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>
49 #include "llite_internal.h"
51 /* should NOT be called with the dcache lock, see fs/dcache.c */
52 static void ll_release(struct dentry *de)
54 struct ll_dentry_data *lld;
58 if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
62 ll_intent_release(lld->lld_it);
63 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
65 LASSERT(lld->lld_cwd_count == 0);
66 LASSERT(lld->lld_mnt_count == 0);
67 OBD_FREE(de->d_fsdata, sizeof(*lld));
72 /* Compare if two dentries are the same. Don't match if the existing dentry
73 * is marked DCACHE_LUSTRE_INVALID. Returns 1 if different, 0 if the same.
75 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
76 * an AST before calling d_revalidate_it(). The dentry still exists (marked
77 * INVALID) so d_lookup() matches it, but we have no lock on it (so
78 * lock_match() fails) and we spin around real_lookup(). */
79 #ifdef HAVE_D_COMPARE_7ARGS
80 int ll_dcompare(const struct dentry *parent, const struct inode *pinode,
81 const struct dentry *dentry, const struct inode *inode,
82 unsigned int len, const char *str, const struct qstr *name)
84 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
87 #ifdef HAVE_D_COMPARE_7ARGS
93 if (memcmp(str, name->name, len))
96 struct dentry *dentry;
99 if (d_name->len != name->len)
102 if (memcmp(d_name->name, name->name, name->len))
105 /* XXX: d_name must be in-dentry structure */
106 dentry = container_of(d_name, struct dentry, d_name); /* ugh */
109 CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n",
110 name->len, name->name, dentry, dentry->d_flags,
113 /* mountpoint is always valid */
114 if (d_mountpoint((struct dentry *)dentry))
117 if (d_lustre_invalid(dentry))
123 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
126 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
127 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
128 return LDLM_ITER_CONTINUE;
129 return LDLM_ITER_STOP;
132 /* find any ldlm lock of the inode in mdc and lov
136 static int find_cbdata(struct inode *inode)
138 struct ll_inode_info *lli = ll_i2info(inode);
139 struct ll_sb_info *sbi = ll_i2sbi(inode);
144 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
145 return_if_equal, NULL);
150 rc = obd_find_cbdata(sbi->ll_dt_exp, lli->lli_smd,
151 return_if_equal, NULL);
157 * Called when last reference to a dentry is dropped and dcache wants to know
158 * whether or not it should cache it:
159 * - return 1 to delete the dentry immediately
160 * - return 0 to cache the dentry
161 * Should NOT be called with the dcache lock, see fs/dcache.c
163 static int ll_ddelete(HAVE_D_DELETE_CONST struct dentry *de)
168 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
169 d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
170 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
171 d_unhashed((struct dentry *)de) ? "" : "hashed,",
172 list_empty(&de->d_subdirs) ? "" : "subdirs");
174 #ifdef HAVE_DCACHE_LOCK
175 LASSERT(d_refcount(de) == 0);
177 /* kernel >= 2.6.38 last refcount is decreased after this function. */
178 LASSERT(d_refcount(de) == 1);
181 /* if not ldlm lock for this inode, set i_nlink to 0 so that
182 * this inode can be recycled later b=20433 */
183 if (de->d_inode && !find_cbdata(de->d_inode))
184 de->d_inode->i_nlink = 0;
186 if (d_lustre_invalid((struct dentry *)de))
192 static int ll_set_dd(struct dentry *de)
197 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
198 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
201 if (de->d_fsdata == NULL) {
202 struct ll_dentry_data *lld;
205 if (likely(lld != NULL)) {
206 spin_lock(&de->d_lock);
207 if (likely(de->d_fsdata == NULL))
211 spin_unlock(&de->d_lock);
220 int ll_dops_init(struct dentry *de, int block, int init_sa)
222 struct ll_dentry_data *lld = ll_d2d(de);
225 if (lld == NULL && block != 0) {
233 if (lld != NULL && init_sa != 0)
234 lld->lld_sa_generation = 0;
236 #ifdef HAVE_DCACHE_LOCK
