2 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License version 2 as
6 * published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11 * GNU General Public License for more details.
13 * You should have received a copy of the GNU General Public License
14 * along with this program; if not, write to the Free Software
15 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 #include <linux/kernel.h>
19 #include <linux/blkdev.h>
20 #include <linux/blktrace_api.h>
21 #include <linux/percpu.h>
22 #include <linux/init.h>
23 #include <linux/mutex.h>
24 #include <linux/debugfs.h>
25 #include <linux/time.h>
26 #include <trace/block.h>
27 #include <linux/uaccess.h>
28 #include <../kernel/trace/trace_output.h>
30 static unsigned int blktrace_seq __read_mostly
= 1;
32 static struct trace_array
*blk_tr
;
33 static int __read_mostly blk_tracer_enabled
;
35 /* Select an alternative, minimalistic output than the original one */
36 #define TRACE_BLK_OPT_CLASSIC 0x1
38 static struct tracer_opt blk_tracer_opts
[] = {
39 /* Default disable the minimalistic output */
40 { TRACER_OPT(blk_classic
, TRACE_BLK_OPT_CLASSIC
) },
44 static struct tracer_flags blk_tracer_flags
= {
46 .opts
= blk_tracer_opts
,
49 /* Global reference count of probes */
50 static DEFINE_MUTEX(blk_probe_mutex
);
51 static atomic_t blk_probes_ref
= ATOMIC_INIT(0);
53 static int blk_register_tracepoints(void);
54 static void blk_unregister_tracepoints(void);
57 * Send out a notify message.
59 static void trace_note(struct blk_trace
*bt
, pid_t pid
, int action
,
60 const void *data
, size_t len
)
62 struct blk_io_trace
*t
;
67 t
= relay_reserve(bt
->rchan
, sizeof(*t
) + len
);
69 const int cpu
= smp_processor_id();
71 t
->magic
= BLK_IO_TRACE_MAGIC
| BLK_IO_TRACE_VERSION
;
72 t
->time
= ktime_to_ns(ktime_get());
78 memcpy((void *) t
+ sizeof(*t
), data
, len
);
83 * Send out a notify for this process, if we haven't done so since a trace
86 static void trace_note_tsk(struct blk_trace
*bt
, struct task_struct
*tsk
)
88 tsk
->btrace_seq
= blktrace_seq
;
89 trace_note(bt
, tsk
->pid
, BLK_TN_PROCESS
, tsk
->comm
, sizeof(tsk
->comm
));
92 static void trace_note_time(struct blk_trace
*bt
)
99 words
[0] = now
.tv_sec
;
100 words
[1] = now
.tv_nsec
;
102 local_irq_save(flags
);
103 trace_note(bt
, 0, BLK_TN_TIMESTAMP
, words
, sizeof(words
));
104 local_irq_restore(flags
);
107 void __trace_note_message(struct blk_trace
*bt
, const char *fmt
, ...)
116 ftrace_vprintk(fmt
, args
);
124 local_irq_save(flags
);
125 buf
= per_cpu_ptr(bt
->msg_data
, smp_processor_id());
127 n
= vscnprintf(buf
, BLK_TN_MAX_MSG
, fmt
, args
);
130 trace_note(bt
, 0, BLK_TN_MESSAGE
, buf
, n
);
131 local_irq_restore(flags
);
133 EXPORT_SYMBOL_GPL(__trace_note_message
);
135 static int act_log_check(struct blk_trace
*bt
, u32 what
, sector_t sector
,
138 if (((bt
->act_mask
<< BLK_TC_SHIFT
) & what
) == 0)
140 if (sector
< bt
->start_lba
|| sector
> bt
->end_lba
)
142 if (bt
->pid
&& pid
!= bt
->pid
)
149 * Data direction bit lookup
151 static u32 ddir_act
[2] __read_mostly
= { BLK_TC_ACT(BLK_TC_READ
),
152 BLK_TC_ACT(BLK_TC_WRITE
) };
154 /* The ilog2() calls fall out because they're constant */
155 #define MASK_TC_BIT(rw, __name) ((rw & (1 << BIO_RW_ ## __name)) << \
156 (ilog2(BLK_TC_ ## __name) + BLK_TC_SHIFT - BIO_RW_ ## __name))
159 * The worker for the various blk_add_trace*() types. Fills out a
160 * blk_io_trace structure and places it in a per-cpu subbuffer.
162 static void __blk_add_trace(struct blk_trace
*bt
, sector_t sector
, int bytes
,
163 int rw
, u32 what
, int error
, int pdu_len
, void *pdu_data
)
165 struct task_struct
*tsk
= current
;
166 struct ring_buffer_event
*event
= NULL
;
167 struct blk_io_trace
*t
;
169 unsigned long *sequence
;
173 if (unlikely(bt
->trace_state
!= Blktrace_running
||
174 !blk_tracer_enabled
))
177 what
|= ddir_act
[rw
& WRITE
];
178 what
|= MASK_TC_BIT(rw
, BARRIER
);
179 what
|= MASK_TC_BIT(rw
, SYNC
);
180 what
|= MASK_TC_BIT(rw
, AHEAD
);
181 what
|= MASK_TC_BIT(rw
, META
);
182 what
|= MASK_TC_BIT(rw
, DISCARD
);
185 if (unlikely(act_log_check(bt
, what
, sector
, pid
)))
187 cpu
= raw_smp_processor_id();
190 struct trace_entry
*ent
;
191 tracing_record_cmdline(current
);
193 event
= ring_buffer_lock_reserve(blk_tr
->buffer
,
194 sizeof(*t
) + pdu_len
, &flags
);
198 ent
= ring_buffer_event_data(event
);
199 t
= (struct blk_io_trace
*)ent
;
200 pc
= preempt_count();
201 tracing_generic_entry_update(ent
, 0, pc
);
202 ent
->type
= TRACE_BLK
;
207 * A word about the locking here - we disable interrupts to reserve
208 * some space in the relay per-cpu buffer, to prevent an irq
209 * from coming in and stepping on our toes.
