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
19 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
21 #include <linux/kernel.h>
22 #include <linux/blkdev.h>
23 #include <linux/blktrace_api.h>
24 #include <linux/percpu.h>
25 #include <linux/init.h>
26 #include <linux/mutex.h>
27 #include <linux/slab.h>
28 #include <linux/debugfs.h>
29 #include <linux/export.h>
30 #include <linux/time.h>
31 #include <linux/uaccess.h>
32 #include <linux/list.h>
33 #include <linux/blk-cgroup.h>
35 #include "../../block/blk.h"
37 #include <trace/events/block.h>
39 #include "trace_output.h"
41 #ifdef CONFIG_BLK_DEV_IO_TRACE
43 static unsigned int blktrace_seq __read_mostly = 1;
45 static struct trace_array *blk_tr;
46 static bool blk_tracer_enabled __read_mostly;
48 static LIST_HEAD(running_trace_list);
49 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(running_trace_lock);
51 /* Select an alternative, minimalistic output than the original one */
52 #define TRACE_BLK_OPT_CLASSIC 0x1
53 #define TRACE_BLK_OPT_CGROUP 0x2
54 #define TRACE_BLK_OPT_CGNAME 0x4
56 static struct tracer_opt blk_tracer_opts[] = {
57 /* Default disable the minimalistic output */
58 { TRACER_OPT(blk_classic, TRACE_BLK_OPT_CLASSIC) },
59 #ifdef CONFIG_BLK_CGROUP
60 { TRACER_OPT(blk_cgroup, TRACE_BLK_OPT_CGROUP) },
61 { TRACER_OPT(blk_cgname, TRACE_BLK_OPT_CGNAME) },
66 static struct tracer_flags blk_tracer_flags = {
68 .opts = blk_tracer_opts,
71 /* Global reference count of probes */
72 static DEFINE_MUTEX(blk_probe_mutex);
73 static int blk_probes_ref;
75 static void blk_register_tracepoints(void);
76 static void blk_unregister_tracepoints(void);
79 * Send out a notify message.
81 static void trace_note(struct blk_trace *bt, pid_t pid, int action,
82 const void *data, size_t len,
83 union kernfs_node_id *cgid)
85 struct blk_io_trace *t;
86 struct ring_buffer_event *event = NULL;
87 struct ring_buffer *buffer = NULL;
89 int cpu = smp_processor_id();
90 bool blk_tracer = blk_tracer_enabled;
91 ssize_t cgid_len = cgid ? sizeof(*cgid) : 0;
94 buffer = blk_tr->trace_buffer.buffer;
96 event = trace_buffer_lock_reserve(buffer, TRACE_BLK,
97 sizeof(*t) + len + cgid_len,
101 t = ring_buffer_event_data(event);
108 t = relay_reserve(bt->rchan, sizeof(*t) + len + cgid_len);
110 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
111 t->time = ktime_to_ns(ktime_get());
114 t->action = action | (cgid ? __BLK_TN_CGROUP : 0);
117 t->pdu_len = len + cgid_len;
119 memcpy((void *)t + sizeof(*t), cgid, cgid_len);
120 memcpy((void *) t + sizeof(*t) + cgid_len, data, len);
123 trace_buffer_unlock_commit(blk_tr, buffer, event, 0, pc);
128 * Send out a notify for this process, if we haven't done so since a trace
131 static void trace_note_tsk(struct task_struct *tsk)
134 struct blk_trace *bt;
136 tsk->btrace_seq = blktrace_seq;
137 spin_lock_irqsave(&running_trace_lock, flags);
138 list_for_each_entry(bt, &running_trace_list, running_list) {
139 trace_note(bt, tsk->pid, BLK_TN_PROCESS, tsk->comm,
140 sizeof(tsk->comm), NULL);
142 spin_unlock_irqrestore(&running_trace_lock, flags);
145 static void trace_note_time(struct blk_trace *bt)
147 struct timespec64 now;
151 /* need to check user space to see if this breaks in y2038 or y2106 */
152 ktime_get_real_ts64(&now);
153 words[0] = (u32)now.tv_sec;
154 words[1] = now.tv_nsec;
156 local_irq_save(flags);
157 trace_note(bt, 0, BLK_TN_TIMESTAMP, words, sizeof(words), NULL);
158 local_irq_restore(flags);
161 void __trace_note_message(struct blk_trace *bt, struct blkcg *blkcg,
162 const char *fmt, ...)
169 if (unlikely(bt->trace_state != Blktrace_running &&
170 !blk_tracer_enabled))
174 * If the BLK_TC_NOTIFY action mask isn't set, don't send any note
175 * message to the trace.
177 if (!(bt->act_mask & BLK_TC_NOTIFY))
180 local_irq_save(flags);
181 buf = this_cpu_ptr(bt->msg_data);
183 n = vscnprintf(buf, BLK_TN_MAX_MSG, fmt, args);
186 if (!(blk_tracer_flags.val & TRACE_BLK_OPT_CGROUP))
188 #ifdef CONFIG_BLK_CGROUP
189 trace_note(bt, 0, BLK_TN_MESSAGE, buf, n,
190 blkcg ? cgroup_get_kernfs_id(blkcg->css.cgroup) : NULL);
192 trace_note(bt, 0, BLK_TN_MESSAGE, buf, n, NULL);
194 local_irq_restore(flags);
196 EXPORT_SYMBOL_GPL(__trace_note_message);
198 static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
201 if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
203 if (sector && (sector < bt->start_lba || sector > bt->end_lba))
205 if (bt->pid && pid != bt->pid)
212 * Data direction bit lookup
214 static const u32 ddir_act[2] = { BLK_TC_ACT(BLK_TC_READ),
215 BLK_TC_ACT(BLK_TC_WRITE) };
217 #define BLK_TC_RAHEAD BLK_TC_AHEAD
218 #define BLK_TC_PREFLUSH BLK_TC_FLUSH
220 /* The ilog2() calls fall out because they're constant */
221 #define MASK_TC_BIT(rw, __name) ((rw & REQ_ ## __name) << \
222 (ilog2(BLK_TC_ ## __name) + BLK_TC_SHIFT - __REQ_ ## __name))
225 * The worker for the various blk_add_trace*() types. Fills out a
226 * blk_io_trace structure and places it in a per-cpu subbuffer.
