aboutsummaryrefslogtreecommitdiffstats
path: root/arch/ia64/kernel/perfmon.c
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 18:20:36 -0400
committerLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 18:20:36 -0400
commit1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 (patch)
tree0bba044c4ce775e45a88a51686b5d9f90697ea9d /arch/ia64/kernel/perfmon.c
Linux-2.6.12-rc2v2.6.12-rc2
Initial git repository build. I'm not bothering with the full history, even though we have it. We can create a separate "historical" git archive of that later if we want to, and in the meantime it's about 3.2GB when imported into git - space that would just make the early git days unnecessarily complicated, when we don't have a lot of good infrastructure for it. Let it rip!
Diffstat (limited to 'arch/ia64/kernel/perfmon.c')
-rw-r--r--arch/ia64/kernel/perfmon.c6676
1 files changed, 6676 insertions, 0 deletions
diff --git a/arch/ia64/kernel/perfmon.c b/arch/ia64/kernel/perfmon.c
new file mode 100644
index 000000000000..71147be3279c
--- /dev/null
+++ b/arch/ia64/kernel/perfmon.c
@@ -0,0 +1,6676 @@
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) 1999-2003, 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/config.h>
23#include <linux/module.h>
24#include <linux/kernel.h>
25#include <linux/sched.h>
26#include <linux/interrupt.h>
27#include <linux/smp_lock.h>
28#include <linux/proc_fs.h>
29#include <linux/seq_file.h>
30#include <linux/init.h>
31#include <linux/vmalloc.h>
32#include <linux/mm.h>
33#include <linux/sysctl.h>
34#include <linux/list.h>
35#include <linux/file.h>
36#include <linux/poll.h>
37#include <linux/vfs.h>
38#include <linux/pagemap.h>
39#include <linux/mount.h>
40#include <linux/version.h>
41#include <linux/bitops.h>
42
43#include <asm/errno.h>
44#include <asm/intrinsics.h>
45#include <asm/page.h>
46#include <asm/perfmon.h>
47#include <asm/processor.h>
48#include <asm/signal.h>
49#include <asm/system.h>
50#include <asm/uaccess.h>
51#include <asm/delay.h>
52
53#ifdef CONFIG_PERFMON
54/*
55 * perfmon context state
56 */
57#define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */
58#define PFM_CTX_LOADED 2 /* context is loaded onto a task */
59#define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */
60#define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */
61
62#define PFM_INVALID_ACTIVATION (~0UL)
63
64/*
65 * depth of message queue
66 */
67#define PFM_MAX_MSGS 32
68#define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
69
70/*
71 * type of a PMU register (bitmask).
72 * bitmask structure:
73 * bit0 : register implemented
74 * bit1 : end marker
75 * bit2-3 : reserved
76 * bit4 : pmc has pmc.pm
77 * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter
78 * bit6-7 : register type
79 * bit8-31: reserved
80 */
81#define PFM_REG_NOTIMPL 0x0 /* not implemented at all */
82#define PFM_REG_IMPL 0x1 /* register implemented */
83#define PFM_REG_END 0x2 /* end marker */
84#define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
85#define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
86#define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */
87#define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */
88#define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
89
90#define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END)
91#define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END)
92
93#define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY)
94
95/* i assumed unsigned */
96#define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
97#define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
98
99/* XXX: these assume that register i is implemented */
100#define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
101#define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
102#define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR)
103#define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL)
104
105#define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value
106#define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask
107#define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0]
108#define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0]
109
110#define PFM_NUM_IBRS IA64_NUM_DBG_REGS
111#define PFM_NUM_DBRS IA64_NUM_DBG_REGS
112
113#define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0)
114#define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling)
115#define PFM_CTX_TASK(h) (h)->ctx_task
116
117#define PMU_PMC_OI 5 /* position of pmc.oi bit */
118
119/* XXX: does not support more than 64 PMDs */
120#define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
121#define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
122
123#define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
124
125#define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
126#define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
127#define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
128#define PFM_CODE_RR 0 /* requesting code range restriction */
129#define PFM_DATA_RR 1 /* requestion data range restriction */
130
131#define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v)
132#define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v)
133#define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info)
134
135#define RDEP(x) (1UL<<(x))
136
137/*
138 * context protection macros
139 * in SMP:
140 * - we need to protect against CPU concurrency (spin_lock)
141 * - we need to protect against PMU overflow interrupts (local_irq_disable)
142 * in UP:
143 * - we need to protect against PMU overflow interrupts (local_irq_disable)
144 *
145 * spin_lock_irqsave()/spin_lock_irqrestore():
146 * in SMP: local_irq_disable + spin_lock
147 * in UP : local_irq_disable
148 *
149 * spin_lock()/spin_lock():
150 * in UP : removed automatically
151 * in SMP: protect against context accesses from other CPU. interrupts
152 * are not masked. This is useful for the PMU interrupt handler
153 * because we know we will not get PMU concurrency in that code.
154 */
155#define PROTECT_CTX(c, f) \
156 do { \
157 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \
158 spin_lock_irqsave(&(c)->ctx_lock, f); \
159 DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \
160 } while(0)
161
162#define UNPROTECT_CTX(c, f) \
163 do { \
164 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
165 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
166 } while(0)
167
168#define PROTECT_CTX_NOPRINT(c, f) \
169 do { \
170 spin_lock_irqsave(&(c)->ctx_lock, f); \
171 } while(0)
172
173
174#define UNPROTECT_CTX_NOPRINT(c, f) \
175 do { \
176 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
177 } while(0)
178
179
180#define PROTECT_CTX_NOIRQ(c) \
181 do { \
182 spin_lock(&(c)->ctx_lock); \
183 } while(0)
184
185#define UNPROTECT_CTX_NOIRQ(c) \
186 do { \
187 spin_unlock(&(c)->ctx_lock); \
188 } while(0)
189
190
191#ifdef CONFIG_SMP
192
193#define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)
194#define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++
195#define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION()
196
197#else /* !CONFIG_SMP */
198#define SET_ACTIVATION(t) do {} while(0)
199#define GET_ACTIVATION(t) do {} while(0)
200#define INC_ACTIVATION(t) do {} while(0)
201#endif /* CONFIG_SMP */
202
203#define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
204#define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner)
205#define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx)
206
207#define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
208#define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
209
210#define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
211
212/*
213 * cmp0 must be the value of pmc0
214 */
215#define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL)
216
217#define PFMFS_MAGIC 0xa0b4d889
218
219/*
220 * debugging
221 */
222#define PFM_DEBUGGING 1
223#ifdef PFM_DEBUGGING
224#define DPRINT(a) \
225 do { \
226 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
227 } while (0)
228
229#define DPRINT_ovfl(a) \
230 do { \
231 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
232 } while (0)
233#endif
234
235/*
236 * 64-bit software counter structure
237 *
238 * the next_reset_type is applied to the next call to pfm_reset_regs()
239 */
240typedef struct {
241 unsigned long val; /* virtual 64bit counter value */
242 unsigned long lval; /* last reset value */
243 unsigned long long_reset; /* reset value on sampling overflow */
244 unsigned long short_reset; /* reset value on overflow */
245 unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */
246 unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */
247 unsigned long seed; /* seed for random-number generator */
248 unsigned long mask; /* mask for random-number generator */
249 unsigned int flags; /* notify/do not notify */
250 unsigned long eventid; /* overflow event identifier */
251} pfm_counter_t;
252
253/*
254 * context flags
255 */
256typedef struct {
257 unsigned int block:1; /* when 1, task will blocked on user notifications */
258 unsigned int system:1; /* do system wide monitoring */
259 unsigned int using_dbreg:1; /* using range restrictions (debug registers) */
260 unsigned int is_sampling:1; /* true if using a custom format */
261 unsigned int excl_idle:1; /* exclude idle task in system wide session */
262 unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */
263 unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */
264 unsigned int no_msg:1; /* no message sent on overflow */
265 unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */
266 unsigned int reserved:22;
267} pfm_context_flags_t;
268
269#define PFM_TRAP_REASON_NONE 0x0 /* default value */
270#define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */
271#define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */
272
273
274/*
275 * perfmon context: encapsulates all the state of a monitoring session
276 */
277
278typedef struct pfm_context {
279 spinlock_t ctx_lock; /* context protection */
280
281 pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */
282 unsigned int ctx_state; /* state: active/inactive (no bitfield) */
283
284 struct task_struct *ctx_task; /* task to which context is attached */
285
286 unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
287
288 struct semaphore ctx_restart_sem; /* use for blocking notification mode */
289
290 unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */
291 unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */
292 unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */
293
294 unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */
295 unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */
296 unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */
297
298 unsigned long ctx_pmcs[IA64_NUM_PMC_REGS]; /* saved copies of PMC values */
299
300 unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */
301 unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */
302 unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */
303 unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */
304
305 pfm_counter_t ctx_pmds[IA64_NUM_PMD_REGS]; /* software state for PMDS */
306
307 u64 ctx_saved_psr_up; /* only contains psr.up value */
308
309 unsigned long ctx_last_activation; /* context last activation number for last_cpu */
310 unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */
311 unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */
312
313 int ctx_fd; /* file descriptor used my this context */
314 pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */
315
316 pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */
317 void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */
318 unsigned long ctx_smpl_size; /* size of sampling buffer */
319 void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */
320
321 wait_queue_head_t ctx_msgq_wait;
322 pfm_msg_t ctx_msgq[PFM_MAX_MSGS];
323 int ctx_msgq_head;
324 int ctx_msgq_tail;
325 struct fasync_struct *ctx_async_queue;
326
327 wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */
328} pfm_context_t;
329
330/*
331 * magic number used to verify that structure is really
332 * a perfmon context
333 */
334#define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops)
335
336#define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context)
337
338#ifdef CONFIG_SMP
339#define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v)
340#define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu
341#else
342#define SET_LAST_CPU(ctx, v) do {} while(0)
343#define GET_LAST_CPU(ctx) do {} while(0)
344#endif
345
346
347#define ctx_fl_block ctx_flags.block
348#define ctx_fl_system ctx_flags.system
349#define ctx_fl_using_dbreg ctx_flags.using_dbreg
350#define ctx_fl_is_sampling ctx_flags.is_sampling
351#define ctx_fl_excl_idle ctx_flags.excl_idle
352#define ctx_fl_going_zombie ctx_flags.going_zombie
353#define ctx_fl_trap_reason ctx_flags.trap_reason
354#define ctx_fl_no_msg ctx_flags.no_msg
355#define ctx_fl_can_restart ctx_flags.can_restart
356
357#define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0);
358#define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking
359
360/*
361 * global information about all sessions
362 * mostly used to synchronize between system wide and per-process
363 */
364typedef struct {
365 spinlock_t pfs_lock; /* lock the structure */
366
367 unsigned int pfs_task_sessions; /* number of per task sessions */
368 unsigned int pfs_sys_sessions; /* number of per system wide sessions */
369 unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */
370 unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */
371 struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
372} pfm_session_t;
373
374/*
375 * information about a PMC or PMD.
376 * dep_pmd[]: a bitmask of dependent PMD registers
377 * dep_pmc[]: a bitmask of dependent PMC registers
378 */
379typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
380typedef struct {
381 unsigned int type;
382 int pm_pos;
383 unsigned long default_value; /* power-on default value */
384 unsigned long reserved_mask; /* bitmask of reserved bits */
385 pfm_reg_check_t read_check;
386 pfm_reg_check_t write_check;
387 unsigned long dep_pmd[4];
388 unsigned long dep_pmc[4];
389} pfm_reg_desc_t;
390
391/* assume cnum is a valid monitor */
392#define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
393
394/*
395 * This structure is initialized at boot time and contains
396 * a description of the PMU main characteristics.
397 *
398 * If the probe function is defined, detection is based
399 * on its return value:
400 * - 0 means recognized PMU
401 * - anything else means not supported
402 * When the probe function is not defined, then the pmu_family field
403 * is used and it must match the host CPU family such that:
404 * - cpu->family & config->pmu_family != 0
405 */
406typedef struct {
407 unsigned long ovfl_val; /* overflow value for counters */
408
409 pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */
410 pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */
411
412 unsigned int num_pmcs; /* number of PMCS: computed at init time */
413 unsigned int num_pmds; /* number of PMDS: computed at init time */
414 unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */
415 unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */
416
417 char *pmu_name; /* PMU family name */
418 unsigned int pmu_family; /* cpuid family pattern used to identify pmu */
419 unsigned int flags; /* pmu specific flags */
420 unsigned int num_ibrs; /* number of IBRS: computed at init time */
421 unsigned int num_dbrs; /* number of DBRS: computed at init time */
422 unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */
423 int (*probe)(void); /* customized probe routine */
424 unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */
425} pmu_config_t;
426/*
427 * PMU specific flags
428 */
429#define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */
430
431/*
432 * debug register related type definitions
433 */
434typedef struct {
435 unsigned long ibr_mask:56;
436 unsigned long ibr_plm:4;
437 unsigned long ibr_ig:3;
438 unsigned long ibr_x:1;
439} ibr_mask_reg_t;
440
441typedef struct {
442 unsigned long dbr_mask:56;
443 unsigned long dbr_plm:4;
444 unsigned long dbr_ig:2;
445 unsigned long dbr_w:1;
446 unsigned long dbr_r:1;
447} dbr_mask_reg_t;
448
449typedef union {
450 unsigned long val;
451 ibr_mask_reg_t ibr;
452 dbr_mask_reg_t dbr;
453} dbreg_t;
454
455
456/*
457 * perfmon command descriptions
458 */
459typedef struct {
460 int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
461 char *cmd_name;
462 int cmd_flags;
463 unsigned int cmd_narg;
464 size_t cmd_argsize;
465 int (*cmd_getsize)(void *arg, size_t *sz);
466} pfm_cmd_desc_t;
467
468#define PFM_CMD_FD 0x01 /* command requires a file descriptor */
469#define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */
470#define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */
471#define PFM_CMD_STOP 0x08 /* command does not work on zombie context */
472
473
474#define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name
475#define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
476#define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
477#define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
478#define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
479
480#define PFM_CMD_ARG_MANY -1 /* cannot be zero */
481
482typedef struct {
483 int debug; /* turn on/off debugging via syslog */
484 int debug_ovfl; /* turn on/off debug printk in overflow handler */
485 int fastctxsw; /* turn on/off fast (unsecure) ctxsw */
486 int expert_mode; /* turn on/off value checking */
487 int debug_pfm_read;
488} pfm_sysctl_t;
489
490typedef struct {
491 unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */
492 unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */
493 unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */
494 unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */
495 unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */
496 unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */
497 unsigned long pfm_smpl_handler_calls;
498 unsigned long pfm_smpl_handler_cycles;
499 char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
500} pfm_stats_t;
501
502/*
503 * perfmon internal variables
504 */
505static pfm_stats_t pfm_stats[NR_CPUS];
506static pfm_session_t pfm_sessions; /* global sessions information */
507
508static struct proc_dir_entry *perfmon_dir;
509static pfm_uuid_t pfm_null_uuid = {0,};
510
511static spinlock_t pfm_buffer_fmt_lock;
512static LIST_HEAD(pfm_buffer_fmt_list);
513
514static pmu_config_t *pmu_conf;
515
516/* sysctl() controls */
517static pfm_sysctl_t pfm_sysctl;
518int pfm_debug_var;
519
520static ctl_table pfm_ctl_table[]={
521 {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
522 {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
523 {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
524 {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
525 { 0, },
526};
527static ctl_table pfm_sysctl_dir[] = {
528 {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
529 {0,},
530};
531static ctl_table pfm_sysctl_root[] = {
532 {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
533 {0,},
534};
535static struct ctl_table_header *pfm_sysctl_header;
536
537static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
538static int pfm_flush(struct file *filp);
539
540#define pfm_get_cpu_var(v) __ia64_per_cpu_var(v)
541#define pfm_get_cpu_data(a,b) per_cpu(a, b)
542
543static inline void
544pfm_put_task(struct task_struct *task)
545{
546 if (task != current) put_task_struct(task);
547}
548
549static inline void
550pfm_set_task_notify(struct task_struct *task)
551{
552 struct thread_info *info;
553
554 info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
555 set_bit(TIF_NOTIFY_RESUME, &info->flags);
556}
557
558static inline void
559pfm_clear_task_notify(void)
560{
561 clear_thread_flag(TIF_NOTIFY_RESUME);
562}
563
564static inline void
565pfm_reserve_page(unsigned long a)
566{
567 SetPageReserved(vmalloc_to_page((void *)a));
568}
569static inline void
570pfm_unreserve_page(unsigned long a)
571{
572 ClearPageReserved(vmalloc_to_page((void*)a));
573}
574
575static inline unsigned long
576pfm_protect_ctx_ctxsw(pfm_context_t *x)
577{
578 spin_lock(&(x)->ctx_lock);
579 return 0UL;
580}
581
582static inline unsigned long
583pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
584{
585 spin_unlock(&(x)->ctx_lock);
586}
587
588static inline unsigned int
589pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
590{
591 return do_munmap(mm, addr, len);
592}
593
594static inline unsigned long
595pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
596{
597 return get_unmapped_area(file, addr, len, pgoff, flags);
598}
599
600
601static struct super_block *
602pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
603{
604 return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
605}
606
607static struct file_system_type pfm_fs_type = {
608 .name = "pfmfs",
609 .get_sb = pfmfs_get_sb,
610 .kill_sb = kill_anon_super,
611};
612
613DEFINE_PER_CPU(unsigned long, pfm_syst_info);
614DEFINE_PER_CPU(struct task_struct *, pmu_owner);
615DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
616DEFINE_PER_CPU(unsigned long, pmu_activation_number);
617
618
619/* forward declaration */
620static struct file_operations pfm_file_ops;
621
622/*
623 * forward declarations
624 */
625#ifndef CONFIG_SMP
626static void pfm_lazy_save_regs (struct task_struct *ta);
627#endif
628
629void dump_pmu_state(const char *);
630static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
631
632#include "perfmon_itanium.h"
633#include "perfmon_mckinley.h"
634#include "perfmon_generic.h"
635
636static pmu_config_t *pmu_confs[]={
637 &pmu_conf_mck,
638 &pmu_conf_ita,
639 &pmu_conf_gen, /* must be last */
640 NULL
641};
642
643
644static int pfm_end_notify_user(pfm_context_t *ctx);
645
646static inline void
647pfm_clear_psr_pp(void)
648{
649 ia64_rsm(IA64_PSR_PP);
650 ia64_srlz_i();
651}
652
653static inline void
654pfm_set_psr_pp(void)
655{
656 ia64_ssm(IA64_PSR_PP);
657 ia64_srlz_i();
658}
659
660static inline void
661pfm_clear_psr_up(void)
662{
663 ia64_rsm(IA64_PSR_UP);
664 ia64_srlz_i();
665}
666
667static inline void
668pfm_set_psr_up(void)
669{
670 ia64_ssm(IA64_PSR_UP);
671 ia64_srlz_i();
672}
673
674static inline unsigned long
675pfm_get_psr(void)
676{
677 unsigned long tmp;
678 tmp = ia64_getreg(_IA64_REG_PSR);
679 ia64_srlz_i();
680 return tmp;
681}
682
683static inline void
684pfm_set_psr_l(unsigned long val)
685{
686 ia64_setreg(_IA64_REG_PSR_L, val);
687 ia64_srlz_i();
688}
689
690static inline void
691pfm_freeze_pmu(void)
692{
693 ia64_set_pmc(0,1UL);
694 ia64_srlz_d();
695}
696
697static inline void
698pfm_unfreeze_pmu(void)
699{
700 ia64_set_pmc(0,0UL);
701 ia64_srlz_d();
702}
703
704static inline void
705pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
706{
707 int i;
708
709 for (i=0; i < nibrs; i++) {
710 ia64_set_ibr(i, ibrs[i]);
711 ia64_dv_serialize_instruction();
712 }
713 ia64_srlz_i();
714}
715
716static inline void
717pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
718{
719 int i;
720
721 for (i=0; i < ndbrs; i++) {
722 ia64_set_dbr(i, dbrs[i]);
723 ia64_dv_serialize_data();
724 }
725 ia64_srlz_d();
726}
727
728/*
729 * PMD[i] must be a counter. no check is made
730 */
731static inline unsigned long
732pfm_read_soft_counter(pfm_context_t *ctx, int i)
733{
734 return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
735}
736
737/*
738 * PMD[i] must be a counter. no check is made
739 */
740static inline void
741pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
742{
743 unsigned long ovfl_val = pmu_conf->ovfl_val;
744
745 ctx->ctx_pmds[i].val = val & ~ovfl_val;
746 /*
747 * writing to unimplemented part is ignore, so we do not need to
748 * mask off top part
749 */
750 ia64_set_pmd(i, val & ovfl_val);
751}
752
753static pfm_msg_t *
754pfm_get_new_msg(pfm_context_t *ctx)
755{
756 int idx, next;
757
758 next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
759
760 DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
761 if (next == ctx->ctx_msgq_head) return NULL;
762
763 idx = ctx->ctx_msgq_tail;
764 ctx->ctx_msgq_tail = next;
765
766 DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
767
768 return ctx->ctx_msgq+idx;
769}
770
771static pfm_msg_t *
772pfm_get_next_msg(pfm_context_t *ctx)
773{
774 pfm_msg_t *msg;
775
776 DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
777
778 if (PFM_CTXQ_EMPTY(ctx)) return NULL;
779
780 /*
781 * get oldest message
782 */
783 msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
784
785 /*
786 * and move forward
787 */
788 ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
789
790 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));
791
792 return msg;
793}
794
795static void
796pfm_reset_msgq(pfm_context_t *ctx)
797{
798 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
799 DPRINT(("ctx=%p msgq reset\n", ctx));
800}
801
802static void *
803pfm_rvmalloc(unsigned long size)
804{
805 void *mem;
806 unsigned long addr;
807
808 size = PAGE_ALIGN(size);
809 mem = vmalloc(size);
810 if (mem) {
811 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
812 memset(mem, 0, size);
813 addr = (unsigned long)mem;
814 while (size > 0) {
815 pfm_reserve_page(addr);
816 addr+=PAGE_SIZE;
817 size-=PAGE_SIZE;
818 }
819 }
820 return mem;
821}
822
823static void
824pfm_rvfree(void *mem, unsigned long size)
825{
826 unsigned long addr;
827
828 if (mem) {
829 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
830 addr = (unsigned long) mem;
831 while ((long) size > 0) {
832 pfm_unreserve_page(addr);
833 addr+=PAGE_SIZE;
834 size-=PAGE_SIZE;
835 }
836 vfree(mem);
837 }
838 return;
839}
840
841static pfm_context_t *
842pfm_context_alloc(void)
843{
844 pfm_context_t *ctx;
845
846 /*
847 * allocate context descriptor
848 * must be able to free with interrupts disabled
849 */
850 ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
851 if (ctx) {
852 memset(ctx, 0, sizeof(pfm_context_t));
853 DPRINT(("alloc ctx @%p\n", ctx));
854 }
855 return ctx;
856}
857
858static void
859pfm_context_free(pfm_context_t *ctx)
860{
861 if (ctx) {
862 DPRINT(("free ctx @%p\n", ctx));
863 kfree(ctx);
864 }
865}
866
867static void
868pfm_mask_monitoring(struct task_struct *task)
869{
870 pfm_context_t *ctx = PFM_GET_CTX(task);
871 struct thread_struct *th = &task->thread;
872 unsigned long mask, val, ovfl_mask;
873 int i;
874
875 DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));
876
877 ovfl_mask = pmu_conf->ovfl_val;
878 /*
879 * monitoring can only be masked as a result of a valid
880 * counter overflow. In UP, it means that the PMU still
881 * has an owner. Note that the owner can be different
882 * from the current task. However the PMU state belongs
883 * to the owner.
