4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * libcfs/libcfs/tracefile.c
38 * Author: Zach Brown <zab@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LNET
43 #define LUSTRE_TRACEFILE_PRIVATE
44 #include "tracefile.h"
46 #include "../../include/linux/libcfs/libcfs.h"
48 /* XXX move things up to the top, comment */
49 union cfs_trace_data_union (*cfs_trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
51 char cfs_tracefile[TRACEFILE_NAME_SIZE];
52 long long cfs_tracefile_size = CFS_TRACEFILE_SIZE;
53 static struct tracefiled_ctl trace_tctl;
54 static DEFINE_MUTEX(cfs_trace_thread_mutex);
55 static int thread_running;
57 static atomic_t cfs_tage_allocated = ATOMIC_INIT(0);
59 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
60 struct cfs_trace_cpu_data *tcd);
62 static inline struct cfs_trace_page *
63 cfs_tage_from_list(struct list_head *list)
65 return list_entry(list, struct cfs_trace_page, linkage);
68 static struct cfs_trace_page *cfs_tage_alloc(gfp_t gfp)
71 struct cfs_trace_page *tage;
73 /* My caller is trying to free memory */
74 if (!in_interrupt() && memory_pressure_get())
78 * Don't spam console with allocation failures: they will be reported
79 * by upper layer anyway.
82 page = alloc_page(gfp);
86 tage = kmalloc(sizeof(*tage), gfp);
93 atomic_inc(&cfs_tage_allocated);
97 static void cfs_tage_free(struct cfs_trace_page *tage)
99 __LASSERT(tage != NULL);
100 __LASSERT(tage->page != NULL);
102 __free_page(tage->page);
104 atomic_dec(&cfs_tage_allocated);
107 static void cfs_tage_to_tail(struct cfs_trace_page *tage,
108 struct list_head *queue)
110 __LASSERT(tage != NULL);
111 __LASSERT(queue != NULL);
113 list_move_tail(&tage->linkage, queue);
116 int cfs_trace_refill_stock(struct cfs_trace_cpu_data *tcd, gfp_t gfp,
117 struct list_head *stock)
122 * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
123 * from here: this will lead to infinite recursion.
126 for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++i) {
127 struct cfs_trace_page *tage;
129 tage = cfs_tage_alloc(gfp);
132 list_add_tail(&tage->linkage, stock);
137 /* return a page that has 'len' bytes left at the end */
138 static struct cfs_trace_page *
139 cfs_trace_get_tage_try(struct cfs_trace_cpu_data *tcd, unsigned long len)
141 struct cfs_trace_page *tage;
143 if (tcd->tcd_cur_pages > 0) {
144 __LASSERT(!list_empty(&tcd->tcd_pages));
145 tage = cfs_tage_from_list(tcd->tcd_pages.prev);
146 if (tage->used + len <= PAGE_CACHE_SIZE)
150 if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
151 if (tcd->tcd_cur_stock_pages > 0) {
152 tage = cfs_tage_from_list(tcd->tcd_stock_pages.prev);
153 --tcd->tcd_cur_stock_pages;
154 list_del_init(&tage->linkage);
156 tage = cfs_tage_alloc(GFP_ATOMIC);
157 if (unlikely(tage == NULL)) {
158 if ((!memory_pressure_get() ||
159 in_interrupt()) && printk_ratelimit())
161 "cannot allocate a tage (%ld)\n",
168 tage->cpu = smp_processor_id();
169 tage->type = tcd->tcd_type;
170 list_add_tail(&tage->linkage, &tcd->tcd_pages);
171 tcd->tcd_cur_pages++;
173 if (tcd->tcd_cur_pages > 8 && thread_running) {
174 struct tracefiled_ctl *tctl = &trace_tctl;
176 * wake up tracefiled to process some pages.
178 wake_up(&tctl->tctl_waitq);
185 static void cfs_tcd_shrink(struct cfs_trace_cpu_data *tcd)
187 int pgcount = tcd->tcd_cur_pages / 10;
188 struct page_collection pc;
189 struct cfs_trace_page *tage;
190 struct cfs_trace_page *tmp;
193 * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
194 * from here: this will lead to infinite recursion.
