1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #include <linux/sched.h>
24 #include <linux/smp_lock.h>
25 #include <linux/quotaops.h>
27 #define DEBUG_SUBSYSTEM S_LLITE
29 #include <obd_support.h>
30 #include <lustre_lite.h>
31 #include <lustre/lustre_idl.h>
32 #include <lustre_dlm.h>
33 #include <lustre_mdc.h>
34 //#include <lustre_ver.h>
35 //#include <lustre_version.h>
37 #include "llite_internal.h"
39 /* should NOT be called with the dcache lock, see fs/dcache.c */
40 static void ll_release(struct dentry *de)
42 struct ll_dentry_data *lld;
46 if (lld == NULL) { /* NFS copies the de->d_op methods (bug 4655) */
50 #ifndef LUSTRE_KERNEL_VERSION
52 ll_intent_release(lld->lld_it);
53 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
56 LASSERT(lld->lld_cwd_count == 0);
57 LASSERT(lld->lld_mnt_count == 0);
58 OBD_FREE(de->d_fsdata, sizeof(*lld));
63 #ifdef LUSTRE_KERNEL_VERSION
64 /* Compare if two dentries are the same. Don't match if the existing dentry
65 * is marked DCACHE_LUSTRE_INVALID. Returns 1 if different, 0 if the same.
67 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
68 * an AST before calling d_revalidate_it(). The dentry still exists (marked
69 * INVALID) so d_lookup() matches it, but we have no lock on it (so
70 * lock_match() fails) and we spin around real_lookup(). */
71 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
73 struct dentry *dchild;
76 if (d_name->len != name->len)
79 if (memcmp(d_name->name, name->name, name->len))
82 /* XXX: d_name must be in-dentry structure */
83 dchild = container_of(d_name, struct dentry, d_name); /* ugh */
84 if (dchild->d_flags & DCACHE_LUSTRE_INVALID) {
85 CDEBUG(D_DENTRY,"INVALID dentry %p not matched, was bug 3784\n",
94 /* should NOT be called with the dcache lock, see fs/dcache.c */
95 static int ll_ddelete(struct dentry *de)
99 #ifndef DCACHE_LUSTRE_INVALID
100 #define DCACHE_LUSTRE_INVALID 0
103 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
104 (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
105 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
106 d_unhashed(de) ? "" : "hashed,",
107 list_empty(&de->d_subdirs) ? "" : "subdirs");
108 #if DCACHE_LUSTRE_INVALID == 0
109 #undef DCACHE_LUSTRE_INVALID
115 void ll_set_dd(struct dentry *de)
120 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
121 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
122 atomic_read(&de->d_count));
124 if (de->d_fsdata == NULL) {
125 OBD_ALLOC(de->d_fsdata, sizeof(struct ll_dentry_data));
132 void ll_intent_drop_lock(struct lookup_intent *it)
134 struct lustre_handle *handle;
136 if (it->it_op && it->d.lustre.it_lock_mode) {
137 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
138 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
139 " from it %p\n", handle->cookie, it);
140 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
142 /* bug 494: intent_release may be called multiple times, from
143 * this thread and we don't want to double-decref this lock */
144 it->d.lustre.it_lock_mode = 0;
148 void ll_intent_release(struct lookup_intent *it)
152 CDEBUG(D_INFO, "intent %p released\n", it);
153 ll_intent_drop_lock(it);
154 #ifdef LUSTRE_KERNEL_VERSION
156 it->it_op_release = 0;
158 /* We are still holding extra reference on a request, need to free it */
159 if (it_disposition(it, DISP_ENQ_OPEN_REF)) /* open req for llfile_open*/
160 ptlrpc_req_finished(it->d.lustre.it_data);
161 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
162 ptlrpc_req_finished(it->d.lustre.it_data);
163 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
165 ptlrpc_req_finished(it->d.lustre.it_data);
167 it->d.lustre.it_disposition = 0;
168 it->d.lustre.it_data = NULL;
172 /* Drop dentry if it is not used already, unhash otherwise.
