1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
7 * Author: Andreas Dilger <adilger@clusterfs.com>
9 * This file is part of Lustre, http://www.lustre.org.
11 * Lustre is free software; you can redistribute it and/or
12 * modify it under the terms of version 2 of the GNU General Public
13 * License as published by the Free Software Foundation.
15 * Lustre is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with Lustre; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #define DEBUG_SUBSYSTEM S_LLITE
26 #include <linux/lustre_dlm.h>
27 #include <linux/lustre_lite.h>
28 #include <linux/pagemap.h>
29 #include <linux/file.h>
30 #include <linux/lustre_acl.h>
31 #include <linux/lustre_sec.h>
32 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
33 #include <linux/lustre_compat25.h>
35 #include <linux/obd_lov.h>
36 #include <linux/lustre_audit.h>
37 #include "llite_internal.h"
39 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
40 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
41 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
43 int ll_validate_size(struct inode *inode, __u64 *size, __u64 *blocks)
45 ldlm_policy_data_t extent = { .l_extent = { 0, OBD_OBJECT_EOF } };
46 struct obd_export *exp = ll_i2sbi(inode)->ll_dt_exp;
47 struct ll_inode_info *lli = ll_i2info(inode);
48 struct lustre_handle match_lockh = {0};
52 if (lli->lli_smd == NULL)
55 LASSERT(size != NULL && blocks != NULL);
57 flags = LDLM_FL_TEST_LOCK | LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
58 rc = obd_match(exp, lli->lli_smd, LDLM_EXTENT, &extent,
59 LCK_PR | LCK_PW, &flags, inode, &match_lockh);
61 /* we have no all needed locks,
62 * so we don't know actual size */
66 /* we know actual size! */
67 down(&lli->lli_size_sem);
68 *size = lov_merge_size(lli->lli_smd, 0);
69 *blocks = lov_merge_blocks(lli->lli_smd);
70 up(&lli->lli_size_sem);
76 int ll_md_och_close(struct obd_export *md_exp, struct inode *inode,
77 struct obd_client_handle *och, int dirty)
79 struct ptlrpc_request *req = NULL;
80 struct obdo *obdo = NULL;
81 struct obd_device *obd;
85 obd = class_exp2obd(md_exp);
87 CERROR("Invalid MDC connection handle "LPX64"\n",
88 md_exp->exp_handle.h_cookie);
94 * here we check if this is forced umount. If so this is called on
95 * canceling "open lock" and we do not call md_close() in this case , as
96 * it will not successful, as import is already deactivated.
98 if (obd->obd_no_recov)
101 /* closing opened file */
106 obdo->o_id = inode->i_ino;
107 obdo->o_generation = inode->i_generation;
108 obdo->o_valid = OBD_MD_FLID;
109 obdo_from_inode(obdo, inode, (OBD_MD_FLTYPE | OBD_MD_FLMODE |
110 OBD_MD_FLATIME | OBD_MD_FLMTIME |
112 if (0 /* ll_is_inode_dirty(inode) */) {
113 obdo->o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
114 obdo->o_valid |= OBD_MD_FLFLAGS;
116 obdo->o_fid = id_fid(&ll_i2info(inode)->lli_id);
117 obdo->o_mds = id_group(&ll_i2info(inode)->lli_id);
119 obdo->o_valid |= OBD_MD_FLEPOCH;
120 obdo->o_easize = ll_i2info(inode)->lli_io_epoch;
123 /* we modified data through this handle */
124 obdo->o_flags |= MDS_BFLAG_DIRTY_EPOCH;
125 obdo->o_valid |= OBD_MD_FLFLAGS;
126 if (ll_validate_size(inode, &obdo->o_size, &obdo->o_blocks))
127 obdo->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
130 rc = md_close(md_exp, obdo, och, &req);
135 * we are the last writer, so the MDS has instructed us to get
136 * the file size and any write cookies, then close again.
139 //ll_queue_done_writing(inode);
142 CERROR("inode %lu mdc close failed: rc = %d\n",
143 (unsigned long)inode->i_ino, rc);
146 ptlrpc_req_finished(req);
149 mdc_clear_open_replay_data(md_exp, och);
151 capa_put(och->och_capa, CLIENT_CAPA);
152 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
153 OBD_FREE(och, sizeof *och);
157 int ll_md_real_close(struct obd_export *md_exp,
158 struct inode *inode, int flags)
160 struct ll_inode_info *lli = ll_i2info(inode);
161 int freeing = inode->i_state & I_FREEING;
162 struct obd_client_handle **och_p;
163 struct obd_client_handle *och;
165 int rc = 0, dirty = 0;
168 if (flags & FMODE_WRITE) {
169 och_p = &lli->lli_mds_write_och;
170 och_usecount = &lli->lli_open_fd_write_count;
171 } else if (flags & FMODE_EXEC) {
172 och_p = &lli->lli_mds_exec_och;
173 och_usecount = &lli->lli_open_fd_exec_count;
175 och_p = &lli->lli_mds_read_och;
176 och_usecount = &lli->lli_open_fd_read_count;
179 down(&lli->lli_och_sem);
180 if (*och_usecount) { /* There are still users of this handle, so
182 up(&lli->lli_och_sem);
185 if (ll_is_inode_dirty(inode)) {
186 /* the inode still has dirty pages, let's close later */
187 CDEBUG(D_INODE, "inode %lu/%u still has dirty pages\n",
188 inode->i_ino, inode->i_generation);
189 LASSERT(freeing == 0);
190 ll_queue_done_writing(inode);
191 up(&lli->lli_och_sem);
195 if (LLI_DIRTY_HANDLE(inode) && (flags & FMODE_WRITE)) {
196 clear_bit(LLI_F_DIRTY_HANDLE, &lli->lli_flags);
198 } else if (0 && !(flags & FMODE_SYNC) && !freeing) {
199 /* in order to speed up creation rate we pass
200 * closing to dedicated thread so we don't need
201 * to wait for close reply here -bzzz */
202 ll_queue_done_writing(inode);
203 up(&lli->lli_och_sem);
211 up(&lli->lli_och_sem);
214 * there might be a race and somebody have freed this och
215 * already. Another way to have this twice called is if file closing
216 * will fail due to netwok problems and on umount lock will be canceled
217 * and this will be called from block_ast callack.
