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);
78 /* Compare if two dentries are the same. Don't match if the existing dentry
79 * is marked invalid. Returns 1 if different, 0 if the same.
81 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
82 * an AST before calling d_revalidate_it(). The dentry still exists (marked
83 * INVALID) so d_lookup() matches it, but we have no lock on it (so
84 * lock_match() fails) and we spin around real_lookup(). */
85 int ll_dcompare(const struct dentry *parent, const struct dentry *dentry,
86 unsigned int len, const char *str, const struct qstr *name)
91 if (memcmp(str, name->name, len))
94 CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n",
95 name->len, name->name, dentry, dentry->d_flags,
98 /* mountpoint is always valid */
99 if (d_mountpoint((struct dentry *)dentry))
102 if (d_lustre_invalid(dentry))
108 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
111 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
112 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
113 return LDLM_ITER_CONTINUE;
114 return LDLM_ITER_STOP;
117 /* find any ldlm lock of the inode in mdc and lov
121 static int find_cbdata(struct inode *inode)
123 struct ll_sb_info *sbi = ll_i2sbi(inode);
124 struct lov_stripe_md *lsm;
128 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
129 return_if_equal, NULL);
133 lsm = ccc_inode_lsm_get(inode);
137 rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL);
138 ccc_inode_lsm_put(inode, lsm);
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(const struct dentry *de)
154 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
155 d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
156 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
157 d_unhashed((struct dentry *)de) ? "" : "hashed,",
158 list_empty(&de->d_subdirs) ? "" : "subdirs");
160 /* kernel >= 2.6.38 last refcount is decreased after this function. */
161 LASSERT(d_count(de) == 1);
163 /* Disable this piece of code temproarily because this is called
164 * inside dcache_lock so it's not appropriate to do lots of work
165 * here. ATTENTION: Before this piece of code enabling, LU-2487 must be
168 /* if not ldlm lock for this inode, set i_nlink to 0 so that
169 * this inode can be recycled later b=20433 */
170 if (de->d_inode && !find_cbdata(de->d_inode))
171 clear_nlink(de->d_inode);
174 if (d_lustre_invalid((struct dentry *)de))
179 static int ll_set_dd(struct dentry *de)
183 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
184 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
187 if (de->d_fsdata == NULL) {
188 struct ll_dentry_data *lld;
191 if (likely(lld != NULL)) {
192 spin_lock(&de->d_lock);
193 if (likely(de->d_fsdata == NULL))
197 spin_unlock(&de->d_lock);
206 int ll_dops_init(struct dentry *de, int block, int init_sa)
208 struct ll_dentry_data *lld = ll_d2d(de);
211 if (lld == NULL && block != 0) {
219 if (lld != NULL && init_sa != 0)
220 lld->lld_sa_generation = 0;
222 /* kernel >= 2.6.38 d_op is set in d_alloc() */
223 LASSERT(de->d_op == &ll_d_ops);
227 void ll_intent_drop_lock(struct lookup_intent *it)
229 if (it->it_op && it->d.lustre.it_lock_mode) {
230 struct lustre_handle handle;
232 handle.cookie = it->d.lustre.it_lock_handle;
234 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
235 " from it %p\n", handle.cookie, it);
236 ldlm_lock_decref(&handle, it->d.lustre.it_lock_mode);
238 /* bug 494: intent_release may be called multiple times, from
239 * this thread and we don't want to double-decref this lock */
240 it->d.lustre.it_lock_mode = 0;
241 if (it->d.lustre.it_remote_lock_mode != 0) {
242 handle.cookie = it->d.lustre.it_remote_lock_handle;
244 CDEBUG(D_DLMTRACE, "releasing remote lock with cookie"
245 LPX64" from it %p\n", handle.cookie, it);
246 ldlm_lock_decref(&handle,
247 it->d.lustre.it_remote_lock_mode);
248 it->d.lustre.it_remote_lock_mode = 0;
253 void ll_intent_release(struct lookup_intent *it)
255 CDEBUG(D_INFO, "intent %p released\n", it);
256 ll_intent_drop_lock(it);
257 /* We are still holding extra reference on a request, need to free it */
258 if (it_disposition(it, DISP_ENQ_OPEN_REF))
259 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
260 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
261 ptlrpc_req_finished(it->d.lustre.it_data);
262 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
264 ptlrpc_req_finished(it->d.lustre.it_data);
266 it->d.lustre.it_disposition = 0;
267 it->d.lustre.it_data = NULL;
270 void ll_invalidate_aliases(struct inode *inode)
272 struct dentry *dentry;
273 struct ll_d_hlist_node *p;
275 LASSERT(inode != NULL);
277 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
278 inode->i_ino, inode->i_generation, inode);
280 ll_lock_dcache(inode);
281 ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) {
282 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
283 "inode %p flags %d\n", dentry->d_name.len,
284 dentry->d_name.name, dentry, dentry->d_parent,
285 dentry->d_inode, dentry->d_flags);
287 if (unlikely(dentry == dentry->d_sb->s_root)) {
288 CERROR("%s: called on root dentry=%p, fid="DFID"\n",
289 ll_get_fsname(dentry->d_sb, NULL, 0),
290 dentry, PFID(ll_inode2fid(inode)));
291 lustre_dump_dentry(dentry, 1);
295 d_lustre_invalidate(dentry, 0);
297 ll_unlock_dcache(inode);
300 int ll_revalidate_it_finish(struct ptlrpc_request *request,
301 struct lookup_intent *it,
309 if (it_disposition(it, DISP_LOOKUP_NEG))
312 rc = ll_prep_inode(&de->d_inode, request, NULL, it);
317 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
320 LASSERT(dentry != NULL);
322 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
323 struct inode *inode = dentry->d_inode;
324 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
326 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
327 inode, inode->i_ino, inode->i_generation);
328 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
331 /* drop lookup or getattr locks immediately */
332 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
333 /* on 2.6 there are situation when several lookups and
334 * revalidations may be requested during single operation.