197 if (printk_ratelimit())
198 printk(KERN_WARNING "debug daemon buffer overflowed; discarding 10%% of pages (%d of %ld)\n",
199 pgcount + 1, tcd->tcd_cur_pages);
201 INIT_LIST_HEAD(&pc.pc_pages);
202 spin_lock_init(&pc.pc_lock);
204 list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages, linkage) {
208 list_move_tail(&tage->linkage, &pc.pc_pages);
209 tcd->tcd_cur_pages--;
211 put_pages_on_tcd_daemon_list(&pc, tcd);
214 /* return a page that has 'len' bytes left at the end */
215 static struct cfs_trace_page *cfs_trace_get_tage(struct cfs_trace_cpu_data *tcd,
218 struct cfs_trace_page *tage;
221 * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
222 * from here: this will lead to infinite recursion.
225 if (len > PAGE_CACHE_SIZE) {
226 pr_err("cowardly refusing to write %lu bytes in a page\n", len);
230 tage = cfs_trace_get_tage_try(tcd, len);
235 if (tcd->tcd_cur_pages > 0) {
236 tage = cfs_tage_from_list(tcd->tcd_pages.next);
238 cfs_tage_to_tail(tage, &tcd->tcd_pages);
243 int libcfs_debug_msg(struct libcfs_debug_msg_data *msgdata,
244 const char *format, ...)
249 va_start(args, format);
250 rc = libcfs_debug_vmsg2(msgdata, format, args, NULL);
255 EXPORT_SYMBOL(libcfs_debug_msg);
257 int libcfs_debug_vmsg2(struct libcfs_debug_msg_data *msgdata,
258 const char *format1, va_list args,
259 const char *format2, ...)
261 struct cfs_trace_cpu_data *tcd = NULL;
262 struct ptldebug_header header = {0};
263 struct cfs_trace_page *tage;
264 /* string_buf is used only if tcd != NULL, and is always set then */
265 char *string_buf = NULL;
268 int needed = 85; /* average message length */
274 int mask = msgdata->msg_mask;
275 const char *file = kbasename(msgdata->msg_file);
276 struct cfs_debug_limit_state *cdls = msgdata->msg_cdls;
278 tcd = cfs_trace_get_tcd();
280 /* cfs_trace_get_tcd() grabs a lock, which disables preemption and
281 * pins us to a particular CPU. This avoids an smp_processor_id()
282 * warning on Linux when debugging is enabled. */
283 cfs_set_ptldebug_header(&header, msgdata, CDEBUG_STACK());
285 if (tcd == NULL) /* arch may not log in IRQ context */
288 if (tcd->tcd_cur_pages == 0)
289 header.ph_flags |= PH_FLAG_FIRST_RECORD;
291 if (tcd->tcd_shutting_down) {
292 cfs_trace_put_tcd(tcd);
297 depth = __current_nesting_level();
298 known_size = strlen(file) + 1 + depth;
300 known_size += strlen(msgdata->msg_fn) + 1;
302 if (libcfs_debug_binary)
303 known_size += sizeof(header);
306 * '2' used because vsnprintf return real size required for output
307 * _without_ terminating NULL.
308 * if needed is to small for this format.
