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_sb_info *sbi = ll_i2sbi(inode);
139 struct lov_stripe_md *lsm;
144 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
145 return_if_equal, NULL);
149 lsm = ccc_inode_lsm_get(inode);
153 rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL);
154 ccc_inode_lsm_put(inode, lsm);
160 * Called when last reference to a dentry is dropped and dcache wants to know
161 * whether or not it should cache it:
162 * - return 1 to delete the dentry immediately
163 * - return 0 to cache the dentry
164 * Should NOT be called with the dcache lock, see fs/dcache.c
166 static int ll_ddelete(HAVE_D_DELETE_CONST struct dentry *de)
171 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
172 d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
173 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
174 d_unhashed((struct dentry *)de) ? "" : "hashed,",
175 list_empty(&de->d_subdirs) ? "" : "subdirs");
177 #ifdef HAVE_DCACHE_LOCK
178 LASSERT(d_refcount(de) == 0);
180 /* kernel >= 2.6.38 last refcount is decreased after this function. */
181 LASSERT(d_refcount(de) == 1);
184 /* Disable this piece of code temproarily because this is called
185 * inside dcache_lock so it's not appropriate to do lots of work
188 /* if not ldlm lock for this inode, set i_nlink to 0 so that
189 * this inode can be recycled later b=20433 */
190 if (de->d_inode && !find_cbdata(de->d_inode))
191 de->d_inode->i_nlink = 0;
194 if (d_lustre_invalid((struct dentry *)de))
199 static int ll_set_dd(struct dentry *de)
204 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
205 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
208 if (de->d_fsdata == NULL) {
209 struct ll_dentry_data *lld;
212 if (likely(lld != NULL)) {
213 spin_lock(&de->d_lock);
214 if (likely(de->d_fsdata == NULL))
218 spin_unlock(&de->d_lock);
227 int ll_dops_init(struct dentry *de, int block, int init_sa)
229 struct ll_dentry_data *lld = ll_d2d(de);
232 if (lld == NULL && block != 0) {
240 if (lld != NULL && init_sa != 0)
241 lld->lld_sa_generation = 0;
243 #ifdef HAVE_DCACHE_LOCK
244 de->d_op = &ll_d_ops;
246 /* kernel >= 2.6.38 d_op is set in d_alloc() */
247 LASSERT(de->d_op == &ll_d_ops);
252 void ll_intent_drop_lock(struct lookup_intent *it)
254 struct lustre_handle *handle;
256 if (it->it_op && it->d.lustre.it_lock_mode) {
257 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
258 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
259 " from it %p\n", handle->cookie, it);
260 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
262 /* bug 494: intent_release may be called multiple times, from
263 * this thread and we don't want to double-decref this lock */
264 it->d.lustre.it_lock_mode = 0;
268 void ll_intent_release(struct lookup_intent *it)
272 CDEBUG(D_INFO, "intent %p released\n", it);
273 ll_intent_drop_lock(it);
274 /* We are still holding extra reference on a request, need to free it */
275 if (it_disposition(it, DISP_ENQ_OPEN_REF))
276 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
277 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
278 ptlrpc_req_finished(it->d.lustre.it_data);
279 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
281 ptlrpc_req_finished(it->d.lustre.it_data);
283 it->d.lustre.it_disposition = 0;
284 it->d.lustre.it_data = NULL;
288 void ll_invalidate_aliases(struct inode *inode)
290 struct dentry *dentry;
293 LASSERT(inode != NULL);
295 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
296 inode->i_ino, inode->i_generation, inode);
298 ll_lock_dcache(inode);
299 cfs_list_for_each_entry(dentry, &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 (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
306 CERROR("called on root (?) dentry=%p, inode=%p "
307 "ino=%lu\n", dentry, inode, inode->i_ino);
308 lustre_dump_dentry(dentry, 1);
309 libcfs_debug_dumpstack(NULL);
312 d_lustre_invalidate(dentry);
314 ll_unlock_dcache(inode);
319 int ll_revalidate_it_finish(struct ptlrpc_request *request,
320 struct lookup_intent *it,
329 if (it_disposition(it, DISP_LOOKUP_NEG))
332 rc = ll_prep_inode(&de->d_inode, request, NULL);
337 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
340 LASSERT(dentry != NULL);
342 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
343 struct inode *inode = dentry->d_inode;
344 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
346 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
347 inode, inode->i_ino, inode->i_generation);
348 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
351 /* drop lookup or getattr locks immediately */
352 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
353 /* on 2.6 there are situation when several lookups and
354 * revalidations may be requested during single operation.
