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.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
40 #include <linux/sched.h>
41 #include <linux/smp_lock.h>
42 #include <linux/quotaops.h>
44 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <obd_support.h>
47 #include <lustre_lite.h>
48 #include <lustre/lustre_idl.h>
49 #include <lustre_dlm.h>
50 //#include <lustre_ver.h>
51 //#include <lustre_version.h>
53 #include "llite_internal.h"
55 cfs_spinlock_t ll_lookup_lock = CFS_SPIN_LOCK_UNLOCKED;
57 /* should NOT be called with the dcache lock, see fs/dcache.c */
58 static void ll_release(struct dentry *de)
60 struct ll_dentry_data *lld;
64 if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
68 ll_intent_release(lld->lld_it);
69 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
71 LASSERT(lld->lld_cwd_count == 0);
72 LASSERT(lld->lld_mnt_count == 0);
73 OBD_FREE(de->d_fsdata, sizeof(*lld));
78 /* Compare if two dentries are the same. Don't match if the existing dentry
79 * is marked DCACHE_LUSTRE_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(struct dentry *parent, struct qstr *d_name, struct qstr *name)
87 struct dentry *dchild;
90 if (d_name->len != name->len)
93 if (memcmp(d_name->name, name->name, name->len))
96 /* XXX: d_name must be in-dentry structure */
97 dchild = container_of(d_name, struct dentry, d_name); /* ugh */
99 CDEBUG(D_DENTRY,"found name %.*s(%p) - flags %d/%x - refc %d\n",
100 name->len, name->name, dchild,
101 d_mountpoint(dchild), dchild->d_flags & DCACHE_LUSTRE_INVALID,
102 atomic_read(&dchild->d_count));
104 /* mountpoint is always valid */
105 if (d_mountpoint(dchild))
108 if (dchild->d_flags & DCACHE_LUSTRE_INVALID)
114 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
117 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
118 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
119 return LDLM_ITER_CONTINUE;
120 return LDLM_ITER_STOP;
123 /* find any ldlm lock of the inode in mdc and lov
127 static int find_cbdata(struct inode *inode)
129 struct ll_inode_info *lli = ll_i2info(inode);
130 struct ll_sb_info *sbi = ll_i2sbi(inode);
135 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
136 return_if_equal, NULL);
141 rc = obd_find_cbdata(sbi->ll_dt_exp, lli->lli_smd,
142 return_if_equal, NULL);
148 * Called when last reference to a dentry is dropped and dcache wants to know
149 * whether or not it should cache it:
150 * - return 1 to delete the dentry immediately
151 * - return 0 to cache the dentry
152 * Should NOT be called with the dcache lock, see fs/dcache.c
154 static int ll_ddelete(struct dentry *de)
159 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
160 (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
161 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
162 d_unhashed(de) ? "" : "hashed,",
163 list_empty(&de->d_subdirs) ? "" : "subdirs");
165 /* if not ldlm lock for this inode, set i_nlink to 0 so that
166 * this inode can be recycled later b=20433 */
167 LASSERT(atomic_read(&de->d_count) == 0);
168 if (de->d_inode && !find_cbdata(de->d_inode))
169 de->d_inode->i_nlink = 0;
171 if (de->d_flags & DCACHE_LUSTRE_INVALID)
177 static int ll_set_dd(struct dentry *de)
182 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
183 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
184 atomic_read(&de->d_count));
186 if (de->d_fsdata == NULL) {
187 struct ll_dentry_data *lld;
190 if (likely(lld != NULL)) {
192 if (likely(de->d_fsdata == NULL))
205 int ll_dops_init(struct dentry *de, int block, int init_sa)
207 struct ll_dentry_data *lld = ll_d2d(de);
210 if (lld == NULL && block != 0) {
218 if (lld != NULL && init_sa != 0)
219 lld->lld_sa_generation = 0;
221 de->d_op = &ll_d_ops;
225 void ll_intent_drop_lock(struct lookup_intent *it)
227 struct lustre_handle *handle;
229 if (it->it_op && it->d.lustre.it_lock_mode) {
230 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
231 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
232 " from it %p\n", handle->cookie, it);
233 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
235 /* bug 494: intent_release may be called multiple times, from
236 * this thread and we don't want to double-decref this lock */
237 it->d.lustre.it_lock_mode = 0;
241 void ll_intent_release(struct lookup_intent *it)
245 CDEBUG(D_INFO, "intent %p released\n", it);
246 ll_intent_drop_lock(it);
247 /* We are still holding extra reference on a request, need to free it */
248 if (it_disposition(it, DISP_ENQ_OPEN_REF))
249 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
250 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
251 ptlrpc_req_finished(it->d.lustre.it_data);
252 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
254 ptlrpc_req_finished(it->d.lustre.it_data);
256 it->d.lustre.it_disposition = 0;
257 it->d.lustre.it_data = NULL;
261 /* Drop dentry if it is not used already, unhash otherwise.
