only way in is via ptrace() and this is possible only when the
3490 * process is stopped. Even in the case where the ctxsw out is not totally
3491 * completed by the time we come here, there is no way the 'stopped' process
3492 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3493 * So this is always safe.
3494 */
3495 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3496
3497 LOCK_PFS(flags);
3498
3499 /*
3500 * We cannot allow setting breakpoints when system wide monitoring
3501 * sessions are using the debug registers.
3502 */
3503 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3504 ret = -1;
3505 else
3506 pfm_sessions.pfs_ptrace_use_dbregs++;
3507
3508 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3509 pfm_sessions.pfs_ptrace_use_dbregs,
3510 pfm_sessions.pfs_sys_use_dbregs,
3511 task_pid_nr(task), ret));
3512
3513 UNLOCK_PFS(flags);
3514
3515 return ret;
3516 }
3517
3518 /*
3519 * This function is called for every task that exits with the
3520 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3521 * able to use the debug registers for debugging purposes via
3522 * ptrace(). Therefore we know it was not using them for
3523 * performance monitoring, so we only decrement the number
3524 * of "ptraced" debug register users to keep the count up to date
3525 */
3526 int
3527 pfm_release_debug_registers(struct task_struct *task)
3528 {
3529 unsigned long flags;
3530 int ret;
3531
3532 if (pmu_conf->use_rr_dbregs == 0) return 0;
3533
3534 LOCK_PFS(flags);
3535 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3536 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3537 ret = -1;
3538 } else {
3539 pfm_sessions.pfs_ptrace_use_dbregs--;
3540 ret = 0;
3541 }
3542 UNLOCK_PFS(flags);
3543
3544 return ret;
3545 }
3546
3547 static int
3548 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3549 {
3550 struct task_struct *task;
3551 pfm_buffer_fmt_t *fmt;
3552 pfm_ovfl_ctrl_t rst_ctrl;
3553 int state, is_system;
3554 int ret = 0;
3555
3556 state = ctx->ctx_state;
3557 fmt = ctx->ctx_buf_fmt;
3558 is_system = ctx->ctx_fl_system;
3559 task = PFM_CTX_TASK(ctx);
3560
3561 switch(state) {
3562 case PFM_CTX_MASKED:
3563 break;
3564 case PFM_CTX_LOADED:
3565 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3566 /* fall through */
3567 case PFM_CTX_UNLOADED:
3568 case PFM_CTX_ZOMBIE:
3569 DPRINT(("invalid state=%d\n", state));
3570 return -EBUSY;
3571 default:
3572 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3573 return -EINVAL;
3574 }
3575
3576 /*
3577 * In system wide and when the context is loaded, access can only happen
3578 * when the caller is running on the CPU being monitored by the session.
3579 * It does not have to be the owner (ctx_task) of the context per se.
3580 */
3581 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3582 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3583 return -EBUSY;
3584 }
3585
3586 /* sanity check */
3587 if (unlikely(task == NULL)) {
3588 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3589 return -EINVAL;
3590 }
3591
3592 if (task == current || is_system) {
3593
3594 fmt = ctx->ctx_buf_fmt;
3595
3596 DPRINT(("restarting self %d ovfl=0x%lx\n",
3597 task_pid_nr(task),
3598 ctx->ctx_ovfl_regs[0]));
3599
3600 if (CTX_HAS_SMPL(ctx)) {
3601
3602 prefetch(ctx->ctx_smpl_hdr);
3603
3604 rst_ctrl.bits.mask_monitoring = 0;
3605 rst_ctrl.bits.reset_ovfl_pmds = 0;
3606
3607 if (state == PFM_CTX_LOADED)
3608 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3609 else
3610 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3611 } else {
3612 rst_ctrl.bits.mask_monitoring = 0;
3613 rst_ctrl.bits.reset_ovfl_pmds = 1;
3614 }
3615
3616 if (ret == 0) {
3617 if (rst_ctrl.bits.reset_ovfl_pmds)
3618 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3619
3620 if (rst_ctrl.bits.mask_monitoring == 0) {
3621 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3622
3623 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3624 } else {
3625 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3626
3627 // cannot use pfm_stop_monitoring(task, regs);
3628 }
3629 }
3630 /*
3631 * clear overflowed PMD mask to remove any stale information
3632 */
3633 ctx->ctx_ovfl_regs[0] = 0UL;
3634
3635 /*
3636 * back to LOADED state
3637 */
3638 ctx->ctx_state = PFM_CTX_LOADED;
3639
3640 /*
3641 * XXX: not really useful for self monitoring
3642 */
3643 ctx->ctx_fl_can_restart = 0;
3644
3645 return 0;
3646 }
3647
3648 /*
3649 * restart another task
3650 */
3651
3652 /*
3653 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3654 * one is seen by the task.
3655 */
3656 if (state == PFM_CTX_MASKED) {
3657 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3658 /*
3659 * will prevent subsequent restart before this one is
3660 * seen by other task
3661 */
3662 ctx->ctx_fl_can_restart = 0;
3663 }
3664
3665 /*
3666 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3667 * the task is blocked or on its way to block. That's the normal
3668 * restart path. If the monitoring is not masked, then the task
3669 * can be actively monitoring and we cannot directly intervene.
3670 * Therefore we use the trap mechanism to catch the task and
3671 * force it to reset the buffer/reset PMDs.
3672 *
3673 * if non-blocking, then we ensure that the task will go into
3674 * pfm_handle_work() before returning to user mode.
3675 *
3676 * We cannot explicitly reset another task, it MUST always
3677 * be done by the task itself. This works for system wide because
3678 * the tool that is controlling the session is logically doing
3679 * "self-monitoring".
3680 */
3681 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3682 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3683 complete(&ctx->ctx_restart_done);
3684 } else {
3685 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3686
3687 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3688
3689 PFM_SET_WORK_PENDING(task, 1);
3690
3691 set_notify_resume(task);
3692
3693 /*
3694 * XXX: send reschedule if task runs on another CPU
3695 */
3696 }
3697 return 0;
3698 }
3699
3700 static int
3701 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3702 {
3703 unsigned int m = *(unsigned int *)arg;
3704
3705 pfm_sysctl.debug = m == 0 ? 0 : 1;
3706
3707 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3708
3709 if (m == 0) {
3710 memset(pfm_stats, 0, sizeof(pfm_stats));
3711 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3712 }
3713 return 0;
3714 }
3715
3716 /*
3717 * arg can be NULL and count can be zero for this function
3718 */
3719 static int
3720 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3721 {
3722 struct thread_struct *thread = NULL;
3723 struct task_struct *task;
3724 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3725 unsigned long flags;
3726 dbreg_t dbreg;
3727 unsigned int rnum;
3728 int first_time;
3729 int ret = 0, state;
3730 int i, can_access_pmu = 0;
3731 int is_system, is_loaded;
3732
3733 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3734
3735 state = ctx->ctx_state;
3736 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3737 is_system = ctx->ctx_fl_system;
3738 task = ctx->ctx_task;
3739
3740 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3741
3742 /*
3743 * on both UP and SMP, we can only write to the PMC when the task is
3744 * the owner of the local PMU.
3745 */
3746 if (is_loaded) {
3747 thread = &task->thread;
3748 /*
3749 * In system wide and when the context is loaded, access can only happen
3750 * when the caller is running on the CPU being monitored by the session.
3751 * It does not have to be the owner (ctx_task) of the context per se.
3752 */
3753 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3754 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3755 return -EBUSY;
3756 }
3757 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3758 }
3759
3760 /*
3761 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3762 * ensuring that no real breakpoint can be installed via this call.
3763 *
3764 * IMPORTANT: regs can be NULL in this function
3765 */
3766
3767 first_time = ctx->ctx_fl_using_dbreg == 0;
3768
3769 /*
3770 * don't bother if we are loaded and task is being debugged
3771 */
3772 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3773 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3774 return -EBUSY;
3775 }
3776
3777 /*
3778 * check for debug registers in system wide mode
3779 *
3780 * If though a check is done in pfm_context_load(),
3781 * we must repeat it here, in case the registers are
3782 * written after the context is loaded
3783 */
3784 if (is_loaded) {
3785 LOCK_PFS(flags);
3786
3787 if (first_time && is_system) {
3788 if (pfm_sessions.pfs_ptrace_use_dbregs)
3789 ret = -EBUSY;
3790 else
3791 pfm_sessions.pfs_sys_use_dbregs++;
3792 }
3793 UNLOCK_PFS(flags);
3794 }
3795
3796 if (ret != 0) return ret;
3797
3798 /*
3799 * mark ourself as user of the debug registers for
3800 * perfmon purposes.
3801 */
3802 ctx->ctx_fl_using_dbreg = 1;
3803
3804 /*
3805 * clear hardware registers to make sure we don't
3806 * pick up stale state.
3807 *
3808 * for a system wide session, we do not use
3809 * thread.dbr, thread.ibr because this process
3810 * never leaves the current CPU and the state
3811 * is shared by all processes running on it
3812 */
3813 if (first_time && can_access_pmu) {
3814 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3815 for (i=0; i < pmu_conf->num_ibrs; i++) {
3816 ia64_set_ibr(i, 0UL);
3817 ia64_dv_serialize_instruction();
3818 }
3819 ia64_srlz_i();
3820 for (i=0; i < pmu_conf->num_dbrs; i++) {
3821 ia64_set_dbr(i, 0UL);
3822 ia64_dv_serialize_data();
3823 }
3824 ia64_srlz_d();
3825 }
3826
3827 /*
3828 * Now install the values into the registers
3829 */
3830 for (i = 0; i < count; i++, req++) {
3831
3832 rnum = req->dbreg_num;
3833 dbreg.val = req->dbreg_value;
3834
3835 ret = -EINVAL;
3836
3837 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3838 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3839 rnum, dbreg.val, mode, i, count));
3840
3841 goto abort_mission;
3842 }
3843
3844 /*
3845 * make sure we do not install enabled breakpoint
3846 */
3847 if (rnum & 0x1) {
3848 if (mode == PFM_CODE_RR)
3849 dbreg.ibr.ibr_x = 0;
3850 else
3851 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3852 }
3853
3854 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3855
3856 /*
3857 * Debug registers, just like PMC, can only be modified
3858 * by a kernel call. Moreover, perfmon() access to those
3859 * registers are centralized in this routine. The hardware
3860 * does not modify the value of these registers, therefore,
3861 * if we save them as they are written, we can avoid having
3862 * to save them on context switch out. This is made possible
3863 * by the fact that when perfmon uses debug registers, ptrace()
3864 * won't be able to modify them concurrently.
