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/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/export.h>
27 #include <linux/time.h>
28 #include <linux/uaccess.h>
29 #include <linux/list.h>
30 #include <linux/blk-cgroup.h>
32 #include "../../block/blk.h"
34 #include <trace/events/block.h>
36 #include "trace_output.h"
38 #ifdef CONFIG_BLK_DEV_IO_TRACE
40 static unsigned int blktrace_seq __read_mostly
= 1;
42 static struct trace_array
*blk_tr
;
43 static bool blk_tracer_enabled __read_mostly
;
45 static LIST_HEAD(running_trace_list
);
46 static __cacheline_aligned_in_smp
DEFINE_SPINLOCK(running_trace_lock
);
48 /* Select an alternative, minimalistic output than the original one */
49 #define TRACE_BLK_OPT_CLASSIC 0x1
50 #define TRACE_BLK_OPT_CGROUP 0x2
51 #define TRACE_BLK_OPT_CGNAME 0x4
53 static struct tracer_opt blk_tracer_opts
[] = {
54 /* Default disable the minimalistic output */
55 { TRACER_OPT(blk_classic
, TRACE_BLK_OPT_CLASSIC
) },
56 #ifdef CONFIG_BLK_CGROUP
57 { TRACER_OPT(blk_cgroup
, TRACE_BLK_OPT_CGROUP
) },
58 { TRACER_OPT(blk_cgname
, TRACE_BLK_OPT_CGNAME
) },
63 static struct tracer_flags blk_tracer_flags
= {
65 .opts
= blk_tracer_opts
,
68 /* Global reference count of probes */
69 static atomic_t blk_probes_ref
= ATOMIC_INIT(0);
71 static void blk_register_tracepoints(void);
72 static void blk_unregister_tracepoints(void);
75 * Send out a notify message.
77 static void trace_note(struct blk_trace
*bt
, pid_t pid
, int action
,
78 const void *data
, size_t len
,
79 union kernfs_node_id
*cgid
)
81 struct blk_io_trace
*t
;
82 struct ring_buffer_event
*event
= NULL
;
83 struct ring_buffer
*buffer
= NULL
;
85 int cpu
= smp_processor_id();
86 bool blk_tracer
= blk_tracer_enabled
;
87 ssize_t cgid_len
= cgid
? sizeof(*cgid
) : 0;
90 buffer
= blk_tr
->trace_buffer
.buffer
;
92 event
= trace_buffer_lock_reserve(buffer
, TRACE_BLK
,
93 sizeof(*t
) + len
+ cgid_len
,
97 t
= ring_buffer_event_data(event
);
104 t
= relay_reserve(bt
->rchan
, sizeof(*t
) + len
+ cgid_len
);
106 t
->magic
= BLK_IO_TRACE_MAGIC
| BLK_IO_TRACE_VERSION
;
107 t
->time
= ktime_to_ns(ktime_get());
110 t
->action
= action
| (cgid
? __BLK_TN_CGROUP
: 0);
113 t
->pdu_len
= len
+ cgid_len
;
115 memcpy((void *)t
+ sizeof(*t
), cgid
, cgid_len
);
116 memcpy((void *) t
+ sizeof(*t
) + cgid_len
, data
, len
);
119 trace_buffer_unlock_commit(blk_tr
, buffer
, event
, 0, pc
);
124 * Send out a notify for this process, if we haven't done so since a trace
127 static void trace_note_tsk(struct task_struct
*tsk
)
130 struct blk_trace
*bt
;
132 tsk
->btrace_seq
= blktrace_seq
;
133 spin_lock_irqsave(&running_trace_lock
, flags
);
134 list_for_each_entry(bt
, &running_trace_list
, running_list
) {
135 trace_note(bt
, tsk
->pid
, BLK_TN_PROCESS
, tsk
->comm
,
136 sizeof(tsk
->comm
), NULL
);
138 spin_unlock_irqrestore(&running_trace_lock
, flags
);
141 static void trace_note_time(struct blk_trace
*bt
)
143 struct timespec64 now
;
147 /* need to check user space to see if this breaks in y2038 or y2106 */
148 ktime_get_real_ts64(&now
);
149 words
[0] = (u32
)now
.tv_sec
;
150 words
[1] = now
.tv_nsec
;
152 local_irq_save(flags
);
153 trace_note(bt
, 0, BLK_TN_TIMESTAMP
, words
, sizeof(words
), NULL
);
154 local_irq_restore(flags
);
157 void __trace_note_message(struct blk_trace
*bt
, struct blkcg
*blkcg
,
158 const char *fmt
, ...)
165 if (unlikely(bt
->trace_state
!= Blktrace_running
&&
166 !blk_tracer_enabled
))
170 * If the BLK_TC_NOTIFY action mask isn't set, don't send any note
171 * message to the trace.
173 if (!(bt
->act_mask
& BLK_TC_NOTIFY
))
176 local_irq_save(flags
);
177 buf
= this_cpu_ptr(bt
->msg_data
);
179 n
= vscnprintf(buf
, BLK_TN_MAX_MSG
, fmt
, args
);
182 if (!(blk_tracer_flags
.val
& TRACE_BLK_OPT_CGROUP
))
184 #ifdef CONFIG_BLK_CGROUP
185 trace_note(bt
, 0, BLK_TN_MESSAGE
, buf
, n
,
186 blkcg
? cgroup_get_kernfs_id(blkcg
->css
.cgroup
) : NULL
);
188 trace_note(bt
, 0, BLK_TN_MESSAGE
, buf
, n
, NULL
);
190 local_irq_restore(flags
);
192 EXPORT_SYMBOL_GPL(__trace_note_message
);
194 static int act_log_check(struct blk_trace
*bt
, u32 what
, sector_t sector
,
197 if (((bt
->act_mask
<< BLK_TC_SHIFT
) & what
) == 0)
199 if (sector
&& (sector
< bt
->start_lba
|| sector
> bt
->end_lba
))
201 if (bt
->pid
&& pid
!= bt
->pid
)
208 * Data direction bit lookup
210 static const u32 ddir_act
[2] = { BLK_TC_ACT(BLK_TC_READ
),
211 BLK_TC_ACT(BLK_TC_WRITE
) };
213 #define BLK_TC_RAHEAD BLK_TC_AHEAD
214 #define BLK_TC_PREFLUSH BLK_TC_FLUSH
216 /* The ilog2() calls fall out because they're constant */
217 #define MASK_TC_BIT(rw, __name) ((rw & REQ_ ## __name) << \
218 (ilog2(BLK_TC_ ## __name) + BLK_TC_SHIFT - __REQ_ ## __name))
221 * The worker for the various blk_add_trace*() types. Fills out a
222 * blk_io_trace structure and places it in a per-cpu subbuffer.