884 * In SMP, a valid overflow only happens when task is
885 * current. Therefore if we come here, we know that
886 * the PMU state belongs to the current task, therefore
887 * we can access the live registers.
888 *
889 * So in both cases, the live register contains the owner's
890 * state. We can ONLY touch the PMU registers and NOT the PSR.
891 *
892 * As a consequence to this call, the thread->pmds[] array
893 * contains stale information which must be ignored
894 * when context is reloaded AND monitoring is active (see
895 * pfm_restart).
896 */
897 mask = ctx->ctx_used_pmds[0];
898 for (i = 0; mask; i++, mask>>=1) {
899 /* skip non used pmds */
900 if ((mask & 0x1) == 0) continue;
901 val = ia64_get_pmd(i);
902
903 if (PMD_IS_COUNTING(i)) {
904 /*
905 * we rebuild the full 64 bit value of the counter
906 */
907 ctx->ctx_pmds[i].val += (val & ovfl_mask);
908 } else {
909 ctx->ctx_pmds[i].val = val;
910 }
911 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
912 i,
913 ctx->ctx_pmds[i].val,
914 val & ovfl_mask));
915 }
916 /*
917 * mask monitoring by setting the privilege level to 0
918 * we cannot use psr.pp/psr.up for this, it is controlled by
919 * the user
920 *
921 * if task is current, modify actual registers, otherwise modify
922 * thread save state, i.e., what will be restored in pfm_load_regs()
923 */
924 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
925 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
926 if ((mask & 0x1) == 0UL) continue;
927 ia64_set_pmc(i, th->pmcs[i] & ~0xfUL);
928 th->pmcs[i] &= ~0xfUL;
929 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i]));
930 }
931 /*
932 * make all of this visible
933 */
934 ia64_srlz_d();
935}
936
937/*
938 * must always be done with task == current
939 *
940 * context must be in MASKED state when calling
941 */
942static void
943pfm_restore_monitoring(struct task_struct *task)
944{
945 pfm_context_t *ctx = PFM_GET_CTX(task);
946 struct thread_struct *th = &task->thread;
947 unsigned long mask, ovfl_mask;
948 unsigned long psr, val;
949 int i, is_system;
950
951 is_system = ctx->ctx_fl_system;
952 ovfl_mask = pmu_conf->ovfl_val;
953
954 if (task != current) {
955 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
956 return;
957 }
958 if (ctx->ctx_state != PFM_CTX_MASKED) {
959 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
960 task->pid, current->pid, ctx->ctx_state);
961 return;
962 }
963 psr = pfm_get_psr();
964 /*
965 * monitoring is masked via the PMC.
966 * As we restore their value, we do not want each counter to
967 * restart right away. We stop monitoring using the PSR,
968 * restore the PMC (and PMD) and then re-establish the psr
969 * as it was. Note that there can be no pending overflow at
970 * this point, because monitoring was MASKED.
971 *
972 * system-wide session are pinned and self-monitoring
973 */
974 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
975 /* disable dcr pp */
976 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
977 pfm_clear_psr_pp();
978 } else {
979 pfm_clear_psr_up();
980 }
981 /*
982 * first, we restore the PMD
983 */
984 mask = ctx->ctx_used_pmds[0];
985 for (i = 0; mask; i++, mask>>=1) {
986 /* skip non used pmds */
987 if ((mask & 0x1) == 0) continue;
988
989 if (PMD_IS_COUNTING(i)) {
990 /*
991 * we split the 64bit value according to
992 * counter width
993 */
994 val = ctx->ctx_pmds[i].val & ovfl_mask;
995 ctx->ctx_pmds[i].val &= ~ovfl_mask;
996 } else {
997 val = ctx->ctx_pmds[i].val;
998 }
999 ia64_set_pmd(i, val);
1000
1001 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1002 i,
1003 ctx->ctx_pmds[i].val,
1004 val));
1005 }
1006 /*
1007 * restore the PMCs
1008 */
1009 mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1010 for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1011 if ((mask & 0x1) == 0UL) continue;
1012 th->pmcs[i] = ctx->ctx_pmcs[i];
1013 ia64_set_pmc(i, th->pmcs[i]);
1014 DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
1015 }
1016 ia64_srlz_d();
1017
1018 /*
1019 * must restore DBR/IBR because could be modified while masked
1020 * XXX: need to optimize
1021 */
1022 if (ctx->ctx_fl_using_dbreg) {
1023 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1024 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1025 }
1026
1027 /*
1028 * now restore PSR
1029 */
1030 if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1031 /* enable dcr pp */
1032 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1033 ia64_srlz_i();
1034 }
1035 pfm_set_psr_l(psr);
1036}
1037
1038static inline void
1039pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1040{
1041 int i;
1042
1043 ia64_srlz_d();
1044
1045 for (i=0; mask; i++, mask>>=1) {
1046 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1047 }
1048}
1049
1050/*
1051 * reload from thread state (used for ctxw only)
1052 */
1053static inline void
1054pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1055{
1056 int i;
1057 unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1058
1059 for (i=0; mask; i++, mask>>=1) {
1060 if ((mask & 0x1) == 0) continue;
1061 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1062 ia64_set_pmd(i, val);
1063 }
1064 ia64_srlz_d();
1065}
1066
1067/*
1068 * propagate PMD from context to thread-state
1069 */
1070static inline void
1071pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1072{
1073 struct thread_struct *thread = &task->thread;
1074 unsigned long ovfl_val = pmu_conf->ovfl_val;
1075 unsigned long mask = ctx->ctx_all_pmds[0];
1076 unsigned long val;
1077 int i;
1078
1079 DPRINT(("mask=0x%lx\n", mask));
1080
1081 for (i=0; mask; i++, mask>>=1) {
1082
1083 val = ctx->ctx_pmds[i].val;
1084
1085 /*
1086 * We break up the 64 bit value into 2 pieces
1087 * the lower bits go to the machine state in the
1088 * thread (will be reloaded on ctxsw in).
1089 * The upper part stays in the soft-counter.
1090 */
1091 if (PMD_IS_COUNTING(i)) {
1092 ctx->ctx_pmds[i].val = val & ~ovfl_val;
1093 val &= ovfl_val;
1094 }
1095 thread->pmds[i] = val;
1096
1097 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1098 i,
1099 thread->pmds[i],
1100 ctx->ctx_pmds[i].val));
1101 }
1102}
1103
1104/*
1105 * propagate PMC from context to thread-state
1106 */
1107static inline void
1108pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1109{
1110 struct thread_struct *thread = &task->thread;
1111 unsigned long mask = ctx->ctx_all_pmcs[0];
1112 int i;
1113
1114 DPRINT(("mask=0x%lx\n", mask));
1115
1116 for (i=0; mask; i++, mask>>=1) {
1117 /* masking 0 with ovfl_val yields 0 */
1118 thread->pmcs[i] = ctx->ctx_pmcs[i];
1119 DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
1120 }
1121}
1122
1123
1124
1125static inline void
1126pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1127{
1128 int i;
1129
1130 for (i=0; mask; i++, mask>>=1) {
1131 if ((mask & 0x1) == 0) continue;
1132 ia64_set_pmc(i, pmcs[i]);
1133 }
1134 ia64_srlz_d();
1135}
1136
1137static inline int
1138pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1139{
1140 return memcmp(a, b, sizeof(pfm_uuid_t));
1141}
1142
1143static inline int
1144pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1145{
1146 int ret = 0;
1147 if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1148 return ret;
1149}
1150
1151static inline int
1152pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1153{
1154 int ret = 0;
1155 if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1156 return ret;
1157}
1158
1159
1160static inline int
1161pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1162 int cpu, void *arg)
1163{
1164 int ret = 0;
1165 if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1166 return ret;
1167}
1168
1169static inline int
1170pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1171 int cpu, void *arg)
1172{
1173 int ret = 0;
1174 if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1175 return ret;
1176}
1177
1178static inline int
1179pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1180{
1181 int ret = 0;
1182 if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1183 return ret;
1184}
1185
1186static inline int
1187pfm_buf_fmt_restart_active(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1188{
1189 int ret = 0;
1190 if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1191 return ret;
1192}
1193
1194static pfm_buffer_fmt_t *
1195__pfm_find_buffer_fmt(pfm_uuid_t uuid)
1196{
1197 struct list_head * pos;
1198 pfm_buffer_fmt_t * entry;
1199
1200 list_for_each(pos, &pfm_buffer_fmt_list) {
1201 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1202 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1203 return entry;
1204 }
1205 return NULL;
1206}
1207
1208/*
1209 * find a buffer format based on its uuid
1210 */
1211static pfm_buffer_fmt_t *
1212pfm_find_buffer_fmt(pfm_uuid_t uuid)
1213{
1214 pfm_buffer_fmt_t * fmt;
1215 spin_lock(&pfm_buffer_fmt_lock);
1216 fmt = __pfm_find_buffer_fmt(uuid);
1217 spin_unlock(&pfm_buffer_fmt_lock);
1218 return fmt;
1219}
1220
1221int
1222pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1223{
1224 int ret = 0;
1225
1226 /* some sanity checks */
1227 if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1228
1229 /* we need at least a handler */
1230 if (fmt->fmt_handler == NULL) return -EINVAL;
1231
1232 /*
1233 * XXX: need check validity of fmt_arg_size
1234 */
1235
1236 spin_lock(&pfm_buffer_fmt_lock);
1237
1238 if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1239 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1240 ret = -EBUSY;
1241 goto out;
1242 }
1243 list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1244 printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1245
1246out:
1247 spin_unlock(&pfm_buffer_fmt_lock);
1248 return ret;
1249}
1250EXPORT_SYMBOL(pfm_register_buffer_fmt);
1251
1252int
1253pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1254{
1255 pfm_buffer_fmt_t *fmt;
1256 int ret = 0;
1257
1258 spin_lock(&pfm_buffer_fmt_lock);
1259
1260 fmt = __pfm_find_buffer_fmt(uuid);
1261 if (!fmt) {
1262 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1263 ret = -EINVAL;
1264 goto out;
1265 }
1266 list_del_init(&fmt->fmt_list);
1267 printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1268
1269out:
1270 spin_unlock(&pfm_buffer_fmt_lock);
1271 return ret;
1272
1273}
1274EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1275
1276static int
1277pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1278{
1279 unsigned long flags;
1280 /*
1281 * validy checks on cpu_mask have been done upstream
1282 */
1283 LOCK_PFS(flags);
1284
1285 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1286 pfm_sessions.pfs_sys_sessions,
1287 pfm_sessions.pfs_task_sessions,
1288 pfm_sessions.pfs_sys_use_dbregs,
1289 is_syswide,
1290 cpu));
1291
1292 if (is_syswide) {
1293 /*
1294 * cannot mix system wide and per-task sessions
1295 */
1296 if (pfm_sessions.pfs_task_sessions > 0UL) {
1297 DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1298 pfm_sessions.pfs_task_sessions));
1299 goto abort;
1300 }
1301
1302 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1303
1304 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1305
1306 pfm_sessions.pfs_sys_session[cpu] = task;
1307
1308 pfm_sessions.pfs_sys_sessions++ ;
1309
1310 } else {
1311 if (pfm_sessions.pfs_sys_sessions) goto abort;
1312 pfm_sessions.pfs_task_sessions++;
1313 }
1314
1315 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1316 pfm_sessions.pfs_sys_sessions,
1317 pfm_sessions.pfs_task_sessions,
1318 pfm_sessions.pfs_sys_use_dbregs,
1319 is_syswide,
1320 cpu));
1321
1322 UNLOCK_PFS(flags);
1323
1324 return 0;
1325
1326error_conflict:
1327 DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1328 pfm_sessions.pfs_sys_session[cpu]->pid,
1329 smp_processor_id()));
1330abort:
1331 UNLOCK_PFS(flags);
1332
1333 return -EBUSY;
1334
1335}
1336
1337static int
1338pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1339{
1340 unsigned long flags;
1341 /*
1342 * validy checks on cpu_mask have been done upstream
1343 */
1344 LOCK_PFS(flags);
1345
1346 DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1347 pfm_sessions.pfs_sys_sessions,
1348 pfm_sessions.pfs_task_sessions,
1349 pfm_sessions.pfs_sys_use_dbregs,
1350 is_syswide,
1351 cpu));
1352
1353
1354 if (is_syswide) {
1355 pfm_sessions.pfs_sys_session[cpu] = NULL;
1356 /*
1357 * would not work with perfmon+more than one bit in cpu_mask
1358 */
1359 if (ctx && ctx->ctx_fl_using_dbreg) {
1360 if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1361 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1362 } else {
1363 pfm_sessions.pfs_sys_use_dbregs--;
1364 }
1365 }
1366 pfm_sessions.pfs_sys_sessions--;
1367 } else {
1368 pfm_sessions.pfs_task_sessions--;
1369 }
1370 DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1371 pfm_sessions.pfs_sys_sessions,
1372 pfm_sessions.pfs_task_sessions,
1373 pfm_sessions.pfs_sys_use_dbregs,
1374 is_syswide,
1375 cpu));
1376
1377 UNLOCK_PFS(flags);
1378
1379 return 0;
1380}
1381
1382/*
1383 * removes virtual mapping of the sampling buffer.
1384 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1385 * a PROTECT_CTX() section.
1386 */
1387static int
1388pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1389{
1390 int r;
1391
1392 /* sanity checks */
1393 if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1394 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
1395 return -EINVAL;
1396 }
1397
1398 DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1399
1400 /*
1401 * does the actual unmapping
1402 */
1403 down_write(&task->mm->mmap_sem);
1404
1405 DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1406
1407 r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1408
1409 up_write(&task->mm->mmap_sem);
1410 if (r !=0) {
1411 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
1412 }
1413
1414 DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1415
1416 return 0;
1417}
1418
1419/*
1420 * free actual physical storage used by sampling buffer
1421 */
1422#if 0
1423static int
1424pfm_free_smpl_buffer(pfm_context_t *ctx)
1425{
1426 pfm_buffer_fmt_t *fmt;
1427
1428 if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1429
1430 /*
1431 * we won't use the buffer format anymore
1432 */
1433 fmt = ctx->ctx_buf_fmt;
1434
1435 DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1436 ctx->ctx_smpl_hdr,
1437 ctx->ctx_smpl_size,
1438 ctx->ctx_smpl_vaddr));
1439
1440 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1441
1442 /*
1443 * free the buffer
1444 */
1445 pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1446
1447 ctx->ctx_smpl_hdr = NULL;
1448 ctx->ctx_smpl_size = 0UL;
1449
1450 return 0;
1451
1452invalid_free:
1453 printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
1454 return -EINVAL;
1455}
1456#endif
1457
1458static inline void
1459pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1460{
1461 if (fmt == NULL) return;
1462
1463 pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1464
1465}
1466
1467/*
1468 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1469 * no real gain from having the whole whorehouse mounted. So we don't need
1470 * any operations on the root directory. However, we need a non-trivial
1471 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1472 */
1473static struct vfsmount *pfmfs_mnt;
1474
1475static int __init
1476init_pfm_fs(void)
1477{
1478 int err = register_filesystem(&pfm_fs_type);
1479 if (!err) {
1480 pfmfs_mnt = kern_mount(&pfm_fs_type);
1481 err = PTR_ERR(pfmfs_mnt);
1482 if (IS_ERR(pfmfs_mnt))
1483 unregister_filesystem(&pfm_fs_type);
1484 else
1485 err = 0;
1486 }
1487 return err;
1488}
1489
1490static void __exit
1491exit_pfm_fs(void)
1492{
1493 unregister_filesystem(&pfm_fs_type);
1494 mntput(pfmfs_mnt);
1495}
1496
1497static ssize_t
1498pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1499{
1500 pfm_context_t *ctx;
1501 pfm_msg_t *msg;
1502 ssize_t ret;
1503 unsigned long flags;
1504 DECLARE_WAITQUEUE(wait, current);
1505 if (PFM_IS_FILE(filp) == 0) {
1506 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1507 return -EINVAL;
1508 }
1509
1510 ctx = (pfm_context_t *)filp->private_data;
1511 if (ctx == NULL) {
1512 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
1513 return -EINVAL;
1514 }
1515
1516 /*
1517 * check even when there is no message
1518 */
1519 if (size < sizeof(pfm_msg_t)) {
1520 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1521 return -EINVAL;
1522 }
1523
1524 PROTECT_CTX(ctx, flags);
1525
1526 /*
1527 * put ourselves on the wait queue
1528 */
1529 add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1530
1531
1532 for(;;) {
1533 /*
1534 * check wait queue
1535 */
1536
1537 set_current_state(TASK_INTERRUPTIBLE);
1538
1539 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1540
1541 ret = 0;
1542 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1543
1544 UNPROTECT_CTX(ctx, flags);
1545
1546 /*
1547 * check non-blocking read
1548 */
1549 ret = -EAGAIN;
1550 if(filp->f_flags & O_NONBLOCK) break;
1551
1552 /*
1553 * check pending signals
1554 */
1555 if(signal_pending(current)) {
1556 ret = -EINTR;
1557 break;
1558 }
1559 /*
1560 * no message, so wait
1561 */
1562 schedule();
1563
1564 PROTECT_CTX(ctx, flags);
1565 }
1566 DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
1567 set_current_state(TASK_RUNNING);
1568 remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1569
1570 if (ret < 0) goto abort;
1571
1572 ret = -EINVAL;
1573 msg = pfm_get_next_msg(ctx);
1574 if (msg == NULL) {
1575 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
1576 goto abort_locked;
1577 }
1578
1579 DPRINT(("[%d] fd=%d type=%d\n", current->pid, msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1580
1581 ret = -EFAULT;
1582 if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1583
1584abort_locked:
1585 UNPROTECT_CTX(ctx, flags);
1586abort:
1587 return ret;
1588}
1589
1590static ssize_t
1591pfm_write(struct file *file, const char __user *ubuf,
1592 size_t size, loff_t *ppos)
1593{
1594 DPRINT(("pfm_write called\n"));
1595 return -EINVAL;
1596}
1597
1598static unsigned int
1599pfm_poll(struct file *filp, poll_table * wait)
1600{
1601 pfm_context_t *ctx;
1602 unsigned long flags;
1603 unsigned int mask = 0;
1604
1605 if (PFM_IS_FILE(filp) == 0) {
1606 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
1607 return 0;
1608 }
1609
1610 ctx = (pfm_context_t *)filp->private_data;
1611 if (ctx == NULL) {
1612 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
1613 return 0;
1614 }
1615
1616
1617 DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1618
1619 poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1620
1621 PROTECT_CTX(ctx, flags);
1622
1623 if (PFM_CTXQ_EMPTY(ctx) == 0)
1624 mask = POLLIN | POLLRDNORM;
1625
1626 UNPROTECT_CTX(ctx, flags);
1627
1628 DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1629
1630 return mask;
1631}
1632
1633static int
1634pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
1635{
1636 DPRINT(("pfm_ioctl called\n"));
1637 return -EINVAL;
1638}
1639
1640/*
1641 * interrupt cannot be masked when coming here
1642 */
1643static inline int
1644pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1645{
1646 int ret;
1647
1648 ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1649
1650 DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1651 current->pid,
1652 fd,
1653 on,
1654 ctx->ctx_async_queue, ret));
1655
1656 return ret;
1657}
1658
1659static int
1660pfm_fasync(int fd, struct file *filp, int on)
1661{
1662 pfm_context_t *ctx;
1663 int ret;
1664
1665 if (PFM_IS_FILE(filp) == 0) {
1666 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
1667 return -EBADF;
1668 }
1669
1670 ctx = (pfm_context_t *)filp->private_data;
1671 if (ctx == NULL) {
1672 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
1673 return -EBADF;
1674 }
1675 /*
1676 * we cannot mask interrupts during this call because this may
1677 * may go to sleep if memory is not readily avalaible.
1678 *
1679 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1680 * done in caller. Serialization of this function is ensured by caller.
1681 */
1682 ret = pfm_do_fasync(fd, filp, ctx, on);
1683
1684
1685 DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1686 fd,
1687 on,
1688 ctx->ctx_async_queue, ret));
1689
1690 return ret;
1691}
1692
1693#ifdef CONFIG_SMP
1694/*
1695 * this function is exclusively called from pfm_close().
1696 * The context is not protected at that time, nor are interrupts
1697 * on the remote CPU. That's necessary to avoid deadlocks.
