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, 2012, 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
196 /* if not ldlm lock for this inode, set i_nlink to 0 so that
197 * this inode can be recycled later b=20433 */
198 if (de->d_inode && !find_cbdata(de->d_inode))
199 clear_nlink(de->d_inode);
202 if (d_lustre_invalid((struct dentry *)de))
207 static int ll_set_dd(struct dentry *de)
212 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
213 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
216 if (de->d_fsdata == NULL) {
217 struct ll_dentry_data *lld;
220 if (likely(lld != NULL)) {
221 spin_lock(&de->d_lock);
222 if (likely(de->d_fsdata == NULL))
226 spin_unlock(&de->d_lock);
235 int ll_dops_init(struct dentry *de, int block, int init_sa)
237 struct ll_dentry_data *lld = ll_d2d(de);
240 if (lld == NULL && block != 0) {
248 if (lld != NULL && init_sa != 0)
249 lld->lld_sa_generation = 0;
251 #ifdef HAVE_DCACHE_LOCK
252 de->d_op = &ll_d_ops;
254 /* kernel >= 2.6.38 d_op is set in d_alloc() */
255 LASSERT(de->d_op == &ll_d_ops);
260 void ll_intent_drop_lock(struct lookup_intent *it)
262 if (it->it_op && it->d.lustre.it_lock_mode) {
263 struct lustre_handle handle;
265 handle.cookie = it->d.lustre.it_lock_handle;
267 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
268 " from it %p\n", handle.cookie, it);
269 ldlm_lock_decref(&handle, it->d.lustre.it_lock_mode);
271 /* bug 494: intent_release may be called multiple times, from
272 * this thread and we don't want to double-decref this lock */
273 it->d.lustre.it_lock_mode = 0;
277 void ll_intent_release(struct lookup_intent *it)
281 CDEBUG(D_INFO, "intent %p released\n", it);
282 ll_intent_drop_lock(it);
283 /* We are still holding extra reference on a request, need to free it */
284 if (it_disposition(it, DISP_ENQ_OPEN_REF))
285 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
286 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
287 ptlrpc_req_finished(it->d.lustre.it_data);
288 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
290 ptlrpc_req_finished(it->d.lustre.it_data);
292 it->d.lustre.it_disposition = 0;
293 it->d.lustre.it_data = NULL;
297 void ll_invalidate_aliases(struct inode *inode)
299 struct dentry *dentry;
300 struct ll_d_hlist_node *p;
303 LASSERT(inode != NULL);
305 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
306 inode->i_ino, inode->i_generation, inode);
308 ll_lock_dcache(inode);
309 ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) {
310 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
311 "inode %p flags %d\n", dentry->d_name.len,
312 dentry->d_name.name, dentry, dentry->d_parent,
313 dentry->d_inode, dentry->d_flags);
315 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
316 CERROR("called on root (?) dentry=%p, inode=%p "
317 "ino=%lu\n", dentry, inode, inode->i_ino);
318 lustre_dump_dentry(dentry, 1);
319 libcfs_debug_dumpstack(NULL);
322 d_lustre_invalidate(dentry);
324 ll_unlock_dcache(inode);
329 int ll_revalidate_it_finish(struct ptlrpc_request *request,
330 struct lookup_intent *it,
339 if (it_disposition(it, DISP_LOOKUP_NEG))
342 rc = ll_prep_inode(&de->d_inode, request, NULL, it);
347 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
350 LASSERT(dentry != NULL);
352 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
353 struct inode *inode = dentry->d_inode;
354 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
356 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
357 inode, inode->i_ino, inode->i_generation);
358 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
361 /* drop lookup or getattr locks immediately */
362 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
363 /* on 2.6 there are situation when several lookups and
364 * revalidations may be requested during single operation.
365 * therefore, we don't release intent here -bzzz */
366 ll_intent_drop_lock(it);
370 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
372 struct lookup_intent *it = *itp;
374 if (!it || it->it_op == IT_GETXATTR)
379 int ll_revalidate_it(struct dentry *de, int lookup_flags,
380 struct lookup_intent *it)
382 struct md_op_data *op_data;
383 struct ptlrpc_request *req = NULL;
384 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
385 struct obd_export *exp;
386 struct inode *parent = de->d_parent->d_inode;
390 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
393 if (de->d_inode == NULL) {
396 /* We can only use negative dentries if this is stat or lookup,
397 for opens and stuff we do need to query server. */
398 /* If there is IT_CREAT in intent op set, then we must throw
399 away this negative dentry and actually do the request to
400 kernel to create whatever needs to be created (if possible)*/
401 if (it && (it->it_op & IT_CREAT))
404 if (d_lustre_invalid(de))
407 ibits = MDS_INODELOCK_UPDATE;
408 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
412 /* Never execute intents for mount points.
