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, 2013, Intel Corporation.
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/quotaops.h>
40 #include <linux/kernel.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 static void free_dentry_data(struct rcu_head *head)
53 struct ll_dentry_data *lld;
55 lld = container_of(head, struct ll_dentry_data, lld_rcu_head);
59 /* should NOT be called with the dcache lock, see fs/dcache.c */
60 static void ll_release(struct dentry *de)
62 struct ll_dentry_data *lld;
66 if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
70 ll_intent_release(lld->lld_it);
71 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
73 LASSERT(lld->lld_cwd_count == 0);
74 LASSERT(lld->lld_mnt_count == 0);
76 call_rcu(&lld->lld_rcu_head, free_dentry_data);
81 /* Compare if two dentries are the same. Don't match if the existing dentry
82 * is marked invalid. Returns 1 if different, 0 if the same.
84 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
85 * an AST before calling d_revalidate_it(). The dentry still exists (marked
86 * INVALID) so d_lookup() matches it, but we have no lock on it (so
87 * lock_match() fails) and we spin around real_lookup(). */
88 #ifdef HAVE_D_COMPARE_7ARGS
89 int ll_dcompare(const struct dentry *parent, const struct inode *pinode,
90 const struct dentry *dentry, const struct inode *inode,
91 unsigned int len, const char *str, const struct qstr *name)
93 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
96 #ifdef HAVE_D_COMPARE_7ARGS
102 if (memcmp(str, name->name, len))
105 struct dentry *dentry;
108 if (d_name->len != name->len)
111 if (memcmp(d_name->name, name->name, name->len))
114 /* XXX: d_name must be in-dentry structure */
115 dentry = container_of(d_name, struct dentry, d_name); /* ugh */
118 CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n",
119 name->len, name->name, dentry, dentry->d_flags,
122 /* mountpoint is always valid */
123 if (d_mountpoint((struct dentry *)dentry))
126 if (d_lustre_invalid(dentry))
132 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
134 return (ldlm_is_canceling(lock) && ldlm_is_discard_data(lock)) ?
135 LDLM_ITER_CONTINUE : LDLM_ITER_STOP;
138 /* find any ldlm lock of the inode in mdc and lov
142 static int find_cbdata(struct inode *inode)
144 struct ll_sb_info *sbi = ll_i2sbi(inode);
145 struct lov_stripe_md *lsm;
150 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
151 return_if_equal, NULL);
155 lsm = ccc_inode_lsm_get(inode);
159 rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL);
160 ccc_inode_lsm_put(inode, lsm);
166 * Called when last reference to a dentry is dropped and dcache wants to know
167 * whether or not it should cache it:
168 * - return 1 to delete the dentry immediately
169 * - return 0 to cache the dentry
170 * Should NOT be called with the dcache lock, see fs/dcache.c
172 static int ll_ddelete(HAVE_D_DELETE_CONST struct dentry *de)
177 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
178 d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
179 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
180 d_unhashed((struct dentry *)de) ? "" : "hashed,",
181 list_empty(&de->d_subdirs) ? "" : "subdirs");
183 #ifdef HAVE_DCACHE_LOCK
184 LASSERT(d_refcount(de) == 0);
186 /* kernel >= 2.6.38 last refcount is decreased after this function. */
187 LASSERT(d_refcount(de) == 1);
190 /* Disable this piece of code temproarily because this is called
191 * inside dcache_lock so it's not appropriate to do lots of work
192 * here. ATTENTION: Before this piece of code enabling, LU-2487 must be
195 /* if not ldlm lock for this inode, set i_nlink to 0 so that
196 * this inode can be recycled later b=20433 */
197 if (de->d_inode && !find_cbdata(de->d_inode))
198 clear_nlink(de->d_inode);
201 if (d_lustre_invalid((struct dentry *)de))
206 int ll_d_init(struct dentry *de)
211 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
212 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
215 if (de->d_fsdata == NULL) {
216 struct ll_dentry_data *lld;
219 if (likely(lld != NULL)) {
220 spin_lock(&de->d_lock);
221 if (likely(de->d_fsdata == NULL)) {
223 __d_lustre_invalidate(de);
224 #ifdef HAVE_DCACHE_LOCK
225 /* kernel >= 2.6.38 d_op is set in d_alloc() */
226 de->d_op = &ll_d_ops;
231 spin_unlock(&de->d_lock);
