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 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
32 #include <linux/lustre_compat25.h>
34 #include "llite_internal.h"
35 #include <linux/obd_lov.h>
37 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
38 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
39 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
41 int ll_validate_size(struct inode *inode, __u64 *size, __u64 *blocks)
43 ldlm_policy_data_t extent = { .l_extent = { 0, OBD_OBJECT_EOF } };
44 struct obd_export *exp = ll_i2sbi(inode)->ll_dt_exp;
45 struct ll_inode_info *lli = ll_i2info(inode);
46 struct lustre_handle match_lockh = {0};
50 if (lli->lli_smd == NULL)
53 LASSERT(size != NULL && blocks != NULL);
55 flags = LDLM_FL_TEST_LOCK | LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
56 rc = obd_match(exp, lli->lli_smd, LDLM_EXTENT, &extent,
57 LCK_PR | LCK_PW, &flags, inode, &match_lockh);
59 /* we have no all needed locks,
60 * so we don't know actual size */
64 /* we know actual size! */
65 down(&lli->lli_size_sem);
66 *size = lov_merge_size(lli->lli_smd, 0);
67 *blocks = lov_merge_blocks(lli->lli_smd);
68 up(&lli->lli_size_sem);
74 int ll_md_och_close(struct obd_export *md_exp, struct inode *inode,
75 struct obd_client_handle *och)
77 struct ptlrpc_request *req = NULL;
78 struct obdo *obdo = NULL;
79 struct obd_device *obd;
83 obd = class_exp2obd(md_exp);
85 CERROR("Invalid MDC connection handle "LPX64"\n",
86 md_exp->exp_handle.h_cookie);
92 * here we check if this is forced umount. If so this is called on
93 * canceling "open lock" and we do not call md_close() in this case , as
94 * it will not successful, as import is already deactivated.
96 if (obd->obd_no_recov)
99 /* closing opened file */
104 obdo->o_id = inode->i_ino;
105 obdo->o_valid = OBD_MD_FLID;
106 obdo_from_inode(obdo, inode, (OBD_MD_FLTYPE | OBD_MD_FLMODE |
107 OBD_MD_FLATIME | OBD_MD_FLMTIME |
109 if (0 /* ll_is_inode_dirty(inode) */) {
110 obdo->o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
111 obdo->o_valid |= OBD_MD_FLFLAGS;
113 obdo->o_fid = id_fid(&ll_i2info(inode)->lli_id);
114 obdo->o_mds = id_group(&ll_i2info(inode)->lli_id);
117 obdo->o_valid |= OBD_MD_FLEPOCH;
118 obdo->o_easize = ll_i2info(inode)->lli_io_epoch;
120 if (ll_validate_size(inode, &obdo->o_size, &obdo->o_blocks))
121 obdo->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
123 rc = md_close(md_exp, obdo, och, &req);
128 * we are the last writer, so the MDS has instructed us to get
129 * the file size and any write cookies, then close again.
132 //ll_queue_done_writing(inode);
135 CERROR("inode %lu mdc close failed: rc = %d\n",
136 (unsigned long)inode->i_ino, rc);
139 ptlrpc_req_finished(req);
142 mdc_clear_open_replay_data(md_exp, och);
143 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
144 OBD_FREE(och, sizeof *och);
148 int ll_md_real_close(struct obd_export *md_exp,
149 struct inode *inode, int flags)
151 struct ll_inode_info *lli = ll_i2info(inode);
152 struct obd_client_handle **och_p;
153 struct obd_client_handle *och;
158 if (flags & FMODE_WRITE) {
159 och_p = &lli->lli_mds_write_och;
160 och_usecount = &lli->lli_open_fd_write_count;
161 } else if (flags & FMODE_EXEC) {
162 och_p = &lli->lli_mds_exec_och;
163 och_usecount = &lli->lli_open_fd_exec_count;
165 och_p = &lli->lli_mds_read_och;
166 och_usecount = &lli->lli_open_fd_read_count;
169 down(&lli->lli_och_sem);
170 if (*och_usecount) { /* There are still users of this handle, so
172 up(&lli->lli_och_sem);
178 up(&lli->lli_och_sem);
181 * there might be a race and somebody have freed this och
182 * already. Another way to have this twice called is if file closing
183 * will fail due to netwok problems and on umount lock will be canceled
184 * and this will be called from block_ast callack.
186 if (och && och->och_fh.cookie != DEAD_HANDLE_MAGIC)
187 rc = ll_md_och_close(md_exp, inode, och);
192 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
195 struct ll_file_data *fd = file->private_data;
196 struct ll_inode_info *lli = ll_i2info(inode);
200 /* clear group lock, if present */
201 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
202 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
203 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
204 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
208 /* Let's see if we have good enough OPEN lock on the file and if
209 we can skip talking to MDS */
210 if (file->f_dentry->d_inode) {
212 struct obd_device *obddev;
213 struct lustre_handle lockh;
214 int flags = LDLM_FL_BLOCK_GRANTED;
215 struct ldlm_res_id file_res_id = {.name = {id_fid(&lli->lli_id),
216 id_group(&lli->lli_id)}};
217 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
219 down(&lli->lli_och_sem);
220 if (fd->fd_omode & FMODE_WRITE) {
222 LASSERT(lli->lli_open_fd_write_count);
223 lli->lli_open_fd_write_count--;
224 } else if (fd->fd_omode & FMODE_EXEC) {
226 LASSERT(lli->lli_open_fd_exec_count);
227 lli->lli_open_fd_exec_count--;
230 LASSERT(lli->lli_open_fd_read_count);
231 lli->lli_open_fd_read_count--;
233 up(&lli->lli_och_sem);
235 obddev = md_get_real_obd(md_exp, &lli->lli_id);
236 if (!ldlm_lock_match(obddev->obd_namespace, flags, &file_res_id,
237 LDLM_IBITS, &policy, lockmode, &lockh))
239 rc = ll_md_real_close(md_exp, file->f_dentry->d_inode,
242 ldlm_lock_decref(&lockh, lockmode);
246 file->private_data = NULL;
247 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof(*fd));
251 /* While this returns an error code, fput() the caller does not, so we need
252 * to make every effort to clean up all of our state here. Also, applications
253 * rarely check close errors and even if an error is returned they will not
254 * re-try the close call.
