blob: 3eb159c277c85a7a9d30f3b94d1a68bfa52661b0 [file] [log] [blame]
Jens Axboe2056a782006-03-23 20:00:26 +01001/*
Jens Axboe0fe23472006-09-04 15:41:16 +02002 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
Jens Axboe2056a782006-03-23 20:00:26 +01003 *
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.
7 *
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.
12 *
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
16 *
17 */
Jens Axboe2056a782006-03-23 20:00:26 +010018#include <linux/kernel.h>
19#include <linux/blkdev.h>
20#include <linux/blktrace_api.h>
21#include <linux/percpu.h>
22#include <linux/init.h>
23#include <linux/mutex.h>
24#include <linux/debugfs.h>
Alexey Dobriyan405f5572009-07-11 22:08:37 +040025#include <linux/smp_lock.h>
Olaf Kirchbe1c6342006-12-01 10:39:12 +010026#include <linux/time.h>
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -020027#include <linux/uaccess.h>
Li Zefan55782132009-06-09 13:43:05 +080028
29#include <trace/events/block.h>
30
Frederic Weisbecker2db270a2009-02-07 20:46:45 +010031#include "trace_output.h"
Jens Axboe2056a782006-03-23 20:00:26 +010032
Li Zefan55782132009-06-09 13:43:05 +080033#ifdef CONFIG_BLK_DEV_IO_TRACE
34
Jens Axboe2056a782006-03-23 20:00:26 +010035static unsigned int blktrace_seq __read_mostly = 1;
36
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -020037static struct trace_array *blk_tr;
Li Zefan5006ea72009-03-20 09:48:03 +080038static bool blk_tracer_enabled __read_mostly;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -020039
40/* Select an alternative, minimalistic output than the original one */
Steven Rostedtef180122009-03-10 14:10:56 -040041#define TRACE_BLK_OPT_CLASSIC 0x1
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -020042
43static struct tracer_opt blk_tracer_opts[] = {
44 /* Default disable the minimalistic output */
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -020045 { TRACER_OPT(blk_classic, TRACE_BLK_OPT_CLASSIC) },
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -020046 { }
47};
48
49static struct tracer_flags blk_tracer_flags = {
50 .val = 0,
51 .opts = blk_tracer_opts,
52};
53
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +010054/* Global reference count of probes */
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +010055static atomic_t blk_probes_ref = ATOMIC_INIT(0);
56
Li Zefan3c289ba2009-03-20 09:48:26 +080057static void blk_register_tracepoints(void);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +010058static void blk_unregister_tracepoints(void);
59
Jens Axboe2056a782006-03-23 20:00:26 +010060/*
Olaf Kirchbe1c6342006-12-01 10:39:12 +010061 * Send out a notify message.
62 */
Jens Axboea8630552006-12-04 09:30:58 +010063static void trace_note(struct blk_trace *bt, pid_t pid, int action,
64 const void *data, size_t len)
Olaf Kirchbe1c6342006-12-01 10:39:12 +010065{
66 struct blk_io_trace *t;
Li Zefan18cea452009-03-27 10:21:54 +080067 struct ring_buffer_event *event = NULL;
Steven Rostedte77405a2009-09-02 14:17:06 -040068 struct ring_buffer *buffer = NULL;
Li Zefan18cea452009-03-27 10:21:54 +080069 int pc = 0;
70 int cpu = smp_processor_id();
71 bool blk_tracer = blk_tracer_enabled;
72
73 if (blk_tracer) {
Steven Rostedte77405a2009-09-02 14:17:06 -040074 buffer = blk_tr->buffer;
Li Zefan18cea452009-03-27 10:21:54 +080075 pc = preempt_count();
Steven Rostedte77405a2009-09-02 14:17:06 -040076 event = trace_buffer_lock_reserve(buffer, TRACE_BLK,
Li Zefan18cea452009-03-27 10:21:54 +080077 sizeof(*t) + len,
78 0, pc);
79 if (!event)
80 return;
81 t = ring_buffer_event_data(event);
82 goto record_it;
83 }
Olaf Kirchbe1c6342006-12-01 10:39:12 +010084
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -020085 if (!bt->rchan)
86 return;
87
Olaf Kirchbe1c6342006-12-01 10:39:12 +010088 t = relay_reserve(bt->rchan, sizeof(*t) + len);
Jens Axboed3d9d2a2006-12-04 09:27:41 +010089 if (t) {
Jens Axboed3d9d2a2006-12-04 09:27:41 +010090 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
Ingo Molnar2997c8c2008-01-11 13:35:54 +010091 t->time = ktime_to_ns(ktime_get());
Li Zefan18cea452009-03-27 10:21:54 +080092record_it:
Jens Axboed3d9d2a2006-12-04 09:27:41 +010093 t->device = bt->dev;
94 t->action = action;
95 t->pid = pid;
96 t->cpu = cpu;
97 t->pdu_len = len;
98 memcpy((void *) t + sizeof(*t), data, len);
Li Zefan18cea452009-03-27 10:21:54 +080099
100 if (blk_tracer)
Steven Rostedte77405a2009-09-02 14:17:06 -0400101 trace_buffer_unlock_commit(buffer, event, 0, pc);
Jens Axboed3d9d2a2006-12-04 09:27:41 +0100102 }
Olaf Kirchbe1c6342006-12-01 10:39:12 +0100103}
104
105/*
Jens Axboe2056a782006-03-23 20:00:26 +0100106 * Send out a notify for this process, if we haven't done so since a trace
107 * started
108 */
109static void trace_note_tsk(struct blk_trace *bt, struct task_struct *tsk)
110{
Jens Axboea8630552006-12-04 09:30:58 +0100111 tsk->btrace_seq = blktrace_seq;
112 trace_note(bt, tsk->pid, BLK_TN_PROCESS, tsk->comm, sizeof(tsk->comm));
Olaf Kirchbe1c6342006-12-01 10:39:12 +0100113}
Jens Axboe2056a782006-03-23 20:00:26 +0100114
Olaf Kirchbe1c6342006-12-01 10:39:12 +0100115static void trace_note_time(struct blk_trace *bt)
116{
117 struct timespec now;
118 unsigned long flags;
119 u32 words[2];
120
121 getnstimeofday(&now);
122 words[0] = now.tv_sec;
123 words[1] = now.tv_nsec;
124
125 local_irq_save(flags);
126 trace_note(bt, 0, BLK_TN_TIMESTAMP, words, sizeof(words));
127 local_irq_restore(flags);
Jens Axboe2056a782006-03-23 20:00:26 +0100128}
129
Alan D. Brunelle9d5f09a2008-05-27 14:54:41 +0200130void __trace_note_message(struct blk_trace *bt, const char *fmt, ...)
131{
132 int n;
133 va_list args;
Carl Henrik Lunde14a73f52008-06-12 20:13:58 +0200134 unsigned long flags;
Jens Axboe64565912008-05-28 14:45:33 +0200135 char *buf;
Alan D. Brunelle9d5f09a2008-05-27 14:54:41 +0200136
Li Zefan18cea452009-03-27 10:21:54 +0800137 if (unlikely(bt->trace_state != Blktrace_running &&
138 !blk_tracer_enabled))
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200139 return;
140
Carl Henrik Lunde14a73f52008-06-12 20:13:58 +0200141 local_irq_save(flags);
Jens Axboe64565912008-05-28 14:45:33 +0200142 buf = per_cpu_ptr(bt->msg_data, smp_processor_id());
Alan D. Brunelle9d5f09a2008-05-27 14:54:41 +0200143 va_start(args, fmt);
Jens Axboe64565912008-05-28 14:45:33 +0200144 n = vscnprintf(buf, BLK_TN_MAX_MSG, fmt, args);
Alan D. Brunelle9d5f09a2008-05-27 14:54:41 +0200145 va_end(args);
146
Jens Axboe64565912008-05-28 14:45:33 +0200147 trace_note(bt, 0, BLK_TN_MESSAGE, buf, n);
Carl Henrik Lunde14a73f52008-06-12 20:13:58 +0200148 local_irq_restore(flags);
Alan D. Brunelle9d5f09a2008-05-27 14:54:41 +0200149}
150EXPORT_SYMBOL_GPL(__trace_note_message);
151
Jens Axboe2056a782006-03-23 20:00:26 +0100152static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
153 pid_t pid)
154{
155 if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
156 return 1;
Shawn Dud0deef52009-04-14 13:58:56 +0800157 if (sector && (sector < bt->start_lba || sector > bt->end_lba))
Jens Axboe2056a782006-03-23 20:00:26 +0100158 return 1;
159 if (bt->pid && pid != bt->pid)
160 return 1;
161
162 return 0;
163}
164
165/*
166 * Data direction bit lookup
167 */
Li Zefane4955c92009-03-24 16:04:37 +0800168static const u32 ddir_act[2] = { BLK_TC_ACT(BLK_TC_READ),
169 BLK_TC_ACT(BLK_TC_WRITE) };
Jens Axboe2056a782006-03-23 20:00:26 +0100170
David Woodhouse35ba8f72008-08-10 12:33:00 +0100171/* The ilog2() calls fall out because they're constant */
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200172#define MASK_TC_BIT(rw, __name) ((rw & (1 << BIO_RW_ ## __name)) << \
173 (ilog2(BLK_TC_ ## __name) + BLK_TC_SHIFT - BIO_RW_ ## __name))
Jens Axboe2056a782006-03-23 20:00:26 +0100174
175/*
176 * The worker for the various blk_add_trace*() types. Fills out a
177 * blk_io_trace structure and places it in a per-cpu subbuffer.