236 LASSERT(de->d_op == &ll_d_ops);
241 void ll_intent_drop_lock(struct lookup_intent *it)
243 if (it->it_op && it->d.lustre.it_lock_mode) {
244 struct lustre_handle handle;
246 handle.cookie = it->d.lustre.it_lock_handle;
248 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
249 " from it %p\n", handle.cookie, it);
250 ldlm_lock_decref(&handle, it->d.lustre.it_lock_mode);
252 /* bug 494: intent_release may be called multiple times, from
253 * this thread and we don't want to double-decref this lock */
254 it->d.lustre.it_lock_mode = 0;
255 if (it->d.lustre.it_remote_lock_mode != 0) {
256 handle.cookie = it->d.lustre.it_remote_lock_handle;
258 CDEBUG(D_DLMTRACE, "releasing remote lock with cookie"
259 LPX64" from it %p\n", handle.cookie, it);
260 ldlm_lock_decref(&handle,
261 it->d.lustre.it_remote_lock_mode);
262 it->d.lustre.it_remote_lock_mode = 0;
267 void ll_intent_release(struct lookup_intent *it)
271 CDEBUG(D_INFO, "intent %p released\n", it);
272 ll_intent_drop_lock(it);
273 /* We are still holding extra reference on a request, need to free it */
274 if (it_disposition(it, DISP_ENQ_OPEN_REF))
275 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
276 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
277 ptlrpc_req_finished(it->d.lustre.it_data);
278 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
280 ptlrpc_req_finished(it->d.lustre.it_data);
282 it->d.lustre.it_disposition = 0;
283 it->d.lustre.it_data = NULL;
287 void ll_invalidate_aliases(struct inode *inode)
289 struct dentry *dentry;
290 DECLARE_LL_D_HLIST_NODE_PTR(p);
293 LASSERT(inode != NULL);
295 CDEBUG(D_INODE, "marking dentries for inode "DFID"(%p) invalid\n",
296 PFID(ll_inode2fid(inode)), inode);
298 ll_lock_dcache(inode);
299 ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) {
300 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
301 "inode %p flags %d\n", dentry->d_name.len,
302 dentry->d_name.name, dentry, dentry->d_parent,
303 dentry->d_inode, dentry->d_flags);
305 if (unlikely(dentry == dentry->d_sb->s_root)) {
306 CERROR("%s: called on root dentry=%p, fid="DFID"\n",
307 ll_get_fsname(dentry->d_sb, NULL, 0),
308 dentry, PFID(ll_inode2fid(inode)));
309 lustre_dump_dentry(dentry, 1);
310 libcfs_debug_dumpstack(NULL);
313 d_lustre_invalidate(dentry, 0);
315 ll_unlock_dcache(inode);
320 int ll_revalidate_it_finish(struct ptlrpc_request *request,
321 struct lookup_intent *it,
330 if (it_disposition(it, DISP_LOOKUP_NEG))
333 rc = ll_prep_inode(&de->d_inode, request, NULL, it);
338 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
341 LASSERT(dentry != NULL);
343 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
344 struct inode *inode = dentry->d_inode;
345 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
347 CDEBUG(D_DLMTRACE, "setting l_data to inode "DFID"(%p)\n",
348 PFID(ll_inode2fid(inode)), inode);
349 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
352 /* drop lookup or getattr locks immediately */
353 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
354 /* on 2.6 there are situation when several lookups and
355 * revalidations may be requested during single operation.
356 * therefore, we don't release intent here -bzzz */
357 ll_intent_drop_lock(it);
361 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
363 struct lookup_intent *it = *itp;
365 if (!it || it->it_op == IT_GETXATTR)
370 int ll_revalidate_it(struct dentry *de, int lookup_flags,
371 struct lookup_intent *it)
373 struct md_op_data *op_data;
374 struct ptlrpc_request *req = NULL;
375 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
376 struct obd_export *exp;
377 struct inode *parent = de->d_parent->d_inode;
381 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
384 LASSERT(de != de->d_sb->s_root);
386 if (de->d_inode == NULL) {
389 /* We can only use negative dentries if this is stat or lookup,
390 for opens and stuff we do need to query server. */
391 /* If there is IT_CREAT in intent op set, then we must throw
392 away this negative dentry and actually do the request to
393 kernel to create whatever needs to be created (if possible)*/
394 if (it && (it->it_op & IT_CREAT))
397 if (d_lustre_invalid(de))
400 ibits = MDS_INODELOCK_UPDATE;
401 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
405 /* Never execute intents for mount points.
