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, Whamcloud, Inc.
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/smp_lock.h>
40 #include <linux/quotaops.h>
42 #define DEBUG_SUBSYSTEM S_LLITE
44 #include <obd_support.h>
45 #include <lustre_lite.h>
46 #include <lustre/lustre_idl.h>
47 #include <lustre_dlm.h>
49 #include "llite_internal.h"
51 cfs_spinlock_t ll_lookup_lock = CFS_SPIN_LOCK_UNLOCKED;
53 /* should NOT be called with the dcache lock, see fs/dcache.c */
54 static void ll_release(struct dentry *de)
56 struct ll_dentry_data *lld;
60 if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
64 ll_intent_release(lld->lld_it);
65 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
67 LASSERT(lld->lld_cwd_count == 0);
68 LASSERT(lld->lld_mnt_count == 0);
69 OBD_FREE(de->d_fsdata, sizeof(*lld));
74 /* Compare if two dentries are the same. Don't match if the existing dentry
75 * is marked DCACHE_LUSTRE_INVALID. Returns 1 if different, 0 if the same.
77 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
78 * an AST before calling d_revalidate_it(). The dentry still exists (marked
79 * INVALID) so d_lookup() matches it, but we have no lock on it (so
80 * lock_match() fails) and we spin around real_lookup(). */
81 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
83 struct dentry *dchild;
86 if (d_name->len != name->len)
89 if (memcmp(d_name->name, name->name, name->len))
92 /* XXX: d_name must be in-dentry structure */
93 dchild = container_of(d_name, struct dentry, d_name); /* ugh */
95 CDEBUG(D_DENTRY,"found name %.*s(%p) - flags %d/%x - refc %d\n",
96 name->len, name->name, dchild,
97 d_mountpoint(dchild), dchild->d_flags & DCACHE_LUSTRE_INVALID,
98 atomic_read(&dchild->d_count));
100 /* mountpoint is always valid */
101 if (d_mountpoint(dchild))
104 if (dchild->d_flags & DCACHE_LUSTRE_INVALID)
110 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
113 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
114 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
115 return LDLM_ITER_CONTINUE;
116 return LDLM_ITER_STOP;
119 /* find any ldlm lock of the inode in mdc and lov
123 static int find_cbdata(struct inode *inode)
125 struct ll_inode_info *lli = ll_i2info(inode);
126 struct ll_sb_info *sbi = ll_i2sbi(inode);
131 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
132 return_if_equal, NULL);
137 rc = obd_find_cbdata(sbi->ll_dt_exp, lli->lli_smd,
138 return_if_equal, NULL);
144 * Called when last reference to a dentry is dropped and dcache wants to know
145 * whether or not it should cache it:
146 * - return 1 to delete the dentry immediately
147 * - return 0 to cache the dentry
148 * Should NOT be called with the dcache lock, see fs/dcache.c
150 static int ll_ddelete(struct dentry *de)
155 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
156 (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
157 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
158 d_unhashed(de) ? "" : "hashed,",
159 list_empty(&de->d_subdirs) ? "" : "subdirs");
161 /* if not ldlm lock for this inode, set i_nlink to 0 so that
162 * this inode can be recycled later b=20433 */
163 LASSERT(atomic_read(&de->d_count) == 0);
164 if (de->d_inode && !find_cbdata(de->d_inode))
165 de->d_inode->i_nlink = 0;
167 if (de->d_flags & DCACHE_LUSTRE_INVALID)
173 static int ll_set_dd(struct dentry *de)
178 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
179 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
180 atomic_read(&de->d_count));
182 if (de->d_fsdata == NULL) {
183 struct ll_dentry_data *lld;
186 if (likely(lld != NULL)) {
188 if (likely(de->d_fsdata == NULL))
201 int ll_dops_init(struct dentry *de, int block, int init_sa)
203 struct ll_dentry_data *lld = ll_d2d(de);
206 if (lld == NULL && block != 0) {
214 if (lld != NULL && init_sa != 0)
215 lld->lld_sa_generation = 0;
217 de->d_op = &ll_d_ops;
221 void ll_intent_drop_lock(struct lookup_intent *it)
223 struct lustre_handle *handle;
225 if (it->it_op && it->d.lustre.it_lock_mode) {
226 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
227 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
228 " from it %p\n", handle->cookie, it);
229 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
231 /* bug 494: intent_release may be called multiple times, from
232 * this thread and we don't want to double-decref this lock */
233 it->d.lustre.it_lock_mode = 0;
237 void ll_intent_release(struct lookup_intent *it)
241 CDEBUG(D_INFO, "intent %p released\n", it);
242 ll_intent_drop_lock(it);
243 /* We are still holding extra reference on a request, need to free it */
244 if (it_disposition(it, DISP_ENQ_OPEN_REF))
245 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
246 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
247 ptlrpc_req_finished(it->d.lustre.it_data);
248 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
250 ptlrpc_req_finished(it->d.lustre.it_data);
252 it->d.lustre.it_disposition = 0;
253 it->d.lustre.it_data = NULL;
257 /* Drop dentry if it is not used already, unhash otherwise.
