[PATCH] r/o bind mount prepwork: move open_namei()'s vfs_create()
[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 <asm/uaccess.h>
26
27 static DEFINE_PER_CPU(unsigned long long, blk_trace_cpu_offset) = { 0, };
28 static 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  */
34 static 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
51 static 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  */
67 static 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  */
72 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) };
73
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))
82 #define trace_ahead_bit(rw)     \
83         (((rw) & (1 << BIO_RW_AHEAD)) << (2 - BIO_RW_AHEAD))
84 #define trace_meta_bit(rw)      \
85         (((rw) & (1 << BIO_RW_META)) >> (BIO_RW_META - 3))
86
87 /*
88  * The worker for the various blk_add_trace*() types. Fills out a
89  * blk_io_trace structure and places it in a per-cpu subbuffer.
90  */
91 void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
92                      int rw, u32 what, int error, int pdu_len, void *pdu_data)
93 {
94         struct task_struct *tsk = current;
95         struct blk_io_trace *t;
96         unsigned long flags;
97         unsigned long *sequence;
98         pid_t pid;
99         int cpu;
100
101         if (unlikely(bt->trace_state != Blktrace_running))
102                 return;
103
104         what |= ddir_act[rw & WRITE];
105         what |= bio_act[trace_barrier_bit(rw)];
106         what |= bio_act[trace_sync_bit(rw)];
107         what |= bio_act[trace_ahead_bit(rw)];
108         what |= bio_act[trace_meta_bit(rw)];
109
110         pid = tsk->pid;
111         if (unlikely(act_log_check(bt, what, sector, pid)))
112                 return;
113
114         /*
115          * A word about the locking here - we disable interrupts to reserve
116          * some space in the relay per-cpu buffer, to prevent an irq
117          * from coming in and stepping on our toes. Once reserved, it's
118          * enough to get preemption disabled to prevent read of this data
119          * before we are through filling it. get_cpu()/put_cpu() does this
120          * for us
121          */
122         local_irq_save(flags);
123
124         if (unlikely(tsk->btrace_seq != blktrace_seq))
125                 trace_note_tsk(bt, tsk);
126
127         t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
128         if (t) {
129                 cpu = smp_processor_id();
130                 sequence = per_cpu_ptr(bt->sequence, cpu);
131
132                 t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
133                 t->sequence = ++(*sequence);
134                 t->time = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
135                 t->sector = sector;
136                 t->bytes = bytes;
137                 t->action = what;
138                 t->pid = pid;
139                 t->device = bt->dev;
140                 t->cpu = cpu;
141                 t->error = error;
142                 t->pdu_len = pdu_len;
143
144                 if (pdu_len)
145                         memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
146         }
147
148         local_irq_restore(flags);
149 }
150
151 EXPORT_SYMBOL_GPL(__blk_add_trace);
152
153 static struct dentry *blk_tree_root;
154 static struct mutex blk_tree_mutex;
155 static unsigned int root_users;
156
157 static inline void blk_remove_root(void)
158 {
159         if (blk_tree_root) {
160                 debugfs_remove(blk_tree_root);
161                 blk_tree_root = NULL;
162         }
163 }
164
165 static void blk_remove_tree(struct dentry *dir)
166 {
167         mutex_lock(&blk_tree_mutex);
168         debugfs_remove(dir);
169         if (--root_users == 0)
170                 blk_remove_root();
171         mutex_unlock(&blk_tree_mutex);
172 }
173
174 static struct dentry *blk_create_tree(const char *blk_name)
175 {
176         struct dentry *dir = NULL;
177
178         mutex_lock(&blk_tree_mutex);
179
180         if (!blk_tree_root) {
181                 blk_tree_root = debugfs_create_dir("block", NULL);
182                 if (!blk_tree_root)
183                         goto err;
184         }
185
186         dir = debugfs_create_dir(blk_name, blk_tree_root);
187         if (dir)
188                 root_users++;
189         else
190                 blk_remove_root();
191
192 err:
193         mutex_unlock(&blk_tree_mutex);
194         return dir;
195 }
196
197 static void blk_trace_cleanup(struct blk_trace *bt)
198 {
199         relay_close(bt->rchan);
200         debugfs_remove(bt->dropped_file);
201         blk_remove_tree(bt->dir);
202         free_percpu(bt->sequence);
203         kfree(bt);
204 }
205
206 static int blk_trace_remove(request_queue_t *q)
207 {
208         struct blk_trace *bt;
209
210         bt = xchg(&q->blk_trace, NULL);
211         if (!bt)
212                 return -EINVAL;
213
214         if (bt->trace_state == Blktrace_setup ||
215             bt->trace_state == Blktrace_stopped)
216                 blk_trace_cleanup(bt);
217
218         return 0;
219 }
220
221 static int blk_dropped_open(struct inode *inode, struct file *filp)
222 {
223         filp->private_data = inode->i_private;
224
225         return 0;
226 }
227
228 static ssize_t blk_dropped_read(struct file *filp, char __user *buffer,
229                                 size_t count, loff_t *ppos)
230 {
231         struct blk_trace *bt = filp->private_data;
232         char buf[16];
233
234         snprintf(buf, sizeof(buf), "%u\n", atomic_read(&bt->dropped));
235
236         return simple_read_from_buffer(buffer, count, ppos, buf, strlen(buf));
237 }
238
239 static struct file_operations blk_dropped_fops = {
240         .owner =        THIS_MODULE,
241         .open =         blk_dropped_open,
242         .read =         blk_dropped_read,
243 };
244
245 /*
246  * Keep track of how many times we encountered a full subbuffer, to aid
247  * the user space app in telling how many lost events there were.
