Whamcloud - gitweb
8170f3479c1cb899a936d8ce7c3a5be9c7ef4a54
[fs/lustre-release.git] / libcfs / libcfs / tracefile.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2012, 2014, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * libcfs/libcfs/tracefile.c
37  *
38  * Author: Zach Brown <zab@clusterfs.com>
39  * Author: Phil Schwan <phil@clusterfs.com>
40  */
41
42
43 #define DEBUG_SUBSYSTEM S_LNET
44 #define LUSTRE_TRACEFILE_PRIVATE
45 #include "tracefile.h"
46
47 #include <libcfs/libcfs.h>
48
49 /* XXX move things up to the top, comment */
50 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
51
52 char cfs_tracefile[TRACEFILE_NAME_SIZE];
53 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
54 static struct tracefiled_ctl trace_tctl;
55 static DEFINE_MUTEX(cfs_trace_thread_mutex);
56 static int thread_running = 0;
57
58 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
59
60 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
61                                         struct cfs_trace_cpu_data *tcd);
62
63 static inline struct cfs_trace_page *
64 cfs_tage_from_list(struct list_head *list)
65 {
66         return list_entry(list, struct cfs_trace_page, linkage);
67 }
68
69 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
70 {
71         struct page            *page;
72         struct cfs_trace_page *tage;
73
74         /* My caller is trying to free memory */
75         if (!in_interrupt() && memory_pressure_get())
76                 return NULL;
77
78         /*
79          * Don't spam console with allocation failures: they will be reported
80          * by upper layer anyway.
81          */
82         gfp |= __GFP_NOWARN;
83         page = alloc_page(gfp);
84         if (page == NULL)
85                 return NULL;
86
87         tage = kmalloc(sizeof(*tage), gfp);
88         if (tage == NULL) {
89                 __free_page(page);
90                 return NULL;
91         }
92
93         tage->page = page;
94         atomic_inc(&cfs_tage_allocated);
95         return tage;
96 }
97
98 static void cfs_tage_free(struct cfs_trace_page *tage)
99 {
100         __LASSERT(tage != NULL);
101         __LASSERT(tage->page != NULL);
102
103         __free_page(tage->page);
104         kfree(tage);
105         atomic_dec(&cfs_tage_allocated);
106 }
107
108 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
109                              struct list_head *queue)
110 {
111         __LASSERT(tage != NULL);
112         __LASSERT(queue != NULL);
113
114         list_move_tail(&tage->linkage, queue);
115 }
116
117 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
118                            struct list_head *stock)
119 {
120         int i;
121
122         /*
123          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
124          * from here: this will lead to infinite recursion.
125          */
126
127         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
128                 struct cfs_trace_page *tage;
129
130                 tage = cfs_tage_alloc(gfp);
131                 if (tage == NULL)
132                         break;
133                 list_add_tail(&tage->linkage, stock);
134         }
135         return i;
136 }
137
138 /* return a page that has 'len' bytes left at the end */
139 static struct cfs_trace_page *
140 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
141 {
142         struct cfs_trace_page *tage;
143
144         if (tcd->tcd_cur_pages > 0) {
145                 __LASSERT(!list_empty(&tcd->tcd_pages));
146                 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
147                 if (tage->used + len <= PAGE_CACHE_SIZE)
148                         return tage;
149         }
150
151         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
152                 if (tcd->tcd_cur_stock_pages > 0) {
153                         tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
154                         --tcd->tcd_cur_stock_pages;
155                         list_del_init(&tage->linkage);
156                 } else {
157                         tage = cfs_tage_alloc(GFP_ATOMIC);
158                         if (unlikely(tage == NULL)) {
159                                 if ((!memory_pressure_get() ||
160                                      in_interrupt()) && printk_ratelimit())
161                                         printk(KERN_WARNING
162                                                "cannot allocate a tage (%ld)\n",
163                                                tcd->tcd_cur_pages);
164                                 return NULL;
165                         }
166                 }
167
168                 tage->used = 0;
169                 tage->cpu = smp_processor_id();
170                 tage->type = tcd->tcd_type;
171                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
172                 tcd->tcd_cur_pages++;
173
174                 if (tcd->tcd_cur_pages > 8 && thread_running) {
175                         struct tracefiled_ctl *tctl = &trace_tctl;
176                         /*
177                          * wake up tracefiled to process some pages.