3865 */
3866 if (mode == PFM_CODE_RR) {
3867 CTX_USED_IBR(ctx, rnum);
3868
3869 if (can_access_pmu) {
3870 ia64_set_ibr(rnum, dbreg.val);
3871 ia64_dv_serialize_instruction();
3872 }
3873
3874 ctx->ctx_ibrs[rnum] = dbreg.val;
3875
3876 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3877 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3878 } else {
3879 CTX_USED_DBR(ctx, rnum);
3880
3881 if (can_access_pmu) {
3882 ia64_set_dbr(rnum, dbreg.val);
3883 ia64_dv_serialize_data();
3884 }
3885 ctx->ctx_dbrs[rnum] = dbreg.val;
3886
3887 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3888 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3889 }
3890 }
3891
3892 return 0;
3893
3894 abort_mission:
3895 /*
3896 * in case it was our first attempt, we undo the global modifications
3897 */
3898 if (first_time) {
3899 LOCK_PFS(flags);
3900 if (ctx->ctx_fl_system) {
3901 pfm_sessions.pfs_sys_use_dbregs--;
3902 }
3903 UNLOCK_PFS(flags);
3904 ctx->ctx_fl_using_dbreg = 0;
3905 }
3906 /*
3907 * install error return flag
3908 */
3909 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3910
3911 return ret;
3912 }
3913
3914 static int
3915 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3916 {
3917 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3918 }
3919
3920 static int
3921 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3922 {
3923 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3924 }
3925
3926 int
3927 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3928 {
3929 pfm_context_t *ctx;
3930
3931 if (req == NULL) return -EINVAL;
3932
3933 ctx = GET_PMU_CTX();
3934
3935 if (ctx == NULL) return -EINVAL;
3936
3937 /*
3938 * for now limit to current task, which is enough when calling
3939 * from overflow handler
3940 */
3941 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3942
3943 return pfm_write_ibrs(ctx, req, nreq, regs);
3944 }
3945 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3946
3947 int
3948 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3949 {
3950 pfm_context_t *ctx;
3951
3952 if (req == NULL) return -EINVAL;
3953
3954 ctx = GET_PMU_CTX();
3955
3956 if (ctx == NULL) return -EINVAL;
3957
3958 /*
3959 * for now limit to current task, which is enough when calling
3960 * from overflow handler
3961 */
3962 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3963
3964 return pfm_write_dbrs(ctx, req, nreq, regs);
3965 }
3966 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3967
3968
3969 static int
3970 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3971 {
3972 pfarg_features_t *req = (pfarg_features_t *)arg;
3973
3974 req->ft_version = PFM_VERSION;
3975 return 0;
3976 }
3977
3978 static int
3979 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3980 {
3981 struct pt_regs *tregs;
3982 struct task_struct *task = PFM_CTX_TASK(ctx);
3983 int state, is_system;
3984
3985 state = ctx->ctx_state;
3986 is_system = ctx->ctx_fl_system;
3987
3988 /*
3989 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3990 */
3991 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3992
3993 /*
3994 * In system wide and when the context is loaded, access can only happen
3995 * when the caller is running on the CPU being monitored by the session.
3996 * It does not have to be the owner (ctx_task) of the context per se.
3997 */
3998 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3999 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4000 return -EBUSY;
4001 }
4002 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4003 task_pid_nr(PFM_CTX_TASK(ctx)),
4004 state,
4005 is_system));
4006 /*
4007 * in system mode, we need to update the PMU directly
4008 * and the user level state of the caller, which may not
4009 * necessarily be the creator of the context.
4010 */
4011 if (is_system) {
4012 /*
4013 * Update local PMU first
4014 *
4015 * disable dcr pp
4016 */
4017 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4018 ia64_srlz_i();
4019
4020 /*
4021 * update local cpuinfo
4022 */
4023 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4024
4025 /*
4026 * stop monitoring, does srlz.i
4027 */
4028 pfm_clear_psr_pp();
4029
4030 /*
4031 * stop monitoring in the caller
4032 */
4033 ia64_psr(regs)->pp = 0;
4034
4035 return 0;
4036 }
4037 /*
4038 * per-task mode
4039 */
4040
4041 if (task == current) {
4042 /* stop monitoring at kernel level */
4043 pfm_clear_psr_up();
4044
4045 /*
4046 * stop monitoring at the user level
4047 */
4048 ia64_psr(regs)->up = 0;
4049 } else {
4050 tregs = task_pt_regs(task);
4051
4052 /*
4053 * stop monitoring at the user level
4054 */
4055 ia64_psr(tregs)->up = 0;
4056
4057 /*
4058 * monitoring disabled in kernel at next reschedule
4059 */
4060 ctx->ctx_saved_psr_up = 0;
4061 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4062 }
4063 return 0;
4064 }
4065
4066
4067 static int
4068 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4069 {
4070 struct pt_regs *tregs;
4071 int state, is_system;
4072
4073 state = ctx->ctx_state;
4074 is_system = ctx->ctx_fl_system;
4075
4076 if (state != PFM_CTX_LOADED) return -EINVAL;
4077
4078 /*
4079 * In system wide and when the context is loaded, access can only happen
4080 * when the caller is running on the CPU being monitored by the session.
4081 * It does not have to be the owner (ctx_task) of the context per se.
4082 */
4083 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4084 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4085 return -EBUSY;
4086 }
4087
4088 /*
4089 * in system mode, we need to update the PMU directly
4090 * and the user level state of the caller, which may not
4091 * necessarily be the creator of the context.
4092 */
4093 if (is_system) {
4094
4095 /*
4096 * set user level psr.pp for the caller
4097 */
4098 ia64_psr(regs)->pp = 1;
4099
4100 /*
4101 * now update the local PMU and cpuinfo
4102 */
4103 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4104
4105 /*
4106 * start monitoring at kernel level
4107 */
4108 pfm_set_psr_pp();
4109
4110 /* enable dcr pp */
4111 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4112 ia64_srlz_i();
4113
4114 return 0;
4115 }
4116
4117 /*
4118 * per-process mode
4119 */
4120
4121 if (ctx->ctx_task == current) {
4122
4123 /* start monitoring at kernel level */
4124 pfm_set_psr_up();
4125
4126 /*
4127 * activate monitoring at user level
4128 */
4129 ia64_psr(regs)->up = 1;
4130
4131 } else {
4132 tregs = task_pt_regs(ctx->ctx_task);
4133
4134 /*
4135 * start monitoring at the kernel level the next
4136 * time the task is scheduled
4137 */
4138 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4139
4140 /*
4141 * activate monitoring at user level
4142 */
4143 ia64_psr(tregs)->up = 1;
4144 }
4145 return 0;
4146 }
4147
4148 static int
4149 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4150 {
4151 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4152 unsigned int cnum;
4153 int i;
4154 int ret = -EINVAL;
4155
4156 for (i = 0; i < count; i++, req++) {
4157
4158 cnum = req->reg_num;
4159
4160 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4161
4162 req->reg_value = PMC_DFL_VAL(cnum);
4163
4164 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4165
4166 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4167 }
4168 return 0;
4169
4170 abort_mission:
4171 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4172 return ret;
4173 }
4174
4175 static int
4176 pfm_check_task_exist(pfm_context_t *ctx)
4177 {
4178 struct task_struct *g, *t;
4179 int ret = -ESRCH;
4180
4181 read_lock(&tasklist_lock);
4182
4183 do_each_thread (g, t) {
4184 if (t->thread.pfm_context == ctx) {
4185 ret = 0;
4186 goto out;
4187 }
4188 } while_each_thread (g, t);
4189 out:
4190 read_unlock(&tasklist_lock);
4191
4192 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4193
4194 return ret;
4195 }
4196
4197 static int
4198 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4199 {
4200 struct task_struct *task;
4201 struct thread_struct *thread;
4202 struct pfm_context_t *old;
4203 unsigned long flags;
4204 #ifndef CONFIG_SMP
4205 struct task_struct *owner_task = NULL;
4206 #endif
4207 pfarg_load_t *req = (pfarg_load_t *)arg;
4208 unsigned long *pmcs_source, *pmds_source;
4209 int the_cpu;
4210 int ret = 0;
4211 int state, is_system, set_dbregs = 0;
4212
4213 state = ctx->ctx_state;
4214 is_system = ctx->ctx_fl_system;
4215 /*
4216 * can only load from unloaded or terminated state
4217 */
4218 if (state != PFM_CTX_UNLOADED) {
4219 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4220 req->load_pid,
4221 ctx->ctx_state));
4222 return -EBUSY;
4223 }
4224
4225 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4226
4227 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4228 DPRINT(("cannot use blocking mode on self\n"));
4229 return -EINVAL;
4230 }
4231
4232 ret = pfm_get_task(ctx, req->load_pid, &task);
4233 if (ret) {
4234 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4235 return ret;
4236 }
4237
4238 ret = -EINVAL;
4239
4240 /*
4241 * system wide is self monitoring only
4242 */
4243 if (is_system && task != current) {
4244 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4245 req->load_pid));
4246 goto error;
4247 }
4248
4249 thread = &task->thread;
4250
4251 ret = 0;
4252 /*
4253 * cannot load a context which is using range restrictions,
4254 * into a task that is being debugged.
4255 */
4256 if (ctx->ctx_fl_using_dbreg) {
4257 if (thread->flags & IA64_THREAD_DBG_VALID) {
4258 ret = -EBUSY;
4259 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4260 goto error;
4261 }
4262 LOCK_PFS(flags);
4263
4264 if (is_system) {
4265 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4266 DPRINT(("cannot load [%d] dbregs in use\n",
4267 task_pid_nr(task)));
4268 ret = -EBUSY;
4269 } else {
4270 pfm_sessions.pfs_sys_use_dbregs++;
4271 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4272 set_dbregs = 1;
4273 }
4274 }
4275
4276 UNLOCK_PFS(flags);
4277
4278 if (ret) goto error;
4279 }
4280
4281 /*
4282 * SMP system-wide monitoring implies self-monitoring.
4283 *
4284 * The programming model expects the task to
4285 * be pinned on a CPU throughout the session.
4286 * Here we take note of the current CPU at the
4287 * time the context is loaded. No call from
4288 * another CPU will be allowed.
4289 *
4290 * The pinning via shed_setaffinity()
4291 * must be done by the calling task prior
4292 * to this call.
4293 *
4294 * systemwide: keep track of CPU this session is supposed to run on
4295 */
4296 the_cpu = ctx->ctx_cpu = smp_processor_id();
4297
4298 ret = -EBUSY;
4299 /*
4300 * now reserve the session
4301 */
4302 ret = pfm_reserve_session(current, is_system, the_cpu);
4303 if (ret) goto error;
4304
4305 /*
4306 * task is necessarily stopped at this point.
4307 *
4308 * If the previous context was zombie, then it got removed in
4309 * pfm_save_regs(). Therefore we should not see it here.