413 * Attributes will be fixed up in ll_inode_revalidate_it */
414 if (d_mountpoint(de))
415 GOTO(out_sa, rc = 1);
417 /* need to get attributes in case root got changed from other client */
418 if (de == de->d_sb->s_root) {
419 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
425 exp = ll_i2mdexp(de->d_inode);
427 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
428 ll_frob_intent(&it, &lookup_it);
431 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
434 if (it->it_op == IT_OPEN) {
435 struct inode *inode = de->d_inode;
436 struct ll_inode_info *lli = ll_i2info(inode);
437 struct obd_client_handle **och_p;
442 * We used to check for MDS_INODELOCK_OPEN here, but in fact
443 * just having LOOKUP lock is enough to justify inode is the
444 * same. And if inode is the same and we have suitable
445 * openhandle, then there is no point in doing another OPEN RPC
446 * just to throw away newly received openhandle. There are no
447 * security implications too, if file owner or access mode is
448 * change, LOOKUP lock is revoked.
452 if (it->it_flags & FMODE_WRITE) {
453 och_p = &lli->lli_mds_write_och;
454 och_usecount = &lli->lli_open_fd_write_count;
455 } else if (it->it_flags & FMODE_EXEC) {
456 och_p = &lli->lli_mds_exec_och;
457 och_usecount = &lli->lli_open_fd_exec_count;
459 och_p = &lli->lli_mds_read_och;
460 och_usecount = &lli->lli_open_fd_read_count;
462 /* Check for the proper lock. */
463 ibits = MDS_INODELOCK_LOOKUP;
464 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
466 mutex_lock(&lli->lli_och_mutex);
467 if (*och_p) { /* Everything is open already, do nothing */
468 /*(*och_usecount)++; Do not let them steal our open
469 handle from under us */
470 SET_BUT_UNUSED(och_usecount);
471 /* XXX The code above was my original idea, but in case
472 we have the handle, but we cannot use it due to later
473 checks (e.g. O_CREAT|O_EXCL flags set), nobody
474 would decrement counter increased here. So we just
475 hope the lock won't be invalidated in between. But
476 if it would be, we'll reopen the open request to
477 MDS later during file open path */
478 mutex_unlock(&lli->lli_och_mutex);
481 mutex_unlock(&lli->lli_och_mutex);
485 if (it->it_op == IT_GETATTR) {
486 rc = ll_statahead_enter(parent, &de, 0);
489 else if (rc != -EAGAIN && rc != 0)
494 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
495 de->d_name.name, de->d_name.len,
496 0, LUSTRE_OPC_ANY, NULL);
498 RETURN(PTR_ERR(op_data));
500 if (!IS_POSIXACL(parent) || !exp_connect_umask(exp))
501 it->it_create_mode &= ~cfs_curproc_umask();
502 it->it_create_mode |= M_CHECK_STALE;
503 rc = md_intent_lock(exp, op_data, NULL, 0, it,
505 &req, ll_md_blocking_ast, 0);
506 it->it_create_mode &= ~M_CHECK_STALE;
507 ll_finish_md_op_data(op_data);
509 /* If req is NULL, then md_intent_lock only tried to do a lock match;
510 * if all was well, it will return 1 if it found locks, 0 otherwise. */
511 if (req == NULL && rc >= 0) {
519 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
520 "%d\n", rc, it->d.lustre.it_status);
526 rc = ll_revalidate_it_finish(req, it, de);
528 if (rc != -ESTALE && rc != -ENOENT)
529 ll_intent_release(it);
533 if ((it->it_op & IT_OPEN) && de->d_inode &&
534 !S_ISREG(de->d_inode->i_mode) &&
535 !S_ISDIR(de->d_inode->i_mode)) {
536 ll_release_openhandle(de, it);
541 /* We do not free request as it may be reused during following lookup
542 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
543 * be freed in ll_lookup_it or in ll_intent_release. But if
544 * request was not completed, we need to free it. (bug 5154, 9903) */
545 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
546 ptlrpc_req_finished(req);
548 /* mdt may grant layout lock for the newly created file, so
549 * release the lock to avoid leaking */
550 ll_intent_drop_lock(it);
551 ll_invalidate_aliases(de->d_inode);
555 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
556 "inode %p refc %d\n", de->d_name.len,
557 de->d_name.name, de, de->d_parent, de->d_inode,
559 ll_set_lock_data(exp, de->d_inode, it, &bits);
560 if ((bits & MDS_INODELOCK_LOOKUP) && d_lustre_invalid(de))
561 d_lustre_revalidate(de);
562 ll_lookup_finish_locks(it, de);
566 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
567 ll_statahead_mark(parent, de);
571 * This part is here to combat evil-evil race in real_lookup on 2.6
572 * kernels. The race details are: We enter do_lookup() looking for some
573 * name, there is nothing in dcache for this name yet and d_lookup()
574 * returns NULL. We proceed to real_lookup(), and while we do this,
575 * another process does open on the same file we looking up (most simple
576 * reproducer), open succeeds and the dentry is added. Now back to
577 * us. In real_lookup() we do d_lookup() again and suddenly find the
578 * dentry, so we call d_revalidate on it, but there is no lock, so
579 * without this code we would return 0, but unpatched real_lookup just
580 * returns -ENOENT in such a case instead of retrying the lookup. Once
581 * this is dealt with in real_lookup(), all of this ugly mess can go and
582 * we can just check locks in ->d_revalidate without doing any RPCs
586 if (it != &lookup_it) {
587 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
588 if (it->it_op == IT_GETATTR)
589 lookup_it.it_op = IT_GETATTR;
590 ll_lookup_finish_locks(it, de);
594 /* Do real lookup here. */
595 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
596 de->d_name.len, 0, (it->it_op & IT_CREAT ?