219 if (och && och->och_fh.cookie != DEAD_HANDLE_MAGIC)
220 rc = ll_md_och_close(md_exp, inode, och, dirty);
225 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
228 struct ll_file_data *fd = file->private_data;
229 struct ll_inode_info *lli = ll_i2info(inode);
233 /* clear group lock, if present */
234 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
235 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
236 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
237 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
241 /* Let's see if we have good enough OPEN lock on the file and if
242 we can skip talking to MDS */
243 if (file->f_dentry->d_inode) {
245 struct obd_device *obddev;
246 struct lustre_handle lockh;
247 int flags = LDLM_FL_BLOCK_GRANTED;
248 struct ldlm_res_id file_res_id = {.name = {id_fid(&lli->lli_id),
249 id_group(&lli->lli_id)}};
250 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
252 down(&lli->lli_och_sem);
253 if (fd->fd_omode & FMODE_WRITE) {
255 LASSERT(lli->lli_open_fd_write_count);
256 lli->lli_open_fd_write_count--;
257 } else if (fd->fd_omode & FMODE_EXEC) {
259 LASSERT(lli->lli_open_fd_exec_count);
260 lli->lli_open_fd_exec_count--;
263 LASSERT(lli->lli_open_fd_read_count);
264 lli->lli_open_fd_read_count--;
266 up(&lli->lli_och_sem);
268 obddev = md_get_real_obd(md_exp, &lli->lli_id);
269 if (!ldlm_lock_match(obddev->obd_namespace, flags, &file_res_id,
270 LDLM_IBITS, &policy, lockmode, &lockh))
272 rc = ll_md_real_close(md_exp, file->f_dentry->d_inode,
275 ldlm_lock_decref(&lockh, lockmode);
279 file->private_data = NULL;
280 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof(*fd));
284 /* While this returns an error code, fput() the caller does not, so we need
285 * to make every effort to clean up all of our state here. Also, applications
286 * rarely check close errors and even if an error is returned they will not
287 * re-try the close call.
289 int ll_file_release(struct inode *inode, struct file *file)
291 struct ll_file_data *fd;
292 struct ll_sb_info *sbi = ll_i2sbi(inode);
296 CDEBUG(D_VFSTRACE, "VFS Op:inode="DLID4"(%p)\n",
297 OLID4(&ll_i2info(inode)->lli_id), inode);
299 /* don't do anything for / */
300 if (inode->i_sb->s_root == file->f_dentry)
303 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
304 fd = (struct ll_file_data *)file->private_data;
307 rc = ll_md_close(sbi->ll_md_exp, inode, file);
311 static int ll_intent_file_open(struct file *file, void *lmm,
312 int lmmsize, struct lookup_intent *itp)
314 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
315 const char *name = (char *)file->f_dentry->d_name.name;
316 struct dentry *parent = file->f_dentry->d_parent;
317 const int len = file->f_dentry->d_name.len;
318 struct lustre_handle lockh;
319 struct mdc_op_data *op_data;
325 OBD_ALLOC(op_data, sizeof(*op_data));
329 ll_prepare_mdc_data(op_data, parent->d_inode, NULL,
332 rc = md_enqueue(sbi->ll_md_exp, LDLM_IBITS, itp, LCK_PR, op_data,
333 &lockh, lmm, lmmsize, ldlm_completion_ast,
334 ll_mdc_blocking_ast, NULL);
335 OBD_FREE(op_data, sizeof(*op_data));
337 if (LUSTRE_IT(itp)->it_lock_mode)
338 memcpy(&LUSTRE_IT(itp)->it_lock_handle,
339 &lockh, sizeof(lockh));
342 CERROR("lock enqueue: err: %d\n", rc);
347 int ll_och_fill(struct inode *inode, struct lookup_intent *it,
348 struct obd_client_handle *och)
350 struct ptlrpc_request *req = LUSTRE_IT(it)->it_data;
351 struct ll_inode_info *lli = ll_i2info(inode);
352 struct mds_body *body;
357 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
358 LASSERT (body != NULL); /* reply already checked out */
359 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
361 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
362 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
363 lli->lli_io_epoch = body->io_epoch;
364 mdc_set_open_replay_data(ll_i2mdexp(inode), och,
365 LUSTRE_IT(it)->it_data);
367 if (S_ISREG(inode->i_mode) && (body->valid & OBD_MD_CAPA))
368 rc = ll_set_och_capa(inode, it, och);
372 int ll_local_open(struct file *file, struct lookup_intent *it,
373 struct obd_client_handle *och)
375 struct ll_file_data *fd;
380 rc = ll_och_fill(file->f_dentry->d_inode, it, och);
385 LASSERTF(file->private_data == NULL, "file %.*s/%.*s ino %lu/%u (%o)\n",
386 file->f_dentry->d_name.len, file->f_dentry->d_name.name,
387 file->f_dentry->d_parent->d_name.len,
388 file->f_dentry->d_parent->d_name.name,
389 file->f_dentry->d_inode->i_ino,
390 file->f_dentry->d_inode->i_generation,
391 file->f_dentry->d_inode->i_mode);
393 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
395 /* We can't handle this well without reorganizing ll_file_open and
396 * ll_md_close(), so don't even try right now. */
399 file->private_data = fd;
400 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
401 fd->fd_omode = it->it_flags;
406 /* Open a file, and (for the very first open) create objects on the OSTs at
407 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
408 * creation or open until ll_lov_setstripe() ioctl is called. We grab
409 * lli_open_sem to ensure no other process will create objects, send the
410 * stripe MD to the MDS, or try to destroy the objects if that fails.
412 * If we already have the stripe MD locally then we don't request it in
413 * mdc_open(), by passing a lmm_size = 0.
415 * It is up to the application to ensure no other processes open this file
416 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
417 * used. We might be able to avoid races of that sort by getting lli_open_sem
418 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
419 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
421 int ll_file_open(struct inode *inode, struct file *file)
423 struct ll_inode_info *lli = ll_i2info(inode);
424 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
425 .it_flags = file->f_flags };
426 struct lov_stripe_md *lsm;
427 struct ptlrpc_request *req;
429 struct obd_client_handle **och_p = NULL;
430 __u64 *och_usecount = NULL;
433 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n",
434 inode->i_ino, inode->i_generation, inode, file->f_flags);
436 /* don't do anything for / */
437 if (inode->i_sb->s_root == file->f_dentry)
440 if ((file->f_flags+1) & O_ACCMODE)
442 if (file->f_flags & O_TRUNC)
447 * sometimes LUSTRE_IT(it) may not be allocated like opening file by
448 * dentry_open() from GNS stuff.