1698 */
1699static void
1700pfm_syswide_force_stop(void *info)
1701{
1702 pfm_context_t *ctx = (pfm_context_t *)info;
1703 struct pt_regs *regs = ia64_task_regs(current);
1704 struct task_struct *owner;
1705 unsigned long flags;
1706 int ret;
1707
1708 if (ctx->ctx_cpu != smp_processor_id()) {
1709 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n",
1710 ctx->ctx_cpu,
1711 smp_processor_id());
1712 return;
1713 }
1714 owner = GET_PMU_OWNER();
1715 if (owner != ctx->ctx_task) {
1716 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1717 smp_processor_id(),
1718 owner->pid, ctx->ctx_task->pid);
1719 return;
1720 }
1721 if (GET_PMU_CTX() != ctx) {
1722 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1723 smp_processor_id(),
1724 GET_PMU_CTX(), ctx);
1725 return;
1726 }
1727
1728 DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));
1729 /*
1730 * the context is already protected in pfm_close(), we simply
1731 * need to mask interrupts to avoid a PMU interrupt race on
1732 * this CPU
1733 */
1734 local_irq_save(flags);
1735
1736 ret = pfm_context_unload(ctx, NULL, 0, regs);
1737 if (ret) {
1738 DPRINT(("context_unload returned %d\n", ret));
1739 }
1740
1741 /*
1742 * unmask interrupts, PMU interrupts are now spurious here
1743 */
1744 local_irq_restore(flags);
1745}
1746
1747static void
1748pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1749{
1750 int ret;
1751
1752 DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1753 ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
1754 DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1755}
1756#endif /* CONFIG_SMP */
1757
1758/*
1759 * called for each close(). Partially free resources.
1760 * When caller is self-monitoring, the context is unloaded.
1761 */
1762static int
1763pfm_flush(struct file *filp)
1764{
1765 pfm_context_t *ctx;
1766 struct task_struct *task;
1767 struct pt_regs *regs;
1768 unsigned long flags;
1769 unsigned long smpl_buf_size = 0UL;
1770 void *smpl_buf_vaddr = NULL;
1771 int state, is_system;
1772
1773 if (PFM_IS_FILE(filp) == 0) {
1774 DPRINT(("bad magic for\n"));
1775 return -EBADF;
1776 }
1777
1778 ctx = (pfm_context_t *)filp->private_data;
1779 if (ctx == NULL) {
1780 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
1781 return -EBADF;
1782 }
1783
1784 /*
1785 * remove our file from the async queue, if we use this mode.
1786 * This can be done without the context being protected. We come
1787 * here when the context has become unreacheable by other tasks.
1788 *
1789 * We may still have active monitoring at this point and we may
1790 * end up in pfm_overflow_handler(). However, fasync_helper()
1791 * operates with interrupts disabled and it cleans up the
1792 * queue. If the PMU handler is called prior to entering
1793 * fasync_helper() then it will send a signal. If it is
1794 * invoked after, it will find an empty queue and no
1795 * signal will be sent. In both case, we are safe
1796 */
1797 if (filp->f_flags & FASYNC) {
1798 DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
1799 pfm_do_fasync (-1, filp, ctx, 0);
1800 }
1801
1802 PROTECT_CTX(ctx, flags);
1803
1804 state = ctx->ctx_state;
1805 is_system = ctx->ctx_fl_system;
1806
1807 task = PFM_CTX_TASK(ctx);
1808 regs = ia64_task_regs(task);
1809
1810 DPRINT(("ctx_state=%d is_current=%d\n",
1811 state,
1812 task == current ? 1 : 0));
1813
1814 /*
1815 * if state == UNLOADED, then task is NULL
1816 */
1817
1818 /*
1819 * we must stop and unload because we are losing access to the context.
1820 */
1821 if (task == current) {
1822#ifdef CONFIG_SMP
1823 /*
1824 * the task IS the owner but it migrated to another CPU: that's bad
1825 * but we must handle this cleanly. Unfortunately, the kernel does
1826 * not provide a mechanism to block migration (while the context is loaded).
1827 *
1828 * We need to release the resource on the ORIGINAL cpu.
1829 */
1830 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1831
1832 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1833 /*
1834 * keep context protected but unmask interrupt for IPI
1835 */
1836 local_irq_restore(flags);
1837
1838 pfm_syswide_cleanup_other_cpu(ctx);
1839
1840 /*
1841 * restore interrupt masking
1842 */
1843 local_irq_save(flags);
1844
1845 /*
1846 * context is unloaded at this point
1847 */
1848 } else
1849#endif /* CONFIG_SMP */
1850 {
1851
1852 DPRINT(("forcing unload\n"));
1853 /*
1854 * stop and unload, returning with state UNLOADED
1855 * and session unreserved.
1856 */
1857 pfm_context_unload(ctx, NULL, 0, regs);
1858
1859 DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1860 }
1861 }
1862
1863 /*
1864 * remove virtual mapping, if any, for the calling task.
1865 * cannot reset ctx field until last user is calling close().
1866 *
1867 * ctx_smpl_vaddr must never be cleared because it is needed
1868 * by every task with access to the context
1869 *
1870 * When called from do_exit(), the mm context is gone already, therefore
1871 * mm is NULL, i.e., the VMA is already gone and we do not have to
1872 * do anything here
1873 */
1874 if (ctx->ctx_smpl_vaddr && current->mm) {
1875 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1876 smpl_buf_size = ctx->ctx_smpl_size;
1877 }
1878
1879 UNPROTECT_CTX(ctx, flags);
1880
1881 /*
1882 * if there was a mapping, then we systematically remove it
1883 * at this point. Cannot be done inside critical section
1884 * because some VM function reenables interrupts.
1885 *
1886 */
1887 if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1888
1889 return 0;
1890}
1891/*
1892 * called either on explicit close() or from exit_files().
1893 * Only the LAST user of the file gets to this point, i.e., it is
1894 * called only ONCE.
1895 *
1896 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero
1897 * (fput()),i.e, last task to access the file. Nobody else can access the
1898 * file at this point.
1899 *
1900 * When called from exit_files(), the VMA has been freed because exit_mm()
1901 * is executed before exit_files().
1902 *
1903 * When called from exit_files(), the current task is not yet ZOMBIE but we
1904 * flush the PMU state to the context.
1905 */
1906static int
1907pfm_close(struct inode *inode, struct file *filp)
1908{
1909 pfm_context_t *ctx;
1910 struct task_struct *task;
1911 struct pt_regs *regs;
1912 DECLARE_WAITQUEUE(wait, current);
1913 unsigned long flags;
1914 unsigned long smpl_buf_size = 0UL;
1915 void *smpl_buf_addr = NULL;
1916 int free_possible = 1;
1917 int state, is_system;
1918
1919 DPRINT(("pfm_close called private=%p\n", filp->private_data));
1920
1921 if (PFM_IS_FILE(filp) == 0) {
1922 DPRINT(("bad magic\n"));
1923 return -EBADF;
1924 }
1925
1926 ctx = (pfm_context_t *)filp->private_data;
1927 if (ctx == NULL) {
1928 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
1929 return -EBADF;
1930 }
1931
1932 PROTECT_CTX(ctx, flags);
1933
1934 state = ctx->ctx_state;
1935 is_system = ctx->ctx_fl_system;
1936
1937 task = PFM_CTX_TASK(ctx);
1938 regs = ia64_task_regs(task);
1939
1940 DPRINT(("ctx_state=%d is_current=%d\n",
1941 state,
1942 task == current ? 1 : 0));
1943
1944 /*
1945 * if task == current, then pfm_flush() unloaded the context
1946 */
1947 if (state == PFM_CTX_UNLOADED) goto doit;
1948
1949 /*
1950 * context is loaded/masked and task != current, we need to
1951 * either force an unload or go zombie
1952 */
1953
1954 /*
1955 * The task is currently blocked or will block after an overflow.
1956 * we must force it to wakeup to get out of the
1957 * MASKED state and transition to the unloaded state by itself.
1958 *
1959 * This situation is only possible for per-task mode
1960 */
1961 if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1962
1963 /*
1964 * set a "partial" zombie state to be checked
1965 * upon return from down() in pfm_handle_work().
1966 *
1967 * We cannot use the ZOMBIE state, because it is checked
1968 * by pfm_load_regs() which is called upon wakeup from down().
1969 * In such case, it would free the context and then we would
1970 * return to pfm_handle_work() which would access the
1971 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1972 * but visible to pfm_handle_work().
1973 *
1974 * For some window of time, we have a zombie context with
1975 * ctx_state = MASKED and not ZOMBIE
1976 */
1977 ctx->ctx_fl_going_zombie = 1;
1978
1979 /*
1980 * force task to wake up from MASKED state
1981 */
1982 up(&ctx->ctx_restart_sem);
1983
1984 DPRINT(("waking up ctx_state=%d\n", state));
1985
1986 /*
1987 * put ourself to sleep waiting for the other
1988 * task to report completion
1989 *
1990 * the context is protected by mutex, therefore there
1991 * is no risk of being notified of completion before
1992 * begin actually on the waitq.
1993 */
1994 set_current_state(TASK_INTERRUPTIBLE);
1995 add_wait_queue(&ctx->ctx_zombieq, &wait);
1996
1997 UNPROTECT_CTX(ctx, flags);
1998
1999 /*
2000 * XXX: check for signals :
2001 * - ok for explicit close
2002 * - not ok when coming from exit_files()
2003 */
2004 schedule();
2005
2006
2007 PROTECT_CTX(ctx, flags);
2008
2009
2010 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2011 set_current_state(TASK_RUNNING);
2012
2013 /*
2014 * context is unloaded at this point
2015 */
2016 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2017 }
2018 else if (task != current) {
2019#ifdef CONFIG_SMP
2020 /*
2021 * switch context to zombie state
2022 */
2023 ctx->ctx_state = PFM_CTX_ZOMBIE;
2024
2025 DPRINT(("zombie ctx for [%d]\n", task->pid));
2026 /*
2027 * cannot free the context on the spot. deferred until
2028 * the task notices the ZOMBIE state
2029 */
2030 free_possible = 0;
2031#else
2032 pfm_context_unload(ctx, NULL, 0, regs);
2033#endif
2034 }
2035
2036doit:
2037 /* reload state, may have changed during opening of critical section */
2038 state = ctx->ctx_state;
2039
2040 /*
2041 * the context is still attached to a task (possibly current)
2042 * we cannot destroy it right now
2043 */
2044
2045 /*
2046 * we must free the sampling buffer right here because
2047 * we cannot rely on it being cleaned up later by the
2048 * monitored task. It is not possible to free vmalloc'ed
2049 * memory in pfm_load_regs(). Instead, we remove the buffer
2050 * now. should there be subsequent PMU overflow originally
2051 * meant for sampling, the will be converted to spurious
2052 * and that's fine because the monitoring tools is gone anyway.
2053 */
2054 if (ctx->ctx_smpl_hdr) {
2055 smpl_buf_addr = ctx->ctx_smpl_hdr;
2056 smpl_buf_size = ctx->ctx_smpl_size;
2057 /* no more sampling */
2058 ctx->ctx_smpl_hdr = NULL;
2059 ctx->ctx_fl_is_sampling = 0;
2060 }
2061
2062 DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2063 state,
2064 free_possible,
2065 smpl_buf_addr,
2066 smpl_buf_size));
2067
2068 if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2069
2070 /*
2071 * UNLOADED that the session has already been unreserved.
2072 */
2073 if (state == PFM_CTX_ZOMBIE) {
2074 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2075 }
2076
2077 /*
2078 * disconnect file descriptor from context must be done
2079 * before we unlock.
2080 */
2081 filp->private_data = NULL;
2082
2083 /*
2084 * if we free on the spot, the context is now completely unreacheable
2085 * from the callers side. The monitored task side is also cut, so we
2086 * can freely cut.
2087 *
2088 * If we have a deferred free, only the caller side is disconnected.
2089 */
2090 UNPROTECT_CTX(ctx, flags);
2091
2092 /*
2093 * All memory free operations (especially for vmalloc'ed memory)
2094 * MUST be done with interrupts ENABLED.
2095 */
2096 if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2097
2098 /*
2099 * return the memory used by the context
2100 */
2101 if (free_possible) pfm_context_free(ctx);
2102
2103 return 0;
2104}
2105
2106static int
2107pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2108{
2109 DPRINT(("pfm_no_open called\n"));
2110 return -ENXIO;
2111}
2112
2113
2114
2115static struct file_operations pfm_file_ops = {
2116 .llseek = no_llseek,
2117 .read = pfm_read,
2118 .write = pfm_write,
2119 .poll = pfm_poll,
2120 .ioctl = pfm_ioctl,
2121 .open = pfm_no_open, /* special open code to disallow open via /proc */
2122 .fasync = pfm_fasync,
2123 .release = pfm_close,
2124 .flush = pfm_flush
2125};
2126
2127static int
2128pfmfs_delete_dentry(struct dentry *dentry)
2129{
2130 return 1;
2131}
2132
2133static struct dentry_operations pfmfs_dentry_operations = {
2134 .d_delete = pfmfs_delete_dentry,
2135};
2136
2137
2138static int
2139pfm_alloc_fd(struct file **cfile)
2140{
2141 int fd, ret = 0;
2142 struct file *file = NULL;
2143 struct inode * inode;
2144 char name[32];
2145 struct qstr this;
2146
2147 fd = get_unused_fd();
2148 if (fd < 0) return -ENFILE;
2149
2150 ret = -ENFILE;
2151
2152 file = get_empty_filp();
2153 if (!file) goto out;
2154
2155 /*
2156 * allocate a new inode
2157 */
2158 inode = new_inode(pfmfs_mnt->mnt_sb);
2159 if (!inode) goto out;
2160
2161 DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2162
2163 inode->i_mode = S_IFCHR|S_IRUGO;
2164 inode->i_uid = current->fsuid;
2165 inode->i_gid = current->fsgid;
2166
2167 sprintf(name, "[%lu]", inode->i_ino);
2168 this.name = name;
2169 this.len = strlen(name);
2170 this.hash = inode->i_ino;
2171
2172 ret = -ENOMEM;
2173
2174 /*
2175 * allocate a new dcache entry
2176 */
2177 file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2178 if (!file->f_dentry) goto out;
2179
2180 file->f_dentry->d_op = &pfmfs_dentry_operations;
2181
2182 d_add(file->f_dentry, inode);
2183 file->f_vfsmnt = mntget(pfmfs_mnt);
2184 file->f_mapping = inode->i_mapping;
2185
2186 file->f_op = &pfm_file_ops;
2187 file->f_mode = FMODE_READ;
2188 file->f_flags = O_RDONLY;
2189 file->f_pos = 0;
2190
2191 /*
2192 * may have to delay until context is attached?
2193 */
2194 fd_install(fd, file);
2195
2196 /*
2197 * the file structure we will use
2198 */
2199 *cfile = file;
2200
2201 return fd;
2202out:
2203 if (file) put_filp(file);
2204 put_unused_fd(fd);
2205 return ret;
2206}
2207
2208static void
2209pfm_free_fd(int fd, struct file *file)
2210{
2211 struct files_struct *files = current->files;
2212
2213 /*
2214 * there ie no fd_uninstall(), so we do it here
2215 */
2216 spin_lock(&files->file_lock);
2217 files->fd[fd] = NULL;
2218 spin_unlock(&files->file_lock);
2219
2220 if (file) put_filp(file);
2221 put_unused_fd(fd);
2222}
2223
2224static int
2225pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2226{
2227 DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2228
2229 while (size > 0) {
2230 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2231
2232
2233 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2234 return -ENOMEM;
2235
2236 addr += PAGE_SIZE;
2237 buf += PAGE_SIZE;
2238 size -= PAGE_SIZE;
2239 }
2240 return 0;
2241}
2242
2243/*
2244 * allocate a sampling buffer and remaps it into the user address space of the task
2245 */
2246static int
2247pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2248{
2249 struct mm_struct *mm = task->mm;
2250 struct vm_area_struct *vma = NULL;
2251 unsigned long size;
2252 void *smpl_buf;
2253
2254
2255 /*
2256 * the fixed header + requested size and align to page boundary
2257 */
2258 size = PAGE_ALIGN(rsize);
2259
2260 DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2261
2262 /*
2263 * check requested size to avoid Denial-of-service attacks
2264 * XXX: may have to refine this test
2265 * Check against address space limit.
2266 *
2267 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2268 * return -ENOMEM;
2269 */
2270 if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
2271 return -ENOMEM;
2272
2273 /*
2274 * We do the easy to undo allocations first.
2275 *
2276 * pfm_rvmalloc(), clears the buffer, so there is no leak
2277 */
2278 smpl_buf = pfm_rvmalloc(size);
2279 if (smpl_buf == NULL) {
2280 DPRINT(("Can't allocate sampling buffer\n"));
2281 return -ENOMEM;
2282 }
2283
2284 DPRINT(("smpl_buf @%p\n", smpl_buf));
2285
2286 /* allocate vma */
2287 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2288 if (!vma) {
2289 DPRINT(("Cannot allocate vma\n"));
2290 goto error_kmem;
2291 }
2292 memset(vma, 0, sizeof(*vma));
2293
2294 /*
2295 * partially initialize the vma for the sampling buffer
2296 */
2297 vma->vm_mm = mm;
2298 vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED;
2299 vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */
2300
2301 /*
2302 * Now we have everything we need and we can initialize
2303 * and connect all the data structures
2304 */
2305
2306 ctx->ctx_smpl_hdr = smpl_buf;
2307 ctx->ctx_smpl_size = size; /* aligned size */
2308
2309 /*
2310 * Let's do the difficult operations next.
2311 *
2312 * now we atomically find some area in the address space and
2313 * remap the buffer in it.
2314 */
2315 down_write(&task->mm->mmap_sem);
2316
2317 /* find some free area in address space, must have mmap sem held */
2318 vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2319 if (vma->vm_start == 0UL) {
2320 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2321 up_write(&task->mm->mmap_sem);
2322 goto error;
2323 }
2324 vma->vm_end = vma->vm_start + size;
2325 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2326
2327 DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2328
2329 /* can only be applied to current task, need to have the mm semaphore held when called */
2330 if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2331 DPRINT(("Can't remap buffer\n"));
2332 up_write(&task->mm->mmap_sem);
2333 goto error;
2334 }
2335
2336 /*
2337 * now insert the vma in the vm list for the process, must be
2338 * done with mmap lock held
2339 */
2340 insert_vm_struct(mm, vma);
2341
2342 mm->total_vm += size >> PAGE_SHIFT;
2343 vm_stat_account(vma);
2344 up_write(&task->mm->mmap_sem);
2345
2346 /*
2347 * keep track of user level virtual address
2348 */
2349 ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2350 *(unsigned long *)user_vaddr = vma->vm_start;
2351
2352 return 0;
2353
2354error:
2355 kmem_cache_free(vm_area_cachep, vma);
2356error_kmem:
2357 pfm_rvfree(smpl_buf, size);
2358
2359 return -ENOMEM;
2360}
2361
2362/*
2363 * XXX: do something better here
2364 */
2365static int
2366pfm_bad_permissions(struct task_struct *task)
2367{
2368 /* inspired by ptrace_attach() */
2369 DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2370 current->uid,
2371 current->gid,
2372 task->euid,
2373 task->suid,
2374 task->uid,
2375 task->egid,
2376 task->sgid));
2377
2378 return ((current->uid != task->euid)
2379 || (current->uid != task->suid)
2380 || (current->uid != task->uid)
2381 || (current->gid != task->egid)
2382 || (current->gid != task->sgid)
2383 || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
2384}
2385
2386static int
2387pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2388{
2389 int ctx_flags;
2390
2391 /* valid signal */
2392
2393 ctx_flags = pfx->ctx_flags;
2394
2395 if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2396
2397 /*
2398 * cannot block in this mode
2399 */
2400 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2401 DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2402 return -EINVAL;
2403 }
2404 } else {
2405 }
2406 /* probably more to add here */
2407
2408 return 0;
2409}
2410
2411static int
2412pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
2413 unsigned int cpu, pfarg_context_t *arg)
2414{
2415 pfm_buffer_fmt_t *fmt = NULL;
2416 unsigned long size = 0UL;
2417 void *uaddr = NULL;
2418 void *fmt_arg = NULL;
2419 int ret = 0;
2420#define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1)
2421
2422 /* invoke and lock buffer format, if found */
2423 fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2424 if (fmt == NULL) {
2425 DPRINT(("[%d] cannot find buffer format\n", task->pid));
2426 return -EINVAL;
2427 }
2428
2429 /*
2430 * buffer argument MUST be contiguous to pfarg_context_t
2431 */
2432 if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2433
2434 ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2435
2436 DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));
2437
2438 if (ret) goto error;
2439
2440 /* link buffer format and context */
2441 ctx->ctx_buf_fmt = fmt;
2442
2443 /*
2444 * check if buffer format wants to use perfmon buffer allocation/mapping service
2445 */
2446 ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2447 if (ret) goto error;
2448
2449 if (size) {
2450 /*
2451 * buffer is always remapped into the caller's address space
2452 */
2453 ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
2454 if (ret) goto error;
2455
2456 /* keep track of user address of buffer */
2457 arg->ctx_smpl_vaddr = uaddr;
2458 }
2459 ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2460
2461error:
2462 return ret;
2463}
2464
2465static void
2466pfm_reset_pmu_state(pfm_context_t *ctx)
2467{
2468 int i;
2469
2470 /*
2471 * install reset values for PMC.
2472 */
2473 for (i=1; PMC_IS_LAST(i) == 0; i++) {
2474 if (PMC_IS_IMPL(i) == 0) continue;
2475 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2476 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2477 }
2478 /*
2479 * PMD registers are set to 0UL when the context in memset()
2480 */
2481
2482 /*
2483 * On context switched restore, we must restore ALL pmc and ALL pmd even
2484 * when they are not actively used by the task. In UP, the incoming process
2485 * may otherwise pick up left over PMC, PMD state from the previous process.
2486 * As opposed to PMD, stale PMC can cause harm to the incoming
2487 * process because they may change what is being measured.
2488 * Therefore, we must systematically reinstall the entire
2489 * PMC state. In SMP, the same thing is possible on the
2490 * same CPU but also on between 2 CPUs.
2491 *
2492 * The problem with PMD is information leaking especially
2493 * to user level when psr.sp=0
2494 *
2495 * There is unfortunately no easy way to avoid this problem
2496 * on either UP or SMP. This definitively slows down the
2497 * pfm_load_regs() function.
2498 */
2499
2500 /*
2501 * bitmask of all PMCs accessible to this context
2502 *
2503 * PMC0 is treated differently.
