4 * This provides a low-level interface to the hardware's Debug Store
5 * feature that is used for branch trace store (BTS) and
6 * precise-event based sampling (PEBS).
9 * - per-thread and per-cpu allocation of BTS and PEBS
10 * - buffer memory allocation (optional)
11 * - buffer overflow handling
15 * - get_task_struct on all parameter tasks
16 * - current is allowed to trace parameter tasks
19 * Copyright (C) 2007-2008 Intel Corporation.
20 * Markus Metzger <markus.t.metzger@intel.com>, 2007-2008
28 #include <linux/errno.h>
29 #include <linux/string.h>
30 #include <linux/slab.h>
31 #include <linux/sched.h>
36 * The configuration for a particular DS hardware implementation.
38 struct ds_configuration {
39 /* the size of the DS structure in bytes */
40 unsigned char sizeof_ds;
41 /* the size of one pointer-typed field in the DS structure in bytes;
42 this covers the first 8 fields related to buffer management. */
43 unsigned char sizeof_field;
44 /* the size of a BTS/PEBS record in bytes */
45 unsigned char sizeof_rec[2];
47 static struct ds_configuration ds_cfg;
51 * Debug Store (DS) save area configuration (see Intel64 and IA32
52 * Architectures Software Developer's Manual, section 18.5)
54 * The DS configuration consists of the following fields; different
55 * architetures vary in the size of those fields.
56 * - double-word aligned base linear address of the BTS buffer
57 * - write pointer into the BTS buffer
58 * - end linear address of the BTS buffer (one byte beyond the end of
60 * - interrupt pointer into BTS buffer
61 * (interrupt occurs when write pointer passes interrupt pointer)
62 * - double-word aligned base linear address of the PEBS buffer
63 * - write pointer into the PEBS buffer
64 * - end linear address of the PEBS buffer (one byte beyond the end of
66 * - interrupt pointer into PEBS buffer
67 * (interrupt occurs when write pointer passes interrupt pointer)
68 * - value to which counter is reset following counter overflow
70 * Later architectures use 64bit pointers throughout, whereas earlier
71 * architectures use 32bit pointers in 32bit mode.
74 * We compute the base address for the first 8 fields based on:
75 * - the field size stored in the DS configuration
76 * - the relative field position
77 * - an offset giving the start of the respective region
79 * This offset is further used to index various arrays holding
80 * information for BTS and PEBS at the respective index.
82 * On later 32bit processors, we only access the lower 32bit of the
83 * 64bit pointer fields. The upper halves will be zeroed out.
90 ds_interrupt_threshold,
98 static inline unsigned long ds_get(const unsigned char *base,
99 enum ds_qualifier qual, enum ds_field field)
101 base += (ds_cfg.sizeof_field * (field + (4 * qual)));
102 return *(unsigned long *)base;
105 static inline void ds_set(unsigned char *base, enum ds_qualifier qual,
106 enum ds_field field, unsigned long value)
108 base += (ds_cfg.sizeof_field * (field + (4 * qual)));
109 (*(unsigned long *)base) = value;
114 * Locking is done only for allocating BTS or PEBS resources and for
115 * guarding context and buffer memory allocation.
117 * Most functions require the current task to own the ds context part
118 * they are going to access. All the locking is done when validating
119 * access to the context.
121 static spinlock_t ds_lock = __SPIN_LOCK_UNLOCKED(ds_lock);
124 * Validate that the current task is allowed to access the BTS/PEBS
125 * buffer of the parameter task.
127 * Returns 0, if access is granted; -Eerrno, otherwise.
129 static inline int ds_validate_access(struct ds_context *context,
130 enum ds_qualifier qual)
135 if (context->owner[qual] == current)
143 * We either support (system-wide) per-cpu or per-thread allocation.
144 * We distinguish the two based on the task_struct pointer, where a
145 * NULL pointer indicates per-cpu allocation for the current cpu.