224 static void __blk_add_trace(struct blk_trace
*bt
, sector_t sector
, int bytes
,
225 int op
, int op_flags
, u32 what
, int error
, int pdu_len
,
226 void *pdu_data
, union kernfs_node_id
*cgid
)
228 struct task_struct
*tsk
= current
;
229 struct ring_buffer_event
*event
= NULL
;
230 struct ring_buffer
*buffer
= NULL
;
231 struct blk_io_trace
*t
;
232 unsigned long flags
= 0;
233 unsigned long *sequence
;
236 bool blk_tracer
= blk_tracer_enabled
;
237 ssize_t cgid_len
= cgid
? sizeof(*cgid
) : 0;
239 if (unlikely(bt
->trace_state
!= Blktrace_running
&& !blk_tracer
))
242 what
|= ddir_act
[op_is_write(op
) ? WRITE
: READ
];
243 what
|= MASK_TC_BIT(op_flags
, SYNC
);
244 what
|= MASK_TC_BIT(op_flags
, RAHEAD
);
245 what
|= MASK_TC_BIT(op_flags
, META
);
246 what
|= MASK_TC_BIT(op_flags
, PREFLUSH
);
247 what
|= MASK_TC_BIT(op_flags
, FUA
);
248 if (op
== REQ_OP_DISCARD
|| op
== REQ_OP_SECURE_ERASE
)
249 what
|= BLK_TC_ACT(BLK_TC_DISCARD
);
250 if (op
== REQ_OP_FLUSH
)
251 what
|= BLK_TC_ACT(BLK_TC_FLUSH
);
253 what
|= __BLK_TA_CGROUP
;
256 if (act_log_check(bt
, what
, sector
, pid
))
258 cpu
= raw_smp_processor_id();
261 tracing_record_cmdline(current
);
263 buffer
= blk_tr
->trace_buffer
.buffer
;
264 pc
= preempt_count();
265 event
= trace_buffer_lock_reserve(buffer
, TRACE_BLK
,
266 sizeof(*t
) + pdu_len
+ cgid_len
,
270 t
= ring_buffer_event_data(event
);
274 if (unlikely(tsk
->btrace_seq
!= blktrace_seq
))
278 * A word about the locking here - we disable interrupts to reserve
279 * some space in the relay per-cpu buffer, to prevent an irq
280 * from coming in and stepping on our toes.
282 local_irq_save(flags
);
283 t
= relay_reserve(bt
->rchan
, sizeof(*t
) + pdu_len
+ cgid_len
);
285 sequence
= per_cpu_ptr(bt
->sequence
, cpu
);
287 t
->magic
= BLK_IO_TRACE_MAGIC
| BLK_IO_TRACE_VERSION
;
288 t
->sequence
= ++(*sequence
);
289 t
->time
= ktime_to_ns(ktime_get());
292 * These two are not needed in ftrace as they are in the
293 * generic trace_entry, filled by tracing_generic_entry_update,
294 * but for the trace_event->bin() synthesizer benefit we do it
305 t
->pdu_len
= pdu_len
+ cgid_len
;
308 memcpy((void *)t
+ sizeof(*t
), cgid
, cgid_len
);
310 memcpy((void *)t
+ sizeof(*t
) + cgid_len
, pdu_data
, pdu_len
);
313 trace_buffer_unlock_commit(blk_tr
, buffer
, event
, 0, pc
);
318 local_irq_restore(flags
);
321 static void blk_trace_free(struct blk_trace
*bt
)
323 debugfs_remove(bt
->msg_file
);
324 debugfs_remove(bt
->dropped_file
);
325 relay_close(bt
->rchan
);
326 debugfs_remove(bt
->dir
);
327 free_percpu(bt
->sequence
);
328 free_percpu(bt
->msg_data
);
332 static void blk_trace_cleanup(struct blk_trace
*bt
)
335 if (atomic_dec_and_test(&blk_probes_ref
))
336 blk_unregister_tracepoints();
339 int blk_trace_remove(struct request_queue
*q
)
341 struct blk_trace
*bt
;
343 bt
= xchg(&q
->blk_trace
, NULL
);
347 if (bt
->trace_state
!= Blktrace_running
)
348 blk_trace_cleanup(bt
);
352 EXPORT_SYMBOL_GPL(blk_trace_remove
);
354 static ssize_t
blk_dropped_read(struct file
*filp
, char __user
*buffer
,
355 size_t count
, loff_t
*ppos
)
357 struct blk_trace
*bt
= filp
->private_data
;
360 snprintf(buf
, sizeof(buf
), "%u\n", atomic_read(&bt
->dropped
));
362 return simple_read_from_buffer(buffer
, count
, ppos
, buf
, strlen(buf
));
365 static const struct file_operations blk_dropped_fops
= {
366 .owner
= THIS_MODULE
,
368 .read
= blk_dropped_read
,
369 .llseek
= default_llseek
,
372 static ssize_t
blk_msg_write(struct file
*filp
, const char __user
*buffer
,
373 size_t count
, loff_t
*ppos
)
376 struct blk_trace
*bt
;
378 if (count
>= BLK_TN_MAX_MSG
)
381 msg
= memdup_user_nul(buffer
, count
);
385 bt
= filp
->private_data
;
386 __trace_note_message(bt
, NULL
, "%s", msg
);
392 static const struct file_operations blk_msg_fops
= {
393 .owner
= THIS_MODULE
,
395 .write
= blk_msg_write
,
396 .llseek
= noop_llseek
,
400 * Keep track of how many times we encountered a full subbuffer, to aid
401 * the user space app in telling how many lost events there were.
403 static int blk_subbuf_start_callback(struct rchan_buf
*buf
, void *subbuf
,
404 void *prev_subbuf
, size_t prev_padding
)
406 struct blk_trace
*bt
;
408 if (!relay_buf_full(buf
))
411 bt
= buf
->chan
->private_data
;
412 atomic_inc(&bt
->dropped
);
416 static int blk_remove_buf_file_callback(struct dentry
*dentry
)
418 debugfs_remove(dentry
);
423 static struct dentry
*blk_create_buf_file_callback(const char *filename
,
424 struct dentry
*parent
,
426 struct rchan_buf
*buf
,
429 return debugfs_create_file(filename
, mode
, parent
, buf
,
430 &relay_file_operations
);
433 static struct rchan_callbacks blk_relay_callbacks
= {
434 .subbuf_start
= blk_subbuf_start_callback
,
435 .create_buf_file
= blk_create_buf_file_callback
,
436 .remove_buf_file
= blk_remove_buf_file_callback
,
439 static void blk_trace_setup_lba(struct blk_trace
*bt
,
440 struct block_device
*bdev
)
442 struct hd_struct
*part
= NULL
;
445 part
= bdev
->bd_part
;
448 bt
->start_lba
= part
->start_sect
;
449 bt
->end_lba
= part
->start_sect
+ part
->nr_sects
;
457 * Setup everything required to start tracing
459 static int do_blk_trace_setup(struct request_queue
*q
, char *name
, dev_t dev
,
460 struct block_device
*bdev
,
461 struct blk_user_trace_setup
*buts
)
463 struct blk_trace
*bt
= NULL
;
464 struct dentry
*dir
= NULL
;
467 if (!buts
->buf_size
|| !buts
->buf_nr
)
470 strncpy(buts
->name
, name
, BLKTRACE_BDEV_SIZE
);
471 buts
->name
[BLKTRACE_BDEV_SIZE
- 1] = '\0';
474 * some device names have larger paths - convert the slashes
475 * to underscores for this to work as expected
477 strreplace(buts
->name
, '/', '_');
479 bt
= kzalloc(sizeof(*bt
), GFP_KERNEL
);
484 bt
->sequence
= alloc_percpu(unsigned long);
488 bt
->msg_data
= __alloc_percpu(BLK_TN_MAX_MSG
, __alignof__(char));
494 if (!blk_debugfs_root
)
497 dir
= debugfs_lookup(buts
->name
, blk_debugfs_root
);
499 bt
->dir
= dir