406 * Attributes will be fixed up in ll_inode_revalidate_it */
407 if (d_mountpoint(de))
408 GOTO(out_sa, rc = 1);
410 exp = ll_i2mdexp(de->d_inode);
412 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
413 ll_frob_intent(&it, &lookup_it);
416 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
419 if (it->it_op == IT_OPEN) {
420 struct inode *inode = de->d_inode;
421 struct ll_inode_info *lli = ll_i2info(inode);
422 struct obd_client_handle **och_p;
426 * We used to check for MDS_INODELOCK_OPEN here, but in fact
427 * just having LOOKUP lock is enough to justify inode is the
428 * same. And if inode is the same and we have suitable
429 * openhandle, then there is no point in doing another OPEN RPC
430 * just to throw away newly received openhandle. There are no
431 * security implications too, if file owner or access mode is
432 * change, LOOKUP lock is revoked.
435 if (it->it_flags & FMODE_WRITE)
436 och_p = &lli->lli_mds_write_och;
437 else if (it->it_flags & FMODE_EXEC)
438 och_p = &lli->lli_mds_exec_och;
440 och_p = &lli->lli_mds_read_och;
441 /* Check for the proper lock. */
442 ibits = MDS_INODELOCK_LOOKUP;
443 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
445 mutex_lock(&lli->lli_och_mutex);
446 if (*och_p) { /* Everything is open already, do nothing */
447 /* Originally it was idea to do not let them steal our
448 * open handle from under us by (*och_usecount)++ here.
449 * But in case we have the handle, but we cannot use it
450 * due to later checks (e.g. O_CREAT|O_EXCL flags set),
451 * nobody would decrement counter increased here. So we
452 * just hope the lock won't be invalidated in between.
453 * But if it would be, we'll reopen the open request to
454 * MDS later during file open path. */
455 mutex_unlock(&lli->lli_och_mutex);
458 mutex_unlock(&lli->lli_och_mutex);
461 if (it->it_op == IT_GETATTR) {
462 rc = ll_statahead_enter(parent, &de, 0);
465 else if (rc != -EAGAIN && rc != 0)
470 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
471 de->d_name.name, de->d_name.len,
472 0, LUSTRE_OPC_ANY, NULL);
474 RETURN(PTR_ERR(op_data));
476 if (!IS_POSIXACL(parent) || !exp_connect_umask(exp))
477 it->it_create_mode &= ~current_umask();
478 it->it_create_mode |= M_CHECK_STALE;
479 rc = md_intent_lock(exp, op_data, NULL, 0, it,
481 &req, ll_md_blocking_ast, 0);
482 it->it_create_mode &= ~M_CHECK_STALE;
483 ll_finish_md_op_data(op_data);
485 /* If req is NULL, then md_intent_lock only tried to do a lock match;
486 * if all was well, it will return 1 if it found locks, 0 otherwise. */
487 if (req == NULL && rc >= 0) {
495 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
496 "%d\n", rc, it->d.lustre.it_status);
502 rc = ll_revalidate_it_finish(req, it, de);
504 if (rc != -ESTALE && rc != -ENOENT)
505 ll_intent_release(it);
509 if ((it->it_op & IT_OPEN) && de->d_inode &&
510 !S_ISREG(de->d_inode->i_mode) &&
511 !S_ISDIR(de->d_inode->i_mode)) {
512 ll_release_openhandle(de, it);
517 /* We do not free request as it may be reused during following lookup
518 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
519 * be freed in ll_lookup_it or in ll_intent_release. But if
520 * request was not completed, we need to free it. (bug 5154, 9903) */
521 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
522 ptlrpc_req_finished(req);
524 /* mdt may grant layout lock for the newly created file, so
525 * release the lock to avoid leaking */
526 ll_intent_drop_lock(it);
527 ll_invalidate_aliases(de->d_inode);
530 __u64 matched_bits = 0;
532 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
533 "inode %p refc %d\n", de->d_name.len,
534 de->d_name.name, de, de->d_parent, de->d_inode,
537 ll_set_lock_data(exp, de->d_inode, it, &bits);
539 /* Note: We have to match both LOOKUP and PERM lock
540 * here to make sure the dentry is valid and no one
541 * changing the permission.