211 local_irq_save(flags
);
213 if (unlikely(tsk
->btrace_seq
!= blktrace_seq
))
214 trace_note_tsk(bt
, tsk
);
216 t
= relay_reserve(bt
->rchan
, sizeof(*t
) + pdu_len
);
218 sequence
= per_cpu_ptr(bt
->sequence
, cpu
);
220 t
->magic
= BLK_IO_TRACE_MAGIC
| BLK_IO_TRACE_VERSION
;
221 t
->sequence
= ++(*sequence
);
222 t
->time
= ktime_to_ns(ktime_get());
225 * These two are not needed in ftrace as they are in the
226 * generic trace_entry, filled by tracing_generic_entry_update,
227 * but for the trace_event->bin() synthesizer benefit we do it
238 t
->pdu_len
= pdu_len
;
241 memcpy((void *) t
+ sizeof(*t
), pdu_data
, pdu_len
);
244 ring_buffer_unlock_commit(blk_tr
->buffer
, event
, flags
);
246 !(blk_tracer_flags
.val
& TRACE_BLK_OPT_CLASSIC
) &&
247 (trace_flags
& TRACE_ITER_STACKTRACE
) != 0)
248 __trace_stack(blk_tr
, flags
, 5, pc
);
254 local_irq_restore(flags
);
257 static struct dentry
*blk_tree_root
;
258 static DEFINE_MUTEX(blk_tree_mutex
);
260 static void blk_trace_cleanup(struct blk_trace
*bt
)
262 debugfs_remove(bt
->msg_file
);
263 debugfs_remove(bt
->dropped_file
);
264 relay_close(bt
->rchan
);
265 free_percpu(bt
->sequence
);
266 free_percpu(bt
->msg_data
);
268 mutex_lock(&blk_probe_mutex
);
269 if (atomic_dec_and_test(&blk_probes_ref
))
270 blk_unregister_tracepoints();
271 mutex_unlock(&blk_probe_mutex
);
274 int blk_trace_remove(struct request_queue
*q
)
276 struct blk_trace
*bt
;
278 bt
= xchg(&q
->blk_trace
, NULL
);
282 if (bt
->trace_state
== Blktrace_setup
||
283 bt
->trace_state
== Blktrace_stopped
)
284 blk_trace_cleanup(bt
);
288 EXPORT_SYMBOL_GPL(blk_trace_remove
);
290 static int blk_dropped_open(struct inode
*inode
, struct file
*filp
)
292 filp
->private_data
= inode
->i_private
;
297 static ssize_t
blk_dropped_read(struct file
*filp
, char __user
*buffer
,
298 size_t count
, loff_t
*ppos
)
300 struct blk_trace
*bt
= filp
->private_data
;
303 snprintf(buf
, sizeof(buf
), "%u\n", atomic_read(&bt
->dropped
));
305 return simple_read_from_buffer(buffer
, count
, ppos
, buf
, strlen(buf
));
308 static const struct file_operations blk_dropped_fops
= {
309 .owner
= THIS_MODULE
,
310 .open
= blk_dropped_open
,
311 .read
= blk_dropped_read
,
314 static int blk_msg_open(struct inode
*inode
, struct file
*filp
)
316 filp
->private_data
= inode
->i_private
;
321 static ssize_t
blk_msg_write(struct file
*filp
, const char __user
*buffer
,
322 size_t count
, loff_t
*ppos
)
325 struct blk_trace
*bt
;
327 if (count
> BLK_TN_MAX_MSG
)
330 msg
= kmalloc(count
, GFP_KERNEL
);
334 if (copy_from_user(msg
, buffer
, count
)) {
339 bt
= filp
->private_data
;
340 __trace_note_message(bt
, "%s", msg
);
346 static const struct file_operations blk_msg_fops
= {
347 .owner
= THIS_MODULE
,
348 .open
= blk_msg_open
,
349 .write
= blk_msg_write
,
353 * Keep track of how many times we encountered a full subbuffer, to aid
354 * the user space app in telling how many lost events there were.
356 static int blk_subbuf_start_callback(struct rchan_buf
*buf
, void *subbuf
,
357 void *prev_subbuf
, size_t prev_padding
)
359 struct blk_trace
*bt
;
361 if (!relay_buf_full(buf
))
364 bt
= buf
->chan
->private_data
;
365 atomic_inc(&bt
->dropped
);
369 static int blk_remove_buf_file_callback(struct dentry
*dentry
)
371 struct dentry
*parent
= dentry
->d_parent
;
372 debugfs_remove(dentry
);
375 * this will fail for all but the last file, but that is ok. what we
376 * care about is the top level buts->name directory going away, when
377 * the last trace file is gone. Then we don't have to rmdir() that
378 * manually on trace stop, so it nicely solves the issue with
379 * force killing of running traces.