173 Should be called with dcache lock held!
174 Returns: 1 if dentry was dropped, 0 if unhashed. */
175 int ll_drop_dentry(struct dentry *dentry)
178 if (atomic_read(&dentry->d_count) == 0) {
179 CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p "
180 "inode %p\n", dentry->d_name.len,
181 dentry->d_name.name, dentry, dentry->d_parent,
185 unlock_dentry(dentry);
186 spin_unlock(&dcache_lock);
188 spin_lock(&dcache_lock);
191 /* disconected dentry can not be find without lookup, because we
192 * not need his to unhash or mark invalid. */
193 if (dentry->d_flags & DCACHE_DISCONNECTED) {
194 unlock_dentry(dentry);
198 #ifdef LUSTRE_KERNEL_VERSION
199 if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
201 if (!d_unhashed(dentry)) {
203 CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p "
204 "inode %p refc %d\n", dentry->d_name.len,
205 dentry->d_name.name, dentry, dentry->d_parent,
206 dentry->d_inode, atomic_read(&dentry->d_count));
207 /* actually we don't unhash the dentry, rather just
208 * mark it inaccessible for to __d_lookup(). otherwise
209 * sys_getcwd() could return -ENOENT -bzzz */
210 #ifdef LUSTRE_KERNEL_VERSION
211 dentry->d_flags |= DCACHE_LUSTRE_INVALID;
213 if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
217 unlock_dentry(dentry);
221 void ll_unhash_aliases(struct inode *inode)
223 struct list_head *tmp, *head;
227 CERROR("unexpected NULL inode, tell phil\n");
231 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
232 inode->i_ino, inode->i_generation, inode);
234 head = &inode->i_dentry;
235 spin_lock(&dcache_lock);
238 while ((tmp = tmp->next) != head) {
239 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
241 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
242 "inode %p flags %d\n", dentry->d_name.len,
243 dentry->d_name.name, dentry, dentry->d_parent,
244 dentry->d_inode, dentry->d_flags);
246 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
247 CERROR("called on root (?) dentry=%p, inode=%p "
248 "ino=%lu\n", dentry, inode, inode->i_ino);
249 lustre_dump_dentry(dentry, 1);
250 libcfs_debug_dumpstack(NULL);
251 } else if (d_mountpoint(dentry)) {
252 /* For mountpoints we skip removal of the dentry
253 which happens solely because we have a lock on it
254 obtained when this dentry was not a mountpoint yet */
255 CDEBUG(D_DENTRY, "Skippind mountpoint dentry removal "
256 "%.*s (%p) parent %p\n",
259 dentry, dentry->d_parent);
264 if (ll_drop_dentry(dentry))
267 spin_unlock(&dcache_lock);
271 int ll_revalidate_it_finish(struct ptlrpc_request *request,
272 int offset, struct lookup_intent *it,
281 if (it_disposition(it, DISP_LOOKUP_NEG))
284 rc = ll_prep_inode(&de->d_inode,
285 request, offset, NULL);
290 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
293 LASSERT(dentry != NULL);
295 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
296 struct inode *inode = dentry->d_inode;
297 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
299 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
300 inode, inode->i_ino, inode->i_generation);
301 md_set_lock_data(sbi->ll_md_exp, &it->d.lustre.it_lock_handle,
305 /* drop lookup or getattr locks immediately */
306 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
307 /* on 2.6 there are situation when several lookups and
308 * revalidations may be requested during single operation.
309 * therefore, we don't release intent here -bzzz */
310 ll_intent_drop_lock(it);
314 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
316 struct lookup_intent *it = *itp;
317 #if defined(LUSTRE_KERNEL_VERSION)
319 LASSERTF(it->it_magic == INTENT_MAGIC,
320 "%p has bad intent magic: %x\n",
325 if (!it || it->it_op == IT_GETXATTR)
328 #ifdef LUSTRE_KERNEL_VERSION
329 it->it_op_release = ll_intent_release;
333 int ll_revalidate_it(struct dentry *de, int lookup_flags,
334 struct lookup_intent *it)
337 struct md_op_data *op_data;
338 struct ptlrpc_request *req = NULL;
339 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
340 struct obd_export *exp;
341 struct inode *parent;
344 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
347 if (de->d_inode == NULL) {
348 /* We can only use negative dentries if this is stat or lookup,
349 for opens and stuff we do need to query server. */
350 /* If there is IT_CREAT in intent op set, then we must throw
351 away this negative dentry and actually do the request to
352 kernel to create whatever needs to be created (if possible)*/
353 if (it && (it->it_op & IT_CREAT))
356 #ifdef LUSTRE_KERNEL_VERSION
357 if (de->d_flags & DCACHE_LUSTRE_INVALID)
361 rc = ll_have_md_lock(de->d_parent->d_inode,
362 MDS_INODELOCK_UPDATE);
367 exp = ll_i2mdexp(de->d_inode);
369 /* Never execute intents for mount points.
370 * Attributes will be fixed up in ll_inode_revalidate_it */
371 if (d_mountpoint(de))
374 /* Root of the lustre tree. Always valid.