262 Should be called with dcache lock held!
263 Returns: 1 if dentry was dropped, 0 if unhashed. */
264 int ll_drop_dentry(struct dentry *dentry)
267 if (atomic_read(&dentry->d_count) == 0) {
268 CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p "
269 "inode %p\n", dentry->d_name.len,
270 dentry->d_name.name, dentry, dentry->d_parent,
274 unlock_dentry(dentry);
275 spin_unlock(&dcache_lock);
276 cfs_spin_unlock(&ll_lookup_lock);
278 cfs_spin_lock(&ll_lookup_lock);
279 spin_lock(&dcache_lock);
282 /* disconected dentry can not be find without lookup, because we
283 * not need his to unhash or mark invalid. */
284 if (dentry->d_flags & DCACHE_DISCONNECTED) {
285 unlock_dentry(dentry);
289 if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
290 CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p "
291 "inode %p refc %d\n", dentry->d_name.len,
292 dentry->d_name.name, dentry, dentry->d_parent,
293 dentry->d_inode, atomic_read(&dentry->d_count));
294 /* actually we don't unhash the dentry, rather just
295 * mark it inaccessible for to __d_lookup(). otherwise
296 * sys_getcwd() could return -ENOENT -bzzz */
297 dentry->d_flags |= DCACHE_LUSTRE_INVALID;
298 if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
301 unlock_dentry(dentry);
305 void ll_unhash_aliases(struct inode *inode)
307 struct list_head *tmp, *head;
311 CERROR("unexpected NULL inode, tell phil\n");
315 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
316 inode->i_ino, inode->i_generation, inode);
318 head = &inode->i_dentry;
319 cfs_spin_lock(&ll_lookup_lock);
320 spin_lock(&dcache_lock);
323 while ((tmp = tmp->next) != head) {
324 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
326 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
327 "inode %p flags %d\n", dentry->d_name.len,
328 dentry->d_name.name, dentry, dentry->d_parent,
329 dentry->d_inode, dentry->d_flags);
331 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
332 CERROR("called on root (?) dentry=%p, inode=%p "
333 "ino=%lu\n", dentry, inode, inode->i_ino);
334 lustre_dump_dentry(dentry, 1);
335 libcfs_debug_dumpstack(NULL);
338 if (ll_drop_dentry(dentry))
341 spin_unlock(&dcache_lock);
342 cfs_spin_unlock(&ll_lookup_lock);
347 int ll_revalidate_it_finish(struct ptlrpc_request *request,
348 struct lookup_intent *it,
357 if (it_disposition(it, DISP_LOOKUP_NEG))
360 rc = ll_prep_inode(&de->d_inode, request, NULL);
365 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
368 LASSERT(dentry != NULL);
370 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
371 struct inode *inode = dentry->d_inode;
372 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
374 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
375 inode, inode->i_ino, inode->i_generation);
376 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
379 /* drop lookup or getattr locks immediately */
380 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
381 /* on 2.6 there are situation when several lookups and
382 * revalidations may be requested during single operation.
383 * therefore, we don't release intent here -bzzz */
384 ll_intent_drop_lock(it);
388 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
390 struct lookup_intent *it = *itp;
392 if (!it || it->it_op == IT_GETXATTR)
397 int ll_revalidate_it(struct dentry *de, int lookup_flags,
398 struct lookup_intent *it)
400 struct md_op_data *op_data;
401 struct ptlrpc_request *req = NULL;
402 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
403 struct obd_export *exp;
404 struct inode *parent = de->d_parent->d_inode;
408 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
411 if (de->d_inode == NULL) {
414 /* We can only use negative dentries if this is stat or lookup,
415 for opens and stuff we do need to query server. */
416 /* If there is IT_CREAT in intent op set, then we must throw
417 away this negative dentry and actually do the request to
418 kernel to create whatever needs to be created (if possible)*/
419 if (it && (it->it_op & IT_CREAT))
422 if (de->d_flags & DCACHE_LUSTRE_INVALID)
425 ibits = MDS_INODELOCK_UPDATE;
426 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
430 /* Never execute intents for mount points.