450 if (!it || !LUSTRE_IT(it)) {
452 rc = ll_intent_alloc(it);
457 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
460 * mdc_intent_lock() didn't get a request ref if there was an open
461 * error, so don't do cleanup on the * request here (bug 3430)
463 if (LUSTRE_IT(it)->it_disposition) {
464 rc = it_open_error(DISP_OPEN_OPEN, it);
469 /* Let's see if we have file open on MDS already. */
470 if (it->it_flags & FMODE_WRITE) {
471 och_p = &lli->lli_mds_write_och;
472 och_usecount = &lli->lli_open_fd_write_count;
473 } else if (it->it_flags & FMODE_EXEC) {
474 och_p = &lli->lli_mds_exec_och;
475 och_usecount = &lli->lli_open_fd_exec_count;
477 och_p = &lli->lli_mds_read_och;
478 och_usecount = &lli->lli_open_fd_read_count;
481 rc = ll_crypto_decrypt_key(inode, it);
485 down(&lli->lli_och_sem);
486 if (*och_p) { /* Open handle is present */
487 if (it_disposition(it, DISP_LOOKUP_POS) && /* Positive lookup */
488 it_disposition(it, DISP_OPEN_OPEN)) { /* & OPEN happened */
489 struct obd_client_handle *och;
490 /* Well, there's extra open request that we do not need,
491 let's close it somehow*/
492 OBD_ALLOC(och, sizeof (struct obd_client_handle));
494 up(&lli->lli_och_sem);
498 rc = ll_och_fill(inode, it, och);
500 up(&lli->lli_och_sem);
504 /* ll_md_och_close() will free och */
505 ll_md_och_close(ll_i2mdexp(inode), inode, och, 0);
509 rc = ll_local_open(file, it, NULL);
513 LASSERT(*och_usecount == 0);
514 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
516 GOTO(out, rc = -ENOMEM);
519 if (!it || !LUSTRE_IT(it) || !LUSTRE_IT(it)->it_disposition) {
521 * we are going to replace intent here, and that may
522 * possibly change access mode (FMODE_EXEC can only be
523 * set in intent), but I hope it never happens (I was
524 * not able to trigger it yet at least) -- green
527 /* FIXME: FMODE_EXEC is not covered by O_ACCMODE! */
528 LASSERT(!(it->it_flags & FMODE_EXEC));
529 LASSERTF((it->it_flags & O_ACCMODE) ==
530 (oit.it_flags & O_ACCMODE), "Changing intent "
531 "flags %x to incompatible %x\n", it->it_flags,
534 rc = ll_intent_file_open(file, NULL, 0, it);
537 rc = it_open_error(DISP_OPEN_OPEN, it);
539 GOTO(out_och_free, rc);
541 mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle,
542 file->f_dentry->d_inode);
544 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
545 rc = ll_local_open(file, it, *och_p);
546 LASSERTF(rc == 0, "rc = %d\n", rc);
548 up(&lli->lli_och_sem);
551 * must do this outside lli_och_sem lock to prevent deadlock where
552 * different kind of OPEN lock for this same inode gets cancelled by
555 if (!S_ISREG(inode->i_mode))
560 if (file->f_flags & O_LOV_DELAY_CREATE ||
561 !(file->f_mode & FMODE_WRITE)) {
562 CDEBUG(D_INODE, "object creation was delayed\n");
566 file->f_flags &= ~O_LOV_DELAY_CREATE;
569 /* audit stuff if there was no RPC */
570 if (LUSTRE_IT(it)->it_data == 0) {
571 ll_audit_log(inode, AUDIT_OPEN, rc);
574 req = LUSTRE_IT(it)->it_data;
575 ll_intent_drop_lock(it);
576 ll_intent_release(it);
577 ptlrpc_req_finished(req);
579 ll_open_complete(inode);
583 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
584 *och_p = NULL; /* OBD_FREE writes some magic there */
587 up(&lli->lli_och_sem);
593 /* Fills the obdo with the attributes for the inode defined by lsm */
594 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
597 struct ptlrpc_request_set *set;
601 LASSERT(lsm != NULL);
603 memset(oa, 0, sizeof *oa);
604 oa->o_id = lsm->lsm_object_id;
605 oa->o_gr = lsm->lsm_object_gr;
606 oa->o_mode = S_IFREG;
607 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
608 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
609 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
611 set = ptlrpc_prep_set();
615 rc = obd_getattr_async(exp, oa, lsm, set);
617 rc = ptlrpc_set_wait(set);
618 ptlrpc_set_destroy(set);
623 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
624 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
628 static inline void ll_remove_suid(struct inode *inode)
632 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
633 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
635 /* was any of the uid bits set? */
636 mode &= inode->i_mode;
637 if (mode && !capable(CAP_FSETID)) {
638 inode->i_mode &= ~mode;
639 // XXX careful here - we cannot change the size
643 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
645 struct ll_inode_info *lli = ll_i2info(inode);
646 struct lov_stripe_md *lsm = lli->lli_smd;
647 struct obd_export *exp = ll_i2dtexp(inode);
650 struct ldlm_lock *lock;
651 struct lov_stripe_md *lsm;
652 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
653 __u32 stripe, vallen = sizeof(stripe);
657 if (lsm->lsm_stripe_count == 1)
658 GOTO(check, stripe = 0);
660 /* get our offset in the lov */
661 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
663 CERROR("obd_get_info: rc = %d\n", rc);
666 LASSERT(stripe < lsm->lsm_stripe_count);
669 if (lsm->lsm_oinfo[stripe].loi_id != lock->l_resource->lr_name.name[0]||
670 lsm->lsm_oinfo[stripe].loi_gr != lock->l_resource->lr_name.name[2]){
671 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64
672 " inode=%lu/%u (%p)\n",
673 lsm->lsm_oinfo[stripe].loi_id,
674 lsm->lsm_oinfo[stripe].loi_gr,
675 inode->i_ino, inode->i_generation, inode);
676 return -ELDLM_NO_LOCK_DATA;
682 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
683 * we get a lock cancellation for each stripe, so we have to map the obd's
684 * region back onto the stripes in the file that it held.
686 * No one can dirty the extent until we've finished our work and they can
687 * enqueue another lock. The DLM protects us from ll_file_read/write here,
688 * but other kernel actors could have pages locked.
