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) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/llite/llite_capa.c
38 * Author: Lai Siyao <lsy@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_LLITE
44 #include <linux/version.h>
45 #include <asm/uaccess.h>
46 #include <linux/file.h>
47 #include <linux/kmod.h>
49 #include <lustre_lite.h>
50 #include "llite_internal.h"
52 /* for obd_capa.c_list, client capa might stay in three places:
55 * 3. stand alone: just allocated.
58 /* capas for oss writeback and those failed to renew */
59 static CFS_LIST_HEAD(ll_idle_capas);
60 static struct ptlrpc_thread ll_capa_thread;
61 static cfs_list_t *ll_capa_list = &capa_list[CAPA_SITE_CLIENT];
63 /* llite capa renewal timer */
64 struct timer_list ll_capa_timer;
65 /* for debug: indicate whether capa on llite is enabled or not */
66 static cfs_atomic_t ll_capa_debug = CFS_ATOMIC_INIT(0);
67 static unsigned long long ll_capa_renewed = 0;
68 static unsigned long long ll_capa_renewal_noent = 0;
69 static unsigned long long ll_capa_renewal_failed = 0;
70 static unsigned long long ll_capa_renewal_retries = 0;
72 static inline void update_capa_timer(struct obd_capa *ocapa, cfs_time_t expiry)
74 if (cfs_time_before(expiry, ll_capa_timer.expires) ||
75 !timer_pending(&ll_capa_timer)) {
76 mod_timer(&ll_capa_timer, expiry);
77 DEBUG_CAPA(D_SEC, &ocapa->c_capa,
78 "ll_capa_timer update: %lu/%lu by", expiry, jiffies);
82 static inline cfs_time_t capa_renewal_time(struct obd_capa *ocapa)
84 return cfs_time_sub(ocapa->c_expiry,
85 cfs_time_seconds(ocapa->c_capa.lc_timeout) / 2);
88 static inline int capa_is_to_expire(struct obd_capa *ocapa)
90 return cfs_time_beforeq(capa_renewal_time(ocapa), cfs_time_current());
93 static inline int have_expired_capa(void)
95 struct obd_capa *ocapa = NULL;
98 /* if ll_capa_list has client capa to expire or ll_idle_capas has
99 * expired capa, return 1.
101 spin_lock(&capa_lock);
102 if (!cfs_list_empty(ll_capa_list)) {
103 ocapa = cfs_list_entry(ll_capa_list->next, struct obd_capa,
105 expired = capa_is_to_expire(ocapa);
107 update_capa_timer(ocapa, capa_renewal_time(ocapa));
108 } else if (!cfs_list_empty(&ll_idle_capas)) {
109 ocapa = cfs_list_entry(ll_idle_capas.next, struct obd_capa,
111 expired = capa_is_expired(ocapa);
113 update_capa_timer(ocapa, ocapa->c_expiry);
115 spin_unlock(&capa_lock);
118 DEBUG_CAPA(D_SEC, &ocapa->c_capa, "expired");
122 static void sort_add_capa(struct obd_capa *ocapa, cfs_list_t *head)
124 struct obd_capa *tmp;
125 cfs_list_t *before = NULL;
127 /* TODO: client capa is sorted by expiry, this could be optimized */
128 cfs_list_for_each_entry_reverse(tmp, head, c_list) {
129 if (cfs_time_aftereq(ocapa->c_expiry, tmp->c_expiry)) {
130 before = &tmp->c_list;
135 LASSERT(&ocapa->c_list != before);
136 cfs_list_add(&ocapa->c_list, before ?: head);
139 static inline int obd_capa_open_count(struct obd_capa *oc)
141 struct ll_inode_info *lli = ll_i2info(oc->u.cli.inode);
142 return cfs_atomic_read(&lli->lli_open_count);
145 static void ll_delete_capa(struct obd_capa *ocapa)
147 struct ll_inode_info *lli = ll_i2info(ocapa->u.cli.inode);
149 if (capa_for_mds(&ocapa->c_capa)) {
150 LASSERT(lli->lli_mds_capa == ocapa);
151 lli->lli_mds_capa = NULL;
152 } else if (capa_for_oss(&ocapa->c_capa)) {
153 cfs_list_del_init(&ocapa->u.cli.lli_list);
156 DEBUG_CAPA(D_SEC, &ocapa->c_capa, "free client");
157 cfs_list_del_init(&ocapa->c_list);
158 capa_count[CAPA_SITE_CLIENT]--;
159 /* release the ref when alloc */
163 /* three places where client capa is deleted:
164 * 1. capa_thread_main(), main place to delete expired capa.
