Merge git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi-misc-2.6
[linux-2.6.git] / block / blktrace.c
1 /*
2  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
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  */
18 #include <linux/kernel.h>
19 #include <linux/blkdev.h>
20 #include <linux/blktrace_api.h>
21 #include <linux/percpu.h>
22 #include <linux/init.h>
23 #include <linux/mutex.h>
24 #include <linux/debugfs.h>
25 #include <linux/time.h>
26 #include <asm/uaccess.h>
27
28 static unsigned int blktrace_seq __read_mostly = 1;
29
30 /*
31  * Send out a notify message.
32  */
33 static void trace_note(struct blk_trace *bt, pid_t pid, int action,
34                        const void *data, size_t len)
35 {
36         struct blk_io_trace *t;
37
38         t = relay_reserve(bt->rchan, sizeof(*t) + len);
39         if (t) {
40                 const int cpu = smp_processor_id();
41
42                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
43                 t->time = ktime_to_ns(ktime_get());
44                 t->device = bt->dev;
45                 t->action = action;
46                 t->pid = pid;
47                 t->cpu = cpu;
48                 t->pdu_len = len;
49                 memcpy((void *) t + sizeof(*t), data, len);
50         }
51 }
52
53 /*
54  * Send out a notify for this process, if we haven't done so since a trace
55  * started
56  */
57 static void trace_note_tsk(struct blk_trace *bt, struct task_struct *tsk)
58 {
59         tsk->btrace_seq = blktrace_seq;
60         trace_note(bt, tsk->pid, BLK_TN_PROCESS, tsk->comm, sizeof(tsk->comm));
61 }
62
63 static void trace_note_time(struct blk_trace *bt)
64 {
65         struct timespec now;
66         unsigned long flags;
67         u32 words[2];
68
69         getnstimeofday(&now);
70         words[0] = now.tv_sec;
71         words[1] = now.tv_nsec;
72
73         local_irq_save(flags);
74         trace_note(bt, 0, BLK_TN_TIMESTAMP, words, sizeof(words));
75         local_irq_restore(flags);
76 }
77
78 void __trace_note_message(struct blk_trace *bt, const char *fmt, ...)
79 {
80         int n;
81         va_list args;
82         unsigned long flags;
83         char *buf;
84
85         local_irq_save(flags);
86         buf = per_cpu_ptr(bt->msg_data, smp_processor_id());
87         va_start(args, fmt);
88         n = vscnprintf(buf, BLK_TN_MAX_MSG, fmt, args);
89         va_end(args);
90
91         trace_note(bt, 0, BLK_TN_MESSAGE, buf, n);
92         local_irq_restore(flags);
93 }
94 EXPORT_SYMBOL_GPL(__trace_note_message);
95
96 static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
97                          pid_t pid)
98 {
99         if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
100                 return 1;
101         if (sector < bt->start_lba || sector > bt->end_lba)
102                 return 1;
103         if (bt->pid && pid != bt->pid)
104                 return 1;
105
106         return 0;
107 }
108
109 /*
110  * Data direction bit lookup
111  */
112 static u32 ddir_act[2] __read_mostly = { BLK_TC_ACT(BLK_TC_READ), BLK_TC_ACT(BLK_TC_WRITE) };
113
114 /* The ilog2() calls fall out because they're constant */
115 #define MASK_TC_BIT(rw, __name) ( (rw & (1 << BIO_RW_ ## __name)) << \
116           (ilog2(BLK_TC_ ## __name) + BLK_TC_SHIFT - BIO_RW_ ## __name) )
117
118 /*
119  * The worker for the various blk_add_trace*() types. Fills out a
120  * blk_io_trace structure and places it in a per-cpu subbuffer.
121  */
122 void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
123                      int rw, u32 what, int error, int pdu_len, void *pdu_data)
124 {
125         struct task_struct *tsk = current;
126         struct blk_io_trace *t;
127         unsigned long flags;
128         unsigned long *sequence;
129         pid_t pid;
130         int cpu;
131
132         if (unlikely(bt->trace_state != Blktrace_running))
133                 return;
134
135         what |= ddir_act[rw & WRITE];
136         what |= MASK_TC_BIT(rw, BARRIER);
137         what |= MASK_TC_BIT(rw, SYNC);
138         what |= MASK_TC_BIT(rw, AHEAD);
139         what |= MASK_TC_BIT(rw, META);
140         what |= MASK_TC_BIT(rw, DISCARD);
141
142         pid = tsk->pid;
143         if (unlikely(act_log_check(bt, what, sector, pid)))
144                 return;
145
146         /*
147          * A word about the locking here - we disable interrupts to reserve
148          * some space in the relay per-cpu buffer, to prevent an irq
149          * from coming in and stepping on our toes.