= debugfs_create_dir(buts
->name
, blk_debugfs_root
);
504 atomic_set(&bt
->dropped
, 0);
505 INIT_LIST_HEAD(&bt
->running_list
);
508 bt
->dropped_file
= debugfs_create_file("dropped", 0444, dir
, bt
,
510 if (!bt
->dropped_file
)
513 bt
->msg_file
= debugfs_create_file("msg", 0222, dir
, bt
, &blk_msg_fops
);
517 bt
->rchan
= relay_open("trace", dir
, buts
->buf_size
,
518 buts
->buf_nr
, &blk_relay_callbacks
, bt
);
522 bt
->act_mask
= buts
->act_mask
;
524 bt
->act_mask
= (u16
) -1;
526 blk_trace_setup_lba(bt
, bdev
);
528 /* overwrite with user settings */
530 bt
->start_lba
= buts
->start_lba
;
532 bt
->end_lba
= buts
->end_lba
;
535 bt
->trace_state
= Blktrace_setup
;
538 if (cmpxchg(&q
->blk_trace
, NULL
, bt
))
541 if (atomic_inc_return(&blk_probes_ref
) == 1)
542 blk_register_tracepoints();
553 int blk_trace_setup(struct request_queue
*q
, char *name
, dev_t dev
,
554 struct block_device
*bdev
,
557 struct blk_user_trace_setup buts
;
560 ret
= copy_from_user(&buts
, arg
, sizeof(buts
));
564 ret
= do_blk_trace_setup(q
, name
, dev
, bdev
, &buts
);
568 if (copy_to_user(arg
, &buts
, sizeof(buts
))) {
574 EXPORT_SYMBOL_GPL(blk_trace_setup
);
576 #if defined(CONFIG_COMPAT) && defined(CONFIG_X86_64)
577 static int compat_blk_trace_setup(struct request_queue
*q
, char *name
,
578 dev_t dev
, struct block_device
*bdev
,
581 struct blk_user_trace_setup buts
;
582 struct compat_blk_user_trace_setup cbuts
;
585 if (copy_from_user(&cbuts
, arg
, sizeof(cbuts
)))
588 buts
= (struct blk_user_trace_setup
) {
589 .act_mask
= cbuts
.act_mask
,
590 .buf_size
= cbuts
.buf_size
,
591 .buf_nr
= cbuts
.buf_nr
,
592 .start_lba
= cbuts
.start_lba
,
593 .end_lba
= cbuts
.end_lba
,
597 ret
= do_blk_trace_setup(q
, name
, dev
, bdev
, &buts
);
601 if (copy_to_user(arg
, &buts
.name
, ARRAY_SIZE(buts
.name
))) {
610 int blk_trace_startstop(struct request_queue
*q
, int start
)
613 struct blk_trace
*bt
= q
->blk_trace
;
619 * For starting a trace, we can transition from a setup or stopped
620 * trace. For stopping a trace, the state must be running
624 if (bt
->trace_state
== Blktrace_setup
||
625 bt
->trace_state
== Blktrace_stopped
) {
628 bt
->trace_state
= Blktrace_running
;
629 spin_lock_irq(&running_trace_lock
);
630 list_add(&bt
->running_list
, &running_trace_list
);
631 spin_unlock_irq(&running_trace_lock
);
637 if (bt
->trace_state
== Blktrace_running
) {
638 bt
->trace_state
= Blktrace_stopped
;
639 spin_lock_irq(&running_trace_lock
);
640 list_del_init(&bt
->running_list
);
641 spin_unlock_irq(&running_trace_lock
);
642 relay_flush(bt
->rchan
);
649 EXPORT_SYMBOL_GPL(blk_trace_startstop
);
652 * blk_trace_ioctl: - handle the ioctls associated with tracing
653 * @bdev: the block device
654 * @cmd: the ioctl cmd
655 * @arg: the argument data, if any
658 int blk_trace_ioctl(struct block_device
*bdev
, unsigned cmd
, char __user
*arg
)
660 struct request_queue
*q
;
662 char b
[BDEVNAME_SIZE
];
664 q
= bdev_get_queue(bdev
);
668 mutex_lock(&bdev
->bd_mutex
);
673 ret
= blk_trace_setup(q
, b
, bdev
->bd_dev
, bdev
, arg
);
675 #if defined(CONFIG_COMPAT) && defined(CONFIG_X86_64)
676 case BLKTRACESETUP32
:
678 ret
= compat_blk_trace_setup(q
, b
, bdev
->bd_dev
, bdev
, arg
);
684 ret
= blk_trace_startstop(q
, start
);
686 case BLKTRACETEARDOWN
:
687 ret
= blk_trace_remove(q
);
694 mutex_unlock(&bdev
->bd_mutex
);
699 * blk_trace_shutdown: - stop and cleanup trace structures
700 * @q: the request queue associated with the device
703 void blk_trace_shutdown(struct request_queue
*q
)
706 blk_trace_startstop(q
, 0);
711 #ifdef CONFIG_BLK_CGROUP
712 static union kernfs_node_id
*
713 blk_trace_bio_get_cgid(struct request_queue
*q
, struct bio
*bio
)
715 struct blk_trace
*bt
= q
->blk_trace
;
717 if (!bt
|| !(blk_tracer_flags
.val
& TRACE_BLK_OPT_CGROUP
))
722 return cgroup_get_kernfs_id(bio
->bi_css
->cgroup
);
725 static union kernfs_node_id
*
726 blk_trace_bio_get_cgid(struct request_queue
*q
, struct bio
*bio
)
732 static union kernfs_node_id
*
733 blk_trace_request_get_cgid(struct request_queue
*q
, struct request
*rq
)
737 /* Use the first bio */
738 return blk_trace_bio_get_cgid(q
, rq
->bio
);
746 * blk_add_trace_rq - Add a trace for a request oriented action
747 * @rq: the source request
748 * @error: return status to log
749 * @nr_bytes: number of completed bytes
751 * @cgid: the cgroup info
754 * Records an action against a request. Will log the bio offset + size.
757 static void blk_add_trace_rq(struct request
*rq
, int error
,
758 unsigned int nr_bytes
, u32 what
,
759 union kernfs_node_id
*cgid
)
761 struct blk_trace
*bt
= rq
->q
->blk_trace
;
766 if (blk_rq_is_passthrough(rq
))
767 what
|= BLK_TC_ACT(BLK_TC_PC
);
769 what
|= BLK_TC_ACT(BLK_TC_FS
);
771 __blk_add_trace(bt
, blk_rq_trace_sector(rq
), nr_bytes
, req_op(rq
),
772 rq
->cmd_flags
, what
, error
, 0, NULL
, cgid
);
775 static void blk_add_trace_rq_insert(void *ignore
,
776 struct request_queue
*q
, struct request
*rq
)
778 blk_add_trace_rq(rq
, 0, blk_rq_bytes(rq
), BLK_TA_INSERT
,
779 blk_trace_request_get_cgid(q
, rq
));
782 static void blk_add_trace_rq_issue(void *ignore
,
783 struct request_queue
*q
, struct request
*rq
)
785 blk_add_trace_rq(rq
, 0, blk_rq_bytes(rq
), BLK_TA_ISSUE
,
786 blk_trace_request_get_cgid(q
, rq
));
789 static void blk_add_trace_rq_requeue(void *ignore
,
790 struct request_queue
*q
,
793 blk_add_trace_rq(rq
, 0, blk_rq_bytes(rq
), BLK_TA_REQUEUE
,
794 blk_trace_request_get_cgid(q
, rq
));
797 static void blk_add_trace_rq_complete(void *ignore
, struct request
*rq
,
798 int error
, unsigned int nr_bytes
)
800 blk_add_trace_rq(rq
, error
, nr_bytes
, BLK_TA_COMPLETE
,
801 blk_trace_request_get_cgid(rq
->q
, rq
));
805 * blk_add_trace_bio - Add a trace for a bio oriented action
806 * @q: queue the io is for
807 * @bio: the source bio
809 * @error: error, if any
812 * Records an action against a bio. Will log the bio offset + size.