310 for (i = 0; i < 2; i++) {
311 tage = cfs_trace_get_tage(tcd, needed + known_size + 1);
313 if (needed + known_size > PAGE_CACHE_SIZE)
316 cfs_trace_put_tcd(tcd);
321 string_buf = (char *)page_address(tage->page) +
322 tage->used + known_size;
324 max_nob = PAGE_CACHE_SIZE - tage->used - known_size;
326 printk(KERN_EMERG "negative max_nob: %d\n",
329 cfs_trace_put_tcd(tcd);
337 needed = vsnprintf(string_buf, max_nob, format1, ap);
342 remain = max_nob - needed;
346 va_start(ap, format2);
347 needed += vsnprintf(string_buf + needed, remain,
352 if (needed < max_nob) /* well. printing ok.. */
356 if (*(string_buf+needed-1) != '\n')
357 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
358 file, msgdata->msg_line, msgdata->msg_fn);
360 header.ph_len = known_size + needed;
361 debug_buf = (char *)page_address(tage->page) + tage->used;
363 if (libcfs_debug_binary) {
364 memcpy(debug_buf, &header, sizeof(header));
365 tage->used += sizeof(header);
366 debug_buf += sizeof(header);
369 /* indent message according to the nesting level */
370 while (depth-- > 0) {
371 *(debug_buf++) = '.';
375 strcpy(debug_buf, file);
376 tage->used += strlen(file) + 1;
377 debug_buf += strlen(file) + 1;
379 if (msgdata->msg_fn) {
380 strcpy(debug_buf, msgdata->msg_fn);
381 tage->used += strlen(msgdata->msg_fn) + 1;
382 debug_buf += strlen(msgdata->msg_fn) + 1;
385 __LASSERT(debug_buf == string_buf);
387 tage->used += needed;
388 __LASSERT (tage->used <= PAGE_CACHE_SIZE);
391 if ((mask & libcfs_printk) == 0) {
392 /* no console output requested */
394 cfs_trace_put_tcd(tcd);
399 if (libcfs_console_ratelimit &&
400 cdls->cdls_next != 0 && /* not first time ever */
401 !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
402 /* skipping a console message */
405 cfs_trace_put_tcd(tcd);
409 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
410 libcfs_console_max_delay
411 + cfs_time_seconds(10))) {
412 /* last timeout was a long time ago */
413 cdls->cdls_delay /= libcfs_console_backoff * 4;
415 cdls->cdls_delay *= libcfs_console_backoff;
418 if (cdls->cdls_delay < libcfs_console_min_delay)
419 cdls->cdls_delay = libcfs_console_min_delay;
420 else if (cdls->cdls_delay > libcfs_console_max_delay)
421 cdls->cdls_delay = libcfs_console_max_delay;
423 /* ensure cdls_next is never zero after it's been seen */
424 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
428 cfs_print_to_console(&header, mask, string_buf, needed, file,
430 cfs_trace_put_tcd(tcd);
432 string_buf = cfs_trace_get_console_buffer();
435 if (format1 != NULL) {
437 needed = vsnprintf(string_buf,
438 CFS_TRACE_CONSOLE_BUFFER_SIZE,
442 if (format2 != NULL) {
443 remain = CFS_TRACE_CONSOLE_BUFFER_SIZE - needed;
445 va_start(ap, format2);
446 needed += vsnprintf(string_buf+needed, remain,
451 cfs_print_to_console(&header, mask,
452 string_buf, needed, file, msgdata->msg_fn);
454 cfs_trace_put_console_buffer(string_buf);
457 if (cdls != NULL && cdls->cdls_count != 0) {
458 string_buf = cfs_trace_get_console_buffer();
460 needed = snprintf(string_buf, CFS_TRACE_CONSOLE_BUFFER_SIZE,
461 "Skipped %d previous similar message%s\n",
463 (cdls->cdls_count > 1) ? "s" : "");
465 cfs_print_to_console(&header, mask,
466 string_buf, needed, file, msgdata->msg_fn);
468 cfs_trace_put_console_buffer(string_buf);
469 cdls->cdls_count = 0;
474 EXPORT_SYMBOL(libcfs_debug_vmsg2);
477 cfs_trace_assertion_failed(const char *str,
