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 /* kernel >= 2.6.38 last refcount is decreased after this function. */
175 #ifdef DCACHE_OP_DELETE
176 LASSERT(d_refcount(de) == 1);
178 LASSERT(d_refcount(de) == 0);
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 DCACHE_OP_HASH
237 LASSERT(de->d_op == &ll_d_ops);
239 de->d_op = &ll_d_ops;
244 void ll_intent_drop_lock(struct lookup_intent *it)
246 struct lustre_handle *handle;
248 if (it->it_op && it->d.lustre.it_lock_mode) {
249 handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
250 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
251 " from it %p\n", handle->cookie, it);
252 ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
254 /* bug 494: intent_release may be called multiple times, from
255 * this thread and we don't want to double-decref this lock */
256 it->d.lustre.it_lock_mode = 0;
260 void ll_intent_release(struct lookup_intent *it)
264 CDEBUG(D_INFO, "intent %p released\n", it);
265 ll_intent_drop_lock(it);
266 /* We are still holding extra reference on a request, need to free it */
267 if (it_disposition(it, DISP_ENQ_OPEN_REF))
268 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
269 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
270 ptlrpc_req_finished(it->d.lustre.it_data);
271 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
273 ptlrpc_req_finished(it->d.lustre.it_data);
275 it->d.lustre.it_disposition = 0;
276 it->d.lustre.it_data = NULL;
280 void ll_invalidate_aliases(struct inode *inode)
282 struct dentry *dentry;
285 LASSERT(inode != NULL);
287 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
288 inode->i_ino, inode->i_generation, inode);
290 ll_lock_dcache(inode);
291 cfs_list_for_each_entry(dentry, &inode->i_dentry, d_alias) {
292 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
293 "inode %p flags %d\n", dentry->d_name.len,
294 dentry->d_name.name, dentry, dentry->d_parent,
295 dentry->d_inode, dentry->d_flags);
297 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
298 CERROR("called on root (?) dentry=%p, inode=%p "
299 "ino=%lu\n", dentry, inode, inode->i_ino);
300 lustre_dump_dentry(dentry, 1);
301 libcfs_debug_dumpstack(NULL);
304 d_lustre_invalidate(dentry);
306 ll_unlock_dcache(inode);
311 int ll_revalidate_it_finish(struct ptlrpc_request *request,
312 struct lookup_intent *it,
321 if (it_disposition(it, DISP_LOOKUP_NEG))
324 rc = ll_prep_inode(&de->d_inode, request, NULL);
329 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
332 LASSERT(dentry != NULL);
334 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
335 struct inode *inode = dentry->d_inode;
336 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
338 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
339 inode, inode->i_ino, inode->i_generation);
340 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
343 /* drop lookup or getattr locks immediately */
344 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
345 /* on 2.6 there are situation when several lookups and
346 * revalidations may be requested during single operation.
347 * therefore, we don't release intent here -bzzz */
348 ll_intent_drop_lock(it);
352 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
354 struct lookup_intent *it = *itp;
356 if (!it || it->it_op == IT_GETXATTR)
361 int ll_revalidate_it(struct dentry *de, int lookup_flags,
362 struct lookup_intent *it)
364 struct md_op_data *op_data;
365 struct ptlrpc_request *req = NULL;
366 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
367 struct obd_export *exp;
368 struct inode *parent = de->d_parent->d_inode;
372 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
375 if (de->d_inode == NULL) {
378 /* We can only use negative dentries if this is stat or lookup,
379 for opens and stuff we do need to query server. */
380 /* If there is IT_CREAT in intent op set, then we must throw
381 away this negative dentry and actually do the request to
382 kernel to create whatever needs to be created (if possible)*/
383 if (it && (it->it_op & IT_CREAT))
386 if (d_lustre_invalid(de))
389 ibits = MDS_INODELOCK_UPDATE;
390 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
394 /* Never execute intents for mount points.