150          */
151         local_irq_save(flags);
152
153         if (unlikely(tsk->btrace_seq != blktrace_seq))
154                 trace_note_tsk(bt, tsk);
155
156         t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
157         if (t) {
158                 cpu = smp_processor_id();
159                 sequence = per_cpu_ptr(bt->sequence, cpu);
160
161                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
162                 t->sequence = ++(*sequence);
163                 t->time = ktime_to_ns(ktime_get());
164                 t->sector = sector;
165                 t->bytes = bytes;
166                 t->action = what;
167                 t->pid = pid;
168                 t->device = bt->dev;
169                 t->cpu = cpu;
170                 t->error = error;
171                 t->pdu_len = pdu_len;
172
173                 if (pdu_len)
174                         memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
175         }
176
177         local_irq_restore(flags);
178 }
179
180 EXPORT_SYMBOL_GPL(__blk_add_trace);
181
182 static struct dentry *blk_tree_root;
183 static DEFINE_MUTEX(blk_tree_mutex);
184 static unsigned int root_users;
185
186 static inline void blk_remove_root(void)
187 {
188         if (blk_tree_root) {
189                 debugfs_remove(blk_tree_root);
190                 blk_tree_root = NULL;
191         }
192 }
193
194 static void blk_remove_tree(struct dentry *dir)
195 {
196         mutex_lock(&blk_tree_mutex);
197         debugfs_remove(dir);
198         if (--root_users == 0)
199                 blk_remove_root();
200         mutex_unlock(&blk_tree_mutex);
201 }
202
203 static struct dentry *blk_create_tree(const char *blk_name)
204 {
205         struct dentry *dir = NULL;
206         int created = 0;
207
208         mutex_lock(&blk_tree_mutex);
209
210         if (!blk_tree_root) {
211                 blk_tree_root = debugfs_create_dir("block", NULL);
212                 if (!blk_tree_root)
213                         goto err;
214                 created = 1;
215         }
216
217         dir = debugfs_create_dir(blk_name, blk_tree_root);
218         if (dir)
219                 root_users++;
220         else {
221                 /* Delete root only if we created it */
222                 if (created)
223                         blk_remove_root();
224         }
225
226 err:
227         mutex_unlock(&blk_tree_mutex);
228         return dir;
229 }
230
231 static void blk_trace_cleanup(struct blk_trace *bt)
232 {
233         relay_close(bt->rchan);
234         debugfs_remove(bt->msg_file);
235         debugfs_remove(bt->dropped_file);
236         blk_remove_tree(bt->dir);
237         free_percpu(bt->sequence);
238         free_percpu(bt->msg_data);
239         kfree(bt);
240 }
241
242 int blk_trace_remove(struct request_queue *q)
243 {
244         struct blk_trace *bt;
245
246         bt = xchg(&q->blk_trace, NULL);
247         if (!bt)
248                 return -EINVAL;
249
250         if (bt->trace_state == Blktrace_setup ||
251             bt->trace_state == Blktrace_stopped)
252                 blk_trace_cleanup(bt);
253
254         return 0;
255 }
256 EXPORT_SYMBOL_GPL(blk_trace_remove);
257
258 static int blk_dropped_open(struct inode *inode, struct file *filp)
259 {
260         filp->private_data = inode->i_private;
261
262         return 0;
263 }
264
265 static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
266                                 size_t count, loff_t *ppos)
267 {
268         struct blk_trace *bt = filp->private_data;
269         char buf[16];
270
271         snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
272
273         return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
274 }
275
276 static const struct file_operations blk_dropped_fops = {
277         .owner =        THIS_MODULE,
278         .open =         blk_dropped_open,
279         .read =         blk_dropped_read,
280 };
281
282 static int blk_msg_open(struct inode *inode, struct file *filp)
283 {
284         filp->private_data = inode->i_private;
285
286         return 0;
287 }
288
289 static ssize_t blk_msg_write(struct file *filp, const char __user *buffer,
290                                 size_t count, loff_t *ppos)
291 {
292         char *msg;
293         struct blk_trace *bt;
294
295         if (count > BLK_TN_MAX_MSG)
296                 return -EINVAL;
297
298         msg = kmalloc(count, GFP_KERNEL);
299         if (msg == NULL)
300                 return -ENOMEM;
301
302         if (copy_from_user(msg, buffer, count)) {
303                 kfree(msg);
304                 return -EFAULT;
305         }
306
307         bt = filp->private_data;
308         __trace_note_message(bt, "%s", msg);
309         kfree(msg);
310
311         return count;
312 }
313
314 static const struct file_operations blk_msg_fops = {
315         .owner =        THIS_MODULE,
316         .open =         blk_msg_open,
317         .write =        blk_msg_write,
318 };
319
320 /*
321  * Keep track of how many times we encountered a full subbuffer, to aid
322  * the user space app in telling how many lost events there were.