178                          */
179                         wake_up(&tctl->tctl_waitq);
180                 }
181                 return tage;
182         }
183         return NULL;
184 }
185
186 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
187 {
188         int pgcount = tcd->tcd_cur_pages / 10;
189         struct page_collection pc;
190         struct cfs_trace_page *tage;
191         struct cfs_trace_page *tmp;
192
193         /*
194          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
195          * from here: this will lead to infinite recursion.
196          */
197
198         if (printk_ratelimit())
199                 printk(KERN_WARNING "debug daemon buffer overflowed; "
200                         "discarding 10%% of pages (%d of %ld)\n",
201                         pgcount + 1, tcd->tcd_cur_pages);
202
203         INIT_LIST_HEAD(&pc.pc_pages);
204
205         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
206                 if (pgcount-- == 0)
207                         break;
208
209                 list_move_tail(&tage->linkage, &pc.pc_pages);
210                 tcd->tcd_cur_pages--;
211         }
212         put_pages_on_tcd_daemon_list(&pc, tcd);
213 }
214
215 /* return a page that has 'len' bytes left at the end */
216 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
217                                                  unsigned long len)
218 {
219         struct cfs_trace_page *tage;
220
221         /*
222          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
223          * from here: this will lead to infinite recursion.
224          */
225
226         if (len > PAGE_CACHE_SIZE) {
227                 printk(KERN_ERR
228                        "cowardly refusing to write %lu bytes in a page\n", len);
229                 return NULL;
230         }
231
232         tage = cfs_trace_get_tage_try(tcd, len);
233         if (tage != NULL)
234                 return tage;
235         if (thread_running)
236                 cfs_tcd_shrink(tcd);
237         if (tcd->tcd_cur_pages > 0) {
238                 tage = cfs_tage_from_list(tcd->tcd_pages.next);
239                 tage->used = 0;
240                 cfs_tage_to_tail(tage, &tcd->tcd_pages);
241         }
242         return tage;
243 }
244
245 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
246                      const char *format, ...)
247 {
248         va_list args;
249         int     rc;
250
251         va_start(args, format);
252         rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
253         va_end(args);
254
255         return rc;
256 }
257 EXPORT_SYMBOL(libcfs_debug_msg);
258
259 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
260                        const char *format1, va_list args,
261                        const char *format2, ...)
262 {
263         struct cfs_trace_cpu_data *tcd = NULL;
264         struct ptldebug_header     header = {0};
265         struct cfs_trace_page     *tage;
266         /* string_buf is used only if tcd != NULL, and is always set then */
267         char                      *string_buf = NULL;
268         char                      *debug_buf;
269         int                        known_size;
270         int                        needed = 85; /* average message length */
271         int                        max_nob;
272         va_list                    ap;
273         int                        i;
274         int                        remain;
275         int                        mask = msgdata->msg_mask;
276         char                      *file = (char *)msgdata->msg_file;
277         cfs_debug_limit_state_t   *cdls = msgdata->msg_cdls;
278
279         if (strchr(file, '/'))
280                 file = strrchr(file, '/') + 1;
281
282         tcd = cfs_trace_get_tcd();
283
284         /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
285          * pins us to a particular CPU.  This avoids an smp_processor_id()
286          * warning on Linux when debugging is enabled. */
287         cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
288
289         if (tcd == NULL)                /* arch may not log in IRQ context */
290                 goto console;
291
292         if (tcd->tcd_cur_pages == 0)
293                 header.ph_flags |= PH_FLAG_FIRST_RECORD;
294
295         if (tcd->tcd_shutting_down) {
296                 cfs_trace_put_tcd(tcd);
297                 tcd = NULL;
298                 goto console;
299         }
300
301         known_size = strlen(file) + 1;
302         if (msgdata->msg_fn)
303                 known_size += strlen(msgdata->msg_fn) + 1;
304
305         if (libcfs_debug_binary)
306                 known_size += sizeof(header);
307
308         /*/
309          * '2' used because vsnprintf return real size required for output
310          * _without_ terminating NULL.
311          * if needed is to small for this format.