147 * Allocations are use-counted. As soon as resources are allocated,
148 * further allocations must be of the same type (per-cpu or
149 * per-thread). We model this by counting allocations (i.e. the number
150 * of tracers of a certain type) for one type negatively:
152 * >0 number of per-thread tracers
153 * <0 number of per-cpu tracers
155 * The below functions to get and put tracers and to check the
156 * allocation type require the ds_lock to be held by the caller.
158 * Tracers essentially gives the number of ds contexts for a certain
159 * type of allocation.
163 static inline void get_tracer(struct task_struct *task)
165 tracers += (task ? 1 : -1);
168 static inline void put_tracer(struct task_struct *task)
170 tracers -= (task ? 1 : -1);
173 static inline int check_tracer(struct task_struct *task)
175 return (task ? (tracers >= 0) : (tracers <= 0));
180 * The DS context is either attached to a thread or to a cpu:
181 * - in the former case, the thread_struct contains a pointer to the
183 * - in the latter case, we use a static array of per-cpu context
186 * Contexts are use-counted. They are allocated on first access and
187 * deallocated when the last user puts the context.
189 * We distinguish between an allocating and a non-allocating get of a
191 * - the allocating get is used for requesting BTS/PEBS resources. It
192 * requires the caller to hold the global ds_lock.
193 * - the non-allocating get is used for all other cases. A
194 * non-existing context indicates an error. It acquires and releases
195 * the ds_lock itself for obtaining the context.
197 * A context and its DS configuration are allocated and deallocated
198 * together. A context always has a DS configuration of the
201 static DEFINE_PER_CPU(struct ds_context *, system_context);
203 #define this_system_context per_cpu(system_context, smp_processor_id())
206 * Returns the pointer to the parameter task's context or to the
207 * system-wide context, if task is NULL.
209 * Increases the use count of the returned context, if not NULL.
211 static inline struct ds_context *ds_get_context(struct task_struct *task)
213 struct ds_context *context;
217 context = (task ? task->thread.ds_ctx : this_system_context);
221 spin_unlock(&ds_lock);
227 * Same as ds_get_context, but allocates the context and it's DS
228 * structure, if necessary; returns NULL; if out of memory.
230 * pre: requires ds_lock to be held
232 static inline struct ds_context *ds_alloc_context(struct task_struct *task)
234 struct ds_context **p_context =
235 (task ? &task->thread.ds_ctx : &this_system_context);
236 struct ds_context *context = *p_context;
239 context = kzalloc(sizeof(*context), GFP_KERNEL);
244 context->ds = kzalloc(ds_cfg.sizeof_ds, GFP_KERNEL);
250 *p_context = context;
252 context->this = p_context;
253 context->task = task;
256 set_tsk_thread_flag(task, TIF_DS_AREA_MSR);
258 if (!task || (task == current))
259 wrmsr(MSR_IA32_DS_AREA, (unsigned long)context->ds, 0);
270 * Decreases the use count of the parameter context, if not NULL.
271 * Deallocates the context, if the use count reaches zero.
273 static inline void ds_put_context(struct ds_context *context)
280 if (--context->count)
283 *(context->this) = NULL;
286 clear_tsk_thread_flag(context->task, TIF_DS_AREA_MSR);
288 if (!context->task || (context->task == current))
289 wrmsrl(MSR_IA32_DS_AREA, 0);
291 put_tracer(context->task);
293 /* free any leftover buffers from tracers that did not
294 * deallocate them properly. */
295 kfree(context->buffer[ds_bts]);
296 kfree(context->buffer[ds_pebs]);
300 spin_unlock(&ds_lock);
305 * Handle a buffer overflow
307 * task: the task whose buffers are overflowing;
308 * NULL for a buffer overflow on the current cpu
309 * context: the ds context
310 * qual: the buffer type
312 static void ds_overflow(struct task_struct *task, struct ds_context *context,
313 enum ds_qualifier qual)
318 if (context->callback[qual])
319 (*context->callback[qual])(task);
321 /* todo: do some more overflow handling */
326 * Allocate a non-pageable buffer of the parameter size.