248  */
249 static int blk_subbuf_start_callback(struct rchan_buf *buf, void *subbuf,
250                                      void *prev_subbuf, size_t prev_padding)
251 {
252         struct blk_trace *bt;
253
254         if (!relay_buf_full(buf))
255                 return 1;
256
257         bt = buf->chan->private_data;
258         atomic_inc(&bt->dropped);
259         return 0;
260 }
261
262 static int blk_remove_buf_file_callback(struct dentry *dentry)
263 {
264         debugfs_remove(dentry);
265         return 0;
266 }
267
268 static struct dentry *blk_create_buf_file_callback(const char *filename,
269                                                    struct dentry *parent,
270                                                    int mode,
271                                                    struct rchan_buf *buf,
272                                                    int *is_global)
273 {
274         return debugfs_create_file(filename, mode, parent, buf,
275                                         &relay_file_operations);
276 }
277
278 static struct rchan_callbacks blk_relay_callbacks = {
279         .subbuf_start           = blk_subbuf_start_callback,
280         .create_buf_file        = blk_create_buf_file_callback,
281         .remove_buf_file        = blk_remove_buf_file_callback,
282 };
283
284 /*
285  * Setup everything required to start tracing
286  */
287 static int blk_trace_setup(request_queue_t *q, struct block_device *bdev,
288                            char __user *arg)
289 {
290         struct blk_user_trace_setup buts;
291         struct blk_trace *old_bt, *bt = NULL;
292         struct dentry *dir = NULL;
293         char b[BDEVNAME_SIZE];
294         int ret, i;
295
296         if (copy_from_user(&buts, arg, sizeof(buts)))
297                 return -EFAULT;
298
299         if (!buts.buf_size || !buts.buf_nr)
300                 return -EINVAL;
301
302         strcpy(buts.name, bdevname(bdev, b));
303
304         /*
305          * some device names have larger paths - convert the slashes
306          * to underscores for this to work as expected
307          */
308         for (i = 0; i < strlen(buts.name); i++)
309                 if (buts.name[i] == '/')
310                         buts.name[i] = '_';
311
312         if (copy_to_user(arg, &buts, sizeof(buts)))
313                 return -EFAULT;
314
315         ret = -ENOMEM;
316         bt = kzalloc(sizeof(*bt), GFP_KERNEL);
317         if (!bt)
318                 goto err;
319
320         bt->sequence = alloc_percpu(unsigned long);
321         if (!bt->sequence)
322                 goto err;
323
324         ret = -ENOENT;
325         dir = blk_create_tree(buts.name);
326         if (!dir)
327                 goto err;
328
329         bt->dir = dir;
330         bt->dev = bdev->bd_dev;
331         atomic_set(&bt->dropped, 0);
332
333         ret = -EIO;
334         bt->dropped_file = debugfs_create_file("dropped", 0444, dir, bt, &blk_dropped_fops);
335         if (!bt->dropped_file)
336                 goto err;
337
338         bt->rchan = relay_open("trace", dir, buts.buf_size, buts.buf_nr, &blk_relay_callbacks);
339         if (!bt->rchan)
340                 goto err;
341         bt->rchan->private_data = bt;
342
343         bt->act_mask = buts.act_mask;
344         if (!bt->act_mask)
345                 bt->act_mask = (u16) -1;
346
347         bt->start_lba = buts.start_lba;
348         bt->end_lba = buts.end_lba;
349         if (!bt->end_lba)
350                 bt->end_lba = -1ULL;
351
352         bt->pid = buts.pid;
353         bt->trace_state = Blktrace_setup;
354
355         ret = -EBUSY;
356         old_bt = xchg(&q->blk_trace, bt);
357         if (old_bt) {
358                 (void) xchg(&q->blk_trace, old_bt);
359                 goto err;
360         }
361
362         return 0;
363 err:
364         if (dir)
365                 blk_remove_tree(dir);
366         if (bt) {
367                 if (bt->dropped_file)
368                         debugfs_remove(bt->dropped_file);
369                 if (bt->sequence)
370                         free_percpu(bt->sequence);
371                 if (bt->rchan)
372                         relay_close(bt->rchan);
373                 kfree(bt);
374         }
375         return ret;
376 }
377
378 static int blk_trace_startstop(request_queue_t *q, int start)
379 {
380         struct blk_trace *bt;
381         int ret;