258 Should be called with dcache lock held!
259 Returns: 1 if dentry was dropped, 0 if unhashed. */
260 int ll_drop_dentry(struct dentry *dentry)
263 if (atomic_read(&dentry->d_count) == 0) {
264 CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p "
265 "inode %p\n", dentry->d_name.len,
266 dentry->d_name.name, dentry, dentry->d_parent,
270 unlock_dentry(dentry);
271 spin_unlock(&dcache_lock);
272 cfs_spin_unlock(&ll_lookup_lock);
274 cfs_spin_lock(&ll_lookup_lock);
275 spin_lock(&dcache_lock);
278 /* disconected dentry can not be find without lookup, because we
279 * not need his to unhash or mark invalid. */
280 if (dentry->d_flags & DCACHE_DISCONNECTED) {
281 unlock_dentry(dentry);
285 if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
286 CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p "
287 "inode %p refc %d\n", dentry->d_name.len,
288 dentry->d_name.name, dentry, dentry->d_parent,
289 dentry->d_inode, atomic_read(&dentry->d_count));
290 /* actually we don't unhash the dentry, rather just
291 * mark it inaccessible for to __d_lookup(). otherwise
292 * sys_getcwd() could return -ENOENT -bzzz */
293 dentry->d_flags |= DCACHE_LUSTRE_INVALID;
294 if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
297 unlock_dentry(dentry);
301 void ll_unhash_aliases(struct inode *inode)
303 struct list_head *tmp, *head;
307 CERROR("unexpected NULL inode, tell phil\n");
311 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
312 inode->i_ino, inode->i_generation, inode);
314 head = &inode->i_dentry;
315 cfs_spin_lock(&ll_lookup_lock);
316 spin_lock(&dcache_lock);
319 while ((tmp = tmp->next) != head) {
320 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
322 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
323 "inode %p flags %d\n", dentry->d_name.len,
324 dentry->d_name.name, dentry, dentry->d_parent,
325 dentry->d_inode, dentry->d_flags);
327 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
328 CERROR("called on root (?) dentry=%p, inode=%p "
329 "ino=%lu\n", dentry, inode, inode->i_ino);
330 lustre_dump_dentry(dentry, 1);
331 libcfs_debug_dumpstack(NULL);
334 if (ll_drop_dentry(dentry))
337 spin_unlock(&dcache_lock);
338 cfs_spin_unlock(&ll_lookup_lock);
343 int ll_revalidate_it_finish(struct ptlrpc_request *request,
344 struct lookup_intent *it,
353 if (it_disposition(it, DISP_LOOKUP_NEG))
356 rc = ll_prep_inode(&de->d_inode, request, NULL);
361 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
364 LASSERT(dentry != NULL);
366 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
367 struct inode *inode = dentry->d_inode;
368 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
370 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
371 inode, inode->i_ino, inode->i_generation);
372 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
375 /* drop lookup or getattr locks immediately */
376 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
377 /* on 2.6 there are situation when several lookups and
378 * revalidations may be requested during single operation.
379 * therefore, we don't release intent here -bzzz */
380 ll_intent_drop_lock(it);
384 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
386 struct lookup_intent *it = *itp;
388 if (!it || it->it_op == IT_GETXATTR)
393 int ll_revalidate_it(struct dentry *de, int lookup_flags,
394 struct lookup_intent *it)
396 struct md_op_data *op_data;
397 struct ptlrpc_request *req = NULL;
398 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
399 struct obd_export *exp;
400 struct inode *parent = de->d_parent->d_inode;
404 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
407 if (de->d_inode == NULL) {
410 /* We can only use negative dentries if this is stat or lookup,
411 for opens and stuff we do need to query server. */
412 /* If there is IT_CREAT in intent op set, then we must throw
413 away this negative dentry and actually do the request to
414 kernel to create whatever needs to be created (if possible)*/
415 if (it && (it->it_op & IT_CREAT))
418 if (de->d_flags & DCACHE_LUSTRE_INVALID)
421 ibits = MDS_INODELOCK_UPDATE;
422 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
426 /* Never execute intents for mount points.