4310 * If we see a context, then this is an active context
4311 *
4312 * XXX: needs to be atomic
4313 */
4314 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4315 thread->pfm_context, ctx));
4316
4317 ret = -EBUSY;
4318 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4319 if (old != NULL) {
4320 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4321 goto error_unres;
4322 }
4323
4324 pfm_reset_msgq(ctx);
4325
4326 ctx->ctx_state = PFM_CTX_LOADED;
4327
4328 /*
4329 * link context to task
4330 */
4331 ctx->ctx_task = task;
4332
4333 if (is_system) {
4334 /*
4335 * we load as stopped
4336 */
4337 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4338 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4339
4340 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4341 } else {
4342 thread->flags |= IA64_THREAD_PM_VALID;
4343 }
4344
4345 /*
4346 * propagate into thread-state
4347 */
4348 pfm_copy_pmds(task, ctx);
4349 pfm_copy_pmcs(task, ctx);
4350
4351 pmcs_source = ctx->th_pmcs;
4352 pmds_source = ctx->th_pmds;
4353
4354 /*
4355 * always the case for system-wide
4356 */
4357 if (task == current) {
4358
4359 if (is_system == 0) {
4360
4361 /* allow user level control */
4362 ia64_psr(regs)->sp = 0;
4363 DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4364
4365 SET_LAST_CPU(ctx, smp_processor_id());
4366 INC_ACTIVATION();
4367 SET_ACTIVATION(ctx);
4368 #ifndef CONFIG_SMP
4369 /*
4370 * push the other task out, if any
4371 */
4372 owner_task = GET_PMU_OWNER();
4373 if (owner_task) pfm_lazy_save_regs(owner_task);
4374 #endif
4375 }
4376 /*
4377 * load all PMD from ctx to PMU (as opposed to thread state)
4378 * restore all PMC from ctx to PMU
4379 */
4380 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4381 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4382
4383 ctx->ctx_reload_pmcs[0] = 0UL;
4384 ctx->ctx_reload_pmds[0] = 0UL;
4385
4386 /*
4387 * guaranteed safe by earlier check against DBG_VALID
4388 */
4389 if (ctx->ctx_fl_using_dbreg) {
4390 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4391 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4392 }
4393 /*
4394 * set new ownership
4395 */
4396 SET_PMU_OWNER(task, ctx);
4397
4398 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4399 } else {
4400 /*
4401 * when not current, task MUST be stopped, so this is safe
4402 */
4403 regs = task_pt_regs(task);
4404
4405 /* force a full reload */
4406 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4407 SET_LAST_CPU(ctx, -1);
4408
4409 /* initial saved psr (stopped) */
4410 ctx->ctx_saved_psr_up = 0UL;
4411 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4412 }
4413
4414 ret = 0;
4415
4416 error_unres:
4417 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4418 error:
4419 /*
4420 * we must undo the dbregs setting (for system-wide)
4421 */
4422 if (ret && set_dbregs) {
4423 LOCK_PFS(flags);
4424 pfm_sessions.pfs_sys_use_dbregs--;
4425 UNLOCK_PFS(flags);
4426 }
4427 /*
4428 * release task, there is now a link with the context
4429 */
4430 if (is_system == 0 && task != current) {
4431 pfm_put_task(task);
4432
4433 if (ret == 0) {
4434 ret = pfm_check_task_exist(ctx);
4435 if (ret) {
4436 ctx->ctx_state = PFM_CTX_UNLOADED;
4437 ctx->ctx_task = NULL;
4438 }
4439 }
4440 }
4441 return ret;
4442 }
4443
4444 /*
4445 * in this function, we do not need to increase the use count
4446 * for the task via get_task_struct(), because we hold the
4447 * context lock. If the task were to disappear while having
4448 * a context attached, it would go through pfm_exit_thread()
4449 * which also grabs the context lock and would therefore be blocked
4450 * until we are here.
4451 */
4452 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4453
4454 static int
4455 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4456 {
4457 struct task_struct *task = PFM_CTX_TASK(ctx);
4458 struct pt_regs *tregs;
4459 int prev_state, is_system;
4460 int ret;
4461
4462 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4463
4464 prev_state = ctx->ctx_state;
4465 is_system = ctx->ctx_fl_system;
4466
4467 /*
4468 * unload only when necessary
4469 */
4470 if (prev_state == PFM_CTX_UNLOADED) {
4471 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4472 return 0;
4473 }
4474
4475 /*
4476 * clear psr and dcr bits
4477 */
4478 ret = pfm_stop(ctx, NULL, 0, regs);
4479 if (ret) return ret;
4480
4481 ctx->ctx_state = PFM_CTX_UNLOADED;
4482
4483 /*
4484 * in system mode, we need to update the PMU directly
4485 * and the user level state of the caller, which may not
4486 * necessarily be the creator of the context.
4487 */
4488 if (is_system) {
4489
4490 /*
4491 * Update cpuinfo
4492 *
4493 * local PMU is taken care of in pfm_stop()
4494 */
4495 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4496 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4497
4498 /*
4499 * save PMDs in context
4500 * release ownership
4501 */
4502 pfm_flush_pmds(current, ctx);
4503
4504 /*
4505 * at this point we are done with the PMU
4506 * so we can unreserve the resource.
4507 */
4508 if (prev_state != PFM_CTX_ZOMBIE)
4509 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4510
4511 /*
4512 * disconnect context from task
4513 */
4514 task->thread.pfm_context = NULL;
4515 /*
4516 * disconnect task from context
4517 */
4518 ctx->ctx_task = NULL;
4519
4520 /*
4521 * There is nothing more to cleanup here.
4522 */
4523 return 0;
4524 }
4525
4526 /*
4527 * per-task mode
4528 */
4529 tregs = task == current ? regs : task_pt_regs(task);
4530
4531 if (task == current) {
4532 /*
4533 * cancel user level control
4534 */
4535 ia64_psr(regs)->sp = 1;
4536
4537 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4538 }
4539 /*
4540 * save PMDs to context
4541 * release ownership
4542 */
4543 pfm_flush_pmds(task, ctx);
4544
4545 /*
4546 * at this point we are done with the PMU
4547 * so we can unreserve the resource.
4548 *
4549 * when state was ZOMBIE, we have already unreserved.
4550 */
4551 if (prev_state != PFM_CTX_ZOMBIE)
4552 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4553
4554 /*
4555 * reset activation counter and psr
4556 */
4557 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4558 SET_LAST_CPU(ctx, -1);
4559
4560 /*
4561 * PMU state will not be restored
4562 */
4563 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4564
4565 /*
4566 * break links between context and task
4567 */
4568 task->thread.pfm_context = NULL;
4569 ctx->ctx_task = NULL;
4570
4571 PFM_SET_WORK_PENDING(task, 0);
4572
4573 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4574 ctx->ctx_fl_can_restart = 0;
4575 ctx->ctx_fl_going_zombie = 0;
4576
4577 DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4578
4579 return 0;
4580 }
4581
4582
4583 /*
4584 * called only from exit_thread(): task == current
4585 * we come here only if current has a context attached (loaded or masked)
4586 */
4587 void
4588 pfm_exit_thread(struct task_struct *task)
4589 {
4590 pfm_context_t *ctx;
4591 unsigned long flags;
4592 struct pt_regs *regs = task_pt_regs(task);
4593 int ret, state;
4594 int free_ok = 0;
4595
4596 ctx = PFM_GET_CTX(task);
4597
4598 PROTECT_CTX(ctx, flags);
4599
4600 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4601
4602 state = ctx->ctx_state;
4603 switch(state) {
4604 case PFM_CTX_UNLOADED:
4605 /*
4606 * only comes to this function if pfm_context is not NULL, i.e., cannot
4607 * be in unloaded state
4608 */
4609 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4610 break;
4611 case PFM_CTX_LOADED:
4612 case PFM_CTX_MASKED:
4613 ret = pfm_context_unload(ctx, NULL, 0, regs);
4614 if (ret) {
4615 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4616 }
4617 DPRINT(("ctx unloaded for current state was %d\n", state));
4618
4619 pfm_end_notify_user(ctx);
4620 break;
4621 case PFM_CTX_ZOMBIE:
4622 ret = pfm_context_unload(ctx, NULL, 0, regs);
4623 if (ret) {
4624 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4625 }
4626 free_ok = 1;
4627 break;
4628 default:
4629 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4630 break;
4631 }
4632 UNPROTECT_CTX(ctx, flags);
4633
4634 { u64 psr = pfm_get_psr();
4635 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4636 BUG_ON(GET_PMU_OWNER());
4637 BUG_ON(ia64_psr(regs)->up);
4638 BUG_ON(ia64_psr(regs)->pp);
4639 }
4640
4641 /*
4642 * All memory free operations (especially for vmalloc'ed memory)
4643 * MUST be done with interrupts ENABLED.
4644 */
4645 if (free_ok) pfm_context_free(ctx);
4646 }
4647
4648 /*
4649 * functions MUST be listed in the increasing order of their index (see permfon.h)
4650 */
4651 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4652 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4653 #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4654 #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4655 #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4656
4657 static pfm_cmd_desc_t pfm_cmd_tab[]={
4658 /* 0 */PFM_CMD_NONE,
4659 /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4660 /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4661 /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4662 /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4663 /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4664 /* 6 */PFM_CMD_NONE,
4665 /* 7 */PFM_CMD_NONE,
4666 /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4667 /* 9 */PFM_CMD_NONE,
4668 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4669 /* 11 */PFM_CMD_NONE,
4670 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4671 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4672 /* 14 */PFM_CMD_NONE,
4673 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4674 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4675 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4676 /* 18 */PFM_CMD_NONE,
4677 /* 19 */PFM_CMD_NONE,
4678 /* 20 */PFM_CMD_NONE,
4679 /* 21 */PFM_CMD_NONE,
4680 /* 22 */PFM_CMD_NONE,
4681 /* 23 */PFM_CMD_NONE,
4682 /* 24 */PFM_CMD_NONE,
4683 /* 25 */PFM_CMD_NONE,
4684 /* 26 */PFM_CMD_NONE,
4685 /* 27 */PFM_CMD_NONE,
4686 /* 28 */PFM_CMD_NONE,
4687 /* 29 */PFM_CMD_NONE,
4688 /* 30 */PFM_CMD_NONE,
4689 /* 31 */PFM_CMD_NONE,
4690 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4691 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4692 };
4693 #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4694
4695 static int
4696 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4697 {
4698 struct task_struct *task;
4699 int state, old_state;
4700
4701 recheck:
4702 state = ctx->ctx_state;
4703 task = ctx->ctx_task;
4704
4705 if (task == NULL) {
4706 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4707 return 0;
4708 }
4709
4710 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4711 ctx->ctx_fd,
4712 state,
4713 task_pid_nr(task),
4714 task->state, PFM_CMD_STOPPED(cmd)));
4715
4716 /*
4717 * self-monitoring always ok.
4718 *
4719 * for system-wide the caller can either be the creator of the
4720 * context (to one to which the context is attached to) OR
4721 * a task running on the same CPU as the session.
4722 */
4723 if (task == current || ctx->ctx_fl_system) return 0;
4724
4725 /*
4726 * we are monitoring another thread
4727 */
4728 switch(state) {
4729 case PFM_CTX_UNLOADED:
4730 /*
4731 * if context is UNLOADED we are safe to go
4732 */
4733 return 0;
4734 case PFM_CTX_ZOMBIE:
4735 /*
4736 * no command can operate on a zombie context
4737 */
4738 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4739 return -EINVAL;
4740 case PFM_CTX_MASKED:
4741 /*
4742 * PMU state has been saved to software even though
4743 * the thread may still be running.
4744 */
4745 if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4746 }
4747
4748 /*
4749 * context is LOADED or MASKED. Some commands may need to have
4750 * the task stopped.
4751 *
4752 * We could lift this restriction for UP but it would mean that
4753 * the user has no guarantee the task would not run between
4754 * two successive calls to perfmonctl(). That's probably OK.
4755 * If this user wants to ensure the task does not run, then
4756 * the task must be stopped.
4757 */
4758 if (PFM_CMD_STOPPED(cmd)) {
4759 if (!task_is_stopped_or_traced(task)) {
4760 DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4761 return -EBUSY;
4762 }
4763 /*
4764 * task is now stopped, wait for ctxsw out
4765 *
4766 * This is an interesting point in the code.
