Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6] / arch / x86 / kernel / ds.c
1 /*
2  * Debug Store support
3  *
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).
7  *
8  * It manages:
9  * - DS and BTS hardware configuration
10  * - buffer overflow handling (to be done)
11  * - buffer access
12  *
13  * It does not do:
14  * - security checking (is the caller allowed to trace the task)
15  * - buffer allocation (memory accounting)
16  *
17  *
18  * Copyright (C) 2007-2009 Intel Corporation.
19  * Markus Metzger <markus.t.metzger@intel.com>, 2007-2009
20  */
21
22
23 #include <asm/ds.h>
24
25 #include <linux/errno.h>
26 #include <linux/string.h>
27 #include <linux/slab.h>
28 #include <linux/sched.h>
29 #include <linux/mm.h>
30 #include <linux/kernel.h>
31
32
33 /*
34  * The configuration for a particular DS hardware implementation.
35  */
36 struct ds_configuration {
37         /* the name of the configuration */
38         const char *name;
39         /* the size of one pointer-typed field in the DS structure and
40            in the BTS and PEBS buffers in bytes;
41            this covers the first 8 DS fields related to buffer management. */
42         unsigned char  sizeof_field;
43         /* the size of a BTS/PEBS record in bytes */
44         unsigned char  sizeof_rec[2];
45         /* a series of bit-masks to control various features indexed
46          * by enum ds_feature */
47         unsigned long ctl[dsf_ctl_max];
48 };
49 static DEFINE_PER_CPU(struct ds_configuration, ds_cfg_array);
50
51 #define ds_cfg per_cpu(ds_cfg_array, smp_processor_id())
52
53 #define MAX_SIZEOF_DS (12 * 8)  /* maximal size of a DS configuration */
54 #define MAX_SIZEOF_BTS (3 * 8)  /* maximal size of a BTS record */
55 #define DS_ALIGNMENT (1 << 3)   /* BTS and PEBS buffer alignment */
56
57 #define BTS_CONTROL \
58  (ds_cfg.ctl[dsf_bts] | ds_cfg.ctl[dsf_bts_kernel] | ds_cfg.ctl[dsf_bts_user] |\
59   ds_cfg.ctl[dsf_bts_overflow])
60
61
62 /*
63  * A BTS or PEBS tracer.
64  *
65  * This holds the configuration of the tracer and serves as a handle
66  * to identify tracers.
67  */
68 struct ds_tracer {
69         /* the DS context (partially) owned by this tracer */
70         struct ds_context *context;
71         /* the buffer provided on ds_request() and its size in bytes */
72         void *buffer;
73         size_t size;
74 };
75
76 struct bts_tracer {
77         /* the common DS part */
78         struct ds_tracer ds;
79         /* the trace including the DS configuration */
80         struct bts_trace trace;
81         /* buffer overflow notification function */
82         bts_ovfl_callback_t ovfl;
83 };
84
85 struct pebs_tracer {
86         /* the common DS part */
87         struct ds_tracer ds;
88         /* the trace including the DS configuration */
89         struct pebs_trace trace;
90         /* buffer overflow notification function */
91         pebs_ovfl_callback_t ovfl;
92 };
93
94 /*
95  * Debug Store (DS) save area configuration (see Intel64 and IA32
96  * Architectures Software Developer's Manual, section 18.5)
97  *
98  * The DS configuration consists of the following fields; different
99  * architetures vary in the size of those fields.
100  * - double-word aligned base linear address of the BTS buffer
101  * - write pointer into the BTS buffer
102  * - end linear address of the BTS buffer (one byte beyond the end of
103  *   the buffer)
104  * - interrupt pointer into BTS buffer
105  *   (interrupt occurs when write pointer passes interrupt pointer)
106  * - double-word aligned base linear address of the PEBS buffer
107  * - write pointer into the PEBS buffer
108  * - end linear address of the PEBS buffer (one byte beyond the end of
109  *   the buffer)
110  * - interrupt pointer into PEBS buffer
111  *   (interrupt occurs when write pointer passes interrupt pointer)
112  * - value to which counter is reset following counter overflow
113  *
114  * Later architectures use 64bit pointers throughout, whereas earlier
115  * architectures use 32bit pointers in 32bit mode.