815 static void blk_add_trace_bio(struct request_queue
*q
, struct bio
*bio
,
816 u32 what
, int error
, union kernfs_node_id
*cgid
)
818 struct blk_trace
*bt
= q
->blk_trace
;
823 __blk_add_trace(bt
, bio
->bi_iter
.bi_sector
, bio
->bi_iter
.bi_size
,
824 bio_op(bio
), bio
->bi_opf
, what
, error
, 0, NULL
, cgid
);
827 static void blk_add_trace_bio_bounce(void *ignore
,
828 struct request_queue
*q
, struct bio
*bio
)
830 blk_add_trace_bio(q
, bio
, BLK_TA_BOUNCE
, 0,
831 blk_trace_bio_get_cgid(q
, bio
));
834 static void blk_add_trace_bio_complete(void *ignore
,
835 struct request_queue
*q
, struct bio
*bio
,
838 blk_add_trace_bio(q
, bio
, BLK_TA_COMPLETE
, error
,
839 blk_trace_bio_get_cgid(q
, bio
));
842 static void blk_add_trace_bio_backmerge(void *ignore
,
843 struct request_queue
*q
,
847 blk_add_trace_bio(q
, bio
, BLK_TA_BACKMERGE
, 0,
848 blk_trace_bio_get_cgid(q
, bio
));
851 static void blk_add_trace_bio_frontmerge(void *ignore
,
852 struct request_queue
*q
,
856 blk_add_trace_bio(q
, bio
, BLK_TA_FRONTMERGE
, 0,
857 blk_trace_bio_get_cgid(q
, bio
));
860 static void blk_add_trace_bio_queue(void *ignore
,
861 struct request_queue
*q
, struct bio
*bio
)
863 blk_add_trace_bio(q
, bio
, BLK_TA_QUEUE
, 0,
864 blk_trace_bio_get_cgid(q
, bio
));
867 static void blk_add_trace_getrq(void *ignore
,
868 struct request_queue
*q
,
869 struct bio
*bio
, int rw
)
872 blk_add_trace_bio(q
, bio
, BLK_TA_GETRQ
, 0,
873 blk_trace_bio_get_cgid(q
, bio
));
875 struct blk_trace
*bt
= q
->blk_trace
;
878 __blk_add_trace(bt
, 0, 0, rw
, 0, BLK_TA_GETRQ
, 0, 0,
884 static void blk_add_trace_sleeprq(void *ignore
,
885 struct request_queue
*q
,
886 struct bio
*bio
, int rw
)
889 blk_add_trace_bio(q
, bio
, BLK_TA_SLEEPRQ
, 0,
890 blk_trace_bio_get_cgid(q
, bio
));
892 struct blk_trace
*bt
= q
->blk_trace
;
895 __blk_add_trace(bt
, 0, 0, rw
, 0, BLK_TA_SLEEPRQ
,
900 static void blk_add_trace_plug(void *ignore
, struct request_queue
*q
)
902 struct blk_trace
*bt
= q
->blk_trace
;
905 __blk_add_trace(bt
, 0, 0, 0, 0, BLK_TA_PLUG
, 0, 0, NULL
, NULL
);
908 static void blk_add_trace_unplug(void *ignore
, struct request_queue
*q
,
909 unsigned int depth
, bool explicit)
911 struct blk_trace
*bt
= q
->blk_trace
;
914 __be64 rpdu
= cpu_to_be64(depth
);
918 what
= BLK_TA_UNPLUG_IO
;
920 what
= BLK_TA_UNPLUG_TIMER
;
922 __blk_add_trace(bt
, 0, 0, 0, 0, what
, 0, sizeof(rpdu
), &rpdu
, NULL
);
926 static void blk_add_trace_split(void *ignore
,
927 struct request_queue
*q
, struct bio
*bio
,
930 struct blk_trace
*bt
= q
->blk_trace
;
933 __be64 rpdu
= cpu_to_be64(pdu
);
935 __blk_add_trace(bt
, bio
->bi_iter
.bi_sector
,
936 bio
->bi_iter
.bi_size
, bio_op(bio
), bio
->bi_opf
,
937 BLK_TA_SPLIT
, bio
->bi_status
, sizeof(rpdu
),
938 &rpdu
, blk_trace_bio_get_cgid(q
, bio
));
943 * blk_add_trace_bio_remap - Add a trace for a bio-remap operation
944 * @ignore: trace callback data parameter (not used)
945 * @q: queue the io is for
946 * @bio: the source bio
947 * @dev: target device
948 * @from: source sector
951 * Device mapper or raid target sometimes need to split a bio because
952 * it spans a stripe (or similar). Add a trace for that action.
955 static void blk_add_trace_bio_remap(void *ignore
,
956 struct request_queue
*q
, struct bio
*bio
,
957 dev_t dev
, sector_t from
)
959 struct blk_trace
*bt
= q
->blk_trace
;
960 struct blk_io_trace_remap r
;
965 r
.device_from
= cpu_to_be32(dev
);
966 r
.device_to
= cpu_to_be32(bio_dev(bio
));
967 r
.sector_from
= cpu_to_be64(from
);
969 __blk_add_trace(bt
, bio
->bi_iter
.bi_sector
, bio
->bi_iter
.bi_size
,
970 bio_op(bio
), bio
->bi_opf
, BLK_TA_REMAP
, bio
->bi_status
,
971 sizeof(r
), &r
, blk_trace_bio_get_cgid(q
, bio
));
975 * blk_add_trace_rq_remap - Add a trace for a request-remap operation
976 * @ignore: trace callback data parameter (not used)
977 * @q: queue the io is for
978 * @rq: the source request
979 * @dev: target device
980 * @from: source sector
983 * Device mapper remaps request to other devices.
984 * Add a trace for that action.
987 static void blk_add_trace_rq_remap(void *ignore
,
988 struct request_queue
*q
,
989 struct request
*rq
, dev_t dev
,
992 struct blk_trace
*bt
= q
->blk_trace
;
993 struct blk_io_trace_remap r
;
998 r
.device_from
= cpu_to_be32(dev
);
999 r
.device_to
= cpu_to_be32(disk_devt(rq
->rq_disk
));
1000 r
.sector_from
= cpu_to_be64(from
);
1002 __blk_add_trace(bt
, blk_rq_pos(rq
), blk_rq_bytes(rq
),
1003 rq_data_dir(rq
), 0, BLK_TA_REMAP
, 0,
1004 sizeof(r
), &r
, blk_trace_request_get_cgid(q
, rq
));
1008 * blk_add_driver_data - Add binary message with driver-specific data
1009 * @q: queue the io is for
1011 * @data: driver-specific data
1012 * @len: length of driver-specific data
1015 * Some drivers might want to write driver-specific data per request.