478 struct libcfs_debug_msg_data *msgdata)
480 struct ptldebug_header hdr;
482 libcfs_panic_in_progress = 1;
483 libcfs_catastrophe = 1;
486 cfs_set_ptldebug_header(&hdr, msgdata, CDEBUG_STACK());
488 cfs_print_to_console(&hdr, D_EMERG, str, strlen(str),
489 msgdata->msg_file, msgdata->msg_fn);
491 panic("Lustre debug assertion failure\n");
497 panic_collect_pages(struct page_collection *pc)
499 /* Do the collect_pages job on a single CPU: assumes that all other
500 * CPUs have been stopped during a panic. If this isn't true for some
501 * arch, this will have to be implemented separately in each arch. */
504 struct cfs_trace_cpu_data *tcd;
506 INIT_LIST_HEAD(&pc->pc_pages);
508 cfs_tcd_for_each(tcd, i, j) {
509 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
510 tcd->tcd_cur_pages = 0;
512 if (pc->pc_want_daemon_pages) {
513 list_splice_init(&tcd->tcd_daemon_pages,
515 tcd->tcd_cur_daemon_pages = 0;
520 static void collect_pages_on_all_cpus(struct page_collection *pc)
522 struct cfs_trace_cpu_data *tcd;
525 spin_lock(&pc->pc_lock);
526 for_each_possible_cpu(cpu) {
527 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
528 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
529 tcd->tcd_cur_pages = 0;
530 if (pc->pc_want_daemon_pages) {
531 list_splice_init(&tcd->tcd_daemon_pages,
533 tcd->tcd_cur_daemon_pages = 0;
537 spin_unlock(&pc->pc_lock);
540 static void collect_pages(struct page_collection *pc)
542 INIT_LIST_HEAD(&pc->pc_pages);
544 if (libcfs_panic_in_progress)
545 panic_collect_pages(pc);
547 collect_pages_on_all_cpus(pc);
550 static void put_pages_back_on_all_cpus(struct page_collection *pc)
552 struct cfs_trace_cpu_data *tcd;
553 struct list_head *cur_head;
554 struct cfs_trace_page *tage;
555 struct cfs_trace_page *tmp;
558 spin_lock(&pc->pc_lock);
559 for_each_possible_cpu(cpu) {
560 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
561 cur_head = tcd->tcd_pages.next;
563 list_for_each_entry_safe(tage, tmp, &pc->pc_pages,
566 __LASSERT_TAGE_INVARIANT(tage);
568 if (tage->cpu != cpu || tage->type != i)
571 cfs_tage_to_tail(tage, cur_head);
572 tcd->tcd_cur_pages++;
576 spin_unlock(&pc->pc_lock);
579 static void put_pages_back(struct page_collection *pc)
581 if (!libcfs_panic_in_progress)
582 put_pages_back_on_all_cpus(pc);
585 /* Add pages to a per-cpu debug daemon ringbuffer. This buffer makes sure that
586 * we have a good amount of data at all times for dumping during an LBUG, even
587 * if we have been steadily writing (and otherwise discarding) pages via the
589 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
590 struct cfs_trace_cpu_data *tcd)
592 struct cfs_trace_page *tage;
593 struct cfs_trace_page *tmp;
595 spin_lock(&pc->pc_lock);
596 list_for_each_entry_safe(tage, tmp, &pc->pc_pages, linkage) {
598 __LASSERT_TAGE_INVARIANT(tage);
600 if (tage->cpu != tcd->tcd_cpu || tage->type != tcd->tcd_type)
603 cfs_tage_to_tail(tage, &tcd->tcd_daemon_pages);
604 tcd->tcd_cur_daemon_pages++;
606 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
607 struct cfs_trace_page *victim;
609 __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
610 victim = cfs_tage_from_list(tcd->tcd_daemon_pages.next);
612 __LASSERT_TAGE_INVARIANT(victim);
614 list_del(&victim->linkage);
615 cfs_tage_free(victim);
616 tcd->tcd_cur_daemon_pages--;
619 spin_unlock(&pc->pc_lock);