116  *
117  *
118  * We compute the base address for the first 8 fields based on:
119  * - the field size stored in the DS configuration
120  * - the relative field position
121  * - an offset giving the start of the respective region
122  *
123  * This offset is further used to index various arrays holding
124  * information for BTS and PEBS at the respective index.
125  *
126  * On later 32bit processors, we only access the lower 32bit of the
127  * 64bit pointer fields. The upper halves will be zeroed out.
128  */
129
130 enum ds_field {
131         ds_buffer_base = 0,
132         ds_index,
133         ds_absolute_maximum,
134         ds_interrupt_threshold,
135 };
136
137 enum ds_qualifier {
138         ds_bts  = 0,
139         ds_pebs
140 };
141
142 static inline unsigned long ds_get(const unsigned char *base,
143                                    enum ds_qualifier qual, enum ds_field field)
144 {
145         base += (ds_cfg.sizeof_field * (field + (4 * qual)));
146         return *(unsigned long *)base;
147 }
148
149 static inline void ds_set(unsigned char *base, enum ds_qualifier qual,
150                           enum ds_field field, unsigned long value)
151 {
152         base += (ds_cfg.sizeof_field * (field + (4 * qual)));
153         (*(unsigned long *)base) = value;
154 }
155
156
157 /*
158  * Locking is done only for allocating BTS or PEBS resources.
159  */
160 static DEFINE_SPINLOCK(ds_lock);
161
162
163 /*
164  * We either support (system-wide) per-cpu or per-thread allocation.
165  * We distinguish the two based on the task_struct pointer, where a
166  * NULL pointer indicates per-cpu allocation for the current cpu.
167  *
168  * Allocations are use-counted. As soon as resources are allocated,
169  * further allocations must be of the same type (per-cpu or
170  * per-thread). We model this by counting allocations (i.e. the number
171  * of tracers of a certain type) for one type negatively:
172  *   =0  no tracers
173  *   >0  number of per-thread tracers
174  *   <0  number of per-cpu tracers
175  *
176  * Tracers essentially gives the number of ds contexts for a certain
177  * type of allocation.
178  */
179 static atomic_t tracers = ATOMIC_INIT(0);
180
181 static inline void get_tracer(struct task_struct *task)
182 {
183         if (task)
184                 atomic_inc(&tracers);
185         else
186                 atomic_dec(&tracers);
187 }
188
189 static inline void put_tracer(struct task_struct *task)
190 {
191         if (task)
192                 atomic_dec(&tracers);
193         else
194                 atomic_inc(&tracers);
195 }
196
197 static inline int check_tracer(struct task_struct *task)
198 {
199         return task ?
200                 (atomic_read(&tracers) >= 0) :
201                 (atomic_read(&tracers) <= 0);
202 }
203
204
205 /*
206  * The DS context is either attached to a thread or to a cpu:
207  * - in the former case, the thread_struct contains a pointer to the
208  *   attached context.
209  * - in the latter case, we use a static array of per-cpu context
210  *   pointers.
211  *
212  * Contexts are use-counted. They are allocated on first access and
213  * deallocated when the last user puts the context.
214  */
215 struct ds_context {
216         /* pointer to the DS configuration; goes into MSR_IA32_DS_AREA */
217         unsigned char ds[MAX_SIZEOF_DS];
218         /* the owner of the BTS and PEBS configuration, respectively */
219         struct bts_tracer *bts_master;
220         struct pebs_tracer *pebs_master;
221         /* use count */
222         unsigned long count;
223         /* a pointer to the context location inside the thread_struct
224          * or the per_cpu context array */
225         struct ds_context **this;
226         /* a pointer to the task owning this context, or NULL, if the
227          * context is owned by a cpu */
228         struct task_struct *task;
229 };
230
231 static DEFINE_PER_CPU(struct ds_context *, system_context_array);
232
233 #define system_context per_cpu(system_context_array, smp_processor_id())
234
235
236 static inline struct ds_context *ds_get_context(struct task_struct *task)
237 {
238         struct ds_context **p_context =
239                 (task ? &task->thread.ds_ctx : &system_context);
240         struct ds_context *context = NULL;
241         struct ds_context *new_context = NULL;
242         unsigned long irq;
243
244         /* Chances are small that we already have a context. */
245         new_context = kzalloc(sizeof(*new_context), GFP_KERNEL);
246         if (!new_context)
247                 return NULL;
248
249         spin_lock_irqsave(&ds_lock, irq);
250
251         context = *p_context;
252         if (!context) {
253                 context = new_context;
254
255                 context->this = p_context;
256                 context->task = task;
257                 context->count = 0;