690 * Called with the DLM lock held. */
691 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
692 struct ldlm_lock *lock, __u32 stripe)
694 ldlm_policy_data_t tmpex;
695 unsigned long start, end, count, skip, i, j;
697 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
698 struct lustre_handle lockh;
701 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
702 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
703 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
706 /* our locks are page granular thanks to osc_enqueue, we invalidate the
708 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
709 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
713 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
714 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
715 if (lsm->lsm_stripe_count > 1) {
716 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
717 skip = (lsm->lsm_stripe_count - 1) * count;
718 start += start/count * skip + stripe * count;
720 end += end/count * skip + stripe * count;
722 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
725 i = inode->i_size ? (inode->i_size - 1) >> PAGE_CACHE_SHIFT : 0;
729 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
730 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
731 count, skip, end, discard ? " (DISCARDING)" : "");
733 /* walk through the vmas on the inode and tear down mmaped pages that
734 * intersect with the lock. this stops immediately if there are no
735 * mmap()ed regions of the file. This is not efficient at all and
736 * should be short lived. We'll associate mmap()ed pages with the lock
737 * and will be able to find them directly */
739 for (i = start; i <= end; i += (j + skip)) {
740 j = min(count - (i % count), end - i + 1);
742 LASSERT(inode->i_mapping);
743 if (ll_teardown_mmaps(inode->i_mapping,
744 (__u64)i << PAGE_CACHE_SHIFT,
745 ((__u64)(i+j) << PAGE_CACHE_SHIFT) - 1) )
749 /* this is the simplistic implementation of page eviction at
750 * cancelation. It is careful to get races with other page
751 * lockers handled correctly. fixes from bug 20 will make it
752 * more efficient by associating locks with pages and with
753 * batching writeback under the lock explicitly. */
754 for (i = start, j = start % count; i <= end;
755 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
757 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
763 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
764 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
765 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
768 if (!mapping_has_pages(inode->i_mapping)) {
769 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
775 page = find_get_page(inode->i_mapping, i);
778 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
779 i, tmpex.l_extent.start);
782 /* page->mapping to check with racing against teardown */
783 if (!discard && clear_page_dirty_for_io(page)) {
784 rc = ll_call_writepage(inode, page);
786 CERROR("writepage of page %p failed: %d\n",
788 /* either waiting for io to complete or reacquiring
789 * the lock that the failed writepage released */
793 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
794 /* check to see if another DLM lock covers this page */
795 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
796 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
798 &lock->l_resource->lr_name, LDLM_EXTENT,
799 &tmpex, LCK_PR | LCK_PW, &lockh);
800 if (rc2 == 0 && page->mapping != NULL) {
801 // checking again to account for writeback's lock_page()
802 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
803 ll_ra_accounting(page, inode->i_mapping);
804 ll_truncate_complete_page(page);
807 page_cache_release(page);
809 LASSERTF(tmpex.l_extent.start <=
810 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
811 lock->l_policy_data.l_extent.end + 1),
812 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
813 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
818 static int ll_extent_lock_callback(struct ldlm_lock *lock,
819 struct ldlm_lock_desc *new, void *data,
822 struct lustre_handle lockh = { 0 };
826 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
827 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
832 case LDLM_CB_BLOCKING:
833 ldlm_lock2handle(lock, &lockh);
834 rc = ldlm_cli_cancel(&lockh);
836 CERROR("ldlm_cli_cancel failed: %d\n", rc);
838 case LDLM_CB_CANCELING: {
840 struct ll_inode_info *lli;
841 struct lov_stripe_md *lsm;
845 /* This lock wasn't granted, don't try to evict pages */
846 if (lock->l_req_mode != lock->l_granted_mode)
849 inode = ll_inode_from_lock(lock);
852 lli = ll_i2info(inode);
855 if (lli->lli_smd == NULL)
859 stripe = ll_lock_to_stripe_offset(inode, lock);
862 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
864 down(&lli->lli_size_sem);
865 lock_res_and_lock(lock);
866 kms = ldlm_extent_shift_kms(lock,
867 lsm->lsm_oinfo[stripe].loi_kms);
869 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
870 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
871 lsm->lsm_oinfo[stripe].loi_kms, kms);
872 lsm->lsm_oinfo[stripe].loi_kms = kms;
873 unlock_res_and_lock(lock);
874 up(&lli->lli_size_sem);
875 //ll_try_done_writing(inode);
888 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
890 /* XXX ALLOCATE - 160 bytes */
891 struct inode *inode = ll_inode_from_lock(lock);
892 struct ll_inode_info *lli = ll_i2info(inode);
893 struct lustre_handle lockh = { 0 };
898 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
899 LDLM_FL_BLOCK_CONV)) {
900 LBUG(); /* not expecting any blocked async locks yet */
901 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
903 ldlm_lock_dump(D_OTHER, lock, 0);
904 ldlm_reprocess_all(lock->l_resource);
908 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
910 stripe = ll_lock_to_stripe_offset(inode, lock);
914 if (lock->l_lvb_len) {
915 struct lov_stripe_md *lsm = lli->lli_smd;
917 lvb = lock->l_lvb_data;
918 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
921 lock_res_and_lock(lock);
922 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
923 kms = ldlm_extent_shift_kms(NULL, kms);
924 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
925 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
926 lsm->lsm_oinfo[stripe].loi_kms, kms);
927 lsm->lsm_oinfo[stripe].loi_kms = kms;
928 unlock_res_and_lock(lock);
934 wake_up(&lock->l_waitq);
936 ldlm_lock2handle(lock, &lockh);
937 ldlm_lock_decref(&lockh, LCK_PR);
942 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
944 struct ptlrpc_request *req = reqp;
945 struct inode *inode = ll_inode_from_lock(lock);
946 struct ll_inode_info *lli;
948 struct lov_stripe_md *lsm;
949 int rc, size = sizeof(*lvb), stripe;
953 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
954 lli = ll_i2info(inode);
956 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
960 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
962 /* First, find out which stripe index this lock corresponds to. */
963 stripe = ll_lock_to_stripe_offset(inode, lock);
965 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
967 rc = lustre_pack_reply(req, 1, &size, NULL);
969 CERROR("lustre_pack_reply: %d\n", rc);
973 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
974 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
