4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, Whamcloud, Inc.
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_ver.h>
49 //#include <lustre_version.h>
51 #include "llite_internal.h"
53 cfs_spinlock_t ll_lookup_lock = CFS_SPIN_LOCK_UNLOCKED;
55 /* should NOT be called with the dcache lock, see fs/dcache.c */
56 static void ll_release(struct dentry *de)
58 struct ll_dentry_data *lld;
62 if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
66 ll_intent_release(lld->lld_it);
67 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
69 LASSERT(lld->lld_cwd_count == 0);
70 LASSERT(lld->lld_mnt_count == 0);
71 OBD_FREE(de->d_fsdata, sizeof(*lld));
76 /* Compare if two dentries are the same. Don't match if the existing dentry
77 * is marked DCACHE_LUSTRE_INVALID. Returns 1 if different, 0 if the same.
79 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
80 * an AST before calling d_revalidate_it(). The dentry still exists (marked
81 * INVALID) so d_lookup() matches it, but we have no lock on it (so
82 * lock_match() fails) and we spin around real_lookup(). */
83 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
85 struct dentry *dchild;
88 if (d_name->len != name->len)
91 if (memcmp(d_name->name, name->name, name->len))
94 /* XXX: d_name must be in-dentry structure */
95 dchild = container_of(d_name, struct dentry, d_name); /* ugh */
97 CDEBUG(D_DENTRY,"found name %.*s(%p) - flags %d/%x - refc %d\n",
98 name->len, name->name, dchild,
99 d_mountpoint(dchild), dchild->d_flags & DCACHE_LUSTRE_INVALID,
100 atomic_read(&dchild->d_count));
102 /* mountpoint is always valid */
103 if (d_mountpoint(dchild))
106 if (dchild->d_flags & DCACHE_LUSTRE_INVALID)
112 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
115 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
116 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
117 return LDLM_ITER_CONTINUE;
118 return LDLM_ITER_STOP;
121 /* find any ldlm lock of the inode in mdc and lov
125 static int find_cbdata(struct inode *inode)
127 struct ll_inode_info *lli = ll_i2info(inode);
128 struct ll_sb_info *sbi = ll_i2sbi(inode);
133 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
134 return_if_equal, NULL);
139 rc = obd_find_cbdata(sbi->ll_dt_exp, lli->lli_smd,
140 return_if_equal, NULL);
146 * Called when last reference to a dentry is dropped and dcache wants to know
147 * whether or not it should cache it:
148 * - return 1 to delete the dentry immediately
149 * - return 0 to cache the dentry
150 * Should NOT be called with the dcache lock, see fs/dcache.c
152 static int ll_ddelete(struct dentry *de)
157 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
158 (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
159 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
160 d_unhashed(de) ? "" : "hashed,",
161 list_empty(&de->d_subdirs) ? "" : "subdirs");
163 /* if not ldlm lock for this inode, set i_nlink to 0 so that
164 * this inode can be recycled later b=20433 */
165 LASSERT(atomic_read(&de->d_count) == 0);
166 if (de->d_inode && !find_cbdata(de->d_inode))
167 de->d_inode->i_nlink = 0;
169 if (de->d_flags & DCACHE_LUSTRE_INVALID)
175 static int ll_set_dd(struct dentry *de)
180 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
181 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
182 atomic_read(&de->d_count));
184 if (de->d_fsdata == NULL) {
185 struct ll_dentry_data *lld;
188 if (likely(lld != NULL)) {
190 if (likely(de->d_fsdata == NULL))
203 int ll_dops_init(struct dentry *de, int block, int init_sa)
205 struct ll_dentry_data *lld = ll_d2d(de);
208 if (lld == NULL && block != 0) {
216 if (lld != NULL && init_sa != 0)
217 lld->lld_sa_generation = 0;
219 de->d_op = &ll_d_ops;
223 void ll_intent_drop_lock(struct lookup_intent *it)
225 struct lustre_handle *handle;
227 if (it->it_op && it->d.lustre.it_lock_mode) {
228 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
229 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
230 " from it %p\n", handle->cookie, it);
231 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
233 /* bug 494: intent_release may be called multiple times, from
234 * this thread and we don't want to double-decref this lock */
235 it->d.lustre.it_lock_mode = 0;
239 void ll_intent_release(struct lookup_intent *it)
243 CDEBUG(D_INFO, "intent %p released\n", it);
244 ll_intent_drop_lock(it);
245 /* We are still holding extra reference on a request, need to free it */
246 if (it_disposition(it, DISP_ENQ_OPEN_REF))
247 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
248 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
249 ptlrpc_req_finished(it->d.lustre.it_data);
250 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
252 ptlrpc_req_finished(it->d.lustre.it_data);
254 it->d.lustre.it_disposition = 0;
255 it->d.lustre.it_data = NULL;
259 /* Drop dentry if it is not used already, unhash otherwise.