258
259                 if (task)
260                         set_tsk_thread_flag(task, TIF_DS_AREA_MSR);
261
262                 if (!task || (task == current))
263                         wrmsrl(MSR_IA32_DS_AREA, (unsigned long)context->ds);
264
265                 *p_context = context;
266         }
267
268         context->count++;
269
270         spin_unlock_irqrestore(&ds_lock, irq);
271
272         if (context != new_context)
273                 kfree(new_context);
274
275         return context;
276 }
277
278 static inline void ds_put_context(struct ds_context *context)
279 {
280         unsigned long irq;
281
282         if (!context)
283                 return;
284
285         spin_lock_irqsave(&ds_lock, irq);
286
287         if (--context->count) {
288                 spin_unlock_irqrestore(&ds_lock, irq);
289                 return;
290         }
291
292         *(context->this) = NULL;
293
294         if (context->task)
295                 clear_tsk_thread_flag(context->task, TIF_DS_AREA_MSR);
296
297         if (!context->task || (context->task == current))
298                 wrmsrl(MSR_IA32_DS_AREA, 0);
299
300         spin_unlock_irqrestore(&ds_lock, irq);
301
302         kfree(context);
303 }
304
305
306 /*
307  * Call the tracer's callback on a buffer overflow.
308  *
309  * context: the ds context
310  * qual: the buffer type
311  */
312 static void ds_overflow(struct ds_context *context, enum ds_qualifier qual)
313 {
314         switch (qual) {
315         case ds_bts:
316                 if (context->bts_master &&
317                     context->bts_master->ovfl)
318                         context->bts_master->ovfl(context->bts_master);
319                 break;
320         case ds_pebs:
321                 if (context->pebs_master &&
322                     context->pebs_master->ovfl)
323                         context->pebs_master->ovfl(context->pebs_master);
324                 break;
325         }
326 }
327
328
329 /*
330  * Write raw data into the BTS or PEBS buffer.
331  *
332  * The remainder of any partially written record is zeroed out.
333  *
334  * context: the DS context
335  * qual: the buffer type
336  * record: the data to write
337  * size: the size of the data
338  */
339 static int ds_write(struct ds_context *context, enum ds_qualifier qual,
340                     const void *record, size_t size)
341 {
342         int bytes_written = 0;
343
344         if (!record)
345                 return -EINVAL;
346
347         while (size) {
348                 unsigned long base, index, end, write_end, int_th;
349                 unsigned long write_size, adj_write_size;
350
351                 /*
352                  * write as much as possible without producing an
353                  * overflow interrupt.
354                  *
355                  * interrupt_threshold must either be
356                  * - bigger than absolute_maximum or
357                  * - point to a record between buffer_base and absolute_maximum
358                  *
359                  * index points to a valid record.
360                  */
361                 base   = ds_get(context->ds, qual, ds_buffer_base);
362                 index  = ds_get(context->ds, qual, ds_index);
363                 end    = ds_get(context->ds, qual, ds_absolute_maximum);
364                 int_th = ds_get(context->ds, qual, ds_interrupt_threshold);
365
366                 write_end = min(end, int_th);
367
368                 /* if we are already beyond the interrupt threshold,
369                  * we fill the entire buffer */
370                 if (write_end <= index)
371                         write_end = end;
372
373                 if (write_end <= index)
374                         break;
375
376                 write_size = min((unsigned long) size, write_end - index);
377                 memcpy((void *)index, record, write_size);
378
379                 record = (const char *)record + write_size;
380                 size -= write_size;
381                 bytes_written += write_size;
382
383                 adj_write_size = write_size / ds_cfg.sizeof_rec[qual];
384                 adj_write_size *= ds_cfg.sizeof_rec[qual];
385
386                 /* zero out trailing bytes */
387                 memset((char *)index + write_size, 0,
388                        adj_write_size - write_size);
389                 index += adj_write_size;
390
391                 if (index >= end)
392                         index = base;
393                 ds_set(context->ds, qual, ds_index, index);
394
395                 if (index >= int_th)
396                         ds_overflow(context, qual);
397         }
398
399         return bytes_written;
400 }
401
402
403 /*
404  * Branch Trace Store (BTS) uses the following format. Different
405  * architectures vary in the size of those fields.