382 debugfs_remove(parent
);
386 static struct dentry
*blk_create_buf_file_callback(const char *filename
,
387 struct dentry
*parent
,
389 struct rchan_buf
*buf
,
392 return debugfs_create_file(filename
, mode
, parent
, buf
,
393 &relay_file_operations
);
396 static struct rchan_callbacks blk_relay_callbacks
= {
397 .subbuf_start
= blk_subbuf_start_callback
,
398 .create_buf_file
= blk_create_buf_file_callback
,
399 .remove_buf_file
= blk_remove_buf_file_callback
,
403 * Setup everything required to start tracing
405 int do_blk_trace_setup(struct request_queue
*q
, char *name
, dev_t dev
,
406 struct blk_user_trace_setup
*buts
)
408 struct blk_trace
*old_bt
, *bt
= NULL
;
409 struct dentry
*dir
= NULL
;
412 if (!buts
->buf_size
|| !buts
->buf_nr
)
415 strncpy(buts
->name
, name
, BLKTRACE_BDEV_SIZE
);
416 buts
->name
[BLKTRACE_BDEV_SIZE
- 1] = '\0';
419 * some device names have larger paths - convert the slashes
420 * to underscores for this to work as expected
422 for (i
= 0; i
< strlen(buts
->name
); i
++)
423 if (buts
->name
[i
] == '/')
427 bt
= kzalloc(sizeof(*bt
), GFP_KERNEL
);
431 bt
->sequence
= alloc_percpu(unsigned long);
435 bt
->msg_data
= __alloc_percpu(BLK_TN_MAX_MSG
);
441 if (!blk_tree_root
) {
442 blk_tree_root
= debugfs_create_dir("block", NULL
);
447 dir
= debugfs_create_dir(buts
->name
, blk_tree_root
);
454 atomic_set(&bt
->dropped
, 0);
457 bt
->dropped_file
= debugfs_create_file("dropped", 0444, dir
, bt
,
459 if (!bt
->dropped_file
)
462 bt
->msg_file
= debugfs_create_file("msg", 0222, dir
, bt
, &blk_msg_fops
);
466 bt
->rchan
= relay_open("trace", dir
, buts
->buf_size
,
467 buts
->buf_nr
, &blk_relay_callbacks
, bt
);
471 bt
->act_mask
= buts
->act_mask
;
473 bt
->act_mask
= (u16
) -1;
475 bt
->start_lba
= buts
->start_lba
;
476 bt
->end_lba
= buts
->end_lba
;
481 bt
->trace_state
= Blktrace_setup
;
483 mutex_lock(&blk_probe_mutex
);
484 if (atomic_add_return(1, &blk_probes_ref
) == 1) {
485 ret
= blk_register_tracepoints();
489 mutex_unlock(&blk_probe_mutex
);
492 old_bt
= xchg(&q
->blk_trace
, bt
);
494 (void) xchg(&q
->blk_trace
, old_bt
);
500 atomic_dec(&blk_probes_ref
);
501 mutex_unlock(&blk_probe_mutex
);
505 debugfs_remove(bt
->msg_file
);
506 if (bt
->dropped_file
)
507 debugfs_remove(bt
->dropped_file
);
508 free_percpu(bt
->sequence
);
509 free_percpu(bt
->msg_data
);
511 relay_close(bt
->rchan
);
517 int blk_trace_setup(struct request_queue
*q
, char *name
, dev_t dev
,
520 struct blk_user_trace_setup buts
;
523 ret
= copy_from_user(&buts
, arg
, sizeof(buts
));
527 ret
= do_blk_trace_setup(q
, name
, dev
, &buts
);
531 if (copy_to_user(arg
, &buts
, sizeof(buts
)))
536 EXPORT_SYMBOL_GPL(blk_trace_setup
);
538 int blk_trace_startstop(struct request_queue
*q
, int start
)
541 struct blk_trace
*bt
= q
->blk_trace
;
547 * For starting a trace, we can transition from a setup or stopped
548 * trace. For stopping a trace, the state must be running
552 if (bt
->trace_state
== Blktrace_setup
||
553 bt
->trace_state
== Blktrace_stopped
) {
556 bt
->trace_state
= Blktrace_running
;
562 if (bt
->trace_state
== Blktrace_running
) {
563 bt
->trace_state
= Blktrace_stopped
;
564 relay_flush(bt
->rchan
);
571 EXPORT_SYMBOL_GPL(blk_trace_startstop
);
574 * blk_trace_ioctl: - handle the ioctls associated with tracing
575 * @bdev: the block device
576 * @cmd: the ioctl cmd
577 * @arg: the argument data, if any
580 int blk_trace_ioctl(struct block_device
*bdev
, unsigned cmd
, char __user
*arg
)
582 struct request_queue
*q
;
584 char b
[BDEVNAME_SIZE
];
586 q
= bdev_get_queue(bdev
);
590 mutex_lock(&bdev
->bd_mutex
);
595 ret
= blk_trace_setup(q
, b
, bdev
->bd_dev
, arg
);
600 ret
= blk_trace_startstop(q
, start
);
602 case BLKTRACETEARDOWN
:
603 ret
= blk_trace_remove(q
);
610 mutex_unlock(&bdev
->bd_mutex
);
615 * blk_trace_shutdown: - stop and cleanup trace structures
616 * @q: the request queue associated with the device
619 void blk_trace_shutdown(struct request_queue
*q
)
622 blk_trace_startstop(q
, 0);
632 * blk_add_trace_rq - Add a trace for a request oriented action
633 * @q: queue the io is for
634 * @rq: the source request
638 * Records an action against a request. Will log the bio offset + size.
641 static void blk_add_trace_rq(struct request_queue
*q
, struct request
*rq
,
644 struct blk_trace
*bt
= q
->blk_trace
;
645 int rw
= rq
->cmd_flags
& 0x03;
650 if (blk_discard_rq(rq
))
651 rw
|= (1 << BIO_RW_DISCARD
);
653 if (blk_pc_request(rq
)) {
654 what
|= BLK_TC_ACT(BLK_TC_PC
);
655 __blk_add_trace(bt
, 0, rq
->data_len
, rw
, what
, rq
->errors
,
656 sizeof(rq
->cmd
), rq
->cmd
);
658 what
|= BLK_TC_ACT(BLK_TC_FS
);
659 __blk_add_trace(bt
, rq
->hard_sector
, rq
->hard_nr_sectors
<< 9,
660 rw
, what
, rq
->errors
, 0, NULL
);
664 static void blk_add_trace_rq_abort(struct request_queue
*q
, struct request
*rq
)
666 blk_add_trace_rq(q
, rq
, BLK_TA_ABORT
);
669 static void blk_add_trace_rq_insert(struct request_queue
*q
, struct request
*rq
)
671 blk_add_trace_rq(q
, rq
, BLK_TA_INSERT
);
674 static void blk_add_trace_rq_issue(struct request_queue
*q
, struct request
*rq
)
676 blk_add_trace_rq(q
, rq
, BLK_TA_ISSUE
);
679 static void blk_add_trace_rq_requeue(struct request_queue
*q
,
682 blk_add_trace_rq(q
, rq
, BLK_TA_REQUEUE
);
685 static void blk_add_trace_rq_complete(struct request_queue
*q
,
688 blk_add_trace_rq(q
, rq
, BLK_TA_COMPLETE
);
692 * blk_add_trace_bio - Add a trace for a bio oriented action
693 * @q: queue the io is for
694 * @bio: the source bio
698 * Records an action against a bio. Will log the bio offset + size.
