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 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
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>
48 #include <lustre_mdc.h>
49 //#include <lustre_ver.h>
50 //#include <lustre_version.h>
52 #include "llite_internal.h"
54 cfs_spinlock_t ll_lookup_lock = CFS_SPIN_LOCK_UNLOCKED;
56 /* should NOT be called with the dcache lock, see fs/dcache.c */
57 static void ll_release(struct dentry *de)
59 struct ll_dentry_data *lld;
63 if (lld == NULL) { /* NFS copies the de->d_op methods (bug 4655) */
67 #ifndef HAVE_VFS_INTENT_PATCHES
69 ll_intent_release(lld->lld_it);
70 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)
117 /* should NOT be called with the dcache lock, see fs/dcache.c */
118 static int ll_ddelete(struct dentry *de)
123 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
124 (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
125 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
126 d_unhashed(de) ? "" : "hashed,",
127 list_empty(&de->d_subdirs) ? "" : "subdirs");
132 void ll_set_dd(struct dentry *de)
137 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
138 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
139 atomic_read(&de->d_count));
141 if (de->d_fsdata == NULL) {
142 struct ll_dentry_data *lld;
145 if (likely(lld != NULL)) {
147 if (likely(de->d_fsdata == NULL))
158 void ll_intent_drop_lock(struct lookup_intent *it)
160 struct lustre_handle *handle;
162 if (it->it_op && it->d.lustre.it_lock_mode) {
163 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
164 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
165 " from it %p\n", handle->cookie, it);
166 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
168 /* bug 494: intent_release may be called multiple times, from
169 * this thread and we don't want to double-decref this lock */
170 it->d.lustre.it_lock_mode = 0;
174 void ll_intent_release(struct lookup_intent *it)
178 CDEBUG(D_INFO, "intent %p released\n", it);
179 ll_intent_drop_lock(it);
180 #ifdef HAVE_VFS_INTENT_PATCHES
182 it->it_op_release = 0;
184 /* We are still holding extra reference on a request, need to free it */
185 if (it_disposition(it, DISP_ENQ_OPEN_REF))
186 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
187 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
188 ptlrpc_req_finished(it->d.lustre.it_data);
189 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
191 ptlrpc_req_finished(it->d.lustre.it_data);
193 it->d.lustre.it_disposition = 0;
194 it->d.lustre.it_data = NULL;
198 /* Drop dentry if it is not used already, unhash otherwise.
199 Should be called with dcache lock held!
200 Returns: 1 if dentry was dropped, 0 if unhashed. */
201 int ll_drop_dentry(struct dentry *dentry)
204 if (atomic_read(&dentry->d_count) == 0) {
205 CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p "
206 "inode %p\n", dentry->d_name.len,
207 dentry->d_name.name, dentry, dentry->d_parent,
211 unlock_dentry(dentry);
212 spin_unlock(&dcache_lock);
213 cfs_spin_unlock(&ll_lookup_lock);
215 cfs_spin_lock(&ll_lookup_lock);
216 spin_lock(&dcache_lock);
219 /* disconected dentry can not be find without lookup, because we
220 * not need his to unhash or mark invalid. */
221 if (dentry->d_flags & DCACHE_DISCONNECTED) {
222 unlock_dentry(dentry);
226 if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
227 CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p "
228 "inode %p refc %d\n", dentry->d_name.len,
229 dentry->d_name.name, dentry, dentry->d_parent,
230 dentry->d_inode, atomic_read(&dentry->d_count));
231 /* actually we don't unhash the dentry, rather just
232 * mark it inaccessible for to __d_lookup(). otherwise
233 * sys_getcwd() could return -ENOENT -bzzz */
234 dentry->d_flags |= DCACHE_LUSTRE_INVALID;
235 if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
238 unlock_dentry(dentry);
242 void ll_unhash_aliases(struct inode *inode)
244 struct list_head *tmp, *head;
248 CERROR("unexpected NULL inode, tell phil\n");
252 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
253 inode->i_ino, inode->i_generation, inode);
255 head = &inode->i_dentry;
256 cfs_spin_lock(&ll_lookup_lock);
257 spin_lock(&dcache_lock);
260 while ((tmp = tmp->next) != head) {
261 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
263 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
264 "inode %p flags %d\n", dentry->d_name.len,
265 dentry->d_name.name, dentry, dentry->d_parent,
266 dentry->d_inode, dentry->d_flags);
268 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
269 CERROR("called on root (?) dentry=%p, inode=%p "
270 "ino=%lu\n", dentry, inode, inode->i_ino);
271 lustre_dump_dentry(dentry, 1);
272 libcfs_debug_dumpstack(NULL);
275 if (ll_drop_dentry(dentry))
278 spin_unlock(&dcache_lock);
279 cfs_spin_unlock(&ll_lookup_lock);
284 int ll_revalidate_it_finish(struct ptlrpc_request *request,
285 struct lookup_intent *it,
294 if (it_disposition(it, DISP_LOOKUP_NEG))
297 rc = ll_prep_inode(&de->d_inode, request, NULL);
302 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
305 LASSERT(dentry != NULL);
307 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
308 struct inode *inode = dentry->d_inode;
309 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
311 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
312 inode, inode->i_ino, inode->i_generation);
313 md_set_lock_data(sbi->ll_md_exp, &it->d.lustre.it_lock_handle,
317 /* drop lookup or getattr locks immediately */
318 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
319 /* on 2.6 there are situation when several lookups and
320 * revalidations may be requested during single operation.
