]>
Commit | Line | Data |
---|---|---|
2056a782 | 1 | /* |
0fe23479 | 2 | * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk> |
2056a782 JA |
3 | * |
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 | */ | |
2056a782 JA |
18 | #include <linux/kernel.h> |
19 | #include <linux/blkdev.h> | |
20 | #include <linux/blktrace_api.h> | |
21 | #include <linux/percpu.h> | |
22 | #include <linux/init.h> | |
23 | #include <linux/mutex.h> | |
5a0e3ad6 | 24 | #include <linux/slab.h> |
2056a782 | 25 | #include <linux/debugfs.h> |
6e5fdeed | 26 | #include <linux/export.h> |
be1c6341 | 27 | #include <linux/time.h> |
939b3669 | 28 | #include <linux/uaccess.h> |
a404d557 | 29 | #include <linux/list.h> |
55782138 | 30 | |
18fbda91 OS |
31 | #include "../../block/blk.h" |
32 | ||
55782138 LZ |
33 | #include <trace/events/block.h> |
34 | ||
2db270a8 | 35 | #include "trace_output.h" |
2056a782 | 36 | |
55782138 LZ |
37 | #ifdef CONFIG_BLK_DEV_IO_TRACE |
38 | ||
2056a782 JA |
39 | static unsigned int blktrace_seq __read_mostly = 1; |
40 | ||
c71a8961 | 41 | static struct trace_array *blk_tr; |
5006ea73 | 42 | static bool blk_tracer_enabled __read_mostly; |
c71a8961 | 43 | |
a404d557 JK |
44 | static LIST_HEAD(running_trace_list); |
45 | static __cacheline_aligned_in_smp DEFINE_SPINLOCK(running_trace_lock); | |
46 | ||
c71a8961 | 47 | /* Select an alternative, minimalistic output than the original one */ |
ef18012b | 48 | #define TRACE_BLK_OPT_CLASSIC 0x1 |
c71a8961 ACM |
49 | |
50 | static struct tracer_opt blk_tracer_opts[] = { | |
51 | /* Default disable the minimalistic output */ | |
157f9c00 | 52 | { TRACER_OPT(blk_classic, TRACE_BLK_OPT_CLASSIC) }, |
c71a8961 ACM |
53 | { } |
54 | }; | |
55 | ||
56 | static struct tracer_flags blk_tracer_flags = { | |
57 | .val = 0, | |
58 | .opts = blk_tracer_opts, | |
59 | }; | |
60 | ||
5f3ea37c | 61 | /* Global reference count of probes */ |
5f3ea37c ACM |
62 | static atomic_t blk_probes_ref = ATOMIC_INIT(0); |
63 | ||
3c289ba7 | 64 | static void blk_register_tracepoints(void); |
5f3ea37c ACM |
65 | static void blk_unregister_tracepoints(void); |
66 | ||
be1c6341 OK |
67 | /* |
68 | * Send out a notify message. | |
69 | */ | |
a863055b JA |
70 | static void trace_note(struct blk_trace *bt, pid_t pid, int action, |
71 | const void *data, size_t len) | |
be1c6341 OK |
72 | { |
73 | struct blk_io_trace *t; | |
18cea459 | 74 | struct ring_buffer_event *event = NULL; |
e77405ad | 75 | struct ring_buffer *buffer = NULL; |
18cea459 LZ |
76 | int pc = 0; |
77 | int cpu = smp_processor_id(); | |
78 | bool blk_tracer = blk_tracer_enabled; | |
79 | ||
80 | if (blk_tracer) { | |
12883efb | 81 | buffer = blk_tr->trace_buffer.buffer; |
18cea459 | 82 | pc = preempt_count(); |
e77405ad | 83 | event = trace_buffer_lock_reserve(buffer, TRACE_BLK, |
18cea459 LZ |
84 | sizeof(*t) + len, |
85 | 0, pc); | |
86 | if (!event) | |
87 | return; | |
88 | t = ring_buffer_event_data(event); | |
89 | goto record_it; | |
90 | } | |
be1c6341 | 91 | |
c71a8961 ACM |
92 | if (!bt->rchan) |
93 | return; | |
94 | ||
be1c6341 | 95 | t = relay_reserve(bt->rchan, sizeof(*t) + len); |
d3d9d2a5 | 96 | if (t) { |
d3d9d2a5 | 97 | t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION; |
2997c8c4 | 98 | t->time = ktime_to_ns(ktime_get()); |
18cea459 | 99 | record_it: |
d3d9d2a5 JA |
100 | t->device = bt->dev; |
101 | t->action = action; | |
102 | t->pid = pid; | |
103 | t->cpu = cpu; | |
104 | t->pdu_len = len; | |
105 | memcpy((void *) t + sizeof(*t), data, len); | |
18cea459 LZ |
106 | |
107 | if (blk_tracer) | |
b7f0c959 | 108 | trace_buffer_unlock_commit(blk_tr, buffer, event, 0, pc); |
d3d9d2a5 | 109 | } |
be1c6341 OK |
110 | } |
111 | ||
2056a782 JA |
112 | /* |
113 | * Send out a notify for this process, if we haven't done so since a trace | |
114 | * started | |
115 | */ | |
a404d557 | 116 | static void trace_note_tsk(struct task_struct *tsk) |
2056a782 | 117 | { |
a404d557 JK |
118 | unsigned long flags; |
119 | struct blk_trace *bt; | |
120 | ||
a863055b | 121 | tsk->btrace_seq = blktrace_seq; |
a404d557 JK |
122 | spin_lock_irqsave(&running_trace_lock, flags); |
123 | list_for_each_entry(bt, &running_trace_list, running_list) { | |
124 | trace_note(bt, tsk->pid, BLK_TN_PROCESS, tsk->comm, | |
125 | sizeof(tsk->comm)); | |
126 | } | |
127 | spin_unlock_irqrestore(&running_trace_lock, flags); | |
be1c6341 | 128 | } |
2056a782 | 129 | |
be1c6341 OK |
130 | static void trace_note_time(struct blk_trace *bt) |
131 | { | |
59a37f8b | 132 | struct timespec64 now; |
be1c6341 OK |
133 | unsigned long flags; |
134 | u32 words[2]; | |
135 | ||
59a37f8b AB |
136 | /* need to check user space to see if this breaks in y2038 or y2106 */ |
137 | ktime_get_real_ts64(&now); | |
138 | words[0] = (u32)now.tv_sec; | |
be1c6341 OK |
139 | words[1] = now.tv_nsec; |
140 | ||
141 | local_irq_save(flags); | |
142 | trace_note(bt, 0, BLK_TN_TIMESTAMP, words, sizeof(words)); | |
143 | local_irq_restore(flags); | |
2056a782 JA |
144 | } |
145 | ||
9d5f09a4 AB |
146 | void __trace_note_message(struct blk_trace *bt, const char *fmt, ...) |
147 | { | |
148 | int n; | |
149 | va_list args; | |
14a73f54 | 150 | unsigned long flags; |
64565911 | 151 | char *buf; |
9d5f09a4 | 152 | |
18cea459 LZ |
153 | if (unlikely(bt->trace_state != Blktrace_running && |
154 | !blk_tracer_enabled)) | |
c71a8961 ACM |
155 | return; |
156 | ||
490da40d TM |
157 | /* |
158 | * If the BLK_TC_NOTIFY action mask isn't set, don't send any note | |
159 | * message to the trace. | |
160 | */ | |
161 | if (!(bt->act_mask & BLK_TC_NOTIFY)) | |
162 | return; | |
163 | ||
14a73f54 | 164 | local_irq_save(flags); |
d8a0349c | 165 | buf = this_cpu_ptr(bt->msg_data); |
9d5f09a4 | 166 | va_start(args, fmt); |
64565911 | 167 | n = vscnprintf(buf, BLK_TN_MAX_MSG, fmt, args); |
9d5f09a4 AB |
168 | va_end(args); |
169 | ||
64565911 | 170 | trace_note(bt, 0, BLK_TN_MESSAGE, buf, n); |
14a73f54 | 171 | local_irq_restore(flags); |
9d5f09a4 AB |
172 | } |
173 | EXPORT_SYMBOL_GPL(__trace_note_message); | |
174 | ||
2056a782 JA |
175 | static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector, |
176 | pid_t pid) | |
177 | { | |
178 | if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0) | |
179 | return 1; | |
d0deef5b | 180 | if (sector && (sector < bt->start_lba || sector > bt->end_lba)) |
2056a782 JA |
181 | return 1; |
182 | if (bt->pid && pid != bt->pid) | |
183 | return 1; | |
184 | ||
185 | return 0; | |
186 | } | |
187 | ||
188 | /* | |
189 | * Data direction bit lookup | |
190 | */ | |
e4955c99 LZ |
191 | static const u32 ddir_act[2] = { BLK_TC_ACT(BLK_TC_READ), |
192 | BLK_TC_ACT(BLK_TC_WRITE) }; | |
2056a782 | 193 | |
7b6d91da | 194 | #define BLK_TC_RAHEAD BLK_TC_AHEAD |
28a8f0d3 | 195 | #define BLK_TC_PREFLUSH BLK_TC_FLUSH |
7b6d91da | 196 | |
35ba8f70 | 197 | /* The ilog2() calls fall out because they're constant */ |
7b6d91da CH |
198 | #define MASK_TC_BIT(rw, __name) ((rw & REQ_ ## __name) << \ |
199 | (ilog2(BLK_TC_ ## __name) + BLK_TC_SHIFT - __REQ_ ## __name)) | |
2056a782 JA |
200 | |
201 | /* | |
202 | * The worker for the various blk_add_trace*() types. Fills out a | |
203 | * blk_io_trace structure and places it in a per-cpu subbuffer. | |
204 | */ | |
5f3ea37c | 205 | static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes, |
1b9a9ab7 MC |
206 | int op, int op_flags, u32 what, int error, int pdu_len, |
207 | void *pdu_data) | |
2056a782 JA |
208 | { |
209 | struct task_struct *tsk = current; | |
c71a8961 | 210 | struct ring_buffer_event *event = NULL; |
e77405ad | 211 | struct ring_buffer *buffer = NULL; |
2056a782 | 212 | struct blk_io_trace *t; |
0a987751 | 213 | unsigned long flags = 0; |
2056a782 JA |
214 | unsigned long *sequence; |
215 | pid_t pid; | |
c71a8961 | 216 | int cpu, pc = 0; |
18cea459 | 217 | bool blk_tracer = blk_tracer_enabled; |
2056a782 | 218 | |
18cea459 | 219 | if (unlikely(bt->trace_state != Blktrace_running && !blk_tracer)) |
2056a782 JA |
220 | return; |
221 | ||
1b9a9ab7 MC |
222 | what |= ddir_act[op_is_write(op) ? WRITE : READ]; |
223 | what |= MASK_TC_BIT(op_flags, SYNC); | |
224 | what |= MASK_TC_BIT(op_flags, RAHEAD); | |
225 | what |= MASK_TC_BIT(op_flags, META); | |
