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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 #include <linux/sched.h>
39 #include <linux/quotaops.h>
41 #define DEBUG_SUBSYSTEM S_LLITE
43 #include <obd_support.h>
44 #include <lustre_lite.h>
45 #include <lustre/lustre_idl.h>
46 #include <lustre_dlm.h>
48 #include "llite_internal.h"
50 static void free_dentry_data(struct rcu_head *head)
52 struct ll_dentry_data *lld;
54 lld = container_of(head, struct ll_dentry_data, lld_rcu_head);
58 /* should NOT be called with the dcache lock, see fs/dcache.c */
59 static void ll_release(struct dentry *de)
61 struct ll_dentry_data *lld;
65 if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
69 ll_intent_release(lld->lld_it);
70 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
72 LASSERT(lld->lld_cwd_count == 0);
73 LASSERT(lld->lld_mnt_count == 0);
75 call_rcu(&lld->lld_rcu_head, free_dentry_data);
80 /* Compare if two dentries are the same. Don't match if the existing dentry
81 * is marked invalid. Returns 1 if different, 0 if the same.
83 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
84 * an AST before calling d_revalidate_it(). The dentry still exists (marked
85 * INVALID) so d_lookup() matches it, but we have no lock on it (so
86 * lock_match() fails) and we spin around real_lookup(). */
87 #ifdef HAVE_D_COMPARE_7ARGS
88 int ll_dcompare(const struct dentry *parent, const struct inode *pinode,
89 const struct dentry *dentry, const struct inode *inode,
90 unsigned int len, const char *str, const struct qstr *name)
92 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
95 #ifdef HAVE_D_COMPARE_7ARGS
101 if (memcmp(str, name->name, len))
104 struct dentry *dentry;
107 if (d_name->len != name->len)
110 if (memcmp(d_name->name, name->name, name->len))
113 /* XXX: d_name must be in-dentry structure */
114 dentry = container_of(d_name, struct dentry, d_name); /* ugh */
117 CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n",
118 name->len, name->name, dentry, dentry->d_flags,
121 /* mountpoint is always valid */
122 if (d_mountpoint((struct dentry *)dentry))
125 if (d_lustre_invalid(dentry))
131 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
134 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
135 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
136 return LDLM_ITER_CONTINUE;
137 return LDLM_ITER_STOP;
140 /* find any ldlm lock of the inode in mdc and lov
144 static int find_cbdata(struct inode *inode)
146 struct ll_sb_info *sbi = ll_i2sbi(inode);
147 struct lov_stripe_md *lsm;
152 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
153 return_if_equal, NULL);
157 lsm = ccc_inode_lsm_get(inode);
161 rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL);
162 ccc_inode_lsm_put(inode, lsm);
168 * Called when last reference to a dentry is dropped and dcache wants to know
169 * whether or not it should cache it:
170 * - return 1 to delete the dentry immediately
171 * - return 0 to cache the dentry
172 * Should NOT be called with the dcache lock, see fs/dcache.c
174 static int ll_ddelete(HAVE_D_DELETE_CONST struct dentry *de)
179 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
180 d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
181 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
182 d_unhashed((struct dentry *)de) ? "" : "hashed,",
183 list_empty(&de->d_subdirs) ? "" : "subdirs");
185 #ifdef HAVE_DCACHE_LOCK
186 LASSERT(d_refcount(de) == 0);
188 /* kernel >= 2.6.38 last refcount is decreased after this function. */
189 LASSERT(d_refcount(de) == 1);
192 /* Disable this piece of code temproarily because this is called
193 * inside dcache_lock so it's not appropriate to do lots of work
194 * here. ATTENTION: Before this piece of code enabling, LU-2487 must be
197 /* if not ldlm lock for this inode, set i_nlink to 0 so that
198 * this inode can be recycled later b=20433 */
199 if (de->d_inode && !find_cbdata(de->d_inode))
200 clear_nlink(de->d_inode);
203 if (d_lustre_invalid((struct dentry *)de))
208 static int ll_set_dd(struct dentry *de)
213 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
214 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
217 if (de->d_fsdata == NULL) {
218 struct ll_dentry_data *lld;
221 if (likely(lld != NULL)) {
222 spin_lock(&de->d_lock);
223 if (likely(de->d_fsdata == NULL))
227 spin_unlock(&de->d_lock);
236 int ll_dops_init(struct dentry *de, int block, int init_sa)
238 struct ll_dentry_data *lld = ll_d2d(de);