2504 */
2505 ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2506
2507 /*
2508 * bitmask of all PMDs that are accesible to this context
2509 */
2510 ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2511
2512 DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2513
2514 /*
2515 * useful in case of re-enable after disable
2516 */
2517 ctx->ctx_used_ibrs[0] = 0UL;
2518 ctx->ctx_used_dbrs[0] = 0UL;
2519}
2520
2521static int
2522pfm_ctx_getsize(void *arg, size_t *sz)
2523{
2524 pfarg_context_t *req = (pfarg_context_t *)arg;
2525 pfm_buffer_fmt_t *fmt;
2526
2527 *sz = 0;
2528
2529 if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2530
2531 fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2532 if (fmt == NULL) {
2533 DPRINT(("cannot find buffer format\n"));
2534 return -EINVAL;
2535 }
2536 /* get just enough to copy in user parameters */
2537 *sz = fmt->fmt_arg_size;
2538 DPRINT(("arg_size=%lu\n", *sz));
2539
2540 return 0;
2541}
2542
2543
2544
2545/*
2546 * cannot attach if :
2547 * - kernel task
2548 * - task not owned by caller
2549 * - task incompatible with context mode
2550 */
2551static int
2552pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2553{
2554 /*
2555 * no kernel task or task not owner by caller
2556 */
2557 if (task->mm == NULL) {
2558 DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
2559 return -EPERM;
2560 }
2561 if (pfm_bad_permissions(task)) {
2562 DPRINT(("no permission to attach to [%d]\n", task->pid));
2563 return -EPERM;
2564 }
2565 /*
2566 * cannot block in self-monitoring mode
2567 */
2568 if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2569 DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
2570 return -EINVAL;
2571 }
2572
2573 if (task->exit_state == EXIT_ZOMBIE) {
2574 DPRINT(("cannot attach to zombie task [%d]\n", task->pid));
2575 return -EBUSY;
2576 }
2577
2578 /*
2579 * always ok for self
2580 */
2581 if (task == current) return 0;
2582
2583 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
2584 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
2585 return -EBUSY;
2586 }
2587 /*
2588 * make sure the task is off any CPU
2589 */
2590 wait_task_inactive(task);
2591
2592 /* more to come... */
2593
2594 return 0;
2595}
2596
2597static int
2598pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2599{
2600 struct task_struct *p = current;
2601 int ret;
2602
2603 /* XXX: need to add more checks here */
2604 if (pid < 2) return -EPERM;
2605
2606 if (pid != current->pid) {
2607
2608 read_lock(&tasklist_lock);
2609
2610 p = find_task_by_pid(pid);
2611
2612 /* make sure task cannot go away while we operate on it */
2613 if (p) get_task_struct(p);
2614
2615 read_unlock(&tasklist_lock);
2616
2617 if (p == NULL) return -ESRCH;
2618 }
2619
2620 ret = pfm_task_incompatible(ctx, p);
2621 if (ret == 0) {
2622 *task = p;
2623 } else if (p != current) {
2624 pfm_put_task(p);
2625 }
2626 return ret;
2627}
2628
2629
2630
2631static int
2632pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2633{
2634 pfarg_context_t *req = (pfarg_context_t *)arg;
2635 struct file *filp;
2636 int ctx_flags;
2637 int ret;
2638
2639 /* let's check the arguments first */
2640 ret = pfarg_is_sane(current, req);
2641 if (ret < 0) return ret;
2642
2643 ctx_flags = req->ctx_flags;
2644
2645 ret = -ENOMEM;
2646
2647 ctx = pfm_context_alloc();
2648 if (!ctx) goto error;
2649
2650 ret = pfm_alloc_fd(&filp);
2651 if (ret < 0) goto error_file;
2652
2653 req->ctx_fd = ctx->ctx_fd = ret;
2654
2655 /*
2656 * attach context to file
2657 */
2658 filp->private_data = ctx;
2659
2660 /*
2661 * does the user want to sample?
2662 */
2663 if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2664 ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
2665 if (ret) goto buffer_error;
2666 }
2667
2668 /*
2669 * init context protection lock
2670 */
2671 spin_lock_init(&ctx->ctx_lock);
2672
2673 /*
2674 * context is unloaded
2675 */
2676 ctx->ctx_state = PFM_CTX_UNLOADED;
2677
2678 /*
2679 * initialization of context's flags
2680 */
2681 ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
2682 ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
2683 ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
2684 ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
2685 /*
2686 * will move to set properties
2687 * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
2688 */
2689
2690 /*
2691 * init restart semaphore to locked
2692 */
2693 sema_init(&ctx->ctx_restart_sem, 0);
2694
2695 /*
2696 * activation is used in SMP only
2697 */
2698 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
2699 SET_LAST_CPU(ctx, -1);
2700
2701 /*
2702 * initialize notification message queue
2703 */
2704 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
2705 init_waitqueue_head(&ctx->ctx_msgq_wait);
2706 init_waitqueue_head(&ctx->ctx_zombieq);
2707
2708 DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
2709 ctx,
2710 ctx_flags,
2711 ctx->ctx_fl_system,
2712 ctx->ctx_fl_block,
2713 ctx->ctx_fl_excl_idle,
2714 ctx->ctx_fl_no_msg,
2715 ctx->ctx_fd));
2716
2717 /*
2718 * initialize soft PMU state
2719 */
2720 pfm_reset_pmu_state(ctx);
2721
2722 return 0;
2723
2724buffer_error:
2725 pfm_free_fd(ctx->ctx_fd, filp);
2726
2727 if (ctx->ctx_buf_fmt) {
2728 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2729 }
2730error_file:
2731 pfm_context_free(ctx);
2732
2733error:
2734 return ret;
2735}
2736
2737static inline unsigned long
2738pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2739{
2740 unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2741 unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2742 extern unsigned long carta_random32 (unsigned long seed);
2743
2744 if (reg->flags & PFM_REGFL_RANDOM) {
2745 new_seed = carta_random32(old_seed);
2746 val -= (old_seed & mask); /* counter values are negative numbers! */
2747 if ((mask >> 32) != 0)
2748 /* construct a full 64-bit random value: */
2749 new_seed |= carta_random32(old_seed >> 32) << 32;
2750 reg->seed = new_seed;
2751 }
2752 reg->lval = val;
2753 return val;
2754}
2755
2756static void
2757pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2758{
2759 unsigned long mask = ovfl_regs[0];
2760 unsigned long reset_others = 0UL;
2761 unsigned long val;
2762 int i;
2763
2764 /*
2765 * now restore reset value on sampling overflowed counters
2766 */
2767 mask >>= PMU_FIRST_COUNTER;
2768 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2769
2770 if ((mask & 0x1UL) == 0UL) continue;
2771
2772 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2773 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2774
2775 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2776 }
2777
2778 /*
2779 * Now take care of resetting the other registers
2780 */
2781 for(i = 0; reset_others; i++, reset_others >>= 1) {
2782
2783 if ((reset_others & 0x1) == 0) continue;
2784
2785 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2786
2787 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2788 is_long_reset ? "long" : "short", i, val));
2789 }
2790}
2791
2792static void
2793pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2794{
2795 unsigned long mask = ovfl_regs[0];
2796 unsigned long reset_others = 0UL;
2797 unsigned long val;
2798 int i;
2799
2800 DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2801
2802 if (ctx->ctx_state == PFM_CTX_MASKED) {
2803 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2804 return;
2805 }
2806
2807 /*
2808 * now restore reset value on sampling overflowed counters
2809 */
2810 mask >>= PMU_FIRST_COUNTER;
2811 for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2812
2813 if ((mask & 0x1UL) == 0UL) continue;
2814
2815 val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2816 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2817
2818 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2819
2820 pfm_write_soft_counter(ctx, i, val);
2821 }
2822
2823 /*
2824 * Now take care of resetting the other registers
2825 */
2826 for(i = 0; reset_others; i++, reset_others >>= 1) {
2827
2828 if ((reset_others & 0x1) == 0) continue;
2829
2830 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2831
2832 if (PMD_IS_COUNTING(i)) {
2833 pfm_write_soft_counter(ctx, i, val);
2834 } else {
2835 ia64_set_pmd(i, val);
2836 }
2837 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2838 is_long_reset ? "long" : "short", i, val));
2839 }
2840 ia64_srlz_d();
2841}
2842
2843static int
2844pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2845{
2846 struct thread_struct *thread = NULL;
2847 struct task_struct *task;
2848 pfarg_reg_t *req = (pfarg_reg_t *)arg;
2849 unsigned long value, pmc_pm;
2850 unsigned long smpl_pmds, reset_pmds, impl_pmds;
2851 unsigned int cnum, reg_flags, flags, pmc_type;
2852 int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2853 int is_monitor, is_counting, state;
2854 int ret = -EINVAL;
2855 pfm_reg_check_t wr_func;
2856#define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2857
2858 state = ctx->ctx_state;
2859 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2860 is_system = ctx->ctx_fl_system;
2861 task = ctx->ctx_task;
2862 impl_pmds = pmu_conf->impl_pmds[0];
2863
2864 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2865
2866 if (is_loaded) {
2867 thread = &task->thread;
2868 /*
2869 * In system wide and when the context is loaded, access can only happen
2870 * when the caller is running on the CPU being monitored by the session.
2871 * It does not have to be the owner (ctx_task) of the context per se.
2872 */
2873 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2874 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2875 return -EBUSY;
2876 }
2877 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2878 }
2879 expert_mode = pfm_sysctl.expert_mode;
2880
2881 for (i = 0; i < count; i++, req++) {
2882
2883 cnum = req->reg_num;
2884 reg_flags = req->reg_flags;
2885 value = req->reg_value;
2886 smpl_pmds = req->reg_smpl_pmds[0];
2887 reset_pmds = req->reg_reset_pmds[0];
2888 flags = 0;
2889
2890
2891 if (cnum >= PMU_MAX_PMCS) {
2892 DPRINT(("pmc%u is invalid\n", cnum));
2893 goto error;
2894 }
2895
2896 pmc_type = pmu_conf->pmc_desc[cnum].type;
2897 pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2898 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2899 is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2900
2901 /*
2902 * we reject all non implemented PMC as well
2903 * as attempts to modify PMC[0-3] which are used
2904 * as status registers by the PMU
2905 */
2906 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2907 DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2908 goto error;
2909 }
2910 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2911 /*
2912 * If the PMC is a monitor, then if the value is not the default:
2913 * - system-wide session: PMCx.pm=1 (privileged monitor)
2914 * - per-task : PMCx.pm=0 (user monitor)
2915 */
2916 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2917 DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2918 cnum,
2919 pmc_pm,
2920 is_system));
2921 goto error;
2922 }
2923
2924 if (is_counting) {
2925 /*
2926 * enforce generation of overflow interrupt. Necessary on all
2927 * CPUs.
2928 */
2929 value |= 1 << PMU_PMC_OI;
2930
2931 if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2932 flags |= PFM_REGFL_OVFL_NOTIFY;
2933 }
2934
2935 if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2936
2937 /* verify validity of smpl_pmds */
2938 if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2939 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2940 goto error;
2941 }
2942
2943 /* verify validity of reset_pmds */
2944 if ((reset_pmds & impl_pmds) != reset_pmds) {
2945 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2946 goto error;
2947 }
2948 } else {
2949 if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2950 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2951 goto error;
2952 }
2953 /* eventid on non-counting monitors are ignored */
2954 }
2955
2956 /*
2957 * execute write checker, if any
2958 */
2959 if (likely(expert_mode == 0 && wr_func)) {
2960 ret = (*wr_func)(task, ctx, cnum, &value, regs);
2961 if (ret) goto error;
2962 ret = -EINVAL;
2963 }
2964
2965 /*
2966 * no error on this register
2967 */
2968 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2969
2970 /*
2971 * Now we commit the changes to the software state
2972 */
2973
2974 /*
2975 * update overflow information
2976 */
2977 if (is_counting) {
2978 /*
2979 * full flag update each time a register is programmed
2980 */
2981 ctx->ctx_pmds[cnum].flags = flags;
2982
2983 ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2984 ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds;
2985 ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid;
2986
2987 /*
2988 * Mark all PMDS to be accessed as used.
2989 *
2990 * We do not keep track of PMC because we have to
2991 * systematically restore ALL of them.
2992 *
2993 * We do not update the used_monitors mask, because
2994 * if we have not programmed them, then will be in
2995 * a quiescent state, therefore we will not need to
2996 * mask/restore then when context is MASKED.
2997 */
2998 CTX_USED_PMD(ctx, reset_pmds);
2999 CTX_USED_PMD(ctx, smpl_pmds);
3000 /*
3001 * make sure we do not try to reset on
3002 * restart because we have established new values
3003 */
3004 if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3005 }
3006 /*
3007 * Needed in case the user does not initialize the equivalent
3008 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3009 * possible leak here.
3010 */
3011 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3012
3013 /*
3014 * keep track of the monitor PMC that we are using.
3015 * we save the value of the pmc in ctx_pmcs[] and if
3016 * the monitoring is not stopped for the context we also
3017 * place it in the saved state area so that it will be
3018 * picked up later by the context switch code.
3019 *
3020 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3021 *
3022 * The value in thread->pmcs[] may be modified on overflow, i.e., when
3023 * monitoring needs to be stopped.
3024 */
3025 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3026
3027 /*
3028 * update context state
3029 */
3030 ctx->ctx_pmcs[cnum] = value;
3031
3032 if (is_loaded) {
3033 /*
3034 * write thread state
3035 */
3036 if (is_system == 0) thread->pmcs[cnum] = value;
3037
3038 /*
3039 * write hardware register if we can
3040 */
3041 if (can_access_pmu) {
3042 ia64_set_pmc(cnum, value);
3043 }
3044#ifdef CONFIG_SMP
3045 else {
3046 /*
3047 * per-task SMP only here
3048 *
3049 * we are guaranteed that the task is not running on the other CPU,
3050 * we indicate that this PMD will need to be reloaded if the task
3051 * is rescheduled on the CPU it ran last on.
3052 */
3053 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3054 }
3055#endif
3056 }
3057
3058 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",
3059 cnum,
3060 value,
3061 is_loaded,
3062 can_access_pmu,
3063 flags,
3064 ctx->ctx_all_pmcs[0],
3065 ctx->ctx_used_pmds[0],
3066 ctx->ctx_pmds[cnum].eventid,
3067 smpl_pmds,
3068 reset_pmds,
3069 ctx->ctx_reload_pmcs[0],
3070 ctx->ctx_used_monitors[0],
3071 ctx->ctx_ovfl_regs[0]));
3072 }
3073
3074 /*
3075 * make sure the changes are visible
3076 */
3077 if (can_access_pmu) ia64_srlz_d();
3078
3079 return 0;
3080error:
3081 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3082 return ret;
3083}
3084
3085static int
3086pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3087{
3088 struct thread_struct *thread = NULL;
3089 struct task_struct *task;
3090 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3091 unsigned long value, hw_value, ovfl_mask;
3092 unsigned int cnum;
3093 int i, can_access_pmu = 0, state;
3094 int is_counting, is_loaded, is_system, expert_mode;
3095 int ret = -EINVAL;
3096 pfm_reg_check_t wr_func;
3097
3098
3099 state = ctx->ctx_state;
3100 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3101 is_system = ctx->ctx_fl_system;
3102 ovfl_mask = pmu_conf->ovfl_val;
3103 task = ctx->ctx_task;
3104
3105 if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3106
3107 /*
3108 * on both UP and SMP, we can only write to the PMC when the task is
3109 * the owner of the local PMU.
3110 */
3111 if (likely(is_loaded)) {
3112 thread = &task->thread;
3113 /*
3114 * In system wide and when the context is loaded, access can only happen
3115 * when the caller is running on the CPU being monitored by the session.
3116 * It does not have to be the owner (ctx_task) of the context per se.
3117 */
3118 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3119 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3120 return -EBUSY;
3121 }
3122 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3123 }
3124 expert_mode = pfm_sysctl.expert_mode;
3125
3126 for (i = 0; i < count; i++, req++) {
3127
3128 cnum = req->reg_num;
3129 value = req->reg_value;
3130
3131 if (!PMD_IS_IMPL(cnum)) {
3132 DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3133 goto abort_mission;
3134 }
3135 is_counting = PMD_IS_COUNTING(cnum);
3136 wr_func = pmu_conf->pmd_desc[cnum].write_check;
3137
3138 /*
3139 * execute write checker, if any
3140 */
3141 if (unlikely(expert_mode == 0 && wr_func)) {
3142 unsigned long v = value;
3143
3144 ret = (*wr_func)(task, ctx, cnum, &v, regs);
3145 if (ret) goto abort_mission;
3146
3147 value = v;
3148 ret = -EINVAL;
3149 }
3150
3151 /*
3152 * no error on this register
3153 */
3154 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3155
3156 /*
3157 * now commit changes to software state
3158 */
3159 hw_value = value;
3160
3161 /*
3162 * update virtualized (64bits) counter
3163 */
3164 if (is_counting) {
3165 /*
3166 * write context state
3167 */
3168 ctx->ctx_pmds[cnum].lval = value;
3169
3170 /*
3171 * when context is load we use the split value
3172 */
3173 if (is_loaded) {
3174 hw_value = value & ovfl_mask;
3175 value = value & ~ovfl_mask;
3176 }
3177 }
3178 /*
3179 * update reset values (not just for counters)
3180 */
3181 ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset;
3182 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3183
3184 /*
3185 * update randomization parameters (not just for counters)
3186 */
3187 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3188 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3189
3190 /*
3191 * update context value
3192 */
3193 ctx->ctx_pmds[cnum].val = value;
3194
3195 /*
3196 * Keep track of what we use
3197 *
3198 * We do not keep track of PMC because we have to
3199 * systematically restore ALL of them.
3200 */
3201 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3202
3203 /*
3204 * mark this PMD register used as well
3205 */
3206 CTX_USED_PMD(ctx, RDEP(cnum));
3207
3208 /*
3209 * make sure we do not try to reset on
3210 * restart because we have established new values
3211 */
3212 if (is_counting && state == PFM_CTX_MASKED) {
3213 ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3214 }
3215
3216 if (is_loaded) {
3217 /*
3218 * write thread state
3219 */
3220 if (is_system == 0) thread->pmds[cnum] = hw_value;
3221
3222 /*
3223 * write hardware register if we can
3224 */
3225 if (can_access_pmu) {
3226 ia64_set_pmd(cnum, hw_value);
3227 } else {
3228#ifdef CONFIG_SMP
3229 /*
3230 * we are guaranteed that the task is not running on the other CPU,
3231 * we indicate that this PMD will need to be reloaded if the task
3232 * is rescheduled on the CPU it ran last on.
3233 */
3234 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3235#endif
3236 }
3237 }
3238
3239 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx "
3240 "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",
3241 cnum,
3242 value,
3243 is_loaded,
3244 can_access_pmu,
3245 hw_value,
3246 ctx->ctx_pmds[cnum].val,
3247 ctx->ctx_pmds[cnum].short_reset,
3248 ctx->ctx_pmds[cnum].long_reset,
3249 PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3250 ctx->ctx_pmds[cnum].seed,
3251 ctx->ctx_pmds[cnum].mask,
3252 ctx->ctx_used_pmds[0],
3253 ctx->ctx_pmds[cnum].reset_pmds[0],
3254 ctx->ctx_reload_pmds[0],
3255 ctx->ctx_all_pmds[0],
3256 ctx->ctx_ovfl_regs[0]));
3257 }
3258
3259 /*
3260 * make changes visible
3261 */
3262 if (can_access_pmu) ia64_srlz_d();
3263
3264 return 0;
3265
3266abort_mission:
3267 /*
3268 * for now, we have only one possibility for error
3269 */
3270 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3271 return ret;
3272}
3273
3274/*
3275 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3276 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3277 * interrupt is delivered during the call, it will be kept pending until we leave, making
3278 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3279 * guaranteed to return consistent data to the user, it may simply be old. It is not
3280 * trivial to treat the overflow while inside the call because you may end up in
3281 * some module sampling buffer code causing deadlocks.
3282 */
3283static int
3284pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3285{
3286 struct thread_struct *thread = NULL;
3287 struct task_struct *task;
3288 unsigned long val = 0UL, lval, ovfl_mask, sval;
3289 pfarg_reg_t *req = (pfarg_reg_t *)arg;
3290 unsigned int cnum, reg_flags = 0;
3291 int i, can_access_pmu = 0, state;
3292 int is_loaded, is_system, is_counting, expert_mode;
3293 int ret = -EINVAL;
3294 pfm_reg_check_t rd_func;
3295
3296 /*
3297 * access is possible when loaded only for
3298 * self-monitoring tasks or in UP mode
3299 */
3300
3301 state = ctx->ctx_state;
3302 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3303 is_system = ctx->ctx_fl_system;
3304 ovfl_mask = pmu_conf->ovfl_val;
3305 task = ctx->ctx_task;
3306
3307 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3308
3309 if (likely(is_loaded)) {
3310 thread = &task->thread;
3311 /*
3312 * In system wide and when the context is loaded, access can only happen
3313 * when the caller is running on the CPU being monitored by the session.
3314 * It does not have to be the owner (ctx_task) of the context per se.
3315 */
3316 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3317 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3318 return -EBUSY;
3319 }
3320 /*
3321 * this can be true when not self-monitoring only in UP
3322 */
3323 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3324
3325 if (can_access_pmu) ia64_srlz_d();
3326 }
3327 expert_mode = pfm_sysctl.expert_mode;
3328
3329 DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3330 is_loaded,
3331 can_access_pmu,
3332 state));
3333
3334 /*
3335 * on both UP and SMP, we can only read the PMD from the hardware register when
3336 * the task is the owner of the local PMU.
3337 */
3338
3339 for (i = 0; i < count; i++, req++) {
3340
3341 cnum = req->reg_num;
3342 reg_flags = req->reg_flags;
3343
3344 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3345 /*
3346 * we can only read the register that we use. That includes
3347 * the one we explicitely initialize AND the one we want included
3348 * in the sampling buffer (smpl_regs).
3349 *
3350 * Having this restriction allows optimization in the ctxsw routine
3351 * without compromising security (leaks)
3352 */
3353 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3354
3355 sval = ctx->ctx_pmds[cnum].val;
3356 lval = ctx->ctx_pmds[cnum].lval;
3357 is_counting = PMD_IS_COUNTING(cnum);
3358
3359 /*
3360 * If the task is not the current one, then we check if the
3361 * PMU state is still in the local live register due to lazy ctxsw.
3362 * If true, then we read directly from the registers.
3363 */
3364 if (can_access_pmu){
3365 val = ia64_get_pmd(cnum);
3366 } else {
3367 /*
3368 * context has been saved
3369 * if context is zombie, then task does not exist anymore.
