blob: 265f7a830619ef1e5de60bfe1179fa9d3a002229 [file] [log] [blame]
Jens Axboe2056a782006-03-23 20:00:26 +01001/*
2 * Copyright (C) 2006 Jens Axboe <axboe@suse.de>
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 */
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>
25#include <asm/uaccess.h>
26
27static DEFINE_PER_CPU(unsigned long long, blk_trace_cpu_offset) = { 0, };
28static unsigned int blktrace_seq __read_mostly = 1;
29
30/*
31 * Send out a notify for this process, if we haven't done so since a trace
32 * started
33 */
34static void trace_note_tsk(struct blk_trace *bt, struct task_struct *tsk)
35{
36 struct blk_io_trace *t;
37
38 t = relay_reserve(bt->rchan, sizeof(*t) + sizeof(tsk->comm));
39 if (t) {
40 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
41 t->device = bt->dev;
42 t->action = BLK_TC_ACT(BLK_TC_NOTIFY);
43 t->pid = tsk->pid;
44 t->cpu = smp_processor_id();
45 t->pdu_len = sizeof(tsk->comm);
46 memcpy((void *) t + sizeof(*t), tsk->comm, t->pdu_len);
47 tsk->btrace_seq = blktrace_seq;
48 }
49}
50
51static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
52 pid_t pid)
53{
54 if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
55 return 1;
56 if (sector < bt->start_lba || sector > bt->end_lba)
57 return 1;
58 if (bt->pid && pid != bt->pid)
59 return 1;
60
61 return 0;
62}
63
64/*
65 * Data direction bit lookup
66 */
67static u32 ddir_act[2] __read_mostly = { BLK_TC_ACT(BLK_TC_READ), BLK_TC_ACT(BLK_TC_WRITE) };
68
69/*
70 * Bio action bits of interest
71 */
Nathan Scott40359cc2006-07-06 10:03:28 +020072static u32 bio_act[5] __read_mostly = { 0, BLK_TC_ACT(BLK_TC_BARRIER), BLK_TC_ACT(BLK_TC_SYNC), 0, BLK_TC_ACT(BLK_TC_AHEAD) };
Jens Axboe2056a782006-03-23 20:00:26 +010073
74/*
75 * More could be added as needed, taking care to increment the decrementer
76 * to get correct indexing
77 */
78#define trace_barrier_bit(rw) \
79 (((rw) & (1 << BIO_RW_BARRIER)) >> (BIO_RW_BARRIER - 0))
80#define trace_sync_bit(rw) \
81 (((rw) & (1 << BIO_RW_SYNC)) >> (BIO_RW_SYNC - 1))
Nathan Scott40359cc2006-07-06 10:03:28 +020082#define trace_ahead_bit(rw) \
Milton Millerad01b1c2006-07-25 15:04:13 +020083 (((rw) & (1 << BIO_RW_AHEAD)) << (2 - BIO_RW_AHEAD))
Jens Axboe2056a782006-03-23 20:00:26 +010084
85/*
86 * The worker for the various blk_add_trace*() types. Fills out a
87 * blk_io_trace structure and places it in a per-cpu subbuffer.