312          */
313         for (i = 0; i < 2; i++) {
314                 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
315                 if (tage == NULL) {
316                         if (needed + known_size > PAGE_CACHE_SIZE)
317                                 mask |= D_ERROR;
318
319                         cfs_trace_put_tcd(tcd);
320                         tcd = NULL;
321                         goto console;
322                 }
323
324                 string_buf = (char *)page_address(tage->page) +
325                                         tage->used + known_size;
326
327                 max_nob = PAGE_CACHE_SIZE - tage->used - known_size;
328                 if (max_nob <= 0) {
329                         printk(KERN_EMERG "negative max_nob: %d\n",
330                                max_nob);
331                         mask |= D_ERROR;
332                         cfs_trace_put_tcd(tcd);
333                         tcd = NULL;
334                         goto console;
335                 }
336
337                 needed = 0;
338                 if (format1) {
339                         va_copy(ap, args);
340                         needed = vsnprintf(string_buf, max_nob, format1, ap);
341                         va_end(ap);
342                 }
343
344                 if (format2) {
345                         remain = max_nob - needed;
346                         if (remain < 0)
347                                 remain = 0;
348
349                         va_start(ap, format2);
350                         needed += vsnprintf(string_buf + needed, remain,
351                                             format2, ap);
352                         va_end(ap);
353                 }
354
355                 if (needed < max_nob) /* well. printing ok.. */
356                         break;
357         }
358
359         if (*(string_buf+needed-1) != '\n')
360                 printk(KERN_INFO "format at %s:%d:%s doesn't end in "
361                        "newline\n", file, msgdata->msg_line, msgdata->msg_fn);
362
363         header.ph_len = known_size + needed;
364         debug_buf = (char *)page_address(tage->page) + tage->used;
365
366         if (libcfs_debug_binary) {
367                 memcpy(debug_buf, &header, sizeof(header));
368                 tage->used += sizeof(header);
369                 debug_buf += sizeof(header);
370         }
371
372         strcpy(debug_buf, file);
373         tage->used += strlen(file) + 1;
374         debug_buf += strlen(file) + 1;
375
376         if (msgdata->msg_fn) {
377                 strcpy(debug_buf, msgdata->msg_fn);
378                 tage->used += strlen(msgdata->msg_fn) + 1;
379                 debug_buf += strlen(msgdata->msg_fn) + 1;
380         }
381
382         __LASSERT(debug_buf == string_buf);
383
384         tage->used += needed;
385         __LASSERT(tage->used <= PAGE_CACHE_SIZE);
386
387 console:
388         if ((mask & libcfs_printk) == 0) {
389                 /* no console output requested */
390                 if (tcd != NULL)
391                         cfs_trace_put_tcd(tcd);
392                 return 1;
393         }
394
395         if (cdls != NULL) {
396                 if (libcfs_console_ratelimit &&
397                     cdls->cdls_next != 0 &&     /* not first time ever */
398                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
399                         /* skipping a console message */
400                         cdls->cdls_count++;
401                         if (tcd != NULL)
402                                 cfs_trace_put_tcd(tcd);
403                         return 1;
404                 }
405
406                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
407                                                        libcfs_console_max_delay
408                                                        + cfs_time_seconds(10))) {
409                         /* last timeout was a long time ago */
410                         cdls->cdls_delay /= libcfs_console_backoff * 4;
411                 } else {
412                         cdls->cdls_delay *= libcfs_console_backoff;
413                 }
414
415                 if (cdls->cdls_delay < libcfs_console_min_delay)
416                         cdls->cdls_delay = libcfs_console_min_delay;
417                 else if (cdls->cdls_delay > libcfs_console_max_delay)
418                         cdls->cdls_delay = libcfs_console_max_delay;
419
420                 /* ensure cdls_next is never zero after it's been seen */
421                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
422         }
423
424         if (tcd != NULL) {
425                 cfs_print_to_console(&header, mask, string_buf, needed, file,
426                                      msgdata->msg_fn);
427                 cfs_trace_put_tcd(tcd);
428         } else {
429                 string_buf = cfs_trace_get_console_buffer();
430
431                 needed = 0;
432                 if (format1 != NULL) {
433                         va_copy(ap, args);
434                         needed = vsnprintf(string_buf,
435                                            CFS_TRACE_CONSOLE_BUFFER_SIZE,
436                                            format1, ap);