375 * Attributes will be fixed up in ll_inode_revalidate_it */
376 if (de == de->d_sb->s_root)
379 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
380 ll_frob_intent(&it, &lookup_it);
383 parent = de->d_parent->d_inode;
385 if (it->it_op & IT_CREAT) {
386 op_data = ll_prep_md_op_data(NULL, parent, NULL,
387 de->d_name.name, de->d_name.len,
388 0, LUSTRE_OPC_CREATE, NULL);
390 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
391 de->d_name.name, de->d_name.len,
392 0, LUSTRE_OPC_ANY, NULL);
395 RETURN(PTR_ERR(op_data));
398 if ((it->it_op == IT_OPEN) && de->d_inode) {
399 struct inode *inode = de->d_inode;
400 struct ll_inode_info *lli = ll_i2info(inode);
401 struct obd_client_handle **och_p;
405 * We used to check for MDS_INODELOCK_OPEN here, but in fact
406 * just having LOOKUP lock is enough to justify inode is the
407 * same. And if inode is the same and we have suitable
408 * openhandle, then there is no point in doing another OPEN RPC
409 * just to throw away newly received openhandle. There are no
410 * security implications too, if file owner or access mode is
411 * change, LOOKUP lock is revoked.
415 if (it->it_flags & FMODE_WRITE) {
416 och_p = &lli->lli_mds_write_och;
417 och_usecount = &lli->lli_open_fd_write_count;
418 } else if (it->it_flags & FMODE_EXEC) {
419 och_p = &lli->lli_mds_exec_och;
420 och_usecount = &lli->lli_open_fd_exec_count;
422 och_p = &lli->lli_mds_read_och;
423 och_usecount = &lli->lli_open_fd_read_count;
425 /* Check for the proper lock. */
426 if (!ll_have_md_lock(inode, MDS_INODELOCK_LOOKUP))
428 down(&lli->lli_och_sem);
429 if (*och_p) { /* Everything is open already, do nothing */
430 /*(*och_usecount)++; Do not let them steal our open
431 handle from under us */
432 /* XXX The code above was my original idea, but in case
433 we have the handle, but we cannot use it due to later
434 checks (e.g. O_CREAT|O_EXCL flags set), nobody
435 would decrement counter increased here. So we just
436 hope the lock won't be invalidated in between. But
437 if it would be, we'll reopen the open request to
438 MDS later during file open path */
439 up(&lli->lli_och_sem);
440 ll_finish_md_op_data(op_data);
443 up(&lli->lli_och_sem);
448 it->it_create_mode &= ~current->fs->umask;
449 it->it_flags |= O_CHECK_STALE;
450 rc = md_intent_lock(exp, op_data, NULL, 0, it,
452 &req, ll_md_blocking_ast, 0);
453 it->it_flags &= ~O_CHECK_STALE;
454 ll_finish_md_op_data(op_data);
455 /* If req is NULL, then md_intent_lock only tried to do a lock match;
456 * if all was well, it will return 1 if it found locks, 0 otherwise. */
457 if (req == NULL && rc >= 0) {
465 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
466 "%d\n", rc, it->d.lustre.it_status);
472 rc = ll_revalidate_it_finish(req, DLM_REPLY_REC_OFF, it, de);
474 if (rc != -ESTALE && rc != -ENOENT)
475 ll_intent_release(it);
479 if ((it->it_op & IT_OPEN) && de->d_inode &&
480 !S_ISREG(de->d_inode->i_mode) &&
481 !S_ISDIR(de->d_inode->i_mode)) {
482 ll_release_openhandle(de, it);
486 /* unfortunately ll_intent_lock may cause a callback and revoke our
488 spin_lock(&dcache_lock);
493 spin_unlock(&dcache_lock);
496 /* We do not free request as it may be reused during following lookup
497 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
498 * be freed in ll_lookup_it or in ll_intent_release. But if
499 * request was not completed, we need to free it. (bug 5154, 9903) */
500 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
501 ptlrpc_req_finished(req);
503 #ifdef LUSTRE_KERNEL_VERSION
504 ll_unhash_aliases(de->d_inode);
505 /* done in ll_unhash_aliases()
506 dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
508 /* We do not want d_invalidate to kill all child dentries too */
512 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
513 "inode %p refc %d\n", de->d_name.len,
514 de->d_name.name, de, de->d_parent, de->d_inode,
515 atomic_read(&de->d_count));
516 ll_lookup_finish_locks(it, de);
517 #ifdef LUSTRE_KERNEL_VERSION
519 de->d_flags &= ~DCACHE_LUSTRE_INVALID;
526 * This part is here to combat evil-evil race in real_lookup on 2.6
527 * kernels. The race details are: We enter do_lookup() looking for some
528 * name, there is nothing in dcache for this name yet and d_lookup()
529 * returns NULL. We proceed to real_lookup(), and while we do this,
530 * another process does open on the same file we looking up (most simple
531 * reproducer), open succeeds and the dentry is added. Now back to
532 * us. In real_lookup() we do d_lookup() again and suddenly find the
533 * dentry, so we call d_revalidate on it, but there is no lock, so
534 * without this code we would return 0, but unpatched real_lookup just
535 * returns -ENOENT in such a case instead of retrying the lookup. Once
536 * this is dealt with in real_lookup(), all of this ugly mess can go and
537 * we can just check locks in ->d_revalidate without doing any RPCs
541 if (it != &lookup_it) {
542 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
543 if (it->it_op == IT_GETATTR)
544 lookup_it.it_op = IT_GETATTR;
545 ll_lookup_finish_locks(it, de);
549 /* Do real lookup here. */
550 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
551 de->d_name.len, 0, (it->it_op & IT_CREAT ?