355 * therefore, we don't release intent here -bzzz */
356 ll_intent_drop_lock(it);
360 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
362 struct lookup_intent *it = *itp;
364 if (!it || it->it_op == IT_GETXATTR)
369 int ll_revalidate_it(struct dentry *de, int lookup_flags,
370 struct lookup_intent *it)
372 struct md_op_data *op_data;
373 struct ptlrpc_request *req = NULL;
374 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
375 struct obd_export *exp;
376 struct inode *parent = de->d_parent->d_inode;
380 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
383 if (de->d_inode == NULL) {
386 /* We can only use negative dentries if this is stat or lookup,
387 for opens and stuff we do need to query server. */
388 /* If there is IT_CREAT in intent op set, then we must throw
389 away this negative dentry and actually do the request to
390 kernel to create whatever needs to be created (if possible)*/
391 if (it && (it->it_op & IT_CREAT))
394 if (d_lustre_invalid(de))
397 ibits = MDS_INODELOCK_UPDATE;
398 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
402 /* Never execute intents for mount points.
403 * Attributes will be fixed up in ll_inode_revalidate_it */
404 if (d_mountpoint(de))
405 GOTO(out_sa, rc = 1);
407 /* need to get attributes in case root got changed from other client */
408 if (de == de->d_sb->s_root) {
409 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
415 exp = ll_i2mdexp(de->d_inode);
417 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
418 ll_frob_intent(&it, &lookup_it);
421 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
424 if ((it->it_op == IT_OPEN) && de->d_inode) {
425 struct inode *inode = de->d_inode;
426 struct ll_inode_info *lli = ll_i2info(inode);
427 struct obd_client_handle **och_p;
432 * We used to check for MDS_INODELOCK_OPEN here, but in fact
433 * just having LOOKUP lock is enough to justify inode is the
434 * same. And if inode is the same and we have suitable
435 * openhandle, then there is no point in doing another OPEN RPC
436 * just to throw away newly received openhandle. There are no
437 * security implications too, if file owner or access mode is
438 * change, LOOKUP lock is revoked.
442 if (it->it_flags & FMODE_WRITE) {
443 och_p = &lli->lli_mds_write_och;
444 och_usecount = &lli->lli_open_fd_write_count;
445 } else if (it->it_flags & FMODE_EXEC) {
446 och_p = &lli->lli_mds_exec_och;
447 och_usecount = &lli->lli_open_fd_exec_count;
449 och_p = &lli->lli_mds_read_och;
450 och_usecount = &lli->lli_open_fd_read_count;
452 /* Check for the proper lock. */
453 ibits = MDS_INODELOCK_LOOKUP;
454 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
456 cfs_mutex_lock(&lli->lli_och_mutex);
457 if (*och_p) { /* Everything is open already, do nothing */
458 /*(*och_usecount)++; Do not let them steal our open
459 handle from under us */
460 SET_BUT_UNUSED(och_usecount);
461 /* XXX The code above was my original idea, but in case
462 we have the handle, but we cannot use it due to later
463 checks (e.g. O_CREAT|O_EXCL flags set), nobody
464 would decrement counter increased here. So we just
465 hope the lock won't be invalidated in between. But
466 if it would be, we'll reopen the open request to
467 MDS later during file open path */
468 cfs_mutex_unlock(&lli->lli_och_mutex);
471 cfs_mutex_unlock(&lli->lli_och_mutex);
475 if (it->it_op == IT_GETATTR) {
476 rc = ll_statahead_enter(parent, &de, 0);
479 else if (rc != -EAGAIN && rc != 0)
484 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
485 de->d_name.name, de->d_name.len,
486 0, LUSTRE_OPC_ANY, NULL);
488 RETURN(PTR_ERR(op_data));
490 it->it_create_mode &= ~cfs_curproc_umask();
491 it->it_create_mode |= M_CHECK_STALE;
492 rc = md_intent_lock(exp, op_data, NULL, 0, it,
494 &req, ll_md_blocking_ast, 0);
495 it->it_create_mode &= ~M_CHECK_STALE;
496 ll_finish_md_op_data(op_data);
498 /* If req is NULL, then md_intent_lock only tried to do a lock match;
499 * if all was well, it will return 1 if it found locks, 0 otherwise. */
500 if (req == NULL && rc >= 0) {
508 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
509 "%d\n", rc, it->d.lustre.it_status);
515 rc = ll_revalidate_it_finish(req, it, de);
517 if (rc != -ESTALE && rc != -ENOENT)
518 ll_intent_release(it);
522 if ((it->it_op & IT_OPEN) && de->d_inode &&
523 !S_ISREG(de->d_inode->i_mode) &&
524 !S_ISDIR(de->d_inode->i_mode)) {
525 ll_release_openhandle(de, it);