228 static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
229 int op, int op_flags, u32 what, int error, int pdu_len,
230 void *pdu_data, union kernfs_node_id *cgid)
232 struct task_struct *tsk = current;
233 struct ring_buffer_event *event = NULL;
234 struct ring_buffer *buffer = NULL;
235 struct blk_io_trace *t;
236 unsigned long flags = 0;
237 unsigned long *sequence;
240 bool blk_tracer = blk_tracer_enabled;
241 ssize_t cgid_len = cgid ? sizeof(*cgid) : 0;
243 if (unlikely(bt->trace_state != Blktrace_running && !blk_tracer))
246 what |= ddir_act[op_is_write(op) ? WRITE : READ];
247 what |= MASK_TC_BIT(op_flags, SYNC);
248 what |= MASK_TC_BIT(op_flags, RAHEAD);
249 what |= MASK_TC_BIT(op_flags, META);
250 what |= MASK_TC_BIT(op_flags, PREFLUSH);
251 what |= MASK_TC_BIT(op_flags, FUA);
252 if (op == REQ_OP_DISCARD || op == REQ_OP_SECURE_ERASE)
253 what |= BLK_TC_ACT(BLK_TC_DISCARD);
254 if (op == REQ_OP_FLUSH)
255 what |= BLK_TC_ACT(BLK_TC_FLUSH);
257 what |= __BLK_TA_CGROUP;
260 if (act_log_check(bt, what, sector, pid))
262 cpu = raw_smp_processor_id();
265 tracing_record_cmdline(current);
267 buffer = blk_tr->trace_buffer.buffer;
268 pc = preempt_count();
269 event = trace_buffer_lock_reserve(buffer, TRACE_BLK,
270 sizeof(*t) + pdu_len + cgid_len,
274 t = ring_buffer_event_data(event);
278 if (unlikely(tsk->btrace_seq != blktrace_seq))
282 * A word about the locking here - we disable interrupts to reserve
283 * some space in the relay per-cpu buffer, to prevent an irq
284 * from coming in and stepping on our toes.
286 local_irq_save(flags);
287 t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len + cgid_len);
289 sequence = per_cpu_ptr(bt->sequence, cpu);
291 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
292 t->sequence = ++(*sequence);
293 t->time = ktime_to_ns(ktime_get());
296 * These two are not needed in ftrace as they are in the
297 * generic trace_entry, filled by tracing_generic_entry_update,
298 * but for the trace_event->bin() synthesizer benefit we do it
309 t->pdu_len = pdu_len + cgid_len;
312 memcpy((void *)t + sizeof(*t), cgid, cgid_len);
314 memcpy((void *)t + sizeof(*t) + cgid_len, pdu_data, pdu_len);
317 trace_buffer_unlock_commit(blk_tr, buffer, event, 0, pc);
322 local_irq_restore(flags);
325 static void blk_trace_free(struct blk_trace *bt)
327 debugfs_remove(bt->msg_file);
328 debugfs_remove(bt->dropped_file);
329 relay_close(bt->rchan);
330 debugfs_remove(bt->dir);
331 free_percpu(bt->sequence);
332 free_percpu(bt->msg_data);
336 static void get_probe_ref(void)
338 mutex_lock(&blk_probe_mutex);
339 if (++blk_probes_ref == 1)
340 blk_register_tracepoints();
341 mutex_unlock(&blk_probe_mutex);
344 static void put_probe_ref(void)
346 mutex_lock(&blk_probe_mutex);
347 if (!--blk_probes_ref)
348 blk_unregister_tracepoints();
349 mutex_unlock(&blk_probe_mutex);
352 static void blk_trace_cleanup(struct blk_trace *bt)
359 static int __blk_trace_remove(struct request_queue *q)
361 struct blk_trace *bt;
363 bt = xchg(&q->blk_trace, NULL);
367 if (bt->trace_state != Blktrace_running)
368 blk_trace_cleanup(bt);
373 int blk_trace_remove(struct request_queue *q)
377 mutex_lock(&q->blk_trace_mutex);
378 ret = __blk_trace_remove(q);
379 mutex_unlock(&q->blk_trace_mutex);
383 EXPORT_SYMBOL_GPL(blk_trace_remove);
385 static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
386 size_t count, loff_t *ppos)
388 struct blk_trace *bt = filp->private_data;
391 snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
393 return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
396 static const struct file_operations blk_dropped_fops = {
397 .owner = THIS_MODULE,
399 .read = blk_dropped_read,
400 .llseek = default_llseek,
403 static ssize_t blk_msg_write(struct file *filp, const char __user *buffer,
404 size_t count, loff_t *ppos)
407 struct blk_trace *bt;
409 if (count >= BLK_TN_MAX_MSG)
412 msg = memdup_user_nul(buffer, count);
416 bt = filp->private_data;
417 __trace_note_message(bt, NULL, "%s", msg);
423 static const struct file_operations blk_msg_fops = {
424 .owner = THIS_MODULE,
426 .write = blk_msg_write,
427 .llseek = noop_llseek,
431 * Keep track of how many times we encountered a full subbuffer, to aid
432 * the user space app in telling how many lost events there were.
434 static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
435 void *prev_subbuf, size_t prev_padding)
437 struct blk_trace *bt;
439 if (!relay_buf_full(buf))
442 bt = buf->chan->private_data;
443 atomic_inc(&bt->dropped);
447 static int blk_remove_buf_file_callback(struct dentry *dentry)
449 debugfs_remove(dentry);
454 static struct dentry *blk_create_buf_file_callback(const char *filename,
455 struct dentry *parent,
457 struct rchan_buf *buf,
460 return debugfs_create_file(filename, mode, parent, buf,
461 &relay_file_operations);
464 static struct rchan_callbacks blk_relay_callbacks = {
465 .subbuf_start = blk_subbuf_start_callback,
466 .create_buf_file = blk_create_buf_file_callback,
467 .remove_buf_file = blk_remove_buf_file_callback,
470 static void blk_trace_setup_lba(struct blk_trace *bt,
471 struct block_device *bdev)
473 struct hd_struct *part = NULL;
476 part = bdev->bd_part;
479 bt->start_lba = part->start_sect;
480 bt->end_lba = part->start_sect + part->nr_sects;
488 * Setup everything required to start tracing
490 static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
491 struct block_device *bdev,
492 struct blk_user_trace_setup *buts)
494 struct blk_trace *bt = NULL;
495 struct dentry *dir = NULL;
498 if (!buts->buf_size || !buts->buf_nr)
501 if (!blk_debugfs_root)
504 strncpy(buts->name, name, BLKTRACE_BDEV_SIZE);
505 buts->name[BLKTRACE_BDEV_SIZE - 1] = '\0';
508 * some device names have larger paths - convert the slashes
509 * to underscores for this to work as expected
511 strreplace(buts->name, '/', '_');
514 * bdev can be NULL, as with scsi-generic, this is a helpful as
518 pr_warn("Concurrent blktraces are not allowed on %s\n",
523 bt = kzalloc(sizeof(*bt), GFP_KERNEL);
528 bt->sequence = alloc_percpu(unsigned long);
532 bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG, __alignof__(char));
536 #ifdef CONFIG_BLK_DEBUG_FS
538 * When tracing whole make_request drivers (multiqueue) block devices,
539 * reuse the existing debugfs directory created by the block layer on
540 * init. For request-based block devices, all partitions block devices,
541 * and scsi-generic block devices we create a temporary new debugfs
542 * directory that will be removed once the trace ends.