256 int ll_file_release(struct inode *inode, struct file *file)
258 struct ll_file_data *fd;
259 struct ll_sb_info *sbi = ll_i2sbi(inode);
263 CDEBUG(D_VFSTRACE, "VFS Op:inode="DLID4"(%p)\n",
264 OLID4(&ll_i2info(inode)->lli_id), inode);
266 /* don't do anything for / */
267 if (inode->i_sb->s_root == file->f_dentry)
270 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
271 fd = (struct ll_file_data *)file->private_data;
274 rc = ll_md_close(sbi->ll_md_exp, inode, file);
278 static int ll_intent_file_open(struct file *file, void *lmm,
279 int lmmsize, struct lookup_intent *itp)
281 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
282 const char *name = (char *)file->f_dentry->d_name.name;
283 struct dentry *parent = file->f_dentry->d_parent;
284 const int len = file->f_dentry->d_name.len;
285 struct lustre_handle lockh;
286 struct mdc_op_data *op_data;
292 OBD_ALLOC(op_data, sizeof(*op_data));
296 ll_prepare_mdc_data(op_data, parent->d_inode, NULL,
299 rc = md_enqueue(sbi->ll_md_exp, LDLM_IBITS, itp, LCK_PR, op_data,
300 &lockh, lmm, lmmsize, ldlm_completion_ast,
301 ll_mdc_blocking_ast, NULL);
302 OBD_FREE(op_data, sizeof(*op_data));
304 if (LUSTRE_IT(itp)->it_lock_mode)
305 memcpy(&LUSTRE_IT(itp)->it_lock_handle,
306 &lockh, sizeof(lockh));
309 CERROR("lock enqueue: err: %d\n", rc);
314 void ll_och_fill(struct inode *inode, struct lookup_intent *it,
315 struct obd_client_handle *och)
317 struct ptlrpc_request *req = LUSTRE_IT(it)->it_data;
318 struct ll_inode_info *lli = ll_i2info(inode);
319 struct mds_body *body;
322 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
323 LASSERT (body != NULL); /* reply already checked out */
324 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
326 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
327 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
328 lli->lli_io_epoch = body->io_epoch;
329 mdc_set_open_replay_data(ll_i2mdexp(inode), och,
330 LUSTRE_IT(it)->it_data);
333 int ll_local_open(struct file *file, struct lookup_intent *it,
334 struct obd_client_handle *och)
336 struct ll_file_data *fd;
340 ll_och_fill(file->f_dentry->d_inode, it, och);
342 LASSERTF(file->private_data == NULL, "file %.*s/%.*s ino %lu/%u (%o)\n",
343 file->f_dentry->d_name.len, file->f_dentry->d_name.name,
344 file->f_dentry->d_parent->d_name.len,
345 file->f_dentry->d_parent->d_name.name,
346 file->f_dentry->d_inode->i_ino,
347 file->f_dentry->d_inode->i_generation,
348 file->f_dentry->d_inode->i_mode);
350 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
352 /* We can't handle this well without reorganizing ll_file_open and
353 * ll_md_close(), so don't even try right now. */
356 file->private_data = fd;
357 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
358 fd->fd_omode = it->it_flags;
362 /* Open a file, and (for the very first open) create objects on the OSTs at
363 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
364 * creation or open until ll_lov_setstripe() ioctl is called. We grab
365 * lli_open_sem to ensure no other process will create objects, send the
366 * stripe MD to the MDS, or try to destroy the objects if that fails.
368 * If we already have the stripe MD locally then we don't request it in
369 * mdc_open(), by passing a lmm_size = 0.
371 * It is up to the application to ensure no other processes open this file
372 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
373 * used. We might be able to avoid races of that sort by getting lli_open_sem
374 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
375 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
377 int ll_file_open(struct inode *inode, struct file *file)
379 struct ll_inode_info *lli = ll_i2info(inode);
380 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
381 .it_flags = file->f_flags };
382 struct lov_stripe_md *lsm;
383 struct ptlrpc_request *req;
385 struct obd_client_handle **och_p = NULL;
386 __u64 *och_usecount = NULL;
389 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n",
390 inode->i_ino, inode->i_generation, inode, file->f_flags);
392 /* don't do anything for / */
393 if (inode->i_sb->s_root == file->f_dentry)
396 if ((file->f_flags+1) & O_ACCMODE)
398 if (file->f_flags & O_TRUNC)
404 * sometimes LUSTRE_IT(it) may not be allocated like opening file by
405 * dentry_open() from GNS stuff.