406  * - source linear address
407  * - destination linear address
408  * - flags
409  *
410  * Later architectures use 64bit pointers throughout, whereas earlier
411  * architectures use 32bit pointers in 32bit mode.
412  *
413  * We compute the base address for the first 8 fields based on:
414  * - the field size stored in the DS configuration
415  * - the relative field position
416  *
417  * In order to store additional information in the BTS buffer, we use
418  * a special source address to indicate that the record requires
419  * special interpretation.
420  *
421  * Netburst indicated via a bit in the flags field whether the branch
422  * was predicted; this is ignored.
423  *
424  * We use two levels of abstraction:
425  * - the raw data level defined here
426  * - an arch-independent level defined in ds.h
427  */
428
429 enum bts_field {
430         bts_from,
431         bts_to,
432         bts_flags,
433
434         bts_qual = bts_from,
435         bts_jiffies = bts_to,
436         bts_pid = bts_flags,
437
438         bts_qual_mask = (bts_qual_max - 1),
439         bts_escape = ((unsigned long)-1 & ~bts_qual_mask)
440 };
441
442 static inline unsigned long bts_get(const char *base, enum bts_field field)
443 {
444         base += (ds_cfg.sizeof_field * field);
445         return *(unsigned long *)base;
446 }
447
448 static inline void bts_set(char *base, enum bts_field field, unsigned long val)
449 {
450         base += (ds_cfg.sizeof_field * field);;
451         (*(unsigned long *)base) = val;
452 }
453
454
455 /*
456  * The raw BTS data is architecture dependent.
457  *
458  * For higher-level users, we give an arch-independent view.
459  * - ds.h defines struct bts_struct
460  * - bts_read translates one raw bts record into a bts_struct
461  * - bts_write translates one bts_struct into the raw format and
462  *   writes it into the top of the parameter tracer's buffer.
463  *
464  * return: bytes read/written on success; -Eerrno, otherwise
465  */
466 static int bts_read(struct bts_tracer *tracer, const void *at,
467                     struct bts_struct *out)
468 {
469         if (!tracer)
470                 return -EINVAL;
471
472         if (at < tracer->trace.ds.begin)
473                 return -EINVAL;
474
475         if (tracer->trace.ds.end < (at + tracer->trace.ds.size))
476                 return -EINVAL;
477
478         memset(out, 0, sizeof(*out));
479         if ((bts_get(at, bts_qual) & ~bts_qual_mask) == bts_escape) {
480                 out->qualifier = (bts_get(at, bts_qual) & bts_qual_mask);
481                 out->variant.timestamp.jiffies = bts_get(at, bts_jiffies);
482                 out->variant.timestamp.pid = bts_get(at, bts_pid);
483         } else {
484                 out->qualifier = bts_branch;
485                 out->variant.lbr.from = bts_get(at, bts_from);
486                 out->variant.lbr.to   = bts_get(at, bts_to);
487
488                 if (!out->variant.lbr.from && !out->variant.lbr.to)
489                         out->qualifier = bts_invalid;
490         }
491
492         return ds_cfg.sizeof_rec[ds_bts];
493 }
494
495 static int bts_write(struct bts_tracer *tracer, const struct bts_struct *in)
496 {
497         unsigned char raw[MAX_SIZEOF_BTS];
498
499         if (!tracer)
500                 return -EINVAL;
501
502         if (MAX_SIZEOF_BTS < ds_cfg.sizeof_rec[ds_bts])
503                 return -EOVERFLOW;
504
505         switch (in->qualifier) {
506         case bts_invalid:
507                 bts_set(raw, bts_from, 0);
508                 bts_set(raw, bts_to, 0);
509                 bts_set(raw, bts_flags, 0);
510                 break;
511         case bts_branch:
512                 bts_set(raw, bts_from, in->variant.lbr.from);
513                 bts_set(raw, bts_to,   in->variant.lbr.to);
514                 bts_set(raw, bts_flags, 0);
515                 break;
516         case bts_task_arrives:
517         case bts_task_departs:
518                 bts_set(raw, bts_qual, (bts_escape | in->qualifier));
519                 bts_set(raw, bts_jiffies, in->variant.timestamp.jiffies);