241 if (lld == NULL && block != 0) {
249 if (lld != NULL && init_sa != 0)
250 lld->lld_sa_generation = 0;
252 #ifdef HAVE_DCACHE_LOCK
253 de->d_op = &ll_d_ops;
255 /* kernel >= 2.6.38 d_op is set in d_alloc() */
256 LASSERT(de->d_op == &ll_d_ops);
261 void ll_intent_drop_lock(struct lookup_intent *it)
263 if (it->it_op && it->d.lustre.it_lock_mode) {
264 struct lustre_handle handle;
266 handle.cookie = it->d.lustre.it_lock_handle;
268 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
269 " from it %p\n", handle.cookie, it);
270 ldlm_lock_decref(&handle, it->d.lustre.it_lock_mode);
272 /* bug 494: intent_release may be called multiple times, from
273 * this thread and we don't want to double-decref this lock */
274 it->d.lustre.it_lock_mode = 0;
275 if (it->d.lustre.it_remote_lock_mode != 0) {
276 handle.cookie = it->d.lustre.it_remote_lock_handle;
278 CDEBUG(D_DLMTRACE, "releasing remote lock with cookie"
279 LPX64" from it %p\n", handle.cookie, it);
280 ldlm_lock_decref(&handle,
281 it->d.lustre.it_remote_lock_mode);
282 it->d.lustre.it_remote_lock_mode = 0;
287 void ll_intent_release(struct lookup_intent *it)
291 CDEBUG(D_INFO, "intent %p released\n", it);
292 ll_intent_drop_lock(it);
293 /* We are still holding extra reference on a request, need to free it */
294 if (it_disposition(it, DISP_ENQ_OPEN_REF))
295 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
296 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
297 ptlrpc_req_finished(it->d.lustre.it_data);
298 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
300 ptlrpc_req_finished(it->d.lustre.it_data);
302 it->d.lustre.it_disposition = 0;
303 it->d.lustre.it_data = NULL;
307 void ll_invalidate_aliases(struct inode *inode)
309 struct dentry *dentry;
310 struct ll_d_hlist_node *p;
313 LASSERT(inode != NULL);
315 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
316 inode->i_ino, inode->i_generation, inode);
318 ll_lock_dcache(inode);
319 ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) {
320 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
321 "inode %p flags %d\n", dentry->d_name.len,
322 dentry->d_name.name, dentry, dentry->d_parent,
323 dentry->d_inode, dentry->d_flags);
325 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
326 CERROR("called on root (?) dentry=%p, inode=%p "
327 "ino=%lu\n", dentry, inode, inode->i_ino);
328 lustre_dump_dentry(dentry, 1);
329 libcfs_debug_dumpstack(NULL);
332 d_lustre_invalidate(dentry, 0);
334 ll_unlock_dcache(inode);
339 int ll_revalidate_it_finish(struct ptlrpc_request *request,
340 struct lookup_intent *it,
349 if (it_disposition(it, DISP_LOOKUP_NEG))
352 rc = ll_prep_inode(&de->d_inode, request, NULL, it);
357 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
360 LASSERT(dentry != NULL);
362 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
363 struct inode *inode = dentry->d_inode;
364 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
366 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
367 inode, inode->i_ino, inode->i_generation);
368 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
371 /* drop lookup or getattr locks immediately */
372 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
373 /* on 2.6 there are situation when several lookups and
374 * revalidations may be requested during single operation.
375 * therefore, we don't release intent here -bzzz */
376 ll_intent_drop_lock(it);
380 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
382 struct lookup_intent *it = *itp;
384 if (!it || it->it_op == IT_GETXATTR)
389 int ll_revalidate_it(struct dentry *de, int lookup_flags,
390 struct lookup_intent *it)
392 struct md_op_data *op_data;
393 struct ptlrpc_request *req = NULL;
394 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
395 struct obd_export *exp;
396 struct inode *parent = de->d_parent->d_inode;
400 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
403 if (de->d_inode == NULL) {
406 /* We can only use negative dentries if this is stat or lookup,
407 for opens and stuff we do need to query server. */
408 /* If there is IT_CREAT in intent op set, then we must throw
409 away this negative dentry and actually do the request to
410 kernel to create whatever needs to be created (if possible)*/
411 if (it && (it->it_op & IT_CREAT))
414 if (d_lustre_invalid(de))
417 ibits = MDS_INODELOCK_UPDATE;
418 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
422 /* Never execute intents for mount points.
