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/smp_lock.h>
27 #include <linux/time.h>
28 #include <linux/uaccess.h>
30 #include <trace/events/block.h>
32 #include "trace_output.h"
34 #ifdef CONFIG_BLK_DEV_IO_TRACE
36 static unsigned int blktrace_seq __read_mostly = 1;
38 static struct trace_array *blk_tr;
39 static bool blk_tracer_enabled __read_mostly;
41 /* Select an alternative, minimalistic output than the original one */
42 #define TRACE_BLK_OPT_CLASSIC 0x1
44 static struct tracer_opt blk_tracer_opts[] = {
45 /* Default disable the minimalistic output */
46 { TRACER_OPT(blk_classic, TRACE_BLK_OPT_CLASSIC) },
50 static struct tracer_flags blk_tracer_flags = {
52 .opts = blk_tracer_opts,
55 /* Global reference count of probes */
56 static atomic_t blk_probes_ref = ATOMIC_INIT(0);
58 static void blk_register_tracepoints(void);
59 static void blk_unregister_tracepoints(void);
62 * Send out a notify message.
64 static void trace_note(struct blk_trace *bt, pid_t pid, int action,
65 const void *data, size_t len)
67 struct blk_io_trace *t;
68 struct ring_buffer_event *event = NULL;
69 struct ring_buffer *buffer = NULL;
71 int cpu = smp_processor_id();
72 bool blk_tracer = blk_tracer_enabled;
75 buffer = blk_tr->buffer;
77 event = trace_buffer_lock_reserve(buffer, TRACE_BLK,
82 t = ring_buffer_event_data(event);
89 t = relay_reserve(bt->rchan, sizeof(*t) + len);
91 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
92 t->time = ktime_to_ns(ktime_get());
99 memcpy((void *) t + sizeof(*t), data, len);
102 trace_buffer_unlock_commit(buffer, event, 0, pc);
107 * Send out a notify for this process, if we haven't done so since a trace
110 static void trace_note_tsk(struct blk_trace *bt, struct task_struct *tsk)
112 tsk->btrace_seq = blktrace_seq;
113 trace_note(bt, tsk->pid, BLK_TN_PROCESS, tsk->comm, sizeof(tsk->comm));
116 static void trace_note_time(struct blk_trace *bt)
122 getnstimeofday(&now);
123 words[0] = now.tv_sec;
124 words[1] = now.tv_nsec;
126 local_irq_save(flags);
127 trace_note(bt, 0, BLK_TN_TIMESTAMP, words, sizeof(words));
128 local_irq_restore(flags);
131 void __trace_note_message(struct blk_trace *bt, const char *fmt, ...)
138 if (unlikely(bt->trace_state != Blktrace_running &&
139 !blk_tracer_enabled))
142 local_irq_save(flags);
143 buf = per_cpu_ptr(bt->msg_data, smp_processor_id());
145 n = vscnprintf(buf, BLK_TN_MAX_MSG, fmt, args);
148 trace_note(bt, 0, BLK_TN_MESSAGE, buf, n);
149 local_irq_restore(flags);
151 EXPORT_SYMBOL_GPL(__trace_note_message);
153 static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
156 if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
158 if (sector && (sector < bt->start_lba || sector > bt->end_lba))
160 if (bt->pid && pid != bt->pid)
167 * Data direction bit lookup
169 static const u32 ddir_act[2] = { BLK_TC_ACT(BLK_TC_READ),
170 BLK_TC_ACT(BLK_TC_WRITE) };
172 /* The ilog2() calls fall out because they're constant */
173 #define MASK_TC_BIT(rw, __name) ((rw & (1 << BIO_RW_ ## __name)) << \
174 (ilog2(BLK_TC_ ## __name) + BLK_TC_SHIFT - BIO_RW_ ## __name))
177 * The worker for the various blk_add_trace*() types. Fills out a
178 * blk_io_trace structure and places it in a per-cpu subbuffer.
180 static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
181 int rw, u32 what, int error, int pdu_len, void *pdu_data)
183 struct task_struct *tsk = current;
184 struct ring_buffer_event *event = NULL;
185 struct ring_buffer *buffer = NULL;
186 struct blk_io_trace *t;
187 unsigned long flags = 0;
188 unsigned long *sequence;
191 bool blk_tracer = blk_tracer_enabled;
193 if (unlikely(bt->trace_state != Blktrace_running && !blk_tracer))
196 what |= ddir_act[rw & WRITE];
197 what |= MASK_TC_BIT(rw, BARRIER);
198 what |= MASK_TC_BIT(rw, SYNCIO);
199 what |= MASK_TC_BIT(rw, AHEAD);
200 what |= MASK_TC_BIT(rw, META);
201 what |= MASK_TC_BIT(rw, DISCARD);
204 if (act_log_check(bt, what, sector, pid))
206 cpu = raw_smp_processor_id();
209 tracing_record_cmdline(current);
211 buffer = blk_tr->buffer;
212 pc = preempt_count();
213 event = trace_buffer_lock_reserve(buffer, TRACE_BLK,
214 sizeof(*t) + pdu_len,
218 t = ring_buffer_event_data(event);
223 * A word about the locking here - we disable interrupts to reserve
224 * some space in the relay per-cpu buffer, to prevent an irq
225 * from coming in and stepping on our toes.