28a8f0d3 | 226 | what |= MASK_TC_BIT(op_flags, PREFLUSH); |
1b9a9ab7 | 227 | what |= MASK_TC_BIT(op_flags, FUA); |
7afafc8a | 228 | if (op == REQ_OP_DISCARD || op == REQ_OP_SECURE_ERASE) |
1b9a9ab7 | 229 | what |= BLK_TC_ACT(BLK_TC_DISCARD); |
3a5e02ce MC |
230 | if (op == REQ_OP_FLUSH) |
231 | what |= BLK_TC_ACT(BLK_TC_FLUSH); | |
2056a782 JA |
232 | |
233 | pid = tsk->pid; | |
d0deef5b | 234 | if (act_log_check(bt, what, sector, pid)) |
2056a782 | 235 | return; |
c71a8961 ACM |
236 | cpu = raw_smp_processor_id(); |
237 | ||
18cea459 | 238 | if (blk_tracer) { |
c71a8961 ACM |
239 | tracing_record_cmdline(current); |
240 | ||
12883efb | 241 | buffer = blk_tr->trace_buffer.buffer; |
51a763dd | 242 | pc = preempt_count(); |
e77405ad | 243 | event = trace_buffer_lock_reserve(buffer, TRACE_BLK, |
51a763dd ACM |
244 | sizeof(*t) + pdu_len, |
245 | 0, pc); | |
c71a8961 ACM |
246 | if (!event) |
247 | return; | |
51a763dd | 248 | t = ring_buffer_event_data(event); |
c71a8961 ACM |
249 | goto record_it; |
250 | } | |
2056a782 | 251 | |
a404d557 JK |
252 | if (unlikely(tsk->btrace_seq != blktrace_seq)) |
253 | trace_note_tsk(tsk); | |
254 | ||
2056a782 JA |
255 | /* |
256 | * A word about the locking here - we disable interrupts to reserve | |
257 | * some space in the relay per-cpu buffer, to prevent an irq | |
14a73f54 | 258 | * from coming in and stepping on our toes. |
2056a782 JA |
259 | */ |
260 | local_irq_save(flags); | |
2056a782 JA |
261 | t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len); |
262 | if (t) { | |
2056a782 JA |
263 | sequence = per_cpu_ptr(bt->sequence, cpu); |
264 | ||
265 | t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION; | |
266 | t->sequence = ++(*sequence); | |
2997c8c4 | 267 | t->time = ktime_to_ns(ktime_get()); |
c71a8961 | 268 | record_it: |
08a06b83 | 269 | /* |
939b3669 ACM |
270 | * These two are not needed in ftrace as they are in the |
271 | * generic trace_entry, filled by tracing_generic_entry_update, | |
272 | * but for the trace_event->bin() synthesizer benefit we do it | |
273 | * here too. | |
274 | */ | |
275 | t->cpu = cpu; | |
276 | t->pid = pid; | |
08a06b83 | 277 | |
2056a782 JA |
278 | t->sector = sector; |
279 | t->bytes = bytes; | |
280 | t->action = what; | |
2056a782 | 281 | t->device = bt->dev; |
2056a782 JA |
282 | t->error = error; |
283 | t->pdu_len = pdu_len; | |
284 | ||
285 | if (pdu_len) | |
286 | memcpy((void *) t + sizeof(*t), pdu_data, pdu_len); | |
c71a8961 | 287 | |
18cea459 | 288 | if (blk_tracer) { |
b7f0c959 | 289 | trace_buffer_unlock_commit(blk_tr, buffer, event, 0, pc); |
c71a8961 ACM |
290 | return; |
291 | } | |
2056a782 JA |
292 | } |
293 | ||
294 | local_irq_restore(flags); | |
295 | } | |
296 | ||
ad5dd549 | 297 | static void blk_trace_free(struct blk_trace *bt) |
2056a782 | 298 | { |
02c62304 | 299 | debugfs_remove(bt->msg_file); |
2056a782 | 300 | debugfs_remove(bt->dropped_file); |
f48fc4d3 | 301 | relay_close(bt->rchan); |
39cbb602 | 302 | debugfs_remove(bt->dir); |
2056a782 | 303 | free_percpu(bt->sequence); |
64565911 | 304 | free_percpu(bt->msg_data); |
2056a782 | 305 | kfree(bt); |
ad5dd549 LZ |
306 | } |
307 | ||
308 | static void blk_trace_cleanup(struct blk_trace *bt) | |
309 | { | |
310 | blk_trace_free(bt); | |
5f3ea37c ACM |
311 | if (atomic_dec_and_test(&blk_probes_ref)) |
312 | blk_unregister_tracepoints(); | |
2056a782 JA |
313 | } |
314 | ||
6da127ad | 315 | int blk_trace_remove(struct request_queue *q) |
2056a782 JA |
316 | { |
317 | struct blk_trace *bt; | |
318 | ||
319 | bt = xchg(&q->blk_trace, NULL); | |
320 | if (!bt) | |
321 | return -EINVAL; | |
322 | ||
55547204 | 323 | if (bt->trace_state != Blktrace_running) |
2056a782 JA |
324 | blk_trace_cleanup(bt); |
325 | ||
326 | return 0; | |
327 | } | |
6da127ad | 328 | EXPORT_SYMBOL_GPL(blk_trace_remove); |
2056a782 | 329 | |
2056a782 JA |
330 | static ssize_t blk_dropped_read(struct file *filp, char __user *buffer, |
331 | size_t count, loff_t *ppos) | |
332 | { | |
333 | struct blk_trace *bt = filp->private_data; | |
334 | char buf[16]; | |
335 | ||
336 | snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped)); | |
337 | ||
338 | return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf)); | |
339 | } | |
340 | ||
2b8693c0 | 341 | static const struct file_operations blk_dropped_fops = { |
2056a782 | 342 | .owner = THIS_MODULE, |
234e3405 | 343 | .open = simple_open, |
2056a782 | 344 | .read = blk_dropped_read, |
6038f373 | 345 | .llseek = default_llseek, |
2056a782 JA |
346 | }; |
347 | ||
02c62304 AB |
348 | static ssize_t blk_msg_write(struct file *filp, const char __user *buffer, |
349 | size_t count, loff_t *ppos) | |
350 | { | |
351 | char *msg; | |
352 | struct blk_trace *bt; | |
353 | ||
7635b03a | 354 | if (count >= BLK_TN_MAX_MSG) |
02c62304 AB |
355 | return -EINVAL; |
356 | ||
16e5c1fc AV |
357 | msg = memdup_user_nul(buffer, count); |
358 | if (IS_ERR(msg)) | |
359 | return PTR_ERR(msg); | |
02c62304 | 360 | |
02c62304 AB |
361 | bt = filp->private_data; |
362 | __trace_note_message(bt, "%s", msg); | |
363 | kfree(msg); | |
364 | ||
365 | return count; | |
366 | } | |
367 | ||
368 | static const struct file_operations blk_msg_fops = { | |
369 | .owner = THIS_MODULE, | |
234e3405 | 370 | .open = simple_open, |
02c62304 | 371 | .write = blk_msg_write, |
6038f373 | 372 | .llseek = noop_llseek, |
02c62304 AB |
373 | }; |
374 | ||
2056a782 JA |
375 | /* |
376 | * Keep track of how many times we encountered a full subbuffer, to aid | |
377 | * the user space app in telling how many lost events there were. | |
378 | */ | |
379 | static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf, | |
380 | void *prev_subbuf, size_t prev_padding) | |
381 | { | |
382 | struct blk_trace *bt; | |
383 | ||
384 | if (!relay_buf_full(buf)) | |
385 | return 1; | |
386 | ||
387 | bt = buf->chan->private_data; | |
388 | atomic_inc(&bt->dropped); | |
389 | return 0; | |
390 | } | |
391 | ||
392 | static int blk_remove_buf_file_callback(struct dentry *dentry) | |
393 | { | |
394 | debugfs_remove(dentry); | |
f48fc4d3 | 395 | |
2056a782 JA |
396 | return 0; |
397 | } | |
398 | ||
399 | static struct dentry *blk_create_buf_file_callback(const char *filename, | |
400 | struct dentry *parent, | |
f4ae40a6 | 401 | umode_t mode, |
2056a782 JA |
402 | struct rchan_buf *buf, |
403 | int *is_global) | |
404 | { | |
405 | return debugfs_create_file(filename, mode, parent, buf, | |
406 | &relay_file_operations); | |
407 | } | |
408 | ||
409 | static struct rchan_callbacks blk_relay_callbacks = { | |
410 | .subbuf_start = blk_subbuf_start_callback, | |
411 | .create_buf_file = blk_create_buf_file_callback, | |
412 | .remove_buf_file = blk_remove_buf_file_callback, | |
413 | }; | |
414 | ||
9908c309 LZ |
415 | static void blk_trace_setup_lba(struct blk_trace *bt, |
416 | struct block_device *bdev) | |
417 | { | |
418 | struct hd_struct *part = NULL; | |
419 | ||
420 | if (bdev) | |
421 | part = bdev->bd_part; | |
422 | ||
423 | if (part) { | |
424 | bt->start_lba = part->start_sect; | |
425 | bt->end_lba = part->start_sect + part->nr_sects; | |
426 | } else { | |
427 | bt->start_lba = 0; | |
428 | bt->end_lba = -1ULL; | |
429 | } | |
430 | } | |
431 | ||
2056a782 JA |
432 | /* |
433 | * Setup everything required to start tracing | |
434 | */ | |
a428d314 OS |
435 | static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev, |
436 | struct block_device *bdev, | |
437 | struct blk_user_trace_setup *buts) | |
2056a782 | 438 | { |
cdea01b2 | 439 | struct blk_trace *bt = NULL; |
2056a782 | 440 | struct dentry *dir = NULL; |
ff14417c | 441 | int ret; |
2056a782 | 442 | |
171044d4 | 443 | if (!buts->buf_size || !buts->buf_nr) |
2056a782 JA |
444 | return -EINVAL; |
445 | ||
0497b345 JA |
446 | strncpy(buts->name, name, BLKTRACE_BDEV_SIZE); |
447 | buts->name[BLKTRACE_BDEV_SIZE - 1] = '\0'; | |
2056a782 JA |
448 | |
449 | /* | |
450 | * some device names have larger paths - convert the slashes | |
451 | * to underscores for this to work as expected | |
452 | */ | |
ff14417c | 453 | strreplace(buts->name, '/', '_'); |
2056a782 | 454 | |
2056a782 JA |
455 | bt = kzalloc(sizeof(*bt), GFP_KERNEL); |
456 | if (!bt) | |
ad5dd549 | 457 | return -ENOMEM; |
2056a782 | 458 | |