544 if (q->mq_ops && bdev && bdev == bdev->bd_contains)
545 dir = q->debugfs_dir;
548 bt->dir = dir = debugfs_create_dir(buts->name, blk_debugfs_root);
553 * As blktrace relies on debugfs for its interface the debugfs directory
554 * is required, contrary to the usual mantra of not checking for debugfs
555 * files or directories.
557 if (IS_ERR_OR_NULL(dir)) {
558 pr_warn("debugfs_dir not present for %s so skipping\n",
565 atomic_set(&bt->dropped, 0);
566 INIT_LIST_HEAD(&bt->running_list);
569 bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt,
571 if (!bt->dropped_file)
574 bt->msg_file = debugfs_create_file("msg", 0222, dir, bt, &blk_msg_fops);
578 bt->rchan = relay_open("trace", dir, buts->buf_size,
579 buts->buf_nr, &blk_relay_callbacks, bt);
583 bt->act_mask = buts->act_mask;
585 bt->act_mask = (u16) -1;
587 blk_trace_setup_lba(bt, bdev);
589 /* overwrite with user settings */
591 bt->start_lba = buts->start_lba;
593 bt->end_lba = buts->end_lba;
596 bt->trace_state = Blktrace_setup;
599 if (cmpxchg(&q->blk_trace, NULL, bt))
611 static int __blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
612 struct block_device *bdev, char __user *arg)
614 struct blk_user_trace_setup buts;
617 ret = copy_from_user(&buts, arg, sizeof(buts));
621 ret = do_blk_trace_setup(q, name, dev, bdev, &buts);
625 if (copy_to_user(arg, &buts, sizeof(buts))) {
626 __blk_trace_remove(q);
632 int blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
633 struct block_device *bdev,
638 mutex_lock(&q->blk_trace_mutex);
639 ret = __blk_trace_setup(q, name, dev, bdev, arg);
640 mutex_unlock(&q->blk_trace_mutex);
644 EXPORT_SYMBOL_GPL(blk_trace_setup);
646 #if defined(CONFIG_COMPAT) && defined(CONFIG_X86_64)
647 static int compat_blk_trace_setup(struct request_queue *q, char *name,
648 dev_t dev, struct block_device *bdev,
651 struct blk_user_trace_setup buts;
652 struct compat_blk_user_trace_setup cbuts;
655 if (copy_from_user(&cbuts, arg, sizeof(cbuts)))
658 buts = (struct blk_user_trace_setup) {
659 .act_mask = cbuts.act_mask,
660 .buf_size = cbuts.buf_size,
661 .buf_nr = cbuts.buf_nr,
662 .start_lba = cbuts.start_lba,
663 .end_lba = cbuts.end_lba,
667 ret = do_blk_trace_setup(q, name, dev, bdev, &buts);
671 if (copy_to_user(arg, &buts.name, ARRAY_SIZE(buts.name))) {
672 __blk_trace_remove(q);
680 static int __blk_trace_startstop(struct request_queue *q, int start)
683 struct blk_trace *bt;
685 bt = rcu_dereference_protected(q->blk_trace,
686 lockdep_is_held(&q->blk_trace_mutex));
691 * For starting a trace, we can transition from a setup or stopped
692 * trace. For stopping a trace, the state must be running
696 if (bt->trace_state == Blktrace_setup ||
697 bt->trace_state == Blktrace_stopped) {
700 bt->trace_state = Blktrace_running;
701 spin_lock_irq(&running_trace_lock);
702 list_add(&bt->running_list, &running_trace_list);
703 spin_unlock_irq(&running_trace_lock);
709 if (bt->trace_state == Blktrace_running) {
710 bt->trace_state = Blktrace_stopped;
711 spin_lock_irq(&running_trace_lock);
712 list_del_init(&bt->running_list);
713 spin_unlock_irq(&running_trace_lock);
714 relay_flush(bt->rchan);
722 int blk_trace_startstop(struct request_queue *q, int start)
726 mutex_lock(&q->blk_trace_mutex);
727 ret = __blk_trace_startstop(q, start);
728 mutex_unlock(&q->blk_trace_mutex);
732 EXPORT_SYMBOL_GPL(blk_trace_startstop);
735 * When reading or writing the blktrace sysfs files, the references to the
736 * opened sysfs or device files should prevent the underlying block device
737 * from being removed. So no further delete protection is really needed.
741 * blk_trace_ioctl: - handle the ioctls associated with tracing
742 * @bdev: the block device
743 * @cmd: the ioctl cmd
744 * @arg: the argument data, if any
747 int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
749 struct request_queue *q;
751 char b[BDEVNAME_SIZE];
753 q = bdev_get_queue(bdev);
757 mutex_lock(&q->blk_trace_mutex);
762 ret = __blk_trace_setup(q, b, bdev->bd_dev, bdev, arg);
764 #if defined(CONFIG_COMPAT) && defined(CONFIG_X86_64)
765 case BLKTRACESETUP32:
767 ret = compat_blk_trace_setup(q, b, bdev->bd_dev, bdev, arg);
773 ret = __blk_trace_startstop(q, start);
775 case BLKTRACETEARDOWN:
776 ret = __blk_trace_remove(q);
783 mutex_unlock(&q->blk_trace_mutex);
788 * blk_trace_shutdown: - stop and cleanup trace structures
789 * @q: the request queue associated with the device
792 void blk_trace_shutdown(struct request_queue *q)
794 mutex_lock(&q->blk_trace_mutex);
795 if (rcu_dereference_protected(q->blk_trace,
796 lockdep_is_held(&q->blk_trace_mutex))) {
797 __blk_trace_startstop(q, 0);
798 __blk_trace_remove(q);
801 mutex_unlock(&q->blk_trace_mutex);
804 #ifdef CONFIG_BLK_CGROUP
805 static union kernfs_node_id *
806 blk_trace_bio_get_cgid(struct request_queue *q, struct bio *bio)
808 struct blk_trace *bt;
810 /* We don't use the 'bt' value here except as an optimization... */
811 bt = rcu_dereference_protected(q->blk_trace, 1);
812 if (!bt || !(blk_tracer_flags.val & TRACE_BLK_OPT_CGROUP))
817 return cgroup_get_kernfs_id(bio->bi_css->cgroup);
820 static union kernfs_node_id *
821 blk_trace_bio_get_cgid(struct request_queue *q, struct bio *bio)
827 static union kernfs_node_id *
828 blk_trace_request_get_cgid(struct request_queue *q, struct request *rq)
832 /* Use the first bio */
833 return blk_trace_bio_get_cgid(q, rq->bio);
841 * blk_add_trace_rq - Add a trace for a request oriented action
842 * @rq: the source request
843 * @error: return status to log
844 * @nr_bytes: number of completed bytes
846 * @cgid: the cgroup info
849 * Records an action against a request. Will log the bio offset + size.