431 * Attributes will be fixed up in ll_inode_revalidate_it */
432 if (d_mountpoint(de))
433 GOTO(out_sa, rc = 1);
435 /* need to get attributes in case root got changed from other client */
436 if (de == de->d_sb->s_root) {
437 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
443 exp = ll_i2mdexp(de->d_inode);
445 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
446 ll_frob_intent(&it, &lookup_it);
449 if (it->it_op == IT_LOOKUP && !(de->d_flags & DCACHE_LUSTRE_INVALID))
452 if ((it->it_op == IT_OPEN) && de->d_inode) {
453 struct inode *inode = de->d_inode;
454 struct ll_inode_info *lli = ll_i2info(inode);
455 struct obd_client_handle **och_p;
460 * We used to check for MDS_INODELOCK_OPEN here, but in fact
461 * just having LOOKUP lock is enough to justify inode is the
462 * same. And if inode is the same and we have suitable
463 * openhandle, then there is no point in doing another OPEN RPC
464 * just to throw away newly received openhandle. There are no
465 * security implications too, if file owner or access mode is
466 * change, LOOKUP lock is revoked.
470 if (it->it_flags & FMODE_WRITE) {
471 och_p = &lli->lli_mds_write_och;
472 och_usecount = &lli->lli_open_fd_write_count;
473 } else if (it->it_flags & FMODE_EXEC) {
474 och_p = &lli->lli_mds_exec_och;
475 och_usecount = &lli->lli_open_fd_exec_count;
477 och_p = &lli->lli_mds_read_och;
478 och_usecount = &lli->lli_open_fd_read_count;
480 /* Check for the proper lock. */
481 ibits = MDS_INODELOCK_LOOKUP;
482 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
484 cfs_mutex_lock(&lli->lli_och_mutex);
485 if (*och_p) { /* Everything is open already, do nothing */
486 /*(*och_usecount)++; Do not let them steal our open
487 handle from under us */
488 SET_BUT_UNUSED(och_usecount);
489 /* XXX The code above was my original idea, but in case
490 we have the handle, but we cannot use it due to later
491 checks (e.g. O_CREAT|O_EXCL flags set), nobody
492 would decrement counter increased here. So we just
493 hope the lock won't be invalidated in between. But
494 if it would be, we'll reopen the open request to
495 MDS later during file open path */
496 cfs_mutex_unlock(&lli->lli_och_mutex);
499 cfs_mutex_unlock(&lli->lli_och_mutex);
503 if (it->it_op == IT_GETATTR) {
504 rc = ll_statahead_enter(parent, &de, 0);
510 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
511 de->d_name.name, de->d_name.len,
512 0, LUSTRE_OPC_ANY, NULL);
514 RETURN(PTR_ERR(op_data));
516 it->it_create_mode &= ~cfs_curproc_umask();
517 it->it_create_mode |= M_CHECK_STALE;
518 rc = md_intent_lock(exp, op_data, NULL, 0, it,
520 &req, ll_md_blocking_ast, 0);
521 it->it_create_mode &= ~M_CHECK_STALE;
522 ll_finish_md_op_data(op_data);
524 /* If req is NULL, then md_intent_lock only tried to do a lock match;
525 * if all was well, it will return 1 if it found locks, 0 otherwise. */
526 if (req == NULL && rc >= 0) {
534 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
535 "%d\n", rc, it->d.lustre.it_status);
541 rc = ll_revalidate_it_finish(req, it, de);
543 if (rc != -ESTALE && rc != -ENOENT)
544 ll_intent_release(it);
548 if ((it->it_op & IT_OPEN) && de->d_inode &&
549 !S_ISREG(de->d_inode->i_mode) &&
550 !S_ISDIR(de->d_inode->i_mode)) {
551 ll_release_openhandle(de, it);
555 /* unfortunately ll_intent_lock may cause a callback and revoke our
557 ll_dentry_rehash(de, 0);
560 /* We do not free request as it may be reused during following lookup
561 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
562 * be freed in ll_lookup_it or in ll_intent_release. But if
563 * request was not completed, we need to free it. (bug 5154, 9903) */
564 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
565 ptlrpc_req_finished(req);
567 ll_unhash_aliases(de->d_inode);
568 /* done in ll_unhash_aliases()
569 dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
573 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
574 "inode %p refc %d\n", de->d_name.len,
575 de->d_name.name, de, de->d_parent, de->d_inode,
576 atomic_read(&de->d_count));
577 ll_set_lock_data(exp, de->d_inode, it, &bits);