3370 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3371 */
3372 val = is_loaded ? thread->pmds[cnum] : 0UL;
3373 }
3374 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3375
3376 if (is_counting) {
3377 /*
3378 * XXX: need to check for overflow when loaded
3379 */
3380 val &= ovfl_mask;
3381 val += sval;
3382 }
3383
3384 /*
3385 * execute read checker, if any
3386 */
3387 if (unlikely(expert_mode == 0 && rd_func)) {
3388 unsigned long v = val;
3389 ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3390 if (ret) goto error;
3391 val = v;
3392 ret = -EINVAL;
3393 }
3394
3395 PFM_REG_RETFLAG_SET(reg_flags, 0);
3396
3397 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3398
3399 /*
3400 * update register return value, abort all if problem during copy.
3401 * we only modify the reg_flags field. no check mode is fine because
3402 * access has been verified upfront in sys_perfmonctl().
3403 */
3404 req->reg_value = val;
3405 req->reg_flags = reg_flags;
3406 req->reg_last_reset_val = lval;
3407 }
3408
3409 return 0;
3410
3411error:
3412 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3413 return ret;
3414}
3415
3416int
3417pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3418{
3419 pfm_context_t *ctx;
3420
3421 if (req == NULL) return -EINVAL;
3422
3423 ctx = GET_PMU_CTX();
3424
3425 if (ctx == NULL) return -EINVAL;
3426
3427 /*
3428 * for now limit to current task, which is enough when calling
3429 * from overflow handler
3430 */
3431 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3432
3433 return pfm_write_pmcs(ctx, req, nreq, regs);
3434}
3435EXPORT_SYMBOL(pfm_mod_write_pmcs);
3436
3437int
3438pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3439{
3440 pfm_context_t *ctx;
3441
3442 if (req == NULL) return -EINVAL;
3443
3444 ctx = GET_PMU_CTX();
3445
3446 if (ctx == NULL) return -EINVAL;
3447
3448 /*
3449 * for now limit to current task, which is enough when calling
3450 * from overflow handler
3451 */
3452 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3453
3454 return pfm_read_pmds(ctx, req, nreq, regs);
3455}
3456EXPORT_SYMBOL(pfm_mod_read_pmds);
3457
3458/*
3459 * Only call this function when a process it trying to
3460 * write the debug registers (reading is always allowed)
3461 */
3462int
3463pfm_use_debug_registers(struct task_struct *task)
3464{
3465 pfm_context_t *ctx = task->thread.pfm_context;
3466 unsigned long flags;
3467 int ret = 0;
3468
3469 if (pmu_conf->use_rr_dbregs == 0) return 0;
3470
3471 DPRINT(("called for [%d]\n", task->pid));
3472
3473 /*
3474 * do it only once
3475 */
3476 if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3477
3478 /*
3479 * Even on SMP, we do not need to use an atomic here because
3480 * the only way in is via ptrace() and this is possible only when the
3481 * process is stopped. Even in the case where the ctxsw out is not totally
3482 * completed by the time we come here, there is no way the 'stopped' process
3483 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3484 * So this is always safe.
3485 */
3486 if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3487
3488 LOCK_PFS(flags);
3489
3490 /*
3491 * We cannot allow setting breakpoints when system wide monitoring
3492 * sessions are using the debug registers.
3493 */
3494 if (pfm_sessions.pfs_sys_use_dbregs> 0)
3495 ret = -1;
3496 else
3497 pfm_sessions.pfs_ptrace_use_dbregs++;
3498
3499 DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n",
3500 pfm_sessions.pfs_ptrace_use_dbregs,
3501 pfm_sessions.pfs_sys_use_dbregs,
3502 task->pid, ret));
3503
3504 UNLOCK_PFS(flags);
3505
3506 return ret;
3507}
3508
3509/*
3510 * This function is called for every task that exits with the
3511 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3512 * able to use the debug registers for debugging purposes via
3513 * ptrace(). Therefore we know it was not using them for
3514 * perfmormance monitoring, so we only decrement the number
3515 * of "ptraced" debug register users to keep the count up to date
3516 */
3517int
3518pfm_release_debug_registers(struct task_struct *task)
3519{
3520 unsigned long flags;
3521 int ret;
3522
3523 if (pmu_conf->use_rr_dbregs == 0) return 0;
3524
3525 LOCK_PFS(flags);
3526 if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3527 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
3528 ret = -1;
3529 } else {
3530 pfm_sessions.pfs_ptrace_use_dbregs--;
3531 ret = 0;
3532 }
3533 UNLOCK_PFS(flags);
3534
3535 return ret;
3536}
3537
3538static int
3539pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3540{
3541 struct task_struct *task;
3542 pfm_buffer_fmt_t *fmt;
3543 pfm_ovfl_ctrl_t rst_ctrl;
3544 int state, is_system;
3545 int ret = 0;
3546
3547 state = ctx->ctx_state;
3548 fmt = ctx->ctx_buf_fmt;
3549 is_system = ctx->ctx_fl_system;
3550 task = PFM_CTX_TASK(ctx);
3551
3552 switch(state) {
3553 case PFM_CTX_MASKED:
3554 break;
3555 case PFM_CTX_LOADED:
3556 if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3557 /* fall through */
3558 case PFM_CTX_UNLOADED:
3559 case PFM_CTX_ZOMBIE:
3560 DPRINT(("invalid state=%d\n", state));
3561 return -EBUSY;
3562 default:
3563 DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3564 return -EINVAL;
3565 }
3566
3567 /*
3568 * In system wide and when the context is loaded, access can only happen
3569 * when the caller is running on the CPU being monitored by the session.
3570 * It does not have to be the owner (ctx_task) of the context per se.
3571 */
3572 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3573 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3574 return -EBUSY;
3575 }
3576
3577 /* sanity check */
3578 if (unlikely(task == NULL)) {
3579 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
3580 return -EINVAL;
3581 }
3582
3583 if (task == current || is_system) {
3584
3585 fmt = ctx->ctx_buf_fmt;
3586
3587 DPRINT(("restarting self %d ovfl=0x%lx\n",
3588 task->pid,
3589 ctx->ctx_ovfl_regs[0]));
3590
3591 if (CTX_HAS_SMPL(ctx)) {
3592
3593 prefetch(ctx->ctx_smpl_hdr);
3594
3595 rst_ctrl.bits.mask_monitoring = 0;
3596 rst_ctrl.bits.reset_ovfl_pmds = 0;
3597
3598 if (state == PFM_CTX_LOADED)
3599 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3600 else
3601 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3602 } else {
3603 rst_ctrl.bits.mask_monitoring = 0;
3604 rst_ctrl.bits.reset_ovfl_pmds = 1;
3605 }
3606
3607 if (ret == 0) {
3608 if (rst_ctrl.bits.reset_ovfl_pmds)
3609 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3610
3611 if (rst_ctrl.bits.mask_monitoring == 0) {
3612 DPRINT(("resuming monitoring for [%d]\n", task->pid));
3613
3614 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3615 } else {
3616 DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));
3617
3618 // cannot use pfm_stop_monitoring(task, regs);
3619 }
3620 }
3621 /*
3622 * clear overflowed PMD mask to remove any stale information
3623 */
3624 ctx->ctx_ovfl_regs[0] = 0UL;
3625
3626 /*
3627 * back to LOADED state
3628 */
3629 ctx->ctx_state = PFM_CTX_LOADED;
3630
3631 /*
3632 * XXX: not really useful for self monitoring
3633 */
3634 ctx->ctx_fl_can_restart = 0;
3635
3636 return 0;
3637 }
3638
3639 /*
3640 * restart another task
3641 */
3642
3643 /*
3644 * When PFM_CTX_MASKED, we cannot issue a restart before the previous
3645 * one is seen by the task.
3646 */
3647 if (state == PFM_CTX_MASKED) {
3648 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3649 /*
3650 * will prevent subsequent restart before this one is
3651 * seen by other task
3652 */
3653 ctx->ctx_fl_can_restart = 0;
3654 }
3655
3656 /*
3657 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3658 * the task is blocked or on its way to block. That's the normal
3659 * restart path. If the monitoring is not masked, then the task
3660 * can be actively monitoring and we cannot directly intervene.
3661 * Therefore we use the trap mechanism to catch the task and
3662 * force it to reset the buffer/reset PMDs.
3663 *
3664 * if non-blocking, then we ensure that the task will go into
3665 * pfm_handle_work() before returning to user mode.
3666 *
3667 * We cannot explicitely reset another task, it MUST always
3668 * be done by the task itself. This works for system wide because
3669 * the tool that is controlling the session is logically doing
3670 * "self-monitoring".
3671 */
3672 if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3673 DPRINT(("unblocking [%d] \n", task->pid));
3674 up(&ctx->ctx_restart_sem);
3675 } else {
3676 DPRINT(("[%d] armed exit trap\n", task->pid));
3677
3678 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3679
3680 PFM_SET_WORK_PENDING(task, 1);
3681
3682 pfm_set_task_notify(task);
3683
3684 /*
3685 * XXX: send reschedule if task runs on another CPU
3686 */
3687 }
3688 return 0;
3689}
3690
3691static int
3692pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3693{
3694 unsigned int m = *(unsigned int *)arg;
3695
3696 pfm_sysctl.debug = m == 0 ? 0 : 1;
3697
3698 pfm_debug_var = pfm_sysctl.debug;
3699
3700 printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3701
3702 if (m == 0) {
3703 memset(pfm_stats, 0, sizeof(pfm_stats));
3704 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3705 }
3706 return 0;
3707}
3708
3709/*
3710 * arg can be NULL and count can be zero for this function
3711 */
3712static int
3713pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3714{
3715 struct thread_struct *thread = NULL;
3716 struct task_struct *task;
3717 pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3718 unsigned long flags;
3719 dbreg_t dbreg;
3720 unsigned int rnum;
3721 int first_time;
3722 int ret = 0, state;
3723 int i, can_access_pmu = 0;
3724 int is_system, is_loaded;
3725
3726 if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3727
3728 state = ctx->ctx_state;
3729 is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3730 is_system = ctx->ctx_fl_system;
3731 task = ctx->ctx_task;
3732
3733 if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3734
3735 /*
3736 * on both UP and SMP, we can only write to the PMC when the task is
3737 * the owner of the local PMU.
3738 */
3739 if (is_loaded) {
3740 thread = &task->thread;
3741 /*
3742 * In system wide and when the context is loaded, access can only happen
3743 * when the caller is running on the CPU being monitored by the session.
3744 * It does not have to be the owner (ctx_task) of the context per se.
3745 */
3746 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3747 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3748 return -EBUSY;
3749 }
3750 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3751 }
3752
3753 /*
3754 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3755 * ensuring that no real breakpoint can be installed via this call.
3756 *
3757 * IMPORTANT: regs can be NULL in this function
3758 */
3759
3760 first_time = ctx->ctx_fl_using_dbreg == 0;
3761
3762 /*
3763 * don't bother if we are loaded and task is being debugged
3764 */
3765 if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3766 DPRINT(("debug registers already in use for [%d]\n", task->pid));
3767 return -EBUSY;
3768 }
3769
3770 /*
3771 * check for debug registers in system wide mode
3772 *
3773 * If though a check is done in pfm_context_load(),
3774 * we must repeat it here, in case the registers are
3775 * written after the context is loaded
3776 */
3777 if (is_loaded) {
3778 LOCK_PFS(flags);
3779
3780 if (first_time && is_system) {
3781 if (pfm_sessions.pfs_ptrace_use_dbregs)
3782 ret = -EBUSY;
3783 else
3784 pfm_sessions.pfs_sys_use_dbregs++;
3785 }
3786 UNLOCK_PFS(flags);
3787 }
3788
3789 if (ret != 0) return ret;
3790
3791 /*
3792 * mark ourself as user of the debug registers for
3793 * perfmon purposes.
3794 */
3795 ctx->ctx_fl_using_dbreg = 1;
3796
3797 /*
3798 * clear hardware registers to make sure we don't
3799 * pick up stale state.
3800 *
3801 * for a system wide session, we do not use
3802 * thread.dbr, thread.ibr because this process
3803 * never leaves the current CPU and the state
3804 * is shared by all processes running on it
3805 */
3806 if (first_time && can_access_pmu) {
3807 DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
3808 for (i=0; i < pmu_conf->num_ibrs; i++) {
3809 ia64_set_ibr(i, 0UL);
3810 ia64_dv_serialize_instruction();
3811 }
3812 ia64_srlz_i();
3813 for (i=0; i < pmu_conf->num_dbrs; i++) {
3814 ia64_set_dbr(i, 0UL);
3815 ia64_dv_serialize_data();
3816 }
3817 ia64_srlz_d();
3818 }
3819
3820 /*
3821 * Now install the values into the registers
3822 */
3823 for (i = 0; i < count; i++, req++) {
3824
3825 rnum = req->dbreg_num;
3826 dbreg.val = req->dbreg_value;
3827
3828 ret = -EINVAL;
3829
3830 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3831 DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3832 rnum, dbreg.val, mode, i, count));
3833
3834 goto abort_mission;
3835 }
3836
3837 /*
3838 * make sure we do not install enabled breakpoint
3839 */
3840 if (rnum & 0x1) {
3841 if (mode == PFM_CODE_RR)
3842 dbreg.ibr.ibr_x = 0;
3843 else
3844 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3845 }
3846
3847 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3848
3849 /*
3850 * Debug registers, just like PMC, can only be modified
3851 * by a kernel call. Moreover, perfmon() access to those
3852 * registers are centralized in this routine. The hardware
3853 * does not modify the value of these registers, therefore,
3854 * if we save them as they are written, we can avoid having
3855 * to save them on context switch out. This is made possible
3856 * by the fact that when perfmon uses debug registers, ptrace()
3857 * won't be able to modify them concurrently.
3858 */
3859 if (mode == PFM_CODE_RR) {
3860 CTX_USED_IBR(ctx, rnum);
3861
3862 if (can_access_pmu) {
3863 ia64_set_ibr(rnum, dbreg.val);
3864 ia64_dv_serialize_instruction();
3865 }
3866
3867 ctx->ctx_ibrs[rnum] = dbreg.val;
3868
3869 DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3870 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3871 } else {
3872 CTX_USED_DBR(ctx, rnum);
3873
3874 if (can_access_pmu) {
3875 ia64_set_dbr(rnum, dbreg.val);
3876 ia64_dv_serialize_data();
3877 }
3878 ctx->ctx_dbrs[rnum] = dbreg.val;
3879
3880 DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3881 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3882 }
3883 }
3884
3885 return 0;
3886
3887abort_mission:
3888 /*
3889 * in case it was our first attempt, we undo the global modifications
3890 */
3891 if (first_time) {
3892 LOCK_PFS(flags);
3893 if (ctx->ctx_fl_system) {
3894 pfm_sessions.pfs_sys_use_dbregs--;
3895 }
3896 UNLOCK_PFS(flags);
3897 ctx->ctx_fl_using_dbreg = 0;
3898 }
3899 /*
3900 * install error return flag
3901 */
3902 PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3903
3904 return ret;
3905}
3906
3907static int
3908pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3909{
3910 return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3911}
3912
3913static int
3914pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3915{
3916 return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3917}
3918
3919int
3920pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3921{
3922 pfm_context_t *ctx;
3923
3924 if (req == NULL) return -EINVAL;
3925
3926 ctx = GET_PMU_CTX();
3927
3928 if (ctx == NULL) return -EINVAL;
3929
3930 /*
3931 * for now limit to current task, which is enough when calling
3932 * from overflow handler
3933 */
3934 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3935
3936 return pfm_write_ibrs(ctx, req, nreq, regs);
3937}
3938EXPORT_SYMBOL(pfm_mod_write_ibrs);
3939
3940int
3941pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3942{
3943 pfm_context_t *ctx;
3944
3945 if (req == NULL) return -EINVAL;
3946
3947 ctx = GET_PMU_CTX();
3948
3949 if (ctx == NULL) return -EINVAL;
3950
3951 /*
3952 * for now limit to current task, which is enough when calling
3953 * from overflow handler
3954 */
3955 if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3956
3957 return pfm_write_dbrs(ctx, req, nreq, regs);
3958}
3959EXPORT_SYMBOL(pfm_mod_write_dbrs);
3960
3961
3962static int
3963pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3964{
3965 pfarg_features_t *req = (pfarg_features_t *)arg;
3966
3967 req->ft_version = PFM_VERSION;
3968 return 0;
3969}
3970
3971static int
3972pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3973{
3974 struct pt_regs *tregs;
3975 struct task_struct *task = PFM_CTX_TASK(ctx);
3976 int state, is_system;
3977
3978 state = ctx->ctx_state;
3979 is_system = ctx->ctx_fl_system;
3980
3981 /*
3982 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3983 */
3984 if (state == PFM_CTX_UNLOADED) return -EINVAL;
3985
3986 /*
3987 * In system wide and when the context is loaded, access can only happen
3988 * when the caller is running on the CPU being monitored by the session.
3989 * It does not have to be the owner (ctx_task) of the context per se.
3990 */
3991 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3992 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3993 return -EBUSY;
3994 }
3995 DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3996 PFM_CTX_TASK(ctx)->pid,
3997 state,
3998 is_system));
3999 /*
4000 * in system mode, we need to update the PMU directly
4001 * and the user level state of the caller, which may not
4002 * necessarily be the creator of the context.
4003 */
4004 if (is_system) {
4005 /*
4006 * Update local PMU first
4007 *
4008 * disable dcr pp
4009 */
4010 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4011 ia64_srlz_i();
4012
4013 /*
4014 * update local cpuinfo
4015 */
4016 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4017
4018 /*
4019 * stop monitoring, does srlz.i
4020 */
4021 pfm_clear_psr_pp();
4022
4023 /*
4024 * stop monitoring in the caller
4025 */
4026 ia64_psr(regs)->pp = 0;
4027
4028 return 0;
4029 }
4030 /*
4031 * per-task mode
4032 */
4033
4034 if (task == current) {
4035 /* stop monitoring at kernel level */
4036 pfm_clear_psr_up();
4037
4038 /*
4039 * stop monitoring at the user level
4040 */
4041 ia64_psr(regs)->up = 0;
4042 } else {
4043 tregs = ia64_task_regs(task);
4044
4045 /*
4046 * stop monitoring at the user level
4047 */
4048 ia64_psr(tregs)->up = 0;
4049
4050 /*
4051 * monitoring disabled in kernel at next reschedule
4052 */
4053 ctx->ctx_saved_psr_up = 0;
4054 DPRINT(("task=[%d]\n", task->pid));
4055 }
4056 return 0;
4057}
4058
4059
4060static int
4061pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4062{
4063 struct pt_regs *tregs;
4064 int state, is_system;
4065
4066 state = ctx->ctx_state;
4067 is_system = ctx->ctx_fl_system;
4068
4069 if (state != PFM_CTX_LOADED) return -EINVAL;
4070
4071 /*
4072 * In system wide and when the context is loaded, access can only happen
4073 * when the caller is running on the CPU being monitored by the session.
4074 * It does not have to be the owner (ctx_task) of the context per se.
4075 */
4076 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4077 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4078 return -EBUSY;
4079 }
4080
4081 /*
4082 * in system mode, we need to update the PMU directly
4083 * and the user level state of the caller, which may not
4084 * necessarily be the creator of the context.
4085 */
4086 if (is_system) {
4087
4088 /*
4089 * set user level psr.pp for the caller
4090 */
4091 ia64_psr(regs)->pp = 1;
4092
4093 /*
4094 * now update the local PMU and cpuinfo
4095 */
4096 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4097
4098 /*
4099 * start monitoring at kernel level
4100 */
4101 pfm_set_psr_pp();
4102
4103 /* enable dcr pp */
4104 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4105 ia64_srlz_i();
4106
4107 return 0;
4108 }
4109
4110 /*
4111 * per-process mode
4112 */
4113
4114 if (ctx->ctx_task == current) {
4115
4116 /* start monitoring at kernel level */
4117 pfm_set_psr_up();
4118
4119 /*
4120 * activate monitoring at user level
4121 */
4122 ia64_psr(regs)->up = 1;
4123
4124 } else {
4125 tregs = ia64_task_regs(ctx->ctx_task);
4126
4127 /*
4128 * start monitoring at the kernel level the next
4129 * time the task is scheduled
4130 */
4131 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4132
4133 /*
4134 * activate monitoring at user level
4135 */
4136 ia64_psr(tregs)->up = 1;
4137 }
4138 return 0;
4139}
4140
4141static int
4142pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4143{
4144 pfarg_reg_t *req = (pfarg_reg_t *)arg;
4145 unsigned int cnum;
4146 int i;
4147 int ret = -EINVAL;
4148
4149 for (i = 0; i < count; i++, req++) {
4150
4151 cnum = req->reg_num;
4152
4153 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4154
4155 req->reg_value = PMC_DFL_VAL(cnum);
4156
4157 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4158
4159 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4160 }
4161 return 0;
4162
4163abort_mission:
4164 PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4165 return ret;
4166}
4167
4168static int
4169pfm_check_task_exist(pfm_context_t *ctx)
4170{
4171 struct task_struct *g, *t;
4172 int ret = -ESRCH;
4173
4174 read_lock(&tasklist_lock);
4175
4176 do_each_thread (g, t) {
4177 if (t->thread.pfm_context == ctx) {
4178 ret = 0;
4179 break;
4180 }
4181 } while_each_thread (g, t);
4182
4183 read_unlock(&tasklist_lock);
4184
4185 DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4186
4187 return ret;
4188}
4189
4190static int
4191pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4192{
4193 struct task_struct *task;
4194 struct thread_struct *thread;
4195 struct pfm_context_t *old;
4196 unsigned long flags;
4197#ifndef CONFIG_SMP
4198 struct task_struct *owner_task = NULL;
4199#endif
4200 pfarg_load_t *req = (pfarg_load_t *)arg;
4201 unsigned long *pmcs_source, *pmds_source;
4202 int the_cpu;
4203 int ret = 0;
4204 int state, is_system, set_dbregs = 0;
4205
4206 state = ctx->ctx_state;
4207 is_system = ctx->ctx_fl_system;
4208 /*
4209 * can only load from unloaded or terminated state
4210 */
4211 if (state != PFM_CTX_UNLOADED) {
4212 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4213 req->load_pid,
4214 ctx->ctx_state));
4215 return -EINVAL;
4216 }
4217
4218 DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4219
4220 if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4221 DPRINT(("cannot use blocking mode on self\n"));
4222 return -EINVAL;
4223 }
4224
4225 ret = pfm_get_task(ctx, req->load_pid, &task);
4226 if (ret) {
4227 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4228 return ret;
4229 }
4230
4231 ret = -EINVAL;
4232
4233 /*
4234 * system wide is self monitoring only
4235 */
4236 if (is_system && task != current) {
4237 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4238 req->load_pid));
4239 goto error;
4240 }
4241
4242 thread = &task->thread;
4243
4244 ret = 0;
4245 /*
4246 * cannot load a context which is using range restrictions,
4247 * into a task that is being debugged.