975 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
976 lvb->lvb_atime = LTIME_S(inode->i_atime);
977 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
979 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
980 inode->i_size, stripe, lvb->lvb_size);
986 /* These errors are normal races, so we don't want to fill the console
987 * with messages by calling ptlrpc_error() */
988 if (rc == -ELDLM_NO_LOCK_DATA)
989 lustre_pack_reply(req, 0, NULL, NULL);
995 /* NB: lov_merge_size will prefer locally cached writes if they extend the
996 * file (because it prefers KMS over RSS when larger) */
997 int ll_glimpse_size(struct inode *inode)
999 struct ll_inode_info *lli = ll_i2info(inode);
1000 struct ll_sb_info *sbi = ll_i2sbi(inode);
1001 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1002 struct lustre_handle lockh = { 0 };
1003 int rc, flags = LDLM_FL_HAS_INTENT;
1006 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1008 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
1009 LCK_PR, &flags, ll_extent_lock_callback,
1010 ldlm_completion_ast, ll_glimpse_callback, inode,
1011 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1016 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1017 RETURN(rc > 0 ? -EIO : rc);
1020 down(&lli->lli_size_sem);
1021 inode->i_size = lov_merge_size(lli->lli_smd, 0);
1022 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
1023 up(&lli->lli_size_sem);
1025 LTIME_S(inode->i_mtime) = lov_merge_mtime(lli->lli_smd,
1026 LTIME_S(inode->i_mtime));
1028 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
1029 (__u64)inode->i_size, (__u64)inode->i_blocks);
1031 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
1035 void ll_stime_record(struct ll_sb_info *sbi, struct timeval *start,
1036 struct obd_service_time *stime)
1038 struct timeval stop;
1039 do_gettimeofday(&stop);
1041 spin_lock(&sbi->ll_lock);
1042 lprocfs_stime_record(stime, &stop, start);
1043 spin_unlock(&sbi->ll_lock);
1046 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
1047 struct lov_stripe_md *lsm, int mode,
1048 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
1049 int ast_flags, struct obd_service_time *stime)
1051 struct ll_inode_info *lli = ll_i2info(inode);
1052 struct ll_sb_info *sbi = ll_i2sbi(inode);
1053 struct timeval start;
1057 LASSERT(lockh->cookie == 0);
1059 /* XXX phil: can we do this? won't it screw the file size up? */
1060 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1061 (sbi->ll_flags & LL_SBI_NOLCK))
1064 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
1065 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
1067 do_gettimeofday(&start);
1068 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
1069 &ast_flags, ll_extent_lock_callback,
1070 ldlm_completion_ast, ll_glimpse_callback, inode,
1071 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
1075 ll_stime_record(sbi, &start, stime);
1077 if (policy->l_extent.start == 0 &&
1078 policy->l_extent.end == OBD_OBJECT_EOF) {
1079 /* vmtruncate()->ll_truncate() first sets the i_size and then
1080 * the kms under both a DLM lock and the i_sem. If we don't
1081 * get the i_sem here we can match the DLM lock and reset
1082 * i_size from the kms before the truncating path has updated
1083 * the kms. generic_file_write can then trust the stale i_size
1084 * when doing appending writes and effectively cancel the
1085 * result of the truncate. Getting the i_sem after the enqueue
1086 * maintains the DLM -> i_sem acquiry order. */
1087 down(&lli->lli_size_sem);
1088 inode->i_size = lov_merge_size(lsm, 1);
1089 up(&lli->lli_size_sem);
1093 LTIME_S(inode->i_mtime) =
1094 lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1100 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1101 struct lov_stripe_md *lsm, int mode,
1102 struct lustre_handle *lockh)
1104 struct ll_sb_info *sbi = ll_i2sbi(inode);
1108 /* XXX phil: can we do this? won't it screw the file size up? */
1109 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1110 (sbi->ll_flags & LL_SBI_NOLCK))
1113 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1118 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1121 struct inode *inode = file->f_dentry->d_inode;
1122 struct ll_inode_info *lli = ll_i2info(inode);
1123 struct lov_stripe_md *lsm = lli->lli_smd;
1124 struct ll_lock_tree tree;
1125 struct ll_lock_tree_node *node;
1130 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1131 inode->i_ino, inode->i_generation, inode, count, *ppos);
1133 /* "If nbyte is 0, read() will return 0 and have no other results."
1134 * -- Single Unix Spec */
1138 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
1144 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1147 tree.lt_fd = file->private_data;
1149 rc = ll_tree_lock(&tree, node, inode, buf, count,
1150 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1154 down(&lli->lli_size_sem);
1155 kms = lov_merge_size(lsm, 1);
1156 if (*ppos + count - 1 > kms) {
1157 /* A glimpse is necessary to determine whether we return a short
1158 * read or some zeroes at the end of the buffer */
1159 up(&lli->lli_size_sem);
1160 retval = ll_glimpse_size(inode);
1164 inode->i_size = kms;
1165 up(&lli->lli_size_sem);
1168 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1169 inode->i_ino, count, *ppos, inode->i_size);
1171 /* turn off the kernel's read-ahead */
1172 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1175 file->f_ra.ra_pages = 0;
1177 retval = generic_file_read(file, buf, count, ppos);
1178 ll_audit_log(inode, AUDIT_READ, retval);
1181 ll_tree_unlock(&tree, inode);
1186 * Write to a file (through the page cache).
1188 static ssize_t ll_file_write(struct file *file, const char *buf,
1189 size_t count, loff_t *ppos)
1191 struct inode *inode = file->f_dentry->d_inode;
1192 loff_t maxbytes = ll_file_maxbytes(inode);
1193 struct ll_lock_tree tree;
1194 struct ll_lock_tree_node *node;
1199 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1200 inode->i_ino, inode->i_generation, inode, count, *ppos);
1202 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1204 /* POSIX, but surprised the VFS doesn't check this already */
1208 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1209 * called on the file, don't fail the below assertion (bug 2388). */
1210 if (file->f_flags & O_LOV_DELAY_CREATE &&
1211 ll_i2info(inode)->lli_smd == NULL)
1214 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1216 if (file->f_flags & O_APPEND)
1217 node = ll_node_from_inode(inode, 0, OBD_OBJECT_EOF, LCK_PW);
1219 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1223 RETURN(PTR_ERR(node));
1225 tree.lt_fd = file->private_data;
1227 rc = ll_tree_lock(&tree, node, inode, buf, count,
1228 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1232 /* this is ok, g_f_w will overwrite this under i_sem if it races
1233 * with a local truncate, it just makes our maxbyte checking easier */
1234 if (file->f_flags & O_APPEND)
1235 *ppos = inode->i_size;
1237 if (*ppos >= maxbytes) {
1238 if (count || *ppos > maxbytes) {
1239 send_sig(SIGXFSZ, current, 0);
1240 GOTO(out, retval = -EFBIG);
1243 if (*ppos + count > maxbytes)
1244 count = maxbytes - *ppos;
1246 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1247 inode->i_ino, count, *ppos);