323  */
324 static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
325                                      void *prev_subbuf, size_t prev_padding)
326 {
327         struct blk_trace *bt;
328
329         if (!relay_buf_full(buf))
330                 return 1;
331
332         bt = buf->chan->private_data;
333         atomic_inc(&bt->dropped);
334         return 0;
335 }
336
337 static int blk_remove_buf_file_callback(struct dentry *dentry)
338 {
339         debugfs_remove(dentry);
340         return 0;
341 }
342
343 static struct dentry *blk_create_buf_file_callback(const char *filename,
344                                                    struct dentry *parent,
345                                                    int mode,
346                                                    struct rchan_buf *buf,
347                                                    int *is_global)
348 {
349         return debugfs_create_file(filename, mode, parent, buf,
350                                         &relay_file_operations);
351 }
352
353 static struct rchan_callbacks blk_relay_callbacks = {
354         .subbuf_start           = blk_subbuf_start_callback,
355         .create_buf_file        = blk_create_buf_file_callback,
356         .remove_buf_file        = blk_remove_buf_file_callback,
357 };
358
359 /*
360  * Setup everything required to start tracing
361  */
362 int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
363                         struct blk_user_trace_setup *buts)
364 {
365         struct blk_trace *old_bt, *bt = NULL;
366         struct dentry *dir = NULL;
367         int ret, i;
368
369         if (!buts->buf_size || !buts->buf_nr)
370                 return -EINVAL;
371
372         strncpy(buts->name, name, BLKTRACE_BDEV_SIZE);
373         buts->name[BLKTRACE_BDEV_SIZE - 1] = '\0';
374
375         /*
376          * some device names have larger paths - convert the slashes
377          * to underscores for this to work as expected
378          */
379         for (i = 0; i < strlen(buts->name); i++)
380                 if (buts->name[i] == '/')
381                         buts->name[i] = '_';
382
383         ret = -ENOMEM;
384         bt = kzalloc(sizeof(*bt), GFP_KERNEL);
385         if (!bt)
386                 goto err;
387
388         bt->sequence = alloc_percpu(unsigned long);
389         if (!bt->sequence)
390                 goto err;
391
392         bt->msg_data = __alloc_percpu(BLK_TN_MAX_MSG);
393         if (!bt->msg_data)
394                 goto err;
395
396         ret = -ENOENT;
397         dir = blk_create_tree(buts->name);
398         if (!dir)
399                 goto err;
400
401         bt->dir = dir;
402         bt->dev = dev;
403         atomic_set(&bt->dropped, 0);
404
405         ret = -EIO;
406         bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, &blk_dropped_fops);
407         if (!bt->dropped_file)
408                 goto err;
409
410         bt->msg_file = debugfs_create_file("msg", 0222, dir, bt, &blk_msg_fops);
411         if (!bt->msg_file)
412                 goto err;
413
414         bt->rchan = relay_open("trace", dir, buts->buf_size,
415                                 buts->buf_nr, &blk_relay_callbacks, bt);
416         if (!bt->rchan)
417                 goto err;
418
419         bt->act_mask = buts->act_mask;
420         if (!bt->act_mask)
421                 bt->act_mask = (u16) -1;
422
423         bt->start_lba = buts->start_lba;
424         bt->end_lba = buts->end_lba;
425         if (!bt->end_lba)
426                 bt->end_lba = -1ULL;
427
428         bt->pid = buts->pid;
429         bt->trace_state = Blktrace_setup;
430
431         ret = -EBUSY;
432         old_bt = xchg(&q->blk_trace, bt);
433         if (old_bt) {
434                 (void) xchg(&q->blk_trace, old_bt);
435                 goto err;
436         }
437