165 * 2. ll_clear_inode_capas() in ll_clear_inode().
166 * 3. ll_truncate_free_capa() delete truncate capa explicitly in ll_setattr_ost().
168 static int capa_thread_main(void *unused)
170 struct obd_capa *ocapa, *tmp, *next;
171 struct inode *inode = NULL;
172 struct l_wait_info lwi = { 0 };
176 cfs_daemonize("ll_capa");
178 thread_set_flags(&ll_capa_thread, SVC_RUNNING);
179 cfs_waitq_signal(&ll_capa_thread.t_ctl_waitq);
182 l_wait_event(ll_capa_thread.t_ctl_waitq,
183 !thread_is_running(&ll_capa_thread) ||
187 if (!thread_is_running(&ll_capa_thread))
192 spin_lock(&capa_lock);
193 cfs_list_for_each_entry_safe(ocapa, tmp, ll_capa_list, c_list) {
196 LASSERT(ocapa->c_capa.lc_opc != CAPA_OPC_OSS_TRUNC);
198 if (!capa_is_to_expire(ocapa)) {
203 cfs_list_del_init(&ocapa->c_list);
205 /* for MDS capability, only renew those which belong to
206 * dir, or its inode is opened, or client holds LOOKUP
209 /* ibits may be changed by ll_have_md_lock() so we have
210 * to set it each time */
211 ibits = MDS_INODELOCK_LOOKUP;
212 if (capa_for_mds(&ocapa->c_capa) &&
213 !S_ISDIR(ocapa->u.cli.inode->i_mode) &&
214 obd_capa_open_count(ocapa) == 0 &&
215 !ll_have_md_lock(ocapa->u.cli.inode,
216 &ibits, LCK_MINMODE)) {
217 DEBUG_CAPA(D_SEC, &ocapa->c_capa,
219 sort_add_capa(ocapa, &ll_idle_capas);
223 /* for OSS capability, only renew those whose inode is
226 if (capa_for_oss(&ocapa->c_capa) &&
227 obd_capa_open_count(ocapa) == 0) {
228 /* oss capa with open count == 0 won't renew,
229 * move to idle list */
230 sort_add_capa(ocapa, &ll_idle_capas);
234 /* NB iput() is in ll_update_capa() */
235 inode = igrab(ocapa->u.cli.inode);
237 DEBUG_CAPA(D_ERROR, &ocapa->c_capa,
244 spin_unlock(&capa_lock);
245 rc = md_renew_capa(ll_i2mdexp(inode), ocapa,
247 spin_lock(&capa_lock);
249 DEBUG_CAPA(D_ERROR, &ocapa->c_capa,
250 "renew failed: %d", rc);
251 ll_capa_renewal_failed++;
256 update_capa_timer(next, capa_renewal_time(next));
258 cfs_list_for_each_entry_safe(ocapa, tmp, &ll_idle_capas,
260 if (!capa_is_expired(ocapa)) {
262 update_capa_timer(ocapa,
267 if (cfs_atomic_read(&ocapa->c_refc) > 1) {
268 DEBUG_CAPA(D_SEC, &ocapa->c_capa,
269 "expired(c_refc %d), don't release",
270 cfs_atomic_read(&ocapa->c_refc));
271 /* don't try to renew any more */
272 cfs_list_del_init(&ocapa->c_list);
276 /* expired capa is released. */
277 DEBUG_CAPA(D_SEC, &ocapa->c_capa, "release expired");
278 ll_delete_capa(ocapa);
281 spin_unlock(&capa_lock);
284 thread_set_flags(&ll_capa_thread, SVC_STOPPED);
285 cfs_waitq_signal(&ll_capa_thread.t_ctl_waitq);
289 void ll_capa_timer_callback(unsigned long unused)
291 cfs_waitq_signal(&ll_capa_thread.t_ctl_waitq);