520                 bts_set(raw, bts_pid, in->variant.timestamp.pid);
521                 break;
522         default:
523                 return -EINVAL;
524         }
525
526         return ds_write(tracer->ds.context, ds_bts, raw,
527                         ds_cfg.sizeof_rec[ds_bts]);
528 }
529
530
531 static void ds_write_config(struct ds_context *context,
532                             struct ds_trace *cfg, enum ds_qualifier qual)
533 {
534         unsigned char *ds = context->ds;
535
536         ds_set(ds, qual, ds_buffer_base, (unsigned long)cfg->begin);
537         ds_set(ds, qual, ds_index, (unsigned long)cfg->top);
538         ds_set(ds, qual, ds_absolute_maximum, (unsigned long)cfg->end);
539         ds_set(ds, qual, ds_interrupt_threshold, (unsigned long)cfg->ith);
540 }
541
542 static void ds_read_config(struct ds_context *context,
543                            struct ds_trace *cfg, enum ds_qualifier qual)
544 {
545         unsigned char *ds = context->ds;
546
547         cfg->begin = (void *)ds_get(ds, qual, ds_buffer_base);
548         cfg->top = (void *)ds_get(ds, qual, ds_index);
549         cfg->end = (void *)ds_get(ds, qual, ds_absolute_maximum);
550         cfg->ith = (void *)ds_get(ds, qual, ds_interrupt_threshold);
551 }
552
553 static void ds_init_ds_trace(struct ds_trace *trace, enum ds_qualifier qual,
554                              void *base, size_t size, size_t ith,
555                              unsigned int flags) {
556         unsigned long buffer, adj;
557
558         /* adjust the buffer address and size to meet alignment
559          * constraints:
560          * - buffer is double-word aligned
561          * - size is multiple of record size
562          *
563          * We checked the size at the very beginning; we have enough
564          * space to do the adjustment.
565          */
566         buffer = (unsigned long)base;
567
568         adj = ALIGN(buffer, DS_ALIGNMENT) - buffer;
569         buffer += adj;
570         size   -= adj;
571
572         trace->n = size / ds_cfg.sizeof_rec[qual];
573         trace->size = ds_cfg.sizeof_rec[qual];
574
575         size = (trace->n * trace->size);
576
577         trace->begin = (void *)buffer;
578         trace->top = trace->begin;
579         trace->end = (void *)(buffer + size);
580         /* The value for 'no threshold' is -1, which will set the
581          * threshold outside of the buffer, just like we want it.
582          */
583         trace->ith = (void *)(buffer + size - ith);
584
585         trace->flags = flags;
586 }
587
588
589 static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace,
590                       enum ds_qualifier qual, struct task_struct *task,
591                       void *base, size_t size, size_t th, unsigned int flags)
592 {
593         struct ds_context *context;
594         int error;
595
596         error = -EINVAL;
597         if (!base)
598                 goto out;
599
600         /* we require some space to do alignment adjustments below */
601         error = -EINVAL;
602         if (size < (DS_ALIGNMENT + ds_cfg.sizeof_rec[qual]))
603                 goto out;
604
605         if (th != (size_t)-1) {
606                 th *= ds_cfg.sizeof_rec[qual];
607
608                 error = -EINVAL;
609                 if (size <= th)
610                         goto out;
611         }
612
613         tracer->buffer = base;
614         tracer->size = size;
615
616         error = -ENOMEM;
617         context = ds_get_context(task);
618         if (!context)
619                 goto out;
620         tracer->context = context;
621
622         ds_init_ds_trace(trace, qual, base, size, th, flags);
623
624         error = 0;
625  out:
626         return error;
627 }
628
629 struct bts_tracer *ds_request_bts(struct task_struct *task,
630                                   void *base, size_t size,
631                                   bts_ovfl_callback_t ovfl, size_t th,
632                                   unsigned int flags)
633 {
634         struct bts_tracer *tracer;
635         unsigned long irq;
636         int error;
637
638         error = -EOPNOTSUPP;
