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) */
69 ll_intent_release(lld->lld_it);
70 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
72 LASSERT(lld->lld_cwd_count == 0);
73 LASSERT(lld->lld_mnt_count == 0);
74 OBD_FREE(de->d_fsdata, sizeof(*lld));
79 /* Compare if two dentries are the same. Don't match if the existing dentry
80 * is marked DCACHE_LUSTRE_INVALID. Returns 1 if different, 0 if the same.
82 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
83 * an AST before calling d_revalidate_it(). The dentry still exists (marked
84 * INVALID) so d_lookup() matches it, but we have no lock on it (so
85 * lock_match() fails) and we spin around real_lookup(). */
86 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
88 struct dentry *dchild;
91 if (d_name->len != name->len)
94 if (memcmp(d_name->name, name->name, name->len))
97 /* XXX: d_name must be in-dentry structure */
98 dchild = container_of(d_name, struct dentry, d_name); /* ugh */
100 CDEBUG(D_DENTRY,"found name %.*s(%p) - flags %d/%x - refc %d\n",
101 name->len, name->name, dchild,
102 d_mountpoint(dchild), dchild->d_flags & DCACHE_LUSTRE_INVALID,
103 atomic_read(&dchild->d_count));
105 /* mountpoint is always valid */
106 if (d_mountpoint(dchild))
109 if (dchild->d_flags & DCACHE_LUSTRE_INVALID)
115 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
118 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
119 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
120 return LDLM_ITER_CONTINUE;
121 return LDLM_ITER_STOP;
124 /* find any ldlm lock of the inode in mdc and lov
128 static int find_cbdata(struct inode *inode)
130 struct ll_inode_info *lli = ll_i2info(inode);
131 struct ll_sb_info *sbi = ll_i2sbi(inode);
136 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
137 return_if_equal, NULL);
142 rc = obd_find_cbdata(sbi->ll_dt_exp, lli->lli_smd,
143 return_if_equal, NULL);
149 * Called when last reference to a dentry is dropped and dcache wants to know
150 * whether or not it should cache it:
151 * - return 1 to delete the dentry immediately
152 * - return 0 to cache the dentry
153 * Should NOT be called with the dcache lock, see fs/dcache.c
155 static int ll_ddelete(struct dentry *de)
160 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
161 (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
162 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
163 d_unhashed(de) ? "" : "hashed,",
164 list_empty(&de->d_subdirs) ? "" : "subdirs");
166 /* if not ldlm lock for this inode, set i_nlink to 0 so that
167 * this inode can be recycled later b=20433 */
168 LASSERT(atomic_read(&de->d_count) == 0);
169 if (de->d_inode && !find_cbdata(de->d_inode))
170 de->d_inode->i_nlink = 0;
172 if (de->d_flags & DCACHE_LUSTRE_INVALID)
178 static int ll_set_dd(struct dentry *de)
183 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
184 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
185 atomic_read(&de->d_count));
187 if (de->d_fsdata == NULL) {
188 struct ll_dentry_data *lld;
191 if (likely(lld != NULL)) {
193 if (likely(de->d_fsdata == NULL))
206 int ll_dops_init(struct dentry *de, int block, int init_sa)
208 struct ll_dentry_data *lld = ll_d2d(de);
211 if (lld == NULL && block != 0) {
219 if (lld != NULL && init_sa != 0)
220 lld->lld_sa_generation = 0;
222 de->d_op = &ll_d_ops;
226 void ll_intent_drop_lock(struct lookup_intent *it)
228 struct lustre_handle *handle;
230 if (it->it_op && it->d.lustre.it_lock_mode) {
231 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
232 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
233 " from it %p\n", handle->cookie, it);
234 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
236 /* bug 494: intent_release may be called multiple times, from
237 * this thread and we don't want to double-decref this lock */
238 it->d.lustre.it_lock_mode = 0;
242 void ll_intent_release(struct lookup_intent *it)
246 CDEBUG(D_INFO, "intent %p released\n", it);
247 ll_intent_drop_lock(it);
248 /* We are still holding extra reference on a request, need to free it */
249 if (it_disposition(it, DISP_ENQ_OPEN_REF))
250 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
251 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
252 ptlrpc_req_finished(it->d.lustre.it_data);
253 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
255 ptlrpc_req_finished(it->d.lustre.it_data);
257 it->d.lustre.it_disposition = 0;
258 it->d.lustre.it_data = NULL;
262 /* Drop dentry if it is not used already, unhash otherwise.