227 local_irq_save(flags);
229 if (unlikely(tsk->btrace_seq != blktrace_seq))
230 trace_note_tsk(bt, tsk);
232 t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
234 sequence = per_cpu_ptr(bt->sequence, cpu);
236 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
237 t->sequence = ++(*sequence);
238 t->time = ktime_to_ns(ktime_get());
241 * These two are not needed in ftrace as they are in the
242 * generic trace_entry, filled by tracing_generic_entry_update,
243 * but for the trace_event->bin() synthesizer benefit we do it
254 t->pdu_len = pdu_len;
257 memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
260 trace_buffer_unlock_commit(buffer, event, 0, pc);
265 local_irq_restore(flags);
268 static struct dentry *blk_tree_root;
269 static DEFINE_MUTEX(blk_tree_mutex);
271 static void blk_trace_free(struct blk_trace *bt)
273 debugfs_remove(bt->msg_file);
274 debugfs_remove(bt->dropped_file);
275 relay_close(bt->rchan);
276 debugfs_remove(bt->dir);
277 free_percpu(bt->sequence);
278 free_percpu(bt->msg_data);
282 static void blk_trace_cleanup(struct blk_trace *bt)
285 if (atomic_dec_and_test(&blk_probes_ref))
286 blk_unregister_tracepoints();
289 int blk_trace_remove(struct request_queue *q)
291 struct blk_trace *bt;
293 bt = xchg(&q->blk_trace, NULL);
297 if (bt->trace_state != Blktrace_running)
298 blk_trace_cleanup(bt);
302 EXPORT_SYMBOL_GPL(blk_trace_remove);
304 static int blk_dropped_open(struct inode *inode, struct file *filp)
306 filp->private_data = inode->i_private;
311 static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
312 size_t count, loff_t *ppos)
314 struct blk_trace *bt = filp->private_data;
317 snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
319 return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
322 static const struct file_operations blk_dropped_fops = {
323 .owner = THIS_MODULE,
324 .open = blk_dropped_open,
325 .read = blk_dropped_read,
328 static int blk_msg_open(struct inode *inode, struct file *filp)
330 filp->private_data = inode->i_private;
335 static ssize_t blk_msg_write(struct file *filp, const char __user *buffer,
336 size_t count, loff_t *ppos)
339 struct blk_trace *bt;
341 if (count >= BLK_TN_MAX_MSG)
344 msg = kmalloc(count + 1, GFP_KERNEL);
348 if (copy_from_user(msg, buffer, count)) {
354 bt = filp->private_data;
355 __trace_note_message(bt, "%s", msg);
361 static const struct file_operations blk_msg_fops = {
362 .owner = THIS_MODULE,
363 .open = blk_msg_open,
364 .write = blk_msg_write,
368 * Keep track of how many times we encountered a full subbuffer, to aid
369 * the user space app in telling how many lost events there were.
371 static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
372 void *prev_subbuf, size_t prev_padding)
374 struct blk_trace *bt;
376 if (!relay_buf_full(buf))
379 bt = buf->chan->private_data;
380 atomic_inc(&bt->dropped);
384 static int blk_remove_buf_file_callback(struct dentry *dentry)
386 debugfs_remove(dentry);
391 static struct dentry *blk_create_buf_file_callback(const char *filename,
392 struct dentry *parent,
394 struct rchan_buf *buf,
397 return debugfs_create_file(filename, mode, parent, buf,
398 &relay_file_operations);
401 static struct rchan_callbacks blk_relay_callbacks = {
402 .subbuf_start = blk_subbuf_start_callback,
403 .create_buf_file = blk_create_buf_file_callback,
404 .remove_buf_file = blk_remove_buf_file_callback,
407 static void blk_trace_setup_lba(struct blk_trace *bt,
408 struct block_device *bdev)
410 struct hd_struct *part = NULL;
413 part = bdev->bd_part;
416 bt->start_lba = part->start_sect;
417 bt->end_lba = part->start_sect + part->nr_sects;
425 * Setup everything required to start tracing
427 int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
428 struct block_device *bdev,
429 struct blk_user_trace_setup *buts)
431 struct blk_trace *old_bt, *bt = NULL;
432 struct dentry *dir = NULL;
435 if (!buts->buf_size || !buts->buf_nr)
438 strncpy(buts->name, name, BLKTRACE_BDEV_SIZE);
439 buts->name[BLKTRACE_BDEV_SIZE - 1] = '\0';
442 * some device names have larger paths - convert the slashes
443 * to underscores for this to work as expected
445 for (i = 0; i < strlen(buts->name); i++)
446 if (buts->name[i] == '/')
449 bt = kzalloc(sizeof(*bt), GFP_KERNEL);
454 bt->sequence = alloc_percpu(unsigned long);
458 bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG, __alignof__(char));
464 mutex_lock(&blk_tree_mutex);
465 if (!blk_tree_root) {
466 blk_tree_root = debugfs_create_dir("block", NULL);
467 if (!blk_tree_root) {
468 mutex_unlock(&blk_tree_mutex);
472 mutex_unlock(&blk_tree_mutex);
474 dir = debugfs_create_dir(buts->name, blk_tree_root);
481 atomic_set(&bt->dropped, 0);
484 bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt,
486 if (!bt->dropped_file)
489 bt->msg_file = debugfs_create_file("msg", 0222, dir, bt, &blk_msg_fops);
493 bt->rchan = relay_open("trace", dir, buts->buf_size,
494 buts->buf_nr, &blk_relay_callbacks, bt);
498 bt->act_mask = buts->act_mask;
500 bt->act_mask = (u16) -1;
502 blk_trace_setup_lba(bt, bdev);
504 /* overwrite with user settings */
506 bt->start_lba = buts->start_lba;
508 bt->end_lba = buts->end_lba;
511 bt->trace_state = Blktrace_setup;
514 old_bt = xchg(&q->blk_trace, bt);
516 (void) xchg(&q->blk_trace, old_bt);
520 if (atomic_inc_return(&blk_probes_ref) == 1)
521 blk_register_tracepoints();
529 int blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
530 struct block_device *bdev,
533 struct blk_user_trace_setup buts;
536 ret = copy_from_user(&buts, arg, sizeof(buts));
540 ret = do_blk_trace_setup(q, name, dev, bdev, &buts);
544 if (copy_to_user(arg, &buts, sizeof(buts))) {
550 EXPORT_SYMBOL_GPL(blk_trace_setup);
552 int blk_trace_startstop(struct request_queue *q, int start)