294 int ll_capa_thread_start(void)
299 cfs_waitq_init(&ll_capa_thread.t_ctl_waitq);
301 rc = cfs_create_thread(capa_thread_main, NULL, 0);
303 CERROR("cannot start expired capa thread: rc %d\n", rc);
306 cfs_wait_event(ll_capa_thread.t_ctl_waitq,
307 thread_is_running(&ll_capa_thread));
312 void ll_capa_thread_stop(void)
314 thread_set_flags(&ll_capa_thread, SVC_STOPPING);
315 cfs_waitq_signal(&ll_capa_thread.t_ctl_waitq);
316 cfs_wait_event(ll_capa_thread.t_ctl_waitq,
317 thread_is_stopped(&ll_capa_thread));
320 struct obd_capa *ll_osscapa_get(struct inode *inode, __u64 opc)
322 struct ll_inode_info *lli = ll_i2info(inode);
323 struct obd_capa *ocapa;
328 if ((ll_i2sbi(inode)->ll_flags & LL_SBI_OSS_CAPA) == 0)
331 LASSERT(opc == CAPA_OPC_OSS_WRITE || opc == CAPA_OPC_OSS_RW ||
332 opc == CAPA_OPC_OSS_TRUNC);
334 spin_lock(&capa_lock);
335 cfs_list_for_each_entry(ocapa, &lli->lli_oss_capas, u.cli.lli_list) {
336 if (capa_is_expired(ocapa))
338 if ((opc & CAPA_OPC_OSS_WRITE) &&
339 capa_opc_supported(&ocapa->c_capa, CAPA_OPC_OSS_WRITE)) {
342 } else if ((opc & CAPA_OPC_OSS_READ) &&
343 capa_opc_supported(&ocapa->c_capa,
344 CAPA_OPC_OSS_READ)) {
347 } else if ((opc & CAPA_OPC_OSS_TRUNC) &&
348 capa_opc_supported(&ocapa->c_capa, opc)) {
355 LASSERT(lu_fid_eq(capa_fid(&ocapa->c_capa),
356 ll_inode2fid(inode)));
357 LASSERT(ocapa->c_site == CAPA_SITE_CLIENT);
361 DEBUG_CAPA(D_SEC, &ocapa->c_capa, "found client");
365 if (cfs_atomic_read(&ll_capa_debug)) {
366 CERROR("no capability for "DFID" opc "LPX64"\n",
367 PFID(&lli->lli_fid), opc);
368 cfs_atomic_set(&ll_capa_debug, 0);
371 spin_unlock(&capa_lock);
375 EXPORT_SYMBOL(ll_osscapa_get);
377 struct obd_capa *ll_mdscapa_get(struct inode *inode)
379 struct ll_inode_info *lli = ll_i2info(inode);
380 struct obd_capa *ocapa;
383 LASSERT(inode != NULL);
385 if ((ll_i2sbi(inode)->ll_flags & LL_SBI_MDS_CAPA) == 0)
388 spin_lock(&capa_lock);
389 ocapa = capa_get(lli->lli_mds_capa);
390 spin_unlock(&capa_lock);
391 if (!ocapa && cfs_atomic_read(&ll_capa_debug)) {
392 CERROR("no mds capability for "DFID"\n", PFID(&lli->lli_fid));
393 cfs_atomic_set(&ll_capa_debug, 0);
399 static struct obd_capa *do_add_mds_capa(struct inode *inode,
400 struct obd_capa *ocapa)
402 struct ll_inode_info *lli = ll_i2info(inode);
403 struct obd_capa *old = lli->lli_mds_capa;
404 struct lustre_capa *capa = &ocapa->c_capa;
407 ocapa->u.cli.inode = inode;
408 lli->lli_mds_capa = ocapa;
409 capa_count[CAPA_SITE_CLIENT]++;
411 DEBUG_CAPA(D_SEC, capa, "add MDS");
413 spin_lock(&old->c_lock);
415 spin_unlock(&old->c_lock);
417 DEBUG_CAPA(D_SEC, capa, "update MDS");
425 static struct obd_capa *do_lookup_oss_capa(struct inode *inode, int opc)