335 * therefore, we don't release intent here -bzzz */
336 ll_intent_drop_lock(it);
340 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
342 struct lookup_intent *it = *itp;
344 if (!it || it->it_op == IT_GETXATTR)
349 int ll_revalidate_it(struct dentry *de, int lookup_flags,
350 struct lookup_intent *it)
352 struct md_op_data *op_data;
353 struct ptlrpc_request *req = NULL;
354 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
355 struct obd_export *exp;
356 struct inode *parent = de->d_parent->d_inode;
359 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
362 if (de->d_inode == NULL) {
365 /* We can only use negative dentries if this is stat or lookup,
366 for opens and stuff we do need to query server. */
367 /* If there is IT_CREAT in intent op set, then we must throw
368 away this negative dentry and actually do the request to
369 kernel to create whatever needs to be created (if possible)*/
370 if (it && (it->it_op & IT_CREAT))
373 if (d_lustre_invalid(de))
376 ibits = MDS_INODELOCK_UPDATE;
377 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
381 /* Never execute intents for mount points.
382 * Attributes will be fixed up in ll_inode_revalidate_it */
383 if (d_mountpoint(de))
384 GOTO(out_sa, rc = 1);
386 /* need to get attributes in case root got changed from other client */
387 if (de == de->d_sb->s_root) {
388 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
394 exp = ll_i2mdexp(de->d_inode);
396 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
397 ll_frob_intent(&it, &lookup_it);
400 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
403 if (it->it_op == IT_OPEN) {
404 struct inode *inode = de->d_inode;
405 struct ll_inode_info *lli = ll_i2info(inode);
406 struct obd_client_handle **och_p;
410 * We used to check for MDS_INODELOCK_OPEN here, but in fact
411 * just having LOOKUP lock is enough to justify inode is the
412 * same. And if inode is the same and we have suitable
413 * openhandle, then there is no point in doing another OPEN RPC
414 * just to throw away newly received openhandle. There are no
415 * security implications too, if file owner or access mode is
416 * change, LOOKUP lock is revoked.
420 if (it->it_flags & FMODE_WRITE)
421 och_p = &lli->lli_mds_write_och;
422 else if (it->it_flags & FMODE_EXEC)
423 och_p = &lli->lli_mds_exec_och;
425 och_p = &lli->lli_mds_read_och;
427 /* Check for the proper lock. */
428 ibits = MDS_INODELOCK_LOOKUP;
429 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
431 mutex_lock(&lli->lli_och_mutex);
432 if (*och_p) { /* Everything is open already, do nothing */
433 /* Originally it was idea to do not let them steal our
434 * open handle from under us by (*och_usecount)++ here.
435 * But in case we have the handle, but we cannot use it
436 * due to later checks (e.g. O_CREAT|O_EXCL flags set),
437 * nobody would decrement counter increased here. So we
438 * just hope the lock won't be invalidated in between.
439 * But if it would be, we'll reopen the open request to
440 * MDS later during file open path.