88 */
89void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
90 int rw, u32 what, int error, int pdu_len, void *pdu_data)
91{
92 struct task_struct *tsk = current;
93 struct blk_io_trace *t;
94 unsigned long flags;
95 unsigned long *sequence;
96 pid_t pid;
97 int cpu;
98
99 if (unlikely(bt->trace_state != Blktrace_running))
100 return;
101
102 what |= ddir_act[rw & WRITE];
103 what |= bio_act[trace_barrier_bit(rw)];
104 what |= bio_act[trace_sync_bit(rw)];
Nathan Scott40359cc2006-07-06 10:03:28 +0200105 what |= bio_act[trace_ahead_bit(rw)];
Jens Axboe2056a782006-03-23 20:00:26 +0100106
107 pid = tsk->pid;
108 if (unlikely(act_log_check(bt, what, sector, pid)))
109 return;
110
111 /*
112 * A word about the locking here - we disable interrupts to reserve
113 * some space in the relay per-cpu buffer, to prevent an irq
114 * from coming in and stepping on our toes. Once reserved, it's
115 * enough to get preemption disabled to prevent read of this data
116 * before we are through filling it. get_cpu()/put_cpu() does this
117 * for us
118 */
119 local_irq_save(flags);
120
121 if (unlikely(tsk->btrace_seq != blktrace_seq))
122 trace_note_tsk(bt, tsk);
123
124 t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
125 if (t) {
126 cpu = smp_processor_id();
127 sequence = per_cpu_ptr(bt->sequence, cpu);
128
129 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
130 t->sequence = ++(*sequence);
131 t->time = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
132 t->sector = sector;
133 t->bytes = bytes;
134 t->action = what;
135 t->pid = pid;
136 t->device = bt->dev;
137 t->cpu = cpu;
138 t->error = error;
139 t->pdu_len = pdu_len;
140
141 if (pdu_len)
142 memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
143 }
144
145 local_irq_restore(flags);
146}
147
148EXPORT_SYMBOL_GPL(__blk_add_trace);
149
150static struct dentry *blk_tree_root;
151static struct mutex blk_tree_mutex;
152static unsigned int root_users;
153
154static inline void blk_remove_root(void)
155{
156 if (blk_tree_root) {
157 debugfs_remove(blk_tree_root);
158 blk_tree_root = NULL;
159 }
160}
161
162static void blk_remove_tree(struct dentry *dir)
163{
164 mutex_lock(&blk_tree_mutex);
165 debugfs_remove(dir);
166 if (--root_users == 0)
167 blk_remove_root();
168 mutex_unlock(&blk_tree_mutex);
169}
170
171static struct dentry *blk_create_tree(const char *blk_name)
172{
173 struct dentry *dir = NULL;
174
175 mutex_lock(&blk_tree_mutex);
176
177 if (!blk_tree_root) {
178 blk_tree_root = debugfs_create_dir("block", NULL);
179 if (!blk_tree_root)
180 goto err;
181 }
182
183 dir = debugfs_create_dir(blk_name, blk_tree_root);
184 if (dir)
185 root_users++;
186 else
187 blk_remove_root();
188
189err:
190 mutex_unlock(&blk_tree_mutex);
191 return dir;
192}
193
194static void blk_trace_cleanup(struct blk_trace *bt)
195{
196 relay_close(bt->rchan);
197 debugfs_remove(bt->dropped_file);
198 blk_remove_tree(bt->dir);
199 free_percpu(bt->sequence);
200 kfree(bt);
201}
202
203static int blk_trace_remove(request_queue_t *q)
204{
205 struct blk_trace *bt;
206
207 bt = xchg(&q->blk_trace, NULL);
208 if (!bt)
209 return -EINVAL;
210
211 if (bt->trace_state == Blktrace_setup ||
212 bt->trace_state == Blktrace_stopped)
213 blk_trace_cleanup(bt);
214
215 return 0;
216}
217
218static int blk_dropped_open(struct inode *inode, struct file *filp)
219{
220 filp->private_data = inode->u.generic_ip;
221
222 return 0;
223}
224
225static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
226 size_t count, loff_t *ppos)
227{
228 struct blk_trace *bt = filp->private_data;
229 char buf[16];
230
231 snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
232
233 return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
234}
235
236static struct file_operations blk_dropped_fops = {
237 .owner = THIS_MODULE,
238 .open = blk_dropped_open,
239 .read = blk_dropped_read,
240};
241
242/*
243 * Keep track of how many times we encountered a full subbuffer, to aid
244 * the user space app in telling how many lost events there were.