382
383         if ((bt = q->blk_trace) == NULL)
384                 return -EINVAL;
385
386         /*
387          * For starting a trace, we can transition from a setup or stopped
388          * trace. For stopping a trace, the state must be running
389          */
390         ret = -EINVAL;
391         if (start) {
392                 if (bt->trace_state == Blktrace_setup ||
393                     bt->trace_state == Blktrace_stopped) {
394                         blktrace_seq++;
395                         smp_mb();
396                         bt->trace_state = Blktrace_running;
397                         ret = 0;
398                 }
399         } else {
400                 if (bt->trace_state == Blktrace_running) {
401                         bt->trace_state = Blktrace_stopped;
402                         relay_flush(bt->rchan);
403                         ret = 0;
404                 }
405         }
406
407         return ret;
408 }
409
410 /**
411  * blk_trace_ioctl: - handle the ioctls associated with tracing
412  * @bdev:       the block device
413  * @cmd:        the ioctl cmd
414  * @arg:        the argument data, if any
415  *
416  **/
417 int blk_trace_ioctl(struct block_device *bdev, unsigned cmd, char __user *arg)
418 {
419         request_queue_t *q;
420         int ret, start = 0;
421
422         q = bdev_get_queue(bdev);
423         if (!q)
424                 return -ENXIO;
425
426         mutex_lock(&bdev->bd_mutex);
427
428         switch (cmd) {
429         case BLKTRACESETUP:
430                 ret = blk_trace_setup(q, bdev, arg);
431                 break;
432         case BLKTRACESTART:
433                 start = 1;
434         case BLKTRACESTOP:
435                 ret = blk_trace_startstop(q, start);
436                 break;
437         case BLKTRACETEARDOWN:
438                 ret = blk_trace_remove(q);
439                 break;
440         default:
441                 ret = -ENOTTY;
442                 break;
443         }
444
445         mutex_unlock(&bdev->bd_mutex);
446         return ret;
447 }
448
449 /**
450  * blk_trace_shutdown: - stop and cleanup trace structures
451  * @q:    the request queue associated with the device
452  *
453  **/
454 void blk_trace_shutdown(request_queue_t *q)
455 {
456         if (q->blk_trace) {
457                 blk_trace_startstop(q, 0);
458                 blk_trace_remove(q);
459         }
460 }
461
462 /*
463  * Average offset over two calls to sched_clock() with a gettimeofday()
464  * in the middle
465  */
466 static void blk_check_time(unsigned long long *t)
467 {
468         unsigned long long a, b;
469         struct timeval tv;
470
471         a = sched_clock();
472         do_gettimeofday(&tv);
473         b = sched_clock();
474
475         *t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
476         *t -= (a + b) / 2;
477 }
478
479 /*
480  * calibrate our inter-CPU timings
481  */
482 static void blk_trace_check_cpu_time(void *data)
483 {
484         unsigned long long *t;
485         int cpu = get_cpu();
486
487         t = &per_cpu(blk_trace_cpu_offset, cpu);
488
489         /*
490          * Just call it twice, hopefully the second call will be cache hot
491          * and a little more precise
492          */
493         blk_check_time(t);
494         blk_check_time(t);
495
496         put_cpu();
497 }
498
499 static void blk_trace_set_ht_offsets(void)
500 {
501 #if defined(CONFIG_SCHED_SMT)
502         int cpu, i;
503
504         /*
505          * now make sure HT siblings have the same time offset
506          */
507         preempt_disable();
508         for_each_online_cpu(cpu) {
509                 unsigned long long *cpu_off, *sibling_off;
510
511                 for_each_cpu_mask(i, cpu_sibling_map[cpu]) {
512                         if (i == cpu)
513                                 continue;
514
515                         cpu_off = &per_cpu(blk_trace_cpu_offset, cpu);
516                         sibling_off = &per_cpu(blk_trace_cpu_offset, i);
517                         *sibling_off = *cpu_off;
518                 }
519         }
520         preempt_enable();
521 #endif
522 }
523
524 static __init int blk_trace_init(void)
525 {
526         mutex_init(&blk_tree_mutex);
527         on_each_cpu(blk_trace_check_cpu_time, NULL, 1, 1);
528         blk_trace_set_ht_offsets();
529
530         return 0;
531 }
532
533 module_init(blk_trace_init);
534