327 * Checks the memory and the locked memory rlimit.
329 * Returns the buffer, if successful;
330 * NULL, if out of memory or rlimit exceeded.
332 * size: the requested buffer size in bytes
333 * pages (out): if not NULL, contains the number of pages reserved
335 static inline void *ds_allocate_buffer(size_t size, unsigned int *pages)
337 unsigned long rlim, vm, pgsz;
340 pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
342 rlim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
343 vm = current->mm->total_vm + pgsz;
347 rlim = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
348 vm = current->mm->locked_vm + pgsz;
352 buffer = kzalloc(size, GFP_KERNEL);
356 current->mm->total_vm += pgsz;
357 current->mm->locked_vm += pgsz;
365 static int ds_request(struct task_struct *task, void *base, size_t size,
366 ds_ovfl_callback_t ovfl, enum ds_qualifier qual)
368 struct ds_context *context;
369 unsigned long buffer, adj;
370 const unsigned long alignment = (1 << 3);
373 if (!ds_cfg.sizeof_ds)
376 /* we require some space to do alignment adjustments below */
377 if (size < (alignment + ds_cfg.sizeof_rec[qual]))
380 /* buffer overflow notification is not yet implemented */
387 if (!check_tracer(task))
391 context = ds_alloc_context(task);
396 if (context->owner[qual] == current)
399 if (context->owner[qual] != NULL)
401 context->owner[qual] = current;
403 spin_unlock(&ds_lock);
408 base = ds_allocate_buffer(size, &context->pages[qual]);
412 context->buffer[qual] = base;
416 context->callback[qual] = ovfl;
418 /* adjust the buffer address and size to meet alignment
420 * - buffer is double-word aligned
421 * - size is multiple of record size
423 * We checked the size at the very beginning; we have enough
424 * space to do the adjustment.
426 buffer = (unsigned long)base;
428 adj = ALIGN(buffer, alignment) - buffer;
432 size /= ds_cfg.sizeof_rec[qual];
433 size *= ds_cfg.sizeof_rec[qual];
435 ds_set(context->ds, qual, ds_buffer_base, buffer);
436 ds_set(context->ds, qual, ds_index, buffer);
437 ds_set(context->ds, qual, ds_absolute_maximum, buffer + size);
440 /* todo: select a suitable interrupt threshold */
442 ds_set(context->ds, qual,
443 ds_interrupt_threshold, buffer + size + 1);
445 /* we keep the context until ds_release */
449 context->owner[qual] = NULL;
450 ds_put_context(context);
454 spin_unlock(&ds_lock);
455 ds_put_context(context);
459 int ds_request_bts(struct task_struct *task, void *base, size_t size,
460 ds_ovfl_callback_t ovfl)
462 return ds_request(task, base, size, ovfl, ds_bts);
465 int ds_request_pebs(struct task_struct *task, void *base, size_t size,
466 ds_ovfl_callback_t ovfl)
468 return ds_request(task, base, size, ovfl, ds_pebs);
471 static int ds_release(struct task_struct *task, enum ds_qualifier qual)
473 struct ds_context *context;
476 context = ds_get_context(task);
477 error = ds_validate_access(context, qual);
481 kfree(context->buffer[qual]);
482 context->buffer[qual] = NULL;
484 current->mm->total_vm -= context->pages[qual];
485 current->mm->locked_vm -= context->pages[qual];