555 struct blk_trace *bt = q->blk_trace;
561 * For starting a trace, we can transition from a setup or stopped
562 * trace. For stopping a trace, the state must be running
566 if (bt->trace_state == Blktrace_setup ||
567 bt->trace_state == Blktrace_stopped) {
570 bt->trace_state = Blktrace_running;
576 if (bt->trace_state == Blktrace_running) {
577 bt->trace_state = Blktrace_stopped;
578 relay_flush(bt->rchan);
585 EXPORT_SYMBOL_GPL(blk_trace_startstop);
588 * blk_trace_ioctl: - handle the ioctls associated with tracing
589 * @bdev: the block device
590 * @cmd: the ioctl cmd
591 * @arg: the argument data, if any
594 int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
596 struct request_queue *q;
598 char b[BDEVNAME_SIZE];
600 q = bdev_get_queue(bdev);
604 mutex_lock(&bdev->bd_mutex);
609 ret = blk_trace_setup(q, b, bdev->bd_dev, bdev, arg);
614 ret = blk_trace_startstop(q, start);
616 case BLKTRACETEARDOWN:
617 ret = blk_trace_remove(q);
624 mutex_unlock(&bdev->bd_mutex);
629 * blk_trace_shutdown: - stop and cleanup trace structures
630 * @q: the request queue associated with the device
633 void blk_trace_shutdown(struct request_queue *q)
636 blk_trace_startstop(q, 0);
646 * blk_add_trace_rq - Add a trace for a request oriented action
647 * @q: queue the io is for
648 * @rq: the source request
652 * Records an action against a request. Will log the bio offset + size.
655 static void blk_add_trace_rq(struct request_queue *q, struct request *rq,
658 struct blk_trace *bt = q->blk_trace;
659 int rw = rq->cmd_flags & 0x03;
664 if (blk_discard_rq(rq))
665 rw |= (1 << BIO_RW_DISCARD);
667 if (blk_pc_request(rq)) {
668 what |= BLK_TC_ACT(BLK_TC_PC);
669 __blk_add_trace(bt, 0, blk_rq_bytes(rq), rw,
670 what, rq->errors, rq->cmd_len, rq->cmd);
672 what |= BLK_TC_ACT(BLK_TC_FS);
673 __blk_add_trace(bt, blk_rq_pos(rq), blk_rq_bytes(rq), rw,
674 what, rq->errors, 0, NULL);
678 static void blk_add_trace_rq_abort(void *ignore,
679 struct request_queue *q, struct request *rq)
681 blk_add_trace_rq(q, rq, BLK_TA_ABORT);
684 static void blk_add_trace_rq_insert(void *ignore,
685 struct request_queue *q, struct request *rq)
687 blk_add_trace_rq(q, rq, BLK_TA_INSERT);
690 static void blk_add_trace_rq_issue(void *ignore,
691 struct request_queue *q, struct request *rq)
693 blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
696 static void blk_add_trace_rq_requeue(void *ignore,
697 struct request_queue *q,
700 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
703 static void blk_add_trace_rq_complete(void *ignore,
704 struct request_queue *q,
707 blk_add_trace_rq(q, rq, BLK_TA_COMPLETE);
711 * blk_add_trace_bio - Add a trace for a bio oriented action
712 * @q: queue the io is for
713 * @bio: the source bio
717 * Records an action against a bio. Will log the bio offset + size.
720 static void blk_add_trace_bio(struct request_queue *q, struct bio *bio,
723 struct blk_trace *bt = q->blk_trace;
728 __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what,
729 !bio_flagged(bio, BIO_UPTODATE), 0, NULL);
732 static void blk_add_trace_bio_bounce(void *ignore,
733 struct request_queue *q, struct bio *bio)
735 blk_add_trace_bio(q, bio, BLK_TA_BOUNCE);
738 static void blk_add_trace_bio_complete(void *ignore,
739 struct request_queue *q, struct bio *bio)
741 blk_add_trace_bio(q, bio, BLK_TA_COMPLETE);
744 static void blk_add_trace_bio_backmerge(void *ignore,
745 struct request_queue *q,
748 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
751 static void blk_add_trace_bio_frontmerge(void *ignore,
752 struct request_queue *q,
755 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
758 static void blk_add_trace_bio_queue(void *ignore,
759 struct request_queue *q, struct bio *bio)
761 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
764 static void blk_add_trace_getrq(void *ignore,
765 struct request_queue *q,
766 struct bio *bio, int rw)
769 blk_add_trace_bio(q, bio, BLK_TA_GETRQ);
771 struct blk_trace *bt = q->blk_trace;
774 __blk_add_trace(bt, 0, 0, rw, BLK_TA_GETRQ, 0, 0, NULL);
779 static void blk_add_trace_sleeprq(void *ignore,
780 struct request_queue *q,
781 struct bio *bio, int rw)
784 blk_add_trace_bio(q, bio, BLK_TA_SLEEPRQ);
786 struct blk_trace *bt = q->blk_trace;
789 __blk_add_trace(bt, 0, 0, rw, BLK_TA_SLEEPRQ,
794 static void blk_add_trace_plug(void *ignore, struct request_queue *q)
796 struct blk_trace *bt = q->blk_trace;
799 __blk_add_trace(bt, 0, 0, 0, BLK_TA_PLUG, 0, 0, NULL);
802 static void blk_add_trace_unplug_io(void *ignore, struct request_queue *q)
804 struct blk_trace *bt = q->blk_trace;
807 unsigned int pdu = q->rq.count[READ] + q->rq.count[WRITE];
808 __be64 rpdu = cpu_to_be64(pdu);
810 __blk_add_trace(bt, 0, 0, 0, BLK_TA_UNPLUG_IO, 0,
811 sizeof(rpdu), &rpdu);
815 static void blk_add_trace_unplug_timer(void *ignore, struct request_queue *q)
817 struct blk_trace *bt = q->blk_trace;
820 unsigned int pdu = q->rq.count[READ] + q->rq.count[WRITE];
821 __be64 rpdu = cpu_to_be64(pdu);
823 __blk_add_trace(bt, 0, 0, 0, BLK_TA_UNPLUG_TIMER, 0,
824 sizeof(rpdu), &rpdu);
828 static void blk_add_trace_split(void *ignore,
829 struct request_queue *q, struct bio *bio,
832 struct blk_trace *bt = q->blk_trace;
835 __be64 rpdu = cpu_to_be64(pdu);
837 __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw,
838 BLK_TA_SPLIT, !bio_flagged(bio, BIO_UPTODATE),
839 sizeof(rpdu), &rpdu);
844 * blk_add_trace_remap - Add a trace for a remap operation
845 * @ignore: trace callback data parameter (not used)
846 * @q: queue the io is for
847 * @bio: the source bio
848 * @dev: target device
849 * @from: source sector
852 * Device mapper or raid target sometimes need to split a bio because
853 * it spans a stripe (or similar). Add a trace for that action.