578 ll_dentry_reset_flags(de, bits);
579 ll_lookup_finish_locks(it, de);
583 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
584 ll_statahead_mark(parent, de);
588 * This part is here to combat evil-evil race in real_lookup on 2.6
589 * kernels. The race details are: We enter do_lookup() looking for some
590 * name, there is nothing in dcache for this name yet and d_lookup()
591 * returns NULL. We proceed to real_lookup(), and while we do this,
592 * another process does open on the same file we looking up (most simple
593 * reproducer), open succeeds and the dentry is added. Now back to
594 * us. In real_lookup() we do d_lookup() again and suddenly find the
595 * dentry, so we call d_revalidate on it, but there is no lock, so
596 * without this code we would return 0, but unpatched real_lookup just
597 * returns -ENOENT in such a case instead of retrying the lookup. Once
598 * this is dealt with in real_lookup(), all of this ugly mess can go and
599 * we can just check locks in ->d_revalidate without doing any RPCs
603 if (it != &lookup_it) {
604 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
605 if (it->it_op == IT_GETATTR)
606 lookup_it.it_op = IT_GETATTR;
607 ll_lookup_finish_locks(it, de);
611 /* Do real lookup here. */
612 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
613 de->d_name.len, 0, (it->it_op & IT_CREAT ?
615 LUSTRE_OPC_ANY), NULL);
617 RETURN(PTR_ERR(op_data));
619 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
620 ll_md_blocking_ast, 0);
622 struct mdt_body *mdt_body;
623 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
624 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
627 fid = *ll_inode2fid(de->d_inode);
629 /* see if we got same inode, if not - return error */
630 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
631 ll_finish_md_op_data(op_data);
633 goto revalidate_finish;
635 ll_intent_release(it);
637 ll_finish_md_op_data(op_data);
642 * For rc == 1 case, should not return directly to prevent losing
643 * statahead windows; for rc == 0 case, the "lookup" will be done later.
645 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
646 ll_statahead_enter(parent, &de, 1);
650 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
655 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
656 struct lookup_intent *it;
658 it = ll_convert_intent(&nd->intent.open, nd->flags);
662 if (it->it_op == (IT_OPEN|IT_CREAT) &&
663 nd->intent.open.flags & O_EXCL) {
664 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
669 rc = ll_revalidate_it(dentry, nd->flags, it);
671 if (rc && (nd->flags & LOOKUP_OPEN) &&
672 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
673 #ifdef HAVE_FILE_IN_STRUCT_INTENT
674 // XXX Code duplication with ll_lookup_nd
675 if (S_ISFIFO(dentry->d_inode->i_mode)) {
676 // We cannot call open here as it would
679 (struct ptlrpc_request *)
680 it->d.lustre.it_data);
682 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,17))
683 /* 2.6.1[456] have a bug in open_namei() that forgets to check
684 * nd->intent.open.file for error, so we need to return it as lookup's result
688 nd->intent.open.file->private_data = it;
689 filp = lookup_instantiate_filp(nd, dentry,NULL);
694 nd->intent.open.file->private_data = it;
695 (void)lookup_instantiate_filp(nd, dentry,NULL);
699 ll_release_openhandle(dentry, it);
700 #endif /* HAVE_FILE_IN_STRUCT_INTENT */
702 if (!rc && (nd->flags & LOOKUP_CREATE) &&
703 it_disposition(it, DISP_OPEN_CREATE)) {
704 /* We created something but we may only return
705 * negative dentry here, so save request in dentry,
706 * if lookup will be called later on, it will
707 * pick the request, otherwise it would be freed
709 ll_d2d(dentry)->lld_it = it;
710 it = NULL; /* avoid freeing */
715 ll_intent_release(it);
716 OBD_FREE(it, sizeof(*it));
719 rc = ll_revalidate_it(dentry, 0, NULL);
725 void ll_d_iput(struct dentry *de, struct inode *inode)
728 if (!find_cbdata(inode))
733 struct dentry_operations ll_d_ops = {
734 .d_revalidate = ll_revalidate_nd,
735 .d_release = ll_release,
736 .d_delete = ll_ddelete,
738 .d_compare = ll_dcompare,