407 if (!it || !LUSTRE_IT(it)) {
409 rc = ll_intent_alloc(it);
414 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
417 * mdc_intent_lock() didn't get a request ref if there was an open
418 * error, so don't do cleanup on the * request here (bug 3430)
420 if (LUSTRE_IT(it)->it_disposition) {
421 rc = it_open_error(DISP_OPEN_OPEN, it);
426 /* Let's see if we have file open on MDS already. */
427 if (it->it_flags & FMODE_WRITE) {
428 och_p = &lli->lli_mds_write_och;
429 och_usecount = &lli->lli_open_fd_write_count;
430 } else if (it->it_flags & FMODE_EXEC) {
431 och_p = &lli->lli_mds_exec_och;
432 och_usecount = &lli->lli_open_fd_exec_count;
434 och_p = &lli->lli_mds_read_och;
435 och_usecount = &lli->lli_open_fd_read_count;
438 down(&lli->lli_och_sem);
439 if (*och_p) { /* Open handle is present */
440 if (it_disposition(it, DISP_LOOKUP_POS) && /* Positive lookup */
441 it_disposition(it, DISP_OPEN_OPEN)) { /* & OPEN happened */
442 struct obd_client_handle *och;
443 /* Well, there's extra open request that we do not need,
444 let's close it somehow*/
445 OBD_ALLOC(och, sizeof (struct obd_client_handle));
447 up(&lli->lli_och_sem);
451 ll_och_fill(inode, it, och);
452 /* ll_md_och_close() will free och */
453 ll_md_och_close(ll_i2mdexp(inode), inode, och);
457 rc = ll_local_open(file, it, NULL);
461 LASSERT(*och_usecount == 0);
462 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
464 GOTO(out, rc = -ENOMEM);
467 if (!it || !LUSTRE_IT(it) || !LUSTRE_IT(it)->it_disposition) {
469 * we are going to replace intent here, and that may
470 * possibly change access mode (FMODE_EXEC can only be
471 * set in intent), but I hope it never happens (I was
472 * not able to trigger it yet at least) -- green
475 /* FIXME: FMODE_EXEC is not covered by O_ACCMODE! */
476 LASSERT(!(it->it_flags & FMODE_EXEC));
477 LASSERTF((it->it_flags & O_ACCMODE) ==
478 (oit.it_flags & O_ACCMODE), "Changing intent "
479 "flags %x to incompatible %x\n", it->it_flags,
482 rc = ll_intent_file_open(file, NULL, 0, it);
485 rc = it_open_error(DISP_OPEN_OPEN, it);
487 GOTO(out_och_free, rc);
489 mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle,
490 file->f_dentry->d_inode);
492 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
493 rc = ll_local_open(file, it, *och_p);
494 LASSERTF(rc == 0, "rc = %d\n", rc);
496 up(&lli->lli_och_sem);
499 * must do this outside lli_och_sem lock to prevent deadlock where
500 * different kind of OPEN lock for this same inode gets cancelled by
504 if (!S_ISREG(inode->i_mode))
509 if (file->f_flags & O_LOV_DELAY_CREATE ||
510 !(file->f_mode & FMODE_WRITE)) {
511 CDEBUG(D_INODE, "object creation was delayed\n");
515 file->f_flags &= ~O_LOV_DELAY_CREATE;
518 req = LUSTRE_IT(it)->it_data;
519 ll_intent_drop_lock(it);
520 ll_intent_release(it);
521 ptlrpc_req_finished(req);
523 ll_open_complete(inode);
527 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
528 *och_p = NULL; /* OBD_FREE writes some magic there */
531 up(&lli->lli_och_sem);
537 /* Fills the obdo with the attributes for the inode defined by lsm */
538 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
541 struct ptlrpc_request_set *set;
545 LASSERT(lsm != NULL);
547 memset(oa, 0, sizeof *oa);
548 oa->o_id = lsm->lsm_object_id;
549 oa->o_gr = lsm->lsm_object_gr;
550 oa->o_mode = S_IFREG;
551 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
552 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
553 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
555 set = ptlrpc_prep_set();
559 rc = obd_getattr_async(exp, oa, lsm, set);
561 rc = ptlrpc_set_wait(set);
562 ptlrpc_set_destroy(set);
567 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
568 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
572 static inline void ll_remove_suid(struct inode *inode)
576 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
577 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
579 /* was any of the uid bits set? */
580 mode &= inode->i_mode;
581 if (mode && !capable(CAP_FSETID)) {
582 inode->i_mode &= ~mode;
583 // XXX careful here - we cannot change the size
587 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
589 struct ll_inode_info *lli = ll_i2info(inode);
590 struct lov_stripe_md *lsm = lli->lli_smd;
591 struct obd_export *exp = ll_i2dtexp(inode);
594 struct ldlm_lock *lock;
595 struct lov_stripe_md *lsm;
596 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
597 __u32 stripe, vallen = sizeof(stripe);
601 if (lsm->lsm_stripe_count == 1)
602 GOTO(check, stripe = 0);
604 /* get our offset in the lov */
605 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
607 CERROR("obd_get_info: rc = %d\n", rc);
610 LASSERT(stripe < lsm->lsm_stripe_count);
613 if (lsm->lsm_oinfo[stripe].loi_id != lock->l_resource->lr_name.name[0]||
614 lsm->lsm_oinfo[stripe].loi_gr != lock->l_resource->lr_name.name[2]){
615 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64
616 " inode=%lu/%u (%p)\n",
617 lsm->lsm_oinfo[stripe].loi_id,
618 lsm->lsm_oinfo[stripe].loi_gr,
619 inode->i_ino, inode->i_generation, inode);
620 return -ELDLM_NO_LOCK_DATA;
626 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
627 * we get a lock cancellation for each stripe, so we have to map the obd's
628 * region back onto the stripes in the file that it held.
630 * No one can dirty the extent until we've finished our work and they can
631 * enqueue another lock. The DLM protects us from ll_file_read/write here,
632 * but other kernel actors could have pages locked.
634 * Called with the DLM lock held. */
635 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
636 struct ldlm_lock *lock, __u32 stripe)
638 ldlm_policy_data_t tmpex;
639 unsigned long start, end, count, skip, i, j;
641 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
642 struct lustre_handle lockh;
645 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
646 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
647 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
650 /* our locks are page granular thanks to osc_enqueue, we invalidate the
652 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
653 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
657 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
658 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
659 if (lsm->lsm_stripe_count > 1) {
660 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
661 skip = (lsm->lsm_stripe_count - 1) * count;
662 start += start/count * skip + stripe * count;
664 end += end/count * skip + stripe * count;
666 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
669 i = inode->i_size ? (inode->i_size - 1) >> PAGE_CACHE_SHIFT : 0;