437                         va_end(ap);
438                 }
439                 if (format2 != NULL) {
440                         remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
441                         if (remain > 0) {
442                                 va_start(ap, format2);
443                                 needed += vsnprintf(string_buf+needed, remain,
444                                                     format2, ap);
445                                 va_end(ap);
446                         }
447                 }
448                 cfs_print_to_console(&header, mask,
449                                      string_buf, needed, file, msgdata->msg_fn);
450
451                 cfs_trace_put_console_buffer(string_buf);
452         }
453
454         if (cdls != NULL && cdls->cdls_count != 0) {
455                 string_buf = cfs_trace_get_console_buffer();
456
457                 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
458                                   "Skipped %d previous similar message%s\n",
459                                   cdls->cdls_count,
460                                   (cdls->cdls_count > 1) ? "s" : "");
461
462                 cfs_print_to_console(&header, mask,
463                                      string_buf, needed, file, msgdata->msg_fn);
464
465                 cfs_trace_put_console_buffer(string_buf);
466                 cdls->cdls_count = 0;
467         }
468
469         return 0;
470 }
471 EXPORT_SYMBOL(libcfs_debug_vmsg2);
472
473 void
474 cfs_trace_assertion_failed(const char *str,
475                            struct libcfs_debug_msg_data *msgdata)
476 {
477         struct ptldebug_header hdr;
478
479         libcfs_panic_in_progress = 1;
480         libcfs_catastrophe = 1;
481         smp_mb();
482
483         cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
484
485         cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
486                              msgdata->msg_file, msgdata->msg_fn);
487
488         panic("Lustre debug assertion failure\n");
489
490         /* not reached */
491 }
492
493 static void
494 panic_collect_pages(struct page_collection *pc)
495 {
496         /* Do the collect_pages job on a single CPU: assumes that all other
497          * CPUs have been stopped during a panic.  If this isn't true for some
498          * arch, this will have to be implemented separately in each arch.  */
499         int                        i;
500         int                        j;
501         struct cfs_trace_cpu_data *tcd;
502
503         INIT_LIST_HEAD(&pc->pc_pages);
504
505         cfs_tcd_for_each(tcd, i, j) {
506                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
507                 tcd->tcd_cur_pages = 0;
508
509                 if (pc->pc_want_daemon_pages) {
510                         list_splice_init(&tcd->tcd_daemon_pages,
511                                                 &pc->pc_pages);
512                         tcd->tcd_cur_daemon_pages = 0;
513                 }
514         }
515 }
516
517 static void collect_pages_on_all_cpus(struct page_collection *pc)
518 {
519         struct cfs_trace_cpu_data *tcd;
520         int i, cpu;
521
522         for_each_possible_cpu(cpu) {
523                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
524                         list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
525                         tcd->tcd_cur_pages = 0;
526                         if (pc->pc_want_daemon_pages) {
527                                 list_splice_init(&tcd->tcd_daemon_pages,
528                                                         &pc->pc_pages);
529                                 tcd->tcd_cur_daemon_pages = 0;
530                         }
531                 }
532         }
533 }
534
535 static void collect_pages(struct page_collection *pc)
536 {
537         INIT_LIST_HEAD(&pc->pc_pages);
538
539         if (libcfs_panic_in_progress)
540                 panic_collect_pages(pc);
541         else
542                 collect_pages_on_all_cpus(pc);
543 }
544
545 static void put_pages_back_on_all_cpus(struct page_collection *pc)
546 {
547         struct cfs_trace_cpu_data *tcd;
548         struct list_head *cur_head;
549         struct cfs_trace_page *tage;
550         struct cfs_trace_page *tmp;
551         int i, cpu;
552
553         for_each_possible_cpu(cpu) {
554                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
555                         cur_head = tcd->tcd_pages.next;
556
557                         list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
558                                                  linkage) {
559
560                                 __LASSERT_TAGE_INVARIANT(tage);
561
562                                 if (tage->cpu != cpu || tage->type != i)
563                                         continue;
564
565                                 cfs_tage_to_tail(tage, cur_head);
566                                 tcd->tcd_cur_pages++;
567                         }
568                 }
569         }
570 }
571
572 static void put_pages_back(struct page_collection *pc)
573 {
574         if (!libcfs_panic_in_progress)
575                 put_pages_back_on_all_cpus(pc);
