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/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 /* should NOT be called with the dcache lock, see fs/dcache.c */
51 static void ll_release(struct dentry *de)
53 struct ll_dentry_data *lld;
57 if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
61 ll_intent_release(lld->lld_it);
62 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
64 LASSERT(lld->lld_cwd_count == 0);
65 LASSERT(lld->lld_mnt_count == 0);
66 OBD_FREE(de->d_fsdata, sizeof(*lld));
71 /* Compare if two dentries are the same. Don't match if the existing dentry
72 * is marked DCACHE_LUSTRE_INVALID. Returns 1 if different, 0 if the same.
74 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
75 * an AST before calling d_revalidate_it(). The dentry still exists (marked
76 * INVALID) so d_lookup() matches it, but we have no lock on it (so
77 * lock_match() fails) and we spin around real_lookup(). */
78 #ifdef HAVE_D_COMPARE_7ARGS
79 int ll_dcompare(const struct dentry *parent, const struct inode *pinode,
80 const struct dentry *dentry, const struct inode *inode,
81 unsigned int len, const char *str, const struct qstr *name)
83 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
86 #ifdef HAVE_D_COMPARE_7ARGS
92 if (memcmp(str, name->name, len))
95 struct dentry *dentry;
98 if (d_name->len != name->len)
101 if (memcmp(d_name->name, name->name, name->len))
104 /* XXX: d_name must be in-dentry structure */
105 dentry = container_of(d_name, struct dentry, d_name); /* ugh */
108 CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n",
109 name->len, name->name, dentry, dentry->d_flags,
112 /* mountpoint is always valid */
113 if (d_mountpoint((struct dentry *)dentry))
116 if (d_lustre_invalid(dentry))
122 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
125 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
126 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
127 return LDLM_ITER_CONTINUE;
128 return LDLM_ITER_STOP;
131 /* find any ldlm lock of the inode in mdc and lov
135 static int find_cbdata(struct inode *inode)
137 struct ll_sb_info *sbi = ll_i2sbi(inode);
138 struct lov_stripe_md *lsm;
143 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
144 return_if_equal, NULL);
148 lsm = ccc_inode_lsm_get(inode);
152 rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL);
153 ccc_inode_lsm_put(inode, lsm);
159 * Called when last reference to a dentry is dropped and dcache wants to know
160 * whether or not it should cache it:
161 * - return 1 to delete the dentry immediately
162 * - return 0 to cache the dentry
163 * Should NOT be called with the dcache lock, see fs/dcache.c
165 static int ll_ddelete(HAVE_D_DELETE_CONST struct dentry *de)
170 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
171 d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
172 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
173 d_unhashed((struct dentry *)de) ? "" : "hashed,",
174 list_empty(&de->d_subdirs) ? "" : "subdirs");
176 #ifdef HAVE_DCACHE_LOCK
177 LASSERT(d_refcount(de) == 0);
179 /* kernel >= 2.6.38 last refcount is decreased after this function. */
180 LASSERT(d_refcount(de) == 1);
183 /* Disable this piece of code temproarily because this is called
184 * inside dcache_lock so it's not appropriate to do lots of work
187 /* if not ldlm lock for this inode, set i_nlink to 0 so that
188 * this inode can be recycled later b=20433 */
189 if (de->d_inode && !find_cbdata(de->d_inode))
190 de->d_inode->i_nlink = 0;
193 if (d_lustre_invalid((struct dentry *)de))
198 static int ll_set_dd(struct dentry *de)
203 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
204 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
207 if (de->d_fsdata == NULL) {
208 struct ll_dentry_data *lld;
211 if (likely(lld != NULL)) {
212 spin_lock(&de->d_lock);
213 if (likely(de->d_fsdata == NULL))
217 spin_unlock(&de->d_lock);
226 int ll_dops_init(struct dentry *de, int block, int init_sa)
228 struct ll_dentry_data *lld = ll_d2d(de);
231 if (lld == NULL && block != 0) {