260 Should be called with dcache lock held!
261 Returns: 1 if dentry was dropped, 0 if unhashed. */
262 int ll_drop_dentry(struct dentry *dentry)
265 if (atomic_read(&dentry->d_count) == 0) {
266 CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p "
267 "inode %p\n", dentry->d_name.len,
268 dentry->d_name.name, dentry, dentry->d_parent,
272 unlock_dentry(dentry);
273 spin_unlock(&dcache_lock);
274 cfs_spin_unlock(&ll_lookup_lock);
276 cfs_spin_lock(&ll_lookup_lock);
277 spin_lock(&dcache_lock);
280 /* disconected dentry can not be find without lookup, because we
281 * not need his to unhash or mark invalid. */
282 if (dentry->d_flags & DCACHE_DISCONNECTED) {
283 unlock_dentry(dentry);
287 if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
288 CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p "
289 "inode %p refc %d\n", dentry->d_name.len,
290 dentry->d_name.name, dentry, dentry->d_parent,
291 dentry->d_inode, atomic_read(&dentry->d_count));
292 /* actually we don't unhash the dentry, rather just
293 * mark it inaccessible for to __d_lookup(). otherwise
294 * sys_getcwd() could return -ENOENT -bzzz */
295 dentry->d_flags |= DCACHE_LUSTRE_INVALID;
296 if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
299 unlock_dentry(dentry);
303 void ll_unhash_aliases(struct inode *inode)
305 struct list_head *tmp, *head;
309 CERROR("unexpected NULL inode, tell phil\n");
313 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
314 inode->i_ino, inode->i_generation, inode);
316 head = &inode->i_dentry;
317 cfs_spin_lock(&ll_lookup_lock);
318 spin_lock(&dcache_lock);
321 while ((tmp = tmp->next) != head) {
322 struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
324 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
325 "inode %p flags %d\n", dentry->d_name.len,
326 dentry->d_name.name, dentry, dentry->d_parent,
327 dentry->d_inode, dentry->d_flags);
329 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
330 CERROR("called on root (?) dentry=%p, inode=%p "
331 "ino=%lu\n", dentry, inode, inode->i_ino);
332 lustre_dump_dentry(dentry, 1);
333 libcfs_debug_dumpstack(NULL);
336 if (ll_drop_dentry(dentry))
339 spin_unlock(&dcache_lock);
340 cfs_spin_unlock(&ll_lookup_lock);
345 int ll_revalidate_it_finish(struct ptlrpc_request *request,
346 struct lookup_intent *it,
355 if (it_disposition(it, DISP_LOOKUP_NEG))
358 rc = ll_prep_inode(&de->d_inode, request, NULL);
363 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
366 LASSERT(dentry != NULL);
368 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
369 struct inode *inode = dentry->d_inode;
370 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
372 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
373 inode, inode->i_ino, inode->i_generation);
374 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
377 /* drop lookup or getattr locks immediately */
378 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
379 /* on 2.6 there are situation when several lookups and
380 * revalidations may be requested during single operation.
381 * therefore, we don't release intent here -bzzz */
382 ll_intent_drop_lock(it);
386 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
388 struct lookup_intent *it = *itp;
390 if (!it || it->it_op == IT_GETXATTR)
395 int ll_revalidate_it(struct dentry *de, int lookup_flags,
396 struct lookup_intent *it)
398 struct md_op_data *op_data;
399 struct ptlrpc_request *req = NULL;
400 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
401 struct obd_export *exp;
402 struct inode *parent = de->d_parent->d_inode;
406 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
409 if (de->d_inode == NULL) {
412 /* We can only use negative dentries if this is stat or lookup,
413 for opens and stuff we do need to query server. */
414 /* If there is IT_CREAT in intent op set, then we must throw
415 away this negative dentry and actually do the request to
416 kernel to create whatever needs to be created (if possible)*/
417 if (it && (it->it_op & IT_CREAT))
420 if (de->d_flags & DCACHE_LUSTRE_INVALID)
423 ibits = MDS_INODELOCK_UPDATE;
424 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
428 /* Never execute intents for mount points.
