Merge master.kernel.org:/pub/scm/linux/kernel/git/sfrench/cifs-2.6
[linux-2.6] / 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 DEFINE_PER_CPU(unsigned long long, blk_trace_cpu_offset) = { 0, };
29 static unsigned int blktrace_seq __read_mostly = 1;
30
31 /*
32  * Send out a notify message.
33  */
34 static void trace_note(struct blk_trace *bt, pid_t pid, int action,
35                        const void *data, size_t len)
36 {
37         struct blk_io_trace *t;
38
39         t = relay_reserve(bt->rchan, sizeof(*t) + len);
40         if (t) {
41                 const int cpu = smp_processor_id();
42
43                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
44                 t->time = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
45                 t->device = bt->dev;
46                 t->action = action;
47                 t->pid = pid;
48                 t->cpu = cpu;
49                 t->pdu_len = len;
50                 memcpy((void *) t + sizeof(*t), data, len);
51         }
52 }
53
54 /*
55  * Send out a notify for this process, if we haven't done so since a trace
56  * started
57  */
58 static void trace_note_tsk(struct blk_trace *bt, struct task_struct *tsk)
59 {
60         tsk->btrace_seq = blktrace_seq;
61         trace_note(bt, tsk->pid, BLK_TN_PROCESS, tsk->comm, sizeof(tsk->comm));
62 }
63
64 static void trace_note_time(struct blk_trace *bt)
65 {
66         struct timespec now;
67         unsigned long flags;
68         u32 words[2];
69
70         getnstimeofday(&now);
71         words[0] = now.tv_sec;
72         words[1] = now.tv_nsec;
73
74         local_irq_save(flags);
75         trace_note(bt, 0, BLK_TN_TIMESTAMP, words, sizeof(words));
76         local_irq_restore(flags);
77 }
78
79 static int act_log_check(struct blk_trace *bt, u32 what, sector_t sector,
80                          pid_t pid)
81 {
82         if (((bt->act_mask << BLK_TC_SHIFT) & what) == 0)
83                 return 1;
84         if (sector < bt->start_lba || sector > bt->end_lba)
85                 return 1;
86         if (bt->pid && pid != bt->pid)
87                 return 1;
88
89         return 0;
90 }
91
92 /*
93  * Data direction bit lookup
94  */
95 static u32 ddir_act[2] __read_mostly = { BLK_TC_ACT(BLK_TC_READ), BLK_TC_ACT(BLK_TC_WRITE) };
96
97 /*
98  * Bio action bits of interest
99  */
100 static u32 bio_act[9] __read_mostly = { 0, BLK_TC_ACT(BLK_TC_BARRIER), BLK_TC_ACT(BLK_TC_SYNC), 0, BLK_TC_ACT(BLK_TC_AHEAD), 0, 0, 0, BLK_TC_ACT(BLK_TC_META) };
101
102 /*
103  * More could be added as needed, taking care to increment the decrementer
104  * to get correct indexing
105  */
106 #define trace_barrier_bit(rw)   \
107         (((rw) & (1 << BIO_RW_BARRIER)) >> (BIO_RW_BARRIER - 0))
108 #define trace_sync_bit(rw)      \
109         (((rw) & (1 << BIO_RW_SYNC)) >> (BIO_RW_SYNC - 1))
110 #define trace_ahead_bit(rw)     \
111         (((rw) & (1 << BIO_RW_AHEAD)) << (2 - BIO_RW_AHEAD))
112 #define trace_meta_bit(rw)      \
113         (((rw) & (1 << BIO_RW_META)) >> (BIO_RW_META - 3))
114
115 /*
116  * The worker for the various blk_add_trace*() types. Fills out a
117  * blk_io_trace structure and places it in a per-cpu subbuffer.