576 }
577
578 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
579  * we have a good amount of data at all times for dumping during an LBUG, even
580  * if we have been steadily writing (and otherwise discarding) pages via the
581  * debug daemon. */
582 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
583                                          struct cfs_trace_cpu_data *tcd)
584 {
585         struct cfs_trace_page *tage;
586         struct cfs_trace_page *tmp;
587
588         list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
589                 __LASSERT_TAGE_INVARIANT(tage);
590
591                 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
592                         continue;
593
594                 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
595                 tcd->tcd_cur_daemon_pages++;
596
597                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
598                         struct cfs_trace_page *victim;
599
600                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
601                         victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
602
603                         __LASSERT_TAGE_INVARIANT(victim);
604
605                         list_del(&victim->linkage);
606                         cfs_tage_free(victim);
607                         tcd->tcd_cur_daemon_pages--;
608                 }
609         }
610 }
611
612 static void put_pages_on_daemon_list(struct page_collection *pc)
613 {
614         struct cfs_trace_cpu_data *tcd;
615         int i, cpu;
616
617         for_each_possible_cpu(cpu) {
618                 cfs_tcd_for_each_type_lock(tcd, i, cpu)
619                         put_pages_on_tcd_daemon_list(pc, tcd);
620         }
621 }
622
623 void cfs_trace_debug_print(void)
624 {
625         struct page_collection pc;
626         struct cfs_trace_page *tage;
627         struct cfs_trace_page *tmp;
628
629         pc.pc_want_daemon_pages = 1;
630         collect_pages(&pc);
631         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
632                 char *p, *file, *fn;
633                 struct page *page;
634
635                 __LASSERT_TAGE_INVARIANT(tage);
636
637                 page = tage->page;
638                 p = page_address(page);
639                 while (p < ((char *)page_address(page) + tage->used)) {
640                         struct ptldebug_header *hdr;
641                         int len;
642                         hdr = (void *)p;
643                         p += sizeof(*hdr);
644                         file = p;
645                         p += strlen(file) + 1;
646                         fn = p;
647                         p += strlen(fn) + 1;
648                         len = hdr->ph_len - (int)(p - (char *)hdr);
649
650                         cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
651
652                         p += len;
653                 }
654
655                 list_del(&tage->linkage);
656                 cfs_tage_free(tage);
657         }
658 }
659
660 int cfs_tracefile_dump_all_pages(char *filename)
661 {
662         struct page_collection  pc;
663         struct file             *filp;
664         struct cfs_trace_page   *tage;
665         struct cfs_trace_page   *tmp;
666         int rc;
667
668         DECL_MMSPACE;
669
670         cfs_tracefile_write_lock();
671
672         filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
673         if (IS_ERR(filp)) {
674                 rc = PTR_ERR(filp);
675                 filp = NULL;
676                 printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
677                       filename, rc);
678                 goto out;
679         }
680
681         pc.pc_want_daemon_pages = 1;
682         collect_pages(&pc);
683         if (list_empty(&pc.pc_pages)) {
684                 rc = 0;
685                 goto close;
686         }
687
688         /* ok, for now, just write the pages.  in the future we'll be building
689          * iobufs with the pages and calling generic_direct_IO */
690         MMSPACE_OPEN;
691         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
692
693                 __LASSERT_TAGE_INVARIANT(tage);
694
695                 rc = filp_write(filp, page_address(tage->page),
696                                 tage->used, filp_poff(filp));
697                 if (rc != (int)tage->used) {
698                         printk(KERN_WARNING "wanted to write %u but wrote "
699                                "%d\n", tage->used, rc);
700                         put_pages_back(&pc);
701                         __LASSERT(list_empty(&pc.pc_pages));
702                         break;
703                 }
704                 list_del(&tage->linkage);
705                 cfs_tage_free(tage);
706         }
707         MMSPACE_CLOSE;
708         rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
709         if (rc)
710                 printk(KERN_ERR "sync returns %d\n", rc);
711 close:
712         filp_close(filp, NULL);
713 out:
714         cfs_tracefile_write_unlock();
715         return rc;
716 }
717
718 void cfs_trace_flush_pages(void)
719 {
720         struct page_collection pc;
721         struct cfs_trace_page *tage;