622 static void put_pages_on_daemon_list(struct page_collection *pc)
624 struct cfs_trace_cpu_data *tcd;
627 for_each_possible_cpu(cpu) {
628 cfs_tcd_for_each_type_lock(tcd, i, cpu)
629 put_pages_on_tcd_daemon_list(pc, tcd);
633 void cfs_trace_debug_print(void)
635 struct page_collection pc;
636 struct cfs_trace_page *tage;
637 struct cfs_trace_page *tmp;
639 spin_lock_init(&pc.pc_lock);
641 pc.pc_want_daemon_pages = 1;
643 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
647 __LASSERT_TAGE_INVARIANT(tage);
650 p = page_address(page);
651 while (p < ((char *)page_address(page) + tage->used)) {
652 struct ptldebug_header *hdr;
658 p += strlen(file) + 1;
661 len = hdr->ph_len - (int)(p - (char *)hdr);
663 cfs_print_to_console(hdr, D_EMERG, p, len, file, fn);
668 list_del(&tage->linkage);
673 int cfs_tracefile_dump_all_pages(char *filename)
675 struct page_collection pc;
677 struct cfs_trace_page *tage;
678 struct cfs_trace_page *tmp;
684 cfs_tracefile_write_lock();
686 filp = filp_open(filename, O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600);
690 pr_err("LustreError: can't open %s for dump: rc %d\n",
695 spin_lock_init(&pc.pc_lock);
696 pc.pc_want_daemon_pages = 1;
698 if (list_empty(&pc.pc_pages)) {
703 /* ok, for now, just write the pages. in the future we'll be building
704 * iobufs with the pages and calling generic_direct_IO */
706 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
708 __LASSERT_TAGE_INVARIANT(tage);
710 buf = kmap(tage->page);
711 rc = vfs_write(filp, (__force const char __user *)buf,
712 tage->used, &filp->f_pos);
715 if (rc != (int)tage->used) {
716 printk(KERN_WARNING "wanted to write %u but wrote %d\n",
719 __LASSERT(list_empty(&pc.pc_pages));
722 list_del(&tage->linkage);
726 rc = vfs_fsync(filp, 1);
728 pr_err("sync returns %d\n", rc);
730 filp_close(filp, NULL);
732 cfs_tracefile_write_unlock();
736 void cfs_trace_flush_pages(void)
738 struct page_collection pc;
739 struct cfs_trace_page *tage;
740 struct cfs_trace_page *tmp;
742 spin_lock_init(&pc.pc_lock);
744 pc.pc_want_daemon_pages = 1;
746 list_for_each_entry_safe(tage, tmp, &pc.pc_pages, linkage) {
748 __LASSERT_TAGE_INVARIANT(tage);
750 list_del(&tage->linkage);
755 int cfs_trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
756 const char __user *usr_buffer, int usr_buffer_nob)
760 if (usr_buffer_nob > knl_buffer_nob)
763 if (copy_from_user((void *)knl_buffer,
764 usr_buffer, usr_buffer_nob))
767 nob = strnlen(knl_buffer, usr_buffer_nob);
768 while (nob-- >= 0) /* strip trailing whitespace */
769 if (!isspace(knl_buffer[nob]))
772 if (nob < 0) /* empty string */
775 if (nob == knl_buffer_nob) /* no space to terminate */
778 knl_buffer[nob + 1] = 0; /* terminate */
781 EXPORT_SYMBOL(cfs_trace_copyin_string);
783 int cfs_trace_copyout_string(char __user *usr_buffer, int usr_buffer_nob,
784 const char *knl_buffer, char *append)
786 /* NB if 'append' != NULL, it's a single character to append to the
787 * copied out string - usually "\n", for /proc entries and "" (i.e. a
788 * terminating zero byte) for sysctl entries */
789 int nob = strlen(knl_buffer);
791 if (nob > usr_buffer_nob)
792 nob = usr_buffer_nob;
794 if (copy_to_user(usr_buffer, knl_buffer, nob))
797 if (append != NULL && nob < usr_buffer_nob) {
798 if (copy_to_user(usr_buffer + nob, append, 1))
806 EXPORT_SYMBOL(cfs_trace_copyout_string);
808 int cfs_trace_allocate_string_buffer(char **str, int nob)