ad5dd549 | 459 | ret = -ENOMEM; |
2056a782 JA |
460 | bt->sequence = alloc_percpu(unsigned long); |
461 | if (!bt->sequence) | |
462 | goto err; | |
463 | ||
313e458f | 464 | bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG, __alignof__(char)); |
64565911 JA |
465 | if (!bt->msg_data) |
466 | goto err; | |
467 | ||
2056a782 | 468 | ret = -ENOENT; |
f48fc4d3 | 469 | |
18fbda91 OS |
470 | if (!blk_debugfs_root) |
471 | goto err; | |
f48fc4d3 | 472 | |
6ac93117 OS |
473 | dir = debugfs_lookup(buts->name, blk_debugfs_root); |
474 | if (!dir) | |
475 | bt->dir = dir = debugfs_create_dir(buts->name, blk_debugfs_root); | |
2056a782 JA |
476 | if (!dir) |
477 | goto err; | |
478 | ||
6da127ad | 479 | bt->dev = dev; |
2056a782 | 480 | atomic_set(&bt->dropped, 0); |
a404d557 | 481 | INIT_LIST_HEAD(&bt->running_list); |
2056a782 JA |
482 | |
483 | ret = -EIO; | |
939b3669 ACM |
484 | bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, |
485 | &blk_dropped_fops); | |
2056a782 JA |
486 | if (!bt->dropped_file) |
487 | goto err; | |
488 | ||
02c62304 AB |
489 | bt->msg_file = debugfs_create_file("msg", 0222, dir, bt, &blk_msg_fops); |
490 | if (!bt->msg_file) | |
491 | goto err; | |
492 | ||
171044d4 AB |
493 | bt->rchan = relay_open("trace", dir, buts->buf_size, |
494 | buts->buf_nr, &blk_relay_callbacks, bt); | |
2056a782 JA |
495 | if (!bt->rchan) |
496 | goto err; | |
2056a782 | 497 | |
171044d4 | 498 | bt->act_mask = buts->act_mask; |
2056a782 JA |
499 | if (!bt->act_mask) |
500 | bt->act_mask = (u16) -1; | |
501 | ||
9908c309 | 502 | blk_trace_setup_lba(bt, bdev); |
2056a782 | 503 | |
d0deef5b SD |
504 | /* overwrite with user settings */ |
505 | if (buts->start_lba) | |
506 | bt->start_lba = buts->start_lba; | |
507 | if (buts->end_lba) | |
508 | bt->end_lba = buts->end_lba; | |
509 | ||
171044d4 | 510 | bt->pid = buts->pid; |
2056a782 JA |
511 | bt->trace_state = Blktrace_setup; |
512 | ||
513 | ret = -EBUSY; | |
cdea01b2 | 514 | if (cmpxchg(&q->blk_trace, NULL, bt)) |
2056a782 | 515 | goto err; |
2056a782 | 516 | |
17ba97e3 | 517 | if (atomic_inc_return(&blk_probes_ref) == 1) |
cbe28296 LZ |
518 | blk_register_tracepoints(); |
519 | ||
6ac93117 | 520 | ret = 0; |
2056a782 | 521 | err: |
6ac93117 OS |
522 | if (dir && !bt->dir) |
523 | dput(dir); | |
524 | if (ret) | |
525 | blk_trace_free(bt); | |
2056a782 JA |
526 | return ret; |
527 | } | |
171044d4 | 528 | |
6da127ad | 529 | int blk_trace_setup(struct request_queue *q, char *name, dev_t dev, |
d0deef5b | 530 | struct block_device *bdev, |
6da127ad | 531 | char __user *arg) |
171044d4 AB |
532 | { |
533 | struct blk_user_trace_setup buts; | |
534 | int ret; | |
535 | ||
536 | ret = copy_from_user(&buts, arg, sizeof(buts)); | |
537 | if (ret) | |
538 | return -EFAULT; | |
539 | ||
d0deef5b | 540 | ret = do_blk_trace_setup(q, name, dev, bdev, &buts); |
171044d4 AB |
541 | if (ret) |
542 | return ret; | |
543 | ||
9a8c28c8 DM |
544 | if (copy_to_user(arg, &buts, sizeof(buts))) { |
545 | blk_trace_remove(q); | |
171044d4 | 546 | return -EFAULT; |
9a8c28c8 | 547 | } |
171044d4 AB |
548 | return 0; |
549 | } | |
6da127ad | 550 | EXPORT_SYMBOL_GPL(blk_trace_setup); |
2056a782 | 551 | |
62c2a7d9 AB |
552 | #if defined(CONFIG_COMPAT) && defined(CONFIG_X86_64) |
553 | static int compat_blk_trace_setup(struct request_queue *q, char *name, | |
554 | dev_t dev, struct block_device *bdev, | |
555 | char __user *arg) | |
556 | { | |
557 | struct blk_user_trace_setup buts; | |
558 | struct compat_blk_user_trace_setup cbuts; | |
559 | int ret; | |
560 | ||
561 | if (copy_from_user(&cbuts, arg, sizeof(cbuts))) | |
562 | return -EFAULT; | |
563 | ||
564 | buts = (struct blk_user_trace_setup) { | |
565 | .act_mask = cbuts.act_mask, | |
566 | .buf_size = cbuts.buf_size, | |
567 | .buf_nr = cbuts.buf_nr, | |
568 | .start_lba = cbuts.start_lba, | |
569 | .end_lba = cbuts.end_lba, | |
570 | .pid = cbuts.pid, | |
571 | }; | |
62c2a7d9 AB |
572 | |
573 | ret = do_blk_trace_setup(q, name, dev, bdev, &buts); | |
574 | if (ret) | |
575 | return ret; | |
576 | ||
f8c5e944 | 577 | if (copy_to_user(arg, &buts.name, ARRAY_SIZE(buts.name))) { |
62c2a7d9 AB |
578 | blk_trace_remove(q); |
579 | return -EFAULT; | |
580 | } | |
581 | ||
582 | return 0; | |
583 | } | |
584 | #endif | |
585 | ||
6da127ad | 586 | int blk_trace_startstop(struct request_queue *q, int start) |
2056a782 | 587 | { |
2056a782 | 588 | int ret; |
939b3669 | 589 | struct blk_trace *bt = q->blk_trace; |
2056a782 | 590 | |
939b3669 | 591 | if (bt == NULL) |
2056a782 JA |
592 | return -EINVAL; |
593 | ||
594 | /* | |
595 | * For starting a trace, we can transition from a setup or stopped | |
596 | * trace. For stopping a trace, the state must be running | |
597 | */ | |
598 | ret = -EINVAL; | |
599 | if (start) { | |
600 | if (bt->trace_state == Blktrace_setup || | |
601 | bt->trace_state == Blktrace_stopped) { | |
602 | blktrace_seq++; | |
603 | smp_mb(); | |
604 | bt->trace_state = Blktrace_running; | |
a404d557 JK |
605 | spin_lock_irq(&running_trace_lock); |
606 | list_add(&bt->running_list, &running_trace_list); | |
607 | spin_unlock_irq(&running_trace_lock); | |
be1c6341 OK |
608 | |
609 | trace_note_time(bt); | |
2056a782 JA |
610 | ret = 0; |
611 | } | |
612 | } else { | |
613 | if (bt->trace_state == Blktrace_running) { | |
614 | bt->trace_state = Blktrace_stopped; | |
a404d557 JK |
615 | spin_lock_irq(&running_trace_lock); |
616 | list_del_init(&bt->running_list); | |
617 | spin_unlock_irq(&running_trace_lock); | |
2056a782 JA |
618 | relay_flush(bt->rchan); |
619 | ret = 0; | |
620 | } | |
621 | } | |
622 | ||
623 | return ret; | |
624 | } | |
6da127ad | 625 | EXPORT_SYMBOL_GPL(blk_trace_startstop); |
2056a782 JA |
626 | |
627 | /** | |
628 | * blk_trace_ioctl: - handle the ioctls associated with tracing | |
629 | * @bdev: the block device | |
ef18012b | 630 | * @cmd: the ioctl cmd |
2056a782 JA |
631 | * @arg: the argument data, if any |
632 | * | |
633 | **/ | |
634 | int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg) | |
635 | { | |
165125e1 | 636 | struct request_queue *q; |
2056a782 | 637 | int ret, start = 0; |
6da127ad | 638 | char b[BDEVNAME_SIZE]; |
2056a782 JA |
639 | |
640 | q = bdev_get_queue(bdev); | |
641 | if (!q) | |
642 | return -ENXIO; | |
643 | ||
644 | mutex_lock(&bdev->bd_mutex); | |
645 | ||
646 | switch (cmd) { | |
647 | case BLKTRACESETUP: | |
f36f21ec | 648 | bdevname(bdev, b); |
d0deef5b | 649 | ret = blk_trace_setup(q, b, bdev->bd_dev, bdev, arg); |
2056a782 | 650 | break; |
62c2a7d9 AB |
651 | #if defined(CONFIG_COMPAT) && defined(CONFIG_X86_64) |
652 | case BLKTRACESETUP32: | |
653 | bdevname(bdev, b); | |
654 | ret = compat_blk_trace_setup(q, b, bdev->bd_dev, bdev, arg); | |
655 | break; | |
656 | #endif | |
2056a782 JA |
657 | case BLKTRACESTART: |
658 | start = 1; | |
659 | case BLKTRACESTOP: | |
660 | ret = blk_trace_startstop(q, start); | |
661 | break; | |
662 | case BLKTRACETEARDOWN: | |
663 | ret = blk_trace_remove(q); | |
664 | break; | |
665 | default: | |
666 | ret = -ENOTTY; | |
667 | break; | |
668 | } | |
669 | ||
670 | mutex_unlock(&bdev->bd_mutex); | |
671 | return ret; | |
672 | } | |
673 | ||
674 | /** | |
675 | * blk_trace_shutdown: - stop and cleanup trace structures | |
676 | * @q: the request queue associated with the device | |
677 | * | |
678 | **/ | |
165125e1 | 679 | void blk_trace_shutdown(struct request_queue *q) |
2056a782 | 680 | { |
6c5c9341 AD |
681 | if (q->blk_trace) { |
682 | blk_trace_startstop(q, 0); | |
683 | blk_trace_remove(q); | |
684 | } | |
2056a782 | 685 | } |
5f3ea37c ACM |
686 | |
687 | /* | |
688 | * blktrace probes | |
689 | */ | |
690 | ||
691 | /** | |
692 | * blk_add_trace_rq - Add a trace for a request oriented action | |
693 | * @q: queue the io is for | |
694 | * @rq: the source request | |
af5040da | 695 | * @nr_bytes: number of completed bytes |
5f3ea37c ACM |
696 | * @what: the action |
697 | * | |
698 | * Description: | |
699 | * Records an action against a request. Will log the bio offset + size. | |
700 | * | |
701 | **/ | |
702 | static void blk_add_trace_rq(struct request_queue *q, struct request *rq, | |
af5040da | 703 | unsigned int nr_bytes, u32 what) |
5f3ea37c ACM |
704 | { |