673 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
674 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
675 count, skip, end, discard ? " (DISCARDING)" : "");
677 /* walk through the vmas on the inode and tear down mmaped pages that
678 * intersect with the lock. this stops immediately if there are no
679 * mmap()ed regions of the file. This is not efficient at all and
680 * should be short lived. We'll associate mmap()ed pages with the lock
681 * and will be able to find them directly */
683 for (i = start; i <= end; i += (j + skip)) {
684 j = min(count - (i % count), end - i + 1);
686 LASSERT(inode->i_mapping);
687 if (ll_teardown_mmaps(inode->i_mapping,
688 (__u64)i << PAGE_CACHE_SHIFT,
689 ((__u64)(i+j) << PAGE_CACHE_SHIFT) - 1) )
693 /* this is the simplistic implementation of page eviction at
694 * cancelation. It is careful to get races with other page
695 * lockers handled correctly. fixes from bug 20 will make it
696 * more efficient by associating locks with pages and with
697 * batching writeback under the lock explicitly. */
698 for (i = start, j = start % count; i <= end;
699 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
701 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
707 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
708 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
709 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
712 if (!mapping_has_pages(inode->i_mapping)) {
713 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
719 page = find_get_page(inode->i_mapping, i);
722 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
723 i, tmpex.l_extent.start);
726 /* page->mapping to check with racing against teardown */
727 if (!discard && clear_page_dirty_for_io(page)) {
728 rc = ll_call_writepage(inode, page);
730 CERROR("writepage of page %p failed: %d\n",
732 /* either waiting for io to complete or reacquiring
733 * the lock that the failed writepage released */
737 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
738 /* check to see if another DLM lock covers this page */
739 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
740 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
742 &lock->l_resource->lr_name, LDLM_EXTENT,
743 &tmpex, LCK_PR | LCK_PW, &lockh);
744 if (rc2 == 0 && page->mapping != NULL) {
745 // checking again to account for writeback's lock_page()
746 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
747 ll_ra_accounting(page, inode->i_mapping);
748 ll_truncate_complete_page(page);
751 page_cache_release(page);
753 LASSERTF(tmpex.l_extent.start <=
754 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
755 lock->l_policy_data.l_extent.end + 1),
756 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
757 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
762 static int ll_extent_lock_callback(struct ldlm_lock *lock,
763 struct ldlm_lock_desc *new, void *data,
766 struct lustre_handle lockh = { 0 };
770 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
771 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
776 case LDLM_CB_BLOCKING:
777 ldlm_lock2handle(lock, &lockh);
778 rc = ldlm_cli_cancel(&lockh);
780 CERROR("ldlm_cli_cancel failed: %d\n", rc);
782 case LDLM_CB_CANCELING: {
784 struct ll_inode_info *lli;
785 struct lov_stripe_md *lsm;
789 /* This lock wasn't granted, don't try to evict pages */
790 if (lock->l_req_mode != lock->l_granted_mode)
793 inode = ll_inode_from_lock(lock);
796 lli = ll_i2info(inode);
799 if (lli->lli_smd == NULL)
803 stripe = ll_lock_to_stripe_offset(inode, lock);
806 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
808 down(&lli->lli_size_sem);
809 lock_res(lock->l_resource);
810 kms = ldlm_extent_shift_kms(lock,
811 lsm->lsm_oinfo[stripe].loi_kms);
813 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
814 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
815 lsm->lsm_oinfo[stripe].loi_kms, kms);
816 lsm->lsm_oinfo[stripe].loi_kms = kms;
817 unlock_res(lock->l_resource);
818 up(&lli->lli_size_sem);
819 //ll_try_done_writing(inode);
832 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
834 /* XXX ALLOCATE - 160 bytes */
835 struct inode *inode = ll_inode_from_lock(lock);
836 struct ll_inode_info *lli = ll_i2info(inode);
837 struct lustre_handle lockh = { 0 };
842 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
843 LDLM_FL_BLOCK_CONV)) {
844 LBUG(); /* not expecting any blocked async locks yet */
845 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
847 ldlm_lock_dump(D_OTHER, lock, 0);
848 ldlm_reprocess_all(lock->l_resource);
852 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
854 stripe = ll_lock_to_stripe_offset(inode, lock);
858 if (lock->l_lvb_len) {
859 struct lov_stripe_md *lsm = lli->lli_smd;
861 lvb = lock->l_lvb_data;
862 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
865 lock_res(lock->l_resource);
866 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
867 kms = ldlm_extent_shift_kms(NULL, kms);
868 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
869 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
870 lsm->lsm_oinfo[stripe].loi_kms, kms);
871 lsm->lsm_oinfo[stripe].loi_kms = kms;
872 unlock_res(lock->l_resource);
878 wake_up(&lock->l_waitq);
880 ldlm_lock2handle(lock, &lockh);
881 ldlm_lock_decref(&lockh, LCK_PR);
886 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
888 struct ptlrpc_request *req = reqp;
889 struct inode *inode = ll_inode_from_lock(lock);
890 struct ll_inode_info *lli;
892 struct lov_stripe_md *lsm;
893 int rc, size = sizeof(*lvb), stripe;
897 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
898 lli = ll_i2info(inode);
900 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
904 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
906 /* First, find out which stripe index this lock corresponds to. */
907 stripe = ll_lock_to_stripe_offset(inode, lock);
909 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
911 rc = lustre_pack_reply(req, 1, &size, NULL);
913 CERROR("lustre_pack_reply: %d\n", rc);
917 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
918 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
919 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
920 lvb->lvb_atime = LTIME_S(inode->i_atime);
921 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
923 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
924 inode->i_size, stripe, lvb->lvb_size);
930 /* These errors are normal races, so we don't want to fill the console
931 * with messages by calling ptlrpc_error() */
932 if (rc == -ELDLM_NO_LOCK_DATA)
933 lustre_pack_reply(req, 0, NULL, NULL);
939 /* NB: lov_merge_size will prefer locally cached writes if they extend the