852 static void blk_add_trace_rq(struct request *rq, int error,
853 unsigned int nr_bytes, u32 what,
854 union kernfs_node_id *cgid)
856 struct blk_trace *bt;
859 bt = rcu_dereference(rq->q->blk_trace);
865 if (blk_rq_is_passthrough(rq))
866 what |= BLK_TC_ACT(BLK_TC_PC);
868 what |= BLK_TC_ACT(BLK_TC_FS);
870 __blk_add_trace(bt, blk_rq_trace_sector(rq), nr_bytes, req_op(rq),
871 rq->cmd_flags, what, error, 0, NULL, cgid);
875 static void blk_add_trace_rq_insert(void *ignore,
876 struct request_queue *q, struct request *rq)
878 blk_add_trace_rq(rq, 0, blk_rq_bytes(rq), BLK_TA_INSERT,
879 blk_trace_request_get_cgid(q, rq));
882 static void blk_add_trace_rq_issue(void *ignore,
883 struct request_queue *q, struct request *rq)
885 blk_add_trace_rq(rq, 0, blk_rq_bytes(rq), BLK_TA_ISSUE,
886 blk_trace_request_get_cgid(q, rq));
889 static void blk_add_trace_rq_requeue(void *ignore,
890 struct request_queue *q,
893 blk_add_trace_rq(rq, 0, blk_rq_bytes(rq), BLK_TA_REQUEUE,
894 blk_trace_request_get_cgid(q, rq));
897 static void blk_add_trace_rq_complete(void *ignore, struct request *rq,
898 int error, unsigned int nr_bytes)
900 blk_add_trace_rq(rq, error, nr_bytes, BLK_TA_COMPLETE,
901 blk_trace_request_get_cgid(rq->q, rq));
905 * blk_add_trace_bio - Add a trace for a bio oriented action
906 * @q: queue the io is for
907 * @bio: the source bio
909 * @error: error, if any
912 * Records an action against a bio. Will log the bio offset + size.
915 static void blk_add_trace_bio(struct request_queue *q, struct bio *bio,
916 u32 what, int error, union kernfs_node_id *cgid)
918 struct blk_trace *bt;
921 bt = rcu_dereference(q->blk_trace);
927 __blk_add_trace(bt, bio->bi_iter.bi_sector, bio->bi_iter.bi_size,
928 bio_op(bio), bio->bi_opf, what, error, 0, NULL, cgid);
932 static void blk_add_trace_bio_bounce(void *ignore,
933 struct request_queue *q, struct bio *bio)
935 blk_add_trace_bio(q, bio, BLK_TA_BOUNCE, 0,
936 blk_trace_bio_get_cgid(q, bio));
939 static void blk_add_trace_bio_complete(void *ignore,
940 struct request_queue *q, struct bio *bio,
943 blk_add_trace_bio(q, bio, BLK_TA_COMPLETE, error,
944 blk_trace_bio_get_cgid(q, bio));
947 static void blk_add_trace_bio_backmerge(void *ignore,
948 struct request_queue *q,
952 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE, 0,
953 blk_trace_bio_get_cgid(q, bio));
956 static void blk_add_trace_bio_frontmerge(void *ignore,
957 struct request_queue *q,
961 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE, 0,
962 blk_trace_bio_get_cgid(q, bio));
965 static void blk_add_trace_bio_queue(void *ignore,
966 struct request_queue *q, struct bio *bio)
968 blk_add_trace_bio(q, bio, BLK_TA_QUEUE, 0,
969 blk_trace_bio_get_cgid(q, bio));
972 static void blk_add_trace_getrq(void *ignore,
973 struct request_queue *q,
974 struct bio *bio, int rw)
977 blk_add_trace_bio(q, bio, BLK_TA_GETRQ, 0,
978 blk_trace_bio_get_cgid(q, bio));
980 struct blk_trace *bt;
983 bt = rcu_dereference(q->blk_trace);
985 __blk_add_trace(bt, 0, 0, rw, 0, BLK_TA_GETRQ, 0, 0,
992 static void blk_add_trace_sleeprq(void *ignore,
993 struct request_queue *q,
994 struct bio *bio, int rw)
997 blk_add_trace_bio(q, bio, BLK_TA_SLEEPRQ, 0,
998 blk_trace_bio_get_cgid(q, bio));
1000 struct blk_trace *bt;
1003 bt = rcu_dereference(q->blk_trace);
1005 __blk_add_trace(bt, 0, 0, rw, 0, BLK_TA_SLEEPRQ,
1011 static void blk_add_trace_plug(void *ignore, struct request_queue *q)
1013 struct blk_trace *bt;
1016 bt = rcu_dereference(q->blk_trace);
1018 __blk_add_trace(bt, 0, 0, 0, 0, BLK_TA_PLUG, 0, 0, NULL, NULL);
1022 static void blk_add_trace_unplug(void *ignore, struct request_queue *q,
1023 unsigned int depth, bool explicit)
1025 struct blk_trace *bt;
1028 bt = rcu_dereference(q->blk_trace);
1030 __be64 rpdu = cpu_to_be64(depth);
1034 what = BLK_TA_UNPLUG_IO;
1036 what = BLK_TA_UNPLUG_TIMER;
1038 __blk_add_trace(bt, 0, 0, 0, 0, what, 0, sizeof(rpdu), &rpdu, NULL);
1043 static void blk_add_trace_split(void *ignore,
1044 struct request_queue *q, struct bio *bio,
1047 struct blk_trace *bt;
1050 bt = rcu_dereference(q->blk_trace);
1052 __be64 rpdu = cpu_to_be64(pdu);
1054 __blk_add_trace(bt, bio->bi_iter.bi_sector,
1055 bio->bi_iter.bi_size, bio_op(bio), bio->bi_opf,
1057 blk_status_to_errno(bio->bi_status),
1058 sizeof(rpdu), &rpdu,
1059 blk_trace_bio_get_cgid(q, bio));
1065 * blk_add_trace_bio_remap - Add a trace for a bio-remap operation
1066 * @ignore: trace callback data parameter (not used)
1067 * @q: queue the io is for
1068 * @bio: the source bio
1069 * @dev: target device
1070 * @from: source sector
1073 * Device mapper or raid target sometimes need to split a bio because
1074 * it spans a stripe (or similar). Add a trace for that action.
1077 static void blk_add_trace_bio_remap(void *ignore,
1078 struct request_queue *q, struct bio *bio,
1079 dev_t dev, sector_t from)
1081 struct blk_trace *bt;
1082 struct blk_io_trace_remap r;
1085 bt = rcu_dereference(q->blk_trace);
1091 r.device_from = cpu_to_be32(dev);
1092 r.device_to = cpu_to_be32(bio_dev(bio));
1093 r.sector_from = cpu_to_be64(from);
1095 __blk_add_trace(bt, bio->bi_iter.bi_sector, bio->bi_iter.bi_size,
1096 bio_op(bio), bio->bi_opf, BLK_TA_REMAP,
1097 blk_status_to_errno(bio->bi_status),
1098 sizeof(r), &r, blk_trace_bio_get_cgid(q, bio));
1103 * blk_add_trace_rq_remap - Add a trace for a request-remap operation
1104 * @ignore: trace callback data parameter (not used)
1105 * @q: queue the io is for
1106 * @rq: the source request
1107 * @dev: target device
1108 * @from: source sector
1111 * Device mapper remaps request to other devices.