1249 /* mark open handle dirty */
1250 set_bit(LLI_F_DIRTY_HANDLE, &(ll_i2info(inode)->lli_flags));
1252 /* generic_file_write handles O_APPEND after getting i_sem */
1253 retval = generic_file_write(file, buf, count, ppos);
1257 ll_audit_log(inode, AUDIT_WRITE, retval);
1259 ll_tree_unlock(&tree, inode);
1260 /* serialize with mmap/munmap/mremap */
1261 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
1262 retval > 0 ? retval : 0);
1266 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
1267 int flags, struct lov_user_md *lum,
1270 struct ll_inode_info *lli = ll_i2info(inode);
1272 struct obd_export *exp = ll_i2dtexp(inode);
1273 struct lov_stripe_md *lsm;
1274 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
1275 struct ptlrpc_request *req = NULL;
1277 struct lustre_md md;
1278 struct obd_client_handle *och;
1282 if ((file->f_flags+1) & O_ACCMODE)
1284 if (file->f_flags & O_TRUNC)
1287 down(&lli->lli_open_sem);
1290 up(&lli->lli_open_sem);
1291 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
1296 f = get_empty_filp();
1300 f->f_dentry = file->f_dentry;
1301 f->f_vfsmnt = file->f_vfsmnt;
1304 rc = ll_intent_alloc(&oit);
1308 rc = ll_intent_file_open(f, lum, lum_size, &oit);
1311 if (it_disposition(&oit, DISP_LOOKUP_NEG))
1314 req = LUSTRE_IT(&oit)->it_data;
1315 rc = LUSTRE_IT(&oit)->it_status;
1320 rc = mdc_req2lustre_md(ll_i2mdexp(inode), req, 1, exp, &md);
1323 ll_update_inode(f->f_dentry->d_inode, &md);
1325 OBD_ALLOC(och, sizeof(struct obd_client_handle));
1326 rc = ll_local_open(f, &oit, och);
1327 if (rc) { /* Actually ll_local_open cannot fail! */
1330 if (LUSTRE_IT(&oit)->it_lock_mode) {
1331 ldlm_lock_decref_and_cancel((struct lustre_handle *)
1332 &LUSTRE_IT(&oit)->it_lock_handle,
1333 LUSTRE_IT(&oit)->it_lock_mode);
1334 LUSTRE_IT(&oit)->it_lock_mode = 0;
1337 ll_intent_release(&oit);
1339 /* ll_file_release will decrease the count, but won't free anything
1340 because we have at least one more reference coming from actual open
1342 down(&lli->lli_och_sem);
1343 lli->lli_open_fd_write_count++;
1344 up(&lli->lli_och_sem);
1345 rc = ll_file_release(f->f_dentry->d_inode, f);
1347 /* Now also destroy our supplemental och */
1348 ll_md_och_close(ll_i2mdexp(inode), f->f_dentry->d_inode, och, 0);
1351 ll_intent_release(&oit);
1354 up(&lli->lli_open_sem);
1356 ptlrpc_req_finished(req);
1360 static int ll_lov_setea(struct inode *inode, struct file *file,
1363 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
1364 struct lov_user_md *lump;
1365 int lum_size = sizeof(struct lov_user_md) +
1366 sizeof(struct lov_user_ost_data);
1370 if (!capable (CAP_SYS_ADMIN))
1373 OBD_ALLOC(lump, lum_size);
1377 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
1379 OBD_FREE(lump, lum_size);
1383 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1385 OBD_FREE(lump, lum_size);
1389 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1392 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1394 int flags = FMODE_WRITE;
1397 /* Bug 1152: copy properly when this is no longer true */
1398 LASSERT(sizeof(lum) == sizeof(*lump));
1399 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1400 rc = copy_from_user(&lum, lump, sizeof(lum));
1404 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1406 put_user(0, &lump->lmm_stripe_count);
1407 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
1408 0, ll_i2info(inode)->lli_smd, lump);
1413 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1415 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1420 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
1424 static int ll_get_grouplock(struct inode *inode, struct file *file,
1427 struct ll_file_data *fd = file->private_data;
1428 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1429 .end = OBD_OBJECT_EOF}};
1430 struct lustre_handle lockh = { 0 };
1431 struct ll_inode_info *lli = ll_i2info(inode);
1432 struct lov_stripe_md *lsm = lli->lli_smd;
1436 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1440 policy.l_extent.gid = arg;
1441 if (file->f_flags & O_NONBLOCK)
1442 flags = LDLM_FL_BLOCK_NOWAIT;
1444 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags,
1445 &ll_i2sbi(inode)->ll_grouplock_stime);
1449 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1451 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1456 static int ll_put_grouplock(struct inode *inode, struct file *file,
1459 struct ll_file_data *fd = file->private_data;
1460 struct ll_inode_info *lli = ll_i2info(inode);
1461 struct lov_stripe_md *lsm = lli->lli_smd;
1465 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1466 /* Ugh, it's already unlocked. */
1470 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1473 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1475 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1480 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1485 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1488 struct ll_file_data *fd = file->private_data;
1489 struct ll_sb_info *sbi = ll_i2sbi(inode);
1493 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1494 inode->i_generation, inode, cmd);
1496 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1499 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1501 case LL_IOC_GETFLAGS:
1502 /* Get the current value of the file flags */
1503 return put_user(fd->fd_flags, (int *)arg);
1504 case LL_IOC_SETFLAGS:
1505 case LL_IOC_CLRFLAGS:
1506 /* Set or clear specific file flags */
1507 /* XXX This probably needs checks to ensure the flags are
1508 * not abused, and to handle any flag side effects.
1510 if (get_user(flags, (int *) arg))
1513 if (cmd == LL_IOC_SETFLAGS)
1514 fd->fd_flags |= flags;
1516 fd->fd_flags &= ~flags;
1518 case LL_IOC_LOV_SETSTRIPE:
1519 RETURN(ll_lov_setstripe(inode, file, arg));
1520 case LL_IOC_LOV_SETEA:
1521 RETURN(ll_lov_setea(inode, file, arg));
1522 case IOC_MDC_SHOWFID: {
1523 struct lustre_id *idp = (struct lustre_id *)arg;
1524 struct lustre_id id;
1528 filename = getname((const char *)arg);
1529 if (IS_ERR(filename))
1530 RETURN(PTR_ERR(filename));
1532 ll_inode2id(&id, inode);
1534 rc = ll_get_fid(sbi->ll_md_exp, &id, filename, &id);
1536 GOTO(out_filename, rc);
1538 rc = copy_to_user(idp, &id, sizeof(*idp));
1540 GOTO(out_filename, rc = -EFAULT);
1547 case LL_IOC_KEY_TYPE: {
1548 struct obd_ioctl_data *data;
1551 int typelen, rc, len = 0;
1553 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1558 type = data->ioc_inlbuf1;
1559 typelen = data->ioc_inllen1;
1562 CDEBUG(D_INFO, "LL_IOC_KEY_TYPE missing filename\n");
1563 GOTO(out, rc = -EINVAL);
1565 ll_set_sb_gksinfo(inode->i_sb, type);
1568 obd_ioctl_freedata(buf, len);
1572 case LL_IOC_LOV_GETSTRIPE:
1573 RETURN(ll_lov_getstripe(inode, arg));
1574 case EXT3_IOC_GETFLAGS:
1575 case EXT3_IOC_SETFLAGS:
1576 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1577 case LL_IOC_GROUP_LOCK:
1578 RETURN(ll_get_grouplock(inode, file, arg));
1579 case LL_IOC_GROUP_UNLOCK:
1580 RETURN(ll_put_grouplock(inode, file, arg));
1581 case EXT3_IOC_GETVERSION_OLD:
1582 case EXT3_IOC_GETVERSION:
1583 return put_user(inode->i_generation, (int *) arg);
1584 /* We need to special case any other ioctls we want to handle,
1585 * to send them to the MDS/OST as appropriate and to properly
1586 * network encode the arg field.