4248 */
4249 if (ctx->ctx_fl_using_dbreg) {
4250 if (thread->flags & IA64_THREAD_DBG_VALID) {
4251 ret = -EBUSY;
4252 DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4253 goto error;
4254 }
4255 LOCK_PFS(flags);
4256
4257 if (is_system) {
4258 if (pfm_sessions.pfs_ptrace_use_dbregs) {
4259 DPRINT(("cannot load [%d] dbregs in use\n", task->pid));
4260 ret = -EBUSY;
4261 } else {
4262 pfm_sessions.pfs_sys_use_dbregs++;
4263 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs));
4264 set_dbregs = 1;
4265 }
4266 }
4267
4268 UNLOCK_PFS(flags);
4269
4270 if (ret) goto error;
4271 }
4272
4273 /*
4274 * SMP system-wide monitoring implies self-monitoring.
4275 *
4276 * The programming model expects the task to
4277 * be pinned on a CPU throughout the session.
4278 * Here we take note of the current CPU at the
4279 * time the context is loaded. No call from
4280 * another CPU will be allowed.
4281 *
4282 * The pinning via shed_setaffinity()
4283 * must be done by the calling task prior
4284 * to this call.
4285 *
4286 * systemwide: keep track of CPU this session is supposed to run on
4287 */
4288 the_cpu = ctx->ctx_cpu = smp_processor_id();
4289
4290 ret = -EBUSY;
4291 /*
4292 * now reserve the session
4293 */
4294 ret = pfm_reserve_session(current, is_system, the_cpu);
4295 if (ret) goto error;
4296
4297 /*
4298 * task is necessarily stopped at this point.
4299 *
4300 * If the previous context was zombie, then it got removed in
4301 * pfm_save_regs(). Therefore we should not see it here.
4302 * If we see a context, then this is an active context
4303 *
4304 * XXX: needs to be atomic
4305 */
4306 DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4307 thread->pfm_context, ctx));
4308
4309 old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4310 if (old != NULL) {
4311 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4312 goto error_unres;
4313 }
4314
4315 pfm_reset_msgq(ctx);
4316
4317 ctx->ctx_state = PFM_CTX_LOADED;
4318
4319 /*
4320 * link context to task
4321 */
4322 ctx->ctx_task = task;
4323
4324 if (is_system) {
4325 /*
4326 * we load as stopped
4327 */
4328 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4329 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4330
4331 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4332 } else {
4333 thread->flags |= IA64_THREAD_PM_VALID;
4334 }
4335
4336 /*
4337 * propagate into thread-state
4338 */
4339 pfm_copy_pmds(task, ctx);
4340 pfm_copy_pmcs(task, ctx);
4341
4342 pmcs_source = thread->pmcs;
4343 pmds_source = thread->pmds;
4344
4345 /*
4346 * always the case for system-wide
4347 */
4348 if (task == current) {
4349
4350 if (is_system == 0) {
4351
4352 /* allow user level control */
4353 ia64_psr(regs)->sp = 0;
4354 DPRINT(("clearing psr.sp for [%d]\n", task->pid));
4355
4356 SET_LAST_CPU(ctx, smp_processor_id());
4357 INC_ACTIVATION();
4358 SET_ACTIVATION(ctx);
4359#ifndef CONFIG_SMP
4360 /*
4361 * push the other task out, if any
4362 */
4363 owner_task = GET_PMU_OWNER();
4364 if (owner_task) pfm_lazy_save_regs(owner_task);
4365#endif
4366 }
4367 /*
4368 * load all PMD from ctx to PMU (as opposed to thread state)
4369 * restore all PMC from ctx to PMU
4370 */
4371 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4372 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4373
4374 ctx->ctx_reload_pmcs[0] = 0UL;
4375 ctx->ctx_reload_pmds[0] = 0UL;
4376
4377 /*
4378 * guaranteed safe by earlier check against DBG_VALID
4379 */
4380 if (ctx->ctx_fl_using_dbreg) {
4381 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4382 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4383 }
4384 /*
4385 * set new ownership
4386 */
4387 SET_PMU_OWNER(task, ctx);
4388
4389 DPRINT(("context loaded on PMU for [%d]\n", task->pid));
4390 } else {
4391 /*
4392 * when not current, task MUST be stopped, so this is safe
4393 */
4394 regs = ia64_task_regs(task);
4395
4396 /* force a full reload */
4397 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4398 SET_LAST_CPU(ctx, -1);
4399
4400 /* initial saved psr (stopped) */
4401 ctx->ctx_saved_psr_up = 0UL;
4402 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4403 }
4404
4405 ret = 0;
4406
4407error_unres:
4408 if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4409error:
4410 /*
4411 * we must undo the dbregs setting (for system-wide)
4412 */
4413 if (ret && set_dbregs) {
4414 LOCK_PFS(flags);
4415 pfm_sessions.pfs_sys_use_dbregs--;
4416 UNLOCK_PFS(flags);
4417 }
4418 /*
4419 * release task, there is now a link with the context
4420 */
4421 if (is_system == 0 && task != current) {
4422 pfm_put_task(task);
4423
4424 if (ret == 0) {
4425 ret = pfm_check_task_exist(ctx);
4426 if (ret) {
4427 ctx->ctx_state = PFM_CTX_UNLOADED;
4428 ctx->ctx_task = NULL;
4429 }
4430 }
4431 }
4432 return ret;
4433}
4434
4435/*
4436 * in this function, we do not need to increase the use count
4437 * for the task via get_task_struct(), because we hold the
4438 * context lock. If the task were to disappear while having
4439 * a context attached, it would go through pfm_exit_thread()
4440 * which also grabs the context lock and would therefore be blocked
4441 * until we are here.
4442 */
4443static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4444
4445static int
4446pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4447{
4448 struct task_struct *task = PFM_CTX_TASK(ctx);
4449 struct pt_regs *tregs;
4450 int prev_state, is_system;
4451 int ret;
4452
4453 DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1));
4454
4455 prev_state = ctx->ctx_state;
4456 is_system = ctx->ctx_fl_system;
4457
4458 /*
4459 * unload only when necessary
4460 */
4461 if (prev_state == PFM_CTX_UNLOADED) {
4462 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4463 return 0;
4464 }
4465
4466 /*
4467 * clear psr and dcr bits
4468 */
4469 ret = pfm_stop(ctx, NULL, 0, regs);
4470 if (ret) return ret;
4471
4472 ctx->ctx_state = PFM_CTX_UNLOADED;
4473
4474 /*
4475 * in system mode, we need to update the PMU directly
4476 * and the user level state of the caller, which may not
4477 * necessarily be the creator of the context.
4478 */
4479 if (is_system) {
4480
4481 /*
4482 * Update cpuinfo
4483 *
4484 * local PMU is taken care of in pfm_stop()
4485 */
4486 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4487 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4488
4489 /*
4490 * save PMDs in context
4491 * release ownership
4492 */
4493 pfm_flush_pmds(current, ctx);
4494
4495 /*
4496 * at this point we are done with the PMU
4497 * so we can unreserve the resource.
4498 */
4499 if (prev_state != PFM_CTX_ZOMBIE)
4500 pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4501
4502 /*
4503 * disconnect context from task
4504 */
4505 task->thread.pfm_context = NULL;
4506 /*
4507 * disconnect task from context
4508 */
4509 ctx->ctx_task = NULL;
4510
4511 /*
4512 * There is nothing more to cleanup here.
4513 */
4514 return 0;
4515 }
4516
4517 /*
4518 * per-task mode
4519 */
4520 tregs = task == current ? regs : ia64_task_regs(task);
4521
4522 if (task == current) {
4523 /*
4524 * cancel user level control
4525 */
4526 ia64_psr(regs)->sp = 1;
4527
4528 DPRINT(("setting psr.sp for [%d]\n", task->pid));
4529 }
4530 /*
4531 * save PMDs to context
4532 * release ownership
4533 */
4534 pfm_flush_pmds(task, ctx);
4535
4536 /*
4537 * at this point we are done with the PMU
4538 * so we can unreserve the resource.
4539 *
4540 * when state was ZOMBIE, we have already unreserved.
4541 */
4542 if (prev_state != PFM_CTX_ZOMBIE)
4543 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4544
4545 /*
4546 * reset activation counter and psr
4547 */
4548 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4549 SET_LAST_CPU(ctx, -1);
4550
4551 /*
4552 * PMU state will not be restored
4553 */
4554 task->thread.flags &= ~IA64_THREAD_PM_VALID;
4555
4556 /*
4557 * break links between context and task
4558 */
4559 task->thread.pfm_context = NULL;
4560 ctx->ctx_task = NULL;
4561
4562 PFM_SET_WORK_PENDING(task, 0);
4563
4564 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4565 ctx->ctx_fl_can_restart = 0;
4566 ctx->ctx_fl_going_zombie = 0;
4567
4568 DPRINT(("disconnected [%d] from context\n", task->pid));
4569
4570 return 0;
4571}
4572
4573
4574/*
4575 * called only from exit_thread(): task == current
4576 * we come here only if current has a context attached (loaded or masked)
4577 */
4578void
4579pfm_exit_thread(struct task_struct *task)
4580{
4581 pfm_context_t *ctx;
4582 unsigned long flags;
4583 struct pt_regs *regs = ia64_task_regs(task);
4584 int ret, state;
4585 int free_ok = 0;
4586
4587 ctx = PFM_GET_CTX(task);
4588
4589 PROTECT_CTX(ctx, flags);
4590
4591 DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid));
4592
4593 state = ctx->ctx_state;
4594 switch(state) {
4595 case PFM_CTX_UNLOADED:
4596 /*
4597 * only comes to thios function if pfm_context is not NULL, i.e., cannot
4598 * be in unloaded state
4599 */
4600 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid);
4601 break;
4602 case PFM_CTX_LOADED:
4603 case PFM_CTX_MASKED:
4604 ret = pfm_context_unload(ctx, NULL, 0, regs);
4605 if (ret) {
4606 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret);
4607 }
4608 DPRINT(("ctx unloaded for current state was %d\n", state));
4609
4610 pfm_end_notify_user(ctx);
4611 break;
4612 case PFM_CTX_ZOMBIE:
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, state, ret);
4616 }
4617 free_ok = 1;
4618 break;
4619 default:
4620 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state);
4621 break;
4622 }
4623 UNPROTECT_CTX(ctx, flags);
4624
4625 { u64 psr = pfm_get_psr();
4626 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4627 BUG_ON(GET_PMU_OWNER());
4628 BUG_ON(ia64_psr(regs)->up);
4629 BUG_ON(ia64_psr(regs)->pp);
4630 }
4631
4632 /*
4633 * All memory free operations (especially for vmalloc'ed memory)
4634 * MUST be done with interrupts ENABLED.
4635 */
4636 if (free_ok) pfm_context_free(ctx);
4637}
4638
4639/*
4640 * functions MUST be listed in the increasing order of their index (see permfon.h)
4641 */
4642#define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4643#define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4644#define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4645#define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW)
4646#define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL}
4647
4648static pfm_cmd_desc_t pfm_cmd_tab[]={
4649/* 0 */PFM_CMD_NONE,
4650/* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4651/* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4652/* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4653/* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4654/* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4655/* 6 */PFM_CMD_NONE,
4656/* 7 */PFM_CMD_NONE,
4657/* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4658/* 9 */PFM_CMD_NONE,
4659/* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4660/* 11 */PFM_CMD_NONE,
4661/* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4662/* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4663/* 14 */PFM_CMD_NONE,
4664/* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4665/* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4666/* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4667/* 18 */PFM_CMD_NONE,
4668/* 19 */PFM_CMD_NONE,
4669/* 20 */PFM_CMD_NONE,
4670/* 21 */PFM_CMD_NONE,
4671/* 22 */PFM_CMD_NONE,
4672/* 23 */PFM_CMD_NONE,
4673/* 24 */PFM_CMD_NONE,
4674/* 25 */PFM_CMD_NONE,
4675/* 26 */PFM_CMD_NONE,
4676/* 27 */PFM_CMD_NONE,
4677/* 28 */PFM_CMD_NONE,
4678/* 29 */PFM_CMD_NONE,
4679/* 30 */PFM_CMD_NONE,
4680/* 31 */PFM_CMD_NONE,
4681/* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4682/* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4683};
4684#define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4685
4686static int
4687pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4688{
4689 struct task_struct *task;
4690 int state, old_state;
4691
4692recheck:
4693 state = ctx->ctx_state;
4694 task = ctx->ctx_task;
4695
4696 if (task == NULL) {
4697 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4698 return 0;
4699 }
4700
4701 DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4702 ctx->ctx_fd,
4703 state,
4704 task->pid,
4705 task->state, PFM_CMD_STOPPED(cmd)));
4706
4707 /*
4708 * self-monitoring always ok.
4709 *
4710 * for system-wide the caller can either be the creator of the
4711 * context (to one to which the context is attached to) OR
4712 * a task running on the same CPU as the session.
4713 */
4714 if (task == current || ctx->ctx_fl_system) return 0;
4715
4716 /*
4717 * if context is UNLOADED we are safe to go
4718 */
4719 if (state == PFM_CTX_UNLOADED) return 0;
4720
4721 /*
4722 * no command can operate on a zombie context
4723 */
4724 if (state == PFM_CTX_ZOMBIE) {
4725 DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4726 return -EINVAL;
4727 }
4728
4729 /*
4730 * context is LOADED or MASKED. Some commands may need to have
4731 * the task stopped.
4732 *
4733 * We could lift this restriction for UP but it would mean that
4734 * the user has no guarantee the task would not run between
4735 * two successive calls to perfmonctl(). That's probably OK.
4736 * If this user wants to ensure the task does not run, then
4737 * the task must be stopped.
4738 */
4739 if (PFM_CMD_STOPPED(cmd)) {
4740 if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
4741 DPRINT(("[%d] task not in stopped state\n", task->pid));
4742 return -EBUSY;
4743 }
4744 /*
4745 * task is now stopped, wait for ctxsw out
4746 *
4747 * This is an interesting point in the code.
4748 * We need to unprotect the context because
4749 * the pfm_save_regs() routines needs to grab
4750 * the same lock. There are danger in doing
4751 * this because it leaves a window open for
4752 * another task to get access to the context
4753 * and possibly change its state. The one thing
4754 * that is not possible is for the context to disappear
4755 * because we are protected by the VFS layer, i.e.,
4756 * get_fd()/put_fd().
4757 */
4758 old_state = state;
4759
4760 UNPROTECT_CTX(ctx, flags);
4761
4762 wait_task_inactive(task);
4763
4764 PROTECT_CTX(ctx, flags);
4765
4766 /*
4767 * we must recheck to verify if state has changed
4768 */
4769 if (ctx->ctx_state != old_state) {
4770 DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4771 goto recheck;
4772 }
4773 }
4774 return 0;
4775}
4776
4777/*
4778 * system-call entry point (must return long)
4779 */
4780asmlinkage long
4781sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4782{
4783 struct file *file = NULL;
4784 pfm_context_t *ctx = NULL;
4785 unsigned long flags = 0UL;
4786 void *args_k = NULL;
4787 long ret; /* will expand int return types */
4788 size_t base_sz, sz, xtra_sz = 0;
4789 int narg, completed_args = 0, call_made = 0, cmd_flags;
4790 int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4791 int (*getsize)(void *arg, size_t *sz);
4792#define PFM_MAX_ARGSIZE 4096
4793
4794 /*
4795 * reject any call if perfmon was disabled at initialization
4796 */
4797 if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4798
4799 if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4800 DPRINT(("invalid cmd=%d\n", cmd));
4801 return -EINVAL;
4802 }
4803
4804 func = pfm_cmd_tab[cmd].cmd_func;
4805 narg = pfm_cmd_tab[cmd].cmd_narg;
4806 base_sz = pfm_cmd_tab[cmd].cmd_argsize;
4807 getsize = pfm_cmd_tab[cmd].cmd_getsize;
4808 cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4809
4810 if (unlikely(func == NULL)) {
4811 DPRINT(("invalid cmd=%d\n", cmd));
4812 return -EINVAL;
4813 }
4814
4815 DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4816 PFM_CMD_NAME(cmd),
4817 cmd,
4818 narg,
4819 base_sz,
4820 count));
4821
4822 /*
4823 * check if number of arguments matches what the command expects
4824 */
4825 if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4826 return -EINVAL;
4827
4828restart_args:
4829 sz = xtra_sz + base_sz*count;
4830 /*
4831 * limit abuse to min page size
4832 */
4833 if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4834 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz);
4835 return -E2BIG;
4836 }
4837
4838 /*
4839 * allocate default-sized argument buffer
4840 */
4841 if (likely(count && args_k == NULL)) {
4842 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4843 if (args_k == NULL) return -ENOMEM;
4844 }
4845
4846 ret = -EFAULT;
4847
4848 /*
4849 * copy arguments
4850 *
4851 * assume sz = 0 for command without parameters
4852 */
4853 if (sz && copy_from_user(args_k, arg, sz)) {
4854 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4855 goto error_args;
4856 }
4857
4858 /*
4859 * check if command supports extra parameters
4860 */
4861 if (completed_args == 0 && getsize) {
4862 /*
4863 * get extra parameters size (based on main argument)
4864 */
4865 ret = (*getsize)(args_k, &xtra_sz);
4866 if (ret) goto error_args;
4867
4868 completed_args = 1;
4869
4870 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4871
4872 /* retry if necessary */
4873 if (likely(xtra_sz)) goto restart_args;
4874 }
4875
4876 if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4877
4878 ret = -EBADF;
4879
4880 file = fget(fd);
4881 if (unlikely(file == NULL)) {
4882 DPRINT(("invalid fd %d\n", fd));
4883 goto error_args;
4884 }
4885 if (unlikely(PFM_IS_FILE(file) == 0)) {
4886 DPRINT(("fd %d not related to perfmon\n", fd));
4887 goto error_args;
4888 }
4889
4890 ctx = (pfm_context_t *)file->private_data;
4891 if (unlikely(ctx == NULL)) {
4892 DPRINT(("no context for fd %d\n", fd));
4893 goto error_args;
4894 }
4895 prefetch(&ctx->ctx_state);
4896
4897 PROTECT_CTX(ctx, flags);
4898
4899 /*
4900 * check task is stopped
4901 */
4902 ret = pfm_check_task_state(ctx, cmd, flags);
4903 if (unlikely(ret)) goto abort_locked;
4904
4905skip_fd:
4906 ret = (*func)(ctx, args_k, count, ia64_task_regs(current));
4907
4908 call_made = 1;
4909
4910abort_locked:
4911 if (likely(ctx)) {
4912 DPRINT(("context unlocked\n"));
4913 UNPROTECT_CTX(ctx, flags);
4914 fput(file);
4915 }
4916
4917 /* copy argument back to user, if needed */
4918 if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4919
4920error_args:
4921 if (args_k) kfree(args_k);
4922
4923 DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4924
4925 return ret;
4926}
4927
4928static void
4929pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4930{
4931 pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4932 pfm_ovfl_ctrl_t rst_ctrl;
4933 int state;
4934 int ret = 0;
4935
4936 state = ctx->ctx_state;
4937 /*
4938 * Unlock sampling buffer and reset index atomically
4939 * XXX: not really needed when blocking
4940 */
4941 if (CTX_HAS_SMPL(ctx)) {
4942
4943 rst_ctrl.bits.mask_monitoring = 0;
4944 rst_ctrl.bits.reset_ovfl_pmds = 0;
4945
4946 if (state == PFM_CTX_LOADED)
4947 ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4948 else
4949 ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4950 } else {
4951 rst_ctrl.bits.mask_monitoring = 0;
4952 rst_ctrl.bits.reset_ovfl_pmds = 1;
4953 }
4954
4955 if (ret == 0) {
4956 if (rst_ctrl.bits.reset_ovfl_pmds) {
4957 pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4958 }
4959 if (rst_ctrl.bits.mask_monitoring == 0) {
4960 DPRINT(("resuming monitoring\n"));
4961 if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4962 } else {
4963 DPRINT(("stopping monitoring\n"));
4964 //pfm_stop_monitoring(current, regs);
4965 }
4966 ctx->ctx_state = PFM_CTX_LOADED;
4967 }
4968}
4969
4970/*
4971 * context MUST BE LOCKED when calling
4972 * can only be called for current
4973 */
4974static void
4975pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4976{
4977 int ret;
4978
4979 DPRINT(("entering for [%d]\n", current->pid));
4980
4981 ret = pfm_context_unload(ctx, NULL, 0, regs);
4982 if (ret) {
4983 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", current->pid, ret);
4984 }
4985
4986 /*
4987 * and wakeup controlling task, indicating we are now disconnected
4988 */
4989 wake_up_interruptible(&ctx->ctx_zombieq);
4990
4991 /*
4992 * given that context is still locked, the controlling
4993 * task will only get access when we return from
4994 * pfm_handle_work().
4995 */
4996}
4997
4998static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4999
5000void
5001pfm_handle_work(void)
5002{
5003 pfm_context_t *ctx;
5004 struct pt_regs *regs;
5005 unsigned long flags;
5006 unsigned long ovfl_regs;
5007 unsigned int reason;
5008 int ret;
5009
5010 ctx = PFM_GET_CTX(current);
5011 if (ctx == NULL) {
5012 printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid);
5013 return;
5014 }
5015
5016 PROTECT_CTX(ctx, flags);
5017
5018 PFM_SET_WORK_PENDING(current, 0);
5019
5020 pfm_clear_task_notify();
5021
5022 regs = ia64_task_regs(current);
5023
5024 /*
5025 * extract reason for being here and clear
5026 */
5027 reason = ctx->ctx_fl_trap_reason;
5028 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5029 ovfl_regs = ctx->ctx_ovfl_regs[0];
5030
5031 DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5032
5033 /*
5034 * must be done before we check for simple-reset mode
5035 */
5036 if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie;
5037
5038
5039 //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5040 if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
5041
5042 UNPROTECT_CTX(ctx, flags);
5043
5044 /*
5045 * pfm_handle_work() is currently called with interrupts disabled.