705 | struct blk_trace *bt = q->blk_trace; | |
5f3ea37c ACM |
706 | |
707 | if (likely(!bt)) | |
708 | return; | |
709 | ||
57292b58 | 710 | if (blk_rq_is_passthrough(rq)) |
5f3ea37c | 711 | what |= BLK_TC_ACT(BLK_TC_PC); |
48b77ad6 | 712 | else |
5f3ea37c | 713 | what |= BLK_TC_ACT(BLK_TC_FS); |
48b77ad6 CH |
714 | |
715 | __blk_add_trace(bt, blk_rq_trace_sector(rq), nr_bytes, req_op(rq), | |
716 | rq->cmd_flags, what, rq->errors, 0, NULL); | |
5f3ea37c ACM |
717 | } |
718 | ||
38516ab5 SR |
719 | static void blk_add_trace_rq_abort(void *ignore, |
720 | struct request_queue *q, struct request *rq) | |
5f3ea37c | 721 | { |
af5040da | 722 | blk_add_trace_rq(q, rq, blk_rq_bytes(rq), BLK_TA_ABORT); |
5f3ea37c ACM |
723 | } |
724 | ||
38516ab5 SR |
725 | static void blk_add_trace_rq_insert(void *ignore, |
726 | struct request_queue *q, struct request *rq) | |
5f3ea37c | 727 | { |
af5040da | 728 | blk_add_trace_rq(q, rq, blk_rq_bytes(rq), BLK_TA_INSERT); |
5f3ea37c ACM |
729 | } |
730 | ||
38516ab5 SR |
731 | static void blk_add_trace_rq_issue(void *ignore, |
732 | struct request_queue *q, struct request *rq) | |
5f3ea37c | 733 | { |
af5040da | 734 | blk_add_trace_rq(q, rq, blk_rq_bytes(rq), BLK_TA_ISSUE); |
5f3ea37c ACM |
735 | } |
736 | ||
38516ab5 SR |
737 | static void blk_add_trace_rq_requeue(void *ignore, |
738 | struct request_queue *q, | |
939b3669 | 739 | struct request *rq) |
5f3ea37c | 740 | { |
af5040da | 741 | blk_add_trace_rq(q, rq, blk_rq_bytes(rq), BLK_TA_REQUEUE); |
5f3ea37c ACM |
742 | } |
743 | ||
38516ab5 SR |
744 | static void blk_add_trace_rq_complete(void *ignore, |
745 | struct request_queue *q, | |
af5040da RP |
746 | struct request *rq, |
747 | unsigned int nr_bytes) | |
5f3ea37c | 748 | { |
af5040da | 749 | blk_add_trace_rq(q, rq, nr_bytes, BLK_TA_COMPLETE); |
5f3ea37c ACM |
750 | } |
751 | ||
752 | /** | |
753 | * blk_add_trace_bio - Add a trace for a bio oriented action | |
754 | * @q: queue the io is for | |
755 | * @bio: the source bio | |
756 | * @what: the action | |
797a455d | 757 | * @error: error, if any |
5f3ea37c ACM |
758 | * |
759 | * Description: | |
760 | * Records an action against a bio. Will log the bio offset + size. | |
761 | * | |
762 | **/ | |
763 | static void blk_add_trace_bio(struct request_queue *q, struct bio *bio, | |
797a455d | 764 | u32 what, int error) |
5f3ea37c ACM |
765 | { |
766 | struct blk_trace *bt = q->blk_trace; | |
767 | ||
768 | if (likely(!bt)) | |
769 | return; | |
770 | ||
4f024f37 | 771 | __blk_add_trace(bt, bio->bi_iter.bi_sector, bio->bi_iter.bi_size, |
1eff9d32 | 772 | bio_op(bio), bio->bi_opf, what, error, 0, NULL); |
5f3ea37c ACM |
773 | } |
774 | ||
38516ab5 SR |
775 | static void blk_add_trace_bio_bounce(void *ignore, |
776 | struct request_queue *q, struct bio *bio) | |
5f3ea37c | 777 | { |
797a455d | 778 | blk_add_trace_bio(q, bio, BLK_TA_BOUNCE, 0); |
5f3ea37c ACM |
779 | } |
780 | ||
0a82a8d1 LT |
781 | static void blk_add_trace_bio_complete(void *ignore, |
782 | struct request_queue *q, struct bio *bio, | |
783 | int error) | |
5f3ea37c | 784 | { |
797a455d | 785 | blk_add_trace_bio(q, bio, BLK_TA_COMPLETE, error); |
5f3ea37c ACM |
786 | } |
787 | ||
38516ab5 SR |
788 | static void blk_add_trace_bio_backmerge(void *ignore, |
789 | struct request_queue *q, | |
8c1cf6bb | 790 | struct request *rq, |
939b3669 | 791 | struct bio *bio) |
5f3ea37c | 792 | { |
797a455d | 793 | blk_add_trace_bio(q, bio, BLK_TA_BACKMERGE, 0); |
5f3ea37c ACM |
794 | } |
795 | ||
38516ab5 SR |
796 | static void blk_add_trace_bio_frontmerge(void *ignore, |
797 | struct request_queue *q, | |
8c1cf6bb | 798 | struct request *rq, |
939b3669 | 799 | struct bio *bio) |
5f3ea37c | 800 | { |
797a455d | 801 | blk_add_trace_bio(q, bio, BLK_TA_FRONTMERGE, 0); |
5f3ea37c ACM |
802 | } |
803 | ||
38516ab5 SR |
804 | static void blk_add_trace_bio_queue(void *ignore, |
805 | struct request_queue *q, struct bio *bio) | |
5f3ea37c | 806 | { |
797a455d | 807 | blk_add_trace_bio(q, bio, BLK_TA_QUEUE, 0); |
5f3ea37c ACM |
808 | } |
809 | ||
38516ab5 SR |
810 | static void blk_add_trace_getrq(void *ignore, |
811 | struct request_queue *q, | |
939b3669 | 812 | struct bio *bio, int rw) |
5f3ea37c ACM |
813 | { |
814 | if (bio) | |
797a455d | 815 | blk_add_trace_bio(q, bio, BLK_TA_GETRQ, 0); |
5f3ea37c ACM |
816 | else { |
817 | struct blk_trace *bt = q->blk_trace; | |
818 | ||
819 | if (bt) | |
1b9a9ab7 MC |
820 | __blk_add_trace(bt, 0, 0, rw, 0, BLK_TA_GETRQ, 0, 0, |
821 | NULL); | |
5f3ea37c ACM |
822 | } |
823 | } | |
824 | ||
825 | ||
38516ab5 SR |
826 | static void blk_add_trace_sleeprq(void *ignore, |
827 | struct request_queue *q, | |
939b3669 | 828 | struct bio *bio, int rw) |
5f3ea37c ACM |
829 | { |
830 | if (bio) | |
797a455d | 831 | blk_add_trace_bio(q, bio, BLK_TA_SLEEPRQ, 0); |
5f3ea37c ACM |
832 | else { |
833 | struct blk_trace *bt = q->blk_trace; | |
834 | ||
835 | if (bt) | |
1b9a9ab7 | 836 | __blk_add_trace(bt, 0, 0, rw, 0, BLK_TA_SLEEPRQ, |
939b3669 | 837 | 0, 0, NULL); |
5f3ea37c ACM |
838 | } |
839 | } | |
840 | ||
38516ab5 | 841 | static void blk_add_trace_plug(void *ignore, struct request_queue *q) |
5f3ea37c ACM |
842 | { |
843 | struct blk_trace *bt = q->blk_trace; | |
844 | ||
845 | if (bt) | |
1b9a9ab7 | 846 | __blk_add_trace(bt, 0, 0, 0, 0, BLK_TA_PLUG, 0, 0, NULL); |
5f3ea37c ACM |
847 | } |
848 | ||
49cac01e JA |
849 | static void blk_add_trace_unplug(void *ignore, struct request_queue *q, |
850 | unsigned int depth, bool explicit) | |
5f3ea37c ACM |
851 | { |
852 | struct blk_trace *bt = q->blk_trace; | |
853 | ||
854 | if (bt) { | |
94b5eb28 | 855 | __be64 rpdu = cpu_to_be64(depth); |
49cac01e | 856 | u32 what; |
5f3ea37c | 857 | |
49cac01e JA |
858 | if (explicit) |
859 | what = BLK_TA_UNPLUG_IO; | |
860 | else | |
861 | what = BLK_TA_UNPLUG_TIMER; | |
862 | ||
1b9a9ab7 | 863 | __blk_add_trace(bt, 0, 0, 0, 0, what, 0, sizeof(rpdu), &rpdu); |
5f3ea37c ACM |
864 | } |
865 | } | |
866 | ||
38516ab5 SR |
867 | static void blk_add_trace_split(void *ignore, |
868 | struct request_queue *q, struct bio *bio, | |
5f3ea37c ACM |
869 | unsigned int pdu) |
870 | { | |
871 | struct blk_trace *bt = q->blk_trace; | |
872 | ||
873 | if (bt) { | |
874 | __be64 rpdu = cpu_to_be64(pdu); | |
875 | ||
4f024f37 | 876 | __blk_add_trace(bt, bio->bi_iter.bi_sector, |
1eff9d32 | 877 | bio->bi_iter.bi_size, bio_op(bio), bio->bi_opf, |
1b9a9ab7 MC |
878 | BLK_TA_SPLIT, bio->bi_error, sizeof(rpdu), |
879 | &rpdu); | |
5f3ea37c ACM |
880 | } |
881 | } | |
882 | ||
883 | /** | |
d07335e5 | 884 | * blk_add_trace_bio_remap - Add a trace for a bio-remap operation |
546cf44a | 885 | * @ignore: trace callback data parameter (not used) |
5f3ea37c ACM |
886 | * @q: queue the io is for |
887 | * @bio: the source bio | |
888 | * @dev: target device | |
a42aaa3b | 889 | * @from: source sector |
5f3ea37c ACM |
890 | * |
891 | * Description: | |
892 | * Device mapper or raid target sometimes need to split a bio because | |
893 | * it spans a stripe (or similar). Add a trace for that action. | |
894 | * | |
895 | **/ | |
d07335e5 MS |
896 | static void blk_add_trace_bio_remap(void *ignore, |
897 | struct request_queue *q, struct bio *bio, | |
898 | dev_t dev, sector_t from) | |
5f3ea37c ACM |
899 | { |
900 | struct blk_trace *bt = q->blk_trace; | |
901 | struct blk_io_trace_remap r; | |
902 | ||
903 | if (likely(!bt)) | |
904 | return; | |
905 | ||
a42aaa3b AB |
906 | r.device_from = cpu_to_be32(dev); |
907 | r.device_to = cpu_to_be32(bio->bi_bdev->bd_dev); | |
908 | r.sector_from = cpu_to_be64(from); | |
5f3ea37c | 909 | |
4f024f37 | 910 | __blk_add_trace(bt, bio->bi_iter.bi_sector, bio->bi_iter.bi_size, |
1eff9d32 | 911 | bio_op(bio), bio->bi_opf, BLK_TA_REMAP, bio->bi_error, |
4246a0b6 | 912 | sizeof(r), &r); |
5f3ea37c ACM |
913 | } |
914 | ||
b0da3f0d JN |
915 | /** |
916 | * blk_add_trace_rq_remap - Add a trace for a request-remap operation | |
546cf44a | 917 | * @ignore: trace callback data parameter (not used) |
b0da3f0d JN |
918 | * @q: queue the io is for |
919 | * @rq: the source request | |
920 | * @dev: target device | |
921 | * @from: source sector | |
922 | * | |
923 | * Description: | |
924 | * Device mapper remaps request to other devices. | |