4767 * We need to unprotect the context because
4768 * the pfm_save_regs() routines needs to grab
4769 * the same lock. There are danger in doing
4770 * this because it leaves a window open for
4771 * another task to get access to the context
4772 * and possibly change its state. The one thing
4773 * that is not possible is for the context to disappear
4774 * because we are protected by the VFS layer, i.e.,
4775 * get_fd()/put_fd().
4776 */
4777 old_state = state;
4778
4779 UNPROTECT_CTX(ctx, flags);
4780
4781 wait_task_inactive(task, 0);
4782
4783 PROTECT_CTX(ctx, flags);
4784
4785 /*
4786 * we must recheck to verify if state has changed
4787 */
4788 if (ctx->ctx_state != old_state) {
4789 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4790 goto recheck;
4791 }
4792 }
4793 return 0;
4794 }
4795
4796 /*
4797 * system-call entry point (must return long)
4798 */
4799 asmlinkage long
4800 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4801 {
4802 struct file *file = NULL;
4803 pfm_context_t *ctx = NULL;
4804 unsigned long flags = 0UL;
4805 void *args_k = NULL;
4806 long ret; /* will expand int return types */
4807 size_t base_sz, sz, xtra_sz = 0;
4808 int narg, completed_args = 0, call_made = 0, cmd_flags;
4809 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4810 int (*getsize)(void *arg, size_t *sz);
4811 #define PFM_MAX_ARGSIZE 4096
4812
4813 /*
4814 * reject any call if perfmon was disabled at initialization
4815 */
4816 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4817
4818 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4819 DPRINT(("invalid cmd=%d\n", cmd));
4820 return -EINVAL;
4821 }
4822
4823 func = pfm_cmd_tab[cmd].cmd_func;
4824 narg = pfm_cmd_tab[cmd].cmd_narg;
4825 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4826 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4827 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4828
4829 if (unlikely(func == NULL)) {
4830 DPRINT(("invalid cmd=%d\n", cmd));
4831 return -EINVAL;
4832 }
4833
4834 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4835 PFM_CMD_NAME(cmd),
4836 cmd,
4837 narg,
4838 base_sz,
4839 count));
4840
4841 /*
4842 * check if number of arguments matches what the command expects
4843 */
4844 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4845 return -EINVAL;
4846
4847 restart_args:
4848 sz = xtra_sz + base_sz*count;
4849 /*
4850 * limit abuse to min page size
4851 */
4852 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4853 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4854 return -E2BIG;
4855 }
4856
4857 /*
4858 * allocate default-sized argument buffer
4859 */
4860 if (likely(count && args_k == NULL)) {
4861 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4862 if (args_k == NULL) return -ENOMEM;
4863 }
4864
4865 ret = -EFAULT;
4866
4867 /*
4868 * copy arguments
4869 *
4870 * assume sz = 0 for command without parameters
4871 */
4872 if (sz && copy_from_user(args_k, arg, sz)) {
4873 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4874 goto error_args;
4875 }
4876
4877 /*
4878 * check if command supports extra parameters
4879 */
4880 if (completed_args == 0 && getsize) {
4881 /*
4882 * get extra parameters size (based on main argument)
4883 */
4884 ret = (*getsize)(args_k, &xtra_sz);
4885 if (ret) goto error_args;
4886
4887 completed_args = 1;
4888
4889 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4890
4891 /* retry if necessary */
4892 if (likely(xtra_sz)) goto restart_args;
4893 }
4894
4895 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4896
4897 ret = -EBADF;
4898
4899 file = fget(fd);
4900 if (unlikely(file == NULL)) {
4901 DPRINT(("invalid fd %d\n", fd));
4902 goto error_args;
4903 }
4904 if (unlikely(PFM_IS_FILE(file) == 0)) {
4905 DPRINT(("fd %d not related to perfmon\n", fd));
4906 goto error_args;
4907 }
4908
4909 ctx = (pfm_context_t *)file->private_data;
4910 if (unlikely(ctx == NULL)) {
4911 DPRINT(("no context for fd %d\n", fd));
4912 goto error_args;
4913 }
4914 prefetch(&ctx->ctx_state);
4915
4916 PROTECT_CTX(ctx, flags);
4917
4918 /*
4919 * check task is stopped
4920 */
4921 ret = pfm_check_task_state(ctx, cmd, flags);
4922 if (unlikely(ret)) goto abort_locked;
4923
4924 skip_fd:
4925 ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4926
4927 call_made = 1;
4928
4929 abort_locked:
4930 if (likely(ctx)) {
4931 DPRINT(("context unlocked\n"));
4932 UNPROTECT_CTX(ctx, flags);
4933 }
4934
4935 /* copy argument back to user, if needed */
4936 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4937
4938 error_args:
4939 if (file)
4940 fput(file);
4941
4942 kfree(args_k);
4943
4944 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4945
4946 return ret;
4947 }
4948
4949 static void
4950 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4951 {
4952 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4953 pfm_ovfl_ctrl_t rst_ctrl;
4954 int state;
4955 int ret = 0;
4956
4957 state = ctx->ctx_state;
4958 /*
4959 * Unlock sampling buffer and reset index atomically
4960 * XXX: not really needed when blocking
4961 */
4962 if (CTX_HAS_SMPL(ctx)) {
4963
4964 rst_ctrl.bits.mask_monitoring = 0;
4965 rst_ctrl.bits.reset_ovfl_pmds = 0;
4966
4967 if (state == PFM_CTX_LOADED)
4968 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4969 else
4970 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4971 } else {
4972 rst_ctrl.bits.mask_monitoring = 0;
4973 rst_ctrl.bits.reset_ovfl_pmds = 1;
4974 }
4975
4976 if (ret == 0) {
4977 if (rst_ctrl.bits.reset_ovfl_pmds) {
4978 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4979 }
4980 if (rst_ctrl.bits.mask_monitoring == 0) {
4981 DPRINT(("resuming monitoring\n"));
4982 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4983 } else {
4984 DPRINT(("stopping monitoring\n"));
4985 //pfm_stop_monitoring(current, regs);
4986 }
4987 ctx->ctx_state = PFM_CTX_LOADED;
4988 }
4989 }
4990
4991 /*
4992 * context MUST BE LOCKED when calling
4993 * can only be called for current
4994 */
4995 static void
4996 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4997 {
4998 int ret;
4999
5000 DPRINT(("entering for [%d]\n", task_pid_nr(current)));
5001
5002 ret = pfm_context_unload(ctx, NULL, 0, regs);
5003 if (ret) {
5004 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
5005 }
5006
5007 /*
5008 * and wakeup controlling task, indicating we are now disconnected
5009 */
5010 wake_up_interruptible(&ctx->ctx_zombieq);
5011
5012 /*
5013 * given that context is still locked, the controlling
5014 * task will only get access when we return from
5015 * pfm_handle_work().
5016 */
5017 }
5018
5019 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5020
5021 /*
5022 * pfm_handle_work() can be called with interrupts enabled
5023 * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5024 * call may sleep, therefore we must re-enable interrupts
5025 * to avoid deadlocks. It is safe to do so because this function
5026 * is called ONLY when returning to user level (pUStk=1), in which case
5027 * there is no risk of kernel stack overflow due to deep
5028 * interrupt nesting.
5029 */
5030 void
5031 pfm_handle_work(void)
5032 {
5033 pfm_context_t *ctx;
5034 struct pt_regs *regs;
5035 unsigned long flags, dummy_flags;
5036 unsigned long ovfl_regs;
5037 unsigned int reason;
5038 int ret;
5039
5040 ctx = PFM_GET_CTX(current);
5041 if (ctx == NULL) {
5042 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5043 task_pid_nr(current));
5044 return;
5045 }
5046
5047 PROTECT_CTX(ctx, flags);
5048
5049 PFM_SET_WORK_PENDING(current, 0);
5050
5051 regs = task_pt_regs(current);
5052
5053 /*
5054 * extract reason for being here and clear
5055 */
5056 reason = ctx->ctx_fl_trap_reason;
5057 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5058 ovfl_regs = ctx->ctx_ovfl_regs[0];
5059
5060 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5061
5062 /*
5063 * must be done before we check for simple-reset mode
5064 */
5065 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5066 goto do_zombie;
5067
5068 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5069 if (reason == PFM_TRAP_REASON_RESET)
5070 goto skip_blocking;
5071
5072 /*
5073 * restore interrupt mask to what it was on entry.
5074 * Could be enabled/diasbled.
5075 */
5076 UNPROTECT_CTX(ctx, flags);
5077
5078 /*
5079 * force interrupt enable because of down_interruptible()
5080 */
5081 local_irq_enable();
5082
5083 DPRINT(("before block sleeping\n"));
5084
5085 /*
5086 * may go through without blocking on SMP systems
5087 * if restart has been received already by the time we call down()
5088 */
5089 ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5090
5091 DPRINT(("after block sleeping ret=%d\n", ret));
5092
5093 /*
5094 * lock context and mask interrupts again
5095 * We save flags into a dummy because we may have
5096 * altered interrupts mask compared to entry in this
5097 * function.
5098 */
5099 PROTECT_CTX(ctx, dummy_flags);
5100
5101 /*
5102 * we need to read the ovfl_regs only after wake-up
5103 * because we may have had pfm_write_pmds() in between
5104 * and that can changed PMD values and therefore
5105 * ovfl_regs is reset for these new PMD values.
5106 */
5107 ovfl_regs = ctx->ctx_ovfl_regs[0];
5108
5109 if (ctx->ctx_fl_going_zombie) {
5110 do_zombie:
5111 DPRINT(("context is zombie, bailing out\n"));
5112 pfm_context_force_terminate(ctx, regs);
5113 goto nothing_to_do;
5114 }
5115 /*
5116 * in case of interruption of down() we don't restart anything
5117 */
5118 if (ret < 0)
5119 goto nothing_to_do;
5120
5121 skip_blocking:
5122 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5123 ctx->ctx_ovfl_regs[0] = 0UL;
5124
5125 nothing_to_do:
5126 /*
5127 * restore flags as they were upon entry
5128 */
5129 UNPROTECT_CTX(ctx, flags);
5130 }
5131
5132 static int
5133 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5134 {
5135 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5136 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5137 return 0;
5138 }
5139
5140 DPRINT(("waking up somebody\n"));
5141
5142 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5143
5144 /*
5145 * safe, we are not in intr handler, nor in ctxsw when
5146 * we come here
5147 */
5148 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5149
5150 return 0;
5151 }
5152
5153 static int
5154 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5155 {
5156 pfm_msg_t *msg = NULL;
5157
5158 if (ctx->ctx_fl_no_msg == 0) {
5159 msg = pfm_get_new_msg(ctx);
5160 if (msg == NULL) {
5161 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5162 return -1;
5163 }
5164
5165 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5166 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5167 msg->pfm_ovfl_msg.msg_active_set = 0;
5168 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5169 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5170 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5171 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5172 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5173 }
5174
5175 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5176 msg,
5177 ctx->ctx_fl_no_msg,
5178 ctx->ctx_fd,
5179 ovfl_pmds));
5180
5181 return pfm_notify_user(ctx, msg);
5182 }
5183
5184 static int
5185 pfm_end_notify_user(pfm_context_t *ctx)
5186 {
5187 pfm_msg_t *msg;
5188
5189 msg = pfm_get_new_msg(ctx);
5190 if (msg == NULL) {
5191 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5192 return -1;
5193 }
5194 /* no leak */
5195 memset(msg, 0, sizeof(*msg));
5196
5197 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5198 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5199 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5200
5201 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5202 msg,
5203 ctx->ctx_fl_no_msg,
5204 ctx->ctx_fd));
5205
5206 return pfm_notify_user(ctx, msg);
5207 }
5208
5209 /*
5210 * main overflow processing routine.