178 */
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100179static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
Jens Axboe2056a782006-03-23 20:00:26 +0100180 int rw, u32 what, int error, int pdu_len, void *pdu_data)
181{
182 struct task_struct *tsk = current;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200183 struct ring_buffer_event *event = NULL;
Steven Rostedte77405a2009-09-02 14:17:06 -0400184 struct ring_buffer *buffer = NULL;
Jens Axboe2056a782006-03-23 20:00:26 +0100185 struct blk_io_trace *t;
Arnaldo Carvalho de Melo0a987752009-02-05 16:12:56 -0200186 unsigned long flags = 0;
Jens Axboe2056a782006-03-23 20:00:26 +0100187 unsigned long *sequence;
188 pid_t pid;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200189 int cpu, pc = 0;
Li Zefan18cea452009-03-27 10:21:54 +0800190 bool blk_tracer = blk_tracer_enabled;
Jens Axboe2056a782006-03-23 20:00:26 +0100191
Li Zefan18cea452009-03-27 10:21:54 +0800192 if (unlikely(bt->trace_state != Blktrace_running && !blk_tracer))
Jens Axboe2056a782006-03-23 20:00:26 +0100193 return;
194
195 what |= ddir_act[rw & WRITE];
David Woodhouse35ba8f72008-08-10 12:33:00 +0100196 what |= MASK_TC_BIT(rw, BARRIER);
Ingo Molnar72c26c92009-02-19 09:00:35 +0100197 what |= MASK_TC_BIT(rw, SYNCIO);
David Woodhouse35ba8f72008-08-10 12:33:00 +0100198 what |= MASK_TC_BIT(rw, AHEAD);
199 what |= MASK_TC_BIT(rw, META);
200 what |= MASK_TC_BIT(rw, DISCARD);
Jens Axboe2056a782006-03-23 20:00:26 +0100201
202 pid = tsk->pid;
Shawn Dud0deef52009-04-14 13:58:56 +0800203 if (act_log_check(bt, what, sector, pid))
Jens Axboe2056a782006-03-23 20:00:26 +0100204 return;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200205 cpu = raw_smp_processor_id();
206
Li Zefan18cea452009-03-27 10:21:54 +0800207 if (blk_tracer) {
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200208 tracing_record_cmdline(current);
209
Steven Rostedte77405a2009-09-02 14:17:06 -0400210 buffer = blk_tr->buffer;
Arnaldo Carvalho de Melo51a763d2009-02-05 16:14:13 -0200211 pc = preempt_count();
Steven Rostedte77405a2009-09-02 14:17:06 -0400212 event = trace_buffer_lock_reserve(buffer, TRACE_BLK,
Arnaldo Carvalho de Melo51a763d2009-02-05 16:14:13 -0200213 sizeof(*t) + pdu_len,
214 0, pc);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200215 if (!event)
216 return;
Arnaldo Carvalho de Melo51a763d2009-02-05 16:14:13 -0200217 t = ring_buffer_event_data(event);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200218 goto record_it;
219 }
Jens Axboe2056a782006-03-23 20:00:26 +0100220
221 /*
222 * A word about the locking here - we disable interrupts to reserve
223 * some space in the relay per-cpu buffer, to prevent an irq
Carl Henrik Lunde14a73f52008-06-12 20:13:58 +0200224 * from coming in and stepping on our toes.
Jens Axboe2056a782006-03-23 20:00:26 +0100225 */
226 local_irq_save(flags);
227
228 if (unlikely(tsk->btrace_seq != blktrace_seq))
229 trace_note_tsk(bt, tsk);
230
231 t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
232 if (t) {
Jens Axboe2056a782006-03-23 20:00:26 +0100233 sequence = per_cpu_ptr(bt->sequence, cpu);
234
235 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
236 t->sequence = ++(*sequence);
Ingo Molnar2997c8c2008-01-11 13:35:54 +0100237 t->time = ktime_to_ns(ktime_get());
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200238record_it:
Arnaldo Carvalho de Melo08a06b82009-02-02 20:30:40 -0200239 /*
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200240 * These two are not needed in ftrace as they are in the
241 * generic trace_entry, filled by tracing_generic_entry_update,
242 * but for the trace_event->bin() synthesizer benefit we do it
243 * here too.
244 */
245 t->cpu = cpu;
246 t->pid = pid;
Arnaldo Carvalho de Melo08a06b82009-02-02 20:30:40 -0200247
Jens Axboe2056a782006-03-23 20:00:26 +0100248 t->sector = sector;
249 t->bytes = bytes;
250 t->action = what;
Jens Axboe2056a782006-03-23 20:00:26 +0100251 t->device = bt->dev;
Jens Axboe2056a782006-03-23 20:00:26 +0100252 t->error = error;
253 t->pdu_len = pdu_len;
254
255 if (pdu_len)
256 memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200257
Li Zefan18cea452009-03-27 10:21:54 +0800258 if (blk_tracer) {
Steven Rostedte77405a2009-09-02 14:17:06 -0400259 trace_buffer_unlock_commit(buffer, event, 0, pc);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200260 return;
261 }
Jens Axboe2056a782006-03-23 20:00:26 +0100262 }
263
264 local_irq_restore(flags);
265}
266
Jens Axboe2056a782006-03-23 20:00:26 +0100267static struct dentry *blk_tree_root;
Jens Axboe11a57152008-01-11 13:37:01 +0100268static DEFINE_MUTEX(blk_tree_mutex);
Jens Axboe2056a782006-03-23 20:00:26 +0100269
Li Zefanad5dd542009-03-27 10:20:24 +0800270static void blk_trace_free(struct blk_trace *bt)
Jens Axboe2056a782006-03-23 20:00:26 +0100271{
Alan D. Brunelle02c62302008-06-11 09:12:52 +0200272 debugfs_remove(bt->msg_file);
Jens Axboe2056a782006-03-23 20:00:26 +0100273 debugfs_remove(bt->dropped_file);
Jens Axboef48fc4d2009-01-05 10:17:25 +0100274 relay_close(bt->rchan);
Alan D. Brunelle39cbb602009-08-07 12:01:08 -0400275 debugfs_remove(bt->dir);
Jens Axboe2056a782006-03-23 20:00:26 +0100276 free_percpu(bt->sequence);
Jens Axboe64565912008-05-28 14:45:33 +0200277 free_percpu(bt->msg_data);
Jens Axboe2056a782006-03-23 20:00:26 +0100278 kfree(bt);
Li Zefanad5dd542009-03-27 10:20:24 +0800279}
280
281static void blk_trace_cleanup(struct blk_trace *bt)
282{
283 blk_trace_free(bt);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100284 if (atomic_dec_and_test(&blk_probes_ref))
285 blk_unregister_tracepoints();
Jens Axboe2056a782006-03-23 20:00:26 +0100286}
287
Christof Schmitt6da127a2008-01-11 10:09:43 +0100288int blk_trace_remove(struct request_queue *q)
Jens Axboe2056a782006-03-23 20:00:26 +0100289{
290 struct blk_trace *bt;
291
292 bt = xchg(&q->blk_trace, NULL);
293 if (!bt)
294 return -EINVAL;
295
Li Zefan55547202009-03-25 17:21:26 +0800296 if (bt->trace_state != Blktrace_running)
Jens Axboe2056a782006-03-23 20:00:26 +0100297 blk_trace_cleanup(bt);
298
299 return 0;
300}
Christof Schmitt6da127a2008-01-11 10:09:43 +0100301EXPORT_SYMBOL_GPL(blk_trace_remove);
Jens Axboe2056a782006-03-23 20:00:26 +0100302
303static int blk_dropped_open(struct inode *inode, struct file *filp)
304{
Theodore Ts'o8e18e292006-09-27 01:50:46 -0700305 filp->private_data = inode->i_private;
Jens Axboe2056a782006-03-23 20:00:26 +0100306
307 return 0;
308}
309
310static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
311 size_t count, loff_t *ppos)
312{
313 struct blk_trace *bt = filp->private_data;
314 char buf[16];
315
316 snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
317
318 return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
319}
320
Arjan van de Ven2b8693c2007-02-12 00:55:32 -0800321static const struct file_operations blk_dropped_fops = {
Jens Axboe2056a782006-03-23 20:00:26 +0100322 .owner = THIS_MODULE,
323 .open = blk_dropped_open,
324 .read = blk_dropped_read,
325};
326
Alan D. Brunelle02c62302008-06-11 09:12:52 +0200327static int blk_msg_open(struct inode *inode, struct file *filp)
328{
329 filp->private_data = inode->i_private;
330
331 return 0;
332}
333
334static ssize_t blk_msg_write(struct file *filp, const char __user *buffer,
335 size_t count, loff_t *ppos)
336{
337 char *msg;
338 struct blk_trace *bt;
339
Li Zefan7635b032009-04-03 15:31:34 +0800340 if (count >= BLK_TN_MAX_MSG)
Alan D. Brunelle02c62302008-06-11 09:12:52 +0200341 return -EINVAL;
342
Carl Henrik Lundea4b3ada2009-04-03 14:27:15 +0200343 msg = kmalloc(count + 1, GFP_KERNEL);
Alan D. Brunelle02c62302008-06-11 09:12:52 +0200344 if (msg == NULL)
345 return -ENOMEM;
346
347 if (copy_from_user(msg, buffer, count)) {
348 kfree(msg);
349 return -EFAULT;
350 }
351
Carl Henrik Lundea4b3ada2009-04-03 14:27:15 +0200352 msg[count] = '\0';
Alan D. Brunelle02c62302008-06-11 09:12:52 +0200353 bt = filp->private_data;
354 __trace_note_message(bt, "%s", msg);
355 kfree(msg);
356
357 return count;
358}
359
360static const struct file_operations blk_msg_fops = {
361 .owner = THIS_MODULE,
362 .open = blk_msg_open,
363 .write = blk_msg_write,
364};
365
Jens Axboe2056a782006-03-23 20:00:26 +0100366/*
367 * Keep track of how many times we encountered a full subbuffer, to aid
368 * the user space app in telling how many lost events there were.