925 | * Add a trace for that action. | |
926 | * | |
927 | **/ | |
38516ab5 SR |
928 | static void blk_add_trace_rq_remap(void *ignore, |
929 | struct request_queue *q, | |
b0da3f0d JN |
930 | struct request *rq, dev_t dev, |
931 | sector_t from) | |
932 | { | |
933 | struct blk_trace *bt = q->blk_trace; | |
934 | struct blk_io_trace_remap r; | |
935 | ||
936 | if (likely(!bt)) | |
937 | return; | |
938 | ||
939 | r.device_from = cpu_to_be32(dev); | |
940 | r.device_to = cpu_to_be32(disk_devt(rq->rq_disk)); | |
941 | r.sector_from = cpu_to_be64(from); | |
942 | ||
943 | __blk_add_trace(bt, blk_rq_pos(rq), blk_rq_bytes(rq), | |
1b9a9ab7 | 944 | rq_data_dir(rq), 0, BLK_TA_REMAP, !!rq->errors, |
b0da3f0d JN |
945 | sizeof(r), &r); |
946 | } | |
947 | ||
5f3ea37c ACM |
948 | /** |
949 | * blk_add_driver_data - Add binary message with driver-specific data | |
950 | * @q: queue the io is for | |
951 | * @rq: io request | |
952 | * @data: driver-specific data | |
953 | * @len: length of driver-specific data | |
954 | * | |
955 | * Description: | |
956 | * Some drivers might want to write driver-specific data per request. | |
957 | * | |
958 | **/ | |
959 | void blk_add_driver_data(struct request_queue *q, | |
960 | struct request *rq, | |
961 | void *data, size_t len) | |
962 | { | |
963 | struct blk_trace *bt = q->blk_trace; | |
964 | ||
965 | if (likely(!bt)) | |
966 | return; | |
967 | ||
48b77ad6 | 968 | __blk_add_trace(bt, blk_rq_trace_sector(rq), blk_rq_bytes(rq), 0, 0, |
2e46e8b2 | 969 | BLK_TA_DRV_DATA, rq->errors, len, data); |
5f3ea37c ACM |
970 | } |
971 | EXPORT_SYMBOL_GPL(blk_add_driver_data); | |
972 | ||
3c289ba7 | 973 | static void blk_register_tracepoints(void) |
5f3ea37c ACM |
974 | { |
975 | int ret; | |
976 | ||
38516ab5 | 977 | ret = register_trace_block_rq_abort(blk_add_trace_rq_abort, NULL); |
5f3ea37c | 978 | WARN_ON(ret); |
38516ab5 | 979 | ret = register_trace_block_rq_insert(blk_add_trace_rq_insert, NULL); |
5f3ea37c | 980 | WARN_ON(ret); |
38516ab5 | 981 | ret = register_trace_block_rq_issue(blk_add_trace_rq_issue, NULL); |
5f3ea37c | 982 | WARN_ON(ret); |
38516ab5 | 983 | ret = register_trace_block_rq_requeue(blk_add_trace_rq_requeue, NULL); |
5f3ea37c | 984 | WARN_ON(ret); |
38516ab5 | 985 | ret = register_trace_block_rq_complete(blk_add_trace_rq_complete, NULL); |
5f3ea37c | 986 | WARN_ON(ret); |
38516ab5 | 987 | ret = register_trace_block_bio_bounce(blk_add_trace_bio_bounce, NULL); |
5f3ea37c | 988 | WARN_ON(ret); |
38516ab5 | 989 | ret = register_trace_block_bio_complete(blk_add_trace_bio_complete, NULL); |
5f3ea37c | 990 | WARN_ON(ret); |
38516ab5 | 991 | ret = register_trace_block_bio_backmerge(blk_add_trace_bio_backmerge, NULL); |
5f3ea37c | 992 | WARN_ON(ret); |
38516ab5 | 993 | ret = register_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge, NULL); |
5f3ea37c | 994 | WARN_ON(ret); |
38516ab5 | 995 | ret = register_trace_block_bio_queue(blk_add_trace_bio_queue, NULL); |
5f3ea37c | 996 | WARN_ON(ret); |
38516ab5 | 997 | ret = register_trace_block_getrq(blk_add_trace_getrq, NULL); |
5f3ea37c | 998 | WARN_ON(ret); |
38516ab5 | 999 | ret = register_trace_block_sleeprq(blk_add_trace_sleeprq, NULL); |
5f3ea37c | 1000 | WARN_ON(ret); |
38516ab5 | 1001 | ret = register_trace_block_plug(blk_add_trace_plug, NULL); |
5f3ea37c | 1002 | WARN_ON(ret); |
49cac01e | 1003 | ret = register_trace_block_unplug(blk_add_trace_unplug, NULL); |
5f3ea37c | 1004 | WARN_ON(ret); |
38516ab5 | 1005 | ret = register_trace_block_split(blk_add_trace_split, NULL); |
5f3ea37c | 1006 | WARN_ON(ret); |
d07335e5 | 1007 | ret = register_trace_block_bio_remap(blk_add_trace_bio_remap, NULL); |
5f3ea37c | 1008 | WARN_ON(ret); |
38516ab5 | 1009 | ret = register_trace_block_rq_remap(blk_add_trace_rq_remap, NULL); |
b0da3f0d | 1010 | WARN_ON(ret); |
5f3ea37c ACM |
1011 | } |
1012 | ||
1013 | static void blk_unregister_tracepoints(void) | |
1014 | { | |
38516ab5 | 1015 | unregister_trace_block_rq_remap(blk_add_trace_rq_remap, NULL); |
d07335e5 | 1016 | unregister_trace_block_bio_remap(blk_add_trace_bio_remap, NULL); |
38516ab5 | 1017 | unregister_trace_block_split(blk_add_trace_split, NULL); |
49cac01e | 1018 | unregister_trace_block_unplug(blk_add_trace_unplug, NULL); |
38516ab5 SR |
1019 | unregister_trace_block_plug(blk_add_trace_plug, NULL); |
1020 | unregister_trace_block_sleeprq(blk_add_trace_sleeprq, NULL); | |
1021 | unregister_trace_block_getrq(blk_add_trace_getrq, NULL); | |
1022 | unregister_trace_block_bio_queue(blk_add_trace_bio_queue, NULL); | |
1023 | unregister_trace_block_bio_frontmerge(blk_add_trace_bio_frontmerge, NULL); | |
1024 | unregister_trace_block_bio_backmerge(blk_add_trace_bio_backmerge, NULL); | |
1025 | unregister_trace_block_bio_complete(blk_add_trace_bio_complete, NULL); | |
1026 | unregister_trace_block_bio_bounce(blk_add_trace_bio_bounce, NULL); | |
1027 | unregister_trace_block_rq_complete(blk_add_trace_rq_complete, NULL); | |
1028 | unregister_trace_block_rq_requeue(blk_add_trace_rq_requeue, NULL); | |
1029 | unregister_trace_block_rq_issue(blk_add_trace_rq_issue, NULL); | |
1030 | unregister_trace_block_rq_insert(blk_add_trace_rq_insert, NULL); | |
1031 | unregister_trace_block_rq_abort(blk_add_trace_rq_abort, NULL); | |
5f3ea37c ACM |
1032 | |
1033 | tracepoint_synchronize_unregister(); | |
1034 | } | |
c71a8961 ACM |
1035 | |
1036 | /* | |
1037 | * struct blk_io_tracer formatting routines | |
1038 | */ | |
1039 | ||
1040 | static void fill_rwbs(char *rwbs, const struct blk_io_trace *t) | |
1041 | { | |
157f9c00 | 1042 | int i = 0; |
65796348 | 1043 | int tc = t->action >> BLK_TC_SHIFT; |
157f9c00 | 1044 | |
18cea459 LZ |
1045 | if (t->action == BLK_TN_MESSAGE) { |
1046 | rwbs[i++] = 'N'; | |
1047 | goto out; | |
1048 | } | |
1049 | ||
c09c47ca NK |
1050 | if (tc & BLK_TC_FLUSH) |
1051 | rwbs[i++] = 'F'; | |
1052 | ||
65796348 | 1053 | if (tc & BLK_TC_DISCARD) |
157f9c00 | 1054 | rwbs[i++] = 'D'; |
65796348 | 1055 | else if (tc & BLK_TC_WRITE) |
157f9c00 ACM |
1056 | rwbs[i++] = 'W'; |
1057 | else if (t->bytes) | |
1058 | rwbs[i++] = 'R'; | |
1059 | else | |
1060 | rwbs[i++] = 'N'; | |
1061 | ||
c09c47ca NK |
1062 | if (tc & BLK_TC_FUA) |
1063 | rwbs[i++] = 'F'; | |
65796348 | 1064 | if (tc & BLK_TC_AHEAD) |
157f9c00 | 1065 | rwbs[i++] = 'A'; |
65796348 | 1066 | if (tc & BLK_TC_SYNC) |
157f9c00 | 1067 | rwbs[i++] = 'S'; |
65796348 | 1068 | if (tc & BLK_TC_META) |
157f9c00 | 1069 | rwbs[i++] = 'M'; |
18cea459 | 1070 | out: |
157f9c00 | 1071 | rwbs[i] = '\0'; |
c71a8961 ACM |
1072 | } |
1073 | ||
1074 | static inline | |
1075 | const struct blk_io_trace *te_blk_io_trace(const struct trace_entry *ent) | |
1076 | { | |
1077 | return (const struct blk_io_trace *)ent; | |
1078 | } | |
1079 | ||
1080 | static inline const void *pdu_start(const struct trace_entry *ent) | |
1081 | { | |
1082 | return te_blk_io_trace(ent) + 1; | |
1083 | } | |
1084 | ||
66de7792 LZ |
1085 | static inline u32 t_action(const struct trace_entry *ent) |
1086 | { | |
1087 | return te_blk_io_trace(ent)->action; | |
1088 | } | |
1089 | ||
1090 | static inline u32 t_bytes(const struct trace_entry *ent) | |
1091 | { | |
1092 | return te_blk_io_trace(ent)->bytes; | |
1093 | } | |
1094 | ||
c71a8961 ACM |
1095 | static inline u32 t_sec(const struct trace_entry *ent) |
1096 | { | |
1097 | return te_blk_io_trace(ent)->bytes >> 9; | |
1098 | } | |
1099 | ||
1100 | static inline unsigned long long t_sector(const struct trace_entry *ent) | |
1101 | { | |
1102 | return te_blk_io_trace(ent)->sector; | |
1103 | } | |
1104 | ||
1105 | static inline __u16 t_error(const struct trace_entry *ent) | |
1106 | { | |
e0dc81be | 1107 | return te_blk_io_trace(ent)->error; |
c71a8961 ACM |
1108 | } |
1109 | ||
1110 | static __u64 get_pdu_int(const struct trace_entry *ent) | |
1111 | { | |
1112 | const __u64 *val = pdu_start(ent); | |
1113 | return be64_to_cpu(*val); | |
1114 | } | |
1115 | ||
1116 | static void get_pdu_remap(const struct trace_entry *ent, | |
1117 | struct blk_io_trace_remap *r) | |
1118 | { | |
1119 | const struct blk_io_trace_remap *__r = pdu_start(ent); | |
a42aaa3b | 1120 | __u64 sector_from = __r->sector_from; |
c71a8961 | 1121 | |
c71a8961 | 1122 | r->device_from = be32_to_cpu(__r->device_from); |
a42aaa3b AB |