639         if (!ds_cfg.ctl[dsf_bts])
640                 goto out;
641
642         /* buffer overflow notification is not yet implemented */
643         error = -EOPNOTSUPP;
644         if (ovfl)
645                 goto out;
646
647         error = -ENOMEM;
648         tracer = kzalloc(sizeof(*tracer), GFP_KERNEL);
649         if (!tracer)
650                 goto out;
651         tracer->ovfl = ovfl;
652
653         error = ds_request(&tracer->ds, &tracer->trace.ds,
654                            ds_bts, task, base, size, th, flags);
655         if (error < 0)
656                 goto out_tracer;
657
658
659         spin_lock_irqsave(&ds_lock, irq);
660
661         error = -EPERM;
662         if (!check_tracer(task))
663                 goto out_unlock;
664         get_tracer(task);
665
666         error = -EPERM;
667         if (tracer->ds.context->bts_master)
668                 goto out_put_tracer;
669         tracer->ds.context->bts_master = tracer;
670
671         spin_unlock_irqrestore(&ds_lock, irq);
672
673
674         tracer->trace.read  = bts_read;
675         tracer->trace.write = bts_write;
676
677         ds_write_config(tracer->ds.context, &tracer->trace.ds, ds_bts);
678         ds_resume_bts(tracer);
679
680         return tracer;
681
682  out_put_tracer:
683         put_tracer(task);
684  out_unlock:
685         spin_unlock_irqrestore(&ds_lock, irq);
686         ds_put_context(tracer->ds.context);
687  out_tracer:
688         kfree(tracer);
689  out:
690         return ERR_PTR(error);
691 }
692
693 struct pebs_tracer *ds_request_pebs(struct task_struct *task,
694                                     void *base, size_t size,
695                                     pebs_ovfl_callback_t ovfl, size_t th,
696                                     unsigned int flags)
697 {
698         struct pebs_tracer *tracer;
699         unsigned long irq;
700         int error;
701
702         /* buffer overflow notification is not yet implemented */
703         error = -EOPNOTSUPP;
704         if (ovfl)
705                 goto out;
706
707         error = -ENOMEM;
708         tracer = kzalloc(sizeof(*tracer), GFP_KERNEL);
709         if (!tracer)
710                 goto out;
711         tracer->ovfl = ovfl;
712
713         error = ds_request(&tracer->ds, &tracer->trace.ds,
714                            ds_pebs, task, base, size, th, flags);
715         if (error < 0)
716                 goto out_tracer;
717
718         spin_lock_irqsave(&ds_lock, irq);
719
720         error = -EPERM;
721         if (!check_tracer(task))
722                 goto out_unlock;
723         get_tracer(task);
724
725         error = -EPERM;
726         if (tracer->ds.context->pebs_master)
727                 goto out_put_tracer;
728         tracer->ds.context->pebs_master = tracer;
729
730         spin_unlock_irqrestore(&ds_lock, irq);
731
732         ds_write_config(tracer->ds.context, &tracer->trace.ds, ds_bts);
733         ds_resume_pebs(tracer);
734
735         return tracer;
736
737  out_put_tracer:
738         put_tracer(task);
739  out_unlock:
740         spin_unlock_irqrestore(&ds_lock, irq);
741         ds_put_context(tracer->ds.context);
742  out_tracer:
743         kfree(tracer);
744  out:
745         return ERR_PTR(error);
746 }
747
748 void ds_release_bts(struct bts_tracer *tracer)
749 {
750         if (!tracer)
751                 return;
752
753         ds_suspend_bts(tracer);
754
755         WARN_ON_ONCE(tracer->ds.context->bts_master != tracer);
756         tracer->ds.context->bts_master = NULL;
757
758         put_tracer(tracer->ds.context->task);
759         ds_put_context(tracer->ds.context);
760
761         kfree(tracer);
762 }
763
764 void ds_suspend_bts(struct bts_tracer *tracer)
765 {
766         struct task_struct *task;
767
768         if (!tracer)
769                 return;
770
771         task = tracer->ds.context->task;
772
773         if (!task || (task == current))
774                 update_debugctlmsr(get_debugctlmsr() & ~BTS_CONTROL);
775
776         if (task) {
777                 task->thread.debugctlmsr &= ~BTS_CONTROL;
778