321 * therefore, we don't release intent here -bzzz */
322 ll_intent_drop_lock(it);
326 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
328 struct lookup_intent *it = *itp;
329 #ifdef HAVE_VFS_INTENT_PATCHES
331 LASSERTF(it->it_magic == INTENT_MAGIC,
332 "%p has bad intent magic: %x\n",
337 if (!it || it->it_op == IT_GETXATTR)
340 #ifdef HAVE_VFS_INTENT_PATCHES
341 it->it_op_release = ll_intent_release;
345 int ll_revalidate_it(struct dentry *de, int lookup_flags,
346 struct lookup_intent *it)
348 struct md_op_data *op_data;
349 struct ptlrpc_request *req = NULL;
350 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
351 struct obd_export *exp;
352 struct inode *parent = de->d_parent->d_inode;
356 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
359 if (de->d_inode == NULL) {
360 /* We can only use negative dentries if this is stat or lookup,
361 for opens and stuff we do need to query server. */
362 /* If there is IT_CREAT in intent op set, then we must throw
363 away this negative dentry and actually do the request to
364 kernel to create whatever needs to be created (if possible)*/
365 if (it && (it->it_op & IT_CREAT))
368 if (de->d_flags & DCACHE_LUSTRE_INVALID)
371 rc = ll_have_md_lock(parent, MDS_INODELOCK_UPDATE);
375 /* Never execute intents for mount points.
376 * Attributes will be fixed up in ll_inode_revalidate_it */
377 if (d_mountpoint(de))
378 GOTO(out_sa, rc = 1);
380 /* need to get attributes in case root got changed from other client */
381 if (de == de->d_sb->s_root) {
382 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
388 exp = ll_i2mdexp(de->d_inode);
390 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
391 ll_frob_intent(&it, &lookup_it);
394 if (it->it_op == IT_LOOKUP && !(de->d_flags & DCACHE_LUSTRE_INVALID))
395 GOTO(out_sa, rc = 1);
397 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
398 de->d_name.name, de->d_name.len,
399 0, LUSTRE_OPC_ANY, NULL);
401 RETURN(PTR_ERR(op_data));
403 if ((it->it_op == IT_OPEN) && de->d_inode) {
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 och_usecount = &lli->lli_open_fd_write_count;
423 } else if (it->it_flags & FMODE_EXEC) {
424 och_p = &lli->lli_mds_exec_och;
425 och_usecount = &lli->lli_open_fd_exec_count;
427 och_p = &lli->lli_mds_read_och;
428 och_usecount = &lli->lli_open_fd_read_count;
430 /* Check for the proper lock. */
431 if (!ll_have_md_lock(inode, MDS_INODELOCK_LOOKUP))
433 cfs_down(&lli->lli_och_sem);
434 if (*och_p) { /* Everything is open already, do nothing */
435 /*(*och_usecount)++; Do not let them steal our open
436 handle from under us */
437 /* XXX The code above was my original idea, but in case
438 we have the handle, but we cannot use it due to later
439 checks (e.g. O_CREAT|O_EXCL flags set), nobody
440 would decrement counter increased here. So we just
441 hope the lock won't be invalidated in between. But
442 if it would be, we'll reopen the open request to
443 MDS later during file open path */
444 cfs_up(&lli->lli_och_sem);
445 ll_finish_md_op_data(op_data);
448 cfs_up(&lli->lli_och_sem);
452 if (it->it_op == IT_GETATTR)
453 first = ll_statahead_enter(parent, &de, 0);
456 it->it_create_mode &= ~current->fs->umask;
457 it->it_create_mode |= M_CHECK_STALE;