369 */
370static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
371 void *prev_subbuf, size_t prev_padding)
372{
373 struct blk_trace *bt;
374
375 if (!relay_buf_full(buf))
376 return 1;
377
378 bt = buf->chan->private_data;
379 atomic_inc(&bt->dropped);
380 return 0;
381}
382
383static int blk_remove_buf_file_callback(struct dentry *dentry)
384{
385 debugfs_remove(dentry);
Jens Axboef48fc4d2009-01-05 10:17:25 +0100386
Jens Axboe2056a782006-03-23 20:00:26 +0100387 return 0;
388}
389
390static struct dentry *blk_create_buf_file_callback(const char *filename,
391 struct dentry *parent,
392 int mode,
393 struct rchan_buf *buf,
394 int *is_global)
395{
396 return debugfs_create_file(filename, mode, parent, buf,
397 &relay_file_operations);
398}
399
400static struct rchan_callbacks blk_relay_callbacks = {
401 .subbuf_start = blk_subbuf_start_callback,
402 .create_buf_file = blk_create_buf_file_callback,
403 .remove_buf_file = blk_remove_buf_file_callback,
404};
405
Li Zefan9908c302009-04-14 13:59:34 +0800406static void blk_trace_setup_lba(struct blk_trace *bt,
407 struct block_device *bdev)
408{
409 struct hd_struct *part = NULL;
410
411 if (bdev)
412 part = bdev->bd_part;
413
414 if (part) {
415 bt->start_lba = part->start_sect;
416 bt->end_lba = part->start_sect + part->nr_sects;
417 } else {
418 bt->start_lba = 0;
419 bt->end_lba = -1ULL;
420 }
421}
422
Jens Axboe2056a782006-03-23 20:00:26 +0100423/*
424 * Setup everything required to start tracing
425 */
Christof Schmitt6da127a2008-01-11 10:09:43 +0100426int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
Shawn Dud0deef52009-04-14 13:58:56 +0800427 struct block_device *bdev,
428 struct blk_user_trace_setup *buts)
Jens Axboe2056a782006-03-23 20:00:26 +0100429{
Jens Axboe2056a782006-03-23 20:00:26 +0100430 struct blk_trace *old_bt, *bt = NULL;
431 struct dentry *dir = NULL;
Jens Axboe2056a782006-03-23 20:00:26 +0100432 int ret, i;
433
Arnd Bergmann171044d42007-10-09 13:23:53 +0200434 if (!buts->buf_size || !buts->buf_nr)
Jens Axboe2056a782006-03-23 20:00:26 +0100435 return -EINVAL;
436
Jens Axboe0497b342008-10-01 16:16:25 +0200437 strncpy(buts->name, name, BLKTRACE_BDEV_SIZE);
438 buts->name[BLKTRACE_BDEV_SIZE - 1] = '\0';
Jens Axboe2056a782006-03-23 20:00:26 +0100439
440 /*
441 * some device names have larger paths - convert the slashes
442 * to underscores for this to work as expected
443 */
Arnd Bergmann171044d42007-10-09 13:23:53 +0200444 for (i = 0; i < strlen(buts->name); i++)
445 if (buts->name[i] == '/')
446 buts->name[i] = '_';
Jens Axboe2056a782006-03-23 20:00:26 +0100447
Jens Axboe2056a782006-03-23 20:00:26 +0100448 bt = kzalloc(sizeof(*bt), GFP_KERNEL);
449 if (!bt)
Li Zefanad5dd542009-03-27 10:20:24 +0800450 return -ENOMEM;
Jens Axboe2056a782006-03-23 20:00:26 +0100451
Li Zefanad5dd542009-03-27 10:20:24 +0800452 ret = -ENOMEM;
Jens Axboe2056a782006-03-23 20:00:26 +0100453 bt->sequence = alloc_percpu(unsigned long);
454 if (!bt->sequence)
455 goto err;
456
Ingo Molnarf0ef0392009-03-06 16:44:14 +0100457 bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG, __alignof__(char));
Jens Axboe64565912008-05-28 14:45:33 +0200458 if (!bt->msg_data)
459 goto err;
460
Jens Axboe2056a782006-03-23 20:00:26 +0100461 ret = -ENOENT;
Jens Axboef48fc4d2009-01-05 10:17:25 +0100462
Li Zefanb5230b52009-03-25 17:19:33 +0800463 mutex_lock(&blk_tree_mutex);
Jens Axboef48fc4d2009-01-05 10:17:25 +0100464 if (!blk_tree_root) {
465 blk_tree_root = debugfs_create_dir("block", NULL);
Li Zefanb5230b52009-03-25 17:19:33 +0800466 if (!blk_tree_root) {
467 mutex_unlock(&blk_tree_mutex);
Li Zefan1a176622009-03-20 09:47:30 +0800468 goto err;
Li Zefanb5230b52009-03-25 17:19:33 +0800469 }
Jens Axboef48fc4d2009-01-05 10:17:25 +0100470 }
Li Zefanb5230b52009-03-25 17:19:33 +0800471 mutex_unlock(&blk_tree_mutex);
Jens Axboef48fc4d2009-01-05 10:17:25 +0100472
473 dir = debugfs_create_dir(buts->name, blk_tree_root);
474
Jens Axboe2056a782006-03-23 20:00:26 +0100475 if (!dir)
476 goto err;
477
478 bt->dir = dir;
Christof Schmitt6da127a2008-01-11 10:09:43 +0100479 bt->dev = dev;
Jens Axboe2056a782006-03-23 20:00:26 +0100480 atomic_set(&bt->dropped, 0);
481
482 ret = -EIO;
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200483 bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt,
484 &blk_dropped_fops);
Jens Axboe2056a782006-03-23 20:00:26 +0100485 if (!bt->dropped_file)
486 goto err;
487
Alan D. Brunelle02c62302008-06-11 09:12:52 +0200488 bt->msg_file = debugfs_create_file("msg", 0222, dir, bt, &blk_msg_fops);
489 if (!bt->msg_file)
490 goto err;
491
Arnd Bergmann171044d42007-10-09 13:23:53 +0200492 bt->rchan = relay_open("trace", dir, buts->buf_size,
493 buts->buf_nr, &blk_relay_callbacks, bt);
Jens Axboe2056a782006-03-23 20:00:26 +0100494 if (!bt->rchan)
495 goto err;
Jens Axboe2056a782006-03-23 20:00:26 +0100496
Arnd Bergmann171044d42007-10-09 13:23:53 +0200497 bt->act_mask = buts->act_mask;
Jens Axboe2056a782006-03-23 20:00:26 +0100498 if (!bt->act_mask)
499 bt->act_mask = (u16) -1;
500
Li Zefan9908c302009-04-14 13:59:34 +0800501 blk_trace_setup_lba(bt, bdev);
Jens Axboe2056a782006-03-23 20:00:26 +0100502
Shawn Dud0deef52009-04-14 13:58:56 +0800503 /* overwrite with user settings */
504 if (buts->start_lba)
505 bt->start_lba = buts->start_lba;
506 if (buts->end_lba)
507 bt->end_lba = buts->end_lba;
508
Arnd Bergmann171044d42007-10-09 13:23:53 +0200509 bt->pid = buts->pid;
Jens Axboe2056a782006-03-23 20:00:26 +0100510 bt->trace_state = Blktrace_setup;
511
512 ret = -EBUSY;
513 old_bt = xchg(&q->blk_trace, bt);
514 if (old_bt) {
515 (void) xchg(&q->blk_trace, old_bt);
516 goto err;
517 }
518
Li Zefan17ba97e2009-03-27 10:20:09 +0800519 if (atomic_inc_return(&blk_probes_ref) == 1)
Li Zefancbe28292009-03-20 09:48:47 +0800520 blk_register_tracepoints();
521
Jens Axboe2056a782006-03-23 20:00:26 +0100522 return 0;
523err:
Li Zefanad5dd542009-03-27 10:20:24 +0800524 blk_trace_free(bt);
Jens Axboe2056a782006-03-23 20:00:26 +0100525 return ret;
526}
527
Christof Schmitt6da127a2008-01-11 10:09:43 +0100528int blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
Shawn Dud0deef52009-04-14 13:58:56 +0800529 struct block_device *bdev,
Christof Schmitt6da127a2008-01-11 10:09:43 +0100530 char __user *arg)
Arnd Bergmann171044d42007-10-09 13:23:53 +0200531{
532 struct blk_user_trace_setup buts;
533 int ret;
534
535 ret = copy_from_user(&buts, arg, sizeof(buts));
536 if (ret)
537 return -EFAULT;
538
Shawn Dud0deef52009-04-14 13:58:56 +0800539 ret = do_blk_trace_setup(q, name, dev, bdev, &buts);
Arnd Bergmann171044d42007-10-09 13:23:53 +0200540 if (ret)
541 return ret;
542
543 if (copy_to_user(arg, &buts, sizeof(buts)))
544 return -EFAULT;
545
546 return 0;
547}
Christof Schmitt6da127a2008-01-11 10:09:43 +0100548EXPORT_SYMBOL_GPL(blk_trace_setup);
Arnd Bergmann171044d42007-10-09 13:23:53 +0200549
Christof Schmitt6da127a2008-01-11 10:09:43 +0100550int blk_trace_startstop(struct request_queue *q, int start)
Jens Axboe2056a782006-03-23 20:00:26 +0100551{
Jens Axboe2056a782006-03-23 20:00:26 +0100552 int ret;
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200553 struct blk_trace *bt = q->blk_trace;
Jens Axboe2056a782006-03-23 20:00:26 +0100554
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200555 if (bt == NULL)
Jens Axboe2056a782006-03-23 20:00:26 +0100556 return -EINVAL;
557
558 /*
559 * For starting a trace, we can transition from a setup or stopped
560 * trace. For stopping a trace, the state must be running
561 */
562 ret = -EINVAL;
563 if (start) {
564 if (bt->trace_state == Blktrace_setup ||
565 bt->trace_state == Blktrace_stopped) {
566 blktrace_seq++;
567 smp_mb();
568 bt->trace_state = Blktrace_running;
Olaf Kirchbe1c6342006-12-01 10:39:12 +0100569
570 trace_note_time(bt);
Jens Axboe2056a782006-03-23 20:00:26 +0100571 ret = 0;
572 }
573 } else {
574 if (bt->trace_state == Blktrace_running) {
575 bt->trace_state = Blktrace_stopped;
576 relay_flush(bt->rchan);
577 ret = 0;
578 }
579 }
580
581 return ret;
582}
Christof Schmitt6da127a2008-01-11 10:09:43 +0100583EXPORT_SYMBOL_GPL(blk_trace_startstop);