239 if (lld != NULL && init_sa != 0)
240 lld->lld_sa_generation = 0;
242 #ifdef HAVE_DCACHE_LOCK
243 de->d_op = &ll_d_ops;
245 /* kernel >= 2.6.38 d_op is set in d_alloc() */
246 LASSERT(de->d_op == &ll_d_ops);
251 void ll_intent_drop_lock(struct lookup_intent *it)
253 struct lustre_handle *handle;
255 if (it->it_op && it->d.lustre.it_lock_mode) {
256 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
257 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
258 " from it %p\n", handle->cookie, it);
259 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
261 /* bug 494: intent_release may be called multiple times, from
262 * this thread and we don't want to double-decref this lock */
263 it->d.lustre.it_lock_mode = 0;
267 void ll_intent_release(struct lookup_intent *it)
271 CDEBUG(D_INFO, "intent %p released\n", it);
272 ll_intent_drop_lock(it);
273 /* We are still holding extra reference on a request, need to free it */
274 if (it_disposition(it, DISP_ENQ_OPEN_REF))
275 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
276 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
277 ptlrpc_req_finished(it->d.lustre.it_data);
278 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
280 ptlrpc_req_finished(it->d.lustre.it_data);
282 it->d.lustre.it_disposition = 0;
283 it->d.lustre.it_data = NULL;
287 void ll_invalidate_aliases(struct inode *inode)
289 struct dentry *dentry;
292 LASSERT(inode != NULL);
294 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
295 inode->i_ino, inode->i_generation, inode);
297 ll_lock_dcache(inode);
298 cfs_list_for_each_entry(dentry, &inode->i_dentry, d_alias) {
299 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
300 "inode %p flags %d\n", dentry->d_name.len,
301 dentry->d_name.name, dentry, dentry->d_parent,
302 dentry->d_inode, dentry->d_flags);
304 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
305 CERROR("called on root (?) dentry=%p, inode=%p "
306 "ino=%lu\n", dentry, inode, inode->i_ino);
307 lustre_dump_dentry(dentry, 1);
308 libcfs_debug_dumpstack(NULL);
311 d_lustre_invalidate(dentry);
313 ll_unlock_dcache(inode);
318 int ll_revalidate_it_finish(struct ptlrpc_request *request,
319 struct lookup_intent *it,
328 if (it_disposition(it, DISP_LOOKUP_NEG))
331 rc = ll_prep_inode(&de->d_inode, request, NULL);
336 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
339 LASSERT(dentry != NULL);
341 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
342 struct inode *inode = dentry->d_inode;
343 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
345 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
346 inode, inode->i_ino, inode->i_generation);
347 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
350 /* drop lookup or getattr locks immediately */
351 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
352 /* on 2.6 there are situation when several lookups and
353 * revalidations may be requested during single operation.
354 * therefore, we don't release intent here -bzzz */
355 ll_intent_drop_lock(it);
359 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
361 struct lookup_intent *it = *itp;
363 if (!it || it->it_op == IT_GETXATTR)
368 int ll_revalidate_it(struct dentry *de, int lookup_flags,
369 struct lookup_intent *it)
371 struct md_op_data *op_data;
372 struct ptlrpc_request *req = NULL;
373 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
374 struct obd_export *exp;
375 struct inode *parent = de->d_parent->d_inode;
379 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
382 if (de->d_inode == NULL) {
385 /* We can only use negative dentries if this is stat or lookup,
386 for opens and stuff we do need to query server. */
387 /* If there is IT_CREAT in intent op set, then we must throw
388 away this negative dentry and actually do the request to
389 kernel to create whatever needs to be created (if possible)*/
390 if (it && (it->it_op & IT_CREAT))
393 if (d_lustre_invalid(de))
396 ibits = MDS_INODELOCK_UPDATE;
397 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
401 /* Never execute intents for mount points.