245 */
246static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
247 void *prev_subbuf, size_t prev_padding)
248{
249 struct blk_trace *bt;
250
251 if (!relay_buf_full(buf))
252 return 1;
253
254 bt = buf->chan->private_data;
255 atomic_inc(&bt->dropped);
256 return 0;
257}
258
259static int blk_remove_buf_file_callback(struct dentry *dentry)
260{
261 debugfs_remove(dentry);
262 return 0;
263}
264
265static struct dentry *blk_create_buf_file_callback(const char *filename,
266 struct dentry *parent,
267 int mode,
268 struct rchan_buf *buf,
269 int *is_global)
270{
271 return debugfs_create_file(filename, mode, parent, buf,
272 &relay_file_operations);
273}
274
275static struct rchan_callbacks blk_relay_callbacks = {
276 .subbuf_start = blk_subbuf_start_callback,
277 .create_buf_file = blk_create_buf_file_callback,
278 .remove_buf_file = blk_remove_buf_file_callback,
279};
280
281/*
282 * Setup everything required to start tracing
283 */
284static int blk_trace_setup(request_queue_t *q, struct block_device *bdev,
285 char __user *arg)
286{
287 struct blk_user_trace_setup buts;
288 struct blk_trace *old_bt, *bt = NULL;
289 struct dentry *dir = NULL;
290 char b[BDEVNAME_SIZE];
291 int ret, i;
292
293 if (copy_from_user(&buts, arg, sizeof(buts)))
294 return -EFAULT;
295
296 if (!buts.buf_size || !buts.buf_nr)
297 return -EINVAL;
298
299 strcpy(buts.name, bdevname(bdev, b));
300
301 /*
302 * some device names have larger paths - convert the slashes
303 * to underscores for this to work as expected
304 */
305 for (i = 0; i < strlen(buts.name); i++)
306 if (buts.name[i] == '/')
307 buts.name[i] = '_';
308
309 if (copy_to_user(arg, &buts, sizeof(buts)))
310 return -EFAULT;
311
312 ret = -ENOMEM;
313 bt = kzalloc(sizeof(*bt), GFP_KERNEL);
314 if (!bt)
315 goto err;
316
317 bt->sequence = alloc_percpu(unsigned long);
318 if (!bt->sequence)
319 goto err;
320
321 ret = -ENOENT;
322 dir = blk_create_tree(buts.name);
323 if (!dir)
324 goto err;
325
326 bt->dir = dir;
327 bt->dev = bdev->bd_dev;
328 atomic_set(&bt->dropped, 0);
329
330 ret = -EIO;
331 bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, &blk_dropped_fops);
332 if (!bt->dropped_file)
333 goto err;
334
335 bt->rchan = relay_open("trace", dir, buts.buf_size, buts.buf_nr, &blk_relay_callbacks);
336 if (!bt->rchan)
337 goto err;
338 bt->rchan->private_data = bt;
339
340 bt->act_mask = buts.act_mask;
341 if (!bt->act_mask)
342 bt->act_mask = (u16) -1;
343
344 bt->start_lba = buts.start_lba;
345 bt->end_lba = buts.end_lba;
346 if (!bt->end_lba)
347 bt->end_lba = -1ULL;
348
349 bt->pid = buts.pid;
350 bt->trace_state = Blktrace_setup;
351
352 ret = -EBUSY;
353 old_bt = xchg(&q->blk_trace, bt);
354 if (old_bt) {
355 (void) xchg(&q->blk_trace, old_bt);
356 goto err;
357 }
358
359 return 0;
360err:
361 if (dir)
362 blk_remove_tree(dir);
363 if (bt) {
364 if (bt->dropped_file)
365 debugfs_remove(bt->dropped_file);
366 if (bt->sequence)
367 free_percpu(bt->sequence);
368 if (bt->rchan)
369 relay_close(bt->rchan);
370 kfree(bt);
371 }
372 return ret;
373}
374
375static int blk_trace_startstop(request_queue_t *q, int start)
376{
377 struct blk_trace *bt;
378 int ret;
379
380 if ((bt = q->blk_trace) == NULL)
381 return -EINVAL;
382
383 /*
384 * For starting a trace, we can transition from a setup or stopped