1112 * Add a trace for that action.
1115 static void blk_add_trace_rq_remap(void *ignore,
1116 struct request_queue *q,
1117 struct request *rq, dev_t dev,
1120 struct blk_trace *bt;
1121 struct blk_io_trace_remap r;
1124 bt = rcu_dereference(q->blk_trace);
1130 r.device_from = cpu_to_be32(dev);
1131 r.device_to = cpu_to_be32(disk_devt(rq->rq_disk));
1132 r.sector_from = cpu_to_be64(from);
1134 __blk_add_trace(bt, blk_rq_pos(rq), blk_rq_bytes(rq),
1135 rq_data_dir(rq), 0, BLK_TA_REMAP, 0,
1136 sizeof(r), &r, blk_trace_request_get_cgid(q, rq));
1141 * blk_add_driver_data - Add binary message with driver-specific data
1142 * @q: queue the io is for
1144 * @data: driver-specific data
1145 * @len: length of driver-specific data
1148 * Some drivers might want to write driver-specific data per request.
1151 void blk_add_driver_data(struct request_queue *q,
1153 void *data, size_t len)
1155 struct blk_trace *bt;
1158 bt = rcu_dereference(q->blk_trace);
1164 __blk_add_trace(bt, blk_rq_trace_sector(rq), blk_rq_bytes(rq), 0, 0,
1165 BLK_TA_DRV_DATA, 0, len, data,
1166 blk_trace_request_get_cgid(q, rq));
1169 EXPORT_SYMBOL_GPL(blk_add_driver_data);
1171 static void blk_register_tracepoints(void)
1175 ret = register_trace_block_rq_insert(blk_add_trace_rq_insert, NULL);
1177 ret = register_trace_block_rq_issue(blk_add_trace_rq_issue, NULL);
1179 ret = register_trace_block_rq_requeue(blk_add_trace_rq_requeue, NULL);
1181 ret = register_trace_block_rq_complete(blk_add_trace_rq_complete, NULL);
1183 ret = register_trace_block_bio_bounce(blk_add_trace_bio_bounce, NULL);
1185 ret = register_trace_block_bio_complete(blk_add_trace_bio_complete, NULL);
1187 ret = register_trace_block_bio_backmerge(blk_add_trace_bio_backmerge, NULL);
1189 ret = register_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge, NULL);
1191 ret = register_trace_block_bio_queue(blk_add_trace_bio_queue, NULL);
1193 ret = register_trace_block_getrq(blk_add_trace_getrq, NULL);
1195 ret = register_trace_block_sleeprq(blk_add_trace_sleeprq, NULL);
1197 ret = register_trace_block_plug(blk_add_trace_plug, NULL);
1199 ret = register_trace_block_unplug(blk_add_trace_unplug, NULL);
1201 ret = register_trace_block_split(blk_add_trace_split, NULL);
1203 ret = register_trace_block_bio_remap(blk_add_trace_bio_remap, NULL);
1205 ret = register_trace_block_rq_remap(blk_add_trace_rq_remap, NULL);
1209 static void blk_unregister_tracepoints(void)
1211 unregister_trace_block_rq_remap(blk_add_trace_rq_remap, NULL);
1212 unregister_trace_block_bio_remap(blk_add_trace_bio_remap, NULL);
1213 unregister_trace_block_split(blk_add_trace_split, NULL);
1214 unregister_trace_block_unplug(blk_add_trace_unplug, NULL);
1215 unregister_trace_block_plug(blk_add_trace_plug, NULL);
1216 unregister_trace_block_sleeprq(blk_add_trace_sleeprq, NULL);
1217 unregister_trace_block_getrq(blk_add_trace_getrq, NULL);
1218 unregister_trace_block_bio_queue(blk_add_trace_bio_queue, NULL);
1219 unregister_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge, NULL);
1220 unregister_trace_block_bio_backmerge(blk_add_trace_bio_backmerge, NULL);
1221 unregister_trace_block_bio_complete(blk_add_trace_bio_complete, NULL);
1222 unregister_trace_block_bio_bounce(blk_add_trace_bio_bounce, NULL);
1223 unregister_trace_block_rq_complete(blk_add_trace_rq_complete, NULL);
1224 unregister_trace_block_rq_requeue(blk_add_trace_rq_requeue, NULL);
1225 unregister_trace_block_rq_issue(blk_add_trace_rq_issue, NULL);
1226 unregister_trace_block_rq_insert(blk_add_trace_rq_insert, NULL);
1228 tracepoint_synchronize_unregister();
1232 * struct blk_io_tracer formatting routines
1235 static void fill_rwbs(char *rwbs, const struct blk_io_trace *t)
1238 int tc = t->action >> BLK_TC_SHIFT;
1240 if ((t->action & ~__BLK_TN_CGROUP) == BLK_TN_MESSAGE) {
1245 if (tc & BLK_TC_FLUSH)
1248 if (tc & BLK_TC_DISCARD)
1250 else if (tc & BLK_TC_WRITE)
1257 if (tc & BLK_TC_FUA)
1259 if (tc & BLK_TC_AHEAD)
1261 if (tc & BLK_TC_SYNC)
1263 if (tc & BLK_TC_META)
1270 const struct blk_io_trace *te_blk_io_trace(const struct trace_entry *ent)
1272 return (const struct blk_io_trace *)ent;
1275 static inline const void *pdu_start(const struct trace_entry *ent, bool has_cg)
1277 return (void *)(te_blk_io_trace(ent) + 1) +
1278 (has_cg ? sizeof(union kernfs_node_id) : 0);
1281 static inline const void *cgid_start(const struct trace_entry *ent)
1283 return (void *)(te_blk_io_trace(ent) + 1);
1286 static inline int pdu_real_len(const struct trace_entry *ent, bool has_cg)
1288 return te_blk_io_trace(ent)->pdu_len -
1289 (has_cg ? sizeof(union kernfs_node_id) : 0);
1292 static inline u32 t_action(const struct trace_entry *ent)
1294 return te_blk_io_trace(ent)->action;
1297 static inline u32 t_bytes(const struct trace_entry *ent)
1299 return te_blk_io_trace(ent)->bytes;
1302 static inline u32 t_sec(const struct trace_entry *ent)
1304 return te_blk_io_trace(ent)->bytes >> 9;
1307 static inline unsigned long long t_sector(const struct trace_entry *ent)
1309 return te_blk_io_trace(ent)->sector;
1312 static inline __u16 t_error(const struct trace_entry *ent)