810 if (nob > 2 * PAGE_CACHE_SIZE) /* string must be "sensible" */
813 *str = kmalloc(nob, GFP_KERNEL | __GFP_ZERO);
820 void cfs_trace_free_string_buffer(char *str, int nob)
825 int cfs_trace_dump_debug_buffer_usrstr(void __user *usr_str, int usr_str_nob)
830 rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
834 rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
835 usr_str, usr_str_nob);
843 rc = cfs_tracefile_dump_all_pages(str);
845 cfs_trace_free_string_buffer(str, usr_str_nob + 1);
849 int cfs_trace_daemon_command(char *str)
853 cfs_tracefile_write_lock();
855 if (strcmp(str, "stop") == 0) {
856 cfs_tracefile_write_unlock();
857 cfs_trace_stop_thread();
858 cfs_tracefile_write_lock();
859 memset(cfs_tracefile, 0, sizeof(cfs_tracefile));
861 } else if (strncmp(str, "size=", 5) == 0) {
862 cfs_tracefile_size = simple_strtoul(str + 5, NULL, 0);
863 if (cfs_tracefile_size < 10 || cfs_tracefile_size > 20480)
864 cfs_tracefile_size = CFS_TRACEFILE_SIZE;
866 cfs_tracefile_size <<= 20;
868 } else if (strlen(str) >= sizeof(cfs_tracefile)) {
870 } else if (str[0] != '/') {
873 strcpy(cfs_tracefile, str);
876 "Lustre: debug daemon will attempt to start writing to %s (%lukB max)\n",
878 (long)(cfs_tracefile_size >> 10));
880 cfs_trace_start_thread();
883 cfs_tracefile_write_unlock();
887 int cfs_trace_daemon_command_usrstr(void __user *usr_str, int usr_str_nob)
892 rc = cfs_trace_allocate_string_buffer(&str, usr_str_nob + 1);
896 rc = cfs_trace_copyin_string(str, usr_str_nob + 1,
897 usr_str, usr_str_nob);
899 rc = cfs_trace_daemon_command(str);
901 cfs_trace_free_string_buffer(str, usr_str_nob + 1);
905 int cfs_trace_set_debug_mb(int mb)
910 int limit = cfs_trace_max_debug_mb();
911 struct cfs_trace_cpu_data *tcd;
913 if (mb < num_possible_cpus()) {
915 "Lustre: %d MB is too small for debug buffer size, setting it to %d MB.\n",
916 mb, num_possible_cpus());
917 mb = num_possible_cpus();
922 "Lustre: %d MB is too large for debug buffer size, setting it to %d MB.\n",
927 mb /= num_possible_cpus();
928 pages = mb << (20 - PAGE_CACHE_SHIFT);
930 cfs_tracefile_write_lock();
932 cfs_tcd_for_each(tcd, i, j)
933 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
935 cfs_tracefile_write_unlock();
940 int cfs_trace_get_debug_mb(void)
944 struct cfs_trace_cpu_data *tcd;
947 cfs_tracefile_read_lock();
949 cfs_tcd_for_each(tcd, i, j)
950 total_pages += tcd->tcd_max_pages;
952 cfs_tracefile_read_unlock();
954 return (total_pages >> (20 - PAGE_CACHE_SHIFT)) + 1;
957 static int tracefiled(void *arg)
959 struct page_collection pc;
960 struct tracefiled_ctl *tctl = arg;
961 struct cfs_trace_page *tage;
962 struct cfs_trace_page *tmp;
970 /* we're started late enough that we pick up init's fs context */
971 /* this is so broken in uml? what on earth is going on? */
973 spin_lock_init(&pc.pc_lock);
974 complete(&tctl->tctl_start);
979 pc.pc_want_daemon_pages = 0;
981 if (list_empty(&pc.pc_pages))
985 cfs_tracefile_read_lock();
986 if (cfs_tracefile[0] != 0) {
987 filp = filp_open(cfs_tracefile,
988 O_CREAT | O_RDWR | O_LARGEFILE,
993 printk(KERN_WARNING "couldn't open %s: %d\n",
997 cfs_tracefile_read_unlock();
999 put_pages_on_daemon_list(&pc);
1000 __LASSERT(list_empty(&pc.pc_pages));
1006 list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1008 static loff_t f_pos;
1010 __LASSERT_TAGE_INVARIANT(tage);
1012 if (f_pos >= (off_t)cfs_tracefile_size)