1123 | r->device_to = be32_to_cpu(__r->device_to); |
1124 | r->sector_from = be64_to_cpu(sector_from); | |
c71a8961 ACM |
1125 | } |
1126 | ||
f4a1d08c | 1127 | typedef void (blk_log_action_t) (struct trace_iterator *iter, const char *act); |
b6a4b0c3 | 1128 | |
f4a1d08c | 1129 | static void blk_log_action_classic(struct trace_iterator *iter, const char *act) |
c71a8961 | 1130 | { |
c09c47ca | 1131 | char rwbs[RWBS_LEN]; |
35ac51bf LZ |
1132 | unsigned long long ts = iter->ts; |
1133 | unsigned long nsec_rem = do_div(ts, NSEC_PER_SEC); | |
c71a8961 | 1134 | unsigned secs = (unsigned long)ts; |
b6a4b0c3 | 1135 | const struct blk_io_trace *t = te_blk_io_trace(iter->ent); |
c71a8961 ACM |
1136 | |
1137 | fill_rwbs(rwbs, t); | |
1138 | ||
f4a1d08c SRRH |
1139 | trace_seq_printf(&iter->seq, |
1140 | "%3d,%-3d %2d %5d.%09lu %5u %2s %3s ", | |
1141 | MAJOR(t->device), MINOR(t->device), iter->cpu, | |
1142 | secs, nsec_rem, iter->ent->pid, act, rwbs); | |
c71a8961 ACM |
1143 | } |
1144 | ||
f4a1d08c | 1145 | static void blk_log_action(struct trace_iterator *iter, const char *act) |
c71a8961 | 1146 | { |
c09c47ca | 1147 | char rwbs[RWBS_LEN]; |
b6a4b0c3 LZ |
1148 | const struct blk_io_trace *t = te_blk_io_trace(iter->ent); |
1149 | ||
c71a8961 | 1150 | fill_rwbs(rwbs, t); |
f4a1d08c SRRH |
1151 | trace_seq_printf(&iter->seq, "%3d,%-3d %2s %3s ", |
1152 | MAJOR(t->device), MINOR(t->device), act, rwbs); | |
c71a8961 ACM |
1153 | } |
1154 | ||
f4a1d08c | 1155 | static void blk_log_dump_pdu(struct trace_seq *s, const struct trace_entry *ent) |
66de7792 | 1156 | { |
04986257 | 1157 | const unsigned char *pdu_buf; |
66de7792 | 1158 | int pdu_len; |
f4a1d08c | 1159 | int i, end; |
66de7792 LZ |
1160 | |
1161 | pdu_buf = pdu_start(ent); | |
1162 | pdu_len = te_blk_io_trace(ent)->pdu_len; | |
1163 | ||
1164 | if (!pdu_len) | |
f4a1d08c | 1165 | return; |
66de7792 LZ |
1166 | |
1167 | /* find the last zero that needs to be printed */ | |
1168 | for (end = pdu_len - 1; end >= 0; end--) | |
1169 | if (pdu_buf[end]) | |
1170 | break; | |
1171 | end++; | |
1172 | ||
f4a1d08c | 1173 | trace_seq_putc(s, '('); |
66de7792 LZ |
1174 | |
1175 | for (i = 0; i < pdu_len; i++) { | |
1176 | ||
f4a1d08c SRRH |
1177 | trace_seq_printf(s, "%s%02x", |
1178 | i == 0 ? "" : " ", pdu_buf[i]); | |
66de7792 LZ |
1179 | |
1180 | /* | |
1181 | * stop when the rest is just zeroes and indicate so | |
1182 | * with a ".." appended | |
1183 | */ | |
f4a1d08c SRRH |
1184 | if (i == end && end != pdu_len - 1) { |
1185 | trace_seq_puts(s, " ..) "); | |
1186 | return; | |
1187 | } | |
66de7792 LZ |
1188 | } |
1189 | ||
f4a1d08c | 1190 | trace_seq_puts(s, ") "); |
66de7792 LZ |
1191 | } |
1192 | ||
f4a1d08c | 1193 | static void blk_log_generic(struct trace_seq *s, const struct trace_entry *ent) |
c71a8961 | 1194 | { |
4ca53085 SR |
1195 | char cmd[TASK_COMM_LEN]; |
1196 | ||
1197 | trace_find_cmdline(ent->pid, cmd); | |
c71a8961 | 1198 | |
66de7792 | 1199 | if (t_action(ent) & BLK_TC_ACT(BLK_TC_PC)) { |
f4a1d08c SRRH |
1200 | trace_seq_printf(s, "%u ", t_bytes(ent)); |
1201 | blk_log_dump_pdu(s, ent); | |
1202 | trace_seq_printf(s, "[%s]\n", cmd); | |
66de7792 LZ |
1203 | } else { |
1204 | if (t_sec(ent)) | |
f4a1d08c | 1205 | trace_seq_printf(s, "%llu + %u [%s]\n", |
66de7792 | 1206 | t_sector(ent), t_sec(ent), cmd); |
f4a1d08c SRRH |
1207 | else |
1208 | trace_seq_printf(s, "[%s]\n", cmd); | |
66de7792 | 1209 | } |
c71a8961 ACM |
1210 | } |
1211 | ||
f4a1d08c | 1212 | static void blk_log_with_error(struct trace_seq *s, |
157f9c00 | 1213 | const struct trace_entry *ent) |
c71a8961 | 1214 | { |
66de7792 | 1215 | if (t_action(ent) & BLK_TC_ACT(BLK_TC_PC)) { |
f4a1d08c SRRH |
1216 | blk_log_dump_pdu(s, ent); |
1217 | trace_seq_printf(s, "[%d]\n", t_error(ent)); | |
66de7792 LZ |
1218 | } else { |
1219 | if (t_sec(ent)) | |
f4a1d08c SRRH |
1220 | trace_seq_printf(s, "%llu + %u [%d]\n", |
1221 | t_sector(ent), | |
1222 | t_sec(ent), t_error(ent)); | |
1223 | else | |
1224 | trace_seq_printf(s, "%llu [%d]\n", | |
1225 | t_sector(ent), t_error(ent)); | |
66de7792 | 1226 | } |
c71a8961 ACM |
1227 | } |
1228 | ||
f4a1d08c | 1229 | static void blk_log_remap(struct trace_seq *s, const struct trace_entry *ent) |
c71a8961 | 1230 | { |
a42aaa3b | 1231 | struct blk_io_trace_remap r = { .device_from = 0, }; |
c71a8961 ACM |
1232 | |
1233 | get_pdu_remap(ent, &r); | |
f4a1d08c SRRH |
1234 | trace_seq_printf(s, "%llu + %u <- (%d,%d) %llu\n", |
1235 | t_sector(ent), t_sec(ent), | |
1236 | MAJOR(r.device_from), MINOR(r.device_from), | |
1237 | (unsigned long long)r.sector_from); | |
c71a8961 ACM |
1238 | } |
1239 | ||
f4a1d08c | 1240 | static void blk_log_plug(struct trace_seq *s, const struct trace_entry *ent) |
c71a8961 | 1241 | { |
4ca53085 SR |
1242 | char cmd[TASK_COMM_LEN]; |
1243 | ||
1244 | trace_find_cmdline(ent->pid, cmd); | |
1245 | ||
f4a1d08c | 1246 | trace_seq_printf(s, "[%s]\n", cmd); |
c71a8961 ACM |
1247 | } |
1248 | ||
f4a1d08c | 1249 | static void blk_log_unplug(struct trace_seq *s, const struct trace_entry *ent) |
c71a8961 | 1250 | { |
4ca53085 SR |
1251 | char cmd[TASK_COMM_LEN]; |
1252 | ||
1253 | trace_find_cmdline(ent->pid, cmd); | |
1254 | ||
f4a1d08c | 1255 | trace_seq_printf(s, "[%s] %llu\n", cmd, get_pdu_int(ent)); |
c71a8961 ACM |
1256 | } |
1257 | ||
f4a1d08c | 1258 | static void blk_log_split(struct trace_seq *s, const struct trace_entry *ent) |
c71a8961 | 1259 | { |
4ca53085 SR |
1260 | char cmd[TASK_COMM_LEN]; |
1261 | ||
1262 | trace_find_cmdline(ent->pid, cmd); | |
1263 | ||
f4a1d08c SRRH |
1264 | trace_seq_printf(s, "%llu / %llu [%s]\n", t_sector(ent), |
1265 | get_pdu_int(ent), cmd); | |
c71a8961 ACM |
1266 | } |
1267 | ||
f4a1d08c | 1268 | static void blk_log_msg(struct trace_seq *s, const struct trace_entry *ent) |
18cea459 | 1269 | { |
18cea459 LZ |
1270 | const struct blk_io_trace *t = te_blk_io_trace(ent); |
1271 | ||
f4a1d08c SRRH |
1272 | trace_seq_putmem(s, t + 1, t->pdu_len); |
1273 | trace_seq_putc(s, '\n'); | |
18cea459 LZ |
1274 | } |
1275 | ||
c71a8961 ACM |
1276 | /* |
1277 | * struct tracer operations | |
1278 | */ | |
1279 | ||
1280 | static void blk_tracer_print_header(struct seq_file *m) | |
1281 | { | |
1282 | if (!(blk_tracer_flags.val & TRACE_BLK_OPT_CLASSIC)) | |
1283 | return; | |
1284 | seq_puts(m, "# DEV CPU TIMESTAMP PID ACT FLG\n" | |
1285 | "# | | | | | |\n"); | |
1286 | } | |
1287 | ||
1288 | static void blk_tracer_start(struct trace_array *tr) | |
1289 | { | |
ad5dd549 | 1290 | blk_tracer_enabled = true; |
c71a8961 ACM |
1291 | } |
1292 | ||
1293 | static int blk_tracer_init(struct trace_array *tr) | |
1294 | { | |
1295 | blk_tr = tr; | |
1296 | blk_tracer_start(tr); | |
c71a8961 ACM |
1297 | return 0; |
1298 | } | |
1299 | ||
1300 | static void blk_tracer_stop(struct trace_array *tr) | |
1301 | { | |
ad5dd549 | 1302 | blk_tracer_enabled = false; |
c71a8961 ACM |
1303 | } |
1304 | ||
1305 | static void blk_tracer_reset(struct trace_array *tr) | |
1306 | { | |
c71a8961 ACM |
1307 | blk_tracer_stop(tr); |
1308 | } | |
1309 | ||
e4955c99 | 1310 | static const struct { |
c71a8961 | 1311 | const char *act[2]; |
f4a1d08c | 1312 | void (*print)(struct trace_seq *s, const struct trace_entry *ent); |
e4955c99 | 1313 | } what2act[] = { |
ef18012b | 1314 | [__BLK_TA_QUEUE] = {{ "Q", "queue" }, blk_log_generic }, |
c71a8961 ACM |
1315 | [__BLK_TA_BACKMERGE] = {{ "M", "backmerge" }, blk_log_generic }, |
1316 | [__BLK_TA_FRONTMERGE] = {{ "F", "frontmerge" }, blk_log_generic }, | |
1317 | [__BLK_TA_GETRQ] = {{ "G", "getrq" }, blk_log_generic }, | |
1318 | [__BLK_TA_SLEEPRQ] = {{ "S", "sleeprq" }, blk_log_generic }, | |
1319 | [__BLK_TA_REQUEUE] = {{ "R", "requeue" }, blk_log_with_error }, | |
1320 | [__BLK_TA_ISSUE] = {{ "D", "issue" }, blk_log_generic }, | |
1321 | [__BLK_TA_COMPLETE] = {{ "C", "complete" }, blk_log_with_error }, | |
1322 | [__BLK_TA_PLUG] = {{ "P", "plug" }, blk_log_plug }, | |
1323 | [__BLK_TA_UNPLUG_IO] = {{ "U", "unplug_io" }, blk_log_unplug }, | |
49cac01e | 1324 | [__BLK_TA_UNPLUG_TIMER] = {{ "UT", "unplug_timer" }, blk_log_unplug }, |
c71a8961 ACM |
1325 | [__BLK_TA_INSERT] = {{ "I", "insert" }, blk_log_generic }, |
1326 | [__BLK_TA_SPLIT] = {{ "X", "split" }, blk_log_split }, | |