542 * But if the client connects < 2.4 server, which will
543 * only grant LOOKUP lock, so we can only Match LOOKUP
544 * lock for old server */
545 if (exp_connect_flags(ll_i2mdexp(de->d_inode)) &&
546 OBD_CONNECT_LVB_TYPE)
548 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
550 matched_bits = MDS_INODELOCK_LOOKUP;
552 if (((bits & matched_bits) == matched_bits) &&
553 d_lustre_invalid(de))
554 d_lustre_revalidate(de);
555 ll_lookup_finish_locks(it, de);
559 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
560 ll_statahead_mark(parent, de);
564 * This part is here to combat evil-evil race in real_lookup on 2.6
565 * kernels. The race details are: We enter do_lookup() looking for some
566 * name, there is nothing in dcache for this name yet and d_lookup()
567 * returns NULL. We proceed to real_lookup(), and while we do this,
568 * another process does open on the same file we looking up (most simple
569 * reproducer), open succeeds and the dentry is added. Now back to
570 * us. In real_lookup() we do d_lookup() again and suddenly find the
571 * dentry, so we call d_revalidate on it, but there is no lock, so
572 * without this code we would return 0, but unpatched real_lookup just
573 * returns -ENOENT in such a case instead of retrying the lookup. Once
574 * this is dealt with in real_lookup(), all of this ugly mess can go and
575 * we can just check locks in ->d_revalidate without doing any RPCs
579 if (it != &lookup_it) {
580 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
581 if (it->it_op == IT_GETATTR)
582 lookup_it.it_op = IT_GETATTR;
583 ll_lookup_finish_locks(it, de);
587 /* Do real lookup here. */
588 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
589 de->d_name.len, 0, (it->it_op & IT_CREAT ?
591 LUSTRE_OPC_ANY), NULL);
593 RETURN(PTR_ERR(op_data));
595 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
596 ll_md_blocking_ast, 0);
598 struct mdt_body *mdt_body;
599 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
600 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
603 fid = *ll_inode2fid(de->d_inode);
605 /* see if we got same inode, if not - return error */
606 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
607 ll_finish_md_op_data(op_data);
609 goto revalidate_finish;
611 ll_intent_release(it);
613 ll_finish_md_op_data(op_data);
618 * For rc == 1 case, should not return directly to prevent losing
619 * statahead windows; for rc == 0 case, the "lookup" will be done later.
621 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
622 ll_statahead_enter(parent, &de, 1);
626 #ifdef HAVE_IOP_ATOMIC_OPEN
628 * Always trust cached dentries. Update statahead window if necessary.
630 int ll_revalidate_nd(struct dentry *dentry, unsigned int flags)
632 struct inode *parent = dentry->d_parent->d_inode;
636 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,flags=%u\n",
637 dentry->d_name.name, flags);
639 if (!(flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE)) &&
640 ll_need_statahead(parent, dentry) > 0) {
641 if (flags & LOOKUP_RCU)
644 if (dentry->d_inode == NULL)
646 do_statahead_enter(parent, &dentry, unplug);
647 ll_statahead_mark(parent, dentry);
653 #else /* !HAVE_IOP_ATOMIC_OPEN */
654 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
659 #ifndef HAVE_DCACHE_LOCK
660 /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */
661 if (nd && (nd->flags & LOOKUP_RCU))
665 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
666 struct lookup_intent *it;
668 it = ll_convert_intent(&nd->intent.open, nd->flags);
672 if (it->it_op == (IT_OPEN|IT_CREAT) &&
673 nd->intent.open.flags & O_EXCL) {
674 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
679 rc = ll_revalidate_it(dentry, nd->flags, it);
681 if (rc && (nd->flags & LOOKUP_OPEN) &&
682 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
683 // XXX Code duplication with ll_lookup_nd
684 if (S_ISFIFO(dentry->d_inode->i_mode)) {
685 // We cannot call open here as it would
688 (struct ptlrpc_request *)
689 it->d.lustre.it_data);
693 nd->intent.open.file->private_data = it;
694 filp = lookup_instantiate_filp(nd, dentry,NULL);
699 if (!rc && (nd->flags & LOOKUP_CREATE) &&
700 it_disposition(it, DISP_OPEN_CREATE)) {
701 /* We created something but we may only return
702 * negative dentry here, so save request in dentry,
703 * if lookup will be called later on, it will
704 * pick the request, otherwise it would be freed
706 ll_d2d(dentry)->lld_it = it;
707 it = NULL; /* avoid freeing */
712 ll_intent_release(it);
713 OBD_FREE(it, sizeof(*it));
716 rc = ll_revalidate_it(dentry, 0, NULL);
721 #endif /* HAVE_IOP_ATOMIC_OPEN */
723 void ll_d_iput(struct dentry *de, struct inode *inode)
726 if (!find_cbdata(inode))
731 struct dentry_operations ll_d_ops = {
732 .d_revalidate = ll_revalidate_nd,
733 .d_release = ll_release,
734 .d_delete = ll_ddelete,
736 .d_compare = ll_dcompare,