Jens Axboe2056a782006-03-23 20:00:26 +0100584
585/**
586 * blk_trace_ioctl: - handle the ioctls associated with tracing
587 * @bdev: the block device
Steven Rostedtef180122009-03-10 14:10:56 -0400588 * @cmd: the ioctl cmd
Jens Axboe2056a782006-03-23 20:00:26 +0100589 * @arg: the argument data, if any
590 *
591 **/
592int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
593{
Jens Axboe165125e2007-07-24 09:28:11 +0200594 struct request_queue *q;
Jens Axboe2056a782006-03-23 20:00:26 +0100595 int ret, start = 0;
Christof Schmitt6da127a2008-01-11 10:09:43 +0100596 char b[BDEVNAME_SIZE];
Jens Axboe2056a782006-03-23 20:00:26 +0100597
598 q = bdev_get_queue(bdev);
599 if (!q)
600 return -ENXIO;
601
602 mutex_lock(&bdev->bd_mutex);
603
604 switch (cmd) {
605 case BLKTRACESETUP:
Jean Delvaref36f21e2008-05-12 14:02:33 -0700606 bdevname(bdev, b);
Shawn Dud0deef52009-04-14 13:58:56 +0800607 ret = blk_trace_setup(q, b, bdev->bd_dev, bdev, arg);
Jens Axboe2056a782006-03-23 20:00:26 +0100608 break;
609 case BLKTRACESTART:
610 start = 1;
611 case BLKTRACESTOP:
612 ret = blk_trace_startstop(q, start);
613 break;
614 case BLKTRACETEARDOWN:
615 ret = blk_trace_remove(q);
616 break;
617 default:
618 ret = -ENOTTY;
619 break;
620 }
621
622 mutex_unlock(&bdev->bd_mutex);
623 return ret;
624}
625
626/**
627 * blk_trace_shutdown: - stop and cleanup trace structures
628 * @q: the request queue associated with the device
629 *
630 **/
Jens Axboe165125e2007-07-24 09:28:11 +0200631void blk_trace_shutdown(struct request_queue *q)
Jens Axboe2056a782006-03-23 20:00:26 +0100632{
Alexey Dobriyan6c5c9342006-09-29 01:59:40 -0700633 if (q->blk_trace) {
634 blk_trace_startstop(q, 0);
635 blk_trace_remove(q);
636 }
Jens Axboe2056a782006-03-23 20:00:26 +0100637}
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100638
639/*
640 * blktrace probes
641 */
642
643/**
644 * blk_add_trace_rq - Add a trace for a request oriented action
645 * @q: queue the io is for
646 * @rq: the source request
647 * @what: the action
648 *
649 * Description:
650 * Records an action against a request. Will log the bio offset + size.
651 *
652 **/
653static void blk_add_trace_rq(struct request_queue *q, struct request *rq,
654 u32 what)
655{
656 struct blk_trace *bt = q->blk_trace;
657 int rw = rq->cmd_flags & 0x03;
658
659 if (likely(!bt))
660 return;
661
662 if (blk_discard_rq(rq))
663 rw |= (1 << BIO_RW_DISCARD);
664
665 if (blk_pc_request(rq)) {
666 what |= BLK_TC_ACT(BLK_TC_PC);
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900667 __blk_add_trace(bt, 0, blk_rq_bytes(rq), rw,
668 what, rq->errors, rq->cmd_len, rq->cmd);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100669 } else {
670 what |= BLK_TC_ACT(BLK_TC_FS);
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900671 __blk_add_trace(bt, blk_rq_pos(rq), blk_rq_bytes(rq), rw,
672 what, rq->errors, 0, NULL);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100673 }
674}
675
676static void blk_add_trace_rq_abort(struct request_queue *q, struct request *rq)
677{
678 blk_add_trace_rq(q, rq, BLK_TA_ABORT);
679}
680
681static void blk_add_trace_rq_insert(struct request_queue *q, struct request *rq)
682{
683 blk_add_trace_rq(q, rq, BLK_TA_INSERT);
684}
685
686static void blk_add_trace_rq_issue(struct request_queue *q, struct request *rq)
687{
688 blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
689}
690
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200691static void blk_add_trace_rq_requeue(struct request_queue *q,
692 struct request *rq)
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100693{
694 blk_add_trace_rq(q, rq, BLK_TA_REQUEUE);
695}
696
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200697static void blk_add_trace_rq_complete(struct request_queue *q,
698 struct request *rq)
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100699{
700 blk_add_trace_rq(q, rq, BLK_TA_COMPLETE);
701}
702
703/**
704 * blk_add_trace_bio - Add a trace for a bio oriented action
705 * @q: queue the io is for
706 * @bio: the source bio
707 * @what: the action
708 *
709 * Description:
710 * Records an action against a bio. Will log the bio offset + size.
711 *
712 **/
713static void blk_add_trace_bio(struct request_queue *q, struct bio *bio,
714 u32 what)
715{
716 struct blk_trace *bt = q->blk_trace;
717
718 if (likely(!bt))
719 return;
720
721 __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw, what,
722 !bio_flagged(bio, BIO_UPTODATE), 0, NULL);
723}
724
725static void blk_add_trace_bio_bounce(struct request_queue *q, struct bio *bio)
726{
727 blk_add_trace_bio(q, bio, BLK_TA_BOUNCE);
728}
729
730static void blk_add_trace_bio_complete(struct request_queue *q, struct bio *bio)
731{
732 blk_add_trace_bio(q, bio, BLK_TA_COMPLETE);
733}
734
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200735static void blk_add_trace_bio_backmerge(struct request_queue *q,
736 struct bio *bio)
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100737{
738 blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE);
739}
740
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200741static void blk_add_trace_bio_frontmerge(struct request_queue *q,
742 struct bio *bio)
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100743{
744 blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE);
745}
746
747static void blk_add_trace_bio_queue(struct request_queue *q, struct bio *bio)
748{
749 blk_add_trace_bio(q, bio, BLK_TA_QUEUE);
750}
751
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200752static void blk_add_trace_getrq(struct request_queue *q,
753 struct bio *bio, int rw)
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100754{
755 if (bio)
756 blk_add_trace_bio(q, bio, BLK_TA_GETRQ);
757 else {
758 struct blk_trace *bt = q->blk_trace;
759
760 if (bt)
761 __blk_add_trace(bt, 0, 0, rw, BLK_TA_GETRQ, 0, 0, NULL);
762 }
763}
764
765
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200766static void blk_add_trace_sleeprq(struct request_queue *q,
767 struct bio *bio, int rw)
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100768{
769 if (bio)
770 blk_add_trace_bio(q, bio, BLK_TA_SLEEPRQ);
771 else {
772 struct blk_trace *bt = q->blk_trace;
773
774 if (bt)
Arnaldo Carvalho de Melo939b3662009-02-03 11:58:29 -0200775 __blk_add_trace(bt, 0, 0, rw, BLK_TA_SLEEPRQ,
776 0, 0, NULL);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100777 }
778}
779
780static void blk_add_trace_plug(struct request_queue *q)
781{
782 struct blk_trace *bt = q->blk_trace;
783
784 if (bt)
785 __blk_add_trace(bt, 0, 0, 0, BLK_TA_PLUG, 0, 0, NULL);
786}
787
788static void blk_add_trace_unplug_io(struct request_queue *q)
789{
790 struct blk_trace *bt = q->blk_trace;
791
792 if (bt) {
793 unsigned int pdu = q->rq.count[READ] + q->rq.count[WRITE];
794 __be64 rpdu = cpu_to_be64(pdu);
795
796 __blk_add_trace(bt, 0, 0, 0, BLK_TA_UNPLUG_IO, 0,
797 sizeof(rpdu), &rpdu);
798 }
799}
800
801static void blk_add_trace_unplug_timer(struct request_queue *q)
802{
803 struct blk_trace *bt = q->blk_trace;
804
805 if (bt) {
806 unsigned int pdu = q->rq.count[READ] + q->rq.count[WRITE];
807 __be64 rpdu = cpu_to_be64(pdu);
808
809 __blk_add_trace(bt, 0, 0, 0, BLK_TA_UNPLUG_TIMER, 0,
810 sizeof(rpdu), &rpdu);
811 }
812}
813
814static void blk_add_trace_split(struct request_queue *q, struct bio *bio,
815 unsigned int pdu)
816{
817 struct blk_trace *bt = q->blk_trace;
818
819 if (bt) {
820 __be64 rpdu = cpu_to_be64(pdu);
821
822 __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw,
823 BLK_TA_SPLIT, !bio_flagged(bio, BIO_UPTODATE),
824 sizeof(rpdu), &rpdu);
825 }
826}
827
828/**
829 * blk_add_trace_remap - Add a trace for a remap operation
830 * @q: queue the io is for
831 * @bio: the source bio
832 * @dev: target device
Alan D. Brunellea42aaa32009-05-04 16:27:26 -0400833 * @from: source sector
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100834 *
835 * Description:
836 * Device mapper or raid target sometimes need to split a bio because
837 * it spans a stripe (or similar). Add a trace for that action.