1327 | [__BLK_TA_BOUNCE] = {{ "B", "bounce" }, blk_log_generic }, | |
1328 | [__BLK_TA_REMAP] = {{ "A", "remap" }, blk_log_remap }, | |
1329 | }; | |
1330 | ||
b6a4b0c3 LZ |
1331 | static enum print_line_t print_one_line(struct trace_iterator *iter, |
1332 | bool classic) | |
c71a8961 | 1333 | { |
983f938a | 1334 | struct trace_array *tr = iter->tr; |
2c9b238e | 1335 | struct trace_seq *s = &iter->seq; |
b6a4b0c3 LZ |
1336 | const struct blk_io_trace *t; |
1337 | u16 what; | |
b6a4b0c3 LZ |
1338 | bool long_act; |
1339 | blk_log_action_t *log_action; | |
c71a8961 | 1340 | |
b6a4b0c3 LZ |
1341 | t = te_blk_io_trace(iter->ent); |
1342 | what = t->action & ((1 << BLK_TC_SHIFT) - 1); | |
983f938a | 1343 | long_act = !!(tr->trace_flags & TRACE_ITER_VERBOSE); |
b6a4b0c3 | 1344 | log_action = classic ? &blk_log_action_classic : &blk_log_action; |
08a06b83 | 1345 | |
18cea459 | 1346 | if (t->action == BLK_TN_MESSAGE) { |
f4a1d08c SRRH |
1347 | log_action(iter, long_act ? "message" : "m"); |
1348 | blk_log_msg(s, iter->ent); | |
b7d7641e | 1349 | return trace_handle_return(s); |
18cea459 LZ |
1350 | } |
1351 | ||
eb08f8eb | 1352 | if (unlikely(what == 0 || what >= ARRAY_SIZE(what2act))) |
f4a1d08c | 1353 | trace_seq_printf(s, "Unknown action %x\n", what); |
c71a8961 | 1354 | else { |
f4a1d08c SRRH |
1355 | log_action(iter, what2act[what].act[long_act]); |
1356 | what2act[what].print(s, iter->ent); | |
c71a8961 | 1357 | } |
f4a1d08c SRRH |
1358 | |
1359 | return trace_handle_return(s); | |
c71a8961 ACM |
1360 | } |
1361 | ||
b6a4b0c3 | 1362 | static enum print_line_t blk_trace_event_print(struct trace_iterator *iter, |
a9a57763 | 1363 | int flags, struct trace_event *event) |
b6a4b0c3 | 1364 | { |
b6a4b0c3 LZ |
1365 | return print_one_line(iter, false); |
1366 | } | |
1367 | ||
f4a1d08c | 1368 | static void blk_trace_synthesize_old_trace(struct trace_iterator *iter) |
08a06b83 ACM |
1369 | { |
1370 | struct trace_seq *s = &iter->seq; | |
1371 | struct blk_io_trace *t = (struct blk_io_trace *)iter->ent; | |
1372 | const int offset = offsetof(struct blk_io_trace, sector); | |
1373 | struct blk_io_trace old = { | |
1374 | .magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION, | |
6c051ce0 | 1375 | .time = iter->ts, |
08a06b83 ACM |
1376 | }; |
1377 | ||
f4a1d08c SRRH |
1378 | trace_seq_putmem(s, &old, offset); |
1379 | trace_seq_putmem(s, &t->sector, | |
1380 | sizeof(old) - offset + t->pdu_len); | |
08a06b83 ACM |
1381 | } |
1382 | ||
ae7462b4 | 1383 | static enum print_line_t |
a9a57763 SR |
1384 | blk_trace_event_print_binary(struct trace_iterator *iter, int flags, |
1385 | struct trace_event *event) | |
08a06b83 | 1386 | { |
f4a1d08c SRRH |
1387 | blk_trace_synthesize_old_trace(iter); |
1388 | ||
1389 | return trace_handle_return(&iter->seq); | |
08a06b83 ACM |
1390 | } |
1391 | ||
c71a8961 ACM |
1392 | static enum print_line_t blk_tracer_print_line(struct trace_iterator *iter) |
1393 | { | |
c71a8961 ACM |
1394 | if (!(blk_tracer_flags.val & TRACE_BLK_OPT_CLASSIC)) |
1395 | return TRACE_TYPE_UNHANDLED; | |
1396 | ||
b6a4b0c3 | 1397 | return print_one_line(iter, true); |
c71a8961 ACM |
1398 | } |
1399 | ||
8c1a49ae SRRH |
1400 | static int |
1401 | blk_tracer_set_flag(struct trace_array *tr, u32 old_flags, u32 bit, int set) | |
f3948f88 LZ |
1402 | { |
1403 | /* don't output context-info for blk_classic output */ | |
1404 | if (bit == TRACE_BLK_OPT_CLASSIC) { | |
1405 | if (set) | |
983f938a | 1406 | tr->trace_flags &= ~TRACE_ITER_CONTEXT_INFO; |
f3948f88 | 1407 | else |
983f938a | 1408 | tr->trace_flags |= TRACE_ITER_CONTEXT_INFO; |
f3948f88 LZ |
1409 | } |
1410 | return 0; | |
1411 | } | |
1412 | ||
c71a8961 ACM |
1413 | static struct tracer blk_tracer __read_mostly = { |
1414 | .name = "blk", | |
1415 | .init = blk_tracer_init, | |
1416 | .reset = blk_tracer_reset, | |
1417 | .start = blk_tracer_start, | |
1418 | .stop = blk_tracer_stop, | |
1419 | .print_header = blk_tracer_print_header, | |
1420 | .print_line = blk_tracer_print_line, | |
1421 | .flags = &blk_tracer_flags, | |
f3948f88 | 1422 | .set_flag = blk_tracer_set_flag, |
c71a8961 ACM |
1423 | }; |
1424 | ||
a9a57763 | 1425 | static struct trace_event_functions trace_blk_event_funcs = { |
c71a8961 | 1426 | .trace = blk_trace_event_print, |
08a06b83 | 1427 | .binary = blk_trace_event_print_binary, |
c71a8961 ACM |
1428 | }; |
1429 | ||
a9a57763 SR |
1430 | static struct trace_event trace_blk_event = { |
1431 | .type = TRACE_BLK, | |
1432 | .funcs = &trace_blk_event_funcs, | |
1433 | }; | |
1434 | ||
c71a8961 ACM |
1435 | static int __init init_blk_tracer(void) |
1436 | { | |
9023c930 | 1437 | if (!register_trace_event(&trace_blk_event)) { |
a395d6a7 | 1438 | pr_warn("Warning: could not register block events\n"); |
c71a8961 ACM |
1439 | return 1; |
1440 | } | |
1441 | ||
1442 | if (register_tracer(&blk_tracer) != 0) { | |
a395d6a7 | 1443 | pr_warn("Warning: could not register the block tracer\n"); |
9023c930 | 1444 | unregister_trace_event(&trace_blk_event); |
c71a8961 ACM |
1445 | return 1; |
1446 | } | |
1447 | ||
1448 | return 0; | |
1449 | } | |
1450 | ||
1451 | device_initcall(init_blk_tracer); | |
1452 | ||
1453 | static int blk_trace_remove_queue(struct request_queue *q) | |
1454 | { | |
1455 | struct blk_trace *bt; | |
1456 | ||
1457 | bt = xchg(&q->blk_trace, NULL); | |
1458 | if (bt == NULL) | |
1459 | return -EINVAL; | |
1460 | ||
17ba97e3 LZ |
1461 | if (atomic_dec_and_test(&blk_probes_ref)) |
1462 | blk_unregister_tracepoints(); | |
1463 | ||
ad5dd549 | 1464 | blk_trace_free(bt); |
c71a8961 ACM |
1465 | return 0; |
1466 | } | |
1467 | ||
1468 | /* | |
1469 | * Setup everything required to start tracing | |
1470 | */ | |
9908c309 LZ |
1471 | static int blk_trace_setup_queue(struct request_queue *q, |
1472 | struct block_device *bdev) | |
c71a8961 | 1473 | { |
cdea01b2 | 1474 | struct blk_trace *bt = NULL; |
18cea459 | 1475 | int ret = -ENOMEM; |
c71a8961 | 1476 | |
c71a8961 ACM |
1477 | bt = kzalloc(sizeof(*bt), GFP_KERNEL); |
1478 | if (!bt) | |
15152e44 | 1479 | return -ENOMEM; |
c71a8961 | 1480 | |
18cea459 LZ |
1481 | bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG, __alignof__(char)); |
1482 | if (!bt->msg_data) | |
1483 | goto free_bt; | |
1484 | ||
9908c309 | 1485 | bt->dev = bdev->bd_dev; |
c71a8961 | 1486 | bt->act_mask = (u16)-1; |
9908c309 LZ |
1487 | |
1488 | blk_trace_setup_lba(bt, bdev); | |
c71a8961 | 1489 | |
cdea01b2 DB |
1490 | ret = -EBUSY; |
1491 | if (cmpxchg(&q->blk_trace, NULL, bt)) | |
18cea459 | 1492 | goto free_bt; |
15152e44 | 1493 | |
17ba97e3 LZ |
1494 | if (atomic_inc_return(&blk_probes_ref) == 1) |
1495 | blk_register_tracepoints(); | |
c71a8961 | 1496 | return 0; |
18cea459 LZ |
1497 | |
1498 | free_bt: | |
1499 | blk_trace_free(bt); | |
1500 | return ret; | |
c71a8961 ACM |
1501 | } |
1502 | ||
1503 | /* | |
1504 | * sysfs interface to enable and configure tracing | |
1505 | */ | |
1506 | ||
c71a8961 ACM |
1507 | static ssize_t sysfs_blk_trace_attr_show(struct device *dev, |
1508 | struct device_attribute *attr, | |
1509 | char *buf); | |
1510 | static ssize_t sysfs_blk_trace_attr_store(struct device *dev, | |
1511 | struct device_attribute *attr, | |
1512 | const char *buf, size_t count); | |
1513 | #define BLK_TRACE_DEVICE_ATTR(_name) \ | |
1514 | DEVICE_ATTR(_name, S_IRUGO | S_IWUSR, \ | |
1515 | sysfs_blk_trace_attr_show, \ | |
1516 | sysfs_blk_trace_attr_store) | |
1517 | ||
cd649b8b | 1518 | static BLK_TRACE_DEVICE_ATTR(enable); |
c71a8961 ACM |
1519 | static BLK_TRACE_DEVICE_ATTR(act_mask); |
1520 | static BLK_TRACE_DEVICE_ATTR(pid); | |
1521 | static BLK_TRACE_DEVICE_ATTR(start_lba); | |
1522 | static BLK_TRACE_DEVICE_ATTR(end_lba); | |
1523 | ||
1524 | static struct attribute *blk_trace_attrs[] = { | |
1525 | &dev_attr_enable.attr, | |
1526 | &dev_attr_act_mask.attr, | |
1527 | &dev_attr_pid.attr, | |
1528 | &dev_attr_start_lba.attr, | |
1529 | &dev_attr_end_lba.attr, | |
1530 | NULL | |
1531 | }; | |
1532 | ||
1533 | struct attribute_group blk_trace_attr_group = { | |
1534 | .name = "trace", | |
1535 | .attrs = blk_trace_attrs, | |
1536 | }; | |
1537 | ||
09341997 LZ |
1538 | static const struct { |
1539 | int mask; | |
1540 | const char *str; | |
1541 | } mask_maps[] = { | |