118  */
119 void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
120                      int rw, u32 what, int error, int pdu_len, void *pdu_data)
121 {
122         struct task_struct *tsk = current;
123         struct blk_io_trace *t;
124         unsigned long flags;
125         unsigned long *sequence;
126         pid_t pid;
127         int cpu;
128
129         if (unlikely(bt->trace_state != Blktrace_running))
130                 return;
131
132         what |= ddir_act[rw & WRITE];
133         what |= bio_act[trace_barrier_bit(rw)];
134         what |= bio_act[trace_sync_bit(rw)];
135         what |= bio_act[trace_ahead_bit(rw)];
136         what |= bio_act[trace_meta_bit(rw)];
137
138         pid = tsk->pid;
139         if (unlikely(act_log_check(bt, what, sector, pid)))
140                 return;
141
142         /*
143          * A word about the locking here - we disable interrupts to reserve
144          * some space in the relay per-cpu buffer, to prevent an irq
145          * from coming in and stepping on our toes. Once reserved, it's
146          * enough to get preemption disabled to prevent read of this data
147          * before we are through filling it. get_cpu()/put_cpu() does this
148          * for us
149          */
150         local_irq_save(flags);
151
152         if (unlikely(tsk->btrace_seq != blktrace_seq))
153                 trace_note_tsk(bt, tsk);
154
155         t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
156         if (t) {
157                 cpu = smp_processor_id();
158                 sequence = per_cpu_ptr(bt->sequence, cpu);
159
160                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
161                 t->sequence = ++(*sequence);
162                 t->time = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
163                 t->sector = sector;
164                 t->bytes = bytes;
165                 t->action = what;
166                 t->pid = pid;
167                 t->device = bt->dev;
168                 t->cpu = cpu;
169                 t->error = error;
170                 t->pdu_len = pdu_len;
171
172                 if (pdu_len)
173                         memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
174         }
175
176         local_irq_restore(flags);
177 }
178
179 EXPORT_SYMBOL_GPL(__blk_add_trace);
180
181 static struct dentry *blk_tree_root;
182 static struct mutex blk_tree_mutex;
183 static unsigned int root_users;
184
185 static inline void blk_remove_root(void)
186 {
187         if (blk_tree_root) {
188                 debugfs_remove(blk_tree_root);
189                 blk_tree_root = NULL;
190         }
191 }
192
193 static void blk_remove_tree(struct dentry *dir)
194 {
195         mutex_lock(&blk_tree_mutex);
196         debugfs_remove(dir);
197         if (--root_users == 0)
198                 blk_remove_root();
199         mutex_unlock(&blk_tree_mutex);
200 }
201
202 static struct dentry *blk_create_tree(const char *blk_name)
203 {
204         struct dentry *dir = NULL;
205
206         mutex_lock(&blk_tree_mutex);
207
208         if (!blk_tree_root) {
209                 blk_tree_root = debugfs_create_dir("block", NULL);
210                 if (!blk_tree_root)
211                         goto err;
212         }
213
214         dir = debugfs_create_dir(blk_name, blk_tree_root);
215         if (dir)
216                 root_users++;
217         else
218                 blk_remove_root();
219
220 err:
221         mutex_unlock(&blk_tree_mutex);
222         return dir;
223 }
224
225 static void blk_trace_cleanup(struct blk_trace *bt)
226 {
227         relay_close(bt->rchan);
228         debugfs_remove(bt->dropped_file);
229         blk_remove_tree(bt->dir);
230         free_percpu(bt->sequence);
231         kfree(bt);
232 }
233
234 static int blk_trace_remove(request_queue_t *q)
235 {
236         struct blk_trace *bt;
237
238         bt = xchg(&q->blk_trace, NULL);
239         if (!bt)
240                 return -EINVAL;
241
242         if (bt->trace_state == Blktrace_setup ||
243             bt->trace_state == Blktrace_stopped)
244                 blk_trace_cleanup(bt);
245
246         return 0;
247 }
248
249 static int blk_dropped_open(struct inode *inode, struct file *filp)
250 {
251         filp->private_data = inode->i_private;
252
253         return 0;
254 }
255
256 static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
257                                 size_t count, loff_t *ppos)
258 {
259         struct blk_trace *bt = filp->private_data;
260         char buf[16];
261
262         snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
263
264         return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
265 }
266
267 static const struct file_operations blk_dropped_fops = {
268         .owner =        THIS_MODULE,
269         .open =         blk_dropped_open,
270         .read =         blk_dropped_read,
271 };
272
273 /*
274  * Keep track of how many times we encountered a full subbuffer, to aid
275  * the user space app in telling how many lost events there were.