940 * file (because it prefers KMS over RSS when larger) */
941 int ll_glimpse_size(struct inode *inode)
943 struct ll_inode_info *lli = ll_i2info(inode);
944 struct ll_sb_info *sbi = ll_i2sbi(inode);
945 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
946 struct lustre_handle lockh = { 0 };
947 int rc, flags = LDLM_FL_HAS_INTENT;
950 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
952 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
953 LCK_PR, &flags, ll_extent_lock_callback,
954 ldlm_completion_ast, ll_glimpse_callback, inode,
955 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
960 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
961 RETURN(rc > 0 ? -EIO : rc);
964 down(&lli->lli_size_sem);
965 inode->i_size = lov_merge_size(lli->lli_smd, 0);
966 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
967 up(&lli->lli_size_sem);
969 LTIME_S(inode->i_mtime) = lov_merge_mtime(lli->lli_smd,
970 LTIME_S(inode->i_mtime));
972 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
973 (__u64)inode->i_size, (__u64)inode->i_blocks);
975 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
979 void ll_stime_record(struct ll_sb_info *sbi, struct timeval *start,
980 struct obd_service_time *stime)
983 do_gettimeofday(&stop);
985 spin_lock(&sbi->ll_lock);
986 lprocfs_stime_record(stime, &stop, start);
987 spin_unlock(&sbi->ll_lock);
990 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
991 struct lov_stripe_md *lsm, int mode,
992 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
993 int ast_flags, struct obd_service_time *stime)
995 struct ll_inode_info *lli = ll_i2info(inode);
996 struct ll_sb_info *sbi = ll_i2sbi(inode);
997 struct timeval start;
1001 LASSERT(lockh->cookie == 0);
1003 /* XXX phil: can we do this? won't it screw the file size up? */
1004 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1005 (sbi->ll_flags & LL_SBI_NOLCK))
1008 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
1009 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
1011 do_gettimeofday(&start);
1012 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
1013 &ast_flags, ll_extent_lock_callback,
1014 ldlm_completion_ast, ll_glimpse_callback, inode,
1015 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
1019 ll_stime_record(sbi, &start, stime);
1021 if (policy->l_extent.start == 0 &&
1022 policy->l_extent.end == OBD_OBJECT_EOF) {
1023 /* vmtruncate()->ll_truncate() first sets the i_size and then
1024 * the kms under both a DLM lock and the i_sem. If we don't
1025 * get the i_sem here we can match the DLM lock and reset
1026 * i_size from the kms before the truncating path has updated
1027 * the kms. generic_file_write can then trust the stale i_size
1028 * when doing appending writes and effectively cancel the
1029 * result of the truncate. Getting the i_sem after the enqueue
1030 * maintains the DLM -> i_sem acquiry order. */
1031 down(&lli->lli_size_sem);
1032 inode->i_size = lov_merge_size(lsm, 1);
1033 up(&lli->lli_size_sem);
1037 LTIME_S(inode->i_mtime) =
1038 lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1044 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1045 struct lov_stripe_md *lsm, int mode,
1046 struct lustre_handle *lockh)
1048 struct ll_sb_info *sbi = ll_i2sbi(inode);
1052 /* XXX phil: can we do this? won't it screw the file size up? */
1053 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1054 (sbi->ll_flags & LL_SBI_NOLCK))
1057 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1062 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1065 struct inode *inode = file->f_dentry->d_inode;
1066 struct ll_inode_info *lli = ll_i2info(inode);
1067 struct lov_stripe_md *lsm = lli->lli_smd;
1068 struct ll_lock_tree tree;
1069 struct ll_lock_tree_node *node;
1074 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1075 inode->i_ino, inode->i_generation, inode, count, *ppos);
1077 /* "If nbyte is 0, read() will return 0 and have no other results."
1078 * -- Single Unix Spec */
1082 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
1088 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1091 tree.lt_fd = file->private_data;
1093 rc = ll_tree_lock(&tree, node, inode, buf, count,
1094 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1098 down(&lli->lli_size_sem);
1099 kms = lov_merge_size(lsm, 1);
1100 if (*ppos + count - 1 > kms) {
1101 /* A glimpse is necessary to determine whether we return a short
1102 * read or some zeroes at the end of the buffer */
1103 up(&lli->lli_size_sem);
1104 retval = ll_glimpse_size(inode);
1108 inode->i_size = kms;
1109 up(&lli->lli_size_sem);
1112 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1113 inode->i_ino, count, *ppos, inode->i_size);
1115 /* turn off the kernel's read-ahead */
1116 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1119 file->f_ra.ra_pages = 0;
1121 retval = generic_file_read(file, buf, count, ppos);
1124 ll_tree_unlock(&tree, inode);
1129 * Write to a file (through the page cache).
1131 static ssize_t ll_file_write(struct file *file, const char *buf,
1132 size_t count, loff_t *ppos)
1134 struct inode *inode = file->f_dentry->d_inode;
1135 loff_t maxbytes = ll_file_maxbytes(inode);
1136 struct ll_lock_tree tree;
1137 struct ll_lock_tree_node *node;
1142 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1143 inode->i_ino, inode->i_generation, inode, count, *ppos);
1145 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1147 /* POSIX, but surprised the VFS doesn't check this already */
1151 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1152 * called on the file, don't fail the below assertion (bug 2388). */
1153 if (file->f_flags & O_LOV_DELAY_CREATE &&
1154 ll_i2info(inode)->lli_smd == NULL)
1157 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1159 if (file->f_flags & O_APPEND)
1160 node = ll_node_from_inode(inode, 0, OBD_OBJECT_EOF, LCK_PW);
1162 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1166 RETURN(PTR_ERR(node));
1168 tree.lt_fd = file->private_data;
1170 rc = ll_tree_lock(&tree, node, inode, buf, count,
1171 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1175 /* this is ok, g_f_w will overwrite this under i_sem if it races
1176 * with a local truncate, it just makes our maxbyte checking easier */
1177 if (file->f_flags & O_APPEND)
1178 *ppos = inode->i_size;
1180 if (*ppos >= maxbytes) {
1181 if (count || *ppos > maxbytes) {
1182 send_sig(SIGXFSZ, current, 0);
1183 GOTO(out, retval = -EFBIG);
1186 if (*ppos + count > maxbytes)
1187 count = maxbytes - *ppos;
1189 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1190 inode->i_ino, count, *ppos);
1192 /* generic_file_write handles O_APPEND after getting i_sem */