1587 case EXT2_IOC_GETVERSION_OLD:
1588 case EXT2_IOC_GETVERSION_NEW:
1589 case EXT2_IOC_SETVERSION_OLD:
1590 case EXT2_IOC_SETVERSION_NEW:
1591 case EXT3_IOC_SETVERSION_OLD:
1592 case EXT3_IOC_SETVERSION:
1594 case LL_IOC_FLUSH_CRED:
1595 RETURN(ll_flush_cred(inode));
1597 RETURN(ll_set_audit(inode, arg));
1599 RETURN( obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
1604 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1606 struct inode *inode = file->f_dentry->d_inode;
1607 struct ll_file_data *fd = file->private_data;
1608 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1609 struct lustre_handle lockh = {0};
1612 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1613 inode->i_generation, inode,
1614 offset + ((origin==2) ? inode->i_size : file->f_pos));
1616 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1617 if (origin == 2) { /* SEEK_END */
1618 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1619 struct ll_inode_info *lli = ll_i2info(inode);
1620 int nonblock = 0, rc;
1622 if (file->f_flags & O_NONBLOCK)
1623 nonblock = LDLM_FL_BLOCK_NOWAIT;
1625 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1626 nonblock, &ll_i2sbi(inode)->ll_seek_stime);
1630 down(&lli->lli_size_sem);
1631 offset += inode->i_size;
1632 up(&lli->lli_size_sem);
1633 } else if (origin == 1) { /* SEEK_CUR */
1634 offset += file->f_pos;
1638 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1639 if (offset != file->f_pos) {
1640 file->f_pos = offset;
1641 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1643 file->f_version = ++event;
1650 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1654 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1656 struct inode *inode = dentry->d_inode;
1657 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1658 struct lustre_id id;
1659 struct ptlrpc_request *req;
1662 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1663 inode->i_generation, inode);
1665 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1667 /* fsync's caller has already called _fdata{sync,write}, we want
1668 * that IO to finish before calling the osc and mdc sync methods */
1669 rc = filemap_fdatawait(inode->i_mapping);
1671 ll_inode2id(&id, inode);
1672 err = md_sync(ll_i2sbi(inode)->ll_md_exp, &id, &req);
1676 ptlrpc_req_finished(req);
1679 struct obdo *oa = obdo_alloc();
1682 RETURN(rc ? rc : -ENOMEM);
1684 oa->o_id = lsm->lsm_object_id;
1685 oa->o_gr = lsm->lsm_object_gr;
1686 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1688 obdo_from_inode(oa, inode, (OBD_MD_FLTYPE | OBD_MD_FLATIME |
1689 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1692 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
1702 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1704 struct inode *inode = file->f_dentry->d_inode;
1705 struct ll_inode_info *li = ll_i2info(inode);
1706 struct ll_sb_info *sbi = ll_i2sbi(inode);
1707 struct obd_device *obddev;
1708 struct ldlm_res_id res_id =
1709 { .name = {id_fid(&li->lli_id), id_group(&li->lli_id), LDLM_FLOCK} };
1710 struct lustre_handle lockh = {0};
1711 ldlm_policy_data_t flock;
1712 ldlm_mode_t mode = 0;
1717 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1718 inode->i_ino, file_lock);
1720 flock.l_flock.pid = file_lock->fl_pid;
1721 flock.l_flock.start = file_lock->fl_start;
1722 flock.l_flock.end = file_lock->fl_end;
1724 switch (file_lock->fl_type) {
1729 /* An unlock request may or may not have any relation to
1730 * existing locks so we may not be able to pass a lock handle
1731 * via a normal ldlm_lock_cancel() request. The request may even
1732 * unlock a byte range in the middle of an existing lock. In
1733 * order to process an unlock request we need all of the same
1734 * information that is given with a normal read or write record
1735 * lock request. To avoid creating another ldlm unlock (cancel)
1736 * message we'll treat a LCK_NL flock request as an unlock. */
1743 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1758 flags = LDLM_FL_BLOCK_NOWAIT;
1764 flags = LDLM_FL_TEST_LOCK;
1765 /* Save the old mode so that if the mode in the lock changes we
1766 * can decrement the appropriate reader or writer refcount. */
1767 file_lock->fl_type = mode;
1770 CERROR("unknown fcntl lock command: %d\n", cmd);
1774 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1775 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1776 flags, mode, flock.l_flock.start, flock.l_flock.end);
1778 obddev = md_get_real_obd(sbi->ll_md_exp, &li->lli_id);
1779 rc = ldlm_cli_enqueue(obddev->obd_self_export, NULL,
1780 obddev->obd_namespace,
1781 res_id, LDLM_FLOCK, &flock, mode, &flags,
1782 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1783 NULL, 0, NULL, &lockh);
1787 int ll_inode_revalidate_it(struct dentry *dentry)
1789 struct lookup_intent oit = { .it_op = IT_GETATTR };
1790 struct inode *inode = dentry->d_inode;
1791 struct ptlrpc_request *req = NULL;
1792 struct ll_inode_info *lli;
1793 struct lov_stripe_md *lsm;
1794 struct ll_sb_info *sbi;
1795 struct lustre_id id;
1800 CERROR("REPORT THIS LINE TO PETER\n");
1804 sbi = ll_i2sbi(inode);
1806 ll_inode2id(&id, inode);
1807 lli = ll_i2info(inode);
1808 LASSERT(id_fid(&id) != 0);
1810 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), name=%s(%p)\n",
1811 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1814 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1815 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1818 rc = ll_intent_alloc(&oit);
1822 rc = ll_crypto_init_it_key(inode, &oit);
1826 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1827 &oit, 0, &req, ll_mdc_blocking_ast);
1831 rc = revalidate_it_finish(req, 1, &oit, dentry);
1836 ll_lookup_finish_locks(&oit, dentry);
1838 if (!LLI_HAVE_FLSIZE(inode)) {
1839 /* if object not yet allocated, don't validate size */
1842 /* ll_glimpse_size() will prefer locally cached
1843 * writes if they extend the file */
1844 rc = ll_glimpse_size(inode);
1849 ll_intent_release(&oit);
1851 ptlrpc_req_finished(req);
1855 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1856 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
1859 struct inode *inode = de->d_inode;
1860 struct ll_inode_info *lli = ll_i2info(inode);
1862 res = ll_inode_revalidate_it(de);