385 * trace. For stopping a trace, the state must be running
386 */
387 ret = -EINVAL;
388 if (start) {
389 if (bt->trace_state == Blktrace_setup ||
390 bt->trace_state == Blktrace_stopped) {
391 blktrace_seq++;
392 smp_mb();
393 bt->trace_state = Blktrace_running;
394 ret = 0;
395 }
396 } else {
397 if (bt->trace_state == Blktrace_running) {
398 bt->trace_state = Blktrace_stopped;
399 relay_flush(bt->rchan);
400 ret = 0;
401 }
402 }
403
404 return ret;
405}
406
407/**
408 * blk_trace_ioctl: - handle the ioctls associated with tracing
409 * @bdev: the block device
410 * @cmd: the ioctl cmd
411 * @arg: the argument data, if any
412 *
413 **/
414int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
415{
416 request_queue_t *q;
417 int ret, start = 0;
418
419 q = bdev_get_queue(bdev);
420 if (!q)
421 return -ENXIO;
422
423 mutex_lock(&bdev->bd_mutex);
424
425 switch (cmd) {
426 case BLKTRACESETUP:
427 ret = blk_trace_setup(q, bdev, arg);
428 break;
429 case BLKTRACESTART:
430 start = 1;
431 case BLKTRACESTOP:
432 ret = blk_trace_startstop(q, start);
433 break;
434 case BLKTRACETEARDOWN:
435 ret = blk_trace_remove(q);
436 break;
437 default:
438 ret = -ENOTTY;
439 break;
440 }
441
442 mutex_unlock(&bdev->bd_mutex);
443 return ret;
444}
445
446/**
447 * blk_trace_shutdown: - stop and cleanup trace structures
448 * @q: the request queue associated with the device
449 *
450 **/
451void blk_trace_shutdown(request_queue_t *q)
452{
453 blk_trace_startstop(q, 0);
454 blk_trace_remove(q);
455}
456
457/*
458 * Average offset over two calls to sched_clock() with a gettimeofday()
459 * in the middle
460 */
461static void blk_check_time(unsigned long long *t)
462{
463 unsigned long long a, b;
464 struct timeval tv;
465
466 a = sched_clock();
467 do_gettimeofday(&tv);
468 b = sched_clock();
469
470 *t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
471 *t -= (a + b) / 2;
472}
473
474static void blk_trace_check_cpu_time(void *data)
475{
476 unsigned long long *t;
477 int cpu = get_cpu();
478
479 t = &per_cpu(blk_trace_cpu_offset, cpu);
480
481 /*
482 * Just call it twice, hopefully the second call will be cache hot
483 * and a little more precise
484 */
485 blk_check_time(t);
486 blk_check_time(t);
487
488 put_cpu();
489}
490
491/*
492 * Call blk_trace_check_cpu_time() on each CPU to calibrate our inter-CPU
493 * timings
494 */
495static void blk_trace_calibrate_offsets(void)
496{
497 unsigned long flags;
498
499 smp_call_function(blk_trace_check_cpu_time, NULL, 1, 1);
500 local_irq_save(flags);
501 blk_trace_check_cpu_time(NULL);
502 local_irq_restore(flags);
503}
504
505static void blk_trace_set_ht_offsets(void)
506{
507#if defined(CONFIG_SCHED_SMT)
508 int cpu, i;
509
510 /*
511 * now make sure HT siblings have the same time offset
512 */
513 preempt_disable();
514 for_each_online_cpu(cpu) {
515 unsigned long long *cpu_off, *sibling_off;
516
517 for_each_cpu_mask(i, cpu_sibling_map[cpu]) {
518 if (i == cpu)
519 continue;
520
521 cpu_off = &per_cpu(blk_trace_cpu_offset, cpu);
522 sibling_off = &per_cpu(blk_trace_cpu_offset, i);
523 *sibling_off = *cpu_off;
524 }
525 }
526 preempt_enable();
527#endif
528}
529
530static __init int blk_trace_init(void)
531{
532 mutex_init(&blk_tree_mutex);
533 blk_trace_calibrate_offsets();
534 blk_trace_set_ht_offsets();
535
536 return 0;
537}
538
539module_init(blk_trace_init);
540