701 static void blk_add_trace_bio(struct request_queue
*q
, struct bio
*bio
,
704 struct blk_trace
*bt
= q
->blk_trace
;
709 __blk_add_trace(bt
, bio
->bi_sector
, bio
->bi_size
, bio
->bi_rw
, what
,
710 !bio_flagged(bio
, BIO_UPTODATE
), 0, NULL
);
713 static void blk_add_trace_bio_bounce(struct request_queue
*q
, struct bio
*bio
)
715 blk_add_trace_bio(q
, bio
, BLK_TA_BOUNCE
);
718 static void blk_add_trace_bio_complete(struct request_queue
*q
, struct bio
*bio
)
720 blk_add_trace_bio(q
, bio
, BLK_TA_COMPLETE
);
723 static void blk_add_trace_bio_backmerge(struct request_queue
*q
,
726 blk_add_trace_bio(q
, bio
, BLK_TA_BACKMERGE
);
729 static void blk_add_trace_bio_frontmerge(struct request_queue
*q
,
732 blk_add_trace_bio(q
, bio
, BLK_TA_FRONTMERGE
);
735 static void blk_add_trace_bio_queue(struct request_queue
*q
, struct bio
*bio
)
737 blk_add_trace_bio(q
, bio
, BLK_TA_QUEUE
);
740 static void blk_add_trace_getrq(struct request_queue
*q
,
741 struct bio
*bio
, int rw
)
744 blk_add_trace_bio(q
, bio
, BLK_TA_GETRQ
);
746 struct blk_trace
*bt
= q
->blk_trace
;
749 __blk_add_trace(bt
, 0, 0, rw
, BLK_TA_GETRQ
, 0, 0, NULL
);
754 static void blk_add_trace_sleeprq(struct request_queue
*q
,
755 struct bio
*bio
, int rw
)
758 blk_add_trace_bio(q
, bio
, BLK_TA_SLEEPRQ
);
760 struct blk_trace
*bt
= q
->blk_trace
;
763 __blk_add_trace(bt
, 0, 0, rw
, BLK_TA_SLEEPRQ
,
768 static void blk_add_trace_plug(struct request_queue
*q
)
770 struct blk_trace
*bt
= q
->blk_trace
;
773 __blk_add_trace(bt
, 0, 0, 0, BLK_TA_PLUG
, 0, 0, NULL
);
776 static void blk_add_trace_unplug_io(struct request_queue
*q
)
778 struct blk_trace
*bt
= q
->blk_trace
;
781 unsigned int pdu
= q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
];
782 __be64 rpdu
= cpu_to_be64(pdu
);
784 __blk_add_trace(bt
, 0, 0, 0, BLK_TA_UNPLUG_IO
, 0,
785 sizeof(rpdu
), &rpdu
);
789 static void blk_add_trace_unplug_timer(struct request_queue
*q
)
791 struct blk_trace
*bt
= q
->blk_trace
;
794 unsigned int pdu
= q
->rq
.count
[READ
] + q
->rq
.count
[WRITE
];
795 __be64 rpdu
= cpu_to_be64(pdu
);
797 __blk_add_trace(bt
, 0, 0, 0, BLK_TA_UNPLUG_TIMER
, 0,
798 sizeof(rpdu
), &rpdu
);
802 static void blk_add_trace_split(struct request_queue
*q
, struct bio
*bio
,
805 struct blk_trace
*bt
= q
->blk_trace
;
808 __be64 rpdu
= cpu_to_be64(pdu
);
810 __blk_add_trace(bt
, bio
->bi_sector
, bio
->bi_size
, bio
->bi_rw
,
811 BLK_TA_SPLIT
, !bio_flagged(bio
, BIO_UPTODATE
),
812 sizeof(rpdu
), &rpdu
);
817 * blk_add_trace_remap - Add a trace for a remap operation
818 * @q: queue the io is for
819 * @bio: the source bio
820 * @dev: target device
821 * @from: source sector
825 * Device mapper or raid target sometimes need to split a bio because
826 * it spans a stripe (or similar). Add a trace for that action.
829 static void blk_add_trace_remap(struct request_queue
*q
, struct bio
*bio
,
830 dev_t dev
, sector_t from
, sector_t to
)
832 struct blk_trace
*bt
= q
->blk_trace
;
833 struct blk_io_trace_remap r
;
838 r
.device
= cpu_to_be32(dev
);
839 r
.device_from
= cpu_to_be32(bio
->bi_bdev
->bd_dev
);
840 r
.sector
= cpu_to_be64(to
);
842 __blk_add_trace(bt
, from
, bio
->bi_size
, bio
->bi_rw
, BLK_TA_REMAP
,
843 !bio_flagged(bio
, BIO_UPTODATE
), sizeof(r
), &r
);
847 * blk_add_driver_data - Add binary message with driver-specific data
848 * @q: queue the io is for
850 * @data: driver-specific data
851 * @len: length of driver-specific data
854 * Some drivers might want to write driver-specific data per request.