423 * Attributes will be fixed up in ll_inode_revalidate_it */
424 if (d_mountpoint(de))
425 GOTO(out_sa, rc = 1);
427 /* need to get attributes in case root got changed from other client */
428 if (de == de->d_sb->s_root) {
429 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
435 exp = ll_i2mdexp(de->d_inode);
437 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
438 ll_frob_intent(&it, &lookup_it);
441 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
444 if (it->it_op == IT_OPEN) {
445 struct inode *inode = de->d_inode;
446 struct ll_inode_info *lli = ll_i2info(inode);
447 struct obd_client_handle **och_p;
452 * We used to check for MDS_INODELOCK_OPEN here, but in fact
453 * just having LOOKUP lock is enough to justify inode is the
454 * same. And if inode is the same and we have suitable
455 * openhandle, then there is no point in doing another OPEN RPC
456 * just to throw away newly received openhandle. There are no
457 * security implications too, if file owner or access mode is
458 * change, LOOKUP lock is revoked.
462 if (it->it_flags & FMODE_WRITE) {
463 och_p = &lli->lli_mds_write_och;
464 och_usecount = &lli->lli_open_fd_write_count;
465 } else if (it->it_flags & FMODE_EXEC) {
466 och_p = &lli->lli_mds_exec_och;
467 och_usecount = &lli->lli_open_fd_exec_count;
469 och_p = &lli->lli_mds_read_och;
470 och_usecount = &lli->lli_open_fd_read_count;
472 /* Check for the proper lock. */
473 ibits = MDS_INODELOCK_LOOKUP;
474 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
476 mutex_lock(&lli->lli_och_mutex);
477 if (*och_p) { /* Everything is open already, do nothing */
478 /*(*och_usecount)++; Do not let them steal our open
479 handle from under us */
480 SET_BUT_UNUSED(och_usecount);
481 /* XXX The code above was my original idea, but in case
482 we have the handle, but we cannot use it due to later
483 checks (e.g. O_CREAT|O_EXCL flags set), nobody
484 would decrement counter increased here. So we just
485 hope the lock won't be invalidated in between. But
486 if it would be, we'll reopen the open request to
487 MDS later during file open path */
488 mutex_unlock(&lli->lli_och_mutex);
491 mutex_unlock(&lli->lli_och_mutex);
495 if (it->it_op == IT_GETATTR) {
496 rc = ll_statahead_enter(parent, &de, 0);
499 else if (rc != -EAGAIN && rc != 0)
504 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
505 de->d_name.name, de->d_name.len,
506 0, LUSTRE_OPC_ANY, NULL);
508 RETURN(PTR_ERR(op_data));
510 if (!IS_POSIXACL(parent) || !exp_connect_umask(exp))
511 it->it_create_mode &= ~cfs_curproc_umask();
512 it->it_create_mode |= M_CHECK_STALE;
513 rc = md_intent_lock(exp, op_data, NULL, 0, it,
515 &req, ll_md_blocking_ast, 0);
516 it->it_create_mode &= ~M_CHECK_STALE;
517 ll_finish_md_op_data(op_data);
519 /* If req is NULL, then md_intent_lock only tried to do a lock match;
520 * if all was well, it will return 1 if it found locks, 0 otherwise. */
521 if (req == NULL && rc >= 0) {
529 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
530 "%d\n", rc, it->d.lustre.it_status);
536 rc = ll_revalidate_it_finish(req, it, de);
538 if (rc != -ESTALE && rc != -ENOENT)
539 ll_intent_release(it);
543 if ((it->it_op & IT_OPEN) && de->d_inode &&
544 !S_ISREG(de->d_inode->i_mode) &&
545 !S_ISDIR(de->d_inode->i_mode)) {
546 ll_release_openhandle(de, it);
551 /* We do not free request as it may be reused during following lookup
552 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
553 * be freed in ll_lookup_it or in ll_intent_release. But if
554 * request was not completed, we need to free it. (bug 5154, 9903) */
555 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
556 ptlrpc_req_finished(req);
558 /* mdt may grant layout lock for the newly created file, so
559 * release the lock to avoid leaking */
560 ll_intent_drop_lock(it);
561 ll_invalidate_aliases(de->d_inode);
564 __u64 matched_bits = 0;
566 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
567 "inode %p refc %d\n", de->d_name.len,
568 de->d_name.name, de, de->d_parent, de->d_inode,
571 ll_set_lock_data(exp, de->d_inode, it, &bits);
573 /* Note: We have to match both LOOKUP and PERM lock
574 * here to make sure the dentry is valid and no one
575 * changing the permission.