779                 if (!task->thread.debugctlmsr)
780                         clear_tsk_thread_flag(task, TIF_DEBUGCTLMSR);
781         }
782 }
783
784 void ds_resume_bts(struct bts_tracer *tracer)
785 {
786         struct task_struct *task;
787         unsigned long control;
788
789         if (!tracer)
790                 return;
791
792         task = tracer->ds.context->task;
793
794         control = ds_cfg.ctl[dsf_bts];
795         if (!(tracer->trace.ds.flags & BTS_KERNEL))
796                 control |= ds_cfg.ctl[dsf_bts_kernel];
797         if (!(tracer->trace.ds.flags & BTS_USER))
798                 control |= ds_cfg.ctl[dsf_bts_user];
799
800         if (task) {
801                 task->thread.debugctlmsr |= control;
802                 set_tsk_thread_flag(task, TIF_DEBUGCTLMSR);
803         }
804
805         if (!task || (task == current))
806                 update_debugctlmsr(get_debugctlmsr() | control);
807 }
808
809 void ds_release_pebs(struct pebs_tracer *tracer)
810 {
811         if (!tracer)
812                 return;
813
814         ds_suspend_pebs(tracer);
815
816         WARN_ON_ONCE(tracer->ds.context->pebs_master != tracer);
817         tracer->ds.context->pebs_master = NULL;
818
819         put_tracer(tracer->ds.context->task);
820         ds_put_context(tracer->ds.context);
821
822         kfree(tracer);
823 }
824
825 void ds_suspend_pebs(struct pebs_tracer *tracer)
826 {
827
828 }
829
830 void ds_resume_pebs(struct pebs_tracer *tracer)
831 {
832
833 }
834
835 const struct bts_trace *ds_read_bts(struct bts_tracer *tracer)
836 {
837         if (!tracer)
838                 return NULL;
839
840         ds_read_config(tracer->ds.context, &tracer->trace.ds, ds_bts);
841         return &tracer->trace;
842 }
843
844 const struct pebs_trace *ds_read_pebs(struct pebs_tracer *tracer)
845 {
846         if (!tracer)
847                 return NULL;
848
849         ds_read_config(tracer->ds.context, &tracer->trace.ds, ds_pebs);
850         tracer->trace.reset_value =
851                 *(u64 *)(tracer->ds.context->ds + (ds_cfg.sizeof_field * 8));
852
853         return &tracer->trace;
854 }
855
856 int ds_reset_bts(struct bts_tracer *tracer)
857 {
858         if (!tracer)
859                 return -EINVAL;
860
861         tracer->trace.ds.top = tracer->trace.ds.begin;
862
863         ds_set(tracer->ds.context->ds, ds_bts, ds_index,
864                (unsigned long)tracer->trace.ds.top);
865
866         return 0;
867 }
868
869 int ds_reset_pebs(struct pebs_tracer *tracer)
870 {
871         if (!tracer)
872                 return -EINVAL;
873
874         tracer->trace.ds.top = tracer->trace.ds.begin;
875
876         ds_set(tracer->ds.context->ds, ds_bts, ds_index,
877                (unsigned long)tracer->trace.ds.top);
878
879         return 0;
880 }
881
882 int ds_set_pebs_reset(struct pebs_tracer *tracer, u64 value)
883 {
884         if (!tracer)
885                 return -EINVAL;
886
887         *(u64 *)(tracer->ds.context->ds + (ds_cfg.sizeof_field * 8)) = value;
888
889         return 0;
890 }
891
892 static const struct ds_configuration ds_cfg_netburst = {
893         .name = "Netburst",
894         .ctl[dsf_bts]           = (1 << 2) | (1 << 3),
895         .ctl[dsf_bts_kernel]    = (1 << 5),
896         .ctl[dsf_bts_user]      = (1 << 6),
897
898         .sizeof_field           = sizeof(long),
899         .sizeof_rec[ds_bts]     = sizeof(long) * 3,
900 #ifdef __i386__
901         .sizeof_rec[ds_pebs]    = sizeof(long) * 10,
902 #else
903         .sizeof_rec[ds_pebs]    = sizeof(long) * 18,
904 #endif
905 };
906 static const struct ds_configuration ds_cfg_pentium_m = {
907         .name = "Pentium M",
908         .ctl[dsf_bts]           = (1 << 6) | (1 << 7),
909
910         .sizeof_field           = sizeof(long),
911         .sizeof_rec[ds_bts]     = sizeof(long) * 3,
912 #ifdef __i386__
913         .sizeof_rec[ds_pebs]    = sizeof(long) * 10,
914 #else
915         .sizeof_rec[ds_pebs]    = sizeof(long) * 18,
916 #endif
917 };
918 static const struct ds_configuration ds_cfg_core2_atom = {
919         .name = "Core 2/Atom",