857 void blk_add_driver_data(struct request_queue
*q
,
859 void *data
, size_t len
)
861 struct blk_trace
*bt
= q
->blk_trace
;
866 if (blk_pc_request(rq
))
867 __blk_add_trace(bt
, 0, rq
->data_len
, 0, BLK_TA_DRV_DATA
,
868 rq
->errors
, len
, data
);
870 __blk_add_trace(bt
, rq
->hard_sector
, rq
->hard_nr_sectors
<< 9,
871 0, BLK_TA_DRV_DATA
, rq
->errors
, len
, data
);
873 EXPORT_SYMBOL_GPL(blk_add_driver_data
);
875 static int blk_register_tracepoints(void)
879 ret
= register_trace_block_rq_abort(blk_add_trace_rq_abort
);
881 ret
= register_trace_block_rq_insert(blk_add_trace_rq_insert
);
883 ret
= register_trace_block_rq_issue(blk_add_trace_rq_issue
);
885 ret
= register_trace_block_rq_requeue(blk_add_trace_rq_requeue
);
887 ret
= register_trace_block_rq_complete(blk_add_trace_rq_complete
);
889 ret
= register_trace_block_bio_bounce(blk_add_trace_bio_bounce
);
891 ret
= register_trace_block_bio_complete(blk_add_trace_bio_complete
);
893 ret
= register_trace_block_bio_backmerge(blk_add_trace_bio_backmerge
);
895 ret
= register_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge
);
897 ret
= register_trace_block_bio_queue(blk_add_trace_bio_queue
);
899 ret
= register_trace_block_getrq(blk_add_trace_getrq
);
901 ret
= register_trace_block_sleeprq(blk_add_trace_sleeprq
);
903 ret
= register_trace_block_plug(blk_add_trace_plug
);
905 ret
= register_trace_block_unplug_timer(blk_add_trace_unplug_timer
);
907 ret
= register_trace_block_unplug_io(blk_add_trace_unplug_io
);
909 ret
= register_trace_block_split(blk_add_trace_split
);
911 ret
= register_trace_block_remap(blk_add_trace_remap
);
916 static void blk_unregister_tracepoints(void)
918 unregister_trace_block_remap(blk_add_trace_remap
);
919 unregister_trace_block_split(blk_add_trace_split
);
920 unregister_trace_block_unplug_io(blk_add_trace_unplug_io
);
921 unregister_trace_block_unplug_timer(blk_add_trace_unplug_timer
);
922 unregister_trace_block_plug(blk_add_trace_plug
);
923 unregister_trace_block_sleeprq(blk_add_trace_sleeprq
);
924 unregister_trace_block_getrq(blk_add_trace_getrq
);
925 unregister_trace_block_bio_queue(blk_add_trace_bio_queue
);
926 unregister_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge
);
927 unregister_trace_block_bio_backmerge(blk_add_trace_bio_backmerge
);
928 unregister_trace_block_bio_complete(blk_add_trace_bio_complete
);
929 unregister_trace_block_bio_bounce(blk_add_trace_bio_bounce
);
930 unregister_trace_block_rq_complete(blk_add_trace_rq_complete
);
931 unregister_trace_block_rq_requeue(blk_add_trace_rq_requeue
);
932 unregister_trace_block_rq_issue(blk_add_trace_rq_issue
);
933 unregister_trace_block_rq_insert(blk_add_trace_rq_insert
);
934 unregister_trace_block_rq_abort(blk_add_trace_rq_abort
);
936 tracepoint_synchronize_unregister();
940 * struct blk_io_tracer formatting routines
943 static void fill_rwbs(char *rwbs
, const struct blk_io_trace
*t
)
947 if (t
->action
& BLK_TC_DISCARD
)
949 else if (t
->action
& BLK_TC_WRITE
)
956 if (t
->action
& BLK_TC_AHEAD
)
958 if (t
->action
& BLK_TC_BARRIER
)
960 if (t
->action
& BLK_TC_SYNC
)
962 if (t
->action
& BLK_TC_META
)
969 const struct blk_io_trace
*te_blk_io_trace(const struct trace_entry
*ent
)
971 return (const struct blk_io_trace
*)ent
;
974 static inline const void *pdu_start(const struct trace_entry
*ent
)
976 return te_blk_io_trace(ent
) + 1;
979 static inline u32
t_sec(const struct trace_entry
*ent
)
981 return te_blk_io_trace(ent
)->bytes
>> 9;
984 static inline unsigned long long t_sector(const struct trace_entry
*ent
)
986 return te_blk_io_trace(ent
)->sector
;
989 static inline __u16
t_error(const struct trace_entry
*ent
)
991 return te_blk_io_trace(ent
)->sector
;
994 static __u64
get_pdu_int(const struct trace_entry
*ent
)
996 const __u64
*val
= pdu_start(ent
);
997 return be64_to_cpu(*val
);
1000 static void get_pdu_remap(const struct trace_entry
*ent
,
1001 struct blk_io_trace_remap
*r
)
1003 const struct blk_io_trace_remap
*__r
= pdu_start(ent
);
1004 __u64 sector
= __r
->sector
;
1006 r
->device
= be32_to_cpu(__r
->device
);
1007 r
->device_from
= be32_to_cpu(__r
->device_from
);
1008 r
->sector
= be64_to_cpu(sector
);
1011 static int blk_log_action_iter(struct trace_iterator
*iter
, const char *act
)
1014 unsigned long long ts
= ns2usecs(iter
->ts
);
1015 unsigned long usec_rem
= do_div(ts
, USEC_PER_SEC
);
1016 unsigned secs
= (unsigned long)ts
;
1017 const struct trace_entry
*ent
= iter
->ent
;
1018 const struct blk_io_trace
*t
= (const struct blk_io_trace
*)ent
;
1022 return trace_seq_printf(&iter
->seq
,
1023 "%3d,%-3d %2d %5d.%06lu %5u %2s %3s ",
1024 MAJOR(t
->device
), MINOR(t
->device
), iter
->cpu
,
1025 secs
, usec_rem
, ent
->pid
, act
, rwbs
);
1028 static int blk_log_action_seq(struct trace_seq
*s
, const struct blk_io_trace
*t
,
1033 return trace_seq_printf(s
, "%3d,%-3d %2s %3s ",
1034 MAJOR(t
->device
), MINOR(t
->device
), act
, rwbs
);
1037 static int blk_log_generic(struct trace_seq
*s
, const struct trace_entry
*ent
)
1039 const char *cmd
= trace_find_cmdline(ent