427 * Attributes will be fixed up in ll_inode_revalidate_it */
428 if (d_mountpoint(de))
429 GOTO(out_sa, rc = 1);
431 /* need to get attributes in case root got changed from other client */
432 if (de == de->d_sb->s_root) {
433 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
439 exp = ll_i2mdexp(de->d_inode);
441 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
442 ll_frob_intent(&it, &lookup_it);
445 if (it->it_op == IT_LOOKUP && !(de->d_flags & DCACHE_LUSTRE_INVALID))
448 if ((it->it_op == IT_OPEN) && de->d_inode) {
449 struct inode *inode = de->d_inode;
450 struct ll_inode_info *lli = ll_i2info(inode);
451 struct obd_client_handle **och_p;
456 * We used to check for MDS_INODELOCK_OPEN here, but in fact
457 * just having LOOKUP lock is enough to justify inode is the
458 * same. And if inode is the same and we have suitable
459 * openhandle, then there is no point in doing another OPEN RPC
460 * just to throw away newly received openhandle. There are no
461 * security implications too, if file owner or access mode is
462 * change, LOOKUP lock is revoked.
466 if (it->it_flags & FMODE_WRITE) {
467 och_p = &lli->lli_mds_write_och;
468 och_usecount = &lli->lli_open_fd_write_count;
469 } else if (it->it_flags & FMODE_EXEC) {
470 och_p = &lli->lli_mds_exec_och;
471 och_usecount = &lli->lli_open_fd_exec_count;
473 och_p = &lli->lli_mds_read_och;
474 och_usecount = &lli->lli_open_fd_read_count;
476 /* Check for the proper lock. */
477 ibits = MDS_INODELOCK_LOOKUP;
478 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
480 cfs_mutex_lock(&lli->lli_och_mutex);
481 if (*och_p) { /* Everything is open already, do nothing */
482 /*(*och_usecount)++; Do not let them steal our open
483 handle from under us */
484 SET_BUT_UNUSED(och_usecount);
485 /* XXX The code above was my original idea, but in case
486 we have the handle, but we cannot use it due to later
487 checks (e.g. O_CREAT|O_EXCL flags set), nobody
488 would decrement counter increased here. So we just
489 hope the lock won't be invalidated in between. But
490 if it would be, we'll reopen the open request to
491 MDS later during file open path */
492 cfs_mutex_unlock(&lli->lli_och_mutex);
495 cfs_mutex_unlock(&lli->lli_och_mutex);
499 if (it->it_op == IT_GETATTR) {
500 rc = ll_statahead_enter(parent, &de, 0);
503 else if (rc != -EAGAIN && rc != 0)
508 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
509 de->d_name.name, de->d_name.len,
510 0, LUSTRE_OPC_ANY, NULL);
512 RETURN(PTR_ERR(op_data));
514 it->it_create_mode &= ~cfs_curproc_umask();
515 it->it_create_mode |= M_CHECK_STALE;
516 rc = md_intent_lock(exp, op_data, NULL, 0, it,
518 &req, ll_md_blocking_ast, 0);
519 it->it_create_mode &= ~M_CHECK_STALE;
520 ll_finish_md_op_data(op_data);
522 /* If req is NULL, then md_intent_lock only tried to do a lock match;
523 * if all was well, it will return 1 if it found locks, 0 otherwise. */
524 if (req == NULL && rc >= 0) {
532 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
533 "%d\n", rc, it->d.lustre.it_status);
539 rc = ll_revalidate_it_finish(req, it, de);
541 if (rc != -ESTALE && rc != -ENOENT)
542 ll_intent_release(it);
546 if ((it->it_op & IT_OPEN) && de->d_inode &&
547 !S_ISREG(de->d_inode->i_mode) &&
548 !S_ISDIR(de->d_inode->i_mode)) {
549 ll_release_openhandle(de, it);
553 /* unfortunately ll_intent_lock may cause a callback and revoke our
555 ll_dentry_rehash(de, 0);
558 /* We do not free request as it may be reused during following lookup
559 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
560 * be freed in ll_lookup_it or in ll_intent_release. But if
561 * request was not completed, we need to free it. (bug 5154, 9903) */
562 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
563 ptlrpc_req_finished(req);
565 ll_unhash_aliases(de->d_inode);
566 /* done in ll_unhash_aliases()
567 dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