458 rc = md_intent_lock(exp, op_data, NULL, 0, it,
460 &req, ll_md_blocking_ast, 0);
461 it->it_create_mode &= ~M_CHECK_STALE;
462 ll_finish_md_op_data(op_data);
463 if (it->it_op == IT_GETATTR && !first)
464 /* If there are too many locks on client-side, then some
465 * locks taken by statahead maybe dropped automatically
466 * before the real "revalidate" using them. */
467 ll_statahead_exit(parent, de, req == NULL ? rc : 0);
468 else if (first == -EEXIST)
469 ll_statahead_mark(parent, de);
471 /* If req is NULL, then md_intent_lock only tried to do a lock match;
472 * if all was well, it will return 1 if it found locks, 0 otherwise. */
473 if (req == NULL && rc >= 0) {
481 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
482 "%d\n", rc, it->d.lustre.it_status);
488 rc = ll_revalidate_it_finish(req, it, de);
490 if (rc != -ESTALE && rc != -ENOENT)
491 ll_intent_release(it);
495 if ((it->it_op & IT_OPEN) && de->d_inode &&
496 !S_ISREG(de->d_inode->i_mode) &&
497 !S_ISDIR(de->d_inode->i_mode)) {
498 ll_release_openhandle(de, it);
502 /* unfortunately ll_intent_lock may cause a callback and revoke our
504 cfs_spin_lock(&ll_lookup_lock);
505 spin_lock(&dcache_lock);
509 d_rehash_cond(de, 0);
510 spin_unlock(&dcache_lock);
511 cfs_spin_unlock(&ll_lookup_lock);
514 /* We do not free request as it may be reused during following lookup
515 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
516 * be freed in ll_lookup_it or in ll_intent_release. But if
517 * request was not completed, we need to free it. (bug 5154, 9903) */
518 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
519 ptlrpc_req_finished(req);
521 ll_unhash_aliases(de->d_inode);
522 /* done in ll_unhash_aliases()
523 dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
525 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
526 "inode %p refc %d\n", de->d_name.len,
527 de->d_name.name, de, de->d_parent, de->d_inode,
528 atomic_read(&de->d_count));
529 ll_lookup_finish_locks(it, de);
531 de->d_flags &= ~DCACHE_LUSTRE_INVALID;
537 * This part is here to combat evil-evil race in real_lookup on 2.6
538 * kernels. The race details are: We enter do_lookup() looking for some
539 * name, there is nothing in dcache for this name yet and d_lookup()
540 * returns NULL. We proceed to real_lookup(), and while we do this,
541 * another process does open on the same file we looking up (most simple
542 * reproducer), open succeeds and the dentry is added. Now back to
543 * us. In real_lookup() we do d_lookup() again and suddenly find the
544 * dentry, so we call d_revalidate on it, but there is no lock, so
545 * without this code we would return 0, but unpatched real_lookup just
546 * returns -ENOENT in such a case instead of retrying the lookup. Once
547 * this is dealt with in real_lookup(), all of this ugly mess can go and
548 * we can just check locks in ->d_revalidate without doing any RPCs
552 if (it != &lookup_it) {
553 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
554 if (it->it_op == IT_GETATTR)
555 lookup_it.it_op = IT_GETATTR;
556 ll_lookup_finish_locks(it, de);
560 /* Do real lookup here. */
561 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
562 de->d_name.len, 0, (it->it_op & IT_CREAT ?