1018 void blk_add_driver_data(struct request_queue
*q
,
1020 void *data
, size_t len
)
1022 struct blk_trace
*bt
= q
->blk_trace
;
1027 __blk_add_trace(bt
, blk_rq_trace_sector(rq
), blk_rq_bytes(rq
), 0, 0,
1028 BLK_TA_DRV_DATA
, 0, len
, data
,
1029 blk_trace_request_get_cgid(q
, rq
));
1031 EXPORT_SYMBOL_GPL(blk_add_driver_data
);
1033 static void blk_register_tracepoints(void)
1037 ret
= register_trace_block_rq_insert(blk_add_trace_rq_insert
, NULL
);
1039 ret
= register_trace_block_rq_issue(blk_add_trace_rq_issue
, NULL
);
1041 ret
= register_trace_block_rq_requeue(blk_add_trace_rq_requeue
, NULL
);
1043 ret
= register_trace_block_rq_complete(blk_add_trace_rq_complete
, NULL
);
1045 ret
= register_trace_block_bio_bounce(blk_add_trace_bio_bounce
, NULL
);
1047 ret
= register_trace_block_bio_complete(blk_add_trace_bio_complete
, NULL
);
1049 ret
= register_trace_block_bio_backmerge(blk_add_trace_bio_backmerge
, NULL
);
1051 ret
= register_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge
, NULL
);
1053 ret
= register_trace_block_bio_queue(blk_add_trace_bio_queue
, NULL
);
1055 ret
= register_trace_block_getrq(blk_add_trace_getrq
, NULL
);
1057 ret
= register_trace_block_sleeprq(blk_add_trace_sleeprq
, NULL
);
1059 ret
= register_trace_block_plug(blk_add_trace_plug
, NULL
);
1061 ret
= register_trace_block_unplug(blk_add_trace_unplug
, NULL
);
1063 ret
= register_trace_block_split(blk_add_trace_split
, NULL
);
1065 ret
= register_trace_block_bio_remap(blk_add_trace_bio_remap
, NULL
);
1067 ret
= register_trace_block_rq_remap(blk_add_trace_rq_remap
, NULL
);
1071 static void blk_unregister_tracepoints(void)
1073 unregister_trace_block_rq_remap(blk_add_trace_rq_remap
, NULL
);
1074 unregister_trace_block_bio_remap(blk_add_trace_bio_remap
, NULL
);
1075 unregister_trace_block_split(blk_add_trace_split
, NULL
);
1076 unregister_trace_block_unplug(blk_add_trace_unplug
, NULL
);
1077 unregister_trace_block_plug(blk_add_trace_plug
, NULL
);
1078 unregister_trace_block_sleeprq(blk_add_trace_sleeprq
, NULL
);
1079 unregister_trace_block_getrq(blk_add_trace_getrq
, NULL
);
1080 unregister_trace_block_bio_queue(blk_add_trace_bio_queue
, NULL
);
1081 unregister_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge
, NULL
);
1082 unregister_trace_block_bio_backmerge(blk_add_trace_bio_backmerge
, NULL
);
1083 unregister_trace_block_bio_complete(blk_add_trace_bio_complete
, NULL
);
1084 unregister_trace_block_bio_bounce(blk_add_trace_bio_bounce
, NULL
);
1085 unregister_trace_block_rq_complete(blk_add_trace_rq_complete
, NULL
);
1086 unregister_trace_block_rq_requeue(blk_add_trace_rq_requeue
, NULL
);
1087 unregister_trace_block_rq_issue(blk_add_trace_rq_issue
, NULL
);
1088 unregister_trace_block_rq_insert(blk_add_trace_rq_insert
, NULL
);
1090 tracepoint_synchronize_unregister();
1094 * struct blk_io_tracer formatting routines
1097 static void fill_rwbs(char *rwbs
, const struct blk_io_trace
*t
)
1100 int tc
= t
->action
>> BLK_TC_SHIFT
;
1102 if ((t
->action
& ~__BLK_TN_CGROUP
) == BLK_TN_MESSAGE
) {
1107 if (tc
& BLK_TC_FLUSH
)
1110 if (tc
& BLK_TC_DISCARD
)
1112 else if (tc
& BLK_TC_WRITE
)
1119 if (tc
& BLK_TC_FUA
)
1121 if (tc
& BLK_TC_AHEAD
)
1123 if (tc
& BLK_TC_SYNC
)
1125 if (tc
& BLK_TC_META
)
1132 const struct blk_io_trace
*te_blk_io_trace(const struct trace_entry
*ent
)
1134 return (const struct blk_io_trace
*)ent
;
1137 static inline const void *pdu_start(const struct trace_entry
*ent
, bool has_cg
)
1139 return (void *)(te_blk_io_trace(ent
) + 1) +
1140 (has_cg
? sizeof(union kernfs_node_id
) : 0);
1143 static inline const void *cgid_start(const struct trace_entry
*ent
)
1145 return (void *)(te_blk_io_trace(ent
) + 1);
1148 static inline int pdu_real_len(const struct trace_entry
*ent
, bool has_cg
)
1150 return te_blk_io_trace(ent
)->pdu_len
-
1151 (has_cg
? sizeof(union kernfs_node_id
) : 0);
1154 static inline u32
t_action(const struct trace_entry
*ent
)
1156 return te_blk_io_trace(ent
)->action
;
1159 static inline u32
t_bytes(const struct trace_entry
*ent
)
1161 return te_blk_io_trace(ent
)->bytes
;
1164 static inline u32
t_sec(const struct trace_entry
*ent
)
1166 return te_blk_io_trace(ent
)->bytes
>> 9;
1169 static inline unsigned long long t_sector(const struct trace_entry
*ent
)
1171 return te_blk_io_trace(ent
)->sector
;
1174 static inline __u16
t_error(const struct trace_entry
*ent
)
1176 return te_blk_io_trace(ent
)->error
;
1179 static __u64
get_pdu_int(const struct trace_entry
*ent
, bool has_cg
)
1181 const __u64
*val
= pdu_start(ent
, has_cg
);
1182 return be64_to_cpu(*val
);
1185 static void get_pdu_remap(const struct trace_entry
*ent
,
1186 struct blk_io_trace_remap
*r
, bool has_cg
)
1188 const struct blk_io_trace_remap
*__r
= pdu_start(ent
, has_cg
);
1189 __u64 sector_from
= __r
->sector_from
;
1191 r
->device_from
= be32_to_cpu(__r
->device_from
);
1192 r
->device_to
= be32_to_cpu(__r
->device_to
);
1193 r
->sector_from
= be64_to_cpu(sector_from
);
1196 typedef void (blk_log_action_t
) (struct trace_iterator
*iter