530 /* We do not free request as it may be reused during following lookup
531 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
532 * be freed in ll_lookup_it or in ll_intent_release. But if
533 * request was not completed, we need to free it. (bug 5154, 9903) */
534 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
535 ptlrpc_req_finished(req);
537 ll_invalidate_aliases(de->d_inode);
541 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
542 "inode %p refc %d\n", de->d_name.len,
543 de->d_name.name, de, de->d_parent, de->d_inode,
545 ll_set_lock_data(exp, de->d_inode, it, &bits);
546 if ((bits & MDS_INODELOCK_LOOKUP) && d_lustre_invalid(de))
547 d_lustre_revalidate(de);
548 ll_lookup_finish_locks(it, de);
552 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
553 ll_statahead_mark(parent, de);
557 * This part is here to combat evil-evil race in real_lookup on 2.6
558 * kernels. The race details are: We enter do_lookup() looking for some
559 * name, there is nothing in dcache for this name yet and d_lookup()
560 * returns NULL. We proceed to real_lookup(), and while we do this,
561 * another process does open on the same file we looking up (most simple
562 * reproducer), open succeeds and the dentry is added. Now back to
563 * us. In real_lookup() we do d_lookup() again and suddenly find the
564 * dentry, so we call d_revalidate on it, but there is no lock, so
565 * without this code we would return 0, but unpatched real_lookup just
566 * returns -ENOENT in such a case instead of retrying the lookup. Once
567 * this is dealt with in real_lookup(), all of this ugly mess can go and
568 * we can just check locks in ->d_revalidate without doing any RPCs
572 if (it != &lookup_it) {
573 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
574 if (it->it_op == IT_GETATTR)
575 lookup_it.it_op = IT_GETATTR;
576 ll_lookup_finish_locks(it, de);
580 /* Do real lookup here. */
581 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
582 de->d_name.len, 0, (it->it_op & IT_CREAT ?
584 LUSTRE_OPC_ANY), NULL);
586 RETURN(PTR_ERR(op_data));
588 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
589 ll_md_blocking_ast, 0);
591 struct mdt_body *mdt_body;
592 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
593 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
596 fid = *ll_inode2fid(de->d_inode);
598 /* see if we got same inode, if not - return error */
599 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
600 ll_finish_md_op_data(op_data);
602 goto revalidate_finish;
604 ll_intent_release(it);
606 ll_finish_md_op_data(op_data);
611 * For rc == 1 case, should not return directly to prevent losing
612 * statahead windows; for rc == 0 case, the "lookup" will be done later.
614 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
615 ll_statahead_enter(parent, &de, 1);
619 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
624 #ifndef HAVE_DCACHE_LOCK
625 /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */
626 if (nd->flags & LOOKUP_RCU)
630 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
631 struct lookup_intent *it;
633 it = ll_convert_intent(&nd->intent.open, nd->flags);
637 if (it->it_op == (IT_OPEN|IT_CREAT) &&
638 nd->intent.open.flags & O_EXCL) {
639 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
644 rc = ll_revalidate_it(dentry, nd->flags, it);
646 if (rc && (nd->flags & LOOKUP_OPEN) &&
647 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
648 // XXX Code duplication with ll_lookup_nd
649 if (S_ISFIFO(dentry->d_inode->i_mode)) {
650 // We cannot call open here as it would
653 (struct ptlrpc_request *)
654 it->d.lustre.it_data);
658 nd->intent.open.file->private_data = it;
659 filp = lookup_instantiate_filp(nd, dentry,NULL);
664 if (!rc && (nd->flags & LOOKUP_CREATE) &&
665 it_disposition(it, DISP_OPEN_CREATE)) {
666 /* We created something but we may only return
667 * negative dentry here, so save request in dentry,
668 * if lookup will be called later on, it will
669 * pick the request, otherwise it would be freed
671 ll_d2d(dentry)->lld_it = it;
672 it = NULL; /* avoid freeing */
677 ll_intent_release(it);
678 OBD_FREE(it, sizeof(*it));
681 rc = ll_revalidate_it(dentry, 0, NULL);
687 void ll_d_iput(struct dentry *de, struct inode *inode)
690 if (!find_cbdata(inode))
695 struct dentry_operations ll_d_ops = {
696 .d_revalidate = ll_revalidate_nd,
697 .d_release = ll_release,
698 .d_delete = ll_ddelete,
700 .d_compare = ll_dcompare,