1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
41 #include <linux/sched.h>
42 #include <linux/smp_lock.h>
43 #include <linux/quotaops.h>
45 #define DEBUG_SUBSYSTEM S_LLITE
47 #include <obd_support.h>
48 #include <lustre_lite.h>
49 #include <lustre/lustre_idl.h>
50 #include <lustre_dlm.h>
51 //#include <lustre_ver.h>
52 //#include <lustre_version.h>
54 #include "llite_internal.h"
56 cfs_spinlock_t ll_lookup_lock = CFS_SPIN_LOCK_UNLOCKED;
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) */
70 ll_intent_release(lld->lld_it);
71 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
73 LASSERT(lld->lld_cwd_count == 0);
74 LASSERT(lld->lld_mnt_count == 0);
75 OBD_FREE(de->d_fsdata, sizeof(*lld));
80 /* Compare if two dentries are the same. Don't match if the existing dentry
81 * is marked DCACHE_LUSTRE_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 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
89 struct dentry *dchild;
92 if (d_name->len != name->len)
95 if (memcmp(d_name->name, name->name, name->len))
98 /* XXX: d_name must be in-dentry structure */
99 dchild = container_of(d_name, struct dentry, d_name); /* ugh */
101 CDEBUG(D_DENTRY,"found name %.*s(%p) - flags %d/%x - refc %d\n",
102 name->len, name->name, dchild,
103 d_mountpoint(dchild), dchild->d_flags & DCACHE_LUSTRE_INVALID,
104 atomic_read(&dchild->d_count));
106 /* mountpoint is always valid */
107 if (d_mountpoint(dchild))
110 if (dchild->d_flags & DCACHE_LUSTRE_INVALID)
116 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
119 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
120 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
121 return LDLM_ITER_CONTINUE;
122 return LDLM_ITER_STOP;
125 /* find any ldlm lock of the inode in mdc and lov
129 static int find_cbdata(struct inode *inode)
131 struct ll_inode_info *lli = ll_i2info(inode);
132 struct ll_sb_info *sbi = ll_i2sbi(inode);
137 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
138 return_if_equal, NULL);
143 rc = obd_find_cbdata(sbi->ll_dt_exp, lli->lli_smd,
144 return_if_equal, NULL);
150 * Called when last reference to a dentry is dropped and dcache wants to know
151 * whether or not it should cache it:
152 * - return 1 to delete the dentry immediately
153 * - return 0 to cache the dentry
154 * Should NOT be called with the dcache lock, see fs/dcache.c
156 static int ll_ddelete(struct dentry *de)
161 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
162 (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
163 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
164 d_unhashed(de) ? "" : "hashed,",
165 list_empty(&de->d_subdirs) ? "" : "subdirs");
167 /* if not ldlm lock for this inode, set i_nlink to 0 so that
168 * this inode can be recycled later b=20433 */
169 LASSERT(atomic_read(&de->d_count) == 0);
170 if (de->d_inode && !find_cbdata(de->d_inode))
171 de->d_inode->i_nlink = 0;
173 if (de->d_flags & DCACHE_LUSTRE_INVALID)
179 static int ll_set_dd(struct dentry *de)
184 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
185 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
186 atomic_read(&de->d_count));
188 if (de->d_fsdata == NULL) {
189 struct ll_dentry_data *lld;
192 if (likely(lld != NULL)) {
194 if (likely(de->d_fsdata == NULL))
207 int ll_dops_init(struct dentry *de, int block, int init_sa)
209 struct ll_dentry_data *lld = ll_d2d(de);
212 if (lld == NULL && block != 0) {
220 if (lld != NULL && init_sa != 0)
221 lld->lld_sa_generation = 0;
223 de->d_op = &ll_d_ops;
227 void ll_intent_drop_lock(struct lookup_intent *it)
229 struct lustre_handle *handle;
231 if (it->it_op && it->d.lustre.it_lock_mode) {
232 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
233 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
234 " from it %p\n", handle->cookie, it);
235 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
237 /* bug 494: intent_release may be called multiple times, from
238 * this thread and we don't want to double-decref this lock */
239 it->d.lustre.it_lock_mode = 0;
243 void ll_intent_release(struct lookup_intent *it)
247 CDEBUG(D_INFO, "intent %p released\n", it);
248 ll_intent_drop_lock(it);
249 /* We are still holding extra reference on a request, need to free it */
250 if (it_disposition(it, DISP_ENQ_OPEN_REF))
251 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
252 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
253 ptlrpc_req_finished(it->d.lustre.it_data);
254 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
256 ptlrpc_req_finished(it->d.lustre.it_data);
258 it->d.lustre.it_disposition = 0;
259 it->d.lustre.it_data = NULL;
263 /* Drop dentry if it is not used already, unhash otherwise.