722         struct cfs_trace_page *tmp;
723
724         pc.pc_want_daemon_pages = 1;
725         collect_pages(&pc);
726         list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
727
728                 __LASSERT_TAGE_INVARIANT(tage);
729
730                 list_del(&tage->linkage);
731                 cfs_tage_free(tage);
732         }
733 }
734
735 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
736                             const char __user *usr_buffer, int usr_buffer_nob)
737 {
738         int    nob;
739
740         if (usr_buffer_nob > knl_buffer_nob)
741                 return -EOVERFLOW;
742
743         if (copy_from_user(knl_buffer, usr_buffer, usr_buffer_nob))
744                 return -EFAULT;
745
746         nob = strnlen(knl_buffer, usr_buffer_nob);
747         while (nob-- >= 0)                      /* strip trailing whitespace */
748                 if (!isspace(knl_buffer[nob]))
749                         break;
750
751         if (nob < 0)                            /* empty string */
752                 return -EINVAL;
753
754         if (nob == knl_buffer_nob)              /* no space to terminate */
755                 return -EOVERFLOW;
756
757         knl_buffer[nob + 1] = 0;                /* terminate */
758         return 0;
759 }
760 EXPORT_SYMBOL(cfs_trace_copyin_string);
761
762 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
763                              const char *knl_buffer, char *append)
764 {
765         /* NB if 'append' != NULL, it's a single character to append to the
766          * copied out string - usually "\n", for /proc entries and "" (i.e. a
767          * terminating zero byte) for sysctl entries */
768         int   nob = strlen(knl_buffer);
769
770         if (nob > usr_buffer_nob)
771                 nob = usr_buffer_nob;
772
773         if (copy_to_user(usr_buffer, knl_buffer, nob))
774                 return -EFAULT;
775
776         if (append != NULL && nob < usr_buffer_nob) {
777                 if (copy_to_user(usr_buffer + nob, append, 1))
778                         return -EFAULT;
779
780                 nob++;
781         }
782
783         return nob;
784 }
785 EXPORT_SYMBOL(cfs_trace_copyout_string);
786
787 int cfs_trace_allocate_string_buffer(char **str, int nob)
788 {
789         if (nob > 2 * PAGE_CACHE_SIZE)  /* string must be "sensible" */
790                 return -EINVAL;
791
792         *str = kmalloc(nob, GFP_IOFS | __GFP_ZERO);
793         if (*str == NULL)
794                 return -ENOMEM;
795
796         return 0;
797 }
798
799 void cfs_trace_free_string_buffer(char *str, int nob)
800 {
801         kfree(str);
802 }
803
804 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
805 {
806         char         *str;
807         int           rc;
808
809         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
810         if (rc != 0)
811                 return rc;
812
813         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
814                                      usr_str, usr_str_nob);
815         if (rc != 0)
816                 goto out;
817
818         if (str[0] != '/') {
819                 rc = -EINVAL;
820                 goto out;
821         }
822         rc = cfs_tracefile_dump_all_pages(str);
823 out:
824         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
825         return rc;
826 }
827
828 int cfs_trace_daemon_command(char *str)
829 {
830         int       rc = 0;
831
832         cfs_tracefile_write_lock();
833
834         if (strcmp(str, "stop") == 0) {
835                 cfs_tracefile_write_unlock();
836                 cfs_trace_stop_thread();
837                 cfs_tracefile_write_lock();
838                 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
839
840         } else if (strncmp(str, "size=", 5) == 0) {
841                 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
842                 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
843                         cfs_tracefile_size = CFS_TRACEFILE_SIZE;
844                 else
845                         cfs_tracefile_size <<= 20;
846
847         } else if (strlen(str) >= sizeof(cfs_tracefile)) {
848                 rc = -ENAMETOOLONG;
849         } else if (str[0] != '/') {
850                 rc = -EINVAL;
851         } else {
852                 strcpy(cfs_tracefile, str);
853
854                 printk(KERN_INFO
855                        "Lustre: debug daemon will attempt to start writing "
856                        "to %s (%lukB max)\n", cfs_tracefile,
857                        (long)(cfs_tracefile_size >> 10));
858
859                 cfs_trace_start_thread();
860         }
861
862         cfs_tracefile_write_unlock();
863         return rc;
864 }
865
866 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
867 {
868         char *str;
869         int   rc;
870
871         rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
872         if (rc != 0)
873                 return rc;
874
875         rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
876                                  usr_str, usr_str_nob);
877         if (rc == 0)
878                 rc = cfs_trace_daemon_command(str);
879
880         cfs_trace_free_string_buffer(str, usr_str_nob + 1);