438         return 0;
439 err:
440         if (dir)
441                 blk_remove_tree(dir);
442         if (bt) {
443                 if (bt->msg_file)
444                         debugfs_remove(bt->msg_file);
445                 if (bt->dropped_file)
446                         debugfs_remove(bt->dropped_file);
447                 free_percpu(bt->sequence);
448                 free_percpu(bt->msg_data);
449                 if (bt->rchan)
450                         relay_close(bt->rchan);
451                 kfree(bt);
452         }
453         return ret;
454 }
455
456 int blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
457                     char __user *arg)
458 {
459         struct blk_user_trace_setup buts;
460         int ret;
461
462         ret = copy_from_user(&buts, arg, sizeof(buts));
463         if (ret)
464                 return -EFAULT;
465
466         ret = do_blk_trace_setup(q, name, dev, &buts);
467         if (ret)
468                 return ret;
469
470         if (copy_to_user(arg, &buts, sizeof(buts)))
471                 return -EFAULT;
472
473         return 0;
474 }
475 EXPORT_SYMBOL_GPL(blk_trace_setup);
476
477 int blk_trace_startstop(struct request_queue *q, int start)
478 {
479         struct blk_trace *bt;
480         int ret;
481
482         if ((bt = q->blk_trace) == NULL)
483                 return -EINVAL;
484
485         /*
486          * For starting a trace, we can transition from a setup or stopped
487          * trace. For stopping a trace, the state must be running
488          */
489         ret = -EINVAL;
490         if (start) {
491                 if (bt->trace_state == Blktrace_setup ||
492                     bt->trace_state == Blktrace_stopped) {
493                         blktrace_seq++;
494                         smp_mb();
495                         bt->trace_state = Blktrace_running;
496
497                         trace_note_time(bt);
498                         ret = 0;
499                 }
500         } else {
501                 if (bt->trace_state == Blktrace_running) {
502                         bt->trace_state = Blktrace_stopped;
503                         relay_flush(bt->rchan);
504                         ret = 0;
505                 }
506         }
507
508         return ret;
509 }
510 EXPORT_SYMBOL_GPL(blk_trace_startstop);
511
512 /**
513  * blk_trace_ioctl: - handle the ioctls associated with tracing
514  * @bdev:       the block device
515  * @cmd:        the ioctl cmd
516  * @arg:        the argument data, if any
517  *
518  **/
519 int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
520 {
521         struct request_queue *q;
522         int ret, start = 0;
523         char b[BDEVNAME_SIZE];
524
525         q = bdev_get_queue(bdev);
526         if (!q)
527                 return -ENXIO;
528
529         mutex_lock(&bdev->bd_mutex);
530
531         switch (cmd) {
532         case BLKTRACESETUP:
533                 bdevname(bdev, b);
534                 ret = blk_trace_setup(q, b, bdev->bd_dev, arg);
535                 break;
536         case BLKTRACESTART:
537                 start = 1;
538         case BLKTRACESTOP:
539                 ret = blk_trace_startstop(q, start);
540                 break;
541         case BLKTRACETEARDOWN:
542                 ret = blk_trace_remove(q);
543                 break;
544         default:
545                 ret = -ENOTTY;
546                 break;
547         }
548
549         mutex_unlock(&bdev->bd_mutex);
550         return ret;
551 }
552
553 /**
554  * blk_trace_shutdown: - stop and cleanup trace structures
555  * @q:    the request queue associated with the device
556  *
557  **/
558 void blk_trace_shutdown(struct request_queue *q)
559 {
560         if (q->blk_trace) {
561                 blk_trace_startstop(q, 0);
562                 blk_trace_remove(q);
563         }
564 }