486 context->pages[qual] = 0;
487 context->owner[qual] = NULL;
490 * we put the context twice:
491 * once for the ds_get_context
492 * once for the corresponding ds_request
494 ds_put_context(context);
496 ds_put_context(context);
500 int ds_release_bts(struct task_struct *task)
502 return ds_release(task, ds_bts);
505 int ds_release_pebs(struct task_struct *task)
507 return ds_release(task, ds_pebs);
510 static int ds_get_index(struct task_struct *task, size_t *pos,
511 enum ds_qualifier qual)
513 struct ds_context *context;
514 unsigned long base, index;
517 context = ds_get_context(task);
518 error = ds_validate_access(context, qual);
522 base = ds_get(context->ds, qual, ds_buffer_base);
523 index = ds_get(context->ds, qual, ds_index);
525 error = ((index - base) / ds_cfg.sizeof_rec[qual]);
529 ds_put_context(context);
533 int ds_get_bts_index(struct task_struct *task, size_t *pos)
535 return ds_get_index(task, pos, ds_bts);
538 int ds_get_pebs_index(struct task_struct *task, size_t *pos)
540 return ds_get_index(task, pos, ds_pebs);
543 static int ds_get_end(struct task_struct *task, size_t *pos,
544 enum ds_qualifier qual)
546 struct ds_context *context;
547 unsigned long base, end;
550 context = ds_get_context(task);
551 error = ds_validate_access(context, qual);
555 base = ds_get(context->ds, qual, ds_buffer_base);
556 end = ds_get(context->ds, qual, ds_absolute_maximum);
558 error = ((end - base) / ds_cfg.sizeof_rec[qual]);
562 ds_put_context(context);
566 int ds_get_bts_end(struct task_struct *task, size_t *pos)
568 return ds_get_end(task, pos, ds_bts);
571 int ds_get_pebs_end(struct task_struct *task, size_t *pos)
573 return ds_get_end(task, pos, ds_pebs);
576 static int ds_access(struct task_struct *task, size_t index,
577 const void **record, enum ds_qualifier qual)
579 struct ds_context *context;
580 unsigned long base, idx;
586 context = ds_get_context(task);
587 error = ds_validate_access(context, qual);
591 base = ds_get(context->ds, qual, ds_buffer_base);
592 idx = base + (index * ds_cfg.sizeof_rec[qual]);
595 if (idx > ds_get(context->ds, qual, ds_absolute_maximum))
598 *record = (const void *)idx;
599 error = ds_cfg.sizeof_rec[qual];
601 ds_put_context(context);
605 int ds_access_bts(struct task_struct *task, size_t index, const void **record)
607 return ds_access(task, index, record, ds_bts);
610 int ds_access_pebs(struct task_struct *task, size_t index, const void **record)
612 return ds_access(task, index, record, ds_pebs);
615 static int ds_write(struct task_struct *task, const void *record, size_t size,
616 enum ds_qualifier qual, int force)
618 struct ds_context *context;
625 context = ds_get_context(task);
630 error = ds_validate_access(context, qual);
637 unsigned long base, index, end, write_end, int_th;
638 unsigned long write_size, adj_write_size;
641 * write as much as possible without producing an
642 * overflow interrupt.
644 * interrupt_threshold must either be
645 * - bigger than absolute_maximum or
646 * - point to a record between buffer_base and absolute_maximum
648 * index points to a valid record.