1314 return te_blk_io_trace(ent)->error;
1317 static __u64 get_pdu_int(const struct trace_entry *ent, bool has_cg)
1319 const __be64 *val = pdu_start(ent, has_cg);
1320 return be64_to_cpu(*val);
1323 typedef void (blk_log_action_t) (struct trace_iterator *iter, const char *act,
1326 static void blk_log_action_classic(struct trace_iterator *iter, const char *act,
1329 char rwbs[RWBS_LEN];
1330 unsigned long long ts = iter->ts;
1331 unsigned long nsec_rem = do_div(ts, NSEC_PER_SEC);
1332 unsigned secs = (unsigned long)ts;
1333 const struct blk_io_trace *t = te_blk_io_trace(iter->ent);
1337 trace_seq_printf(&iter->seq,
1338 "%3d,%-3d %2d %5d.%09lu %5u %2s %3s ",
1339 MAJOR(t->device), MINOR(t->device), iter->cpu,
1340 secs, nsec_rem, iter->ent->pid, act, rwbs);
1343 static void blk_log_action(struct trace_iterator *iter, const char *act,
1346 char rwbs[RWBS_LEN];
1347 const struct blk_io_trace *t = te_blk_io_trace(iter->ent);
1351 const union kernfs_node_id *id = cgid_start(iter->ent);
1353 if (blk_tracer_flags.val & TRACE_BLK_OPT_CGNAME) {
1354 char blkcg_name_buf[NAME_MAX + 1] = "<...>";
1356 cgroup_path_from_kernfs_id(id, blkcg_name_buf,
1357 sizeof(blkcg_name_buf));
1358 trace_seq_printf(&iter->seq, "%3d,%-3d %s %2s %3s ",
1359 MAJOR(t->device), MINOR(t->device),
1360 blkcg_name_buf, act, rwbs);
1362 trace_seq_printf(&iter->seq,
1363 "%3d,%-3d %x,%-x %2s %3s ",
1364 MAJOR(t->device), MINOR(t->device),
1365 id->ino, id->generation, act, rwbs);
1367 trace_seq_printf(&iter->seq, "%3d,%-3d %2s %3s ",
1368 MAJOR(t->device), MINOR(t->device), act, rwbs);
1371 static void blk_log_dump_pdu(struct trace_seq *s,
1372 const struct trace_entry *ent, bool has_cg)
1374 const unsigned char *pdu_buf;
1378 pdu_buf = pdu_start(ent, has_cg);
1379 pdu_len = pdu_real_len(ent, has_cg);
1384 /* find the last zero that needs to be printed */
1385 for (end = pdu_len - 1; end >= 0; end--)
1390 trace_seq_putc(s, '(');
1392 for (i = 0; i < pdu_len; i++) {
1394 trace_seq_printf(s, "%s%02x",
1395 i == 0 ? "" : " ", pdu_buf[i]);
1398 * stop when the rest is just zeroes and indicate so
1399 * with a ".." appended
1401 if (i == end && end != pdu_len - 1) {
1402 trace_seq_puts(s, " ..) ");
1407 trace_seq_puts(s, ") ");
1410 static void blk_log_generic(struct trace_seq *s, const struct trace_entry *ent, bool has_cg)
1412 char cmd[TASK_COMM_LEN];
1414 trace_find_cmdline(ent->pid, cmd);
1416 if (t_action(ent) & BLK_TC_ACT(BLK_TC_PC)) {
1417 trace_seq_printf(s, "%u ", t_bytes(ent));
1418 blk_log_dump_pdu(s, ent, has_cg);
1419 trace_seq_printf(s, "[%s]\n", cmd);
1422 trace_seq_printf(s, "%llu + %u [%s]\n",
1423 t_sector(ent), t_sec(ent), cmd);
1425 trace_seq_printf(s, "[%s]\n", cmd);
1429 static void blk_log_with_error(struct trace_seq *s,
1430 const struct trace_entry *ent, bool has_cg)
1432 if (t_action(ent) & BLK_TC_ACT(BLK_TC_PC)) {
1433 blk_log_dump_pdu(s, ent, has_cg);
1434 trace_seq_printf(s, "[%d]\n", t_error(ent));
1437 trace_seq_printf(s, "%llu + %u [%d]\n",
1439 t_sec(ent), t_error(ent));
1441 trace_seq_printf(s, "%llu [%d]\n",
1442 t_sector(ent), t_error(ent));
1446 static void blk_log_remap(struct trace_seq *s, const struct trace_entry *ent, bool has_cg)
1448 const struct blk_io_trace_remap *__r = pdu_start(ent, has_cg);
1450 trace_seq_printf(s, "%llu + %u <- (%d,%d) %llu\n",
1451 t_sector(ent), t_sec(ent),
1452 MAJOR(be32_to_cpu(__r->device_from)),
1453 MINOR(be32_to_cpu(__r->device_from)),
1454 be64_to_cpu(__r->sector_from));
1457 static void blk_log_plug(struct trace_seq *s, const struct trace_entry *ent, bool has_cg)
1459 char cmd[TASK_COMM_LEN];
1461 trace_find_cmdline(ent->pid, cmd);
1463 trace_seq_printf(s, "[%s]\n", cmd);
1466 static void blk_log_unplug(struct trace_seq *s, const struct trace_entry *ent, bool has_cg)
1468 char cmd[TASK_COMM_LEN];
1470 trace_find_cmdline(ent->pid, cmd);
1472 trace_seq_printf(s, "[%s] %llu\n", cmd, get_pdu_int(ent, has_cg));
1475 static void blk_log_split(struct trace_seq *s, const struct trace_entry *ent, bool has_cg)
1477 char cmd[TASK_COMM_LEN];
1479 trace_find_cmdline(ent->pid, cmd);
1481 trace_seq_printf(s, "%llu / %llu [%s]\n", t_sector(ent),
1482 get_pdu_int(ent, has_cg), cmd);
1485 static void blk_log_msg(struct trace_seq *s, const struct trace_entry *ent,
1489 trace_seq_putmem(s, pdu_start(ent, has_cg),
1490 pdu_real_len(ent, has_cg));
1491 trace_seq_putc(s, '\n');
1495 * struct tracer operations
1498 static void blk_tracer_print_header(struct seq_file *m)
1500 if (!(blk_tracer_flags.val & TRACE_BLK_OPT_CLASSIC))
1502 seq_puts(m, "# DEV CPU TIMESTAMP PID ACT FLG\n"
1506 static void blk_tracer_start(struct trace_array *tr)
1508 blk_tracer_enabled = true;
1511 static int blk_tracer_init(struct trace_array *tr)
1514 blk_tracer_start(tr);
1518 static void blk_tracer_stop(struct trace_array *tr)
1520 blk_tracer_enabled = false;
1523 static void blk_tracer_reset(struct trace_array *tr)
1525 blk_tracer_stop(tr);
1528 static const struct {
1530 void (*print)(struct trace_seq *s, const struct trace_entry *ent,