237 de->d_op = &ll_d_ops;
239 /* kernel >= 2.6.38 d_op is set in d_alloc() */
240 LASSERT(de->d_op == &ll_d_ops);
245 void ll_intent_drop_lock(struct lookup_intent *it)
247 struct lustre_handle *handle;
249 if (it->it_op && it->d.lustre.it_lock_mode) {
250 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
251 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
252 " from it %p\n", handle->cookie, it);
253 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
255 /* bug 494: intent_release may be called multiple times, from
256 * this thread and we don't want to double-decref this lock */
257 it->d.lustre.it_lock_mode = 0;
261 void ll_intent_release(struct lookup_intent *it)
265 CDEBUG(D_INFO, "intent %p released\n", it);
266 ll_intent_drop_lock(it);
267 /* We are still holding extra reference on a request, need to free it */
268 if (it_disposition(it, DISP_ENQ_OPEN_REF))
269 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
270 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
271 ptlrpc_req_finished(it->d.lustre.it_data);
272 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
274 ptlrpc_req_finished(it->d.lustre.it_data);
276 it->d.lustre.it_disposition = 0;
277 it->d.lustre.it_data = NULL;
281 void ll_invalidate_aliases(struct inode *inode)
283 struct dentry *dentry;
286 LASSERT(inode != NULL);
288 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
289 inode->i_ino, inode->i_generation, inode);
291 ll_lock_dcache(inode);
292 cfs_list_for_each_entry(dentry, &inode->i_dentry, d_alias) {
293 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
294 "inode %p flags %d\n", dentry->d_name.len,
295 dentry->d_name.name, dentry, dentry->d_parent,
296 dentry->d_inode, dentry->d_flags);
298 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
299 CERROR("called on root (?) dentry=%p, inode=%p "
300 "ino=%lu\n", dentry, inode, inode->i_ino);
301 lustre_dump_dentry(dentry, 1);
302 libcfs_debug_dumpstack(NULL);
305 d_lustre_invalidate(dentry);
307 ll_unlock_dcache(inode);
312 int ll_revalidate_it_finish(struct ptlrpc_request *request,
313 struct lookup_intent *it,
322 if (it_disposition(it, DISP_LOOKUP_NEG))
325 rc = ll_prep_inode(&de->d_inode, request, NULL);
330 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
333 LASSERT(dentry != NULL);
335 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
336 struct inode *inode = dentry->d_inode;
337 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
339 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
340 inode, inode->i_ino, inode->i_generation);
341 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
344 /* drop lookup or getattr locks immediately */
345 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
346 /* on 2.6 there are situation when several lookups and
347 * revalidations may be requested during single operation.
348 * therefore, we don't release intent here -bzzz */
349 ll_intent_drop_lock(it);
353 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
355 struct lookup_intent *it = *itp;
357 if (!it || it->it_op == IT_GETXATTR)
362 int ll_revalidate_it(struct dentry *de, int lookup_flags,
363 struct lookup_intent *it)
365 struct md_op_data *op_data;
366 struct ptlrpc_request *req = NULL;
367 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
368 struct obd_export *exp;
369 struct inode *parent = de->d_parent->d_inode;
373 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
376 if (de->d_inode == NULL) {
379 /* We can only use negative dentries if this is stat or lookup,
380 for opens and stuff we do need to query server. */
381 /* If there is IT_CREAT in intent op set, then we must throw
382 away this negative dentry and actually do the request to
383 kernel to create whatever needs to be created (if possible)*/
384 if (it && (it->it_op & IT_CREAT))
387 if (d_lustre_invalid(de))
390 ibits = MDS_INODELOCK_UPDATE;
391 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
395 /* Never execute intents for mount points.