->pid
);
1042 return trace_seq_printf(s
, "%llu + %u [%s]\n",
1043 t_sector(ent
), t_sec(ent
), cmd
);
1044 return trace_seq_printf(s
, "[%s]\n", cmd
);
1047 static int blk_log_with_error(struct trace_seq
*s
,
1048 const struct trace_entry
*ent
)
1051 return trace_seq_printf(s
, "%llu + %u [%d]\n", t_sector(ent
),
1052 t_sec(ent
), t_error(ent
));
1053 return trace_seq_printf(s
, "%llu [%d]\n", t_sector(ent
), t_error(ent
));
1056 static int blk_log_remap(struct trace_seq
*s
, const struct trace_entry
*ent
)
1058 struct blk_io_trace_remap r
= { .device
= 0, };
1060 get_pdu_remap(ent
, &r
);
1061 return trace_seq_printf(s
, "%llu + %u <- (%d,%d) %llu\n",
1063 t_sec(ent
), MAJOR(r
.device
), MINOR(r
.device
),
1064 (unsigned long long)r
.sector
);
1067 static int blk_log_plug(struct trace_seq
*s
, const struct trace_entry
*ent
)
1069 return trace_seq_printf(s
, "[%s]\n", trace_find_cmdline(ent
->pid
));
1072 static int blk_log_unplug(struct trace_seq
*s
, const struct trace_entry
*ent
)
1074 return trace_seq_printf(s
, "[%s] %llu\n", trace_find_cmdline(ent
->pid
),
1078 static int blk_log_split(struct trace_seq
*s
, const struct trace_entry
*ent
)
1080 return trace_seq_printf(s
, "%llu / %llu [%s]\n", t_sector(ent
),
1081 get_pdu_int(ent
), trace_find_cmdline(ent
->pid
));
1085 * struct tracer operations
1088 static void blk_tracer_print_header(struct seq_file
*m
)
1090 if (!(blk_tracer_flags
.val
& TRACE_BLK_OPT_CLASSIC
))
1092 seq_puts(m
, "# DEV CPU TIMESTAMP PID ACT FLG\n"
1096 static void blk_tracer_start(struct trace_array
*tr
)
1098 tracing_reset_online_cpus(tr
);
1100 mutex_lock(&blk_probe_mutex
);
1101 if (atomic_add_return(1, &blk_probes_ref
) == 1)
1102 if (blk_register_tracepoints())
1103 atomic_dec(&blk_probes_ref
);
1104 mutex_unlock(&blk_probe_mutex
);
1105 trace_flags
&= ~TRACE_ITER_CONTEXT_INFO
;
1108 static int blk_tracer_init(struct trace_array
*tr
)
1111 blk_tracer_start(tr
);
1112 mutex_lock(&blk_probe_mutex
);
1113 blk_tracer_enabled
++;
1114 mutex_unlock(&blk_probe_mutex
);
1118 static void blk_tracer_stop(struct trace_array
*tr
)
1120 trace_flags
|= TRACE_ITER_CONTEXT_INFO
;
1121 mutex_lock(&blk_probe_mutex
);
1122 if (atomic_dec_and_test(&blk_probes_ref
))
1123 blk_unregister_tracepoints();
1124 mutex_unlock(&blk_probe_mutex
);
1127 static void blk_tracer_reset(struct trace_array
*tr
)
1129 if (!atomic_read(&blk_probes_ref
))
1132 mutex_lock(&blk_probe_mutex
);
1133 blk_tracer_enabled
--;
1134 WARN_ON(blk_tracer_enabled
< 0);
1135 mutex_unlock(&blk_probe_mutex
);
1137 blk_tracer_stop(tr
);
1142 int (*print
)(struct trace_seq
*s
, const struct trace_entry
*ent
);
1143 } what2act
[] __read_mostly
= {
1144 [__BLK_TA_QUEUE
] = {{ "Q", "queue" }, blk_log_generic
},
1145 [__BLK_TA_BACKMERGE
] = {{ "M", "backmerge" }, blk_log_generic
},
1146 [__BLK_TA_FRONTMERGE
] = {{ "F", "frontmerge" }, blk_log_generic
},
1147 [__BLK_TA_GETRQ
] = {{ "G", "getrq" }, blk_log_generic
},
1148 [__BLK_TA_SLEEPRQ
] = {{ "S", "sleeprq" }, blk_log_generic
},
1149 [__BLK_TA_REQUEUE
] = {{ "R", "requeue" }, blk_log_with_error
},
1150 [__BLK_TA_ISSUE
] = {{ "D", "issue" }, blk_log_generic
},
1151 [__BLK_TA_COMPLETE
] = {{ "C", "complete" }, blk_log_with_error
},
1152 [__BLK_TA_PLUG
] = {{ "P", "plug" }, blk_log_plug
},
1153 [__BLK_TA_UNPLUG_IO
] = {{ "U", "unplug_io" }, blk_log_unplug
},
1154 [__BLK_TA_UNPLUG_TIMER
] = {{ "UT", "unplug_timer" }, blk_log_unplug
},
1155 [__BLK_TA_INSERT
] = {{ "I", "insert" }, blk_log_generic
},
1156 [__BLK_TA_SPLIT
] = {{ "X", "split" }, blk_log_split
},
1157 [__BLK_TA_BOUNCE
] = {{ "B", "bounce" }, blk_log_generic
},
1158 [__BLK_TA_REMAP
] = {{ "A", "remap" }, blk_log_remap
},
1161 static enum print_line_t
blk_trace_event_print(struct trace_iterator
*iter
,
1164 struct trace_seq
*s
= &iter
->seq
;
1165 const struct blk_io_trace
*t
= (struct blk_io_trace
*)iter
->ent
;
1166 const u16 what
= t
->action
& ((1 << BLK_TC_SHIFT
) - 1);
1169 if (!trace_print_context(iter
))
1170 return TRACE_TYPE_PARTIAL_LINE
;
1172 if (unlikely(what
== 0 || what
> ARRAY_SIZE(what2act
)))
1173 ret
= trace_seq_printf(s
, "Bad pc action %x\n", what
);
1175 const bool long_act
= !!(trace_flags
& TRACE_ITER_VERBOSE
);
1176 ret
= blk_log_action_seq(s
, t
, what2act
[what
].act
[long_act
]);
1178 ret
= what2act
[what
].print(s
, iter
->ent
);
1181 return ret
? TRACE_TYPE_HANDLED
: TRACE_TYPE_PARTIAL_LINE
;
1184 static int blk_trace_synthesize_old_trace(struct trace_iterator
*iter
)
1186 struct trace_seq
*s
= &iter
->seq
;
1187 struct blk_io_trace
*t
= (struct blk_io_trace
*)iter
->ent
;
1188 const int offset
= offsetof(struct blk_io_trace
, sector
);
1189 struct blk_io_trace old
= {
1190 .magic
= BLK_IO_TRACE_MAGIC
| BLK_IO_TRACE_VERSION
,
1191 .time
= ns2usecs(iter
->ts
),
1194 if (!trace_seq_putmem(s
, &old
, offset
))
1196 return trace_seq_putmem(s
, &t
->sector
,
1197 sizeof(old
) - offset
+ t
->pdu_len
);
1200 static enum print_line_t
1201 blk_trace_event_print_binary(struct trace_iterator
*iter
, int flags
)
1203 return blk_trace_synthesize_old_trace(iter
) ?