881         return rc;
882 }
883
884 int cfs_trace_set_debug_mb(int mb)
885 {
886         int i;
887         int j;
888         int pages;
889         int limit = cfs_trace_max_debug_mb();
890         struct cfs_trace_cpu_data *tcd;
891
892         if (mb < num_possible_cpus()) {
893                 printk(KERN_WARNING
894                        "Lustre: %d MB is too small for debug buffer size, "
895                        "setting it to %d MB.\n", mb, num_possible_cpus());
896                 mb = num_possible_cpus();
897         }
898
899         if (mb > limit) {
900                 printk(KERN_WARNING
901                        "Lustre: %d MB is too large for debug buffer size, "
902                        "setting it to %d MB.\n", mb, limit);
903                 mb = limit;
904         }
905
906         mb /= num_possible_cpus();
907         pages = mb << (20 - PAGE_CACHE_SHIFT);
908
909         cfs_tracefile_write_lock();
910
911         cfs_tcd_for_each(tcd, i, j)
912                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
913
914         cfs_tracefile_write_unlock();
915
916         return 0;
917 }
918
919 int cfs_trace_set_debug_mb_usrstr(void __user *usr_str, int usr_str_nob)
920 {
921         char     str[32];
922         int      rc;
923
924         rc = cfs_trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
925         if (rc < 0)
926                 return rc;
927
928         return cfs_trace_set_debug_mb(simple_strtoul(str, NULL, 0));
929 }
930
931 int cfs_trace_get_debug_mb(void)
932 {
933         int i;
934         int j;
935         struct cfs_trace_cpu_data *tcd;
936         int total_pages = 0;
937
938         cfs_tracefile_read_lock();
939
940         cfs_tcd_for_each(tcd, i, j)
941                 total_pages += tcd->tcd_max_pages;
942
943         cfs_tracefile_read_unlock();
944
945         return (total_pages >> (20 - PAGE_CACHE_SHIFT)) + 1;
946 }
947
948 static int tracefiled(void *arg)
949 {
950         struct page_collection pc;
951         struct tracefiled_ctl *tctl = arg;
952         struct cfs_trace_page *tage;
953         struct cfs_trace_page *tmp;
954         struct file *filp;
955         int last_loop = 0;
956         int rc;
957
958         DECL_MMSPACE;
959
960         /* we're started late enough that we pick up init's fs context */
961         /* this is so broken in uml?  what on earth is going on? */
962
963         complete(&tctl->tctl_start);
964
965         while (1) {
966                 wait_queue_t __wait;
967
968                 pc.pc_want_daemon_pages = 0;
969                 collect_pages(&pc);
970                 if (list_empty(&pc.pc_pages))
971                         goto end_loop;
972
973                 filp = NULL;
974                 cfs_tracefile_read_lock();
975                 if (cfs_tracefile[0] != 0) {
976                         filp = filp_open(cfs_tracefile,
977                                          O_CREAT | O_RDWR | O_LARGEFILE,
978                                          0600);
979                         if (IS_ERR(filp)) {
980                                 rc = PTR_ERR(filp);
981                                 filp = NULL;
982                                 printk(KERN_WARNING "couldn't open %s: "
983                                        "%d\n", cfs_tracefile, rc);
984                         }
985                 }
986                 cfs_tracefile_read_unlock();
987                 if (filp == NULL) {
988                         put_pages_on_daemon_list(&pc);
989                         __LASSERT(list_empty(&pc.pc_pages));
990                         goto end_loop;
991                 }
992
993                 MMSPACE_OPEN;
994
995                 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
996                         static loff_t f_pos;
997
998                         __LASSERT_TAGE_INVARIANT(tage);
999
1000                         if (f_pos >= (off_t)cfs_tracefile_size)
1001                                 f_pos = 0;
1002                         else if (f_pos > (off_t)filp_size(filp))
1003                                 f_pos = filp_size(filp);
1004
1005                         rc = filp_write(filp, page_address(tage->page),
1006                                         tage->used, &f_pos);
1007                         if (rc != (int)tage->used) {
1008                                 printk(KERN_WARNING "wanted to write %u "
1009                                        "but wrote %d\n", tage->used, rc);
1010                                 put_pages_back(&pc);
1011                                 __LASSERT(list_empty(&pc.pc_pages));
1012                                 break;
1013                         }
1014                 }
1015                 MMSPACE_CLOSE;
1016
1017                 filp_close(filp, NULL);
1018                 put_pages_on_daemon_list(&pc);
1019                 if (!list_empty(&pc.pc_pages)) {
1020                         int i;
1021
1022                         printk(KERN_ALERT "Lustre: trace pages aren't "
1023                                " empty\n");
1024                         printk(KERN_ERR "total cpus(%d): ",
1025                                num_possible_cpus());
1026                         for (i = 0; i < num_possible_cpus(); i++)