1193 retval = generic_file_write(file, buf, count, ppos);
1196 ll_tree_unlock(&tree, inode);
1197 /* serialize with mmap/munmap/mremap */
1198 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
1199 retval > 0 ? retval : 0);
1203 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
1204 int flags, struct lov_user_md *lum,
1207 struct ll_inode_info *lli = ll_i2info(inode);
1209 struct obd_export *exp = ll_i2dtexp(inode);
1210 struct lov_stripe_md *lsm;
1211 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
1212 struct ptlrpc_request *req = NULL;
1214 struct lustre_md md;
1215 struct obd_client_handle *och;
1219 if ((file->f_flags+1) & O_ACCMODE)
1221 if (file->f_flags & O_TRUNC)
1224 down(&lli->lli_open_sem);
1227 up(&lli->lli_open_sem);
1228 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
1233 f = get_empty_filp();
1237 f->f_dentry = file->f_dentry;
1238 f->f_vfsmnt = file->f_vfsmnt;
1241 rc = ll_intent_alloc(&oit);
1245 rc = ll_intent_file_open(f, lum, lum_size, &oit);
1248 if (it_disposition(&oit, DISP_LOOKUP_NEG))
1251 req = LUSTRE_IT(&oit)->it_data;
1252 rc = LUSTRE_IT(&oit)->it_status;
1257 rc = mdc_req2lustre_md(ll_i2mdexp(inode), req, 1, exp, &md);
1260 ll_update_inode(f->f_dentry->d_inode, &md);
1262 OBD_ALLOC(och, sizeof(struct obd_client_handle));
1263 rc = ll_local_open(f, &oit, och);
1264 if (rc) { /* Actually ll_local_open cannot fail! */
1267 if (LUSTRE_IT(&oit)->it_lock_mode) {
1268 ldlm_lock_decref_and_cancel((struct lustre_handle *)
1269 &LUSTRE_IT(&oit)->it_lock_handle,
1270 LUSTRE_IT(&oit)->it_lock_mode);
1271 LUSTRE_IT(&oit)->it_lock_mode = 0;
1274 ll_intent_release(&oit);
1276 /* ll_file_release will decrease the count, but won't free anything
1277 because we have at least one more reference coming from actual open
1279 down(&lli->lli_och_sem);
1280 lli->lli_open_fd_write_count++;
1281 up(&lli->lli_och_sem);
1282 rc = ll_file_release(f->f_dentry->d_inode, f);
1284 /* Now also destroy our supplemental och */
1285 ll_md_och_close(ll_i2mdexp(inode), f->f_dentry->d_inode, och);
1288 ll_intent_release(&oit);
1291 up(&lli->lli_open_sem);
1293 ptlrpc_req_finished(req);
1297 static int ll_lov_setea(struct inode *inode, struct file *file,
1300 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
1301 struct lov_user_md *lump;
1302 int lum_size = sizeof(struct lov_user_md) +
1303 sizeof(struct lov_user_ost_data);
1307 if (!capable (CAP_SYS_ADMIN))
1310 OBD_ALLOC(lump, lum_size);
1314 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
1316 OBD_FREE(lump, lum_size);
1320 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1322 OBD_FREE(lump, lum_size);
1326 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1329 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1331 int flags = FMODE_WRITE;
1334 /* Bug 1152: copy properly when this is no longer true */
1335 LASSERT(sizeof(lum) == sizeof(*lump));
1336 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1337 rc = copy_from_user(&lum, lump, sizeof(lum));
1341 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1343 put_user(0, &lump->lmm_stripe_count);
1344 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
1345 0, ll_i2info(inode)->lli_smd, lump);
1350 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1352 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1357 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
1361 static int ll_get_grouplock(struct inode *inode, struct file *file,
1364 struct ll_file_data *fd = file->private_data;
1365 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1366 .end = OBD_OBJECT_EOF}};
1367 struct lustre_handle lockh = { 0 };
1368 struct ll_inode_info *lli = ll_i2info(inode);
1369 struct lov_stripe_md *lsm = lli->lli_smd;
1373 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1377 policy.l_extent.gid = arg;
1378 if (file->f_flags & O_NONBLOCK)
1379 flags = LDLM_FL_BLOCK_NOWAIT;
1381 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags,
1382 &ll_i2sbi(inode)->ll_grouplock_stime);
1386 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1388 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1393 static int ll_put_grouplock(struct inode *inode, struct file *file,
1396 struct ll_file_data *fd = file->private_data;
1397 struct ll_inode_info *lli = ll_i2info(inode);
1398 struct lov_stripe_md *lsm = lli->lli_smd;
1402 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1403 /* Ugh, it's already unlocked. */
1407 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1410 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1412 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1417 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1422 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1425 struct ll_file_data *fd = file->private_data;
1426 struct ll_sb_info *sbi = ll_i2sbi(inode);
1430 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1431 inode->i_generation, inode, cmd);
1433 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1436 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1438 case LL_IOC_GETFLAGS:
1439 /* Get the current value of the file flags */
1440 return put_user(fd->fd_flags, (int *)arg);
1441 case LL_IOC_SETFLAGS:
1442 case LL_IOC_CLRFLAGS:
1443 /* Set or clear specific file flags */
1444 /* XXX This probably needs checks to ensure the flags are
1445 * not abused, and to handle any flag side effects.
1447 if (get_user(flags, (int *) arg))
1450 if (cmd == LL_IOC_SETFLAGS)
1451 fd->fd_flags |= flags;
1453 fd->fd_flags &= ~flags;
1455 case LL_IOC_LOV_SETSTRIPE:
1456 RETURN(ll_lov_setstripe(inode, file, arg));
1457 case LL_IOC_LOV_SETEA:
1458 RETURN(ll_lov_setea(inode, file, arg));
1459 case IOC_MDC_SHOWFID: {
1460 struct lustre_id *idp = (struct lustre_id *)arg;
1461 struct lustre_id id;
1465 filename = getname((const char *)arg);
1466 if (IS_ERR(filename))
1467 RETURN(PTR_ERR(filename));
1469 ll_inode2id(&id, inode);
1471 rc = ll_get_fid(sbi->ll_md_exp, &id, filename, &id);
1473 GOTO(out_filename, rc);
1475 rc = copy_to_user(idp, &id, sizeof(*idp));
1477 GOTO(out_filename, rc = -EFAULT);
1484 case LL_IOC_LOV_GETSTRIPE:
1485 RETURN(ll_lov_getstripe(inode, arg));
1486 case EXT3_IOC_GETFLAGS:
1487 case EXT3_IOC_SETFLAGS:
1488 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1489 case LL_IOC_GROUP_LOCK:
1490 RETURN(ll_get_grouplock(inode, file, arg));
1491 case LL_IOC_GROUP_UNLOCK:
1492 RETURN(ll_put_grouplock(inode, file, arg));
1493 case EXT3_IOC_GETVERSION_OLD:
1494 case EXT3_IOC_GETVERSION:
1495 return put_user(inode->i_generation, (int *) arg);
1496 /* We need to special case any other ioctls we want to handle,
1497 * to send them to the MDS/OST as appropriate and to properly
1498 * network encode the arg field.