598 LUSTRE_OPC_ANY), NULL);
600 RETURN(PTR_ERR(op_data));
602 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
603 ll_md_blocking_ast, 0);
605 struct mdt_body *mdt_body;
606 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
607 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
610 fid = *ll_inode2fid(de->d_inode);
612 /* see if we got same inode, if not - return error */
613 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
614 ll_finish_md_op_data(op_data);
616 goto revalidate_finish;
618 ll_intent_release(it);
620 ll_finish_md_op_data(op_data);
625 * For rc == 1 case, should not return directly to prevent losing
626 * statahead windows; for rc == 0 case, the "lookup" will be done later.
628 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
629 ll_statahead_enter(parent, &de, 1);
633 #ifdef HAVE_IOP_ATOMIC_OPEN
635 * Always trust cached dentries. Update statahead window if necessary.
637 int ll_revalidate_nd(struct dentry *dentry, unsigned int flags)
639 struct inode *parent = dentry->d_parent->d_inode;
643 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,flags=%u\n",
644 dentry->d_name.name, flags);
646 if (!(flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE)) &&
647 ll_need_statahead(parent, dentry) > 0) {
648 if (flags & LOOKUP_RCU)
651 if (dentry->d_inode == NULL)
653 do_statahead_enter(parent, &dentry, unplug);
654 ll_statahead_mark(parent, dentry);
660 #else /* !HAVE_IOP_ATOMIC_OPEN */
661 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
666 #ifndef HAVE_DCACHE_LOCK
667 /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */
668 if (nd->flags & LOOKUP_RCU)
672 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
673 struct lookup_intent *it;
675 it = ll_convert_intent(&nd->intent.open, nd->flags);
679 if (it->it_op == (IT_OPEN|IT_CREAT) &&
680 nd->intent.open.flags & O_EXCL) {
681 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
686 rc = ll_revalidate_it(dentry, nd->flags, it);
688 if (rc && (nd->flags & LOOKUP_OPEN) &&
689 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
690 // XXX Code duplication with ll_lookup_nd
691 if (S_ISFIFO(dentry->d_inode->i_mode)) {
692 // We cannot call open here as it would
695 (struct ptlrpc_request *)
696 it->d.lustre.it_data);
700 nd->intent.open.file->private_data = it;
701 filp = lookup_instantiate_filp(nd, dentry,NULL);
706 if (!rc && (nd->flags & LOOKUP_CREATE) &&
707 it_disposition(it, DISP_OPEN_CREATE)) {
708 /* We created something but we may only return
709 * negative dentry here, so save request in dentry,
710 * if lookup will be called later on, it will
711 * pick the request, otherwise it would be freed
713 ll_d2d(dentry)->lld_it = it;
714 it = NULL; /* avoid freeing */
719 ll_intent_release(it);
720 OBD_FREE(it, sizeof(*it));
723 rc = ll_revalidate_it(dentry, 0, NULL);
728 #endif /* HAVE_IOP_ATOMIC_OPEN */
730 void ll_d_iput(struct dentry *de, struct inode *inode)
733 if (!find_cbdata(inode))
738 struct dentry_operations ll_d_ops = {
739 .d_revalidate = ll_revalidate_nd,
740 .d_release = ll_release,
741 .d_delete = ll_ddelete,
743 .d_compare = ll_dcompare,