395 * Attributes will be fixed up in ll_inode_revalidate_it */
396 if (d_mountpoint(de))
397 GOTO(out_sa, rc = 1);
399 /* need to get attributes in case root got changed from other client */
400 if (de == de->d_sb->s_root) {
401 rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
407 exp = ll_i2mdexp(de->d_inode);
409 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
410 ll_frob_intent(&it, &lookup_it);
413 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
416 if ((it->it_op == IT_OPEN) && de->d_inode) {
417 struct inode *inode = de->d_inode;
418 struct ll_inode_info *lli = ll_i2info(inode);
419 struct obd_client_handle **och_p;
424 * We used to check for MDS_INODELOCK_OPEN here, but in fact
425 * just having LOOKUP lock is enough to justify inode is the
426 * same. And if inode is the same and we have suitable
427 * openhandle, then there is no point in doing another OPEN RPC
428 * just to throw away newly received openhandle. There are no
429 * security implications too, if file owner or access mode is
430 * change, LOOKUP lock is revoked.
434 if (it->it_flags & FMODE_WRITE) {
435 och_p = &lli->lli_mds_write_och;
436 och_usecount = &lli->lli_open_fd_write_count;
437 } else if (it->it_flags & FMODE_EXEC) {
438 och_p = &lli->lli_mds_exec_och;
439 och_usecount = &lli->lli_open_fd_exec_count;
441 och_p = &lli->lli_mds_read_och;
442 och_usecount = &lli->lli_open_fd_read_count;
444 /* Check for the proper lock. */
445 ibits = MDS_INODELOCK_LOOKUP;
446 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
448 cfs_mutex_lock(&lli->lli_och_mutex);
449 if (*och_p) { /* Everything is open already, do nothing */
450 /*(*och_usecount)++; Do not let them steal our open
451 handle from under us */
452 SET_BUT_UNUSED(och_usecount);
453 /* XXX The code above was my original idea, but in case
454 we have the handle, but we cannot use it due to later
455 checks (e.g. O_CREAT|O_EXCL flags set), nobody
456 would decrement counter increased here. So we just
457 hope the lock won't be invalidated in between. But
458 if it would be, we'll reopen the open request to
459 MDS later during file open path */
460 cfs_mutex_unlock(&lli->lli_och_mutex);
463 cfs_mutex_unlock(&lli->lli_och_mutex);
467 if (it->it_op == IT_GETATTR) {
468 rc = ll_statahead_enter(parent, &de, 0);
471 else if (rc != -EAGAIN && rc != 0)
476 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
477 de->d_name.name, de->d_name.len,
478 0, LUSTRE_OPC_ANY, NULL);
480 RETURN(PTR_ERR(op_data));
482 it->it_create_mode &= ~cfs_curproc_umask();
483 it->it_create_mode |= M_CHECK_STALE;
484 rc = md_intent_lock(exp, op_data, NULL, 0, it,
486 &req, ll_md_blocking_ast, 0);
487 it->it_create_mode &= ~M_CHECK_STALE;
488 ll_finish_md_op_data(op_data);
490 /* If req is NULL, then md_intent_lock only tried to do a lock match;
491 * if all was well, it will return 1 if it found locks, 0 otherwise. */
492 if (req == NULL && rc >= 0) {
500 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
501 "%d\n", rc, it->d.lustre.it_status);
507 rc = ll_revalidate_it_finish(req, it, de);
509 if (rc != -ESTALE && rc != -ENOENT)
510 ll_intent_release(it);
514 if ((it->it_op & IT_OPEN) && de->d_inode &&
515 !S_ISREG(de->d_inode->i_mode) &&
516 !S_ISDIR(de->d_inode->i_mode)) {
517 ll_release_openhandle(de, it);
522 /* We do not free request as it may be reused during following lookup
523 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
524 * be freed in ll_lookup_it or in ll_intent_release. But if
525 * request was not completed, we need to free it. (bug 5154, 9903) */
526 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
527 ptlrpc_req_finished(req);
529 ll_invalidate_aliases(de->d_inode);
533 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
534 "inode %p refc %d\n", de->d_name.len,
535 de->d_name.name, de, de->d_parent, de->d_inode,