856 static void blk_add_trace_remap(void *ignore,
857 struct request_queue *q, struct bio *bio,
858 dev_t dev, sector_t from)
860 struct blk_trace *bt = q->blk_trace;
861 struct blk_io_trace_remap r;
866 r.device_from = cpu_to_be32(dev);
867 r.device_to = cpu_to_be32(bio->bi_bdev->bd_dev);
868 r.sector_from = cpu_to_be64(from);
870 __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw,
871 BLK_TA_REMAP, !bio_flagged(bio, BIO_UPTODATE),
876 * blk_add_trace_rq_remap - Add a trace for a request-remap operation
877 * @ignore: trace callback data parameter (not used)
878 * @q: queue the io is for
879 * @rq: the source request
880 * @dev: target device
881 * @from: source sector
884 * Device mapper remaps request to other devices.
885 * Add a trace for that action.
888 static void blk_add_trace_rq_remap(void *ignore,
889 struct request_queue *q,
890 struct request *rq, dev_t dev,
893 struct blk_trace *bt = q->blk_trace;
894 struct blk_io_trace_remap r;
899 r.device_from = cpu_to_be32(dev);
900 r.device_to = cpu_to_be32(disk_devt(rq->rq_disk));
901 r.sector_from = cpu_to_be64(from);
903 __blk_add_trace(bt, blk_rq_pos(rq), blk_rq_bytes(rq),
904 rq_data_dir(rq), BLK_TA_REMAP, !!rq->errors,
909 * blk_add_driver_data - Add binary message with driver-specific data
910 * @q: queue the io is for
912 * @data: driver-specific data
913 * @len: length of driver-specific data
916 * Some drivers might want to write driver-specific data per request.
919 void blk_add_driver_data(struct request_queue *q,
921 void *data, size_t len)
923 struct blk_trace *bt = q->blk_trace;
928 if (blk_pc_request(rq))
929 __blk_add_trace(bt, 0, blk_rq_bytes(rq), 0,
930 BLK_TA_DRV_DATA, rq->errors, len, data);
932 __blk_add_trace(bt, blk_rq_pos(rq), blk_rq_bytes(rq), 0,
933 BLK_TA_DRV_DATA, rq->errors, len, data);
935 EXPORT_SYMBOL_GPL(blk_add_driver_data);
937 static void blk_register_tracepoints(void)
941 ret = register_trace_block_rq_abort(blk_add_trace_rq_abort, NULL);
943 ret = register_trace_block_rq_insert(blk_add_trace_rq_insert, NULL);
945 ret = register_trace_block_rq_issue(blk_add_trace_rq_issue, NULL);
947 ret = register_trace_block_rq_requeue(blk_add_trace_rq_requeue, NULL);
949 ret = register_trace_block_rq_complete(blk_add_trace_rq_complete, NULL);
951 ret = register_trace_block_bio_bounce(blk_add_trace_bio_bounce, NULL);
953 ret = register_trace_block_bio_complete(blk_add_trace_bio_complete, NULL);
955 ret = register_trace_block_bio_backmerge(blk_add_trace_bio_backmerge, NULL);
957 ret = register_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge, NULL);
959 ret = register_trace_block_bio_queue(blk_add_trace_bio_queue, NULL);
961 ret = register_trace_block_getrq(blk_add_trace_getrq, NULL);
963 ret = register_trace_block_sleeprq(blk_add_trace_sleeprq, NULL);
965 ret = register_trace_block_plug(blk_add_trace_plug, NULL);
967 ret = register_trace_block_unplug_timer(blk_add_trace_unplug_timer, NULL);
969 ret = register_trace_block_unplug_io(blk_add_trace_unplug_io, NULL);
971 ret = register_trace_block_split(blk_add_trace_split, NULL);
973 ret = register_trace_block_remap(blk_add_trace_remap, NULL);
975 ret = register_trace_block_rq_remap(blk_add_trace_rq_remap, NULL);
979 static void blk_unregister_tracepoints(void)
981 unregister_trace_block_rq_remap(blk_add_trace_rq_remap, NULL);
982 unregister_trace_block_remap(blk_add_trace_remap, NULL);
983 unregister_trace_block_split(blk_add_trace_split, NULL);
984 unregister_trace_block_unplug_io(blk_add_trace_unplug_io, NULL);
985 unregister_trace_block_unplug_timer(blk_add_trace_unplug_timer, NULL);
986 unregister_trace_block_plug(blk_add_trace_plug, NULL);
987 unregister_trace_block_sleeprq(blk_add_trace_sleeprq, NULL);
988 unregister_trace_block_getrq(blk_add_trace_getrq, NULL);
989 unregister_trace_block_bio_queue(blk_add_trace_bio_queue, NULL);
990 unregister_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge, NULL);
991 unregister_trace_block_bio_backmerge(blk_add_trace_bio_backmerge, NULL);
992 unregister_trace_block_bio_complete(blk_add_trace_bio_complete, NULL);
993 unregister_trace_block_bio_bounce(blk_add_trace_bio_bounce, NULL);
994 unregister_trace_block_rq_complete(blk_add_trace_rq_complete, NULL);
995 unregister_trace_block_rq_requeue(blk_add_trace_rq_requeue, NULL);
996 unregister_trace_block_rq_issue(blk_add_trace_rq_issue, NULL);