5211 * it can be called from the interrupt path or explicitly during the context switch code
5212 */
5213 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5214 unsigned long pmc0, struct pt_regs *regs)
5215 {
5216 pfm_ovfl_arg_t *ovfl_arg;
5217 unsigned long mask;
5218 unsigned long old_val, ovfl_val, new_val;
5219 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5220 unsigned long tstamp;
5221 pfm_ovfl_ctrl_t ovfl_ctrl;
5222 unsigned int i, has_smpl;
5223 int must_notify = 0;
5224
5225 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5226
5227 /*
5228 * sanity test. Should never happen
5229 */
5230 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5231
5232 tstamp = ia64_get_itc();
5233 mask = pmc0 >> PMU_FIRST_COUNTER;
5234 ovfl_val = pmu_conf->ovfl_val;
5235 has_smpl = CTX_HAS_SMPL(ctx);
5236
5237 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5238 "used_pmds=0x%lx\n",
5239 pmc0,
5240 task ? task_pid_nr(task): -1,
5241 (regs ? regs->cr_iip : 0),
5242 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5243 ctx->ctx_used_pmds[0]));
5244
5245
5246 /*
5247 * first we update the virtual counters
5248 * assume there was a prior ia64_srlz_d() issued
5249 */
5250 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5251
5252 /* skip pmd which did not overflow */
5253 if ((mask & 0x1) == 0) continue;
5254
5255 /*
5256 * Note that the pmd is not necessarily 0 at this point as qualified events
5257 * may have happened before the PMU was frozen. The residual count is not
5258 * taken into consideration here but will be with any read of the pmd via
5259 * pfm_read_pmds().
5260 */
5261 old_val = new_val = ctx->ctx_pmds[i].val;
5262 new_val += 1 + ovfl_val;
5263 ctx->ctx_pmds[i].val = new_val;
5264
5265 /*
5266 * check for overflow condition
5267 */
5268 if (likely(old_val > new_val)) {
5269 ovfl_pmds |= 1UL << i;
5270 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5271 }
5272
5273 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5274 i,
5275 new_val,
5276 old_val,
5277 ia64_get_pmd(i) & ovfl_val,
5278 ovfl_pmds,
5279 ovfl_notify));
5280 }
5281
5282 /*
5283 * there was no 64-bit overflow, nothing else to do
5284 */
5285 if (ovfl_pmds == 0UL) return;
5286
5287 /*
5288 * reset all control bits
5289 */
5290 ovfl_ctrl.val = 0;
5291 reset_pmds = 0UL;
5292
5293 /*
5294 * if a sampling format module exists, then we "cache" the overflow by
5295 * calling the module's handler() routine.
5296 */
5297 if (has_smpl) {
5298 unsigned long start_cycles, end_cycles;
5299 unsigned long pmd_mask;
5300 int j, k, ret = 0;
5301 int this_cpu = smp_processor_id();
5302
5303 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5304 ovfl_arg = &ctx->ctx_ovfl_arg;
5305
5306 prefetch(ctx->ctx_smpl_hdr);
5307
5308 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5309
5310 mask = 1UL << i;
5311
5312 if ((pmd_mask & 0x1) == 0) continue;
5313
5314 ovfl_arg->ovfl_pmd = (unsigned char )i;
5315 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5316 ovfl_arg->active_set = 0;
5317 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5318 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5319
5320 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5321 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5322 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5323
5324 /*
5325 * copy values of pmds of interest. Sampling format may copy them
5326 * into sampling buffer.
5327 */
5328 if (smpl_pmds) {
5329 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5330 if ((smpl_pmds & 0x1) == 0) continue;
5331 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5332 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5333 }
5334 }
5335
5336 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5337
5338 start_cycles = ia64_get_itc();
5339
5340 /*
5341 * call custom buffer format record (handler) routine
5342 */
5343 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5344
5345 end_cycles = ia64_get_itc();
5346
5347 /*
5348 * For those controls, we take the union because they have
5349 * an all or nothing behavior.
5350 */
5351 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5352 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5353 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5354 /*
5355 * build the bitmask of pmds to reset now
5356 */
5357 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5358
5359 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5360 }
5361 /*
5362 * when the module cannot handle the rest of the overflows, we abort right here
5363 */
5364 if (ret && pmd_mask) {
5365 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5366 pmd_mask<<PMU_FIRST_COUNTER));
5367 }
5368 /*
5369 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5370 */
5371 ovfl_pmds &= ~reset_pmds;
5372 } else {
5373 /*
5374 * when no sampling module is used, then the default
5375 * is to notify on overflow if requested by user
5376 */
5377 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5378 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5379 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5380 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5381 /*
5382 * if needed, we reset all overflowed pmds
5383 */
5384 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5385 }
5386
5387 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5388
5389 /*
5390 * reset the requested PMD registers using the short reset values
5391 */
5392 if (reset_pmds) {
5393 unsigned long bm = reset_pmds;
5394 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5395 }
5396
5397 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5398 /*
5399 * keep track of what to reset when unblocking
5400 */
5401 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5402
5403 /*
5404 * check for blocking context
5405 */
5406 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5407
5408 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5409
5410 /*
5411 * set the perfmon specific checking pending work for the task
5412 */
5413 PFM_SET_WORK_PENDING(task, 1);
5414
5415 /*
5416 * when coming from ctxsw, current still points to the
5417 * previous task, therefore we must work with task and not current.
5418 */
5419 set_notify_resume(task);
5420 }
5421 /*
5422 * defer until state is changed (shorten spin window). the context is locked
5423 * anyway, so the signal receiver would come spin for nothing.
5424 */
5425 must_notify = 1;
5426 }
5427
5428 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5429 GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5430 PFM_GET_WORK_PENDING(task),
5431 ctx->ctx_fl_trap_reason,
5432 ovfl_pmds,
5433 ovfl_notify,
5434 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5435 /*
5436 * in case monitoring must be stopped, we toggle the psr bits
5437 */
5438 if (ovfl_ctrl.bits.mask_monitoring) {
5439 pfm_mask_monitoring(task);
5440 ctx->ctx_state = PFM_CTX_MASKED;
5441 ctx->ctx_fl_can_restart = 1;
5442 }
5443
5444 /*
5445 * send notification now
5446 */
5447 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5448
5449 return;
5450
5451 sanity_check:
5452 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5453 smp_processor_id(),
5454 task ? task_pid_nr(task) : -1,
5455 pmc0);
5456 return;
5457
5458 stop_monitoring:
5459 /*
5460 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5461 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5462 * come here as zombie only if the task is the current task. In which case, we
5463 * can access the PMU hardware directly.
5464 *
5465 * Note that zombies do have PM_VALID set. So here we do the minimal.
5466 *
5467 * In case the context was zombified it could not be reclaimed at the time
5468 * the monitoring program exited. At this point, the PMU reservation has been
5469 * returned, the sampiing buffer has been freed. We must convert this call
5470 * into a spurious interrupt. However, we must also avoid infinite overflows
5471 * by stopping monitoring for this task. We can only come here for a per-task
5472 * context. All we need to do is to stop monitoring using the psr bits which
5473 * are always task private. By re-enabling secure montioring, we ensure that
5474 * the monitored task will not be able to re-activate monitoring.
5475 * The task will eventually be context switched out, at which point the context
5476 * will be reclaimed (that includes releasing ownership of the PMU).
5477 *
5478 * So there might be a window of time where the number of per-task session is zero
5479 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5480 * context. This is safe because if a per-task session comes in, it will push this one
5481 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5482 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5483 * also push our zombie context out.
5484 *
5485 * Overall pretty hairy stuff....