571 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
572 "inode %p refc %d\n", de->d_name.len,
573 de->d_name.name, de, de->d_parent, de->d_inode,
574 atomic_read(&de->d_count));
575 ll_set_lock_data(exp, de->d_inode, it, &bits);
576 ll_dentry_reset_flags(de, bits);
577 ll_lookup_finish_locks(it, de);
581 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
582 ll_statahead_mark(parent, de);
586 * This part is here to combat evil-evil race in real_lookup on 2.6
587 * kernels. The race details are: We enter do_lookup() looking for some
588 * name, there is nothing in dcache for this name yet and d_lookup()
589 * returns NULL. We proceed to real_lookup(), and while we do this,
590 * another process does open on the same file we looking up (most simple
591 * reproducer), open succeeds and the dentry is added. Now back to
592 * us. In real_lookup() we do d_lookup() again and suddenly find the
593 * dentry, so we call d_revalidate on it, but there is no lock, so
594 * without this code we would return 0, but unpatched real_lookup just
595 * returns -ENOENT in such a case instead of retrying the lookup. Once
596 * this is dealt with in real_lookup(), all of this ugly mess can go and
597 * we can just check locks in ->d_revalidate without doing any RPCs
601 if (it != &lookup_it) {
602 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
603 if (it->it_op == IT_GETATTR)
604 lookup_it.it_op = IT_GETATTR;
605 ll_lookup_finish_locks(it, de);
609 /* Do real lookup here. */
610 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
611 de->d_name.len, 0, (it->it_op & IT_CREAT ?
613 LUSTRE_OPC_ANY), NULL);
615 RETURN(PTR_ERR(op_data));
617 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
618 ll_md_blocking_ast, 0);
620 struct mdt_body *mdt_body;
621 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
622 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
625 fid = *ll_inode2fid(de->d_inode);
627 /* see if we got same inode, if not - return error */
628 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
629 ll_finish_md_op_data(op_data);
631 goto revalidate_finish;
633 ll_intent_release(it);
635 ll_finish_md_op_data(op_data);
640 * For rc == 1 case, should not return directly to prevent losing
641 * statahead windows; for rc == 0 case, the "lookup" will be done later.
643 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
644 ll_statahead_enter(parent, &de, 1);
648 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
653 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
654 struct lookup_intent *it;
656 it = ll_convert_intent(&nd->intent.open, nd->flags);
660 if (it->it_op == (IT_OPEN|IT_CREAT) &&
661 nd->intent.open.flags & O_EXCL) {
662 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
667 rc = ll_revalidate_it(dentry, nd->flags, it);
669 if (rc && (nd->flags & LOOKUP_OPEN) &&
670 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
671 // XXX Code duplication with ll_lookup_nd
672 if (S_ISFIFO(dentry->d_inode->i_mode)) {
673 // We cannot call open here as it would
676 (struct ptlrpc_request *)
677 it->d.lustre.it_data);
681 nd->intent.open.file->private_data = it;
682 filp = lookup_instantiate_filp(nd, dentry,NULL);
687 if (!rc && (nd->flags & LOOKUP_CREATE) &&
688 it_disposition(it, DISP_OPEN_CREATE)) {
689 /* We created something but we may only return
690 * negative dentry here, so save request in dentry,
691 * if lookup will be called later on, it will
692 * pick the request, otherwise it would be freed
694 ll_d2d(dentry)->lld_it = it;
695 it = NULL; /* avoid freeing */
700 ll_intent_release(it);
701 OBD_FREE(it, sizeof(*it));
704 rc = ll_revalidate_it(dentry, 0, NULL);
710 void ll_d_iput(struct dentry *de, struct inode *inode)
713 if (!find_cbdata(inode))
718 struct dentry_operations ll_d_ops = {
719 .d_revalidate = ll_revalidate_nd,
720 .d_release = ll_release,
721 .d_delete = ll_ddelete,
723 .d_compare = ll_dcompare,