838 *
839 **/
840static void blk_add_trace_remap(struct request_queue *q, struct bio *bio,
Alan D. Brunelle22a7c312009-05-04 16:35:08 -0400841 dev_t dev, sector_t from)
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100842{
843 struct blk_trace *bt = q->blk_trace;
844 struct blk_io_trace_remap r;
845
846 if (likely(!bt))
847 return;
848
Alan D. Brunellea42aaa32009-05-04 16:27:26 -0400849 r.device_from = cpu_to_be32(dev);
850 r.device_to = cpu_to_be32(bio->bi_bdev->bd_dev);
851 r.sector_from = cpu_to_be64(from);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100852
Alan D. Brunelle22a7c312009-05-04 16:35:08 -0400853 __blk_add_trace(bt, bio->bi_sector, bio->bi_size, bio->bi_rw,
854 BLK_TA_REMAP, !bio_flagged(bio, BIO_UPTODATE),
855 sizeof(r), &r);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100856}
857
858/**
859 * blk_add_driver_data - Add binary message with driver-specific data
860 * @q: queue the io is for
861 * @rq: io request
862 * @data: driver-specific data
863 * @len: length of driver-specific data
864 *
865 * Description:
866 * Some drivers might want to write driver-specific data per request.
867 *
868 **/
869void blk_add_driver_data(struct request_queue *q,
870 struct request *rq,
871 void *data, size_t len)
872{
873 struct blk_trace *bt = q->blk_trace;
874
875 if (likely(!bt))
876 return;
877
878 if (blk_pc_request(rq))
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900879 __blk_add_trace(bt, 0, blk_rq_bytes(rq), 0,
880 BLK_TA_DRV_DATA, rq->errors, len, data);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100881 else
Tejun Heo2e46e8b2009-05-07 22:24:41 +0900882 __blk_add_trace(bt, blk_rq_pos(rq), blk_rq_bytes(rq), 0,
883 BLK_TA_DRV_DATA, rq->errors, len, data);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100884}
885EXPORT_SYMBOL_GPL(blk_add_driver_data);
886
Li Zefan3c289ba2009-03-20 09:48:26 +0800887static void blk_register_tracepoints(void)
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100888{
889 int ret;
890
891 ret = register_trace_block_rq_abort(blk_add_trace_rq_abort);
892 WARN_ON(ret);
893 ret = register_trace_block_rq_insert(blk_add_trace_rq_insert);
894 WARN_ON(ret);
895 ret = register_trace_block_rq_issue(blk_add_trace_rq_issue);
896 WARN_ON(ret);
897 ret = register_trace_block_rq_requeue(blk_add_trace_rq_requeue);
898 WARN_ON(ret);
899 ret = register_trace_block_rq_complete(blk_add_trace_rq_complete);
900 WARN_ON(ret);
901 ret = register_trace_block_bio_bounce(blk_add_trace_bio_bounce);
902 WARN_ON(ret);
903 ret = register_trace_block_bio_complete(blk_add_trace_bio_complete);
904 WARN_ON(ret);
905 ret = register_trace_block_bio_backmerge(blk_add_trace_bio_backmerge);
906 WARN_ON(ret);
907 ret = register_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge);
908 WARN_ON(ret);
909 ret = register_trace_block_bio_queue(blk_add_trace_bio_queue);
910 WARN_ON(ret);
911 ret = register_trace_block_getrq(blk_add_trace_getrq);
912 WARN_ON(ret);
913 ret = register_trace_block_sleeprq(blk_add_trace_sleeprq);
914 WARN_ON(ret);
915 ret = register_trace_block_plug(blk_add_trace_plug);
916 WARN_ON(ret);
917 ret = register_trace_block_unplug_timer(blk_add_trace_unplug_timer);
918 WARN_ON(ret);
919 ret = register_trace_block_unplug_io(blk_add_trace_unplug_io);
920 WARN_ON(ret);
921 ret = register_trace_block_split(blk_add_trace_split);
922 WARN_ON(ret);
923 ret = register_trace_block_remap(blk_add_trace_remap);
924 WARN_ON(ret);
Arnaldo Carvalho de Melo5f3ea372008-10-30 08:34:33 +0100925}
926
927static void blk_unregister_tracepoints(void)
928{
929 unregister_trace_block_remap(blk_add_trace_remap);
930 unregister_trace_block_split(blk_add_trace_split);
931 unregister_trace_block_unplug_io(blk_add_trace_unplug_io);
932 unregister_trace_block_unplug_timer(blk_add_trace_unplug_timer);
933 unregister_trace_block_plug(blk_add_trace_plug);
934 unregister_trace_block_sleeprq(blk_add_trace_sleeprq);
935 unregister_trace_block_getrq(blk_add_trace_getrq);
936 unregister_trace_block_bio_queue(blk_add_trace_bio_queue);
937 unregister_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge);
938 unregister_trace_block_bio_backmerge(blk_add_trace_bio_backmerge);
939 unregister_trace_block_bio_complete(blk_add_trace_bio_complete);
940 unregister_trace_block_bio_bounce(blk_add_trace_bio_bounce);
941 unregister_trace_block_rq_complete(blk_add_trace_rq_complete);
942 unregister_trace_block_rq_requeue(blk_add_trace_rq_requeue);
943 unregister_trace_block_rq_issue(blk_add_trace_rq_issue);
944 unregister_trace_block_rq_insert(blk_add_trace_rq_insert);
945 unregister_trace_block_rq_abort(blk_add_trace_rq_abort);
946
947 tracepoint_synchronize_unregister();
948}
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200949
950/*
951 * struct blk_io_tracer formatting routines
952 */
953
954static void fill_rwbs(char *rwbs, const struct blk_io_trace *t)
955{
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -0200956 int i = 0;
Li Zefan65796342009-03-24 16:05:06 +0800957 int tc = t->action >> BLK_TC_SHIFT;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200958
Li Zefan18cea452009-03-27 10:21:54 +0800959 if (t->action == BLK_TN_MESSAGE) {
960 rwbs[i++] = 'N';
961 goto out;
962 }
963
Li Zefan65796342009-03-24 16:05:06 +0800964 if (tc & BLK_TC_DISCARD)
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -0200965 rwbs[i++] = 'D';
Li Zefan65796342009-03-24 16:05:06 +0800966 else if (tc & BLK_TC_WRITE)
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -0200967 rwbs[i++] = 'W';
968 else if (t->bytes)
969 rwbs[i++] = 'R';
970 else
971 rwbs[i++] = 'N';
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200972
Li Zefan65796342009-03-24 16:05:06 +0800973 if (tc & BLK_TC_AHEAD)
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -0200974 rwbs[i++] = 'A';
Li Zefan65796342009-03-24 16:05:06 +0800975 if (tc & BLK_TC_BARRIER)
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -0200976 rwbs[i++] = 'B';
Li Zefan65796342009-03-24 16:05:06 +0800977 if (tc & BLK_TC_SYNC)
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -0200978 rwbs[i++] = 'S';
Li Zefan65796342009-03-24 16:05:06 +0800979 if (tc & BLK_TC_META)
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -0200980 rwbs[i++] = 'M';
Li Zefan18cea452009-03-27 10:21:54 +0800981out:
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -0200982 rwbs[i] = '\0';
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -0200983}
984
985static inline
986const struct blk_io_trace *te_blk_io_trace(const struct trace_entry *ent)
987{
988 return (const struct blk_io_trace *)ent;
989}
990
991static inline const void *pdu_start(const struct trace_entry *ent)
992{
993 return te_blk_io_trace(ent) + 1;
994}
995
Li Zefan66de7792009-04-01 16:19:19 +0800996static inline u32 t_action(const struct trace_entry *ent)
997{
998 return te_blk_io_trace(ent)->action;
999}
1000
1001static inline u32 t_bytes(const struct trace_entry *ent)
1002{
1003 return te_blk_io_trace(ent)->bytes;
1004}
1005
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001006static inline u32 t_sec(const struct trace_entry *ent)
1007{
1008 return te_blk_io_trace(ent)->bytes >> 9;
1009}
1010
1011static inline unsigned long long t_sector(const struct trace_entry *ent)
1012{
1013 return te_blk_io_trace(ent)->sector;
1014}
1015
1016static inline __u16 t_error(const struct trace_entry *ent)
1017{
Li Zefane0dc81b2009-03-24 16:05:51 +08001018 return te_blk_io_trace(ent)->error;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001019}
1020
1021static __u64 get_pdu_int(const struct trace_entry *ent)
1022{
1023 const __u64 *val = pdu_start(ent);
1024 return be64_to_cpu(*val);
1025}
1026
1027static void get_pdu_remap(const struct trace_entry *ent,
1028 struct blk_io_trace_remap *r)
1029{
1030 const struct blk_io_trace_remap *__r = pdu_start(ent);
Alan D. Brunellea42aaa32009-05-04 16:27:26 -04001031 __u64 sector_from = __r->sector_from;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001032
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001033 r->device_from = be32_to_cpu(__r->device_from);
Alan D. Brunellea42aaa32009-05-04 16:27:26 -04001034 r->device_to = be32_to_cpu(__r->device_to);
1035 r->sector_from = be64_to_cpu(sector_from);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001036}
1037
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001038typedef int (blk_log_action_t) (struct trace_iterator *iter, const char *act);
1039
1040static int blk_log_action_classic(struct trace_iterator *iter, const char *act)
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001041{
1042 char rwbs[6];
Li Zefan35ac51b2009-03-27 10:19:46 +08001043 unsigned long long ts = iter->ts;
1044 unsigned long nsec_rem = do_div(ts, NSEC_PER_SEC);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001045 unsigned secs = (unsigned long)ts;
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001046 const struct blk_io_trace *t = te_blk_io_trace(iter->ent);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001047
1048 fill_rwbs(rwbs, t);
1049
1050 return trace_seq_printf(&iter->seq,
Li Zefan35ac51b2009-03-27 10:19:46 +08001051 "%3d,%-3d %2d %5d.%09lu %5u %2s %3s ",
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001052 MAJOR(t->device), MINOR(t->device), iter->cpu,
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001053 secs, nsec_rem, iter->ent->pid, act, rwbs);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001054}
1055
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001056static int blk_log_action(struct trace_iterator *iter, const char *act)
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001057{
1058 char rwbs[6];
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001059 const struct blk_io_trace *t = te_blk_io_trace(iter->ent);
1060
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001061 fill_rwbs(rwbs, t);
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001062 return trace_seq_printf(&iter->seq, "%3d,%-3d %2s %3s ",
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001063 MAJOR(t->device), MINOR(t->device), act, rwbs);
1064}
1065
Li Zefan66de7792009-04-01 16:19:19 +08001066static int blk_log_dump_pdu(struct trace_seq *s, const struct trace_entry *ent)
1067{