5046 * The down_interruptible call may sleep, therefore we
5047 * must re-enable interrupts to avoid deadlocks. It is
5048 * safe to do so because this function is called ONLY
5049 * when returning to user level (PUStk=1), in which case
5050 * there is no risk of kernel stack overflow due to deep
5051 * interrupt nesting.
5052 */
5053 BUG_ON(flags & IA64_PSR_I);
5054 local_irq_enable();
5055
5056 DPRINT(("before block sleeping\n"));
5057
5058 /*
5059 * may go through without blocking on SMP systems
5060 * if restart has been received already by the time we call down()
5061 */
5062 ret = down_interruptible(&ctx->ctx_restart_sem);
5063
5064 DPRINT(("after block sleeping ret=%d\n", ret));
5065
5066 /*
5067 * disable interrupts to restore state we had upon entering
5068 * this function
5069 */
5070 local_irq_disable();
5071
5072 PROTECT_CTX(ctx, flags);
5073
5074 /*
5075 * we need to read the ovfl_regs only after wake-up
5076 * because we may have had pfm_write_pmds() in between
5077 * and that can changed PMD values and therefore
5078 * ovfl_regs is reset for these new PMD values.
5079 */
5080 ovfl_regs = ctx->ctx_ovfl_regs[0];
5081
5082 if (ctx->ctx_fl_going_zombie) {
5083do_zombie:
5084 DPRINT(("context is zombie, bailing out\n"));
5085 pfm_context_force_terminate(ctx, regs);
5086 goto nothing_to_do;
5087 }
5088 /*
5089 * in case of interruption of down() we don't restart anything
5090 */
5091 if (ret < 0) goto nothing_to_do;
5092
5093skip_blocking:
5094 pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5095 ctx->ctx_ovfl_regs[0] = 0UL;
5096
5097nothing_to_do:
5098
5099 UNPROTECT_CTX(ctx, flags);
5100}
5101
5102static int
5103pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5104{
5105 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5106 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5107 return 0;
5108 }
5109
5110 DPRINT(("waking up somebody\n"));
5111
5112 if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5113
5114 /*
5115 * safe, we are not in intr handler, nor in ctxsw when
5116 * we come here
5117 */
5118 kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5119
5120 return 0;
5121}
5122
5123static int
5124pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5125{
5126 pfm_msg_t *msg = NULL;
5127
5128 if (ctx->ctx_fl_no_msg == 0) {
5129 msg = pfm_get_new_msg(ctx);
5130 if (msg == NULL) {
5131 printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5132 return -1;
5133 }
5134
5135 msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL;
5136 msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd;
5137 msg->pfm_ovfl_msg.msg_active_set = 0;
5138 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5139 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5140 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5141 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5142 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5143 }
5144
5145 DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5146 msg,
5147 ctx->ctx_fl_no_msg,
5148 ctx->ctx_fd,
5149 ovfl_pmds));
5150
5151 return pfm_notify_user(ctx, msg);
5152}
5153
5154static int
5155pfm_end_notify_user(pfm_context_t *ctx)
5156{
5157 pfm_msg_t *msg;
5158
5159 msg = pfm_get_new_msg(ctx);
5160 if (msg == NULL) {
5161 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5162 return -1;
5163 }
5164 /* no leak */
5165 memset(msg, 0, sizeof(*msg));
5166
5167 msg->pfm_end_msg.msg_type = PFM_MSG_END;
5168 msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd;
5169 msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5170
5171 DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5172 msg,
5173 ctx->ctx_fl_no_msg,
5174 ctx->ctx_fd));
5175
5176 return pfm_notify_user(ctx, msg);
5177}
5178
5179/*
5180 * main overflow processing routine.
5181 * it can be called from the interrupt path or explicitely during the context switch code
5182 */
5183static void
5184pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs)
5185{
5186 pfm_ovfl_arg_t *ovfl_arg;
5187 unsigned long mask;
5188 unsigned long old_val, ovfl_val, new_val;
5189 unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5190 unsigned long tstamp;
5191 pfm_ovfl_ctrl_t ovfl_ctrl;
5192 unsigned int i, has_smpl;
5193 int must_notify = 0;
5194
5195 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5196
5197 /*
5198 * sanity test. Should never happen
5199 */
5200 if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5201
5202 tstamp = ia64_get_itc();
5203 mask = pmc0 >> PMU_FIRST_COUNTER;
5204 ovfl_val = pmu_conf->ovfl_val;
5205 has_smpl = CTX_HAS_SMPL(ctx);
5206
5207 DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5208 "used_pmds=0x%lx\n",
5209 pmc0,
5210 task ? task->pid: -1,
5211 (regs ? regs->cr_iip : 0),
5212 CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5213 ctx->ctx_used_pmds[0]));
5214
5215
5216 /*
5217 * first we update the virtual counters
5218 * assume there was a prior ia64_srlz_d() issued
5219 */
5220 for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5221
5222 /* skip pmd which did not overflow */
5223 if ((mask & 0x1) == 0) continue;
5224
5225 /*
5226 * Note that the pmd is not necessarily 0 at this point as qualified events
5227 * may have happened before the PMU was frozen. The residual count is not
5228 * taken into consideration here but will be with any read of the pmd via
5229 * pfm_read_pmds().
5230 */
5231 old_val = new_val = ctx->ctx_pmds[i].val;
5232 new_val += 1 + ovfl_val;
5233 ctx->ctx_pmds[i].val = new_val;
5234
5235 /*
5236 * check for overflow condition
5237 */
5238 if (likely(old_val > new_val)) {
5239 ovfl_pmds |= 1UL << i;
5240 if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5241 }
5242
5243 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5244 i,
5245 new_val,
5246 old_val,
5247 ia64_get_pmd(i) & ovfl_val,
5248 ovfl_pmds,
5249 ovfl_notify));
5250 }
5251
5252 /*
5253 * there was no 64-bit overflow, nothing else to do
5254 */
5255 if (ovfl_pmds == 0UL) return;
5256
5257 /*
5258 * reset all control bits
5259 */
5260 ovfl_ctrl.val = 0;
5261 reset_pmds = 0UL;
5262
5263 /*
5264 * if a sampling format module exists, then we "cache" the overflow by
5265 * calling the module's handler() routine.
5266 */
5267 if (has_smpl) {
5268 unsigned long start_cycles, end_cycles;
5269 unsigned long pmd_mask;
5270 int j, k, ret = 0;
5271 int this_cpu = smp_processor_id();
5272
5273 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5274 ovfl_arg = &ctx->ctx_ovfl_arg;
5275
5276 prefetch(ctx->ctx_smpl_hdr);
5277
5278 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5279
5280 mask = 1UL << i;
5281
5282 if ((pmd_mask & 0x1) == 0) continue;
5283
5284 ovfl_arg->ovfl_pmd = (unsigned char )i;
5285 ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0;
5286 ovfl_arg->active_set = 0;
5287 ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5288 ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5289
5290 ovfl_arg->pmd_value = ctx->ctx_pmds[i].val;
5291 ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5292 ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid;
5293
5294 /*
5295 * copy values of pmds of interest. Sampling format may copy them
5296 * into sampling buffer.
5297 */
5298 if (smpl_pmds) {
5299 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5300 if ((smpl_pmds & 0x1) == 0) continue;
5301 ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5302 DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5303 }
5304 }
5305
5306 pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5307
5308 start_cycles = ia64_get_itc();
5309
5310 /*
5311 * call custom buffer format record (handler) routine
5312 */
5313 ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5314
5315 end_cycles = ia64_get_itc();
5316
5317 /*
5318 * For those controls, we take the union because they have
5319 * an all or nothing behavior.
5320 */
5321 ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5322 ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task;
5323 ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5324 /*
5325 * build the bitmask of pmds to reset now
5326 */
5327 if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5328
5329 pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5330 }
5331 /*
5332 * when the module cannot handle the rest of the overflows, we abort right here
5333 */
5334 if (ret && pmd_mask) {
5335 DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5336 pmd_mask<<PMU_FIRST_COUNTER));
5337 }
5338 /*
5339 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5340 */
5341 ovfl_pmds &= ~reset_pmds;
5342 } else {
5343 /*
5344 * when no sampling module is used, then the default
5345 * is to notify on overflow if requested by user
5346 */
5347 ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0;
5348 ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0;
5349 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5350 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5351 /*
5352 * if needed, we reset all overflowed pmds
5353 */
5354 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5355 }
5356
5357 DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5358
5359 /*
5360 * reset the requested PMD registers using the short reset values
5361 */
5362 if (reset_pmds) {
5363 unsigned long bm = reset_pmds;
5364 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5365 }
5366
5367 if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5368 /*
5369 * keep track of what to reset when unblocking
5370 */
5371 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5372
5373 /*
5374 * check for blocking context
5375 */
5376 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5377
5378 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5379
5380 /*
5381 * set the perfmon specific checking pending work for the task
5382 */
5383 PFM_SET_WORK_PENDING(task, 1);
5384
5385 /*
5386 * when coming from ctxsw, current still points to the
5387 * previous task, therefore we must work with task and not current.
5388 */
5389 pfm_set_task_notify(task);
5390 }
5391 /*
5392 * defer until state is changed (shorten spin window). the context is locked
5393 * anyway, so the signal receiver would come spin for nothing.
5394 */
5395 must_notify = 1;
5396 }
5397
5398 DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5399 GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1,
5400 PFM_GET_WORK_PENDING(task),
5401 ctx->ctx_fl_trap_reason,
5402 ovfl_pmds,
5403 ovfl_notify,
5404 ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5405 /*
5406 * in case monitoring must be stopped, we toggle the psr bits
5407 */
5408 if (ovfl_ctrl.bits.mask_monitoring) {
5409 pfm_mask_monitoring(task);
5410 ctx->ctx_state = PFM_CTX_MASKED;
5411 ctx->ctx_fl_can_restart = 1;
5412 }
5413
5414 /*
5415 * send notification now
5416 */
5417 if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5418
5419 return;
5420
5421sanity_check:
5422 printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5423 smp_processor_id(),
5424 task ? task->pid : -1,
5425 pmc0);
5426 return;
5427
5428stop_monitoring:
5429 /*
5430 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5431 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5432 * come here as zombie only if the task is the current task. In which case, we
5433 * can access the PMU hardware directly.
5434 *
5435 * Note that zombies do have PM_VALID set. So here we do the minimal.
5436 *
5437 * In case the context was zombified it could not be reclaimed at the time
5438 * the monitoring program exited. At this point, the PMU reservation has been
5439 * returned, the sampiing buffer has been freed. We must convert this call
5440 * into a spurious interrupt. However, we must also avoid infinite overflows
5441 * by stopping monitoring for this task. We can only come here for a per-task
5442 * context. All we need to do is to stop monitoring using the psr bits which
5443 * are always task private. By re-enabling secure montioring, we ensure that
5444 * the monitored task will not be able to re-activate monitoring.
5445 * The task will eventually be context switched out, at which point the context
5446 * will be reclaimed (that includes releasing ownership of the PMU).
5447 *
5448 * So there might be a window of time where the number of per-task session is zero
5449 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5450 * context. This is safe because if a per-task session comes in, it will push this one
5451 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5452 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5453 * also push our zombie context out.
5454 *
5455 * Overall pretty hairy stuff....
5456 */
5457 DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1));
5458 pfm_clear_psr_up();
5459 ia64_psr(regs)->up = 0;
5460 ia64_psr(regs)->sp = 1;
5461 return;
5462}
5463
5464static int
5465pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5466{
5467 struct task_struct *task;
5468 pfm_context_t *ctx;
5469 unsigned long flags;
5470 u64 pmc0;
5471 int this_cpu = smp_processor_id();
5472 int retval = 0;
5473
5474 pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5475
5476 /*
5477 * srlz.d done before arriving here
5478 */
5479 pmc0 = ia64_get_pmc(0);
5480
5481 task = GET_PMU_OWNER();
5482 ctx = GET_PMU_CTX();
5483
5484 /*
5485 * if we have some pending bits set
5486 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5487 */
5488 if (PMC0_HAS_OVFL(pmc0) && task) {
5489 /*
5490 * we assume that pmc0.fr is always set here
5491 */
5492
5493 /* sanity check */
5494 if (!ctx) goto report_spurious1;
5495
5496 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0)
5497 goto report_spurious2;
5498
5499 PROTECT_CTX_NOPRINT(ctx, flags);
5500
5501 pfm_overflow_handler(task, ctx, pmc0, regs);
5502
5503 UNPROTECT_CTX_NOPRINT(ctx, flags);
5504
5505 } else {
5506 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5507 retval = -1;
5508 }
5509 /*
5510 * keep it unfrozen at all times
5511 */
5512 pfm_unfreeze_pmu();
5513
5514 return retval;
5515
5516report_spurious1:
5517 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5518 this_cpu, task->pid);
5519 pfm_unfreeze_pmu();
5520 return -1;
5521report_spurious2:
5522 printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n",
5523 this_cpu,
5524 task->pid);
5525 pfm_unfreeze_pmu();
5526 return -1;
5527}
5528
5529static irqreturn_t
5530pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
5531{
5532 unsigned long start_cycles, total_cycles;
5533 unsigned long min, max;
5534 int this_cpu;
5535 int ret;
5536
5537 this_cpu = get_cpu();
5538 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5539 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5540
5541 start_cycles = ia64_get_itc();
5542
5543 ret = pfm_do_interrupt_handler(irq, arg, regs);
5544
5545 total_cycles = ia64_get_itc();
5546
5547 /*
5548 * don't measure spurious interrupts
5549 */
5550 if (likely(ret == 0)) {
5551 total_cycles -= start_cycles;
5552
5553 if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5554 if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5555
5556 pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5557 }
5558 put_cpu_no_resched();
5559 return IRQ_HANDLED;
5560}
5561
5562/*
5563 * /proc/perfmon interface, for debug only
5564 */
5565
5566#define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1)
5567
5568static void *
5569pfm_proc_start(struct seq_file *m, loff_t *pos)
5570{
5571 if (*pos == 0) {
5572 return PFM_PROC_SHOW_HEADER;
5573 }
5574
5575 while (*pos <= NR_CPUS) {
5576 if (cpu_online(*pos - 1)) {
5577 return (void *)*pos;
5578 }
5579 ++*pos;
5580 }
5581 return NULL;
5582}
5583
5584static void *
5585pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5586{
5587 ++*pos;
5588 return pfm_proc_start(m, pos);
5589}
5590
5591static void
5592pfm_proc_stop(struct seq_file *m, void *v)
5593{
5594}
5595
5596static void
5597pfm_proc_show_header(struct seq_file *m)
5598{
5599 struct list_head * pos;
5600 pfm_buffer_fmt_t * entry;
5601 unsigned long flags;
5602
5603 seq_printf(m,
5604 "perfmon version : %u.%u\n"
5605 "model : %s\n"
5606 "fastctxsw : %s\n"
5607 "expert mode : %s\n"
5608 "ovfl_mask : 0x%lx\n"
5609 "PMU flags : 0x%x\n",
5610 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5611 pmu_conf->pmu_name,
5612 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5613 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5614 pmu_conf->ovfl_val,
5615 pmu_conf->flags);
5616
5617 LOCK_PFS(flags);
5618
5619 seq_printf(m,
5620 "proc_sessions : %u\n"
5621 "sys_sessions : %u\n"
5622 "sys_use_dbregs : %u\n"
5623 "ptrace_use_dbregs : %u\n",
5624 pfm_sessions.pfs_task_sessions,
5625 pfm_sessions.pfs_sys_sessions,
5626 pfm_sessions.pfs_sys_use_dbregs,
5627 pfm_sessions.pfs_ptrace_use_dbregs);
5628
5629 UNLOCK_PFS(flags);
5630
5631 spin_lock(&pfm_buffer_fmt_lock);
5632
5633 list_for_each(pos, &pfm_buffer_fmt_list) {
5634 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5635 seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5636 entry->fmt_uuid[0],
5637 entry->fmt_uuid[1],
5638 entry->fmt_uuid[2],
5639 entry->fmt_uuid[3],
5640 entry->fmt_uuid[4],
5641 entry->fmt_uuid[5],
5642 entry->fmt_uuid[6],
5643 entry->fmt_uuid[7],
5644 entry->fmt_uuid[8],
5645 entry->fmt_uuid[9],
5646 entry->fmt_uuid[10],
5647 entry->fmt_uuid[11],
5648 entry->fmt_uuid[12],
5649 entry->fmt_uuid[13],
5650 entry->fmt_uuid[14],
5651 entry->fmt_uuid[15],
5652 entry->fmt_name);
5653 }
5654 spin_unlock(&pfm_buffer_fmt_lock);
5655
5656}
5657
5658static int
5659pfm_proc_show(struct seq_file *m, void *v)
5660{
5661 unsigned long psr;
5662 unsigned int i;
5663 int cpu;
5664
5665 if (v == PFM_PROC_SHOW_HEADER) {
5666 pfm_proc_show_header(m);
5667 return 0;
5668 }
5669
5670 /* show info for CPU (v - 1) */
5671
5672 cpu = (long)v - 1;
5673 seq_printf(m,
5674 "CPU%-2d overflow intrs : %lu\n"
5675 "CPU%-2d overflow cycles : %lu\n"
5676 "CPU%-2d overflow min : %lu\n"
5677 "CPU%-2d overflow max : %lu\n"
5678 "CPU%-2d smpl handler calls : %lu\n"
5679 "CPU%-2d smpl handler cycles : %lu\n"
5680 "CPU%-2d spurious intrs : %lu\n"
5681 "CPU%-2d replay intrs : %lu\n"
5682 "CPU%-2d syst_wide : %d\n"
5683 "CPU%-2d dcr_pp : %d\n"
5684 "CPU%-2d exclude idle : %d\n"
5685 "CPU%-2d owner : %d\n"
5686 "CPU%-2d context : %p\n"
5687 "CPU%-2d activations : %lu\n",
5688 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5689 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5690 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5691 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5692 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5693 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5694 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5695 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5696 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5697 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5698 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5699 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5700 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5701 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5702
5703 if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5704
5705 psr = pfm_get_psr();
5706
5707 ia64_srlz_d();
5708
5709 seq_printf(m,
5710 "CPU%-2d psr : 0x%lx\n"
5711 "CPU%-2d pmc0 : 0x%lx\n",
5712 cpu, psr,
5713 cpu, ia64_get_pmc(0));
5714
5715 for (i=0; PMC_IS_LAST(i) == 0; i++) {
5716 if (PMC_IS_COUNTING(i) == 0) continue;
5717 seq_printf(m,
5718 "CPU%-2d pmc%u : 0x%lx\n"
5719 "CPU%-2d pmd%u : 0x%lx\n",
5720 cpu, i, ia64_get_pmc(i),
5721 cpu, i, ia64_get_pmd(i));
5722 }
5723 }
5724 return 0;
5725}
5726
5727struct seq_operations pfm_seq_ops = {
5728 .start = pfm_proc_start,
5729 .next = pfm_proc_next,
5730 .stop = pfm_proc_stop,
5731 .show = pfm_proc_show
5732};
5733
5734static int
5735pfm_proc_open(struct inode *inode, struct file *file)
5736{
5737 return seq_open(file, &pfm_seq_ops);
5738}
5739
5740
5741/*
5742 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5743 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5744 * is active or inactive based on mode. We must rely on the value in
5745 * local_cpu_data->pfm_syst_info
5746 */
5747void
5748pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5749{
5750 struct pt_regs *regs;
5751 unsigned long dcr;
5752 unsigned long dcr_pp;
5753
5754 dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5755
5756 /*
5757 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5758 * on every CPU, so we can rely on the pid to identify the idle task.
5759 */
5760 if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5761 regs = ia64_task_regs(task);
5762 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5763 return;
5764 }
5765 /*
5766 * if monitoring has started
5767 */
5768 if (dcr_pp) {
5769 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5770 /*
5771 * context switching in?
5772 */
5773 if (is_ctxswin) {
5774 /* mask monitoring for the idle task */
5775 ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5776 pfm_clear_psr_pp();
5777 ia64_srlz_i();
5778 return;
5779 }
5780 /*
5781 * context switching out
5782 * restore monitoring for next task
5783 *
5784 * Due to inlining this odd if-then-else construction generates
5785 * better code.
5786 */
5787 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5788 pfm_set_psr_pp();
5789 ia64_srlz_i();
5790 }
5791}
5792
5793#ifdef CONFIG_SMP
5794
5795static void
5796pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5797{
5798 struct task_struct *task = ctx->ctx_task;
5799
5800 ia64_psr(regs)->up = 0;
5801 ia64_psr(regs)->sp = 1;
5802
5803 if (GET_PMU_OWNER() == task) {
5804 DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid));
5805 SET_PMU_OWNER(NULL, NULL);
5806 }
5807
5808 /*
5809 * disconnect the task from the context and vice-versa
5810 */
5811 PFM_SET_WORK_PENDING(task, 0);
5812
5813 task->thread.pfm_context = NULL;
5814 task->thread.flags &= ~IA64_THREAD_PM_VALID;
5815
5816 DPRINT(("force cleanup for [%d]\n", task->pid));
5817}
5818
5819
5820/*
5821 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5822 */
5823void
5824pfm_save_regs(struct task_struct *task)
5825{
5826 pfm_context_t *ctx;
5827 struct thread_struct *t;
5828 unsigned long flags;
5829 u64 psr;
5830
5831
5832 ctx = PFM_GET_CTX(task);
5833 if (ctx == NULL) return;
5834 t = &task->thread;
5835
5836 /*
5837 * we always come here with interrupts ALREADY disabled by
5838 * the scheduler. So we simply need to protect against concurrent
5839 * access, not CPU concurrency.
5840 */
5841 flags = pfm_protect_ctx_ctxsw(ctx);
5842
5843 if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5844 struct pt_regs *regs = ia64_task_regs(task);
5845
5846 pfm_clear_psr_up();
5847
5848 pfm_force_cleanup(ctx, regs);
5849
5850 BUG_ON(ctx->ctx_smpl_hdr);
5851
5852 pfm_unprotect_ctx_ctxsw(ctx, flags);
5853
5854 pfm_context_free(ctx);
5855 return;
5856 }
5857
5858 /*
5859 * save current PSR: needed because we modify it
5860 */
5861 ia64_srlz_d();
5862 psr = pfm_get_psr();
5863
5864 BUG_ON(psr & (IA64_PSR_I));
5865
5866 /*
5867 * stop monitoring:
5868 * This is the last instruction which may generate an overflow
5869 *
5870 * We do not need to set psr.sp because, it is irrelevant in kernel.