537 ll_set_lock_data(exp, de->d_inode, it, &bits);
538 if ((bits & MDS_INODELOCK_LOOKUP) && d_lustre_invalid(de))
539 d_lustre_revalidate(de);
540 ll_lookup_finish_locks(it, de);
544 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
545 ll_statahead_mark(parent, de);
549 * This part is here to combat evil-evil race in real_lookup on 2.6
550 * kernels. The race details are: We enter do_lookup() looking for some
551 * name, there is nothing in dcache for this name yet and d_lookup()
552 * returns NULL. We proceed to real_lookup(), and while we do this,
553 * another process does open on the same file we looking up (most simple
554 * reproducer), open succeeds and the dentry is added. Now back to
555 * us. In real_lookup() we do d_lookup() again and suddenly find the
556 * dentry, so we call d_revalidate on it, but there is no lock, so
557 * without this code we would return 0, but unpatched real_lookup just
558 * returns -ENOENT in such a case instead of retrying the lookup. Once
559 * this is dealt with in real_lookup(), all of this ugly mess can go and
560 * we can just check locks in ->d_revalidate without doing any RPCs
564 if (it != &lookup_it) {
565 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
566 if (it->it_op == IT_GETATTR)
567 lookup_it.it_op = IT_GETATTR;
568 ll_lookup_finish_locks(it, de);
572 /* Do real lookup here. */
573 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
574 de->d_name.len, 0, (it->it_op & IT_CREAT ?
576 LUSTRE_OPC_ANY), NULL);
578 RETURN(PTR_ERR(op_data));
580 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
581 ll_md_blocking_ast, 0);
583 struct mdt_body *mdt_body;
584 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
585 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
588 fid = *ll_inode2fid(de->d_inode);
590 /* see if we got same inode, if not - return error */
591 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
592 ll_finish_md_op_data(op_data);
594 goto revalidate_finish;
596 ll_intent_release(it);
598 ll_finish_md_op_data(op_data);
603 * For rc == 1 case, should not return directly to prevent losing
604 * statahead windows; for rc == 0 case, the "lookup" will be done later.
606 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
607 ll_statahead_enter(parent, &de, 1);
611 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
617 if (nd->flags & LOOKUP_RCU)
621 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
622 struct lookup_intent *it;
624 it = ll_convert_intent(&nd->intent.open, nd->flags);
628 if (it->it_op == (IT_OPEN|IT_CREAT) &&
629 nd->intent.open.flags & O_EXCL) {
630 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
635 rc = ll_revalidate_it(dentry, nd->flags, it);
637 if (rc && (nd->flags & LOOKUP_OPEN) &&
638 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
639 // XXX Code duplication with ll_lookup_nd
640 if (S_ISFIFO(dentry->d_inode->i_mode)) {
641 // We cannot call open here as it would
644 (struct ptlrpc_request *)
645 it->d.lustre.it_data);
649 nd->intent.open.file->private_data = it;
650 filp = lookup_instantiate_filp(nd, dentry,NULL);
655 if (!rc && (nd->flags & LOOKUP_CREATE) &&
656 it_disposition(it, DISP_OPEN_CREATE)) {
657 /* We created something but we may only return
658 * negative dentry here, so save request in dentry,
659 * if lookup will be called later on, it will
660 * pick the request, otherwise it would be freed
662 ll_d2d(dentry)->lld_it = it;
663 it = NULL; /* avoid freeing */
668 ll_intent_release(it);
669 OBD_FREE(it, sizeof(*it));
672 rc = ll_revalidate_it(dentry, 0, NULL);
678 void ll_d_iput(struct dentry *de, struct inode *inode)
681 if (!find_cbdata(inode))
686 struct dentry_operations ll_d_ops = {
687 .d_revalidate = ll_revalidate_nd,
688 .d_release = ll_release,
689 .d_delete = ll_ddelete,
691 .d_compare = ll_dcompare,