553 LUSTRE_OPC_ANY), NULL);
555 RETURN(PTR_ERR(op_data));
557 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
558 ll_md_blocking_ast, 0);
560 struct mdt_body *mdt_body = lustre_msg_buf(req->rq_repmsg,
563 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
566 fid = *ll_inode2fid(de->d_inode);
568 /* see if we got same inode, if not - return error */
569 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
570 ll_finish_md_op_data(op_data);
572 goto revalidate_finish;
574 ll_intent_release(it);
576 ll_finish_md_op_data(op_data);
580 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
582 struct inode *inode= de->d_inode;
583 struct ll_sb_info *sbi = ll_i2sbi(inode);
584 struct ll_dentry_data *ldd = ll_d2d(de);
585 struct obd_client_handle *handle;
592 /* Strictly speaking this introduces an additional race: the
593 * increments should wait until the rpc has returned.
594 * However, given that at present the function is void, this
596 if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
602 if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
609 handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
610 oc = ll_mdscapa_get(inode);
611 rc = obd_pin(sbi->ll_md_exp, ll_inode2fid(inode), oc, handle, flag);
615 memset(handle, 0, sizeof(*handle));
617 ldd->lld_cwd_count--;
619 ldd->lld_mnt_count--;
627 /*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
629 struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
630 struct ll_dentry_data *ldd = ll_d2d(de);
631 struct obd_client_handle handle;
637 /* Strictly speaking this introduces an additional race: the
638 * increments should wait until the rpc has returned.
639 * However, given that at present the function is void, this
641 handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
642 if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
643 /* the "pin" failed */
650 count = --ldd->lld_mnt_count;
652 count = --ldd->lld_cwd_count;
660 rc = obd_unpin(sbi->ll_md_exp, &handle, flag);
665 #ifdef LUSTRE_KERNEL_VERSION
666 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
671 if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
672 rc = ll_revalidate_it(dentry, nd->flags, &nd->intent);
674 rc = ll_revalidate_it(dentry, 0, NULL);
679 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
684 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
685 struct lookup_intent *it;
686 it = ll_convert_intent(&nd->intent.open, nd->flags);
689 if (it->it_op == (IT_OPEN|IT_CREAT))
690 if (nd->intent.open.flags & O_EXCL) {
691 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
696 rc = ll_revalidate_it(dentry, nd->flags, it);
698 if (rc && (nd->flags & LOOKUP_OPEN) &&
699 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
700 #ifdef HAVE_FILE_IN_STRUCT_INTENT
701 // XXX Code duplication with ll_lookup_nd
702 if (S_ISFIFO(dentry->d_inode->i_mode)) {
703 // We cannot call open here as it would
706 (struct ptlrpc_request *)
707 it->d.lustre.it_data);
711 nd->intent.open.file->private_data = it;
712 filp = lookup_instantiate_filp(nd, dentry,NULL);
713 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,17))
714 /* 2.6.1[456] have a bug in open_namei() that forgets to check
715 * nd->intent.open.file for error, so we need to return it as lookup's result
722 ll_release_openhandle(dentry, it);
723 #endif /* HAVE_FILE_IN_STRUCT_INTENT */
725 if (!rc && (nd->flags & LOOKUP_CREATE) &&
726 it_disposition(it, DISP_OPEN_CREATE)) {
727 /* We created something but we may only return
728 * negative dentry here, so save request in dentry,
729 * if lookup will be called later on, it will
730 * pick the request, otherwise it would be freed
732 ll_d2d(dentry)->lld_it = it;
733 it = NULL; /* avoid freeing */
738 ll_intent_release(it);
739 OBD_FREE(it, sizeof(*it));
742 rc = ll_revalidate_it(dentry, 0, NULL);
749 struct dentry_operations ll_d_ops = {
750 .d_revalidate = ll_revalidate_nd,
751 .d_release = ll_release,
752 .d_delete = ll_ddelete,
753 #ifdef LUSTRE_KERNEL_VERSION
754 .d_compare = ll_dcompare,