5486 */
5487 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5488 pfm_clear_psr_up();
5489 ia64_psr(regs)->up = 0;
5490 ia64_psr(regs)->sp = 1;
5491 return;
5492 }
5493
5494 static int
5495 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5496 {
5497 struct task_struct *task;
5498 pfm_context_t *ctx;
5499 unsigned long flags;
5500 u64 pmc0;
5501 int this_cpu = smp_processor_id();
5502 int retval = 0;
5503
5504 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5505
5506 /*
5507 * srlz.d done before arriving here
5508 */
5509 pmc0 = ia64_get_pmc(0);
5510
5511 task = GET_PMU_OWNER();
5512 ctx = GET_PMU_CTX();
5513
5514 /*
5515 * if we have some pending bits set
5516 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5517 */
5518 if (PMC0_HAS_OVFL(pmc0) && task) {
5519 /*
5520 * we assume that pmc0.fr is always set here
5521 */
5522
5523 /* sanity check */
5524 if (!ctx) goto report_spurious1;
5525
5526 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5527 goto report_spurious2;
5528
5529 PROTECT_CTX_NOPRINT(ctx, flags);
5530
5531 pfm_overflow_handler(task, ctx, pmc0, regs);
5532
5533 UNPROTECT_CTX_NOPRINT(ctx, flags);
5534
5535 } else {
5536 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5537 retval = -1;
5538 }
5539 /*
5540 * keep it unfrozen at all times
5541 */
5542 pfm_unfreeze_pmu();
5543
5544 return retval;
5545
5546 report_spurious1:
5547 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5548 this_cpu, task_pid_nr(task));
5549 pfm_unfreeze_pmu();
5550 return -1;
5551 report_spurious2:
5552 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5553 this_cpu,
5554 task_pid_nr(task));
5555 pfm_unfreeze_pmu();
5556 return -1;
5557 }
5558
5559 static irqreturn_t
5560 pfm_interrupt_handler(int irq, void *arg)
5561 {
5562 unsigned long start_cycles, total_cycles;
5563 unsigned long min, max;
5564 int this_cpu;
5565 int ret;
5566 struct pt_regs *regs = get_irq_regs();
5567
5568 this_cpu = get_cpu();
5569 if (likely(!pfm_alt_intr_handler)) {
5570 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5571 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5572
5573 start_cycles = ia64_get_itc();
5574
5575 ret = pfm_do_interrupt_handler(arg, regs);
5576
5577 total_cycles = ia64_get_itc();
5578
5579 /*
5580 * don't measure spurious interrupts
5581 */
5582 if (likely(ret == 0)) {
5583 total_cycles -= start_cycles;
5584
5585 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5586 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5587
5588 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5589 }
5590 }
5591 else {
5592 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5593 }
5594
5595 put_cpu();
5596 return IRQ_HANDLED;
5597 }
5598
5599 /*
5600 * /proc/perfmon interface, for debug only
5601 */
5602
5603 #define PFM_PROC_SHOW_HEADER ((void *)(long)nr_cpu_ids+1)
5604
5605 static void *
5606 pfm_proc_start(struct seq_file *m, loff_t *pos)
5607 {
5608 if (*pos == 0) {
5609 return PFM_PROC_SHOW_HEADER;
5610 }
5611
5612 while (*pos <= nr_cpu_ids) {
5613 if (cpu_online(*pos - 1)) {
5614 return (void *)*pos;
5615 }
5616 ++*pos;
5617 }
5618 return NULL;
5619 }
5620
5621 static void *
5622 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5623 {
5624 ++*pos;
5625 return pfm_proc_start(m, pos);
5626 }
5627
5628 static void
5629 pfm_proc_stop(struct seq_file *m, void *v)
5630 {
5631 }
5632
5633 static void
5634 pfm_proc_show_header(struct seq_file *m)
5635 {
5636 struct list_head * pos;
5637 pfm_buffer_fmt_t * entry;
5638 unsigned long flags;
5639
5640 seq_printf(m,
5641 "perfmon version : %u.%u\n"
5642 "model : %s\n"
5643 "fastctxsw : %s\n"
5644 "expert mode : %s\n"
5645 "ovfl_mask : 0x%lx\n"
5646 "PMU flags : 0x%x\n",
5647 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5648 pmu_conf->pmu_name,
5649 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5650 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5651 pmu_conf->ovfl_val,
5652 pmu_conf->flags);
5653
5654 LOCK_PFS(flags);
5655
5656 seq_printf(m,
5657 "proc_sessions : %u\n"
5658 "sys_sessions : %u\n"
5659 "sys_use_dbregs : %u\n"
5660 "ptrace_use_dbregs : %u\n",
5661 pfm_sessions.pfs_task_sessions,
5662 pfm_sessions.pfs_sys_sessions,
5663 pfm_sessions.pfs_sys_use_dbregs,
5664 pfm_sessions.pfs_ptrace_use_dbregs);
5665
5666 UNLOCK_PFS(flags);
5667
5668 spin_lock(&pfm_buffer_fmt_lock);
5669
5670 list_for_each(pos, &pfm_buffer_fmt_list) {
5671 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5672 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5673 entry->fmt_uuid[0],
5674 entry->fmt_uuid[1],
5675 entry->fmt_uuid[2],
5676 entry->fmt_uuid[3],
5677 entry->fmt_uuid[4],
5678 entry->fmt_uuid[5],
5679 entry->fmt_uuid[6],
5680 entry->fmt_uuid[7],
5681 entry->fmt_uuid[8],
5682 entry->fmt_uuid[9],
5683 entry->fmt_uuid[10],
5684 entry->fmt_uuid[11],
5685 entry->fmt_uuid[12],
5686 entry->fmt_uuid[13],
5687 entry->fmt_uuid[14],
5688 entry->fmt_uuid[15],
5689 entry->fmt_name);
5690 }
5691 spin_unlock(&pfm_buffer_fmt_lock);
5692
5693 }
5694
5695 static int
5696 pfm_proc_show(struct seq_file *m, void *v)
5697 {
5698 unsigned long psr;
5699 unsigned int i;
5700 int cpu;
5701
5702 if (v == PFM_PROC_SHOW_HEADER) {
5703 pfm_proc_show_header(m);
5704 return 0;
5705 }
5706
5707 /* show info for CPU (v - 1) */
5708
5709 cpu = (long)v - 1;
5710 seq_printf(m,
5711 "CPU%-2d overflow intrs : %lu\n"
5712 "CPU%-2d overflow cycles : %lu\n"
5713 "CPU%-2d overflow min : %lu\n"
5714 "CPU%-2d overflow max : %lu\n"
5715 "CPU%-2d smpl handler calls : %lu\n"
5716 "CPU%-2d smpl handler cycles : %lu\n"
5717 "CPU%-2d spurious intrs : %lu\n"
5718 "CPU%-2d replay intrs : %lu\n"
5719 "CPU%-2d syst_wide : %d\n"
5720 "CPU%-2d dcr_pp : %d\n"
5721 "CPU%-2d exclude idle : %d\n"
5722 "CPU%-2d owner : %d\n"
5723 "CPU%-2d context : %p\n"
5724 "CPU%-2d activations : %lu\n",
5725 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5726 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5727 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5728 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5729 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5730 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5731 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5732 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5733 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5734 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5735 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5736 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5737 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5738 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5739
5740 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5741
5742 psr = pfm_get_psr();
5743
5744 ia64_srlz_d();
5745
5746 seq_printf(m,
5747 "CPU%-2d psr : 0x%lx\n"
5748 "CPU%-2d pmc0 : 0x%lx\n",
5749 cpu, psr,
5750 cpu, ia64_get_pmc(0));
5751
5752 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5753 if (PMC_IS_COUNTING(i) == 0) continue;
5754 seq_printf(m,
5755 "CPU%-2d pmc%u : 0x%lx\n"
5756 "CPU%-2d pmd%u : 0x%lx\n",
5757 cpu, i, ia64_get_pmc(i),
5758 cpu, i, ia64_get_pmd(i));
5759 }
5760 }
5761 return 0;
5762 }
5763
5764 const struct seq_operations pfm_seq_ops = {
5765 .start = pfm_proc_start,
5766 .next = pfm_proc_next,
5767 .stop = pfm_proc_stop,
5768 .show = pfm_proc_show
5769 };
5770
5771 static int
5772 pfm_proc_open(struct inode *inode, struct file *file)
5773 {
5774 return seq_open(file, &pfm_seq_ops);
5775 }
5776
5777
5778 /*
5779 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5780 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5781 * is active or inactive based on mode. We must rely on the value in
5782 * local_cpu_data->pfm_syst_info
5783 */
5784 void
5785 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5786 {
5787 struct pt_regs *regs;
5788 unsigned long dcr;
5789 unsigned long dcr_pp;
5790
5791 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5792
5793 /*
5794 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5795 * on every CPU, so we can rely on the pid to identify the idle task.
5796 */
5797 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5798 regs = task_pt_regs(task);
5799 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5800 return;
5801 }
5802 /*
5803 * if monitoring has started
5804 */
5805 if (dcr_pp) {
5806 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5807 /*
5808 * context switching in?
5809 */
5810 if (is_ctxswin) {
5811 /* mask monitoring for the idle task */
5812 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5813 pfm_clear_psr_pp();
5814 ia64_srlz_i();
5815 return;
5816 }
5817 /*
5818 * context switching out
5819 * restore monitoring for next task
5820 *
5821 * Due to inlining this odd if-then-else construction generates
5822 * better code.
5823 */
5824 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5825 pfm_set_psr_pp();
5826 ia64_srlz_i();
5827 }
5828 }
5829
5830 #ifdef CONFIG_SMP
5831
5832 static void
5833 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5834 {
5835 struct task_struct *task = ctx->ctx_task;
5836
5837 ia64_psr(regs)->up = 0;
5838 ia64_psr(regs)->sp = 1;
5839
5840 if (GET_PMU_OWNER() == task) {
5841 DPRINT(("cleared ownership for [%d]\n",
5842 task_pid_nr(ctx->ctx_task)));
5843 SET_PMU_OWNER(NULL, NULL);
5844 }
5845
5846 /*
5847 * disconnect the task from the context and vice-versa
5848 */
5849 PFM_SET_WORK_PENDING(task, 0);
5850
5851 task->thread.pfm_context = NULL;
5852 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5853
5854 DPRINT(("force cleanup for [%d]\n", task_pid_nr(task)));
5855 }
5856
5857
5858 /*
5859 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5860 */
5861 void
5862 pfm_save_regs(struct task_struct *task)
5863 {
5864 pfm_context_t *ctx;
5865 unsigned long flags;
5866 u64 psr;
5867
5868
5869 ctx = PFM_GET_CTX(task);
5870 if (ctx == NULL) return;
5871
5872 /*
5873 * we always come here with interrupts ALREADY disabled by
5874 * the scheduler. So we simply need to protect against concurrent
5875 * access, not CPU concurrency.
5876 */
5877 flags = pfm_protect_ctx_ctxsw(ctx);
5878
5879 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5880 struct pt_regs *regs = task_pt_regs(task);
5881
5882 pfm_clear_psr_up();
5883
5884 pfm_force_cleanup(ctx, regs);
5885
5886 BUG_ON(ctx->ctx_smpl_hdr);
5887
5888 pfm_unprotect_ctx_ctxsw(ctx, flags);
5889
5890 pfm_context_free(ctx);
5891 return;
5892 }
5893
5894 /*
5895 * save current PSR: needed because we modify it
5896 */
5897 ia64_srlz_d();
5898 psr = pfm_get_psr();
5899
5900 BUG_ON(psr & (IA64_PSR_I));
5901
5902 /*
5903 * stop monitoring:
5904 * This is the last instruction which may generate an overflow
5905 *
5906 * We do not need to set psr.sp because, it is irrelevant in kernel.
5907 * It will be restored from ipsr when going back to user level
5908 */
5909 pfm_clear_psr_up();
5910
5911 /*
5912 * keep a copy of psr.up (for reload)
5913 */
5914 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5915
5916 /*
5917 * release ownership of this PMU.
5918 * PM interrupts are masked, so nothing
5919 * can happen.
5920 */
5921 SET_PMU_OWNER(NULL, NULL);
5922
5923 /*
5924 * we systematically save the PMD as we have no
5925 * guarantee we will be schedule at that same
5926 * CPU again.
5927 */
5928 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5929
5930 /*
5931 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5932 * we will need it on the restore path to check
5933 * for pending overflow.
5934 */
5935 ctx->th_pmcs[0] = ia64_get_pmc(0);
5936
5937 /*
5938 * unfreeze PMU if had pending overflows
5939 */
5940 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5941
5942 /*
5943 * finally, allow context access.
5944 * interrupts will still be masked after this call.
5945 */
5946 pfm_unprotect_ctx_ctxsw(ctx, flags);
5947 }
5948
5949 #else /* !CONFIG_SMP */
5950 void
5951 pfm_save_regs(struct task_struct *task)
5952 {
5953 pfm_context_t *ctx;
5954 u64 psr;
5955
5956 ctx = PFM_GET_CTX(task);
5957 if (ctx == NULL) return;
5958
5959 /*
5960 * save current PSR: needed because we modify it
5961 */
5962 psr = pfm_get_psr();
5963
5964 BUG_ON(psr & (IA64_PSR_I));
5965
5966 /*
5967 * stop monitoring:
5968 * This is the last instruction which may generate an overflow
5969 *
5970 * We do not need to set psr.sp because, it is irrelevant in kernel.
5971 * It will be restored from ipsr when going back to user level
5972 */
5973 pfm_clear_psr_up();
5974
5975 /*
5976 * keep a copy of psr.up (for reload)
5977 */
5978 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5979 }
5980
5981 static void
5982 pfm_lazy_save_regs (struct task_struct *task)
5983 {
5984 pfm_context_t *ctx;
5985 unsigned long flags;
5986
5987 { u64 psr = pfm_get_psr();
5988 BUG_ON(psr & IA64_PSR_UP);
5989 }
5990
5991 ctx = PFM_GET_CTX(task);
5992
5993 /*
5994 * we need to mask PMU overflow here to
5995 * make sure that we maintain pmc0 until
5996 * we save it. overflow interrupts are
5997 * treated as spurious if there is no
5998 * owner.
5999 *
6000 * XXX: I don't think this is necessary
6001 */
6002 PROTECT_CTX(ctx,flags);
6003
6004 /*
6005 * release ownership of this PMU.
6006 * must be done before we save the registers.