1014 else if (f_pos > i_size_read(file_inode(filp)))
1015 f_pos = i_size_read(file_inode(filp));
1017 buf = kmap(tage->page);
1018 rc = vfs_write(filp, (__force const char __user *)buf,
1019 tage->used, &f_pos);
1022 if (rc != (int)tage->used) {
1023 printk(KERN_WARNING "wanted to write %u but wrote %d\n",
1025 put_pages_back(&pc);
1026 __LASSERT(list_empty(&pc.pc_pages));
1032 filp_close(filp, NULL);
1033 put_pages_on_daemon_list(&pc);
1034 if (!list_empty(&pc.pc_pages)) {
1037 printk(KERN_ALERT "Lustre: trace pages aren't empty\n");
1038 pr_err("total cpus(%d): ",
1039 num_possible_cpus());
1040 for (i = 0; i < num_possible_cpus(); i++)
1042 pr_cont("%d(on) ", i);
1044 pr_cont("%d(off) ", i);
1048 list_for_each_entry_safe(tage, tmp, &pc.pc_pages,
1050 pr_err("page %d belongs to cpu %d\n",
1052 pr_err("There are %d pages unwritten\n", i);
1054 __LASSERT(list_empty(&pc.pc_pages));
1056 if (atomic_read(&tctl->tctl_shutdown)) {
1057 if (last_loop == 0) {
1064 init_waitqueue_entry(&__wait, current);
1065 add_wait_queue(&tctl->tctl_waitq, &__wait);
1066 set_current_state(TASK_INTERRUPTIBLE);
1067 schedule_timeout(cfs_time_seconds(1));
1068 remove_wait_queue(&tctl->tctl_waitq, &__wait);
1070 complete(&tctl->tctl_stop);
1074 int cfs_trace_start_thread(void)
1076 struct tracefiled_ctl *tctl = &trace_tctl;
1079 mutex_lock(&cfs_trace_thread_mutex);
1083 init_completion(&tctl->tctl_start);
1084 init_completion(&tctl->tctl_stop);
1085 init_waitqueue_head(&tctl->tctl_waitq);
1086 atomic_set(&tctl->tctl_shutdown, 0);
1088 if (IS_ERR(kthread_run(tracefiled, tctl, "ktracefiled"))) {
1093 wait_for_completion(&tctl->tctl_start);
1096 mutex_unlock(&cfs_trace_thread_mutex);
1100 void cfs_trace_stop_thread(void)
1102 struct tracefiled_ctl *tctl = &trace_tctl;
1104 mutex_lock(&cfs_trace_thread_mutex);
1105 if (thread_running) {
1107 "Lustre: shutting down debug daemon thread...\n");
1108 atomic_set(&tctl->tctl_shutdown, 1);
1109 wait_for_completion(&tctl->tctl_stop);
1112 mutex_unlock(&cfs_trace_thread_mutex);
1115 int cfs_tracefile_init(int max_pages)
1117 struct cfs_trace_cpu_data *tcd;
1123 rc = cfs_tracefile_init_arch();
1127 cfs_tcd_for_each(tcd, i, j) {
1128 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1129 factor = tcd->tcd_pages_factor;
1130 INIT_LIST_HEAD(&tcd->tcd_pages);
1131 INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1132 INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1133 tcd->tcd_cur_pages = 0;
1134 tcd->tcd_cur_stock_pages = 0;
1135 tcd->tcd_cur_daemon_pages = 0;
1136 tcd->tcd_max_pages = (max_pages * factor) / 100;
1137 LASSERT(tcd->tcd_max_pages > 0);
1138 tcd->tcd_shutting_down = 0;
1144 static void trace_cleanup_on_all_cpus(void)
1146 struct cfs_trace_cpu_data *tcd;
1147 struct cfs_trace_page *tage;
1148 struct cfs_trace_page *tmp;
1151 for_each_possible_cpu(cpu) {
1152 cfs_tcd_for_each_type_lock(tcd, i, cpu) {
1153 tcd->tcd_shutting_down = 1;
1155 list_for_each_entry_safe(tage, tmp, &tcd->tcd_pages,
1157 __LASSERT_TAGE_INVARIANT(tage);
1159 list_del(&tage->linkage);
1160 cfs_tage_free(tage);
1163 tcd->tcd_cur_pages = 0;
1168 static void cfs_trace_cleanup(void)
1170 struct page_collection pc;
1172 INIT_LIST_HEAD(&pc.pc_pages);
1173 spin_lock_init(&pc.pc_lock);
1175 trace_cleanup_on_all_cpus();
1177 cfs_tracefile_fini_arch();
1180 void cfs_tracefile_exit(void)
1182 cfs_trace_stop_thread();
1183 cfs_trace_cleanup();