997 unregister_trace_block_rq_insert(blk_add_trace_rq_insert, NULL);
998 unregister_trace_block_rq_abort(blk_add_trace_rq_abort, NULL);
1000 tracepoint_synchronize_unregister();
1004 * struct blk_io_tracer formatting routines
1007 static void fill_rwbs(char *rwbs, const struct blk_io_trace *t)
1010 int tc = t->action >> BLK_TC_SHIFT;
1012 if (t->action == BLK_TN_MESSAGE) {
1017 if (tc & BLK_TC_DISCARD)
1019 else if (tc & BLK_TC_WRITE)
1026 if (tc & BLK_TC_AHEAD)
1028 if (tc & BLK_TC_BARRIER)
1030 if (tc & BLK_TC_SYNC)
1032 if (tc & BLK_TC_META)
1039 const struct blk_io_trace *te_blk_io_trace(const struct trace_entry *ent)
1041 return (const struct blk_io_trace *)ent;
1044 static inline const void *pdu_start(const struct trace_entry *ent)
1046 return te_blk_io_trace(ent) + 1;
1049 static inline u32 t_action(const struct trace_entry *ent)
1051 return te_blk_io_trace(ent)->action;
1054 static inline u32 t_bytes(const struct trace_entry *ent)
1056 return te_blk_io_trace(ent)->bytes;
1059 static inline u32 t_sec(const struct trace_entry *ent)
1061 return te_blk_io_trace(ent)->bytes >> 9;
1064 static inline unsigned long long t_sector(const struct trace_entry *ent)
1066 return te_blk_io_trace(ent)->sector;
1069 static inline __u16 t_error(const struct trace_entry *ent)
1071 return te_blk_io_trace(ent)->error;
1074 static __u64 get_pdu_int(const struct trace_entry *ent)
1076 const __u64 *val = pdu_start(ent);
1077 return be64_to_cpu(*val);
1080 static void get_pdu_remap(const struct trace_entry *ent,
1081 struct blk_io_trace_remap *r)
1083 const struct blk_io_trace_remap *__r = pdu_start(ent);
1084 __u64 sector_from = __r->sector_from;
1086 r->device_from = be32_to_cpu(__r->device_from);
1087 r->device_to = be32_to_cpu(__r->device_to);
1088 r->sector_from = be64_to_cpu(sector_from);
1091 typedef int (blk_log_action_t) (struct trace_iterator *iter, const char *act);
1093 static int blk_log_action_classic(struct trace_iterator *iter, const char *act)
1096 unsigned long long ts = iter->ts;
1097 unsigned long nsec_rem = do_div(ts, NSEC_PER_SEC);
1098 unsigned secs = (unsigned long)ts;
1099 const struct blk_io_trace *t = te_blk_io_trace(iter->ent);
1103 return trace_seq_printf(&iter->seq,
1104 "%3d,%-3d %2d %5d.%09lu %5u %2s %3s ",
1105 MAJOR(t->device), MINOR(t->device), iter->cpu,
1106 secs, nsec_rem, iter->ent->pid, act, rwbs);
1109 static int blk_log_action(struct trace_iterator *iter, const char *act)
1112 const struct blk_io_trace *t = te_blk_io_trace(iter->ent);
1115 return trace_seq_printf(&iter->seq, "%3d,%-3d %2s %3s ",
1116 MAJOR(t->device), MINOR(t->device), act, rwbs);
1119 static int blk_log_dump_pdu(struct trace_seq *s, const struct trace_entry *ent)
1121 const unsigned char *pdu_buf;
1125 pdu_buf = pdu_start(ent);
1126 pdu_len = te_blk_io_trace(ent)->pdu_len;
1131 /* find the last zero that needs to be printed */
1132 for (end = pdu_len - 1; end >= 0; end--)
1137 if (!trace_seq_putc(s, '('))
1140 for (i = 0; i < pdu_len; i++) {
1142 ret = trace_seq_printf(s, "%s%02x",
1143 i == 0 ? "" : " ", pdu_buf[i]);
1148 * stop when the rest is just zeroes and indicate so
1149 * with a ".." appended
1151 if (i == end && end != pdu_len - 1)
1152 return trace_seq_puts(s, " ..) ");
1155 return trace_seq_puts(s, ") ");
1158 static int blk_log_generic(struct trace_seq *s, const struct trace_entry *ent)
1160 char cmd[TASK_COMM_LEN];
1162 trace_find_cmdline(ent->pid, cmd);
1164 if (t_action(ent) & BLK_TC_ACT(BLK_TC_PC)) {
1167 ret = trace_seq_printf(s, "%u ", t_bytes(ent));
1170 ret = blk_log_dump_pdu(s, ent);
1173 return trace_seq_printf(s, "[%s]\n", cmd);
1176 return trace_seq_printf(s, "%llu + %u [%s]\n",
1177 t_sector(ent), t_sec(ent), cmd);
1178 return trace_seq_printf(s, "[%s]\n", cmd);
1182 static int blk_log_with_error(struct trace_seq *s,
1183 const struct trace_entry *ent)
1185 if (t_action(ent) & BLK_TC_ACT(BLK_TC_PC)) {
1188 ret = blk_log_dump_pdu(s, ent);
1190 return trace_seq_printf(s, "[%d]\n", t_error(ent));
1194 return trace_seq_printf(s, "%llu + %u [%d]\n",
1196 t_sec(ent), t_error(ent));
1197 return trace_seq_printf(s, "%llu [%d]\n",
1198 t_sector(ent), t_error(ent));
1202 static int blk_log_remap(struct trace_seq *s, const struct trace_entry *ent)
1204 struct blk_io_trace_remap r = { .device_from = 0, };
1206 get_pdu_remap(ent, &r);