6007 *
6008 * after this call any PMU interrupt is treated
6009 * as spurious.
6010 */
6011 SET_PMU_OWNER(NULL, NULL);
6012
6013 /*
6014 * save all the pmds we use
6015 */
6016 pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6017
6018 /*
6019 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6020 * it is needed to check for pended overflow
6021 * on the restore path
6022 */
6023 ctx->th_pmcs[0] = ia64_get_pmc(0);
6024
6025 /*
6026 * unfreeze PMU if had pending overflows
6027 */
6028 if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6029
6030 /*
6031 * now get can unmask PMU interrupts, they will
6032 * be treated as purely spurious and we will not
6033 * lose any information
6034 */
6035 UNPROTECT_CTX(ctx,flags);
6036 }
6037 #endif /* CONFIG_SMP */
6038
6039 #ifdef CONFIG_SMP
6040 /*
6041 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6042 */
6043 void
6044 pfm_load_regs (struct task_struct *task)
6045 {
6046 pfm_context_t *ctx;
6047 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6048 unsigned long flags;
6049 u64 psr, psr_up;
6050 int need_irq_resend;
6051
6052 ctx = PFM_GET_CTX(task);
6053 if (unlikely(ctx == NULL)) return;
6054
6055 BUG_ON(GET_PMU_OWNER());
6056
6057 /*
6058 * possible on unload
6059 */
6060 if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6061
6062 /*
6063 * we always come here with interrupts ALREADY disabled by
6064 * the scheduler. So we simply need to protect against concurrent
6065 * access, not CPU concurrency.
6066 */
6067 flags = pfm_protect_ctx_ctxsw(ctx);
6068 psr = pfm_get_psr();
6069
6070 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6071
6072 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6073 BUG_ON(psr & IA64_PSR_I);
6074
6075 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6076 struct pt_regs *regs = task_pt_regs(task);
6077
6078 BUG_ON(ctx->ctx_smpl_hdr);
6079
6080 pfm_force_cleanup(ctx, regs);
6081
6082 pfm_unprotect_ctx_ctxsw(ctx, flags);
6083
6084 /*
6085 * this one (kmalloc'ed) is fine with interrupts disabled
6086 */
6087 pfm_context_free(ctx);
6088
6089 return;
6090 }
6091
6092 /*
6093 * we restore ALL the debug registers to avoid picking up
6094 * stale state.
6095 */
6096 if (ctx->ctx_fl_using_dbreg) {
6097 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6098 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6099 }
6100 /*
6101 * retrieve saved psr.up
6102 */
6103 psr_up = ctx->ctx_saved_psr_up;
6104
6105 /*
6106 * if we were the last user of the PMU on that CPU,
6107 * then nothing to do except restore psr
6108 */
6109 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6110
6111 /*
6112 * retrieve partial reload masks (due to user modifications)
6113 */
6114 pmc_mask = ctx->ctx_reload_pmcs[0];
6115 pmd_mask = ctx->ctx_reload_pmds[0];
6116
6117 } else {
6118 /*
6119 * To avoid leaking information to the user level when psr.sp=0,
6120 * we must reload ALL implemented pmds (even the ones we don't use).
6121 * In the kernel we only allow PFM_READ_PMDS on registers which
6122 * we initialized or requested (sampling) so there is no risk there.
6123 */
6124 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6125
6126 /*
6127 * ALL accessible PMCs are systematically reloaded, unused registers
6128 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6129 * up stale configuration.
6130 *
6131 * PMC0 is never in the mask. It is always restored separately.
6132 */
6133 pmc_mask = ctx->ctx_all_pmcs[0];
6134 }
6135 /*
6136 * when context is MASKED, we will restore PMC with plm=0
6137 * and PMD with stale information, but that's ok, nothing
6138 * will be captured.
6139 *
6140 * XXX: optimize here
6141 */
6142 if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6143 if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6144
6145 /*
6146 * check for pending overflow at the time the state
6147 * was saved.
6148 */
6149 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6150 /*
6151 * reload pmc0 with the overflow information
6152 * On McKinley PMU, this will trigger a PMU interrupt
6153 */
6154 ia64_set_pmc(0, ctx->th_pmcs[0]);
6155 ia64_srlz_d();
6156 ctx->th_pmcs[0] = 0UL;
6157
6158 /*
6159 * will replay the PMU interrupt
6160 */
6161 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6162
6163 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6164 }
6165
6166 /*
6167 * we just did a reload, so we reset the partial reload fields
6168 */
6169 ctx->ctx_reload_pmcs[0] = 0UL;
6170 ctx->ctx_reload_pmds[0] = 0UL;
6171
6172 SET_LAST_CPU(ctx, smp_processor_id());
6173
6174 /*
6175 * dump activation value for this PMU
6176 */
6177 INC_ACTIVATION();
6178 /*
6179 * record current activation for this context
6180 */
6181 SET_ACTIVATION(ctx);
6182
6183 /*
6184 * establish new ownership.
6185 */
6186 SET_PMU_OWNER(task, ctx);
6187
6188 /*
6189 * restore the psr.up bit. measurement
6190 * is active again.
6191 * no PMU interrupt can happen at this point
6192 * because we still have interrupts disabled.
6193 */
6194 if (likely(psr_up)) pfm_set_psr_up();
6195
6196 /*
6197 * allow concurrent access to context
6198 */
6199 pfm_unprotect_ctx_ctxsw(ctx, flags);
6200 }
6201 #else /* !CONFIG_SMP */
6202 /*
6203 * reload PMU state for UP kernels
6204 * in 2.5 we come here with interrupts disabled
6205 */
6206 void
6207 pfm_load_regs (struct task_struct *task)
6208 {
6209 pfm_context_t *ctx;
6210 struct task_struct *owner;
6211 unsigned long pmd_mask, pmc_mask;
6212 u64 psr, psr_up;
6213 int need_irq_resend;
6214
6215 owner = GET_PMU_OWNER();
6216 ctx = PFM_GET_CTX(task);
6217 psr = pfm_get_psr();
6218
6219 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6220 BUG_ON(psr & IA64_PSR_I);
6221
6222 /*
6223 * we restore ALL the debug registers to avoid picking up
6224 * stale state.
6225 *
6226 * This must be done even when the task is still the owner
6227 * as the registers may have been modified via ptrace()
6228 * (not perfmon) by the previous task.
6229 */
6230 if (ctx->ctx_fl_using_dbreg) {
6231 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6232 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6233 }
6234
6235 /*
6236 * retrieved saved psr.up
6237 */
6238 psr_up = ctx->ctx_saved_psr_up;
6239 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6240
6241 /*
6242 * short path, our state is still there, just
6243 * need to restore psr and we go
6244 *
6245 * we do not touch either PMC nor PMD. the psr is not touched
6246 * by the overflow_handler. So we are safe w.r.t. to interrupt
6247 * concurrency even without interrupt masking.
6248 */
6249 if (likely(owner == task)) {
6250 if (likely(psr_up)) pfm_set_psr_up();
6251 return;
6252 }
6253
6254 /*
6255 * someone else is still using the PMU, first push it out and
6256 * then we'll be able to install our stuff !
6257 *
6258 * Upon return, there will be no owner for the current PMU
6259 */
6260 if (owner) pfm_lazy_save_regs(owner);
6261
6262 /*
6263 * To avoid leaking information to the user level when psr.sp=0,
6264 * we must reload ALL implemented pmds (even the ones we don't use).
6265 * In the kernel we only allow PFM_READ_PMDS on registers which
6266 * we initialized or requested (sampling) so there is no risk there.
6267 */
6268 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6269
6270 /*
6271 * ALL accessible PMCs are systematically reloaded, unused registers
6272 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6273 * up stale configuration.
6274 *
6275 * PMC0 is never in the mask. It is always restored separately
6276 */
6277 pmc_mask = ctx->ctx_all_pmcs[0];
6278
6279 pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6280 pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6281
6282 /*
6283 * check for pending overflow at the time the state
6284 * was saved.
6285 */
6286 if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6287 /*
6288 * reload pmc0 with the overflow information
6289 * On McKinley PMU, this will trigger a PMU interrupt
6290 */
6291 ia64_set_pmc(0, ctx->th_pmcs[0]);
6292 ia64_srlz_d();
6293
6294 ctx->th_pmcs[0] = 0UL;
6295
6296 /*
6297 * will replay the PMU interrupt
6298 */
6299 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6300
6301 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6302 }
6303
6304 /*
6305 * establish new ownership.
6306 */
6307 SET_PMU_OWNER(task, ctx);
6308
6309 /*
6310 * restore the psr.up bit. measurement
6311 * is active again.
6312 * no PMU interrupt can happen at this point
6313 * because we still have interrupts disabled.
6314 */
6315 if (likely(psr_up)) pfm_set_psr_up();
6316 }
6317 #endif /* CONFIG_SMP */
6318
6319 /*
6320 * this function assumes monitoring is stopped
6321 */
6322 static void
6323 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6324 {
6325 u64 pmc0;
6326 unsigned long mask2, val, pmd_val, ovfl_val;
6327 int i, can_access_pmu = 0;
6328 int is_self;
6329
6330 /*
6331 * is the caller the task being monitored (or which initiated the
6332 * session for system wide measurements)
6333 */
6334 is_self = ctx->ctx_task == task ? 1 : 0;
6335
6336 /*
6337 * can access PMU is task is the owner of the PMU state on the current CPU
6338 * or if we are running on the CPU bound to the context in system-wide mode
6339 * (that is not necessarily the task the context is attached to in this mode).
6340 * In system-wide we always have can_access_pmu true because a task running on an
6341 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6342 */
6343 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6344 if (can_access_pmu) {
6345 /*
6346 * Mark the PMU as not owned
6347 * This will cause the interrupt handler to do nothing in case an overflow
6348 * interrupt was in-flight
6349 * This also guarantees that pmc0 will contain the final state
6350 * It virtually gives us full control on overflow processing from that point
6351 * on.