1499 case EXT2_IOC_GETVERSION_OLD:
1500 case EXT2_IOC_GETVERSION_NEW:
1501 case EXT2_IOC_SETVERSION_OLD:
1502 case EXT2_IOC_SETVERSION_NEW:
1503 case EXT3_IOC_SETVERSION_OLD:
1504 case EXT3_IOC_SETVERSION:
1506 case LL_IOC_FLUSH_CRED:
1507 RETURN(ll_flush_cred(inode));
1509 RETURN( obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
1514 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1516 struct inode *inode = file->f_dentry->d_inode;
1517 struct ll_file_data *fd = file->private_data;
1518 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1519 struct lustre_handle lockh = {0};
1522 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1523 inode->i_generation, inode,
1524 offset + ((origin==2) ? inode->i_size : file->f_pos));
1526 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1527 if (origin == 2) { /* SEEK_END */
1528 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1529 struct ll_inode_info *lli = ll_i2info(inode);
1530 int nonblock = 0, rc;
1532 if (file->f_flags & O_NONBLOCK)
1533 nonblock = LDLM_FL_BLOCK_NOWAIT;
1535 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1536 nonblock, &ll_i2sbi(inode)->ll_seek_stime);
1540 down(&lli->lli_size_sem);
1541 offset += inode->i_size;
1542 up(&lli->lli_size_sem);
1543 } else if (origin == 1) { /* SEEK_CUR */
1544 offset += file->f_pos;
1548 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1549 if (offset != file->f_pos) {
1550 file->f_pos = offset;
1551 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1553 file->f_version = ++event;
1560 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1564 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1566 struct inode *inode = dentry->d_inode;
1567 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1568 struct lustre_id id;
1569 struct ptlrpc_request *req;
1572 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1573 inode->i_generation, inode);
1575 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1577 /* fsync's caller has already called _fdata{sync,write}, we want
1578 * that IO to finish before calling the osc and mdc sync methods */
1579 rc = filemap_fdatawait(inode->i_mapping);
1581 ll_inode2id(&id, inode);
1582 err = md_sync(ll_i2sbi(inode)->ll_md_exp, &id, &req);
1586 ptlrpc_req_finished(req);
1589 struct obdo *oa = obdo_alloc();
1592 RETURN(rc ? rc : -ENOMEM);
1594 oa->o_id = lsm->lsm_object_id;
1595 oa->o_gr = lsm->lsm_object_gr;
1596 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1598 obdo_from_inode(oa, inode, (OBD_MD_FLTYPE | OBD_MD_FLATIME |
1599 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1602 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
1612 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1614 struct inode *inode = file->f_dentry->d_inode;
1615 struct ll_inode_info *li = ll_i2info(inode);
1616 struct ll_sb_info *sbi = ll_i2sbi(inode);
1617 struct obd_device *obddev;
1618 struct ldlm_res_id res_id =
1619 { .name = {id_fid(&li->lli_id), id_group(&li->lli_id), LDLM_FLOCK} };
1620 struct lustre_handle lockh = {0};
1621 ldlm_policy_data_t flock;
1622 ldlm_mode_t mode = 0;
1627 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1628 inode->i_ino, file_lock);
1630 flock.l_flock.pid = file_lock->fl_pid;
1631 flock.l_flock.start = file_lock->fl_start;
1632 flock.l_flock.end = file_lock->fl_end;
1634 switch (file_lock->fl_type) {
1639 /* An unlock request may or may not have any relation to
1640 * existing locks so we may not be able to pass a lock handle
1641 * via a normal ldlm_lock_cancel() request. The request may even
1642 * unlock a byte range in the middle of an existing lock. In
1643 * order to process an unlock request we need all of the same
1644 * information that is given with a normal read or write record
1645 * lock request. To avoid creating another ldlm unlock (cancel)
1646 * message we'll treat a LCK_NL flock request as an unlock. */
1653 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1668 flags = LDLM_FL_BLOCK_NOWAIT;
1674 flags = LDLM_FL_TEST_LOCK;
1675 /* Save the old mode so that if the mode in the lock changes we
1676 * can decrement the appropriate reader or writer refcount. */
1677 file_lock->fl_type = mode;
1680 CERROR("unknown fcntl lock command: %d\n", cmd);
1684 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1685 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1686 flags, mode, flock.l_flock.start, flock.l_flock.end);
1688 obddev = md_get_real_obd(sbi->ll_md_exp, &li->lli_id);
1689 rc = ldlm_cli_enqueue(obddev->obd_self_export, NULL,
1690 obddev->obd_namespace,
1691 res_id, LDLM_FLOCK, &flock, mode, &flags,
1692 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1693 NULL, 0, NULL, &lockh);
1697 int ll_inode_revalidate_it(struct dentry *dentry)
1699 struct lookup_intent oit = { .it_op = IT_GETATTR };
1700 struct inode *inode = dentry->d_inode;
1701 struct ptlrpc_request *req = NULL;
1702 struct ll_inode_info *lli;
1703 struct lov_stripe_md *lsm;
1704 struct ll_sb_info *sbi;
1705 struct lustre_id id;
1710 CERROR("REPORT THIS LINE TO PETER\n");
1714 sbi = ll_i2sbi(inode);
1716 ll_inode2id(&id, inode);
1717 lli = ll_i2info(inode);
1718 LASSERT(id_fid(&id) != 0);
1720 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), name=%s(%p)\n",
1721 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1724 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1725 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1728 rc = ll_intent_alloc(&oit);
1732 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1733 &oit, 0, &req, ll_mdc_blocking_ast);
1737 rc = revalidate_it_finish(req, 1, &oit, dentry);
1742 ll_lookup_finish_locks(&oit, dentry);
1744 if (!LLI_HAVE_FLSIZE(inode)) {
1745 /* if object not yet allocated, don't validate size */
1748 /* ll_glimpse_size() will prefer locally cached
1749 * writes if they extend the file */
1750 rc = ll_glimpse_size(inode);
1755 ll_intent_release(&oit);