1863 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1868 stat->ino = inode->i_ino;
1869 stat->mode = inode->i_mode;
1870 stat->nlink = inode->i_nlink;
1871 stat->uid = inode->i_uid;
1872 stat->gid = inode->i_gid;
1873 stat->atime = inode->i_atime;
1874 stat->mtime = inode->i_mtime;
1875 stat->ctime = inode->i_ctime;
1876 stat->blksize = inode->i_blksize;
1878 down(&lli->lli_size_sem);
1879 stat->size = inode->i_size;
1880 stat->blocks = inode->i_blocks;
1881 up(&lli->lli_size_sem);
1883 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1884 stat->dev = id_group(&ll_i2info(inode)->lli_id);
1890 int ll_setxattr_internal(struct inode *inode, const char *name,
1891 const void *value, size_t size, int flags,
1894 struct ll_sb_info *sbi = ll_i2sbi(inode);
1895 struct ptlrpc_request *request = NULL;
1896 struct mdc_op_data op_data;
1899 int rc = 0, key_size = 0;
1902 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu\n", inode->i_ino);
1903 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_SETXATTR);
1905 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
1906 RETURN(-EOPNOTSUPP);
1908 memset(&attr, 0x0, sizeof(attr));
1909 attr.ia_valid |= valid;
1910 attr.ia_attr_flags = flags;
1912 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1913 rc = ll_crypto_get_mac(inode, NULL, (void *)value, size,
1917 ll_prepare_mdc_data(&op_data, inode, NULL, NULL, 0, 0);
1919 rc = md_setattr(sbi->ll_md_exp, &op_data, &attr,
1920 (void*) name, strnlen(name, XATTR_NAME_MAX) + 1,
1921 (void*) value, size, key, key_size, &request);
1923 if (key && key_size)
1924 OBD_FREE(key, key_size);
1926 CERROR("md_setattr fails: rc = %d\n", rc);
1930 ptlrpc_req_finished(request);
1934 int ll_setxattr(struct dentry *dentry, const char *name, const void *value,
1935 size_t size, int flags)
1938 struct posix_acl *acl;
1939 struct ll_inode_info *lli;
1942 rc = ll_setxattr_internal(dentry->d_inode, name, value, size,
1945 /* update inode's acl info */
1946 if (rc == 0 && strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1948 acl = posix_acl_from_xattr(value, size);
1950 CERROR("convert from xattr to acl error: %ld",
1954 error = posix_acl_valid(acl);
1956 CERROR("acl valid error: %d", error);
1957 posix_acl_release(acl);
1965 lli = ll_i2info(dentry->d_inode);
1966 spin_lock(&lli->lli_lock);
1967 if (lli->lli_posix_acl != NULL)
1968 posix_acl_release(lli->lli_posix_acl);
1969 lli->lli_posix_acl = acl;
1970 spin_unlock(&lli->lli_lock);
1977 int ll_removexattr(struct dentry *dentry, const char *name)
1979 return ll_setxattr_internal(dentry->d_inode, name, NULL, 0, 0,
1983 int ll_getxattr_internal(struct inode *inode, const char *name,
1984 void *value, size_t size, __u64 valid)
1986 struct ptlrpc_request *request = NULL;
1987 struct ll_sb_info *sbi = ll_i2sbi(inode);
1988 struct lustre_id id;
1989 struct mds_body *body;
1994 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETXATTR);
1996 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
1997 RETURN(-EOPNOTSUPP);
1999 ll_inode2id(&id, inode);
2000 rc = md_getattr(sbi->ll_md_exp, &id, valid, name, NULL, 0,
2001 size, NULL, &request);
2003 if (rc != -ENODATA && rc != -EOPNOTSUPP)
2004 CERROR("rc = %d\n", rc);
2008 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
2009 LASSERT(body != NULL);
2010 LASSERT_REPSWABBED(request, 0);
2012 ea_size = body->eadatasize;
2014 GOTO(out, rc = ea_size);
2016 LASSERT(ea_size <= request->rq_repmsg->buflens[1]);
2017 ea_data = lustre_msg_buf(request->rq_repmsg, 1, ea_size);
2018 LASSERT(ea_data != NULL);
2019 LASSERT_REPSWABBED(request, 1);
2022 memcpy(value, ea_data, ea_size);
2026 ptlrpc_req_finished(request);
2030 int ll_getxattr(struct dentry *dentry, const char *name, void *value,
2033 return ll_getxattr_internal(dentry->d_inode, name,
2034 value, size, OBD_MD_FLXATTR);
2037 int ll_listxattr(struct dentry *dentry, char *list, size_t size)
2039 return ll_getxattr_internal(dentry->d_inode, NULL, list, size,
2040 OBD_MD_FLXATTRLIST);
2044 * Here we hold DLM lock across permission check, to get better
2045 * conformance especially for remote acl. will introduce certain
2046 * overhead, maybe need a profile.
2049 lustre_check_acl(struct inode *inode, int mask)
2051 struct lookup_intent it = { .it_op = IT_GETATTR };
2052 struct dentry de = { .d_inode = inode };
2053 struct ll_sb_info *sbi;
2054 struct lustre_id id;
2055 struct ptlrpc_request *req = NULL;
2056 struct ll_inode_info *lli = ll_i2info(inode);
2057 struct posix_acl *acl;
2061 sbi = ll_i2sbi(inode);
2062 ll_inode2id(&id, inode);
2064 if (ll_intent_alloc(&it))
2067 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
2068 &it, 0, &req, ll_mdc_blocking_ast);
2070 ll_intent_free(&it);
2074 rc = revalidate_it_finish(req, 1, &it, &de);
2076 ll_intent_release(&it);
2080 if (sbi->ll_remote) {
2081 rc = ll_remote_acl_permission(inode, mask);
2083 spin_lock(&lli->lli_lock);
2084 acl = posix_acl_dup(ll_i2info(inode)->lli_posix_acl);
2085 spin_unlock(&lli->lli_lock);
2090 rc = posix_acl_permission(inode, acl, mask);
2091 posix_acl_release(acl);
2095 ll_lookup_finish_locks(&it, &de);
2096 ll_intent_free(&it);
2100 ptlrpc_req_finished(req);
2105 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
2107 struct ll_sb_info *sbi = ll_i2sbi(inode);
2109 /* for remote client, permission bits in inode doesn't
2110 * play a role anymore.
2113 return lustre_check_acl(inode, mask);
2115 return generic_permission(inode, mask, lustre_check_acl);
2118 struct file_operations ll_file_operations = {
2119 .read = ll_file_read,
2120 .write = ll_file_write,
2121 .ioctl = ll_file_ioctl,
2122 .open = ll_file_open,
2123 .release = ll_file_release,
2124 .mmap = ll_file_mmap,
2125 .llseek = ll_file_seek,
2126 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
2127 .sendfile = generic_file_sendfile,
2130 .lock = ll_file_flock
2133 struct inode_operations ll_file_inode_operations = {
2134 .setattr = ll_setattr,
2135 .truncate = ll_truncate,
2136 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
2137 .getattr = ll_getattr,
2139 .revalidate_it = ll_inode_revalidate_it,
2141 .setxattr = ll_setxattr,
2142 .getxattr = ll_getxattr,
2143 .listxattr = ll_listxattr,
2144 .removexattr = ll_removexattr,
2145 .permission = ll_inode_permission,