263 Should be called with dcache lock held!
264 Returns: 1 if dentry was dropped, 0 if unhashed. */
265 int ll_drop_dentry(struct dentry *dentry)
268 if (atomic_read(&dentry->d_count) == 0) {
269 CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p "
270 "inode %p\n", dentry->d_name.len,
271 dentry->d_name.name, dentry, dentry->d_parent,
275 unlock_dentry(dentry);
276 spin_unlock(&dcache_lock);
277 cfs_spin_unlock(&ll_lookup_lock);
279 cfs_spin_lock(&ll_lookup_lock);
280 spin_lock(&dcache_lock);
283 /* disconected dentry can not be find without lookup, because we
284 * not need his to unhash or mark invalid. */
285 if (dentry->d_flags & DCACHE_DISCONNECTED) {
286 unlock_dentry(dentry);
290 if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
291 CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p "
292 "inode %p refc %d\n", dentry->d_name.len,
293 dentry->d_name.name, dentry, dentry->d_parent,
294 dentry->d_inode, atomic_read(&dentry->d_count));
295 /* actually we don't unhash the dentry, rather just
296 * mark it inaccessible for to __d_lookup(). otherwise
297 * sys_getcwd() could return -ENOENT -bzzz */
298 dentry->d_flags |= DCACHE_LUSTRE_INVALID;
299 if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
302 unlock_dentry(dentry);
306 void ll_unhash_aliases(struct inode *inode)
308 struct list_head *tmp, *head;
312 CERROR("unexpected NULL inode, tell phil\n");
316 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
317 inode->i_ino, inode->i_generation, inode);
319 head = &inode->i_dentry;
320 cfs_spin_lock(&ll_lookup_lock);
321 spin_lock(&dcache_lock);
324 while ((tmp = tmp->next) != head) {
325 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
327 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
328 "inode %p flags %d\n", dentry->d_name.len,
329 dentry->d_name.name, dentry, dentry->d_parent,
330 dentry->d_inode, dentry->d_flags);
332 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
333 CERROR("called on root (?) dentry=%p, inode=%p "
334 "ino=%lu\n", dentry, inode, inode->i_ino);
335 lustre_dump_dentry(dentry, 1);
336 libcfs_debug_dumpstack(NULL);
339 if (ll_drop_dentry(dentry))
342 spin_unlock(&dcache_lock);
343 cfs_spin_unlock(&ll_lookup_lock);
348 int ll_revalidate_it_finish(struct ptlrpc_request *request,
349 struct lookup_intent *it,
358 if (it_disposition(it, DISP_LOOKUP_NEG))
361 rc = ll_prep_inode(&de->d_inode, request, NULL);
366 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
369 LASSERT(dentry != NULL);
371 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
372 struct inode *inode = dentry->d_inode;
373 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
375 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
376 inode, inode->i_ino, inode->i_generation);
377 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
380 /* drop lookup or getattr locks immediately */
381 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
382 /* on 2.6 there are situation when several lookups and
383 * revalidations may be requested during single operation.
384 * therefore, we don't release intent here -bzzz */
385 ll_intent_drop_lock(it);
389 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
391 struct lookup_intent *it = *itp;
393 if (!it || it->it_op == IT_GETXATTR)
398 int ll_revalidate_it(struct dentry *de, int lookup_flags,
399 struct lookup_intent *it)
401 struct md_op_data *op_data;
402 struct ptlrpc_request *req = NULL;
403 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
404 struct obd_export *exp;
405 struct inode *parent = de->d_parent->d_inode;
409 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
412 if (de->d_inode == NULL) {
415 /* We can only use negative dentries if this is stat or lookup,
416 for opens and stuff we do need to query server. */
417 /* If there is IT_CREAT in intent op set, then we must throw
418 away this negative dentry and actually do the request to
419 kernel to create whatever needs to be created (if possible)*/
420 if (it && (it->it_op & IT_CREAT))
423 if (de->d_flags & DCACHE_LUSTRE_INVALID)
426 ibits = MDS_INODELOCK_UPDATE;
427 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
431 /* Never execute intents for mount points.
432 * Attributes will be fixed up in ll_inode_revalidate_it */
433 if (d_mountpoint(de))
434 GOTO(out_sa, rc = 1);
436 /* need to get attributes in case root got changed from other client */
437 if (de == de->d_sb->s_root) {
438 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
444 exp = ll_i2mdexp(de->d_inode);
446 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
447 ll_frob_intent(&it, &lookup_it);
450 if (it->it_op == IT_LOOKUP && !(de->d_flags & DCACHE_LUSTRE_INVALID))
453 if ((it->it_op == IT_OPEN) && de->d_inode) {
454 struct inode *inode = de->d_inode;
455 struct ll_inode_info *lli = ll_i2info(inode);
456 struct obd_client_handle **och_p;
461 * We used to check for MDS_INODELOCK_OPEN here, but in fact
462 * just having LOOKUP lock is enough to justify inode is the
463 * same. And if inode is the same and we have suitable
464 * openhandle, then there is no point in doing another OPEN RPC
465 * just to throw away newly received openhandle. There are no
466 * security implications too, if file owner or access mode is
467 * change, LOOKUP lock is revoked.
471 if (it->it_flags & FMODE_WRITE) {
472 och_p = &lli->lli_mds_write_och;
473 och_usecount = &lli->lli_open_fd_write_count;
474 } else if (it->it_flags & FMODE_EXEC) {
475 och_p = &lli->lli_mds_exec_och;
476 och_usecount = &lli->lli_open_fd_exec_count;
478 och_p = &lli->lli_mds_read_och;
479 och_usecount = &lli->lli_open_fd_read_count;
481 /* Check for the proper lock. */
482 ibits = MDS_INODELOCK_LOOKUP;
483 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
485 cfs_down(&lli->lli_och_sem);
486 if (*och_p) { /* Everything is open already, do nothing */
487 /*(*och_usecount)++; Do not let them steal our open
488 handle from under us */
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_up(&lli->lli_och_sem);
499 cfs_up(&lli->lli_och_sem);
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,