442 mutex_unlock(&lli->lli_och_mutex);
445 mutex_unlock(&lli->lli_och_mutex);
448 if (it->it_op == IT_GETATTR) {
449 rc = ll_statahead_enter(parent, &de, 0);
452 else if (rc != -EAGAIN && rc != 0)
457 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
458 de->d_name.name, de->d_name.len,
459 0, LUSTRE_OPC_ANY, NULL);
461 return PTR_ERR(op_data);
463 if (!IS_POSIXACL(parent) || !exp_connect_umask(exp))
464 it->it_create_mode &= ~current_umask();
465 it->it_create_mode |= M_CHECK_STALE;
466 rc = md_intent_lock(exp, op_data, NULL, 0, it,
468 &req, ll_md_blocking_ast, 0);
469 it->it_create_mode &= ~M_CHECK_STALE;
470 ll_finish_md_op_data(op_data);
472 /* If req is NULL, then md_intent_lock only tried to do a lock match;
473 * if all was well, it will return 1 if it found locks, 0 otherwise. */
474 if (req == NULL && rc >= 0) {
482 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
483 "%d\n", rc, it->d.lustre.it_status);
489 rc = ll_revalidate_it_finish(req, it, de);
491 if (rc != -ESTALE && rc != -ENOENT)
492 ll_intent_release(it);
496 if ((it->it_op & IT_OPEN) && de->d_inode &&
497 !S_ISREG(de->d_inode->i_mode) &&
498 !S_ISDIR(de->d_inode->i_mode)) {
499 ll_release_openhandle(de, it);
504 /* We do not free request as it may be reused during following lookup
505 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
506 * be freed in ll_lookup_it or in ll_intent_release. But if
507 * request was not completed, we need to free it. (bug 5154, 9903) */
508 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
509 ptlrpc_req_finished(req);
511 /* mdt may grant layout lock for the newly created file, so
512 * release the lock to avoid leaking */
513 ll_intent_drop_lock(it);
514 ll_invalidate_aliases(de->d_inode);
517 __u64 matched_bits = 0;
519 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
520 "inode %p refc %d\n", de->d_name.len,
521 de->d_name.name, de, de->d_parent, de->d_inode,
524 ll_set_lock_data(exp, de->d_inode, it, &bits);
526 /* Note: We have to match both LOOKUP and PERM lock
527 * here to make sure the dentry is valid and no one
528 * changing the permission.
529 * But if the client connects < 2.4 server, which will
530 * only grant LOOKUP lock, so we can only Match LOOKUP
531 * lock for old server */
532 if (exp_connect_flags(ll_i2mdexp(de->d_inode)) &&
533 OBD_CONNECT_LVB_TYPE)
535 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
537 matched_bits = MDS_INODELOCK_LOOKUP;
539 if (((bits & matched_bits) == matched_bits) &&
540 d_lustre_invalid(de))
541 d_lustre_revalidate(de);
542 ll_lookup_finish_locks(it, de);
546 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
547 ll_statahead_mark(parent, de);
551 * This part is here to combat evil-evil race in real_lookup on 2.6
552 * kernels. The race details are: We enter do_lookup() looking for some
553 * name, there is nothing in dcache for this name yet and d_lookup()
554 * returns NULL. We proceed to real_lookup(), and while we do this,
555 * another process does open on the same file we looking up (most simple
556 * reproducer), open succeeds and the dentry is added. Now back to
557 * us. In real_lookup() we do d_lookup() again and suddenly find the
558 * dentry, so we call d_revalidate on it, but there is no lock, so
559 * without this code we would return 0, but unpatched real_lookup just
560 * returns -ENOENT in such a case instead of retrying the lookup. Once
561 * this is dealt with in real_lookup(), all of this ugly mess can go and
562 * we can just check locks in ->d_revalidate without doing any RPCs
566 if (it != &lookup_it) {
567 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
568 if (it->it_op == IT_GETATTR)
569 lookup_it.it_op = IT_GETATTR;
570 ll_lookup_finish_locks(it, de);
574 /* Do real lookup here. */
575 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
576 de->d_name.len, 0, (it->it_op & IT_CREAT ?
578 LUSTRE_OPC_ANY), NULL);
580 return PTR_ERR(op_data);
582 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
583 ll_md_blocking_ast, 0);
585 struct mdt_body *mdt_body;
586 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
587 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
590 fid = *ll_inode2fid(de->d_inode);
592 /* see if we got same inode, if not - return error */
593 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
594 ll_finish_md_op_data(op_data);
596 goto revalidate_finish;
598 ll_intent_release(it);
600 ll_finish_md_op_data(op_data);
605 * For rc == 1 case, should not return directly to prevent losing
606 * statahead windows; for rc == 0 case, the "lookup" will be done later.
608 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
609 ll_statahead_enter(parent, &de, 1);
614 * Always trust cached dentries. Update statahead window if necessary.
616 int ll_revalidate_nd(struct dentry *dentry, unsigned int flags)
618 struct inode *parent = dentry->d_parent->d_inode;
621 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,flags=%u\n",
622 dentry->d_name.name, flags);
624 if (!(flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE)) &&
625 ll_need_statahead(parent, dentry) > 0) {
626 if (flags & LOOKUP_RCU)
629 if (dentry->d_inode == NULL)
631 do_statahead_enter(parent, &dentry, unplug);
632 ll_statahead_mark(parent, dentry);
639 void ll_d_iput(struct dentry *de, struct inode *inode)
642 if (!find_cbdata(inode))
647 struct dentry_operations ll_d_ops = {
648 .d_revalidate = ll_revalidate_nd,
649 .d_release = ll_release,
650 .d_delete = ll_ddelete,
652 .d_compare = ll_dcompare,