1533 [__BLK_TA_QUEUE] = {{ "Q", "queue" }, blk_log_generic },
1534 [__BLK_TA_BACKMERGE] = {{ "M", "backmerge" }, blk_log_generic },
1535 [__BLK_TA_FRONTMERGE] = {{ "F", "frontmerge" }, blk_log_generic },
1536 [__BLK_TA_GETRQ] = {{ "G", "getrq" }, blk_log_generic },
1537 [__BLK_TA_SLEEPRQ] = {{ "S", "sleeprq" }, blk_log_generic },
1538 [__BLK_TA_REQUEUE] = {{ "R", "requeue" }, blk_log_with_error },
1539 [__BLK_TA_ISSUE] = {{ "D", "issue" }, blk_log_generic },
1540 [__BLK_TA_COMPLETE] = {{ "C", "complete" }, blk_log_with_error },
1541 [__BLK_TA_PLUG] = {{ "P", "plug" }, blk_log_plug },
1542 [__BLK_TA_UNPLUG_IO] = {{ "U", "unplug_io" }, blk_log_unplug },
1543 [__BLK_TA_UNPLUG_TIMER] = {{ "UT", "unplug_timer" }, blk_log_unplug },
1544 [__BLK_TA_INSERT] = {{ "I", "insert" }, blk_log_generic },
1545 [__BLK_TA_SPLIT] = {{ "X", "split" }, blk_log_split },
1546 [__BLK_TA_BOUNCE] = {{ "B", "bounce" }, blk_log_generic },
1547 [__BLK_TA_REMAP] = {{ "A", "remap" }, blk_log_remap },
1550 static enum print_line_t print_one_line(struct trace_iterator *iter,
1553 struct trace_array *tr = iter->tr;
1554 struct trace_seq *s = &iter->seq;
1555 const struct blk_io_trace *t;
1558 blk_log_action_t *log_action;
1561 t = te_blk_io_trace(iter->ent);
1562 what = (t->action & ((1 << BLK_TC_SHIFT) - 1)) & ~__BLK_TA_CGROUP;
1563 long_act = !!(tr->trace_flags & TRACE_ITER_VERBOSE);
1564 log_action = classic ? &blk_log_action_classic : &blk_log_action;
1565 has_cg = t->action & __BLK_TA_CGROUP;
1567 if ((t->action & ~__BLK_TN_CGROUP) == BLK_TN_MESSAGE) {
1568 log_action(iter, long_act ? "message" : "m", has_cg);
1569 blk_log_msg(s, iter->ent, has_cg);
1570 return trace_handle_return(s);
1573 if (unlikely(what == 0 || what >= ARRAY_SIZE(what2act)))
1574 trace_seq_printf(s, "Unknown action %x\n", what);
1576 log_action(iter, what2act[what].act[long_act], has_cg);
1577 what2act[what].print(s, iter->ent, has_cg);
1580 return trace_handle_return(s);
1583 static enum print_line_t blk_trace_event_print(struct trace_iterator *iter,
1584 int flags, struct trace_event *event)
1586 return print_one_line(iter, false);
1589 static void blk_trace_synthesize_old_trace(struct trace_iterator *iter)
1591 struct trace_seq *s = &iter->seq;
1592 struct blk_io_trace *t = (struct blk_io_trace *)iter->ent;
1593 const int offset = offsetof(struct blk_io_trace, sector);
1594 struct blk_io_trace old = {
1595 .magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION,
1599 trace_seq_putmem(s, &old, offset);
1600 trace_seq_putmem(s, &t->sector,
1601 sizeof(old) - offset + t->pdu_len);
1604 static enum print_line_t
1605 blk_trace_event_print_binary(struct trace_iterator *iter, int flags,
1606 struct trace_event *event)
1608 blk_trace_synthesize_old_trace(iter);
1610 return trace_handle_return(&iter->seq);
1613 static enum print_line_t blk_tracer_print_line(struct trace_iterator *iter)
1615 if (!(blk_tracer_flags.val & TRACE_BLK_OPT_CLASSIC))
1616 return TRACE_TYPE_UNHANDLED;
1618 return print_one_line(iter, true);
1622 blk_tracer_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set)
1624 /* don't output context-info for blk_classic output */
1625 if (bit == TRACE_BLK_OPT_CLASSIC) {
1627 tr->trace_flags &= ~TRACE_ITER_CONTEXT_INFO;
1629 tr->trace_flags |= TRACE_ITER_CONTEXT_INFO;
1634 static struct tracer blk_tracer __read_mostly = {
1636 .init = blk_tracer_init,
1637 .reset = blk_tracer_reset,
1638 .start = blk_tracer_start,
1639 .stop = blk_tracer_stop,
1640 .print_header = blk_tracer_print_header,
1641 .print_line = blk_tracer_print_line,
1642 .flags = &blk_tracer_flags,
1643 .set_flag = blk_tracer_set_flag,
1646 static struct trace_event_functions trace_blk_event_funcs = {
1647 .trace = blk_trace_event_print,
1648 .binary = blk_trace_event_print_binary,
1651 static struct trace_event trace_blk_event = {
1653 .funcs = &trace_blk_event_funcs,
1656 static int __init init_blk_tracer(void)
1658 if (!register_trace_event(&trace_blk_event)) {
1659 pr_warn("Warning: could not register block events\n");
1663 if (register_tracer(&blk_tracer) != 0) {
1664 pr_warn("Warning: could not register the block tracer\n");
1665 unregister_trace_event(&trace_blk_event);
1672 device_initcall(init_blk_tracer);
1674 static int blk_trace_remove_queue(struct request_queue *q)
1676 struct blk_trace *bt;
1678 bt = xchg(&q->blk_trace, NULL);
1682 if (bt->trace_state == Blktrace_running) {
1683 bt->trace_state = Blktrace_stopped;
1684 spin_lock_irq(&running_trace_lock);
1685 list_del_init(&bt->running_list);
1686 spin_unlock_irq(&running_trace_lock);
1687 relay_flush(bt->rchan);
1697 * Setup everything required to start tracing
1699 static int blk_trace_setup_queue(struct request_queue *q,
1700 struct block_device *bdev)
1702 struct blk_trace *bt = NULL;
1705 bt = kzalloc(sizeof(*bt), GFP_KERNEL);
1709 bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG, __alignof__(char));
1713 bt->dev = bdev->bd_dev;
1714 bt->act_mask = (u16)-1;
1716 blk_trace_setup_lba(bt, bdev);
1719 if (cmpxchg(&q->blk_trace, NULL, bt))
1731 * sysfs interface to enable and configure tracing