576 * But if the client connects < 2.4 server, which will
577 * only grant LOOKUP lock, so we can only Match LOOKUP
578 * lock for old server */
579 if (exp_connect_flags(ll_i2mdexp(de->d_inode)) &&
580 OBD_CONNECT_LVB_TYPE)
582 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
584 matched_bits = MDS_INODELOCK_LOOKUP;
586 if (((bits & matched_bits) == matched_bits) &&
587 d_lustre_invalid(de))
588 d_lustre_revalidate(de);
589 ll_lookup_finish_locks(it, de);
593 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
594 ll_statahead_mark(parent, de);
598 * This part is here to combat evil-evil race in real_lookup on 2.6
599 * kernels. The race details are: We enter do_lookup() looking for some
600 * name, there is nothing in dcache for this name yet and d_lookup()
601 * returns NULL. We proceed to real_lookup(), and while we do this,
602 * another process does open on the same file we looking up (most simple
603 * reproducer), open succeeds and the dentry is added. Now back to
604 * us. In real_lookup() we do d_lookup() again and suddenly find the
605 * dentry, so we call d_revalidate on it, but there is no lock, so
606 * without this code we would return 0, but unpatched real_lookup just
607 * returns -ENOENT in such a case instead of retrying the lookup. Once
608 * this is dealt with in real_lookup(), all of this ugly mess can go and
609 * we can just check locks in ->d_revalidate without doing any RPCs
613 if (it != &lookup_it) {
614 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
615 if (it->it_op == IT_GETATTR)
616 lookup_it.it_op = IT_GETATTR;
617 ll_lookup_finish_locks(it, de);
621 /* Do real lookup here. */
622 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
623 de->d_name.len, 0, (it->it_op & IT_CREAT ?
625 LUSTRE_OPC_ANY), NULL);
627 RETURN(PTR_ERR(op_data));
629 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
630 ll_md_blocking_ast, 0);
632 struct mdt_body *mdt_body;
633 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
634 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
637 fid = *ll_inode2fid(de->d_inode);
639 /* see if we got same inode, if not - return error */
640 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
641 ll_finish_md_op_data(op_data);
643 goto revalidate_finish;
645 ll_intent_release(it);
647 ll_finish_md_op_data(op_data);
652 * For rc == 1 case, should not return directly to prevent losing
653 * statahead windows; for rc == 0 case, the "lookup" will be done later.
655 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
656 ll_statahead_enter(parent, &de, 1);
660 #ifdef HAVE_IOP_ATOMIC_OPEN
662 * Always trust cached dentries. Update statahead window if necessary.
664 int ll_revalidate_nd(struct dentry *dentry, unsigned int flags)
666 struct inode *parent = dentry->d_parent->d_inode;
670 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,flags=%u\n",
671 dentry->d_name.name, flags);
673 if (!(flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE)) &&
674 ll_need_statahead(parent, dentry) > 0) {
675 if (flags & LOOKUP_RCU)
678 if (dentry->d_inode == NULL)
680 do_statahead_enter(parent, &dentry, unplug);
681 ll_statahead_mark(parent, dentry);
687 #else /* !HAVE_IOP_ATOMIC_OPEN */
688 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
693 #ifndef HAVE_DCACHE_LOCK
694 /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */
695 if (nd->flags & LOOKUP_RCU)
699 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
700 struct lookup_intent *it;
702 it = ll_convert_intent(&nd->intent.open, nd->flags);
706 if (it->it_op == (IT_OPEN|IT_CREAT) &&
707 nd->intent.open.flags & O_EXCL) {
708 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
713 rc = ll_revalidate_it(dentry, nd->flags, it);
715 if (rc && (nd->flags & LOOKUP_OPEN) &&
716 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
717 // XXX Code duplication with ll_lookup_nd
718 if (S_ISFIFO(dentry->d_inode->i_mode)) {
719 // We cannot call open here as it would
722 (struct ptlrpc_request *)
723 it->d.lustre.it_data);
727 nd->intent.open.file->private_data = it;
728 filp = lookup_instantiate_filp(nd, dentry,NULL);
733 if (!rc && (nd->flags & LOOKUP_CREATE) &&
734 it_disposition(it, DISP_OPEN_CREATE)) {
735 /* We created something but we may only return
736 * negative dentry here, so save request in dentry,
737 * if lookup will be called later on, it will
738 * pick the request, otherwise it would be freed
740 ll_d2d(dentry)->lld_it = it;
741 it = NULL; /* avoid freeing */
746 ll_intent_release(it);
747 OBD_FREE(it, sizeof(*it));
750 rc = ll_revalidate_it(dentry, 0, NULL);
755 #endif /* HAVE_IOP_ATOMIC_OPEN */
757 void ll_d_iput(struct dentry *de, struct inode *inode)
760 if (!find_cbdata(inode))
765 struct dentry_operations ll_d_ops = {
766 .d_revalidate = ll_revalidate_nd,
767 .d_release = ll_release,
768 .d_delete = ll_ddelete,
770 .d_compare = ll_dcompare,