Li Zefan049862572009-05-11 14:33:23 +08001068 const unsigned char *pdu_buf;
Li Zefan66de7792009-04-01 16:19:19 +08001069 int pdu_len;
1070 int i, end, ret;
1071
1072 pdu_buf = pdu_start(ent);
1073 pdu_len = te_blk_io_trace(ent)->pdu_len;
1074
1075 if (!pdu_len)
1076 return 1;
1077
1078 /* find the last zero that needs to be printed */
1079 for (end = pdu_len - 1; end >= 0; end--)
1080 if (pdu_buf[end])
1081 break;
1082 end++;
1083
1084 if (!trace_seq_putc(s, '('))
1085 return 0;
1086
1087 for (i = 0; i < pdu_len; i++) {
1088
1089 ret = trace_seq_printf(s, "%s%02x",
1090 i == 0 ? "" : " ", pdu_buf[i]);
1091 if (!ret)
1092 return ret;
1093
1094 /*
1095 * stop when the rest is just zeroes and indicate so
1096 * with a ".." appended
1097 */
1098 if (i == end && end != pdu_len - 1)
1099 return trace_seq_puts(s, " ..) ");
1100 }
1101
1102 return trace_seq_puts(s, ") ");
1103}
1104
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001105static int blk_log_generic(struct trace_seq *s, const struct trace_entry *ent)
1106{
Steven Rostedt4ca530852009-03-16 19:20:15 -04001107 char cmd[TASK_COMM_LEN];
1108
1109 trace_find_cmdline(ent->pid, cmd);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001110
Li Zefan66de7792009-04-01 16:19:19 +08001111 if (t_action(ent) & BLK_TC_ACT(BLK_TC_PC)) {
1112 int ret;
1113
1114 ret = trace_seq_printf(s, "%u ", t_bytes(ent));
1115 if (!ret)
1116 return 0;
1117 ret = blk_log_dump_pdu(s, ent);
1118 if (!ret)
1119 return 0;
1120 return trace_seq_printf(s, "[%s]\n", cmd);
1121 } else {
1122 if (t_sec(ent))
1123 return trace_seq_printf(s, "%llu + %u [%s]\n",
1124 t_sector(ent), t_sec(ent), cmd);
1125 return trace_seq_printf(s, "[%s]\n", cmd);
1126 }
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001127}
1128
Arnaldo Carvalho de Melo157f9c02009-01-26 15:00:56 -02001129static int blk_log_with_error(struct trace_seq *s,
1130 const struct trace_entry *ent)
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001131{
Li Zefan66de7792009-04-01 16:19:19 +08001132 if (t_action(ent) & BLK_TC_ACT(BLK_TC_PC)) {
1133 int ret;
1134
1135 ret = blk_log_dump_pdu(s, ent);
1136 if (ret)
1137 return trace_seq_printf(s, "[%d]\n", t_error(ent));
1138 return 0;
1139 } else {
1140 if (t_sec(ent))
1141 return trace_seq_printf(s, "%llu + %u [%d]\n",
1142 t_sector(ent),
1143 t_sec(ent), t_error(ent));
1144 return trace_seq_printf(s, "%llu [%d]\n",
1145 t_sector(ent), t_error(ent));
1146 }
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001147}
1148
1149static int blk_log_remap(struct trace_seq *s, const struct trace_entry *ent)
1150{
Alan D. Brunellea42aaa32009-05-04 16:27:26 -04001151 struct blk_io_trace_remap r = { .device_from = 0, };
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001152
1153 get_pdu_remap(ent, &r);
1154 return trace_seq_printf(s, "%llu + %u <- (%d,%d) %llu\n",
Alan D. Brunellea42aaa32009-05-04 16:27:26 -04001155 t_sector(ent), t_sec(ent),
1156 MAJOR(r.device_from), MINOR(r.device_from),
1157 (unsigned long long)r.sector_from);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001158}
1159
1160static int blk_log_plug(struct trace_seq *s, const struct trace_entry *ent)
1161{
Steven Rostedt4ca530852009-03-16 19:20:15 -04001162 char cmd[TASK_COMM_LEN];
1163
1164 trace_find_cmdline(ent->pid, cmd);
1165
1166 return trace_seq_printf(s, "[%s]\n", cmd);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001167}
1168
1169static int blk_log_unplug(struct trace_seq *s, const struct trace_entry *ent)
1170{
Steven Rostedt4ca530852009-03-16 19:20:15 -04001171 char cmd[TASK_COMM_LEN];
1172
1173 trace_find_cmdline(ent->pid, cmd);
1174
1175 return trace_seq_printf(s, "[%s] %llu\n", cmd, get_pdu_int(ent));
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001176}
1177
1178static int blk_log_split(struct trace_seq *s, const struct trace_entry *ent)
1179{
Steven Rostedt4ca530852009-03-16 19:20:15 -04001180 char cmd[TASK_COMM_LEN];
1181
1182 trace_find_cmdline(ent->pid, cmd);
1183
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001184 return trace_seq_printf(s, "%llu / %llu [%s]\n", t_sector(ent),
Steven Rostedt4ca530852009-03-16 19:20:15 -04001185 get_pdu_int(ent), cmd);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001186}
1187
Li Zefan18cea452009-03-27 10:21:54 +08001188static int blk_log_msg(struct trace_seq *s, const struct trace_entry *ent)
1189{
1190 int ret;
1191 const struct blk_io_trace *t = te_blk_io_trace(ent);
1192
1193 ret = trace_seq_putmem(s, t + 1, t->pdu_len);
1194 if (ret)
1195 return trace_seq_putc(s, '\n');
1196 return ret;
1197}
1198
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001199/*
1200 * struct tracer operations
1201 */
1202
1203static void blk_tracer_print_header(struct seq_file *m)
1204{
1205 if (!(blk_tracer_flags.val & TRACE_BLK_OPT_CLASSIC))
1206 return;
1207 seq_puts(m, "# DEV CPU TIMESTAMP PID ACT FLG\n"
1208 "# | | | | | |\n");
1209}
1210
1211static void blk_tracer_start(struct trace_array *tr)
1212{
Li Zefanad5dd542009-03-27 10:20:24 +08001213 blk_tracer_enabled = true;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001214}
1215
1216static int blk_tracer_init(struct trace_array *tr)
1217{
1218 blk_tr = tr;
1219 blk_tracer_start(tr);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001220 return 0;
1221}
1222
1223static void blk_tracer_stop(struct trace_array *tr)
1224{
Li Zefanad5dd542009-03-27 10:20:24 +08001225 blk_tracer_enabled = false;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001226}
1227
1228static void blk_tracer_reset(struct trace_array *tr)
1229{
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001230 blk_tracer_stop(tr);
1231}
1232
Li Zefane4955c92009-03-24 16:04:37 +08001233static const struct {
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001234 const char *act[2];
Steven Rostedtef180122009-03-10 14:10:56 -04001235 int (*print)(struct trace_seq *s, const struct trace_entry *ent);
Li Zefane4955c92009-03-24 16:04:37 +08001236} what2act[] = {
Steven Rostedtef180122009-03-10 14:10:56 -04001237 [__BLK_TA_QUEUE] = {{ "Q", "queue" }, blk_log_generic },
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001238 [__BLK_TA_BACKMERGE] = {{ "M", "backmerge" }, blk_log_generic },
1239 [__BLK_TA_FRONTMERGE] = {{ "F", "frontmerge" }, blk_log_generic },
1240 [__BLK_TA_GETRQ] = {{ "G", "getrq" }, blk_log_generic },
1241 [__BLK_TA_SLEEPRQ] = {{ "S", "sleeprq" }, blk_log_generic },
1242 [__BLK_TA_REQUEUE] = {{ "R", "requeue" }, blk_log_with_error },
1243 [__BLK_TA_ISSUE] = {{ "D", "issue" }, blk_log_generic },
1244 [__BLK_TA_COMPLETE] = {{ "C", "complete" }, blk_log_with_error },
1245 [__BLK_TA_PLUG] = {{ "P", "plug" }, blk_log_plug },
1246 [__BLK_TA_UNPLUG_IO] = {{ "U", "unplug_io" }, blk_log_unplug },
1247 [__BLK_TA_UNPLUG_TIMER] = {{ "UT", "unplug_timer" }, blk_log_unplug },
1248 [__BLK_TA_INSERT] = {{ "I", "insert" }, blk_log_generic },
1249 [__BLK_TA_SPLIT] = {{ "X", "split" }, blk_log_split },
1250 [__BLK_TA_BOUNCE] = {{ "B", "bounce" }, blk_log_generic },
1251 [__BLK_TA_REMAP] = {{ "A", "remap" }, blk_log_remap },
1252};
1253
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001254static enum print_line_t print_one_line(struct trace_iterator *iter,
1255 bool classic)
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001256{
Arnaldo Carvalho de Melo2c9b2382009-02-02 20:30:12 -02001257 struct trace_seq *s = &iter->seq;
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001258 const struct blk_io_trace *t;
1259 u16 what;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001260 int ret;
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001261 bool long_act;
1262 blk_log_action_t *log_action;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001263
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001264 t = te_blk_io_trace(iter->ent);
1265 what = t->action & ((1 << BLK_TC_SHIFT) - 1);
1266 long_act = !!(trace_flags & TRACE_ITER_VERBOSE);
1267 log_action = classic ? &blk_log_action_classic : &blk_log_action;
Arnaldo Carvalho de Melo08a06b82009-02-02 20:30:40 -02001268
Li Zefan18cea452009-03-27 10:21:54 +08001269 if (t->action == BLK_TN_MESSAGE) {
1270 ret = log_action(iter, long_act ? "message" : "m");
1271 if (ret)
1272 ret = blk_log_msg(s, iter->ent);
1273 goto out;
1274 }
1275
Li Zefaneb08f8e2009-03-24 16:05:27 +08001276 if (unlikely(what == 0 || what >= ARRAY_SIZE(what2act)))
Li Zefanb78825d2009-04-01 16:18:53 +08001277 ret = trace_seq_printf(s, "Unknown action %x\n", what);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001278 else {
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001279 ret = log_action(iter, what2act[what].act[long_act]);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001280 if (ret)
Arnaldo Carvalho de Melo2c9b2382009-02-02 20:30:12 -02001281 ret = what2act[what].print(s, iter->ent);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001282 }
Li Zefan18cea452009-03-27 10:21:54 +08001283out:
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001284 return ret ? TRACE_TYPE_HANDLED : TRACE_TYPE_PARTIAL_LINE;
1285}
1286
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001287static enum print_line_t blk_trace_event_print(struct trace_iterator *iter,
1288 int flags)
1289{
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001290 return print_one_line(iter, false);
1291}
1292
Arnaldo Carvalho de Melo08a06b82009-02-02 20:30:40 -02001293static int blk_trace_synthesize_old_trace(struct trace_iterator *iter)
1294{
1295 struct trace_seq *s = &iter->seq;
1296 struct blk_io_trace *t = (struct blk_io_trace *)iter->ent;
1297 const int offset = offsetof(struct blk_io_trace, sector);
1298 struct blk_io_trace old = {
1299 .magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION,
Li Zefan6c051ce2009-03-25 17:18:56 +08001300 .time = iter->ts,
Arnaldo Carvalho de Melo08a06b82009-02-02 20:30:40 -02001301 };
1302
1303 if (!trace_seq_putmem(s, &old, offset))
1304 return 0;
1305 return trace_seq_putmem(s, &t->sector,
1306 sizeof(old) - offset + t->pdu_len);
1307}
1308
Arnaldo Carvalho de Meloae7462b2009-02-03 22:05:50 -02001309static enum print_line_t
1310blk_trace_event_print_binary(struct trace_iterator *iter, int flags)
Arnaldo Carvalho de Melo08a06b82009-02-02 20:30:40 -02001311{
1312 return blk_trace_synthesize_old_trace(iter) ?