, const char *act
,
1199 static void blk_log_action_classic(struct trace_iterator
*iter
, const char *act
,
1202 char rwbs
[RWBS_LEN
];
1203 unsigned long long ts
= iter
->ts
;
1204 unsigned long nsec_rem
= do_div(ts
, NSEC_PER_SEC
);
1205 unsigned secs
= (unsigned long)ts
;
1206 const struct blk_io_trace
*t
= te_blk_io_trace(iter
->ent
);
1210 trace_seq_printf(&iter
->seq
,
1211 "%3d,%-3d %2d %5d.%09lu %5u %2s %3s ",
1212 MAJOR(t
->device
), MINOR(t
->device
), iter
->cpu
,
1213 secs
, nsec_rem
, iter
->ent
->pid
, act
, rwbs
);
1216 static void blk_log_action(struct trace_iterator
*iter
, const char *act
,
1219 char rwbs
[RWBS_LEN
];
1220 const struct blk_io_trace
*t
= te_blk_io_trace(iter
->ent
);
1224 const union kernfs_node_id
*id
= cgid_start(iter
->ent
);
1226 if (blk_tracer_flags
.val
& TRACE_BLK_OPT_CGNAME
) {
1227 char blkcg_name_buf
[NAME_MAX
+ 1] = "<...>";
1229 cgroup_path_from_kernfs_id(id
, blkcg_name_buf
,
1230 sizeof(blkcg_name_buf
));
1231 trace_seq_printf(&iter
->seq
, "%3d,%-3d %s %2s %3s ",
1232 MAJOR(t
->device
), MINOR(t
->device
),
1233 blkcg_name_buf
, act
, rwbs
);
1235 trace_seq_printf(&iter
->seq
,
1236 "%3d,%-3d %x,%-x %2s %3s ",
1237 MAJOR(t
->device
), MINOR(t
->device
),
1238 id
->ino
, id
->generation
, act
, rwbs
);
1240 trace_seq_printf(&iter
->seq
, "%3d,%-3d %2s %3s ",
1241 MAJOR(t
->device
), MINOR(t
->device
), act
, rwbs
);
1244 static void blk_log_dump_pdu(struct trace_seq
*s
,
1245 const struct trace_entry
*ent
, bool has_cg
)
1247 const unsigned char *pdu_buf
;
1251 pdu_buf
= pdu_start(ent
, has_cg
);
1252 pdu_len
= pdu_real_len(ent
, has_cg
);
1257 /* find the last zero that needs to be printed */
1258 for (end
= pdu_len
- 1; end
>= 0; end
--)
1263 trace_seq_putc(s
, '(');
1265 for (i
= 0; i
< pdu_len
; i
++) {
1267 trace_seq_printf(s
, "%s%02x",
1268 i
== 0 ? "" : " ", pdu_buf
[i
]);
1271 * stop when the rest is just zeroes and indicate so
1272 * with a ".." appended
1274 if (i
== end
&& end
!= pdu_len
- 1) {
1275 trace_seq_puts(s
, " ..) ");
1280 trace_seq_puts(s
, ") ");
1283 static void blk_log_generic(struct trace_seq
*s
, const struct trace_entry
*ent
, bool has_cg
)
1285 char cmd
[TASK_COMM_LEN
];
1287 trace_find_cmdline(ent
->pid
, cmd
);
1289 if (t_action(ent
) & BLK_TC_ACT(BLK_TC_PC
)) {
1290 trace_seq_printf(s
, "%u ", t_bytes(ent
));
1291 blk_log_dump_pdu(s
, ent
, has_cg
);
1292 trace_seq_printf(s
, "[%s]\n", cmd
);
1295 trace_seq_printf(s
, "%llu + %u [%s]\n",
1296 t_sector(ent
), t_sec(ent
), cmd
);
1298 trace_seq_printf(s
, "[%s]\n", cmd
);
1302 static void blk_log_with_error(struct trace_seq
*s
,
1303 const struct trace_entry
*ent
, bool has_cg
)
1305 if (t_action(ent
) & BLK_TC_ACT(BLK_TC_PC
)) {
1306 blk_log_dump_pdu(s
, ent
, has_cg
);
1307 trace_seq_printf(s
, "[%d]\n", t_error(ent
));
1310 trace_seq_printf(s
, "%llu + %u [%d]\n",
1312 t_sec(ent
), t_error(ent
));
1314 trace_seq_printf(s
, "%llu [%d]\n",
1315 t_sector(ent
), t_error(ent
));
1319 static void blk_log_remap(struct trace_seq
*s
, const struct trace_entry
*ent
, bool has_cg
)
1321 struct blk_io_trace_remap r
= { .device_from
= 0, };
1323 get_pdu_remap(ent
, &r
, has_cg
);
1324 trace_seq_printf(s
, "%llu + %u <- (%d,%d) %llu\n",
1325 t_sector(ent
), t_sec(ent
),
1326 MAJOR(r
.device_from
), MINOR(r
.device_from
),
1327 (unsigned long long)r
.sector_from
);
1330 static void blk_log_plug(struct trace_seq
*s
, const struct trace_entry
*ent
, bool has_cg
)
1332 char cmd
[TASK_COMM_LEN
];
1334 trace_find_cmdline(ent
->pid
, cmd
);
1336 trace_seq_printf(s
, "[%s]\n", cmd
);
1339 static void blk_log_unplug(struct trace_seq
*s
, const struct trace_entry
*ent
, bool has_cg
)
1341 char cmd
[TASK_COMM_LEN
];
1343 trace_find_cmdline(ent
->pid
, cmd
);
1345 trace_seq_printf(s
, "[%s] %llu\n", cmd
, get_pdu_int(ent
, has_cg
));
1348 static void blk_log_split(struct trace_seq
*s
, const struct trace_entry
*ent
, bool has_cg
)
1350 char cmd
[TASK_COMM_LEN
];
1352 trace_find_cmdline(ent
->pid
, cmd
);
1354 trace_seq_printf(s
, "%llu / %llu [%s]\n", t_sector(ent
),
1355 get_pdu_int(ent
, has_cg
), cmd
);
1358 static void blk_log_msg(struct trace_seq
*s
, const struct trace_entry
*ent
,
1362 trace_seq_putmem(s
, pdu_start(ent
, has_cg
),
1363 pdu_real_len(ent
, has_cg
));
1364 trace_seq_putc(s
, '\n');
1368 * struct tracer operations
1371 static void blk_tracer_print_header(struct seq_file
*m
)
1373 if (!(blk_tracer_flags
.val
& TRACE_BLK_OPT_CLASSIC
))
1375 seq_puts(m
, "# DEV CPU TIMESTAMP PID ACT FLG\n"
1379 static void blk_tracer_start(struct trace_array
*tr
)
1381 blk_tracer_enabled
= true;
1384 static int blk_tracer_init(struct trace_array
*tr
)
1387 blk_tracer_start(tr
);
1391 static void blk_tracer_stop(struct trace_array
*tr
)
1393 blk_tracer_enabled
= false;
1396 static void blk_tracer_reset(struct trace_array
*tr
)
1398 blk_tracer_stop(tr
);
1401 static const struct {
1403 void (*print
)(struct trace_seq
*s
, const struct trace_entry
*ent
,
1406 [__BLK_TA_QUEUE
] = {{ "Q", "queue" }, blk_log_generic
},
1407 [__BLK_TA_BACKMERGE