1027                                 if (cpu_online(i))
1028                                         printk(KERN_ERR "%d(on) ", i);
1029                                 else
1030                                         printk(KERN_ERR "%d(off) ", i);
1031                         printk(KERN_ERR "\n");
1032
1033                         i = 0;
1034                         list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1035                                                      linkage)
1036                                 printk(KERN_ERR "page %d belongs to cpu "
1037                                        "%d\n", ++i, tage->cpu);
1038                         printk(KERN_ERR "There are %d pages unwritten\n",
1039                                i);
1040                 }
1041                 __LASSERT(list_empty(&pc.pc_pages));
1042 end_loop:
1043                 if (atomic_read(&tctl->tctl_shutdown)) {
1044                         if (last_loop == 0) {
1045                                 last_loop = 1;
1046                                 continue;
1047                         } else {
1048                                 break;
1049                         }
1050                 }
1051                 init_waitqueue_entry_current(&__wait);
1052                 add_wait_queue(&tctl->tctl_waitq, &__wait);
1053                 set_current_state(TASK_INTERRUPTIBLE);
1054                 waitq_timedwait(&__wait, TASK_INTERRUPTIBLE,
1055                                 cfs_time_seconds(1));
1056                 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1057         }
1058         complete(&tctl->tctl_stop);
1059         return 0;
1060 }
1061
1062 int cfs_trace_start_thread(void)
1063 {
1064         struct tracefiled_ctl *tctl = &trace_tctl;
1065         int rc = 0;
1066
1067         mutex_lock(&cfs_trace_thread_mutex);
1068         if (thread_running)
1069                 goto out;
1070
1071         init_completion(&tctl->tctl_start);
1072         init_completion(&tctl->tctl_stop);
1073         init_waitqueue_head(&tctl->tctl_waitq);
1074         atomic_set(&tctl->tctl_shutdown, 0);
1075
1076         if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1077                 rc = -ECHILD;
1078                 goto out;
1079         }
1080
1081         wait_for_completion(&tctl->tctl_start);
1082         thread_running = 1;
1083 out:
1084         mutex_unlock(&cfs_trace_thread_mutex);
1085         return rc;
1086 }
1087
1088 void cfs_trace_stop_thread(void)
1089 {
1090         struct tracefiled_ctl *tctl = &trace_tctl;
1091
1092         mutex_lock(&cfs_trace_thread_mutex);
1093         if (thread_running) {
1094                 printk(KERN_INFO
1095                        "Lustre: shutting down debug daemon thread...\n");
1096                 atomic_set(&tctl->tctl_shutdown, 1);
1097                 wait_for_completion(&tctl->tctl_stop);
1098                 thread_running = 0;
1099         }
1100         mutex_unlock(&cfs_trace_thread_mutex);
1101 }
1102
1103 int cfs_tracefile_init(int max_pages)
1104 {
1105         struct cfs_trace_cpu_data *tcd;
1106         int     i;
1107         int     j;
1108         int     rc;
1109         int     factor;
1110
1111         rc = cfs_tracefile_init_arch();
1112         if (rc != 0)
1113                 return rc;
1114
1115         cfs_tcd_for_each(tcd, i, j) {
1116                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1117                 factor = tcd->tcd_pages_factor;
1118                 INIT_LIST_HEAD(&tcd->tcd_pages);
1119                 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1120                 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1121                 tcd->tcd_cur_pages = 0;
1122                 tcd->tcd_cur_stock_pages = 0;
1123                 tcd->tcd_cur_daemon_pages = 0;
1124                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1125                 LASSERT(tcd->tcd_max_pages > 0);
1126                 tcd->tcd_shutting_down = 0;
1127         }
1128         return 0;
1129 }
1130
1131 static void trace_cleanup_on_all_cpus(void)
1132 {
1133         struct cfs_trace_cpu_data *tcd;
1134         struct cfs_trace_page *tage;
1135         struct cfs_trace_page *tmp;
1136         int i, cpu;
1137
1138         for_each_possible_cpu(cpu) {
1139                 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1140                         tcd->tcd_shutting_down = 1;
1141
1142                         list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
1143                                 __LASSERT_TAGE_INVARIANT(tage);
1144
1145                                 list_del(&tage->linkage);
1146                                 cfs_tage_free(tage);
1147                         }
1148                         tcd->tcd_cur_pages = 0;
1149                 }
1150         }
1151 }
1152
1153 static void cfs_trace_cleanup(void)
1154 {
1155         struct page_collection pc;
1156
1157         INIT_LIST_HEAD(&pc.pc_pages);
1158
1159         trace_cleanup_on_all_cpus();
1160
1161         cfs_tracefile_fini_arch();
1162 }
1163
1164 void cfs_tracefile_exit(void)
1165 {
1166         cfs_trace_stop_thread();
1167         cfs_trace_cleanup();
1168 }