427 struct ll_inode_info *lli = ll_i2info(inode);
428 struct obd_capa *ocapa;
430 /* inside capa_lock */
431 cfs_list_for_each_entry(ocapa, &lli->lli_oss_capas, u.cli.lli_list) {
432 if ((capa_opc(&ocapa->c_capa) & opc) != opc)
435 LASSERT(lu_fid_eq(capa_fid(&ocapa->c_capa),
436 ll_inode2fid(inode)));
437 LASSERT(ocapa->c_site == CAPA_SITE_CLIENT);
439 DEBUG_CAPA(D_SEC, &ocapa->c_capa, "found client");
446 static inline void inode_add_oss_capa(struct inode *inode,
447 struct obd_capa *ocapa)
449 struct ll_inode_info *lli = ll_i2info(inode);
450 struct obd_capa *tmp;
451 cfs_list_t *next = NULL;
453 /* capa is sorted in lli_oss_capas so lookup can always find the
455 cfs_list_for_each_entry(tmp, &lli->lli_oss_capas, u.cli.lli_list) {
456 if (cfs_time_after(ocapa->c_expiry, tmp->c_expiry)) {
457 next = &tmp->u.cli.lli_list;
461 LASSERT(&ocapa->u.cli.lli_list != next);
462 cfs_list_move_tail(&ocapa->u.cli.lli_list, next ?: &lli->lli_oss_capas);
465 static struct obd_capa *do_add_oss_capa(struct inode *inode,
466 struct obd_capa *ocapa)
468 struct obd_capa *old;
469 struct lustre_capa *capa = &ocapa->c_capa;
471 LASSERTF(S_ISREG(inode->i_mode),
472 "inode has oss capa, but not regular file, mode: %d\n",
475 /* FIXME: can't replace it so easily with fine-grained opc */
476 old = do_lookup_oss_capa(inode, capa_opc(capa) & CAPA_OPC_OSS_ONLY);
478 ocapa->u.cli.inode = inode;
479 CFS_INIT_LIST_HEAD(&ocapa->u.cli.lli_list);
480 capa_count[CAPA_SITE_CLIENT]++;
482 DEBUG_CAPA(D_SEC, capa, "add OSS");
484 spin_lock(&old->c_lock);
486 spin_unlock(&old->c_lock);
488 DEBUG_CAPA(D_SEC, capa, "update OSS");
494 inode_add_oss_capa(inode, ocapa);
498 struct obd_capa *ll_add_capa(struct inode *inode, struct obd_capa *ocapa)
500 spin_lock(&capa_lock);
501 ocapa = capa_for_mds(&ocapa->c_capa) ? do_add_mds_capa(inode, ocapa) :
502 do_add_oss_capa(inode, ocapa);
504 /* truncate capa won't renew */
505 if (ocapa->c_capa.lc_opc != CAPA_OPC_OSS_TRUNC) {
506 set_capa_expiry(ocapa);
507 cfs_list_del_init(&ocapa->c_list);
508 sort_add_capa(ocapa, ll_capa_list);
510 update_capa_timer(ocapa, capa_renewal_time(ocapa));
513 spin_unlock(&capa_lock);
515 cfs_atomic_set(&ll_capa_debug, 1);
519 static inline void delay_capa_renew(struct obd_capa *oc, cfs_time_t delay)
521 /* NB: set a fake expiry for this capa to prevent it renew too soon */
522 oc->c_expiry = cfs_time_add(oc->c_expiry, cfs_time_seconds(delay));
525 int ll_update_capa(struct obd_capa *ocapa, struct lustre_capa *capa)
527 struct inode *inode = ocapa->u.cli.inode;
536 spin_lock(&capa_lock);
538 DEBUG_CAPA(D_SEC, &ocapa->c_capa,
539 "renewal canceled because object removed");