1313 TRACE_TYPE_HANDLED : TRACE_TYPE_PARTIAL_LINE;
1314}
1315
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001316static enum print_line_t blk_tracer_print_line(struct trace_iterator *iter)
1317{
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001318 if (!(blk_tracer_flags.val & TRACE_BLK_OPT_CLASSIC))
1319 return TRACE_TYPE_UNHANDLED;
1320
Li Zefanb6a4b0c2009-03-27 10:21:23 +08001321 return print_one_line(iter, true);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001322}
1323
Li Zefanf3948f82009-04-15 11:02:56 +08001324static int blk_tracer_set_flag(u32 old_flags, u32 bit, int set)
1325{
1326 /* don't output context-info for blk_classic output */
1327 if (bit == TRACE_BLK_OPT_CLASSIC) {
1328 if (set)
1329 trace_flags &= ~TRACE_ITER_CONTEXT_INFO;
1330 else
1331 trace_flags |= TRACE_ITER_CONTEXT_INFO;
1332 }
1333 return 0;
1334}
1335
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001336static struct tracer blk_tracer __read_mostly = {
1337 .name = "blk",
1338 .init = blk_tracer_init,
1339 .reset = blk_tracer_reset,
1340 .start = blk_tracer_start,
1341 .stop = blk_tracer_stop,
1342 .print_header = blk_tracer_print_header,
1343 .print_line = blk_tracer_print_line,
1344 .flags = &blk_tracer_flags,
Li Zefanf3948f82009-04-15 11:02:56 +08001345 .set_flag = blk_tracer_set_flag,
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001346};
1347
1348static struct trace_event trace_blk_event = {
Steven Rostedtef180122009-03-10 14:10:56 -04001349 .type = TRACE_BLK,
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001350 .trace = blk_trace_event_print,
Arnaldo Carvalho de Melo08a06b82009-02-02 20:30:40 -02001351 .binary = blk_trace_event_print_binary,
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001352};
1353
1354static int __init init_blk_tracer(void)
1355{
1356 if (!register_ftrace_event(&trace_blk_event)) {
1357 pr_warning("Warning: could not register block events\n");
1358 return 1;
1359 }
1360
1361 if (register_tracer(&blk_tracer) != 0) {
1362 pr_warning("Warning: could not register the block tracer\n");
1363 unregister_ftrace_event(&trace_blk_event);
1364 return 1;
1365 }
1366
1367 return 0;
1368}
1369
1370device_initcall(init_blk_tracer);
1371
1372static int blk_trace_remove_queue(struct request_queue *q)
1373{
1374 struct blk_trace *bt;
1375
1376 bt = xchg(&q->blk_trace, NULL);
1377 if (bt == NULL)
1378 return -EINVAL;
1379
Li Zefan17ba97e2009-03-27 10:20:09 +08001380 if (atomic_dec_and_test(&blk_probes_ref))
1381 blk_unregister_tracepoints();
1382
Li Zefanad5dd542009-03-27 10:20:24 +08001383 blk_trace_free(bt);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001384 return 0;
1385}
1386
1387/*
1388 * Setup everything required to start tracing
1389 */
Li Zefan9908c302009-04-14 13:59:34 +08001390static int blk_trace_setup_queue(struct request_queue *q,
1391 struct block_device *bdev)
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001392{
1393 struct blk_trace *old_bt, *bt = NULL;
Li Zefan18cea452009-03-27 10:21:54 +08001394 int ret = -ENOMEM;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001395
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001396 bt = kzalloc(sizeof(*bt), GFP_KERNEL);
1397 if (!bt)
Li Zefan15152e42009-03-20 09:49:08 +08001398 return -ENOMEM;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001399
Li Zefan18cea452009-03-27 10:21:54 +08001400 bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG, __alignof__(char));
1401 if (!bt->msg_data)
1402 goto free_bt;
1403
Li Zefan9908c302009-04-14 13:59:34 +08001404 bt->dev = bdev->bd_dev;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001405 bt->act_mask = (u16)-1;
Li Zefan9908c302009-04-14 13:59:34 +08001406
1407 blk_trace_setup_lba(bt, bdev);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001408
1409 old_bt = xchg(&q->blk_trace, bt);
1410 if (old_bt != NULL) {
1411 (void)xchg(&q->blk_trace, old_bt);
Li Zefan18cea452009-03-27 10:21:54 +08001412 ret = -EBUSY;
1413 goto free_bt;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001414 }
Li Zefan15152e42009-03-20 09:49:08 +08001415
Li Zefan17ba97e2009-03-27 10:20:09 +08001416 if (atomic_inc_return(&blk_probes_ref) == 1)
1417 blk_register_tracepoints();
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001418 return 0;
Li Zefan18cea452009-03-27 10:21:54 +08001419
1420free_bt:
1421 blk_trace_free(bt);
1422 return ret;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001423}
1424
1425/*
1426 * sysfs interface to enable and configure tracing
1427 */
1428
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001429static ssize_t sysfs_blk_trace_attr_show(struct device *dev,
1430 struct device_attribute *attr,
1431 char *buf);
1432static ssize_t sysfs_blk_trace_attr_store(struct device *dev,
1433 struct device_attribute *attr,
1434 const char *buf, size_t count);
1435#define BLK_TRACE_DEVICE_ATTR(_name) \
1436 DEVICE_ATTR(_name, S_IRUGO | S_IWUSR, \
1437 sysfs_blk_trace_attr_show, \
1438 sysfs_blk_trace_attr_store)
1439
Li Zefancd649b82009-03-20 11:33:55 +08001440static BLK_TRACE_DEVICE_ATTR(enable);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001441static BLK_TRACE_DEVICE_ATTR(act_mask);
1442static BLK_TRACE_DEVICE_ATTR(pid);
1443static BLK_TRACE_DEVICE_ATTR(start_lba);
1444static BLK_TRACE_DEVICE_ATTR(end_lba);
1445
1446static struct attribute *blk_trace_attrs[] = {
1447 &dev_attr_enable.attr,
1448 &dev_attr_act_mask.attr,
1449 &dev_attr_pid.attr,
1450 &dev_attr_start_lba.attr,
1451 &dev_attr_end_lba.attr,
1452 NULL
1453};
1454
1455struct attribute_group blk_trace_attr_group = {
1456 .name = "trace",
1457 .attrs = blk_trace_attrs,
1458};
1459
Li Zefan09341992009-03-24 17:43:30 +08001460static const struct {
1461 int mask;
1462 const char *str;
1463} mask_maps[] = {
1464 { BLK_TC_READ, "read" },
1465 { BLK_TC_WRITE, "write" },
1466 { BLK_TC_BARRIER, "barrier" },
1467 { BLK_TC_SYNC, "sync" },
1468 { BLK_TC_QUEUE, "queue" },
1469 { BLK_TC_REQUEUE, "requeue" },
1470 { BLK_TC_ISSUE, "issue" },
1471 { BLK_TC_COMPLETE, "complete" },
1472 { BLK_TC_FS, "fs" },
1473 { BLK_TC_PC, "pc" },
1474 { BLK_TC_AHEAD, "ahead" },
1475 { BLK_TC_META, "meta" },
1476 { BLK_TC_DISCARD, "discard" },
1477 { BLK_TC_DRV_DATA, "drv_data" },
1478};
1479
1480static int blk_trace_str2mask(const char *str)
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001481{
Li Zefan09341992009-03-24 17:43:30 +08001482 int i;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001483 int mask = 0;
Li Zefan9eb851252009-04-09 11:19:40 +08001484 char *buf, *s, *token;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001485
Li Zefan9eb851252009-04-09 11:19:40 +08001486 buf = kstrdup(str, GFP_KERNEL);
1487 if (buf == NULL)
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001488 return -ENOMEM;
Li Zefan9eb851252009-04-09 11:19:40 +08001489 s = strstrip(buf);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001490
1491 while (1) {
Li Zefan09341992009-03-24 17:43:30 +08001492 token = strsep(&s, ",");
1493 if (token == NULL)
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001494 break;
1495
Li Zefan09341992009-03-24 17:43:30 +08001496 if (*token == '\0')
1497 continue;
1498
1499 for (i = 0; i < ARRAY_SIZE(mask_maps); i++) {