920         .ctl[dsf_bts]           = (1 << 6) | (1 << 7),
921         .ctl[dsf_bts_kernel]    = (1 << 9),
922         .ctl[dsf_bts_user]      = (1 << 10),
923
924         .sizeof_field           = 8,
925         .sizeof_rec[ds_bts]     = 8 * 3,
926         .sizeof_rec[ds_pebs]    = 8 * 18,
927 };
928
929 static void
930 ds_configure(const struct ds_configuration *cfg)
931 {
932         memset(&ds_cfg, 0, sizeof(ds_cfg));
933         ds_cfg = *cfg;
934
935         printk(KERN_INFO "[ds] using %s configuration\n", ds_cfg.name);
936
937         if (!cpu_has_bts) {
938                 ds_cfg.ctl[dsf_bts] = 0;
939                 printk(KERN_INFO "[ds] bts not available\n");
940         }
941         if (!cpu_has_pebs)
942                 printk(KERN_INFO "[ds] pebs not available\n");
943
944         WARN_ON_ONCE(MAX_SIZEOF_DS < (12 * ds_cfg.sizeof_field));
945 }
946
947 void __cpuinit ds_init_intel(struct cpuinfo_x86 *c)
948 {
949         switch (c->x86) {
950         case 0x6:
951                 switch (c->x86_model) {
952                 case 0x9:
953                 case 0xd: /* Pentium M */
954                         ds_configure(&ds_cfg_pentium_m);
955                         break;
956                 case 0xf:
957                 case 0x17: /* Core2 */
958                 case 0x1c: /* Atom */
959                         ds_configure(&ds_cfg_core2_atom);
960                         break;
961                 case 0x1a: /* i7 */
962                 default:
963                         /* sorry, don't know about them */
964                         break;
965                 }
966                 break;
967         case 0xf:
968                 switch (c->x86_model) {
969                 case 0x0:
970                 case 0x1:
971                 case 0x2: /* Netburst */
972                         ds_configure(&ds_cfg_netburst);
973                         break;
974                 default:
975                         /* sorry, don't know about them */
976                         break;
977                 }
978                 break;
979         default:
980                 /* sorry, don't know about them */
981                 break;
982         }
983 }
984
985 /*
986  * Change the DS configuration from tracing prev to tracing next.
987  */
988 void ds_switch_to(struct task_struct *prev, struct task_struct *next)
989 {
990         struct ds_context *prev_ctx = prev->thread.ds_ctx;
991         struct ds_context *next_ctx = next->thread.ds_ctx;
992
993         if (prev_ctx) {
994                 update_debugctlmsr(0);
995
996                 if (prev_ctx->bts_master &&
997                     (prev_ctx->bts_master->trace.ds.flags & BTS_TIMESTAMPS)) {
998                         struct bts_struct ts = {
999                                 .qualifier = bts_task_departs,
1000                                 .variant.timestamp.jiffies = jiffies_64,
1001                                 .variant.timestamp.pid = prev->pid
1002                         };
1003                         bts_write(prev_ctx->bts_master, &ts);
1004                 }
1005         }
1006
1007         if (next_ctx) {
1008                 if (next_ctx->bts_master &&
1009                     (next_ctx->bts_master->trace.ds.flags & BTS_TIMESTAMPS)) {
1010                         struct bts_struct ts = {
1011                                 .qualifier = bts_task_arrives,
1012                                 .variant.timestamp.jiffies = jiffies_64,
1013                                 .variant.timestamp.pid = next->pid
1014                         };
1015                         bts_write(next_ctx->bts_master, &ts);
1016                 }
1017
1018                 wrmsrl(MSR_IA32_DS_AREA, (unsigned long)next_ctx->ds);
1019         }
1020
1021         update_debugctlmsr(next->thread.debugctlmsr);
1022 }
1023
1024 void ds_copy_thread(struct task_struct *tsk, struct task_struct *father)
1025 {
1026         clear_tsk_thread_flag(tsk, TIF_DS_AREA_MSR);
1027         tsk->thread.ds_ctx = NULL;
1028 }
1029
1030 void ds_exit_thread(struct task_struct *tsk)
1031 {
1032         WARN_ON(tsk->thread.ds_ctx);
1033 }