264 Should be called with dcache lock held!
265 Returns: 1 if dentry was dropped, 0 if unhashed. */
266 int ll_drop_dentry(struct dentry *dentry)
269 if (atomic_read(&dentry->d_count) == 0) {
270 CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p "
271 "inode %p\n", dentry->d_name.len,
272 dentry->d_name.name, dentry, dentry->d_parent,
276 unlock_dentry(dentry);
277 spin_unlock(&dcache_lock);
278 cfs_spin_unlock(&ll_lookup_lock);
280 cfs_spin_lock(&ll_lookup_lock);
281 spin_lock(&dcache_lock);
284 /* disconected dentry can not be find without lookup, because we
285 * not need his to unhash or mark invalid. */
286 if (dentry->d_flags & DCACHE_DISCONNECTED) {
287 unlock_dentry(dentry);
291 if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
292 CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p "
293 "inode %p refc %d\n", dentry->d_name.len,
294 dentry->d_name.name, dentry, dentry->d_parent,
295 dentry->d_inode, atomic_read(&dentry->d_count));
296 /* actually we don't unhash the dentry, rather just
297 * mark it inaccessible for to __d_lookup(). otherwise
298 * sys_getcwd() could return -ENOENT -bzzz */
299 dentry->d_flags |= DCACHE_LUSTRE_INVALID;
300 if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
303 unlock_dentry(dentry);
307 void ll_unhash_aliases(struct inode *inode)
309 struct list_head *tmp, *head;
313 CERROR("unexpected NULL inode, tell phil\n");
317 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
318 inode->i_ino, inode->i_generation, inode);
320 head = &inode->i_dentry;
321 cfs_spin_lock(&ll_lookup_lock);
322 spin_lock(&dcache_lock);
325 while ((tmp = tmp->next) != head) {
326 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
328 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
329 "inode %p flags %d\n", dentry->d_name.len,
330 dentry->d_name.name, dentry, dentry->d_parent,
331 dentry->d_inode, dentry->d_flags);
333 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
334 CERROR("called on root (?) dentry=%p, inode=%p "
335 "ino=%lu\n", dentry, inode, inode->i_ino);
336 lustre_dump_dentry(dentry, 1);
337 libcfs_debug_dumpstack(NULL);
340 if (ll_drop_dentry(dentry))
343 spin_unlock(&dcache_lock);
344 cfs_spin_unlock(&ll_lookup_lock);
349 int ll_revalidate_it_finish(struct ptlrpc_request *request,
350 struct lookup_intent *it,
359 if (it_disposition(it, DISP_LOOKUP_NEG))
362 rc = ll_prep_inode(&de->d_inode, request, NULL);
367 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
370 LASSERT(dentry != NULL);
372 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
373 struct inode *inode = dentry->d_inode;
374 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
376 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
377 inode, inode->i_ino, inode->i_generation);
378 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
381 /* drop lookup or getattr locks immediately */
382 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
383 /* on 2.6 there are situation when several lookups and
384 * revalidations may be requested during single operation.
385 * therefore, we don't release intent here -bzzz */
386 ll_intent_drop_lock(it);
390 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
392 struct lookup_intent *it = *itp;
394 if (!it || it->it_op == IT_GETXATTR)
399 int ll_revalidate_it(struct dentry *de, int lookup_flags,
400 struct lookup_intent *it)
402 struct md_op_data *op_data;
403 struct ptlrpc_request *req = NULL;
404 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
405 struct obd_export *exp;
406 struct inode *parent = de->d_parent->d_inode;
410 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
413 if (de->d_inode == NULL) {
416 /* We can only use negative dentries if this is stat or lookup,
417 for opens and stuff we do need to query server. */
418 /* If there is IT_CREAT in intent op set, then we must throw
419 away this negative dentry and actually do the request to
420 kernel to create whatever needs to be created (if possible)*/
421 if (it && (it->it_op & IT_CREAT))
424 if (de->d_flags & DCACHE_LUSTRE_INVALID)
427 ibits = MDS_INODELOCK_UPDATE;
428 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
432 /* Never execute intents for mount points.