1500 if (strcasecmp(token, mask_maps[i].str) == 0) {
1501 mask |= mask_maps[i].mask;
1502 break;
1503 }
1504 }
1505 if (i == ARRAY_SIZE(mask_maps)) {
1506 mask = -EINVAL;
1507 break;
1508 }
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001509 }
Li Zefan9eb851252009-04-09 11:19:40 +08001510 kfree(buf);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001511
1512 return mask;
1513}
1514
Li Zefan09341992009-03-24 17:43:30 +08001515static ssize_t blk_trace_mask2str(char *buf, int mask)
1516{
1517 int i;
1518 char *p = buf;
1519
1520 for (i = 0; i < ARRAY_SIZE(mask_maps); i++) {
1521 if (mask & mask_maps[i].mask) {
1522 p += sprintf(p, "%s%s",
1523 (p == buf) ? "" : ",", mask_maps[i].str);
1524 }
1525 }
1526 *p++ = '\n';
1527
1528 return p - buf;
1529}
1530
Li Zefanb1251302009-03-20 10:34:00 +08001531static struct request_queue *blk_trace_get_queue(struct block_device *bdev)
1532{
1533 if (bdev->bd_disk == NULL)
1534 return NULL;
1535
1536 return bdev_get_queue(bdev);
1537}
1538
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001539static ssize_t sysfs_blk_trace_attr_show(struct device *dev,
1540 struct device_attribute *attr,
1541 char *buf)
1542{
1543 struct hd_struct *p = dev_to_part(dev);
1544 struct request_queue *q;
1545 struct block_device *bdev;
1546 ssize_t ret = -ENXIO;
1547
1548 lock_kernel();
1549 bdev = bdget(part_devt(p));
1550 if (bdev == NULL)
1551 goto out_unlock_kernel;
1552
Li Zefanb1251302009-03-20 10:34:00 +08001553 q = blk_trace_get_queue(bdev);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001554 if (q == NULL)
1555 goto out_bdput;
Li Zefanb1251302009-03-20 10:34:00 +08001556
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001557 mutex_lock(&bdev->bd_mutex);
Li Zefancd649b82009-03-20 11:33:55 +08001558
1559 if (attr == &dev_attr_enable) {
1560 ret = sprintf(buf, "%u\n", !!q->blk_trace);
1561 goto out_unlock_bdev;
1562 }
1563
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001564 if (q->blk_trace == NULL)
1565 ret = sprintf(buf, "disabled\n");
1566 else if (attr == &dev_attr_act_mask)
Li Zefan09341992009-03-24 17:43:30 +08001567 ret = blk_trace_mask2str(buf, q->blk_trace->act_mask);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001568 else if (attr == &dev_attr_pid)
1569 ret = sprintf(buf, "%u\n", q->blk_trace->pid);
1570 else if (attr == &dev_attr_start_lba)
1571 ret = sprintf(buf, "%llu\n", q->blk_trace->start_lba);
1572 else if (attr == &dev_attr_end_lba)
1573 ret = sprintf(buf, "%llu\n", q->blk_trace->end_lba);
Li Zefancd649b82009-03-20 11:33:55 +08001574
1575out_unlock_bdev:
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001576 mutex_unlock(&bdev->bd_mutex);
1577out_bdput:
1578 bdput(bdev);
1579out_unlock_kernel:
1580 unlock_kernel();
1581 return ret;
1582}
1583
1584static ssize_t sysfs_blk_trace_attr_store(struct device *dev,
1585 struct device_attribute *attr,
1586 const char *buf, size_t count)
1587{
1588 struct block_device *bdev;
1589 struct request_queue *q;
1590 struct hd_struct *p;
1591 u64 value;
Li Zefan09341992009-03-24 17:43:30 +08001592 ssize_t ret = -EINVAL;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001593
1594 if (count == 0)
1595 goto out;
1596
1597 if (attr == &dev_attr_act_mask) {
1598 if (sscanf(buf, "%llx", &value) != 1) {
1599 /* Assume it is a list of trace category names */
Li Zefan09341992009-03-24 17:43:30 +08001600 ret = blk_trace_str2mask(buf);
1601 if (ret < 0)
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001602 goto out;
Li Zefan09341992009-03-24 17:43:30 +08001603 value = ret;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001604 }
1605 } else if (sscanf(buf, "%llu", &value) != 1)
1606 goto out;
1607
Li Zefan09341992009-03-24 17:43:30 +08001608 ret = -ENXIO;
1609
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001610 lock_kernel();
1611 p = dev_to_part(dev);
1612 bdev = bdget(part_devt(p));
1613 if (bdev == NULL)
1614 goto out_unlock_kernel;
1615
Li Zefanb1251302009-03-20 10:34:00 +08001616 q = blk_trace_get_queue(bdev);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001617 if (q == NULL)
1618 goto out_bdput;
1619
1620 mutex_lock(&bdev->bd_mutex);
Li Zefancd649b82009-03-20 11:33:55 +08001621
1622 if (attr == &dev_attr_enable) {
1623 if (value)
Li Zefan9908c302009-04-14 13:59:34 +08001624 ret = blk_trace_setup_queue(q, bdev);
Li Zefancd649b82009-03-20 11:33:55 +08001625 else
1626 ret = blk_trace_remove_queue(q);
1627 goto out_unlock_bdev;
1628 }
1629
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001630 ret = 0;
1631 if (q->blk_trace == NULL)
Li Zefan9908c302009-04-14 13:59:34 +08001632 ret = blk_trace_setup_queue(q, bdev);
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001633
1634 if (ret == 0) {
1635 if (attr == &dev_attr_act_mask)
1636 q->blk_trace->act_mask = value;
1637 else if (attr == &dev_attr_pid)
1638 q->blk_trace->pid = value;
1639 else if (attr == &dev_attr_start_lba)
1640 q->blk_trace->start_lba = value;
1641 else if (attr == &dev_attr_end_lba)
1642 q->blk_trace->end_lba = value;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001643 }
Li Zefancd649b82009-03-20 11:33:55 +08001644
1645out_unlock_bdev:
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001646 mutex_unlock(&bdev->bd_mutex);
1647out_bdput:
1648 bdput(bdev);
1649out_unlock_kernel:
1650 unlock_kernel();
1651out:
Li Zefancd649b82009-03-20 11:33:55 +08001652 return ret ? ret : count;
Arnaldo Carvalho de Meloc71a8962009-01-23 12:06:27 -02001653}
Li Zefancd649b82009-03-20 11:33:55 +08001654
Li Zefan1d54ad62009-04-14 14:00:05 +08001655int blk_trace_init_sysfs(struct device *dev)
1656{
1657 return sysfs_create_group(&dev->kobj, &blk_trace_attr_group);
1658}
1659
Li Zefan55782132009-06-09 13:43:05 +08001660#endif /* CONFIG_BLK_DEV_IO_TRACE */
1661
1662#ifdef CONFIG_EVENT_TRACING
1663
1664void blk_dump_cmd(char *buf, struct request *rq)
1665{
1666 int i, end;
1667 int len = rq->cmd_len;
1668 unsigned char *cmd = rq->cmd;
1669
1670 if (!blk_pc_request(rq)) {
1671 buf[0] = '\0';
1672 return;
1673 }
1674
1675 for (end = len - 1; end >= 0; end--)
1676 if (cmd[end])
1677 break;
1678 end++;
1679
1680 for (i = 0; i < len; i++) {
1681 buf += sprintf(buf, "%s%02x", i == 0 ? "" : " ", cmd[i]);
1682 if (i == end && end != len - 1) {
1683 sprintf(buf, " ..");
1684 break;
1685 }
1686 }
1687}
1688
1689void blk_fill_rwbs(char *rwbs, u32 rw, int bytes)
1690{
1691 int i = 0;
1692
1693 if (rw & WRITE)
1694 rwbs[i++] = 'W';
1695 else if (rw & 1 << BIO_RW_DISCARD)
1696 rwbs[i++] = 'D';
1697 else if (bytes)
1698 rwbs[i++] = 'R';
1699 else
1700 rwbs[i++] = 'N';
1701
1702 if (rw & 1 << BIO_RW_AHEAD)
1703 rwbs[i++] = 'A';
1704 if (rw & 1 << BIO_RW_BARRIER)
1705 rwbs[i++] = 'B';
1706 if (rw & 1 << BIO_RW_SYNCIO)
1707 rwbs[i++] = 'S';
1708 if (rw & 1 << BIO_RW_META)
1709 rwbs[i++] = 'M';
1710
1711 rwbs[i] = '\0';
1712}
1713
1714void blk_fill_rwbs_rq(char *rwbs, struct request *rq)
1715{
1716 int rw = rq->cmd_flags & 0x03;
1717 int bytes;
1718
1719 if (blk_discard_rq(rq))
1720 rw |= (1 << BIO_RW_DISCARD);
1721
Linus Torvaldsc9059592009-06-11 10:52:27 -07001722 bytes = blk_rq_bytes(rq);
Li Zefan55782132009-06-09 13:43:05 +08001723
1724 blk_fill_rwbs(rwbs, rw, bytes);
1725}
1726
1727#endif /* CONFIG_EVENT_TRACING */
1728