1542 | { BLK_TC_READ, "read" }, | |
1543 | { BLK_TC_WRITE, "write" }, | |
c09c47ca | 1544 | { BLK_TC_FLUSH, "flush" }, |
09341997 LZ |
1545 | { BLK_TC_SYNC, "sync" }, |
1546 | { BLK_TC_QUEUE, "queue" }, | |
1547 | { BLK_TC_REQUEUE, "requeue" }, | |
1548 | { BLK_TC_ISSUE, "issue" }, | |
1549 | { BLK_TC_COMPLETE, "complete" }, | |
1550 | { BLK_TC_FS, "fs" }, | |
1551 | { BLK_TC_PC, "pc" }, | |
8d1547e0 | 1552 | { BLK_TC_NOTIFY, "notify" }, |
09341997 LZ |
1553 | { BLK_TC_AHEAD, "ahead" }, |
1554 | { BLK_TC_META, "meta" }, | |
1555 | { BLK_TC_DISCARD, "discard" }, | |
1556 | { BLK_TC_DRV_DATA, "drv_data" }, | |
c09c47ca | 1557 | { BLK_TC_FUA, "fua" }, |
09341997 LZ |
1558 | }; |
1559 | ||
1560 | static int blk_trace_str2mask(const char *str) | |
c71a8961 | 1561 | { |
09341997 | 1562 | int i; |
c71a8961 | 1563 | int mask = 0; |
9eb85125 | 1564 | char *buf, *s, *token; |
c71a8961 | 1565 | |
9eb85125 LZ |
1566 | buf = kstrdup(str, GFP_KERNEL); |
1567 | if (buf == NULL) | |
c71a8961 | 1568 | return -ENOMEM; |
9eb85125 | 1569 | s = strstrip(buf); |
c71a8961 ACM |
1570 | |
1571 | while (1) { | |
09341997 LZ |
1572 | token = strsep(&s, ","); |
1573 | if (token == NULL) | |
c71a8961 ACM |
1574 | break; |
1575 | ||
09341997 LZ |
1576 | if (*token == '\0') |
1577 | continue; | |
1578 | ||
1579 | for (i = 0; i < ARRAY_SIZE(mask_maps); i++) { | |
1580 | if (strcasecmp(token, mask_maps[i].str) == 0) { | |
1581 | mask |= mask_maps[i].mask; | |
1582 | break; | |
1583 | } | |
1584 | } | |
1585 | if (i == ARRAY_SIZE(mask_maps)) { | |
1586 | mask = -EINVAL; | |
1587 | break; | |
1588 | } | |
c71a8961 | 1589 | } |
9eb85125 | 1590 | kfree(buf); |
c71a8961 ACM |
1591 | |
1592 | return mask; | |
1593 | } | |
1594 | ||
09341997 LZ |
1595 | static ssize_t blk_trace_mask2str(char *buf, int mask) |
1596 | { | |
1597 | int i; | |
1598 | char *p = buf; | |
1599 | ||
1600 | for (i = 0; i < ARRAY_SIZE(mask_maps); i++) { | |
1601 | if (mask & mask_maps[i].mask) { | |
1602 | p += sprintf(p, "%s%s", | |
1603 | (p == buf) ? "" : ",", mask_maps[i].str); | |
1604 | } | |
1605 | } | |
1606 | *p++ = '\n'; | |
1607 | ||
1608 | return p - buf; | |
1609 | } | |
1610 | ||
b125130b LZ |
1611 | static struct request_queue *blk_trace_get_queue(struct block_device *bdev) |
1612 | { | |
1613 | if (bdev->bd_disk == NULL) | |
1614 | return NULL; | |
1615 | ||
1616 | return bdev_get_queue(bdev); | |
1617 | } | |
1618 | ||
c71a8961 ACM |
1619 | static ssize_t sysfs_blk_trace_attr_show(struct device *dev, |
1620 | struct device_attribute *attr, | |
1621 | char *buf) | |
1622 | { | |
1623 | struct hd_struct *p = dev_to_part(dev); | |
1624 | struct request_queue *q; | |
1625 | struct block_device *bdev; | |
1626 | ssize_t ret = -ENXIO; | |
1627 | ||
c71a8961 ACM |
1628 | bdev = bdget(part_devt(p)); |
1629 | if (bdev == NULL) | |
01b284f9 | 1630 | goto out; |
c71a8961 | 1631 | |
b125130b | 1632 | q = blk_trace_get_queue(bdev); |
c71a8961 ACM |
1633 | if (q == NULL) |
1634 | goto out_bdput; | |
b125130b | 1635 | |
c71a8961 | 1636 | mutex_lock(&bdev->bd_mutex); |
cd649b8b LZ |
1637 | |
1638 | if (attr == &dev_attr_enable) { | |
1639 | ret = sprintf(buf, "%u\n", !!q->blk_trace); | |
1640 | goto out_unlock_bdev; | |
1641 | } | |
1642 | ||
c71a8961 ACM |
1643 | if (q->blk_trace == NULL) |
1644 | ret = sprintf(buf, "disabled\n"); | |
1645 | else if (attr == &dev_attr_act_mask) | |
09341997 | 1646 | ret = blk_trace_mask2str(buf, q->blk_trace->act_mask); |
c71a8961 ACM |
1647 | else if (attr == &dev_attr_pid) |
1648 | ret = sprintf(buf, "%u\n", q->blk_trace->pid); | |
1649 | else if (attr == &dev_attr_start_lba) | |
1650 | ret = sprintf(buf, "%llu\n", q->blk_trace->start_lba); | |
1651 | else if (attr == &dev_attr_end_lba) | |
1652 | ret = sprintf(buf, "%llu\n", q->blk_trace->end_lba); | |
cd649b8b LZ |
1653 | |
1654 | out_unlock_bdev: | |
c71a8961 ACM |
1655 | mutex_unlock(&bdev->bd_mutex); |
1656 | out_bdput: | |
1657 | bdput(bdev); | |
01b284f9 | 1658 | out: |
c71a8961 ACM |
1659 | return ret; |
1660 | } | |
1661 | ||
1662 | static ssize_t sysfs_blk_trace_attr_store(struct device *dev, | |
1663 | struct device_attribute *attr, | |
1664 | const char *buf, size_t count) | |
1665 | { | |
1666 | struct block_device *bdev; | |
1667 | struct request_queue *q; | |
1668 | struct hd_struct *p; | |
1669 | u64 value; | |
09341997 | 1670 | ssize_t ret = -EINVAL; |
c71a8961 ACM |
1671 | |
1672 | if (count == 0) | |
1673 | goto out; | |
1674 | ||
1675 | if (attr == &dev_attr_act_mask) { | |
1676 | if (sscanf(buf, "%llx", &value) != 1) { | |
1677 | /* Assume it is a list of trace category names */ | |
09341997 LZ |
1678 | ret = blk_trace_str2mask(buf); |
1679 | if (ret < 0) | |
c71a8961 | 1680 | goto out; |
09341997 | 1681 | value = ret; |
c71a8961 ACM |
1682 | } |
1683 | } else if (sscanf(buf, "%llu", &value) != 1) | |
1684 | goto out; | |
1685 | ||
09341997 LZ |
1686 | ret = -ENXIO; |
1687 | ||
c71a8961 ACM |
1688 | p = dev_to_part(dev); |
1689 | bdev = bdget(part_devt(p)); | |
1690 | if (bdev == NULL) | |
01b284f9 | 1691 | goto out; |
c71a8961 | 1692 | |
b125130b | 1693 | q = blk_trace_get_queue(bdev); |
c71a8961 ACM |
1694 | if (q == NULL) |
1695 | goto out_bdput; | |
1696 | ||
1697 | mutex_lock(&bdev->bd_mutex); | |
cd649b8b LZ |
1698 | |
1699 | if (attr == &dev_attr_enable) { | |
1700 | if (value) | |
9908c309 | 1701 | ret = blk_trace_setup_queue(q, bdev); |
cd649b8b LZ |
1702 | else |
1703 | ret = blk_trace_remove_queue(q); | |
1704 | goto out_unlock_bdev; | |
1705 | } | |
1706 | ||
c71a8961 ACM |
1707 | ret = 0; |
1708 | if (q->blk_trace == NULL) | |
9908c309 | 1709 | ret = blk_trace_setup_queue(q, bdev); |
c71a8961 ACM |
1710 | |
1711 | if (ret == 0) { | |
1712 | if (attr == &dev_attr_act_mask) | |
1713 | q->blk_trace->act_mask = value; | |
1714 | else if (attr == &dev_attr_pid) | |
1715 | q->blk_trace->pid = value; | |
1716 | else if (attr == &dev_attr_start_lba) | |
1717 | q->blk_trace->start_lba = value; | |
1718 | else if (attr == &dev_attr_end_lba) | |
1719 | q->blk_trace->end_lba = value; | |
c71a8961 | 1720 | } |
cd649b8b LZ |
1721 | |
1722 | out_unlock_bdev: | |
c71a8961 ACM |
1723 | mutex_unlock(&bdev->bd_mutex); |
1724 | out_bdput: | |
1725 | bdput(bdev); | |
c71a8961 | 1726 | out: |
cd649b8b | 1727 | return ret ? ret : count; |
c71a8961 | 1728 | } |
cd649b8b | 1729 | |
1d54ad6d LZ |
1730 | int blk_trace_init_sysfs(struct device *dev) |
1731 | { | |
1732 | return sysfs_create_group(&dev->kobj, &blk_trace_attr_group); | |
1733 | } | |
1734 | ||
48c0d4d4 ZK |
1735 | void blk_trace_remove_sysfs(struct device *dev) |
1736 | { | |
1737 | sysfs_remove_group(&dev->kobj, &blk_trace_attr_group); | |
1738 | } | |
1739 | ||
55782138 LZ |
1740 | #endif /* CONFIG_BLK_DEV_IO_TRACE */ |
1741 | ||
1742 | #ifdef CONFIG_EVENT_TRACING | |
1743 | ||
ef295ecf | 1744 | void blk_fill_rwbs(char *rwbs, unsigned int op, int bytes) |
55782138 LZ |
1745 | { |
1746 | int i = 0; | |
1747 | ||
ef295ecf | 1748 | if (op & REQ_PREFLUSH) |
c09c47ca NK |
1749 | rwbs[i++] = 'F'; |
1750 | ||
ef295ecf | 1751 | switch (op & REQ_OP_MASK) { |
1b9a9ab7 MC |
1752 | case REQ_OP_WRITE: |
1753 | case REQ_OP_WRITE_SAME: | |
55782138 | 1754 | rwbs[i++] = 'W'; |
1b9a9ab7 MC |
1755 | break; |
1756 | case REQ_OP_DISCARD: | |
55782138 | 1757 | rwbs[i++] = 'D'; |
1b9a9ab7 | 1758 | break; |
288dab8a CH |
1759 | case REQ_OP_SECURE_ERASE: |
1760 | rwbs[i++] = 'D'; | |
1761 | rwbs[i++] = 'E'; | |
1762 | break; | |
3a5e02ce MC |
1763 | case REQ_OP_FLUSH: |
1764 | rwbs[i++] = 'F'; | |
1765 | break; | |
1b9a9ab7 | 1766 | case REQ_OP_READ: |
55782138 | 1767 | rwbs[i++] = 'R'; |
1b9a9ab7 MC |
1768 | break; |
1769 | default: | |
55782138 | 1770 | rwbs[i++] = 'N'; |
1b9a9ab7 | 1771 | } |
55782138 | 1772 | |
ef295ecf | 1773 | if (op & REQ_FUA) |
c09c47ca | 1774 | rwbs[i++] = 'F'; |
ef295ecf | 1775 | if (op & REQ_RAHEAD) |
55782138 | 1776 | rwbs[i++] = 'A'; |
ef295ecf | 1777 | if (op & REQ_SYNC) |
55782138 | 1778 | rwbs[i++] = 'S'; |
ef295ecf | 1779 | if (op & REQ_META) |
55782138 LZ |
1780 | rwbs[i++] = 'M'; |
1781 | ||
1782 | rwbs[i] = '\0'; | |
1783 | } | |
9ca8f8e5 | 1784 | EXPORT_SYMBOL_GPL(blk_fill_rwbs); |
55782138 | 1785 | |
55782138 LZ |
1786 | #endif /* CONFIG_EVENT_TRACING */ |
1787 |