1207 return trace_seq_printf(s, "%llu + %u <- (%d,%d) %llu\n",
1208 t_sector(ent), t_sec(ent),
1209 MAJOR(r.device_from), MINOR(r.device_from),
1210 (unsigned long long)r.sector_from);
1213 static int blk_log_plug(struct trace_seq *s, const struct trace_entry *ent)
1215 char cmd[TASK_COMM_LEN];
1217 trace_find_cmdline(ent->pid, cmd);
1219 return trace_seq_printf(s, "[%s]\n", cmd);
1222 static int blk_log_unplug(struct trace_seq *s, const struct trace_entry *ent)
1224 char cmd[TASK_COMM_LEN];
1226 trace_find_cmdline(ent->pid, cmd);
1228 return trace_seq_printf(s, "[%s] %llu\n", cmd, get_pdu_int(ent));
1231 static int blk_log_split(struct trace_seq *s, const struct trace_entry *ent)
1233 char cmd[TASK_COMM_LEN];
1235 trace_find_cmdline(ent->pid, cmd);
1237 return trace_seq_printf(s, "%llu / %llu [%s]\n", t_sector(ent),
1238 get_pdu_int(ent), cmd);
1241 static int blk_log_msg(struct trace_seq *s, const struct trace_entry *ent)
1244 const struct blk_io_trace *t = te_blk_io_trace(ent);
1246 ret = trace_seq_putmem(s, t + 1, t->pdu_len);
1248 return trace_seq_putc(s, '\n');
1253 * struct tracer operations
1256 static void blk_tracer_print_header(struct seq_file *m)
1258 if (!(blk_tracer_flags.val & TRACE_BLK_OPT_CLASSIC))
1260 seq_puts(m, "# DEV CPU TIMESTAMP PID ACT FLG\n"
1264 static void blk_tracer_start(struct trace_array *tr)
1266 blk_tracer_enabled = true;
1269 static int blk_tracer_init(struct trace_array *tr)
1272 blk_tracer_start(tr);
1276 static void blk_tracer_stop(struct trace_array *tr)
1278 blk_tracer_enabled = false;
1281 static void blk_tracer_reset(struct trace_array *tr)
1283 blk_tracer_stop(tr);
1286 static const struct {
1288 int (*print)(struct trace_seq *s, const struct trace_entry *ent);
1290 [__BLK_TA_QUEUE] = {{ "Q", "queue" }, blk_log_generic },
1291 [__BLK_TA_BACKMERGE] = {{ "M", "backmerge" }, blk_log_generic },
1292 [__BLK_TA_FRONTMERGE] = {{ "F", "frontmerge" }, blk_log_generic },
1293 [__BLK_TA_GETRQ] = {{ "G", "getrq" }, blk_log_generic },
1294 [__BLK_TA_SLEEPRQ] = {{ "S", "sleeprq" }, blk_log_generic },
1295 [__BLK_TA_REQUEUE] = {{ "R", "requeue" }, blk_log_with_error },
1296 [__BLK_TA_ISSUE] = {{ "D", "issue" }, blk_log_generic },
1297 [__BLK_TA_COMPLETE] = {{ "C", "complete" }, blk_log_with_error },
1298 [__BLK_TA_PLUG] = {{ "P", "plug" }, blk_log_plug },
1299 [__BLK_TA_UNPLUG_IO] = {{ "U", "unplug_io" }, blk_log_unplug },
1300 [__BLK_TA_UNPLUG_TIMER] = {{ "UT", "unplug_timer" }, blk_log_unplug },
1301 [__BLK_TA_INSERT] = {{ "I", "insert" }, blk_log_generic },
1302 [__BLK_TA_SPLIT] = {{ "X", "split" }, blk_log_split },
1303 [__BLK_TA_BOUNCE] = {{ "B", "bounce" }, blk_log_generic },
1304 [__BLK_TA_REMAP] = {{ "A", "remap" }, blk_log_remap },
1307 static enum print_line_t print_one_line(struct trace_iterator *iter,
1310 struct trace_seq *s = &iter->seq;
1311 const struct blk_io_trace *t;
1315 blk_log_action_t *log_action;
1317 t = te_blk_io_trace(iter->ent);
1318 what = t->action & ((1 << BLK_TC_SHIFT) - 1);
1319 long_act = !!(trace_flags & TRACE_ITER_VERBOSE);
1320 log_action = classic ? &blk_log_action_classic : &blk_log_action;
1322 if (t->action == BLK_TN_MESSAGE) {
1323 ret = log_action(iter, long_act ? "message" : "m");
1325 ret = blk_log_msg(s, iter->ent);
1329 if (unlikely(what == 0 || what >= ARRAY_SIZE(what2act)))
1330 ret = trace_seq_printf(s, "Unknown action %x\n", what);
1332 ret = log_action(iter, what2act[what].act[long_act]);
1334 ret = what2act[what].print(s, iter->ent);
1337 return ret ? TRACE_TYPE_HANDLED : TRACE_TYPE_PARTIAL_LINE;
1340 static enum print_line_t blk_trace_event_print(struct trace_iterator *iter,
1341 int flags, struct trace_event *event)
1343 return print_one_line(iter, false);
1346 static int blk_trace_synthesize_old_trace(struct trace_iterator *iter)
1348 struct trace_seq *s = &iter->seq;
1349 struct blk_io_trace *t = (struct blk_io_trace *)iter->ent;
1350 const int offset = offsetof(struct blk_io_trace, sector);
1351 struct blk_io_trace old = {
1352 .magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION,
1356 if (!trace_seq_putmem(s, &old, offset))
1358 return trace_seq_putmem(s, &t->sector,
1359 sizeof(old) - offset + t->pdu_len);
1362 static enum print_line_t
1363 blk_trace_event_print_binary(struct trace_iterator *iter, int flags,
1364 struct trace_event *event)
1366 return blk_trace_synthesize_old_trace(iter) ?