429 * Attributes will be fixed up in ll_inode_revalidate_it */
430 if (d_mountpoint(de))
431 GOTO(out_sa, rc = 1);
433 /* need to get attributes in case root got changed from other client */
434 if (de == de->d_sb->s_root) {
435 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
441 exp = ll_i2mdexp(de->d_inode);
443 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
444 ll_frob_intent(&it, &lookup_it);
447 if (it->it_op == IT_LOOKUP && !(de->d_flags & DCACHE_LUSTRE_INVALID))
450 if ((it->it_op == IT_OPEN) && de->d_inode) {
451 struct inode *inode = de->d_inode;
452 struct ll_inode_info *lli = ll_i2info(inode);
453 struct obd_client_handle **och_p;
458 * We used to check for MDS_INODELOCK_OPEN here, but in fact
459 * just having LOOKUP lock is enough to justify inode is the
460 * same. And if inode is the same and we have suitable
461 * openhandle, then there is no point in doing another OPEN RPC
462 * just to throw away newly received openhandle. There are no
463 * security implications too, if file owner or access mode is
464 * change, LOOKUP lock is revoked.
468 if (it->it_flags & FMODE_WRITE) {
469 och_p = &lli->lli_mds_write_och;
470 och_usecount = &lli->lli_open_fd_write_count;
471 } else if (it->it_flags & FMODE_EXEC) {
472 och_p = &lli->lli_mds_exec_och;
473 och_usecount = &lli->lli_open_fd_exec_count;
475 och_p = &lli->lli_mds_read_och;
476 och_usecount = &lli->lli_open_fd_read_count;
478 /* Check for the proper lock. */
479 ibits = MDS_INODELOCK_LOOKUP;
480 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
482 cfs_mutex_lock(&lli->lli_och_mutex);
483 if (*och_p) { /* Everything is open already, do nothing */
484 /*(*och_usecount)++; Do not let them steal our open
485 handle from under us */
486 SET_BUT_UNUSED(och_usecount);
487 /* XXX The code above was my original idea, but in case
488 we have the handle, but we cannot use it due to later
489 checks (e.g. O_CREAT|O_EXCL flags set), nobody
490 would decrement counter increased here. So we just
491 hope the lock won't be invalidated in between. But
492 if it would be, we'll reopen the open request to
493 MDS later during file open path */
494 cfs_mutex_unlock(&lli->lli_och_mutex);
497 cfs_mutex_unlock(&lli->lli_och_mutex);
501 if (it->it_op == IT_GETATTR) {
502 rc = ll_statahead_enter(parent, &de, 0);
505 else if (rc != -EAGAIN && rc != 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 // XXX Code duplication with ll_lookup_nd
674 if (S_ISFIFO(dentry->d_inode->i_mode)) {
675 // We cannot call open here as it would
678 (struct ptlrpc_request *)
679 it->d.lustre.it_data);
681 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,17))
682 /* 2.6.1[456] have a bug in open_namei() that forgets to check
683 * nd->intent.open.file for error, so we need to return it as lookup's result
687 nd->intent.open.file->private_data = it;
688 filp = lookup_instantiate_filp(nd, dentry,NULL);
693 nd->intent.open.file->private_data = it;
694 (void)lookup_instantiate_filp(nd, dentry,NULL);
698 if (!rc && (nd->flags & LOOKUP_CREATE) &&
699 it_disposition(it, DISP_OPEN_CREATE)) {
700 /* We created something but we may only return
701 * negative dentry here, so save request in dentry,
702 * if lookup will be called later on, it will
703 * pick the request, otherwise it would be freed
705 ll_d2d(dentry)->lld_it = it;
706 it = NULL; /* avoid freeing */
711 ll_intent_release(it);
712 OBD_FREE(it, sizeof(*it));
715 rc = ll_revalidate_it(dentry, 0, NULL);
721 void ll_d_iput(struct dentry *de, struct inode *inode)
724 if (!find_cbdata(inode))
729 struct dentry_operations ll_d_ops = {
730 .d_revalidate = ll_revalidate_nd,
731 .d_release = ll_release,
732 .d_delete = ll_ddelete,
734 .d_compare = ll_dcompare,