1734 static ssize_t sysfs_blk_trace_attr_show(struct device *dev,
1735 struct device_attribute *attr,
1737 static ssize_t sysfs_blk_trace_attr_store(struct device *dev,
1738 struct device_attribute *attr,
1739 const char *buf, size_t count);
1740 #define BLK_TRACE_DEVICE_ATTR(_name) \
1741 DEVICE_ATTR(_name, S_IRUGO | S_IWUSR, \
1742 sysfs_blk_trace_attr_show, \
1743 sysfs_blk_trace_attr_store)
1745 static BLK_TRACE_DEVICE_ATTR(enable);
1746 static BLK_TRACE_DEVICE_ATTR(act_mask);
1747 static BLK_TRACE_DEVICE_ATTR(pid);
1748 static BLK_TRACE_DEVICE_ATTR(start_lba);
1749 static BLK_TRACE_DEVICE_ATTR(end_lba);
1751 static struct attribute *blk_trace_attrs[] = {
1752 &dev_attr_enable.attr,
1753 &dev_attr_act_mask.attr,
1755 &dev_attr_start_lba.attr,
1756 &dev_attr_end_lba.attr,
1760 struct attribute_group blk_trace_attr_group = {
1762 .attrs = blk_trace_attrs,
1765 static const struct {
1769 { BLK_TC_READ, "read" },
1770 { BLK_TC_WRITE, "write" },
1771 { BLK_TC_FLUSH, "flush" },
1772 { BLK_TC_SYNC, "sync" },
1773 { BLK_TC_QUEUE, "queue" },
1774 { BLK_TC_REQUEUE, "requeue" },
1775 { BLK_TC_ISSUE, "issue" },
1776 { BLK_TC_COMPLETE, "complete" },
1777 { BLK_TC_FS, "fs" },
1778 { BLK_TC_PC, "pc" },
1779 { BLK_TC_NOTIFY, "notify" },
1780 { BLK_TC_AHEAD, "ahead" },
1781 { BLK_TC_META, "meta" },
1782 { BLK_TC_DISCARD, "discard" },
1783 { BLK_TC_DRV_DATA, "drv_data" },
1784 { BLK_TC_FUA, "fua" },
1787 static int blk_trace_str2mask(const char *str)
1791 char *buf, *s, *token;
1793 buf = kstrdup(str, GFP_KERNEL);
1799 token = strsep(&s, ",");
1806 for (i = 0; i < ARRAY_SIZE(mask_maps); i++) {
1807 if (strcasecmp(token, mask_maps[i].str) == 0) {
1808 mask |= mask_maps[i].mask;
1812 if (i == ARRAY_SIZE(mask_maps)) {
1822 static ssize_t blk_trace_mask2str(char *buf, int mask)
1827 for (i = 0; i < ARRAY_SIZE(mask_maps); i++) {
1828 if (mask & mask_maps[i].mask) {
1829 p += sprintf(p, "%s%s",
1830 (p == buf) ? "" : ",", mask_maps[i].str);
1838 static struct request_queue *blk_trace_get_queue(struct block_device *bdev)
1840 if (bdev->bd_disk == NULL)
1843 return bdev_get_queue(bdev);
1846 static ssize_t sysfs_blk_trace_attr_show(struct device *dev,
1847 struct device_attribute *attr,
1850 struct hd_struct *p = dev_to_part(dev);
1851 struct request_queue *q;
1852 struct block_device *bdev;
1853 struct blk_trace *bt;
1854 ssize_t ret = -ENXIO;
1856 bdev = bdget(part_devt(p));
1860 q = blk_trace_get_queue(bdev);
1864 mutex_lock(&q->blk_trace_mutex);
1866 bt = rcu_dereference_protected(q->blk_trace,
1867 lockdep_is_held(&q->blk_trace_mutex));
1868 if (attr == &dev_attr_enable) {
1869 ret = sprintf(buf, "%u\n", !!bt);
1870 goto out_unlock_bdev;
1874 ret = sprintf(buf, "disabled\n");
1875 else if (attr == &dev_attr_act_mask)
1876 ret = blk_trace_mask2str(buf, bt->act_mask);
1877 else if (attr == &dev_attr_pid)
1878 ret = sprintf(buf, "%u\n", bt->pid);
1879 else if (attr == &dev_attr_start_lba)
1880 ret = sprintf(buf, "%llu\n", bt->start_lba);
1881 else if (attr == &dev_attr_end_lba)
1882 ret = sprintf(buf, "%llu\n", bt->end_lba);
1885 mutex_unlock(&q->blk_trace_mutex);
1892 static ssize_t sysfs_blk_trace_attr_store(struct device *dev,
1893 struct device_attribute *attr,
1894 const char *buf, size_t count)
1896 struct block_device *bdev;
1897 struct request_queue *q;
1898 struct hd_struct *p;
1899 struct blk_trace *bt;
1901 ssize_t ret = -EINVAL;
1906 if (attr == &dev_attr_act_mask) {
1907 if (kstrtoull(buf, 0, &value)) {
1908 /* Assume it is a list of trace category names */
1909 ret = blk_trace_str2mask(buf);
1914 } else if (kstrtoull(buf, 0, &value))
1919 p = dev_to_part(dev);
1920 bdev = bdget(part_devt(p));
1924 q = blk_trace_get_queue(bdev);
1928 mutex_lock(&q->blk_trace_mutex);
1930 bt = rcu_dereference_protected(q->blk_trace,
1931 lockdep_is_held(&q->blk_trace_mutex));
1932 if (attr == &dev_attr_enable) {
1933 if (!!value == !!bt) {
1935 goto out_unlock_bdev;
1938 ret = blk_trace_setup_queue(q, bdev);
1940 ret = blk_trace_remove_queue(q);
1941 goto out_unlock_bdev;
1946 ret = blk_trace_setup_queue(q, bdev);
1947 bt = rcu_dereference_protected(q->blk_trace,
1948 lockdep_is_held(&q->blk_trace_mutex));
1952 if (attr == &dev_attr_act_mask)
1953 bt->act_mask = value;
1954 else if (attr == &dev_attr_pid)
1956 else if (attr == &dev_attr_start_lba)
1957 bt->start_lba = value;
1958 else if (attr == &dev_attr_end_lba)
1959 bt->end_lba = value;
1963 mutex_unlock(&q->blk_trace_mutex);
1967 return ret ? ret : count;
1970 int blk_trace_init_sysfs(struct device *dev)
1972 return sysfs_create_group(&dev->kobj, &blk_trace_attr_group);
1975 void blk_trace_remove_sysfs(struct device *dev)
1977 sysfs_remove_group(&dev->kobj, &blk_trace_attr_group);
1980 #endif /* CONFIG_BLK_DEV_IO_TRACE */
1982 #ifdef CONFIG_EVENT_TRACING
1984 void blk_fill_rwbs(char *rwbs, unsigned int op, int bytes)
1988 if (op & REQ_PREFLUSH)
1991 switch (op & REQ_OP_MASK) {
1993 case REQ_OP_WRITE_SAME:
1996 case REQ_OP_DISCARD:
1999 case REQ_OP_SECURE_ERASE:
2015 if (op & REQ_RAHEAD)
2024 EXPORT_SYMBOL_GPL(blk_fill_rwbs);
2026 #endif /* CONFIG_EVENT_TRACING */