1367 TRACE_TYPE_HANDLED : TRACE_TYPE_PARTIAL_LINE;
1370 static enum print_line_t blk_tracer_print_line(struct trace_iterator *iter)
1372 if (!(blk_tracer_flags.val & TRACE_BLK_OPT_CLASSIC))
1373 return TRACE_TYPE_UNHANDLED;
1375 return print_one_line(iter, true);
1378 static int blk_tracer_set_flag(u32 old_flags, u32 bit, int set)
1380 /* don't output context-info for blk_classic output */
1381 if (bit == TRACE_BLK_OPT_CLASSIC) {
1383 trace_flags &= ~TRACE_ITER_CONTEXT_INFO;
1385 trace_flags |= TRACE_ITER_CONTEXT_INFO;
1390 static struct tracer blk_tracer __read_mostly = {
1392 .init = blk_tracer_init,
1393 .reset = blk_tracer_reset,
1394 .start = blk_tracer_start,
1395 .stop = blk_tracer_stop,
1396 .print_header = blk_tracer_print_header,
1397 .print_line = blk_tracer_print_line,
1398 .flags = &blk_tracer_flags,
1399 .set_flag = blk_tracer_set_flag,
1402 static struct trace_event_functions trace_blk_event_funcs = {
1403 .trace = blk_trace_event_print,
1404 .binary = blk_trace_event_print_binary,
1407 static struct trace_event trace_blk_event = {
1409 .funcs = &trace_blk_event_funcs,
1412 static int __init init_blk_tracer(void)
1414 if (!register_ftrace_event(&trace_blk_event)) {
1415 pr_warning("Warning: could not register block events\n");
1419 if (register_tracer(&blk_tracer) != 0) {
1420 pr_warning("Warning: could not register the block tracer\n");
1421 unregister_ftrace_event(&trace_blk_event);
1428 device_initcall(init_blk_tracer);
1430 static int blk_trace_remove_queue(struct request_queue *q)
1432 struct blk_trace *bt;
1434 bt = xchg(&q->blk_trace, NULL);
1438 if (atomic_dec_and_test(&blk_probes_ref))
1439 blk_unregister_tracepoints();
1446 * Setup everything required to start tracing
1448 static int blk_trace_setup_queue(struct request_queue *q,
1449 struct block_device *bdev)
1451 struct blk_trace *old_bt, *bt = NULL;
1454 bt = kzalloc(sizeof(*bt), GFP_KERNEL);
1458 bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG, __alignof__(char));
1462 bt->dev = bdev->bd_dev;
1463 bt->act_mask = (u16)-1;
1465 blk_trace_setup_lba(bt, bdev);
1467 old_bt = xchg(&q->blk_trace, bt);
1468 if (old_bt != NULL) {
1469 (void)xchg(&q->blk_trace, old_bt);
1474 if (atomic_inc_return(&blk_probes_ref) == 1)
1475 blk_register_tracepoints();
1484 * sysfs interface to enable and configure tracing
1487 static ssize_t sysfs_blk_trace_attr_show(struct device *dev,
1488 struct device_attribute *attr,
1490 static ssize_t sysfs_blk_trace_attr_store(struct device *dev,
1491 struct device_attribute *attr,
1492 const char *buf, size_t count);
1493 #define BLK_TRACE_DEVICE_ATTR(_name) \
1494 DEVICE_ATTR(_name, S_IRUGO | S_IWUSR, \
1495 sysfs_blk_trace_attr_show, \
1496 sysfs_blk_trace_attr_store)
1498 static BLK_TRACE_DEVICE_ATTR(enable);
1499 static BLK_TRACE_DEVICE_ATTR(act_mask);
1500 static BLK_TRACE_DEVICE_ATTR(pid);
1501 static BLK_TRACE_DEVICE_ATTR(start_lba);
1502 static BLK_TRACE_DEVICE_ATTR(end_lba);
1504 static struct attribute *blk_trace_attrs[] = {
1505 &dev_attr_enable.attr,
1506 &dev_attr_act_mask.attr,
1508 &dev_attr_start_lba.attr,
1509 &dev_attr_end_lba.attr,
1513 struct attribute_group blk_trace_attr_group = {
1515 .attrs = blk_trace_attrs,
1518 static const struct {
1522 { BLK_TC_READ, "read" },
1523 { BLK_TC_WRITE, "write" },
1524 { BLK_TC_BARRIER, "barrier" },
1525 { BLK_TC_SYNC, "sync" },
1526 { BLK_TC_QUEUE, "queue" },
1527 { BLK_TC_REQUEUE, "requeue" },
1528 { BLK_TC_ISSUE, "issue" },
1529 { BLK_TC_COMPLETE, "complete" },
1530 { BLK_TC_FS, "fs" },
1531 { BLK_TC_PC, "pc" },
1532 { BLK_TC_AHEAD, "ahead" },
1533 { BLK_TC_META, "meta" },
1534 { BLK_TC_DISCARD, "discard" },
1535 { BLK_TC_DRV_DATA, "drv_data" },
1538 static int blk_trace_str2mask(const char *str)
1542 char *buf, *s, *token;
1544 buf = kstrdup(str, GFP_KERNEL);
1550 token = strsep(&s, ",");
1557 for (i = 0; i < ARRAY_SIZE(mask_maps); i++) {