1204 TRACE_TYPE_HANDLED
: TRACE_TYPE_PARTIAL_LINE
;
1207 static enum print_line_t
blk_tracer_print_line(struct trace_iterator
*iter
)
1209 const struct blk_io_trace
*t
;
1213 if (!(blk_tracer_flags
.val
& TRACE_BLK_OPT_CLASSIC
))
1214 return TRACE_TYPE_UNHANDLED
;
1216 t
= (const struct blk_io_trace
*)iter
->ent
;
1217 what
= t
->action
& ((1 << BLK_TC_SHIFT
) - 1);
1219 if (unlikely(what
== 0 || what
> ARRAY_SIZE(what2act
)))
1220 ret
= trace_seq_printf(&iter
->seq
, "Bad pc action %x\n", what
);
1222 const bool long_act
= !!(trace_flags
& TRACE_ITER_VERBOSE
);
1223 ret
= blk_log_action_iter(iter
, what2act
[what
].act
[long_act
]);
1225 ret
= what2act
[what
].print(&iter
->seq
, iter
->ent
);
1228 return ret
? TRACE_TYPE_HANDLED
: TRACE_TYPE_PARTIAL_LINE
;
1231 static struct tracer blk_tracer __read_mostly
= {
1233 .init
= blk_tracer_init
,
1234 .reset
= blk_tracer_reset
,
1235 .start
= blk_tracer_start
,
1236 .stop
= blk_tracer_stop
,
1237 .print_header
= blk_tracer_print_header
,
1238 .print_line
= blk_tracer_print_line
,
1239 .flags
= &blk_tracer_flags
,
1242 static struct trace_event trace_blk_event
= {
1244 .trace
= blk_trace_event_print
,
1245 .latency_trace
= blk_trace_event_print
,
1246 .binary
= blk_trace_event_print_binary
,
1249 static int __init
init_blk_tracer(void)
1251 if (!register_ftrace_event(&trace_blk_event
)) {
1252 pr_warning("Warning: could not register block events\n");
1256 if (register_tracer(&blk_tracer
) != 0) {
1257 pr_warning("Warning: could not register the block tracer\n");
1258 unregister_ftrace_event(&trace_blk_event
);
1265 device_initcall(init_blk_tracer
);
1267 static int blk_trace_remove_queue(struct request_queue
*q
)
1269 struct blk_trace
*bt
;
1271 bt
= xchg(&q
->blk_trace
, NULL
);
1280 * Setup everything required to start tracing
1282 static int blk_trace_setup_queue(struct request_queue
*q
, dev_t dev
)
1284 struct blk_trace
*old_bt
, *bt
= NULL
;
1288 bt
= kzalloc(sizeof(*bt
), GFP_KERNEL
);
1293 bt
->act_mask
= (u16
)-1;
1294 bt
->end_lba
= -1ULL;
1295 bt
->trace_state
= Blktrace_running
;
1297 old_bt
= xchg(&q
->blk_trace
, bt
);
1298 if (old_bt
!= NULL
) {
1299 (void)xchg(&q
->blk_trace
, old_bt
);
1309 * sysfs interface to enable and configure tracing
1312 static ssize_t
sysfs_blk_trace_enable_show(struct device
*dev
,
1313 struct device_attribute
*attr
,
1316 struct hd_struct
*p
= dev_to_part(dev
);
1317 struct block_device
*bdev
;
1318 ssize_t ret
= -ENXIO
;
1321 bdev
= bdget(part_devt(p
));
1323 struct request_queue
*q
= bdev_get_queue(bdev
);
1326 mutex_lock(&bdev
->bd_mutex
);
1327 ret
= sprintf(buf
, "%u\n", !!q
->blk_trace
);
1328 mutex_unlock(&bdev
->bd_mutex
);
1338 static ssize_t
sysfs_blk_trace_enable_store(struct device
*dev
,
1339 struct device_attribute
*attr
,
1340 const char *buf
, size_t count
)
1342 struct block_device
*bdev
;
1343 struct request_queue
*q
;
1344 struct hd_struct
*p
;
1346 ssize_t ret
= -ENXIO
;
1348 if (count
== 0 || sscanf(buf
, "%d", &value
) != 1)
1352 p
= dev_to_part(dev
);
1353 bdev
= bdget(part_devt(p
));
1355 goto out_unlock_kernel
;
1357 q
= bdev_get_queue(bdev
);
1361 mutex_lock(&bdev
->bd_mutex
);
1363 ret
= blk_trace_setup_queue(q
, bdev
->bd_dev
);
1365 ret
= blk_trace_remove_queue(q
);
1366 mutex_unlock(&bdev
->bd_mutex
);
1378 static ssize_t
sysfs_blk_trace_attr_show(struct device
*dev
,
1379 struct device_attribute
*attr
,
1381 static ssize_t
sysfs_blk_trace_attr_store(struct device
*dev
,
1382 struct device_attribute
*attr
,
1383 const char *buf
, size_t count