540 ll_capa_renewal_noent++;
542 ll_capa_renewal_failed++;
544 /* failed capa won't be renewed any longer, but if -EIO,
545 * client might be doing recovery, retry in 2 min. */
546 if (rc == -EIO && !capa_is_expired(ocapa)) {
547 delay_capa_renew(ocapa, 120);
548 DEBUG_CAPA(D_ERROR, &ocapa->c_capa,
549 "renewal failed: -EIO, "
551 ll_capa_renewal_retries++;
554 DEBUG_CAPA(D_ERROR, &ocapa->c_capa,
555 "renewal failed(rc: %d) for", rc);
559 cfs_list_del_init(&ocapa->c_list);
560 sort_add_capa(ocapa, &ll_idle_capas);
561 spin_unlock(&capa_lock);
568 spin_lock(&ocapa->c_lock);
569 LASSERT(!memcmp(&ocapa->c_capa, capa,
570 offsetof(struct lustre_capa, lc_opc)));
571 ocapa->c_capa = *capa;
572 set_capa_expiry(ocapa);
573 spin_unlock(&ocapa->c_lock);
575 spin_lock(&capa_lock);
576 if (capa_for_oss(capa))
577 inode_add_oss_capa(inode, ocapa);
578 DEBUG_CAPA(D_SEC, capa, "renew");
581 cfs_list_del_init(&ocapa->c_list);
582 sort_add_capa(ocapa, ll_capa_list);
583 update_capa_timer(ocapa, capa_renewal_time(ocapa));
584 spin_unlock(&capa_lock);
591 void ll_capa_open(struct inode *inode)
593 struct ll_inode_info *lli = ll_i2info(inode);
595 if ((ll_i2sbi(inode)->ll_flags & (LL_SBI_MDS_CAPA | LL_SBI_OSS_CAPA))
599 if (!S_ISREG(inode->i_mode))
602 cfs_atomic_inc(&lli->lli_open_count);
605 void ll_capa_close(struct inode *inode)
607 struct ll_inode_info *lli = ll_i2info(inode);
609 if ((ll_i2sbi(inode)->ll_flags & (LL_SBI_MDS_CAPA | LL_SBI_OSS_CAPA))
613 if (!S_ISREG(inode->i_mode))
616 cfs_atomic_dec(&lli->lli_open_count);
619 /* delete CAPA_OPC_OSS_TRUNC only */
620 void ll_truncate_free_capa(struct obd_capa *ocapa)
625 LASSERT(ocapa->c_capa.lc_opc & CAPA_OPC_OSS_TRUNC);
626 DEBUG_CAPA(D_SEC, &ocapa->c_capa, "free truncate");
628 /* release ref when find */
630 if (likely(ocapa->c_capa.lc_opc == CAPA_OPC_OSS_TRUNC)) {
631 spin_lock(&capa_lock);
632 ll_delete_capa(ocapa);
633 spin_unlock(&capa_lock);
637 void ll_clear_inode_capas(struct inode *inode)
639 struct ll_inode_info *lli = ll_i2info(inode);
640 struct obd_capa *ocapa, *tmp;
642 spin_lock(&capa_lock);
643 ocapa = lli->lli_mds_capa;
645 ll_delete_capa(ocapa);
647 cfs_list_for_each_entry_safe(ocapa, tmp, &lli->lli_oss_capas,
649 ll_delete_capa(ocapa);
650 spin_unlock(&capa_lock);
653 void ll_print_capa_stat(struct ll_sb_info *sbi)
655 if (sbi->ll_flags & (LL_SBI_MDS_CAPA | LL_SBI_OSS_CAPA))
656 LCONSOLE_INFO("Fid capabilities renewed: %llu\n"
657 "Fid capabilities renewal ENOENT: %llu\n"
658 "Fid capabilities failed to renew: %llu\n"
659 "Fid capabilities renewal retries: %llu\n",
660 ll_capa_renewed, ll_capa_renewal_noent,
661 ll_capa_renewal_failed, ll_capa_renewal_retries);