402 * Attributes will be fixed up in ll_inode_revalidate_it */
403 if (d_mountpoint(de))
404 GOTO(out_sa, rc = 1);
406 /* need to get attributes in case root got changed from other client */
407 if (de == de->d_sb->s_root) {
408 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
414 exp = ll_i2mdexp(de->d_inode);
416 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
417 ll_frob_intent(&it, &lookup_it);
420 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
423 if ((it->it_op == IT_OPEN) && de->d_inode) {
424 struct inode *inode = de->d_inode;
425 struct ll_inode_info *lli = ll_i2info(inode);
426 struct obd_client_handle **och_p;
431 * We used to check for MDS_INODELOCK_OPEN here, but in fact
432 * just having LOOKUP lock is enough to justify inode is the
433 * same. And if inode is the same and we have suitable
434 * openhandle, then there is no point in doing another OPEN RPC
435 * just to throw away newly received openhandle. There are no
436 * security implications too, if file owner or access mode is
437 * change, LOOKUP lock is revoked.
441 if (it->it_flags & FMODE_WRITE) {
442 och_p = &lli->lli_mds_write_och;
443 och_usecount = &lli->lli_open_fd_write_count;
444 } else if (it->it_flags & FMODE_EXEC) {
445 och_p = &lli->lli_mds_exec_och;
446 och_usecount = &lli->lli_open_fd_exec_count;
448 och_p = &lli->lli_mds_read_och;
449 och_usecount = &lli->lli_open_fd_read_count;
451 /* Check for the proper lock. */
452 ibits = MDS_INODELOCK_LOOKUP;
453 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
455 cfs_mutex_lock(&lli->lli_och_mutex);
456 if (*och_p) { /* Everything is open already, do nothing */
457 /*(*och_usecount)++; Do not let them steal our open
458 handle from under us */
459 SET_BUT_UNUSED(och_usecount);
460 /* XXX The code above was my original idea, but in case
461 we have the handle, but we cannot use it due to later
462 checks (e.g. O_CREAT|O_EXCL flags set), nobody
463 would decrement counter increased here. So we just
464 hope the lock won't be invalidated in between. But
465 if it would be, we'll reopen the open request to
466 MDS later during file open path */
467 cfs_mutex_unlock(&lli->lli_och_mutex);
470 cfs_mutex_unlock(&lli->lli_och_mutex);
474 if (it->it_op == IT_GETATTR) {
475 rc = ll_statahead_enter(parent, &de, 0);
478 else if (rc != -EAGAIN && rc != 0)
483 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
484 de->d_name.name, de->d_name.len,
485 0, LUSTRE_OPC_ANY, NULL);
487 RETURN(PTR_ERR(op_data));
489 it->it_create_mode &= ~cfs_curproc_umask();
490 it->it_create_mode |= M_CHECK_STALE;
491 rc = md_intent_lock(exp, op_data, NULL, 0, it,
493 &req, ll_md_blocking_ast, 0);
494 it->it_create_mode &= ~M_CHECK_STALE;
495 ll_finish_md_op_data(op_data);
497 /* If req is NULL, then md_intent_lock only tried to do a lock match;
498 * if all was well, it will return 1 if it found locks, 0 otherwise. */
499 if (req == NULL && rc >= 0) {
507 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
508 "%d\n", rc, it->d.lustre.it_status);
514 rc = ll_revalidate_it_finish(req, it, de);
516 if (rc != -ESTALE && rc != -ENOENT)
517 ll_intent_release(it);
521 if ((it->it_op & IT_OPEN) && de->d_inode &&
522 !S_ISREG(de->d_inode->i_mode) &&
523 !S_ISDIR(de->d_inode->i_mode)) {
524 ll_release_openhandle(de, it);
529 /* We do not free request as it may be reused during following lookup
530 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
531 * be freed in ll_lookup_it or in ll_intent_release. But if
532 * request was not completed, we need to free it. (bug 5154, 9903) */
533 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
534 ptlrpc_req_finished(req);
536 ll_invalidate_aliases(de->d_inode);
540 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
541 "inode %p refc %d\n", de->d_name.len,
542 de->d_name.name, de, de->d_parent, de->d_inode,
544 ll_set_lock_data(exp, de->d_inode, it, &bits);