276  */
277 static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
278                                      void *prev_subbuf, size_t prev_padding)
279 {
280         struct blk_trace *bt;
281
282         if (!relay_buf_full(buf))
283                 return 1;
284
285         bt = buf->chan->private_data;
286         atomic_inc(&bt->dropped);
287         return 0;
288 }
289
290 static int blk_remove_buf_file_callback(struct dentry *dentry)
291 {
292         debugfs_remove(dentry);
293         return 0;
294 }
295
296 static struct dentry *blk_create_buf_file_callback(const char *filename,
297                                                    struct dentry *parent,
298                                                    int mode,
299                                                    struct rchan_buf *buf,
300                                                    int *is_global)
301 {
302         return debugfs_create_file(filename, mode, parent, buf,
303                                         &relay_file_operations);
304 }
305
306 static struct rchan_callbacks blk_relay_callbacks = {
307         .subbuf_start           = blk_subbuf_start_callback,
308         .create_buf_file        = blk_create_buf_file_callback,
309         .remove_buf_file        = blk_remove_buf_file_callback,
310 };
311
312 /*
313  * Setup everything required to start tracing
314  */
315 static int blk_trace_setup(request_queue_t *q, struct block_device *bdev,
316                            char __user *arg)
317 {
318         struct blk_user_trace_setup buts;
319         struct blk_trace *old_bt, *bt = NULL;
320         struct dentry *dir = NULL;
321         char b[BDEVNAME_SIZE];
322         int ret, i;
323
324         if (copy_from_user(&buts, arg, sizeof(buts)))
325                 return -EFAULT;
326
327         if (!buts.buf_size || !buts.buf_nr)
328                 return -EINVAL;
329
330         strcpy(buts.name, bdevname(bdev, b));
331
332         /*
333          * some device names have larger paths - convert the slashes
334          * to underscores for this to work as expected
335          */
336         for (i = 0; i < strlen(buts.name); i++)
337                 if (buts.name[i] == '/')
338                         buts.name[i] = '_';
339
340         if (copy_to_user(arg, &buts, sizeof(buts)))
341                 return -EFAULT;
342
343         ret = -ENOMEM;
344         bt = kzalloc(sizeof(*bt), GFP_KERNEL);
345         if (!bt)
346                 goto err;
347
348         bt->sequence = alloc_percpu(unsigned long);
349         if (!bt->sequence)
350                 goto err;
351
352         ret = -ENOENT;
353         dir = blk_create_tree(buts.name);
354         if (!dir)
355                 goto err;
356
357         bt->dir = dir;
358         bt->dev = bdev->bd_dev;
359         atomic_set(&bt->dropped, 0);
360
361         ret = -EIO;
362         bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, &blk_dropped_fops);
363         if (!bt->dropped_file)
364                 goto err;
365
366         bt->rchan = relay_open("trace", dir, buts.buf_size, buts.buf_nr, &blk_relay_callbacks, bt);
367         if (!bt->rchan)
368                 goto err;
369
370         bt->act_mask = buts.act_mask;
371         if (!bt->act_mask)
372                 bt->act_mask = (u16) -1;
373
374         bt->start_lba = buts.start_lba;
375         bt->end_lba = buts.end_lba;
376         if (!bt->end_lba)
377                 bt->end_lba = -1ULL;
378
379         bt->pid = buts.pid;
380         bt->trace_state = Blktrace_setup;
381
382         ret = -EBUSY;
383         old_bt = xchg(&q->blk_trace, bt);
384         if (old_bt) {
385                 (void) xchg(&q->blk_trace, old_bt);
386                 goto err;
387         }
388
389         return 0;
390 err:
391         if (dir)
392                 blk_remove_tree(dir);
393         if (bt) {
394                 if (bt->dropped_file)
395                         debugfs_remove(bt->dropped_file);
396                 free_percpu(bt->sequence);
397                 if (bt->rchan)
398                         relay_close(bt->rchan);
399                 kfree(bt);
400         }
401         return ret;
402 }
403
404 static int blk_trace_startstop(request_queue_t *q, int start)
405 {
406         struct blk_trace *bt;
407         int ret;
408
409         if ((bt = q->blk_trace) == NULL)
410                 return -EINVAL;