1558 if (strcasecmp(token, mask_maps[i].str) == 0) {
1559 mask |= mask_maps[i].mask;
1563 if (i == ARRAY_SIZE(mask_maps)) {
1573 static ssize_t blk_trace_mask2str(char *buf, int mask)
1578 for (i = 0; i < ARRAY_SIZE(mask_maps); i++) {
1579 if (mask & mask_maps[i].mask) {
1580 p += sprintf(p, "%s%s",
1581 (p == buf) ? "" : ",", mask_maps[i].str);
1589 static struct request_queue *blk_trace_get_queue(struct block_device *bdev)
1591 if (bdev->bd_disk == NULL)
1594 return bdev_get_queue(bdev);
1597 static ssize_t sysfs_blk_trace_attr_show(struct device *dev,
1598 struct device_attribute *attr,
1601 struct hd_struct *p = dev_to_part(dev);
1602 struct request_queue *q;
1603 struct block_device *bdev;
1604 ssize_t ret = -ENXIO;
1607 bdev = bdget(part_devt(p));
1609 goto out_unlock_kernel;
1611 q = blk_trace_get_queue(bdev);
1615 mutex_lock(&bdev->bd_mutex);
1617 if (attr == &dev_attr_enable) {
1618 ret = sprintf(buf, "%u\n", !!q->blk_trace);
1619 goto out_unlock_bdev;
1622 if (q->blk_trace == NULL)
1623 ret = sprintf(buf, "disabled\n");
1624 else if (attr == &dev_attr_act_mask)
1625 ret = blk_trace_mask2str(buf, q->blk_trace->act_mask);
1626 else if (attr == &dev_attr_pid)
1627 ret = sprintf(buf, "%u\n", q->blk_trace->pid);
1628 else if (attr == &dev_attr_start_lba)
1629 ret = sprintf(buf, "%llu\n", q->blk_trace->start_lba);
1630 else if (attr == &dev_attr_end_lba)
1631 ret = sprintf(buf, "%llu\n", q->blk_trace->end_lba);
1634 mutex_unlock(&bdev->bd_mutex);
1642 static ssize_t sysfs_blk_trace_attr_store(struct device *dev,
1643 struct device_attribute *attr,
1644 const char *buf, size_t count)
1646 struct block_device *bdev;
1647 struct request_queue *q;
1648 struct hd_struct *p;
1650 ssize_t ret = -EINVAL;
1655 if (attr == &dev_attr_act_mask) {
1656 if (sscanf(buf, "%llx", &value) != 1) {
1657 /* Assume it is a list of trace category names */
1658 ret = blk_trace_str2mask(buf);
1663 } else if (sscanf(buf, "%llu", &value) != 1)
1669 p = dev_to_part(dev);
1670 bdev = bdget(part_devt(p));
1672 goto out_unlock_kernel;
1674 q = blk_trace_get_queue(bdev);
1678 mutex_lock(&bdev->bd_mutex);
1680 if (attr == &dev_attr_enable) {
1682 ret = blk_trace_setup_queue(q, bdev);
1684 ret = blk_trace_remove_queue(q);
1685 goto out_unlock_bdev;
1689 if (q->blk_trace == NULL)
1690 ret = blk_trace_setup_queue(q, bdev);
1693 if (attr == &dev_attr_act_mask)
1694 q->blk_trace->act_mask = value;
1695 else if (attr == &dev_attr_pid)
1696 q->blk_trace->pid = value;
1697 else if (attr == &dev_attr_start_lba)
1698 q->blk_trace->start_lba = value;
1699 else if (attr == &dev_attr_end_lba)
1700 q->blk_trace->end_lba = value;
1704 mutex_unlock(&bdev->bd_mutex);
1710 return ret ? ret : count;
1713 int blk_trace_init_sysfs(struct device *dev)
1715 return sysfs_create_group(&dev->kobj, &blk_trace_attr_group);
1718 void blk_trace_remove_sysfs(struct device *dev)
1720 sysfs_remove_group(&dev->kobj, &blk_trace_attr_group);
1723 #endif /* CONFIG_BLK_DEV_IO_TRACE */
1725 #ifdef CONFIG_EVENT_TRACING
1727 void blk_dump_cmd(char *buf, struct request *rq)
1730 int len = rq->cmd_len;
1731 unsigned char *cmd = rq->cmd;
1733 if (!blk_pc_request(rq)) {
1738 for (end = len - 1; end >= 0; end--)
1743 for (i = 0; i < len; i++) {
1744 buf += sprintf(buf, "%s%02x", i == 0 ? "" : " ", cmd[i]);
1745 if (i == end && end != len - 1) {
1746 sprintf(buf, " ..");
1752 void blk_fill_rwbs(char *rwbs, u32 rw, int bytes)
1758 else if (rw & 1 << BIO_RW_DISCARD)
1765 if (rw & 1 << BIO_RW_AHEAD)
1767 if (rw & 1 << BIO_RW_BARRIER)
1769 if (rw & 1 << BIO_RW_SYNCIO)
1771 if (rw & 1 << BIO_RW_META)
1777 void blk_fill_rwbs_rq(char *rwbs, struct request *rq)
1779 int rw = rq->cmd_flags & 0x03;
1782 if (blk_discard_rq(rq))
1783 rw |= (1 << BIO_RW_DISCARD);
1785 bytes = blk_rq_bytes(rq);
1787 blk_fill_rwbs(rwbs, rw, bytes);
1790 #endif /* CONFIG_EVENT_TRACING */