1757 ptlrpc_req_finished(req);
1761 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1762 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
1765 struct inode *inode = de->d_inode;
1766 struct ll_inode_info *lli = ll_i2info(inode);
1768 res = ll_inode_revalidate_it(de);
1769 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1774 stat->ino = inode->i_ino;
1775 stat->mode = inode->i_mode;
1776 stat->nlink = inode->i_nlink;
1777 stat->uid = inode->i_uid;
1778 stat->gid = inode->i_gid;
1779 stat->atime = inode->i_atime;
1780 stat->mtime = inode->i_mtime;
1781 stat->ctime = inode->i_ctime;
1782 stat->blksize = inode->i_blksize;
1784 down(&lli->lli_size_sem);
1785 stat->size = inode->i_size;
1786 stat->blocks = inode->i_blocks;
1787 up(&lli->lli_size_sem);
1789 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1790 stat->dev = id_group(&ll_i2info(inode)->lli_id);
1796 int ll_setxattr_internal(struct inode *inode, const char *name,
1797 const void *value, size_t size, int flags,
1800 struct ll_sb_info *sbi = ll_i2sbi(inode);
1801 struct ptlrpc_request *request = NULL;
1802 struct mdc_op_data op_data;
1807 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu\n", inode->i_ino);
1808 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_SETXATTR);
1810 memset(&attr, 0x0, sizeof(attr));
1811 attr.ia_valid |= valid;
1812 attr.ia_attr_flags = flags;
1814 ll_prepare_mdc_data(&op_data, inode, NULL, NULL, 0, 0);
1816 rc = md_setattr(sbi->ll_md_exp, &op_data, &attr,
1817 (void*) name, strnlen(name, XATTR_NAME_MAX)+1,
1818 (void*) value, size, &request);
1820 CERROR("md_setattr fails: rc = %d\n", rc);
1825 ptlrpc_req_finished(request);
1829 int ll_setxattr(struct dentry *dentry, const char *name, const void *value,
1830 size_t size, int flags)
1833 struct posix_acl *acl;
1834 struct ll_inode_info *lli;
1837 rc = ll_setxattr_internal(dentry->d_inode, name, value, size,
1840 /* update inode's acl info */
1841 if (rc == 0 && strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1843 acl = posix_acl_from_xattr(value, size);
1845 CERROR("convert from xattr to acl error: %ld",
1849 error = posix_acl_valid(acl);
1851 CERROR("acl valid error: %d", error);
1852 posix_acl_release(acl);
1860 lli = ll_i2info(dentry->d_inode);
1861 spin_lock(&lli->lli_lock);
1862 if (lli->lli_posix_acl != NULL)
1863 posix_acl_release(lli->lli_posix_acl);
1864 lli->lli_posix_acl = acl;
1865 spin_unlock(&lli->lli_lock);
1872 int ll_removexattr(struct dentry *dentry, const char *name)
1874 return ll_setxattr_internal(dentry->d_inode, name, NULL, 0, 0,
1879 int ll_getxattr_internal(struct inode *inode, const char *name, int namelen,
1880 void *value, size_t size, __u64 valid)
1882 struct ptlrpc_request *request = NULL;
1883 struct ll_sb_info *sbi = ll_i2sbi(inode);
1884 struct lustre_id id;
1885 struct mds_body *body;
1890 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETXATTR);
1892 ll_inode2id(&id, inode);
1893 rc = md_getattr(sbi->ll_md_exp, &id, valid, name, namelen,
1896 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1897 CERROR("md_getattr fails: rc = %d\n", rc);
1901 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
1902 LASSERT(body != NULL);
1903 LASSERT_REPSWABBED(request, 0);
1905 ea_size = body->eadatasize;
1906 LASSERT(ea_size <= request->rq_repmsg->buflens[0]);
1909 GOTO(out, rc = ea_size);
1911 ea_data = lustre_msg_buf(request->rq_repmsg, 1, ea_size);
1912 LASSERT(ea_data != NULL);
1913 LASSERT_REPSWABBED(request, 1);
1916 memcpy(value, ea_data, ea_size);
1919 ptlrpc_req_finished(request);
1923 int ll_getxattr(struct dentry *dentry, const char *name, void *value,
1926 return ll_getxattr_internal(dentry->d_inode, name, strlen(name) + 1,
1927 value, size, OBD_MD_FLEA);
1930 int ll_listxattr(struct dentry *dentry, char *list, size_t size)
1932 return ll_getxattr_internal(dentry->d_inode, NULL, 0, list, size,
1937 * XXX We could choose not to check DLM lock. Leave the decision
1938 * to remote acl handling.
1941 lustre_check_acl(struct inode *inode, int mask)
1943 struct lookup_intent it = { .it_op = IT_GETATTR };
1944 struct dentry de = { .d_inode = inode };
1945 struct ll_sb_info *sbi;
1946 struct lustre_id id;
1947 struct ptlrpc_request *req = NULL;
1948 struct ll_inode_info *lli = ll_i2info(inode);
1949 struct posix_acl *acl;
1953 sbi = ll_i2sbi(inode);
1954 ll_inode2id(&id, inode);
1956 if (ll_intent_alloc(&it))
1959 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1960 &it, 0, &req, ll_mdc_blocking_ast);
1962 ll_intent_free(&it);
1966 rc = revalidate_it_finish(req, 1, &it, &de);
1968 ll_intent_release(&it);
1972 if (sbi->ll_remote) {
1973 rc = ll_remote_acl_permission(inode, mask);
1975 spin_lock(&lli->lli_lock);
1976 acl = posix_acl_dup(ll_i2info(inode)->lli_posix_acl);
1977 spin_unlock(&lli->lli_lock);
1982 rc = posix_acl_permission(inode, acl, mask);
1983 posix_acl_release(acl);
1987 ll_lookup_finish_locks(&it, &de);
1988 ll_intent_free(&it);
1992 ptlrpc_req_finished(req);
1997 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
1999 return generic_permission(inode, mask, lustre_check_acl);
2002 struct file_operations ll_file_operations = {
2003 .read = ll_file_read,
2004 .write = ll_file_write,
2005 .ioctl = ll_file_ioctl,
2006 .open = ll_file_open,
2007 .release = ll_file_release,
2008 .mmap = ll_file_mmap,
2009 .llseek = ll_file_seek,
2010 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
2011 .sendfile = generic_file_sendfile,
2014 .lock = ll_file_flock
2017 struct inode_operations ll_file_inode_operations = {
2018 .setattr = ll_setattr,
2019 .truncate = ll_truncate,
2020 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
2021 .getattr = ll_getattr,
2023 .revalidate_it = ll_inode_revalidate_it,
2025 .setxattr = ll_setxattr,
2026 .getxattr = ll_getxattr,
2027 .listxattr = ll_listxattr,
2028 .removexattr = ll_removexattr,
2029 .permission = ll_inode_permission,