564 LUSTRE_OPC_ANY), NULL);
566 RETURN(PTR_ERR(op_data));
568 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
569 ll_md_blocking_ast, 0);
571 struct mdt_body *mdt_body;
572 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
573 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
576 fid = *ll_inode2fid(de->d_inode);
578 /* see if we got same inode, if not - return error */
579 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
580 ll_finish_md_op_data(op_data);
582 goto revalidate_finish;
584 ll_intent_release(it);
586 ll_finish_md_op_data(op_data);
591 * For rc == 1 case, should not return directly to prevent losing
592 * statahead windows; for rc == 0 case, the "lookup" will be done later.
594 if (it && it->it_op == IT_GETATTR && rc == 1) {
595 first = ll_statahead_enter(parent, &de, 0);
597 ll_statahead_exit(parent, de, 1);
598 else if (first == -EEXIST)
599 ll_statahead_mark(parent, de);
606 static void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
608 struct inode *inode= de->d_inode;
609 struct ll_sb_info *sbi = ll_i2sbi(inode);
610 struct ll_dentry_data *ldd = ll_d2d(de);
611 struct obd_client_handle *handle;
618 /* Strictly speaking this introduces an additional race: the
619 * increments should wait until the rpc has returned.
620 * However, given that at present the function is void, this
622 if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
628 if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
635 handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
636 oc = ll_mdscapa_get(inode);
637 rc = obd_pin(sbi->ll_md_exp, ll_inode2fid(inode), oc, handle, flag);
641 memset(handle, 0, sizeof(*handle));
643 ldd->lld_cwd_count--;
645 ldd->lld_mnt_count--;
653 static void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
655 struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
656 struct ll_dentry_data *ldd = ll_d2d(de);
657 struct obd_client_handle handle;
663 /* Strictly speaking this introduces an additional race: the
664 * increments should wait until the rpc has returned.
665 * However, given that at present the function is void, this
667 handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
668 if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
669 /* the "pin" failed */
676 count = --ldd->lld_mnt_count;
678 count = --ldd->lld_cwd_count;
686 rc = obd_unpin(sbi->ll_md_exp, &handle, flag);
692 #ifdef HAVE_VFS_INTENT_PATCHES
693 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
698 if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
699 rc = ll_revalidate_it(dentry, nd->flags, &nd->intent);
701 rc = ll_revalidate_it(dentry, 0, NULL);
706 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
711 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
712 struct lookup_intent *it;
713 it = ll_convert_intent(&nd->intent.open, nd->flags);
716 if (it->it_op == (IT_OPEN|IT_CREAT))
717 if (nd->intent.open.flags & O_EXCL) {
718 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
723 rc = ll_revalidate_it(dentry, nd->flags, it);
725 if (rc && (nd->flags & LOOKUP_OPEN) &&
726 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
727 #ifdef HAVE_FILE_IN_STRUCT_INTENT
728 // XXX Code duplication with ll_lookup_nd
729 if (S_ISFIFO(dentry->d_inode->i_mode)) {
730 // We cannot call open here as it would
733 (struct ptlrpc_request *)
734 it->d.lustre.it_data);
736 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,17))
737 /* 2.6.1[456] have a bug in open_namei() that forgets to check
738 * nd->intent.open.file for error, so we need to return it as lookup's result
742 nd->intent.open.file->private_data = it;
743 filp = lookup_instantiate_filp(nd, dentry,NULL);
748 nd->intent.open.file->private_data = it;
749 (void)lookup_instantiate_filp(nd, dentry,NULL);
753 ll_release_openhandle(dentry, it);
754 #endif /* HAVE_FILE_IN_STRUCT_INTENT */
756 if (!rc && (nd->flags & LOOKUP_CREATE) &&
757 it_disposition(it, DISP_OPEN_CREATE)) {
758 /* We created something but we may only return
759 * negative dentry here, so save request in dentry,
760 * if lookup will be called later on, it will
761 * pick the request, otherwise it would be freed
763 ll_d2d(dentry)->lld_it = it;
764 it = NULL; /* avoid freeing */
769 ll_intent_release(it);
770 OBD_FREE(it, sizeof(*it));
773 rc = ll_revalidate_it(dentry, 0, NULL);
780 struct dentry_operations ll_d_ops = {
781 .d_revalidate = ll_revalidate_nd,
782 .d_release = ll_release,
783 .d_delete = ll_ddelete,
784 .d_compare = ll_dcompare,