433 * Attributes will be fixed up in ll_inode_revalidate_it */
434 if (d_mountpoint(de))
435 GOTO(out_sa, rc = 1);
437 /* need to get attributes in case root got changed from other client */
438 if (de == de->d_sb->s_root) {
439 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
445 exp = ll_i2mdexp(de->d_inode);
447 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
448 ll_frob_intent(&it, &lookup_it);
451 if (it->it_op == IT_LOOKUP && !(de->d_flags & DCACHE_LUSTRE_INVALID))
452 GOTO(out_sa, rc = 1);
454 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
455 de->d_name.name, de->d_name.len,
456 0, LUSTRE_OPC_ANY, NULL);
458 RETURN(PTR_ERR(op_data));
460 if ((it->it_op == IT_OPEN) && de->d_inode) {
461 struct inode *inode = de->d_inode;
462 struct ll_inode_info *lli = ll_i2info(inode);
463 struct obd_client_handle **och_p;
468 * We used to check for MDS_INODELOCK_OPEN here, but in fact
469 * just having LOOKUP lock is enough to justify inode is the
470 * same. And if inode is the same and we have suitable
471 * openhandle, then there is no point in doing another OPEN RPC
472 * just to throw away newly received openhandle. There are no
473 * security implications too, if file owner or access mode is
474 * change, LOOKUP lock is revoked.
478 if (it->it_flags & FMODE_WRITE) {
479 och_p = &lli->lli_mds_write_och;
480 och_usecount = &lli->lli_open_fd_write_count;
481 } else if (it->it_flags & FMODE_EXEC) {
482 och_p = &lli->lli_mds_exec_och;
483 och_usecount = &lli->lli_open_fd_exec_count;
485 och_p = &lli->lli_mds_read_och;
486 och_usecount = &lli->lli_open_fd_read_count;
488 /* Check for the proper lock. */
489 ibits = MDS_INODELOCK_LOOKUP;
490 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
492 cfs_down(&lli->lli_och_sem);
493 if (*och_p) { /* Everything is open already, do nothing */
494 /*(*och_usecount)++; Do not let them steal our open
495 handle from under us */
496 /* XXX The code above was my original idea, but in case
497 we have the handle, but we cannot use it due to later
498 checks (e.g. O_CREAT|O_EXCL flags set), nobody
499 would decrement counter increased here. So we just
500 hope the lock won't be invalidated in between. But
501 if it would be, we'll reopen the open request to
502 MDS later during file open path */
503 cfs_up(&lli->lli_och_sem);
504 ll_finish_md_op_data(op_data);
507 cfs_up(&lli->lli_och_sem);
511 if (it->it_op == IT_GETATTR)
512 first = ll_statahead_enter(parent, &de, 0);
515 it->it_create_mode &= ~cfs_curproc_umask();
516 it->it_create_mode |= M_CHECK_STALE;
517 rc = md_intent_lock(exp, op_data, NULL, 0, it,
519 &req, ll_md_blocking_ast, 0);
520 it->it_create_mode &= ~M_CHECK_STALE;
521 ll_finish_md_op_data(op_data);
522 if (it->it_op == IT_GETATTR && !first)
523 /* If there are too many locks on client-side, then some
524 * locks taken by statahead maybe dropped automatically
525 * before the real "revalidate" using them. */
526 ll_statahead_exit(parent, de, req == NULL ? rc : 0);
527 else if (first == -EEXIST)
528 ll_statahead_mark(parent, de);
530 /* If req is NULL, then md_intent_lock only tried to do a lock match;
531 * if all was well, it will return 1 if it found locks, 0 otherwise. */
532 if (req == NULL && rc >= 0) {
540 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
541 "%d\n", rc, it->d.lustre.it_status);
547 rc = ll_revalidate_it_finish(req, it, de);
549 if (rc != -ESTALE && rc != -ENOENT)
550 ll_intent_release(it);
554 if ((it->it_op & IT_OPEN) && de->d_inode &&
555 !S_ISREG(de->d_inode->i_mode) &&
556 !S_ISDIR(de->d_inode->i_mode)) {
557 ll_release_openhandle(de, it);
561 /* unfortunately ll_intent_lock may cause a callback and revoke our
563 cfs_spin_lock(&ll_lookup_lock);
564 spin_lock(&dcache_lock);
568 d_rehash_cond(de, 0);
569 spin_unlock(&dcache_lock);
570 cfs_spin_unlock(&ll_lookup_lock);
573 /* We do not free request as it may be reused during following lookup
574 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
575 * be freed in ll_lookup_it or in ll_intent_release. But if
576 * request was not completed, we need to free it. (bug 5154, 9903) */
577 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
578 ptlrpc_req_finished(req);
580 ll_unhash_aliases(de->d_inode);
581 /* done in ll_unhash_aliases()
582 dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
586 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