5871 * It will be restored from ipsr when going back to user level
5872 */
5873 pfm_clear_psr_up();
5874
5875 /*
5876 * keep a copy of psr.up (for reload)
5877 */
5878 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5879
5880 /*
5881 * release ownership of this PMU.
5882 * PM interrupts are masked, so nothing
5883 * can happen.
5884 */
5885 SET_PMU_OWNER(NULL, NULL);
5886
5887 /*
5888 * we systematically save the PMD as we have no
5889 * guarantee we will be schedule at that same
5890 * CPU again.
5891 */
5892 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5893
5894 /*
5895 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5896 * we will need it on the restore path to check
5897 * for pending overflow.
5898 */
5899 t->pmcs[0] = ia64_get_pmc(0);
5900
5901 /*
5902 * unfreeze PMU if had pending overflows
5903 */
5904 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5905
5906 /*
5907 * finally, allow context access.
5908 * interrupts will still be masked after this call.
5909 */
5910 pfm_unprotect_ctx_ctxsw(ctx, flags);
5911}
5912
5913#else /* !CONFIG_SMP */
5914void
5915pfm_save_regs(struct task_struct *task)
5916{
5917 pfm_context_t *ctx;
5918 u64 psr;
5919
5920 ctx = PFM_GET_CTX(task);
5921 if (ctx == NULL) return;
5922
5923 /*
5924 * save current PSR: needed because we modify it
5925 */
5926 psr = pfm_get_psr();
5927
5928 BUG_ON(psr & (IA64_PSR_I));
5929
5930 /*
5931 * stop monitoring:
5932 * This is the last instruction which may generate an overflow
5933 *
5934 * We do not need to set psr.sp because, it is irrelevant in kernel.
5935 * It will be restored from ipsr when going back to user level
5936 */
5937 pfm_clear_psr_up();
5938
5939 /*
5940 * keep a copy of psr.up (for reload)
5941 */
5942 ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5943}
5944
5945static void
5946pfm_lazy_save_regs (struct task_struct *task)
5947{
5948 pfm_context_t *ctx;
5949 struct thread_struct *t;
5950 unsigned long flags;
5951
5952 { u64 psr = pfm_get_psr();
5953 BUG_ON(psr & IA64_PSR_UP);
5954 }
5955
5956 ctx = PFM_GET_CTX(task);
5957 t = &task->thread;
5958
5959 /*
5960 * we need to mask PMU overflow here to
5961 * make sure that we maintain pmc0 until
5962 * we save it. overflow interrupts are
5963 * treated as spurious if there is no
5964 * owner.
5965 *
5966 * XXX: I don't think this is necessary
5967 */
5968 PROTECT_CTX(ctx,flags);
5969
5970 /*
5971 * release ownership of this PMU.
5972 * must be done before we save the registers.
5973 *
5974 * after this call any PMU interrupt is treated
5975 * as spurious.
5976 */
5977 SET_PMU_OWNER(NULL, NULL);
5978
5979 /*
5980 * save all the pmds we use
5981 */
5982 pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
5983
5984 /*
5985 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5986 * it is needed to check for pended overflow
5987 * on the restore path
5988 */
5989 t->pmcs[0] = ia64_get_pmc(0);
5990
5991 /*
5992 * unfreeze PMU if had pending overflows
5993 */
5994 if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5995
5996 /*
5997 * now get can unmask PMU interrupts, they will
5998 * be treated as purely spurious and we will not
5999 * lose any information
6000 */
6001 UNPROTECT_CTX(ctx,flags);
6002}
6003#endif /* CONFIG_SMP */
6004
6005#ifdef CONFIG_SMP
6006/*
6007 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6008 */
6009void
6010pfm_load_regs (struct task_struct *task)
6011{
6012 pfm_context_t *ctx;
6013 struct thread_struct *t;
6014 unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6015 unsigned long flags;
6016 u64 psr, psr_up;
6017 int need_irq_resend;
6018
6019 ctx = PFM_GET_CTX(task);
6020 if (unlikely(ctx == NULL)) return;
6021
6022 BUG_ON(GET_PMU_OWNER());
6023
6024 t = &task->thread;
6025 /*
6026 * possible on unload
6027 */
6028 if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
6029
6030 /*
6031 * we always come here with interrupts ALREADY disabled by
6032 * the scheduler. So we simply need to protect against concurrent
6033 * access, not CPU concurrency.
6034 */
6035 flags = pfm_protect_ctx_ctxsw(ctx);
6036 psr = pfm_get_psr();
6037
6038 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6039
6040 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6041 BUG_ON(psr & IA64_PSR_I);
6042
6043 if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6044 struct pt_regs *regs = ia64_task_regs(task);
6045
6046 BUG_ON(ctx->ctx_smpl_hdr);
6047
6048 pfm_force_cleanup(ctx, regs);
6049
6050 pfm_unprotect_ctx_ctxsw(ctx, flags);
6051
6052 /*
6053 * this one (kmalloc'ed) is fine with interrupts disabled
6054 */
6055 pfm_context_free(ctx);
6056
6057 return;
6058 }
6059
6060 /*
6061 * we restore ALL the debug registers to avoid picking up
6062 * stale state.
6063 */
6064 if (ctx->ctx_fl_using_dbreg) {
6065 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6066 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6067 }
6068 /*
6069 * retrieve saved psr.up
6070 */
6071 psr_up = ctx->ctx_saved_psr_up;
6072
6073 /*
6074 * if we were the last user of the PMU on that CPU,
6075 * then nothing to do except restore psr
6076 */
6077 if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6078
6079 /*
6080 * retrieve partial reload masks (due to user modifications)
6081 */
6082 pmc_mask = ctx->ctx_reload_pmcs[0];
6083 pmd_mask = ctx->ctx_reload_pmds[0];
6084
6085 } else {
6086 /*
6087 * To avoid leaking information to the user level when psr.sp=0,
6088 * we must reload ALL implemented pmds (even the ones we don't use).
6089 * In the kernel we only allow PFM_READ_PMDS on registers which
6090 * we initialized or requested (sampling) so there is no risk there.
6091 */
6092 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6093
6094 /*
6095 * ALL accessible PMCs are systematically reloaded, unused registers
6096 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6097 * up stale configuration.
6098 *
6099 * PMC0 is never in the mask. It is always restored separately.
6100 */
6101 pmc_mask = ctx->ctx_all_pmcs[0];
6102 }
6103 /*
6104 * when context is MASKED, we will restore PMC with plm=0
6105 * and PMD with stale information, but that's ok, nothing
6106 * will be captured.
6107 *
6108 * XXX: optimize here
6109 */
6110 if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
6111 if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
6112
6113 /*
6114 * check for pending overflow at the time the state
6115 * was saved.
6116 */
6117 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6118 /*
6119 * reload pmc0 with the overflow information
6120 * On McKinley PMU, this will trigger a PMU interrupt
6121 */
6122 ia64_set_pmc(0, t->pmcs[0]);
6123 ia64_srlz_d();
6124 t->pmcs[0] = 0UL;
6125
6126 /*
6127 * will replay the PMU interrupt
6128 */
6129 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6130
6131 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6132 }
6133
6134 /*
6135 * we just did a reload, so we reset the partial reload fields
6136 */
6137 ctx->ctx_reload_pmcs[0] = 0UL;
6138 ctx->ctx_reload_pmds[0] = 0UL;
6139
6140 SET_LAST_CPU(ctx, smp_processor_id());
6141
6142 /*
6143 * dump activation value for this PMU
6144 */
6145 INC_ACTIVATION();
6146 /*
6147 * record current activation for this context
6148 */
6149 SET_ACTIVATION(ctx);
6150
6151 /*
6152 * establish new ownership.
6153 */
6154 SET_PMU_OWNER(task, ctx);
6155
6156 /*
6157 * restore the psr.up bit. measurement
6158 * is active again.
6159 * no PMU interrupt can happen at this point
6160 * because we still have interrupts disabled.
6161 */
6162 if (likely(psr_up)) pfm_set_psr_up();
6163
6164 /*
6165 * allow concurrent access to context
6166 */
6167 pfm_unprotect_ctx_ctxsw(ctx, flags);
6168}
6169#else /* !CONFIG_SMP */
6170/*
6171 * reload PMU state for UP kernels
6172 * in 2.5 we come here with interrupts disabled
6173 */
6174void
6175pfm_load_regs (struct task_struct *task)
6176{
6177 struct thread_struct *t;
6178 pfm_context_t *ctx;
6179 struct task_struct *owner;
6180 unsigned long pmd_mask, pmc_mask;
6181 u64 psr, psr_up;
6182 int need_irq_resend;
6183
6184 owner = GET_PMU_OWNER();
6185 ctx = PFM_GET_CTX(task);
6186 t = &task->thread;
6187 psr = pfm_get_psr();
6188
6189 BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6190 BUG_ON(psr & IA64_PSR_I);
6191
6192 /*
6193 * we restore ALL the debug registers to avoid picking up
6194 * stale state.
6195 *
6196 * This must be done even when the task is still the owner
6197 * as the registers may have been modified via ptrace()
6198 * (not perfmon) by the previous task.
6199 */
6200 if (ctx->ctx_fl_using_dbreg) {
6201 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6202 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6203 }
6204
6205 /*
6206 * retrieved saved psr.up
6207 */
6208 psr_up = ctx->ctx_saved_psr_up;
6209 need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6210
6211 /*
6212 * short path, our state is still there, just
6213 * need to restore psr and we go
6214 *
6215 * we do not touch either PMC nor PMD. the psr is not touched
6216 * by the overflow_handler. So we are safe w.r.t. to interrupt
6217 * concurrency even without interrupt masking.
6218 */
6219 if (likely(owner == task)) {
6220 if (likely(psr_up)) pfm_set_psr_up();
6221 return;
6222 }
6223
6224 /*
6225 * someone else is still using the PMU, first push it out and
6226 * then we'll be able to install our stuff !
6227 *
6228 * Upon return, there will be no owner for the current PMU
6229 */
6230 if (owner) pfm_lazy_save_regs(owner);
6231
6232 /*
6233 * To avoid leaking information to the user level when psr.sp=0,
6234 * we must reload ALL implemented pmds (even the ones we don't use).
6235 * In the kernel we only allow PFM_READ_PMDS on registers which
6236 * we initialized or requested (sampling) so there is no risk there.
6237 */
6238 pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6239
6240 /*
6241 * ALL accessible PMCs are systematically reloaded, unused registers
6242 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6243 * up stale configuration.
6244 *
6245 * PMC0 is never in the mask. It is always restored separately
6246 */
6247 pmc_mask = ctx->ctx_all_pmcs[0];
6248
6249 pfm_restore_pmds(t->pmds, pmd_mask);
6250 pfm_restore_pmcs(t->pmcs, pmc_mask);
6251
6252 /*
6253 * check for pending overflow at the time the state
6254 * was saved.
6255 */
6256 if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
6257 /*
6258 * reload pmc0 with the overflow information
6259 * On McKinley PMU, this will trigger a PMU interrupt
6260 */
6261 ia64_set_pmc(0, t->pmcs[0]);
6262 ia64_srlz_d();
6263
6264 t->pmcs[0] = 0UL;
6265
6266 /*
6267 * will replay the PMU interrupt
6268 */
6269 if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
6270
6271 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6272 }
6273
6274 /*
6275 * establish new ownership.
6276 */
6277 SET_PMU_OWNER(task, ctx);
6278
6279 /*
6280 * restore the psr.up bit. measurement
6281 * is active again.
6282 * no PMU interrupt can happen at this point
6283 * because we still have interrupts disabled.
6284 */
6285 if (likely(psr_up)) pfm_set_psr_up();
6286}
6287#endif /* CONFIG_SMP */
6288
6289/*
6290 * this function assumes monitoring is stopped
6291 */
6292static void
6293pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6294{
6295 u64 pmc0;
6296 unsigned long mask2, val, pmd_val, ovfl_val;
6297 int i, can_access_pmu = 0;
6298 int is_self;
6299
6300 /*
6301 * is the caller the task being monitored (or which initiated the
6302 * session for system wide measurements)
6303 */
6304 is_self = ctx->ctx_task == task ? 1 : 0;
6305
6306 /*
6307 * can access PMU is task is the owner of the PMU state on the current CPU
6308 * or if we are running on the CPU bound to the context in system-wide mode
6309 * (that is not necessarily the task the context is attached to in this mode).
6310 * In system-wide we always have can_access_pmu true because a task running on an
6311 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6312 */
6313 can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6314 if (can_access_pmu) {
6315 /*
6316 * Mark the PMU as not owned
6317 * This will cause the interrupt handler to do nothing in case an overflow
6318 * interrupt was in-flight
6319 * This also guarantees that pmc0 will contain the final state
6320 * It virtually gives us full control on overflow processing from that point
6321 * on.
6322 */
6323 SET_PMU_OWNER(NULL, NULL);
6324 DPRINT(("releasing ownership\n"));
6325
6326 /*
6327 * read current overflow status:
6328 *
6329 * we are guaranteed to read the final stable state
6330 */
6331 ia64_srlz_d();
6332 pmc0 = ia64_get_pmc(0); /* slow */
6333
6334 /*
6335 * reset freeze bit, overflow status information destroyed
6336 */
6337 pfm_unfreeze_pmu();
6338 } else {
6339 pmc0 = task->thread.pmcs[0];
6340 /*
6341 * clear whatever overflow status bits there were
6342 */
6343 task->thread.pmcs[0] = 0;
6344 }
6345 ovfl_val = pmu_conf->ovfl_val;
6346 /*
6347 * we save all the used pmds
6348 * we take care of overflows for counting PMDs
6349 *
6350 * XXX: sampling situation is not taken into account here
6351 */
6352 mask2 = ctx->ctx_used_pmds[0];
6353
6354 DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6355
6356 for (i = 0; mask2; i++, mask2>>=1) {
6357
6358 /* skip non used pmds */
6359 if ((mask2 & 0x1) == 0) continue;
6360
6361 /*
6362 * can access PMU always true in system wide mode
6363 */
6364 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
6365
6366 if (PMD_IS_COUNTING(i)) {
6367 DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6368 task->pid,
6369 i,
6370 ctx->ctx_pmds[i].val,
6371 val & ovfl_val));
6372
6373 /*
6374 * we rebuild the full 64 bit value of the counter
6375 */
6376 val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6377
6378 /*
6379 * now everything is in ctx_pmds[] and we need
6380 * to clear the saved context from save_regs() such that
6381 * pfm_read_pmds() gets the correct value
6382 */
6383 pmd_val = 0UL;
6384
6385 /*
6386 * take care of overflow inline
6387 */
6388 if (pmc0 & (1UL << i)) {
6389 val += 1 + ovfl_val;
6390 DPRINT(("[%d] pmd[%d] overflowed\n", task->pid, i));
6391 }
6392 }
6393
6394 DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, i, val, pmd_val));
6395
6396 if (is_self) task->thread.pmds[i] = pmd_val;
6397
6398 ctx->ctx_pmds[i].val = val;
6399 }
6400}
6401
6402static struct irqaction perfmon_irqaction = {
6403 .handler = pfm_interrupt_handler,
6404 .flags = SA_INTERRUPT,
6405 .name = "perfmon"
6406};
6407
6408/*
6409 * perfmon initialization routine, called from the initcall() table
6410 */
6411static int init_pfm_fs(void);
6412
6413static int __init
6414pfm_probe_pmu(void)
6415{
6416 pmu_config_t **p;
6417 int family;
6418
6419 family = local_cpu_data->family;
6420 p = pmu_confs;
6421
6422 while(*p) {
6423 if ((*p)->probe) {
6424 if ((*p)->probe() == 0) goto found;
6425 } else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6426 goto found;
6427 }
6428 p++;
6429 }
6430 return -1;
6431found:
6432 pmu_conf = *p;
6433 return 0;
6434}
6435
6436static struct file_operations pfm_proc_fops = {
6437 .open = pfm_proc_open,
6438 .read = seq_read,
6439 .llseek = seq_lseek,
6440 .release = seq_release,
6441};
6442
6443int __init
6444pfm_init(void)
6445{
6446 unsigned int n, n_counters, i;
6447
6448 printk("perfmon: version %u.%u IRQ %u\n",
6449 PFM_VERSION_MAJ,
6450 PFM_VERSION_MIN,
6451 IA64_PERFMON_VECTOR);
6452
6453 if (pfm_probe_pmu()) {
6454 printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n",
6455 local_cpu_data->family);
6456 return -ENODEV;
6457 }
6458
6459 /*
6460 * compute the number of implemented PMD/PMC from the
6461 * description tables
6462 */
6463 n = 0;
6464 for (i=0; PMC_IS_LAST(i) == 0; i++) {
6465 if (PMC_IS_IMPL(i) == 0) continue;
6466 pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6467 n++;
6468 }
6469 pmu_conf->num_pmcs = n;
6470
6471 n = 0; n_counters = 0;
6472 for (i=0; PMD_IS_LAST(i) == 0; i++) {
6473 if (PMD_IS_IMPL(i) == 0) continue;
6474 pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6475 n++;
6476 if (PMD_IS_COUNTING(i)) n_counters++;
6477 }
6478 pmu_conf->num_pmds = n;
6479 pmu_conf->num_counters = n_counters;
6480
6481 /*
6482 * sanity checks on the number of debug registers
6483 */
6484 if (pmu_conf->use_rr_dbregs) {
6485 if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6486 printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6487 pmu_conf = NULL;
6488 return -1;
6489 }
6490 if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6491 printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6492 pmu_conf = NULL;
6493 return -1;
6494 }
6495 }
6496
6497 printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6498 pmu_conf->pmu_name,
6499 pmu_conf->num_pmcs,
6500 pmu_conf->num_pmds,
6501 pmu_conf->num_counters,
6502 ffz(pmu_conf->ovfl_val));
6503
6504 /* sanity check */
6505 if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
6506 printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6507 pmu_conf = NULL;
6508 return -1;
6509 }
6510
6511 /*
6512 * create /proc/perfmon (mostly for debugging purposes)
6513 */
6514 perfmon_dir = create_proc_entry("perfmon", S_IRUGO, NULL);
6515 if (perfmon_dir == NULL) {
6516 printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6517 pmu_conf = NULL;
6518 return -1;
6519 }
6520 /*
6521 * install customized file operations for /proc/perfmon entry
6522 */
6523 perfmon_dir->proc_fops = &pfm_proc_fops;
6524
6525 /*
6526 * create /proc/sys/kernel/perfmon (for debugging purposes)
6527 */
6528 pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root, 0);
6529
6530 /*
6531 * initialize all our spinlocks
6532 */
6533 spin_lock_init(&pfm_sessions.pfs_lock);
6534 spin_lock_init(&pfm_buffer_fmt_lock);
6535
6536 init_pfm_fs();
6537
6538 for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6539
6540 return 0;
6541}
6542
6543__initcall(pfm_init);
6544
6545/*
6546 * this function is called before pfm_init()
6547 */
6548void
6549pfm_init_percpu (void)
6550{
6551 /*
6552 * make sure no measurement is active
6553 * (may inherit programmed PMCs from EFI).
6554 */
6555 pfm_clear_psr_pp();
6556 pfm_clear_psr_up();
6557
6558 /*
6559 * we run with the PMU not frozen at all times
6560 */
6561 pfm_unfreeze_pmu();
6562
6563 if (smp_processor_id() == 0)
6564 register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6565
6566 ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6567 ia64_srlz_d();
6568}
6569
6570/*
6571 * used for debug purposes only
6572 */
6573void
6574dump_pmu_state(const char *from)
6575{
6576 struct task_struct *task;
6577 struct thread_struct *t;
6578 struct pt_regs *regs;
6579 pfm_context_t *ctx;
6580 unsigned long psr, dcr, info, flags;
6581 int i, this_cpu;
6582
6583 local_irq_save(flags);
6584
6585 this_cpu = smp_processor_id();
6586 regs = ia64_task_regs(current);
6587 info = PFM_CPUINFO_GET();
6588 dcr = ia64_getreg(_IA64_REG_CR_DCR);
6589
6590 if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6591 local_irq_restore(flags);
6592 return;
6593 }
6594
6595 printk("CPU%d from %s() current [%d] iip=0x%lx %s\n",
6596 this_cpu,
6597 from,
6598 current->pid,
6599 regs->cr_iip,
6600 current->comm);
6601
6602 task = GET_PMU_OWNER();
6603 ctx = GET_PMU_CTX();
6604
6605 printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task->pid : -1, ctx);
6606
6607 psr = pfm_get_psr();
6608
6609 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",
6610 this_cpu,
6611 ia64_get_pmc(0),
6612 psr & IA64_PSR_PP ? 1 : 0,
6613 psr & IA64_PSR_UP ? 1 : 0,
6614 dcr & IA64_DCR_PP ? 1 : 0,
6615 info,
6616 ia64_psr(regs)->up,
6617 ia64_psr(regs)->pp);
6618
6619 ia64_psr(regs)->up = 0;
6620 ia64_psr(regs)->pp = 0;
6621
6622 t = &current->thread;
6623
6624 for (i=1; PMC_IS_LAST(i) == 0; i++) {
6625 if (PMC_IS_IMPL(i) == 0) continue;
6626 printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
6627 }
6628
6629 for (i=1; PMD_IS_LAST(i) == 0; i++) {
6630 if (PMD_IS_IMPL(i) == 0) continue;
6631 printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
6632 }
6633
6634 if (ctx) {
6635 printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6636 this_cpu,
6637 ctx->ctx_state,
6638 ctx->ctx_smpl_vaddr,
6639 ctx->ctx_smpl_hdr,
6640 ctx->ctx_msgq_head,
6641 ctx->ctx_msgq_tail,
6642 ctx->ctx_saved_psr_up);
6643 }
6644 local_irq_restore(flags);
6645}
6646
6647/*
6648 * called from process.c:copy_thread(). task is new child.
6649 */
6650void
6651pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6652{
6653 struct thread_struct *thread;
6654
6655 DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task->pid));
6656
6657 thread = &task->thread;
6658
6659 /*
6660 * cut links inherited from parent (current)
6661 */
6662 thread->pfm_context = NULL;
6663
6664 PFM_SET_WORK_PENDING(task, 0);
6665
6666 /*
6667 * the psr bits are already set properly in copy_threads()
6668 */
6669}
6670#else /* !CONFIG_PERFMON */
6671asmlinkage long
6672sys_perfmonctl (int fd, int cmd, void *arg, int count)
6673{
6674 return -ENOSYS;
6675}
6676#endif /* CONFIG_PERFMON */