396 * Attributes will be fixed up in ll_inode_revalidate_it */
397 if (d_mountpoint(de))
398 GOTO(out_sa, rc = 1);
400 /* need to get attributes in case root got changed from other client */
401 if (de == de->d_sb->s_root) {
402 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
408 exp = ll_i2mdexp(de->d_inode);
410 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
411 ll_frob_intent(&it, &lookup_it);
414 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
417 if ((it->it_op == IT_OPEN) && de->d_inode) {
418 struct inode *inode = de->d_inode;
419 struct ll_inode_info *lli = ll_i2info(inode);
420 struct obd_client_handle **och_p;
425 * We used to check for MDS_INODELOCK_OPEN here, but in fact
426 * just having LOOKUP lock is enough to justify inode is the
427 * same. And if inode is the same and we have suitable
428 * openhandle, then there is no point in doing another OPEN RPC
429 * just to throw away newly received openhandle. There are no
430 * security implications too, if file owner or access mode is
431 * change, LOOKUP lock is revoked.
435 if (it->it_flags & FMODE_WRITE) {
436 och_p = &lli->lli_mds_write_och;
437 och_usecount = &lli->lli_open_fd_write_count;
438 } else if (it->it_flags & FMODE_EXEC) {
439 och_p = &lli->lli_mds_exec_och;
440 och_usecount = &lli->lli_open_fd_exec_count;
442 och_p = &lli->lli_mds_read_och;
443 och_usecount = &lli->lli_open_fd_read_count;
445 /* Check for the proper lock. */
446 ibits = MDS_INODELOCK_LOOKUP;
447 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
449 cfs_mutex_lock(&lli->lli_och_mutex);
450 if (*och_p) { /* Everything is open already, do nothing */
451 /*(*och_usecount)++; Do not let them steal our open
452 handle from under us */
453 SET_BUT_UNUSED(och_usecount);
454 /* XXX The code above was my original idea, but in case
455 we have the handle, but we cannot use it due to later
456 checks (e.g. O_CREAT|O_EXCL flags set), nobody
457 would decrement counter increased here. So we just
458 hope the lock won't be invalidated in between. But
459 if it would be, we'll reopen the open request to
460 MDS later during file open path */
461 cfs_mutex_unlock(&lli->lli_och_mutex);
464 cfs_mutex_unlock(&lli->lli_och_mutex);
468 if (it->it_op == IT_GETATTR) {
469 rc = ll_statahead_enter(parent, &de, 0);
472 else if (rc != -EAGAIN && rc != 0)
477 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
478 de->d_name.name, de->d_name.len,
479 0, LUSTRE_OPC_ANY, NULL);
481 RETURN(PTR_ERR(op_data));
483 it->it_create_mode &= ~cfs_curproc_umask();
484 it->it_create_mode |= M_CHECK_STALE;
485 rc = md_intent_lock(exp, op_data, NULL, 0, it,
487 &req, ll_md_blocking_ast, 0);
488 it->it_create_mode &= ~M_CHECK_STALE;
489 ll_finish_md_op_data(op_data);
491 /* If req is NULL, then md_intent_lock only tried to do a lock match;
492 * if all was well, it will return 1 if it found locks, 0 otherwise. */
493 if (req == NULL && rc >= 0) {
501 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
502 "%d\n", rc, it->d.lustre.it_status);
508 rc = ll_revalidate_it_finish(req, it, de);
510 if (rc != -ESTALE && rc != -ENOENT)
511 ll_intent_release(it);
515 if ((it->it_op & IT_OPEN) && de->d_inode &&
516 !S_ISREG(de->d_inode->i_mode) &&
517 !S_ISDIR(de->d_inode->i_mode)) {
518 ll_release_openhandle(de, it);
523 /* We do not free request as it may be reused during following lookup
524 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
525 * be freed in ll_lookup_it or in ll_intent_release. But if
526 * request was not completed, we need to free it. (bug 5154, 9903) */
527 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
528 ptlrpc_req_finished(req);
530 ll_invalidate_aliases(de->d_inode);
534 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
535 "inode %p refc %d\n", de->d_name.len,
536 de->d_name.name, de, de->d_parent, de->d_inode,
538 ll_set_lock_data(exp, de->d_inode, it, &bits);