] = {{ "M", "backmerge" }, blk_log_generic
},
1408 [__BLK_TA_FRONTMERGE
] = {{ "F", "frontmerge" }, blk_log_generic
},
1409 [__BLK_TA_GETRQ
] = {{ "G", "getrq" }, blk_log_generic
},
1410 [__BLK_TA_SLEEPRQ
] = {{ "S", "sleeprq" }, blk_log_generic
},
1411 [__BLK_TA_REQUEUE
] = {{ "R", "requeue" }, blk_log_with_error
},
1412 [__BLK_TA_ISSUE
] = {{ "D", "issue" }, blk_log_generic
},
1413 [__BLK_TA_COMPLETE
] = {{ "C", "complete" }, blk_log_with_error
},
1414 [__BLK_TA_PLUG
] = {{ "P", "plug" }, blk_log_plug
},
1415 [__BLK_TA_UNPLUG_IO
] = {{ "U", "unplug_io" }, blk_log_unplug
},
1416 [__BLK_TA_UNPLUG_TIMER
] = {{ "UT", "unplug_timer" }, blk_log_unplug
},
1417 [__BLK_TA_INSERT
] = {{ "I", "insert" }, blk_log_generic
},
1418 [__BLK_TA_SPLIT
] = {{ "X", "split" }, blk_log_split
},
1419 [__BLK_TA_BOUNCE
] = {{ "B", "bounce" }, blk_log_generic
},
1420 [__BLK_TA_REMAP
] = {{ "A", "remap" }, blk_log_remap
},
1423 static enum print_line_t
print_one_line(struct trace_iterator
*iter
,
1426 struct trace_array
*tr
= iter
->tr
;
1427 struct trace_seq
*s
= &iter
->seq
;
1428 const struct blk_io_trace
*t
;
1431 blk_log_action_t
*log_action
;
1434 t
= te_blk_io_trace(iter
->ent
);
1435 what
= (t
->action
& ((1 << BLK_TC_SHIFT
) - 1)) & ~__BLK_TA_CGROUP
;
1436 long_act
= !!(tr
->trace_flags
& TRACE_ITER_VERBOSE
);
1437 log_action
= classic
? &blk_log_action_classic
: &blk_log_action
;
1438 has_cg
= t
->action
& __BLK_TA_CGROUP
;
1440 if ((t
->action
& ~__BLK_TN_CGROUP
) == BLK_TN_MESSAGE
) {
1441 log_action(iter
, long_act
? "message" : "m", has_cg
);
1442 blk_log_msg(s
, iter
->ent
, has_cg
);
1443 return trace_handle_return(s
);
1446 if (unlikely(what
== 0 || what
>= ARRAY_SIZE(what2act
)))
1447 trace_seq_printf(s
, "Unknown action %x\n", what
);
1449 log_action(iter
, what2act
[what
].act
[long_act
], has_cg
);
1450 what2act
[what
].print(s
, iter
->ent
, has_cg
);
1453 return trace_handle_return(s
);
1456 static enum print_line_t
blk_trace_event_print(struct trace_iterator
*iter
,
1457 int flags
, struct trace_event
*event
)
1459 return print_one_line(iter
, false);
1462 static void blk_trace_synthesize_old_trace(struct trace_iterator
*iter
)
1464 struct trace_seq
*s
= &iter
->seq
;
1465 struct blk_io_trace
*t
= (struct blk_io_trace
*)iter
->ent
;
1466 const int offset
= offsetof(struct blk_io_trace
, sector
);
1467 struct blk_io_trace old
= {
1468 .magic
= BLK_IO_TRACE_MAGIC
| BLK_IO_TRACE_VERSION
,
1472 trace_seq_putmem(s
, &old
, offset
);
1473 trace_seq_putmem(s
, &t
->sector
,
1474 sizeof(old
) - offset
+ t
->pdu_len
);
1477 static enum print_line_t
1478 blk_trace_event_print_binary(struct trace_iterator
*iter
, int flags
,
1479 struct trace_event
*event
)
1481 blk_trace_synthesize_old_trace(iter
);
1483 return trace_handle_return(&iter
->seq
);
1486 static enum print_line_t
blk_tracer_print_line(struct trace_iterator
*iter
)
1488 if (!(blk_tracer_flags
.val
& TRACE_BLK_OPT_CLASSIC
))
1489 return TRACE_TYPE_UNHANDLED
;
1491 return print_one_line(iter
, true);
1495 blk_tracer_set_flag(struct trace_array
*tr
, u32 old_flags
, u32 bit
, int set
)
1497 /* don't output context-info for blk_classic output */
1498 if (bit
== TRACE_BLK_OPT_CLASSIC
) {
1500 tr
->trace_flags
&= ~TRACE_ITER_CONTEXT_INFO
;
1502 tr
->trace_flags
|= TRACE_ITER_CONTEXT_INFO
;
1507 static struct tracer blk_tracer __read_mostly
= {
1509 .init
= blk_tracer_init
,
1510 .reset
= blk_tracer_reset
,
1511 .start
= blk_tracer_start
,
1512 .stop
= blk_tracer_stop
,
1513 .print_header
= blk_tracer_print_header
,
1514 .print_line
= blk_tracer_print_line
,
1515 .flags
= &blk_tracer_flags
,
1516 .set_flag
= blk_tracer_set_flag
,
1519 static struct trace_event_functions trace_blk_event_funcs
= {
1520 .trace
= blk_trace_event_print
,
1521 .binary
= blk_trace_event_print_binary
,
1524 static struct trace_event trace_blk_event
= {
1526 .funcs
= &trace_blk_event_funcs
,
1529 static int __init
init_blk_tracer(void)
1531 if (!register_trace_event(&trace_blk_event
)) {
1532 pr_warn("Warning: could not register block events\n");
1536 if (register_tracer(&blk_tracer
) != 0) {
1537 pr_warn("Warning: could not register the block tracer\n");
1538 unregister_trace_event(&trace_blk_event
);
1545 device_initcall(init_blk_tracer
);
1547 static int blk_trace_remove_queue(struct request_queue
*q
)
1549 struct blk_trace
*bt
;
1551 bt
= xchg(&q
->blk_trace
, NULL
);
1555 if (atomic_dec_and_test(&blk_probes_ref
))
1556 blk_unregister_tracepoints();
1563 * Setup everything required to start tracing
1565 static int blk_trace_setup_queue(struct request_queue
*q
,
1566 struct block_device
*bdev
)
1568 struct blk_trace
*bt
= NULL
;
1571 bt
= kzalloc(sizeof(*bt
), GFP_KERNEL
);
1575 bt
->msg_data
= __alloc_percpu(BLK_TN_MAX_MSG
, __alignof__(char));
1579 bt
->dev
= bdev
->bd_dev
;
1580 bt
->act_mask
= (u16
)-1;
1582 blk_trace_setup_lba(bt
, bdev
);
1585 if (cmpxchg(&q
->blk_trace
, NULL
, bt
))
1588 if (atomic_inc_return(&blk_probes_ref
) == 1)
1589 blk_register_tracepoints();
1598 * sysfs interface to enable and configure tracing