6352 */
6353 SET_PMU_OWNER(NULL, NULL);
6354 DPRINT(("releasing ownership\n"));
6355
6356 /*
6357 * read current overflow status:
6358 *
6359 * we are guaranteed to read the final stable state
6360 */
6361 ia64_srlz_d();
6362 pmc0 = ia64_get_pmc(0); /* slow */
6363
6364 /*
6365 * reset freeze bit, overflow status information destroyed
6366 */
6367 pfm_unfreeze_pmu();
6368 } else {
6369 pmc0 = ctx->th_pmcs[0];
6370 /*
6371 * clear whatever overflow status bits there were
6372 */
6373 ctx->th_pmcs[0] = 0;
6374 }
6375 ovfl_val = pmu_conf->ovfl_val;
6376 /*
6377 * we save all the used pmds
6378 * we take care of overflows for counting PMDs
6379 *
6380 * XXX: sampling situation is not taken into account here
6381 */
6382 mask2 = ctx->ctx_used_pmds[0];
6383
6384 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6385
6386 for (i = 0; mask2; i++, mask2>>=1) {
6387
6388 /* skip non used pmds */
6389 if ((mask2 & 0x1) == 0) continue;
6390
6391 /*
6392 * can access PMU always true in system wide mode
6393 */
6394 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6395
6396 if (PMD_IS_COUNTING(i)) {
6397 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6398 task_pid_nr(task),
6399 i,
6400 ctx->ctx_pmds[i].val,
6401 val & ovfl_val));
6402
6403 /*
6404 * we rebuild the full 64 bit value of the counter
6405 */
6406 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6407
6408 /*
6409 * now everything is in ctx_pmds[] and we need
6410 * to clear the saved context from save_regs() such that
6411 * pfm_read_pmds() gets the correct value
6412 */
6413 pmd_val = 0UL;
6414
6415 /*
6416 * take care of overflow inline
6417 */
6418 if (pmc0 & (1UL << i)) {
6419 val += 1 + ovfl_val;
6420 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6421 }
6422 }
6423
6424 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6425
6426 if (is_self) ctx->th_pmds[i] = pmd_val;
6427
6428 ctx->ctx_pmds[i].val = val;
6429 }
6430 }
6431
6432 static struct irqaction perfmon_irqaction = {
6433 .handler = pfm_interrupt_handler,
6434 .flags = IRQF_DISABLED,
6435 .name = "perfmon"
6436 };
6437
6438 static void
6439 pfm_alt_save_pmu_state(void *data)
6440 {
6441 struct pt_regs *regs;
6442
6443 regs = task_pt_regs(current);
6444
6445 DPRINT(("called\n"));
6446
6447 /*
6448 * should not be necessary but
6449 * let's take not risk
6450 */
6451 pfm_clear_psr_up();
6452 pfm_clear_psr_pp();
6453 ia64_psr(regs)->pp = 0;
6454
6455 /*
6456 * This call is required
6457 * May cause a spurious interrupt on some processors
6458 */
6459 pfm_freeze_pmu();
6460
6461 ia64_srlz_d();
6462 }
6463
6464 void
6465 pfm_alt_restore_pmu_state(void *data)
6466 {
6467 struct pt_regs *regs;
6468
6469 regs = task_pt_regs(current);
6470
6471 DPRINT(("called\n"));
6472
6473 /*
6474 * put PMU back in state expected
6475 * by perfmon
6476 */
6477 pfm_clear_psr_up();
6478 pfm_clear_psr_pp();
6479 ia64_psr(regs)->pp = 0;
6480
6481 /*
6482 * perfmon runs with PMU unfrozen at all times
6483 */
6484 pfm_unfreeze_pmu();
6485
6486 ia64_srlz_d();
6487 }
6488
6489 int
6490 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6491 {
6492 int ret, i;
6493 int reserve_cpu;
6494
6495 /* some sanity checks */
6496 if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6497
6498 /* do the easy test first */
6499 if (pfm_alt_intr_handler) return -EBUSY;
6500
6501 /* one at a time in the install or remove, just fail the others */
6502 if (!spin_trylock(&pfm_alt_install_check)) {
6503 return -EBUSY;
6504 }
6505
6506 /* reserve our session */
6507 for_each_online_cpu(reserve_cpu) {
6508 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6509 if (ret) goto cleanup_reserve;
6510 }
6511
6512 /* save the current system wide pmu states */
6513 ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6514 if (ret) {
6515 DPRINT(("on_each_cpu() failed: %d\n", ret));
6516 goto cleanup_reserve;
6517 }
6518
6519 /* officially change to the alternate interrupt handler */
6520 pfm_alt_intr_handler = hdl;
6521
6522 spin_unlock(&pfm_alt_install_check);
6523
6524 return 0;
6525
6526 cleanup_reserve:
6527 for_each_online_cpu(i) {
6528 /* don't unreserve more than we reserved */
6529 if (i >= reserve_cpu) break;
6530
6531 pfm_unreserve_session(NULL, 1, i);
6532 }
6533
6534 spin_unlock(&pfm_alt_install_check);
6535
6536 return ret;
6537 }
6538 EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6539
6540 int
6541 pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6542 {
6543 int i;
6544 int ret;
6545
6546 if (hdl == NULL) return -EINVAL;
6547
6548 /* cannot remove someone else's handler! */
6549 if (pfm_alt_intr_handler != hdl) return -EINVAL;
6550
6551 /* one at a time in the install or remove, just fail the others */
6552 if (!spin_trylock(&pfm_alt_install_check)) {
6553 return -EBUSY;
6554 }
6555
6556 pfm_alt_intr_handler = NULL;
6557
6558 ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6559 if (ret) {
6560 DPRINT(("on_each_cpu() failed: %d\n", ret));
6561 }
6562
6563 for_each_online_cpu(i) {
6564 pfm_unreserve_session(NULL, 1, i);
6565 }
6566
6567 spin_unlock(&pfm_alt_install_check);
6568
6569 return 0;
6570 }
6571 EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6572
6573 /*
6574 * perfmon initialization routine, called from the initcall() table
6575 */
6576 static int init_pfm_fs(void);
6577
6578 static int __init
6579 pfm_probe_pmu(void)
6580 {
6581 pmu_config_t **p;
6582 int family;
6583
6584 family = local_cpu_data->family;
6585 p = pmu_confs;
6586
6587 while(*p) {
6588 if ((*p)->probe) {
6589 if ((*p)->probe() == 0) goto found;
6590 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6591 goto found;
6592 }
6593 p++;
6594 }
6595 return -1;
6596 found:
6597 pmu_conf = *p;
6598 return 0;
6599 }
6600
6601 static const struct file_operations pfm_proc_fops = {
6602 .open = pfm_proc_open,
6603 .read = seq_read,
6604 .llseek = seq_lseek,
6605 .release = seq_release,
6606 };
6607
6608 int __init
6609 pfm_init(void)
6610 {
6611 unsigned int n, n_counters, i;
6612
6613 printk("perfmon: version %u.%u IRQ %u\n",
6614 PFM_VERSION_MAJ,
6615 PFM_VERSION_MIN,
6616 IA64_PERFMON_VECTOR);
6617
6618 if (pfm_probe_pmu()) {
6619 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6620 local_cpu_data->family);
6621 return -ENODEV;
6622 }
6623
6624 /*
6625 * compute the number of implemented PMD/PMC from the
6626 * description tables
6627 */
6628 n = 0;
6629 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6630 if (PMC_IS_IMPL(i) == 0) continue;
6631 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6632 n++;
6633 }
6634 pmu_conf->num_pmcs = n;
6635
6636 n = 0; n_counters = 0;
6637 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6638 if (PMD_IS_IMPL(i) == 0) continue;
6639 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6640 n++;
6641 if (PMD_IS_COUNTING(i)) n_counters++;
6642 }
6643 pmu_conf->num_pmds = n;
6644 pmu_conf->num_counters = n_counters;
6645
6646 /*
6647 * sanity checks on the number of debug registers
6648 */
6649 if (pmu_conf->use_rr_dbregs) {
6650 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6651 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6652 pmu_conf = NULL;
6653 return -1;
6654 }
6655 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6656 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6657 pmu_conf = NULL;
6658 return -1;
6659 }
6660 }
6661
6662 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6663 pmu_conf->pmu_name,
6664 pmu_conf->num_pmcs,
6665 pmu_conf->num_pmds,
6666 pmu_conf->num_counters,
6667 ffz(pmu_conf->ovfl_val));
6668
6669 /* sanity check */
6670 if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6671 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6672 pmu_conf = NULL;
6673 return -1;
6674 }
6675
6676 /*
6677 * create /proc/perfmon (mostly for debugging purposes)
6678 */
6679 perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6680 if (perfmon_dir == NULL) {
6681 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6682 pmu_conf = NULL;
6683 return -1;
6684 }
6685
6686 /*
6687 * create /proc/sys/kernel/perfmon (for debugging purposes)
6688 */
6689 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6690
6691 /*
6692 * initialize all our spinlocks
6693 */
6694 spin_lock_init(&pfm_sessions.pfs_lock);
6695 spin_lock_init(&pfm_buffer_fmt_lock);
6696
6697 init_pfm_fs();
6698
6699 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6700
6701 return 0;
6702 }
6703
6704 __initcall(pfm_init);
6705
6706 /*
6707 * this function is called before pfm_init()
6708 */
6709 void
6710 pfm_init_percpu (void)
6711 {
6712 static int first_time=1;
6713 /*
6714 * make sure no measurement is active
6715 * (may inherit programmed PMCs from EFI).
6716 */
6717 pfm_clear_psr_pp();
6718 pfm_clear_psr_up();
6719
6720 /*
6721 * we run with the PMU not frozen at all times
6722 */
6723 pfm_unfreeze_pmu();
6724
6725 if (first_time) {
6726 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6727 first_time=0;
6728 }
6729
6730 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6731 ia64_srlz_d();
6732 }
6733
6734 /*
6735 * used for debug purposes only
6736 */
6737 void
6738 dump_pmu_state(const char *from)
6739 {
6740 struct task_struct *task;
6741 struct pt_regs *regs;
6742 pfm_context_t *ctx;
6743 unsigned long psr, dcr, info, flags;
6744 int i, this_cpu;
6745
6746 local_irq_save(flags);
6747
6748 this_cpu = smp_processor_id();
6749 regs = task_pt_regs(current);
6750 info = PFM_CPUINFO_GET();
6751 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6752
6753 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6754 local_irq_restore(flags);
6755 return;
6756 }
6757
6758 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6759 this_cpu,
6760 from,
6761 task_pid_nr(current),
6762 regs->cr_iip,
6763 current->comm);
6764
6765 task = GET_PMU_OWNER();
6766 ctx = GET_PMU_CTX();
6767
6768 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6769
6770 psr = pfm_get_psr();
6771
6772 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",
6773 this_cpu,
6774 ia64_get_pmc(0),
6775 psr & IA64_PSR_PP ? 1 : 0,
6776 psr & IA64_PSR_UP ? 1 : 0,
6777 dcr & IA64_DCR_PP ? 1 : 0,
6778 info,
6779 ia64_psr(regs)->up,
6780 ia64_psr(regs)->pp);
6781
6782 ia64_psr(regs)->up = 0;
6783 ia64_psr(regs)->pp = 0;
6784
6785 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6786 if (PMC_IS_IMPL(i) == 0) continue;
6787 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]);
6788 }
6789
6790 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6791 if (PMD_IS_IMPL(i) == 0) continue;
6792 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]);
6793 }
6794
6795 if (ctx) {
6796 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6797 this_cpu,
6798 ctx->ctx_state,
6799 ctx->ctx_smpl_vaddr,
6800 ctx->ctx_smpl_hdr,
6801 ctx->ctx_msgq_head,
6802 ctx->ctx_msgq_tail,
6803 ctx->ctx_saved_psr_up);
6804 }
6805 local_irq_restore(flags);
6806 }
6807
6808 /*
6809 * called from process.c:copy_thread(). task is new child.
6810 */
6811 void
6812 pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6813 {
6814 struct thread_struct *thread;
6815
6816 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6817
6818 thread = &task->thread;
6819
6820 /*
6821 * cut links inherited from parent (current)
6822 */
6823 thread->pfm_context = NULL;
6824
6825 PFM_SET_WORK_PENDING(task, 0);
6826
6827 /*
6828 * the psr bits are already set properly in copy_threads()
6829 */
6830 }
6831 #else /* !CONFIG_PERFMON */
6832 asmlinkage long
6833 sys_perfmonctl (int fd, int cmd, void *arg, int count)
6834 {
6835 return -ENOSYS;
6836 }
6837 #endif /* CONFIG_PERFMON */