587 "inode %p refc %d\n", de->d_name.len,
588 de->d_name.name, de, de->d_parent, de->d_inode,
589 atomic_read(&de->d_count));
590 ll_set_lock_data(exp, de->d_inode, it, &bits);
591 if (de->d_flags & DCACHE_LUSTRE_INVALID &&
592 bits & MDS_INODELOCK_LOOKUP) {
594 de->d_flags &= ~DCACHE_LUSTRE_INVALID;
597 ll_lookup_finish_locks(it, de);
602 * This part is here to combat evil-evil race in real_lookup on 2.6
603 * kernels. The race details are: We enter do_lookup() looking for some
604 * name, there is nothing in dcache for this name yet and d_lookup()
605 * returns NULL. We proceed to real_lookup(), and while we do this,
606 * another process does open on the same file we looking up (most simple
607 * reproducer), open succeeds and the dentry is added. Now back to
608 * us. In real_lookup() we do d_lookup() again and suddenly find the
609 * dentry, so we call d_revalidate on it, but there is no lock, so
610 * without this code we would return 0, but unpatched real_lookup just
611 * returns -ENOENT in such a case instead of retrying the lookup. Once
612 * this is dealt with in real_lookup(), all of this ugly mess can go and
613 * we can just check locks in ->d_revalidate without doing any RPCs
617 if (it != &lookup_it) {
618 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
619 if (it->it_op == IT_GETATTR)
620 lookup_it.it_op = IT_GETATTR;
621 ll_lookup_finish_locks(it, de);
625 /* Do real lookup here. */
626 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
627 de->d_name.len, 0, (it->it_op & IT_CREAT ?
629 LUSTRE_OPC_ANY), NULL);
631 RETURN(PTR_ERR(op_data));
633 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
634 ll_md_blocking_ast, 0);
636 struct mdt_body *mdt_body;
637 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
638 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
641 fid = *ll_inode2fid(de->d_inode);
643 /* see if we got same inode, if not - return error */
644 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
645 ll_finish_md_op_data(op_data);
647 goto revalidate_finish;
649 ll_intent_release(it);
651 ll_finish_md_op_data(op_data);
656 * For rc == 1 case, should not return directly to prevent losing
657 * statahead windows; for rc == 0 case, the "lookup" will be done later.
659 if (it && it->it_op == IT_GETATTR && rc == 1) {
660 first = ll_statahead_enter(parent, &de, 0);
662 ll_statahead_exit(parent, de, 1);
663 else if (first == -EEXIST)
664 ll_statahead_mark(parent, de);
670 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
675 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
676 struct lookup_intent *it;
678 it = ll_convert_intent(&nd->intent.open, nd->flags);
682 if (it->it_op == (IT_OPEN|IT_CREAT) &&
683 nd->intent.open.flags & O_EXCL) {
684 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
689 rc = ll_revalidate_it(dentry, nd->flags, it);
691 if (rc && (nd->flags & LOOKUP_OPEN) &&
692 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
693 #ifdef HAVE_FILE_IN_STRUCT_INTENT
694 // XXX Code duplication with ll_lookup_nd
695 if (S_ISFIFO(dentry->d_inode->i_mode)) {
696 // We cannot call open here as it would
699 (struct ptlrpc_request *)
700 it->d.lustre.it_data);
702 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,17))
703 /* 2.6.1[456] have a bug in open_namei() that forgets to check
704 * nd->intent.open.file for error, so we need to return it as lookup's result
708 nd->intent.open.file->private_data = it;
709 filp = lookup_instantiate_filp(nd, dentry,NULL);
714 nd->intent.open.file->private_data = it;
715 (void)lookup_instantiate_filp(nd, dentry,NULL);
719 ll_release_openhandle(dentry, it);
720 #endif /* HAVE_FILE_IN_STRUCT_INTENT */
722 if (!rc && (nd->flags & LOOKUP_CREATE) &&
723 it_disposition(it, DISP_OPEN_CREATE)) {
724 /* We created something but we may only return
725 * negative dentry here, so save request in dentry,
726 * if lookup will be called later on, it will
727 * pick the request, otherwise it would be freed
729 ll_d2d(dentry)->lld_it = it;
730 it = NULL; /* avoid freeing */
735 ll_intent_release(it);
736 OBD_FREE(it, sizeof(*it));
739 rc = ll_revalidate_it(dentry, 0, NULL);
745 void ll_d_iput(struct dentry *de, struct inode *inode)
748 if (!find_cbdata(inode))
753 struct dentry_operations ll_d_ops = {
754 .d_revalidate = ll_revalidate_nd,
755 .d_release = ll_release,
756 .d_delete = ll_ddelete,
758 .d_compare = ll_dcompare,