);
1384 #define BLK_TRACE_DEVICE_ATTR(_name) \
1385 DEVICE_ATTR(_name, S_IRUGO | S_IWUSR, \
1386 sysfs_blk_trace_attr_show, \
1387 sysfs_blk_trace_attr_store)
1389 static DEVICE_ATTR(enable
, S_IRUGO
| S_IWUSR
,
1390 sysfs_blk_trace_enable_show
, sysfs_blk_trace_enable_store
);
1391 static BLK_TRACE_DEVICE_ATTR(act_mask
);
1392 static BLK_TRACE_DEVICE_ATTR(pid
);
1393 static BLK_TRACE_DEVICE_ATTR(start_lba
);
1394 static BLK_TRACE_DEVICE_ATTR(end_lba
);
1396 static struct attribute
*blk_trace_attrs
[] = {
1397 &dev_attr_enable
.attr
,
1398 &dev_attr_act_mask
.attr
,
1400 &dev_attr_start_lba
.attr
,
1401 &dev_attr_end_lba
.attr
,
1405 struct attribute_group blk_trace_attr_group
= {
1407 .attrs
= blk_trace_attrs
,
1410 static int blk_str2act_mask(const char *str
)
1413 char *copy
= kstrdup(str
, GFP_KERNEL
), *s
;
1421 char *sep
= strchr(s
, ',');
1426 if (strcasecmp(s
, "barrier") == 0)
1427 mask
|= BLK_TC_BARRIER
;
1428 else if (strcasecmp(s
, "complete") == 0)
1429 mask
|= BLK_TC_COMPLETE
;
1430 else if (strcasecmp(s
, "fs") == 0)
1432 else if (strcasecmp(s
, "issue") == 0)
1433 mask
|= BLK_TC_ISSUE
;
1434 else if (strcasecmp(s
, "pc") == 0)
1436 else if (strcasecmp(s
, "queue") == 0)
1437 mask
|= BLK_TC_QUEUE
;
1438 else if (strcasecmp(s
, "read") == 0)
1439 mask
|= BLK_TC_READ
;
1440 else if (strcasecmp(s
, "requeue") == 0)
1441 mask
|= BLK_TC_REQUEUE
;
1442 else if (strcasecmp(s
, "sync") == 0)
1443 mask
|= BLK_TC_SYNC
;
1444 else if (strcasecmp(s
, "write") == 0)
1445 mask
|= BLK_TC_WRITE
;
1457 static ssize_t
sysfs_blk_trace_attr_show(struct device
*dev
,
1458 struct device_attribute
*attr
,
1461 struct hd_struct
*p
= dev_to_part(dev
);
1462 struct request_queue
*q
;
1463 struct block_device
*bdev
;
1464 ssize_t ret
= -ENXIO
;
1467 bdev
= bdget(part_devt(p
));
1469 goto out_unlock_kernel
;
1471 q
= bdev_get_queue(bdev
);
1474 mutex_lock(&bdev
->bd_mutex
);
1475 if (q
->blk_trace
== NULL
)
1476 ret
= sprintf(buf
, "disabled\n");
1477 else if (attr
== &dev_attr_act_mask
)
1478 ret
= sprintf(buf
, "%#x\n", q
->blk_trace
->act_mask
);
1479 else if (attr
== &dev_attr_pid
)
1480 ret
= sprintf(buf
, "%u\n", q
->blk_trace
->pid
);
1481 else if (attr
== &dev_attr_start_lba
)
1482 ret
= sprintf(buf
, "%llu\n", q
->blk_trace
->start_lba
);
1483 else if (attr
== &dev_attr_end_lba
)
1484 ret
= sprintf(buf
, "%llu\n", q
->blk_trace
->end_lba
);
1485 mutex_unlock(&bdev
->bd_mutex
);
1493 static ssize_t
sysfs_blk_trace_attr_store(struct device
*dev
,
1494 struct device_attribute
*attr
,
1495 const char *buf
, size_t count
)
1497 struct block_device
*bdev
;
1498 struct request_queue
*q
;
1499 struct hd_struct
*p
;
1501 ssize_t ret
= -ENXIO
;
1506 if (attr
== &dev_attr_act_mask
) {
1507 if (sscanf(buf
, "%llx", &value
) != 1) {
1508 /* Assume it is a list of trace category names */
1509 value
= blk_str2act_mask(buf
);
1513 } else if (sscanf(buf
, "%llu", &value
) != 1)
1517 p
= dev_to_part(dev
);
1518 bdev
= bdget(part_devt(p
));
1520 goto out_unlock_kernel
;
1522 q
= bdev_get_queue(bdev
);
1526 mutex_lock(&bdev
->bd_mutex
);
1528 if (q
->blk_trace
== NULL
)
1529 ret
= blk_trace_setup_queue(q
, bdev
->bd_dev
);
1532 if (attr
== &dev_attr_act_mask
)
1533 q
->blk_trace
->act_mask
= value
;
1534 else if (attr
== &dev_attr_pid
)
1535 q
->blk_trace
->pid
= value
;
1536 else if (attr
== &dev_attr_start_lba
)
1537 q
->blk_trace
->start_lba
= value
;
1538 else if (attr
== &dev_attr_end_lba
)
1539 q
->blk_trace
->end_lba
= value
;
1542 mutex_unlock(&bdev
->bd_mutex
);