411
412         /*
413          * For starting a trace, we can transition from a setup or stopped
414          * trace. For stopping a trace, the state must be running
415          */
416         ret = -EINVAL;
417         if (start) {
418                 if (bt->trace_state == Blktrace_setup ||
419                     bt->trace_state == Blktrace_stopped) {
420                         blktrace_seq++;
421                         smp_mb();
422                         bt->trace_state = Blktrace_running;
423
424                         trace_note_time(bt);
425                         ret = 0;
426                 }
427         } else {
428                 if (bt->trace_state == Blktrace_running) {
429                         bt->trace_state = Blktrace_stopped;
430                         relay_flush(bt->rchan);
431                         ret = 0;
432                 }
433         }
434
435         return ret;
436 }
437
438 /**
439  * blk_trace_ioctl: - handle the ioctls associated with tracing
440  * @bdev:       the block device
441  * @cmd:        the ioctl cmd
442  * @arg:        the argument data, if any
443  *
444  **/
445 int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
446 {
447         request_queue_t *q;
448         int ret, start = 0;
449
450         q = bdev_get_queue(bdev);
451         if (!q)
452                 return -ENXIO;
453
454         mutex_lock(&bdev->bd_mutex);
455
456         switch (cmd) {
457         case BLKTRACESETUP:
458                 ret = blk_trace_setup(q, bdev, arg);
459                 break;
460         case BLKTRACESTART:
461                 start = 1;
462         case BLKTRACESTOP:
463                 ret = blk_trace_startstop(q, start);
464                 break;
465         case BLKTRACETEARDOWN:
466                 ret = blk_trace_remove(q);
467                 break;
468         default:
469                 ret = -ENOTTY;
470                 break;
471         }
472
473         mutex_unlock(&bdev->bd_mutex);
474         return ret;
475 }
476
477 /**
478  * blk_trace_shutdown: - stop and cleanup trace structures
479  * @q:    the request queue associated with the device
480  *
481  **/
482 void blk_trace_shutdown(request_queue_t *q)
483 {
484         if (q->blk_trace) {
485                 blk_trace_startstop(q, 0);
486                 blk_trace_remove(q);
487         }
488 }
489
490 /*
491  * Average offset over two calls to sched_clock() with a gettimeofday()
492  * in the middle
493  */
494 static void blk_check_time(unsigned long long *t)
495 {
496         unsigned long long a, b;
497         struct timeval tv;
498
499         a = sched_clock();
500         do_gettimeofday(&tv);
501         b = sched_clock();
502
503         *t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
504         *t -= (a + b) / 2;
505 }
506
507 /*
508  * calibrate our inter-CPU timings
509  */
510 static void blk_trace_check_cpu_time(void *data)
511 {
512         unsigned long long *t;
513         int cpu = get_cpu();
514
515         t = &per_cpu(blk_trace_cpu_offset, cpu);
516
517         /*
518          * Just call it twice, hopefully the second call will be cache hot
519          * and a little more precise
520          */
521         blk_check_time(t);
522         blk_check_time(t);
523
524         put_cpu();
525 }
526
527 static void blk_trace_set_ht_offsets(void)
528 {
529 #if defined(CONFIG_SCHED_SMT)
530         int cpu, i;
531
532         /*
533          * now make sure HT siblings have the same time offset
534          */
535         preempt_disable();
536         for_each_online_cpu(cpu) {
537                 unsigned long long *cpu_off, *sibling_off;
538
539                 for_each_cpu_mask(i, cpu_sibling_map[cpu]) {
540                         if (i == cpu)
541                                 continue;
542
543                         cpu_off = &per_cpu(blk_trace_cpu_offset, cpu);
544                         sibling_off = &per_cpu(blk_trace_cpu_offset, i);
545                         *sibling_off = *cpu_off;
546                 }
547         }
548         preempt_enable();
549 #endif
550 }
551
552 static __init int blk_trace_init(void)
553 {
554         mutex_init(&blk_tree_mutex);
555         on_each_cpu(blk_trace_check_cpu_time, NULL, 1, 1);
556         blk_trace_set_ht_offsets();
557
558         return 0;
559 }
560
561 module_init(blk_trace_init);
562