650 base = ds_get(context->ds, qual, ds_buffer_base);
651 index = ds_get(context->ds, qual, ds_index);
652 end = ds_get(context->ds, qual, ds_absolute_maximum);
653 int_th = ds_get(context->ds, qual, ds_interrupt_threshold);
655 write_end = min(end, int_th);
657 /* if we are already beyond the interrupt threshold,
658 * we fill the entire buffer */
659 if (write_end <= index)
662 if (write_end <= index)
665 write_size = min((unsigned long) size, write_end - index);
666 memcpy((void *)index, record, write_size);
668 record = (const char *)record + write_size;
672 adj_write_size = write_size / ds_cfg.sizeof_rec[qual];
673 adj_write_size *= ds_cfg.sizeof_rec[qual];
675 /* zero out trailing bytes */
676 memset((char *)index + write_size, 0,
677 adj_write_size - write_size);
678 index += adj_write_size;
682 ds_set(context->ds, qual, ds_index, index);
685 ds_overflow(task, context, qual);
689 ds_put_context(context);
693 int ds_write_bts(struct task_struct *task, const void *record, size_t size)
695 return ds_write(task, record, size, ds_bts, /* force = */ 0);
698 int ds_write_pebs(struct task_struct *task, const void *record, size_t size)
700 return ds_write(task, record, size, ds_pebs, /* force = */ 0);
703 int ds_unchecked_write_bts(struct task_struct *task,
704 const void *record, size_t size)
706 return ds_write(task, record, size, ds_bts, /* force = */ 1);
709 int ds_unchecked_write_pebs(struct task_struct *task,
710 const void *record, size_t size)
712 return ds_write(task, record, size, ds_pebs, /* force = */ 1);
715 static int ds_reset_or_clear(struct task_struct *task,
716 enum ds_qualifier qual, int clear)
718 struct ds_context *context;
719 unsigned long base, end;
722 context = ds_get_context(task);
723 error = ds_validate_access(context, qual);
727 base = ds_get(context->ds, qual, ds_buffer_base);
728 end = ds_get(context->ds, qual, ds_absolute_maximum);
731 memset((void *)base, 0, end - base);
733 ds_set(context->ds, qual, ds_index, base);
737 ds_put_context(context);
741 int ds_reset_bts(struct task_struct *task)
743 return ds_reset_or_clear(task, ds_bts, /* clear = */ 0);
746 int ds_reset_pebs(struct task_struct *task)
748 return ds_reset_or_clear(task, ds_pebs, /* clear = */ 0);
751 int ds_clear_bts(struct task_struct *task)
753 return ds_reset_or_clear(task, ds_bts, /* clear = */ 1);
756 int ds_clear_pebs(struct task_struct *task)
758 return ds_reset_or_clear(task, ds_pebs, /* clear = */ 1);
761 int ds_get_pebs_reset(struct task_struct *task, u64 *value)
763 struct ds_context *context;
769 context = ds_get_context(task);
770 error = ds_validate_access(context, ds_pebs);
774 *value = *(u64 *)(context->ds + (ds_cfg.sizeof_field * 8));
778 ds_put_context(context);
782 int ds_set_pebs_reset(struct task_struct *task, u64 value)
784 struct ds_context *context;
787 context = ds_get_context(task);
788 error = ds_validate_access(context, ds_pebs);
792 *(u64 *)(context->ds + (ds_cfg.sizeof_field * 8)) = value;
796 ds_put_context(context);
800 static const struct ds_configuration ds_cfg_var = {
801 .sizeof_ds = sizeof(long) * 12,
802 .sizeof_field = sizeof(long),
803 .sizeof_rec[ds_bts] = sizeof(long) * 3,
804 .sizeof_rec[ds_pebs] = sizeof(long) * 10
806 static const struct ds_configuration ds_cfg_64 = {
809 .sizeof_rec[ds_bts] = 8 * 3,
810 .sizeof_rec[ds_pebs] = 8 * 10
814 ds_configure(const struct ds_configuration *cfg)
819 void __cpuinit ds_init_intel(struct cpuinfo_x86 *c)
823 switch (c->x86_model) {
825 case 0xE: /* Pentium M */
826 ds_configure(&ds_cfg_var);
828 case 0xF: /* Core2 */
829 case 0x1C: /* Atom */
830 ds_configure(&ds_cfg_64);
833 /* sorry, don't know about them */
838 switch (c->x86_model) {
841 case 0x2: /* Netburst */
842 ds_configure(&ds_cfg_var);
845 /* sorry, don't know about them */
850 /* sorry, don't know about them */
855 void ds_free(struct ds_context *context)
857 /* This is called when the task owning the parameter context
858 * is dying. There should not be any user of that context left
859 * to disturb us, anymore. */
860 unsigned long leftovers = context->count;
862 ds_put_context(context);
864 #endif /* CONFIG_X86_DS */