539 if ((bits & MDS_INODELOCK_LOOKUP) && d_lustre_invalid(de))
540 d_lustre_revalidate(de);
541 ll_lookup_finish_locks(it, de);
545 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
546 ll_statahead_mark(parent, de);
550 * This part is here to combat evil-evil race in real_lookup on 2.6
551 * kernels. The race details are: We enter do_lookup() looking for some
552 * name, there is nothing in dcache for this name yet and d_lookup()
553 * returns NULL. We proceed to real_lookup(), and while we do this,
554 * another process does open on the same file we looking up (most simple
555 * reproducer), open succeeds and the dentry is added. Now back to
556 * us. In real_lookup() we do d_lookup() again and suddenly find the
557 * dentry, so we call d_revalidate on it, but there is no lock, so
558 * without this code we would return 0, but unpatched real_lookup just
559 * returns -ENOENT in such a case instead of retrying the lookup. Once
560 * this is dealt with in real_lookup(), all of this ugly mess can go and
561 * we can just check locks in ->d_revalidate without doing any RPCs
565 if (it != &lookup_it) {
566 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
567 if (it->it_op == IT_GETATTR)
568 lookup_it.it_op = IT_GETATTR;
569 ll_lookup_finish_locks(it, de);
573 /* Do real lookup here. */
574 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
575 de->d_name.len, 0, (it->it_op & IT_CREAT ?
577 LUSTRE_OPC_ANY), NULL);
579 RETURN(PTR_ERR(op_data));
581 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
582 ll_md_blocking_ast, 0);
584 struct mdt_body *mdt_body;
585 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
586 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
589 fid = *ll_inode2fid(de->d_inode);
591 /* see if we got same inode, if not - return error */
592 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
593 ll_finish_md_op_data(op_data);
595 goto revalidate_finish;
597 ll_intent_release(it);
599 ll_finish_md_op_data(op_data);
604 * For rc == 1 case, should not return directly to prevent losing
605 * statahead windows; for rc == 0 case, the "lookup" will be done later.
607 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
608 ll_statahead_enter(parent, &de, 1);
612 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
617 #ifndef HAVE_DCACHE_LOCK
618 /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */
619 if (nd->flags & LOOKUP_RCU)
623 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
624 struct lookup_intent *it;
626 it = ll_convert_intent(&nd->intent.open, nd->flags);
630 if (it->it_op == (IT_OPEN|IT_CREAT) &&
631 nd->intent.open.flags & O_EXCL) {
632 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
637 rc = ll_revalidate_it(dentry, nd->flags, it);
639 if (rc && (nd->flags & LOOKUP_OPEN) &&
640 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
641 // XXX Code duplication with ll_lookup_nd
642 if (S_ISFIFO(dentry->d_inode->i_mode)) {
643 // We cannot call open here as it would
646 (struct ptlrpc_request *)
647 it->d.lustre.it_data);
651 nd->intent.open.file->private_data = it;
652 filp = lookup_instantiate_filp(nd, dentry,NULL);
657 if (!rc && (nd->flags & LOOKUP_CREATE) &&
658 it_disposition(it, DISP_OPEN_CREATE)) {
659 /* We created something but we may only return
660 * negative dentry here, so save request in dentry,
661 * if lookup will be called later on, it will
662 * pick the request, otherwise it would be freed
664 ll_d2d(dentry)->lld_it = it;
665 it = NULL; /* avoid freeing */
670 ll_intent_release(it);
671 OBD_FREE(it, sizeof(*it));
674 rc = ll_revalidate_it(dentry, 0, NULL);
680 void ll_d_iput(struct dentry *de, struct inode *inode)
683 if (!find_cbdata(inode))
688 struct dentry_operations ll_d_ops = {
689 .d_revalidate = ll_revalidate_nd,
690 .d_release = ll_release,
691 .d_delete = ll_ddelete,
693 .d_compare = ll_dcompare,