1601 static ssize_t
sysfs_blk_trace_attr_show(struct device
*dev
,
1602 struct device_attribute
*attr
,
1604 static ssize_t
sysfs_blk_trace_attr_store(struct device
*dev
,
1605 struct device_attribute
*attr
,
1606 const char *buf
, size_t count
);
1607 #define BLK_TRACE_DEVICE_ATTR(_name) \
1608 DEVICE_ATTR(_name, S_IRUGO | S_IWUSR, \
1609 sysfs_blk_trace_attr_show, \
1610 sysfs_blk_trace_attr_store)
1612 static BLK_TRACE_DEVICE_ATTR(enable
);
1613 static BLK_TRACE_DEVICE_ATTR(act_mask
);
1614 static BLK_TRACE_DEVICE_ATTR(pid
);
1615 static BLK_TRACE_DEVICE_ATTR(start_lba
);
1616 static BLK_TRACE_DEVICE_ATTR(end_lba
);
1618 static struct attribute
*blk_trace_attrs
[] = {
1619 &dev_attr_enable
.attr
,
1620 &dev_attr_act_mask
.attr
,
1622 &dev_attr_start_lba
.attr
,
1623 &dev_attr_end_lba
.attr
,
1627 struct attribute_group blk_trace_attr_group
= {
1629 .attrs
= blk_trace_attrs
,
1632 static const struct {
1636 { BLK_TC_READ
, "read" },
1637 { BLK_TC_WRITE
, "write" },
1638 { BLK_TC_FLUSH
, "flush" },
1639 { BLK_TC_SYNC
, "sync" },
1640 { BLK_TC_QUEUE
, "queue" },
1641 { BLK_TC_REQUEUE
, "requeue" },
1642 { BLK_TC_ISSUE
, "issue" },
1643 { BLK_TC_COMPLETE
, "complete" },
1644 { BLK_TC_FS
, "fs" },
1645 { BLK_TC_PC
, "pc" },
1646 { BLK_TC_NOTIFY
, "notify" },
1647 { BLK_TC_AHEAD
, "ahead" },
1648 { BLK_TC_META
, "meta" },
1649 { BLK_TC_DISCARD
, "discard" },
1650 { BLK_TC_DRV_DATA
, "drv_data" },
1651 { BLK_TC_FUA
, "fua" },
1654 static int blk_trace_str2mask(const char *str
)
1658 char *buf
, *s
, *token
;
1660 buf
= kstrdup(str
, GFP_KERNEL
);
1666 token
= strsep(&s
, ",");
1673 for (i
= 0; i
< ARRAY_SIZE(mask_maps
); i
++) {
1674 if (strcasecmp(token
, mask_maps
[i
].str
) == 0) {
1675 mask
|= mask_maps
[i
].mask
;
1679 if (i
== ARRAY_SIZE(mask_maps
)) {
1689 static ssize_t
blk_trace_mask2str(char *buf
, int mask
)
1694 for (i
= 0; i
< ARRAY_SIZE(mask_maps
); i
++) {
1695 if (mask
& mask_maps
[i
].mask
) {
1696 p
+= sprintf(p
, "%s%s",
1697 (p
== buf
) ? "" : ",", mask_maps
[i
].str
);
1705 static struct request_queue
*blk_trace_get_queue(struct block_device
*bdev
)
1707 if (bdev
->bd_disk
== NULL
)
1710 return bdev_get_queue(bdev
);
1713 static ssize_t
sysfs_blk_trace_attr_show(struct device
*dev
,
1714 struct device_attribute
*attr
,
1717 struct hd_struct
*p
= dev_to_part(dev
);
1718 struct request_queue
*q
;
1719 struct block_device
*bdev
;
1720 ssize_t ret
= -ENXIO
;
1722 bdev
= bdget(part_devt(p
));
1726 q
= blk_trace_get_queue(bdev
);
1730 mutex_lock(&bdev
->bd_mutex
);
1732 if (attr
== &dev_attr_enable
) {
1733 ret
= sprintf(buf
, "%u\n", !!q
->blk_trace
);
1734 goto out_unlock_bdev
;
1737 if (q
->blk_trace
== NULL
)
1738 ret
= sprintf(buf
, "disabled\n");
1739 else if (attr
== &dev_attr_act_mask
)
1740 ret
= blk_trace_mask2str(buf
, q
->blk_trace
->act_mask
);
1741 else if (attr
== &dev_attr_pid
)
1742 ret
= sprintf(buf
, "%u\n", q
->blk_trace
->pid
);
1743 else if (attr
== &dev_attr_start_lba
)
1744 ret
= sprintf(buf
, "%llu\n", q
->blk_trace
->start_lba
);
1745 else if (attr
== &dev_attr_end_lba
)
1746 ret
= sprintf(buf
, "%llu\n", q
->blk_trace
->end_lba
);
1749 mutex_unlock(&bdev
->bd_mutex
);
1756 static ssize_t
sysfs_blk_trace_attr_store(struct device
*dev
,
1757 struct device_attribute
*attr
,
1758 const char *buf
, size_t count
)
1760 struct block_device
*bdev
;
1761 struct request_queue
*q
;
1762 struct hd_struct
*p
;
1764 ssize_t ret
= -EINVAL
;
1769 if (attr
== &dev_attr_act_mask
) {
1770 if (kstrtoull(buf
, 0, &value
)) {
1771 /* Assume it is a list of trace category names */
1772 ret
= blk_trace_str2mask(buf
);
1777 } else if (kstrtoull(buf
, 0, &value
))
1782 p
= dev_to_part(dev
);
1783 bdev
= bdget(part_devt(p
));
1787 q
= blk_trace_get_queue(bdev
);
1791 mutex_lock(&bdev
->bd_mutex
);
1793 if (attr
== &dev_attr_enable
) {
1795 ret
= blk_trace_setup_queue(q
, bdev
);
1797 ret
= blk_trace_remove_queue(q
);
1798 goto out_unlock_bdev
;
1802 if (q
->blk_trace
== NULL
)
1803 ret
= blk_trace_setup_queue(q
, bdev
);
1806 if (attr
== &dev_attr_act_mask
)
1807 q
->blk_trace
->act_mask
= value
;
1808 else if (attr
== &dev_attr_pid
)
1809 q
->blk_trace
->pid
= value
;
1810 else if (attr
== &dev_attr_start_lba
)
1811 q
->blk_trace
->start_lba
= value
;
1812 else if (attr
== &dev_attr_end_lba
)
1813 q
->blk_trace
->end_lba
= value
;
1817 mutex_unlock(&bdev
->bd_mutex
);
1821 return ret
? ret
: count
;
1824 int blk_trace_init_sysfs(struct device
*dev
)
1826 return sysfs_create_group(&dev
->kobj
, &blk_trace_attr_group
);
1829 void blk_trace_remove_sysfs(struct device
*dev
)
1831 sysfs_remove_group(&dev
->kobj
, &blk_trace_attr_group
);
1834 #endif /* CONFIG_BLK_DEV_IO_TRACE */
1836 #ifdef CONFIG_EVENT_TRACING
1838 void blk_fill_rwbs(char *rwbs
, unsigned int op
, int bytes
)
1842 if (op
& REQ_PREFLUSH
)
1845 switch (op
& REQ_OP_MASK
) {
1847 case REQ_OP_WRITE_SAME
:
1850 case REQ_OP_DISCARD
:
1853 case REQ_OP_SECURE_ERASE
:
1869 if (op
& REQ_RAHEAD
)
1878 EXPORT_SYMBOL_GPL(blk_fill_rwbs
);
1880 #endif /* CONFIG_EVENT_TRACING */