545 if ((bits & MDS_INODELOCK_LOOKUP) && d_lustre_invalid(de))
546 d_lustre_revalidate(de);
547 ll_lookup_finish_locks(it, de);
551 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
552 ll_statahead_mark(parent, de);
556 * This part is here to combat evil-evil race in real_lookup on 2.6
557 * kernels. The race details are: We enter do_lookup() looking for some
558 * name, there is nothing in dcache for this name yet and d_lookup()
559 * returns NULL. We proceed to real_lookup(), and while we do this,
560 * another process does open on the same file we looking up (most simple
561 * reproducer), open succeeds and the dentry is added. Now back to
562 * us. In real_lookup() we do d_lookup() again and suddenly find the
563 * dentry, so we call d_revalidate on it, but there is no lock, so
564 * without this code we would return 0, but unpatched real_lookup just
565 * returns -ENOENT in such a case instead of retrying the lookup. Once
566 * this is dealt with in real_lookup(), all of this ugly mess can go and
567 * we can just check locks in ->d_revalidate without doing any RPCs
571 if (it != &lookup_it) {
572 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
573 if (it->it_op == IT_GETATTR)
574 lookup_it.it_op = IT_GETATTR;
575 ll_lookup_finish_locks(it, de);
579 /* Do real lookup here. */
580 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
581 de->d_name.len, 0, (it->it_op & IT_CREAT ?
583 LUSTRE_OPC_ANY), NULL);
585 RETURN(PTR_ERR(op_data));
587 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
588 ll_md_blocking_ast, 0);
590 struct mdt_body *mdt_body;
591 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
592 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
595 fid = *ll_inode2fid(de->d_inode);
597 /* see if we got same inode, if not - return error */
598 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
599 ll_finish_md_op_data(op_data);
601 goto revalidate_finish;
603 ll_intent_release(it);
605 ll_finish_md_op_data(op_data);
610 * For rc == 1 case, should not return directly to prevent losing
611 * statahead windows; for rc == 0 case, the "lookup" will be done later.
613 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
614 ll_statahead_enter(parent, &de, 1);
618 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
623 #ifndef HAVE_DCACHE_LOCK
624 /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */
625 if (nd->flags & LOOKUP_RCU)
629 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
630 struct lookup_intent *it;
632 it = ll_convert_intent(&nd->intent.open, nd->flags);
636 if (it->it_op == (IT_OPEN|IT_CREAT) &&
637 nd->intent.open.flags & O_EXCL) {
638 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
643 rc = ll_revalidate_it(dentry, nd->flags, it);
645 if (rc && (nd->flags & LOOKUP_OPEN) &&
646 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
647 // XXX Code duplication with ll_lookup_nd
648 if (S_ISFIFO(dentry->d_inode->i_mode)) {
649 // We cannot call open here as it would
652 (struct ptlrpc_request *)
653 it->d.lustre.it_data);
657 nd->intent.open.file->private_data = it;
658 filp = lookup_instantiate_filp(nd, dentry,NULL);
663 if (!rc && (nd->flags & LOOKUP_CREATE) &&
664 it_disposition(it, DISP_OPEN_CREATE)) {
665 /* We created something but we may only return
666 * negative dentry here, so save request in dentry,
667 * if lookup will be called later on, it will
668 * pick the request, otherwise it would be freed
670 ll_d2d(dentry)->lld_it = it;
671 it = NULL; /* avoid freeing */
676 ll_intent_release(it);
677 OBD_FREE(it, sizeof(*it));
680 rc = ll_revalidate_it(dentry, 0, NULL);
686 void ll_d_iput(struct dentry *de, struct inode *inode)
689 if (!find_cbdata(inode))
694 struct dentry_operations ll_d_ops = {
695 .d_revalidate = ll_revalidate_nd,
696 .d_release = ll_release,
697 .d_delete = ll_ddelete,
699 .d_compare = ll_dcompare,