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 int ll_md_och_close(struct obd_export *md_exp, struct inode *inode,
38 struct obd_client_handle *och)
40 struct ptlrpc_request *req = NULL;
41 struct obdo *obdo = NULL;
42 struct obd_device *obd;
46 obd = class_exp2obd(md_exp);
48 CERROR("Invalid MDC connection handle "LPX64"\n",
49 md_exp->exp_handle.h_cookie);
55 * here we check if this is forced umount. If so this is called on
56 * canceling "open lock" and we do not call md_close() in this case , as
57 * it will not successful, as import is already deactivated.
59 if (obd->obd_no_recov)
62 /* closing opened file */
67 obdo->o_id = inode->i_ino;
68 obdo->o_valid = OBD_MD_FLID;
69 obdo_from_inode(obdo, inode, (OBD_MD_FLTYPE | OBD_MD_FLMODE |
70 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
71 OBD_MD_FLATIME | OBD_MD_FLMTIME |
73 if (0 /* ll_is_inode_dirty(inode) */) {
74 obdo->o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
75 obdo->o_valid |= OBD_MD_FLFLAGS;
77 obdo->o_fid = id_fid(&ll_i2info(inode)->lli_id);
78 obdo->o_mds = id_group(&ll_i2info(inode)->lli_id);
79 rc = md_close(md_exp, obdo, och, &req);
84 * we are the last writer, so the MDS has instructed us to get
85 * the file size and any write cookies, then close again.
88 //ll_queue_done_writing(inode);
91 CERROR("inode %lu mdc close failed: rc = %d\n",
92 (unsigned long)inode->i_ino, rc);
95 ptlrpc_req_finished(req);
98 mdc_clear_open_replay_data(md_exp, och);
99 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
100 OBD_FREE(och, sizeof *och);
104 int ll_md_real_close(struct obd_export *md_exp,
105 struct inode *inode, int flags)
107 struct ll_inode_info *lli = ll_i2info(inode);
108 struct obd_client_handle **och_p;
109 struct obd_client_handle *och;
114 if (flags & FMODE_WRITE) {
115 och_p = &lli->lli_mds_write_och;
116 och_usecount = &lli->lli_open_fd_write_count;
117 } else if (flags & FMODE_EXEC) {
118 och_p = &lli->lli_mds_exec_och;
119 och_usecount = &lli->lli_open_fd_exec_count;
121 och_p = &lli->lli_mds_read_och;
122 och_usecount = &lli->lli_open_fd_read_count;
125 down(&lli->lli_och_sem);
126 if (*och_usecount) { /* There are still users of this handle, so
128 up(&lli->lli_och_sem);
134 up(&lli->lli_och_sem);
137 * there might be a race and somebody have freed this och
138 * already. Another way to have this twice called is if file closing
139 * will fail due to netwok problems and on umount lock will be canceled
140 * and this will be called from block_ast callack.
142 if (och && och->och_fh.cookie != DEAD_HANDLE_MAGIC)
143 rc = ll_md_och_close(md_exp, inode, och);
148 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
151 struct ll_file_data *fd = file->private_data;
152 struct ll_inode_info *lli = ll_i2info(inode);
156 /* clear group lock, if present */
157 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
158 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
159 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
160 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
164 /* Let's see if we have good enough OPEN lock on the file and if
165 we can skip talking to MDS */
166 if (file->f_dentry->d_inode) {
168 struct obd_device *obddev;
169 struct lustre_handle lockh;
170 int flags = LDLM_FL_BLOCK_GRANTED;
171 struct ldlm_res_id file_res_id = {.name = {id_fid(&lli->lli_id),
172 id_group(&lli->lli_id)}};
173 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
175 down(&lli->lli_och_sem);
176 if (fd->fd_omode & FMODE_WRITE) {
178 LASSERT(lli->lli_open_fd_write_count);
179 lli->lli_open_fd_write_count--;
180 } else if (fd->fd_omode & FMODE_EXEC) {
182 LASSERT(lli->lli_open_fd_exec_count);
183 lli->lli_open_fd_exec_count--;
186 LASSERT(lli->lli_open_fd_read_count);
187 lli->lli_open_fd_read_count--;
189 up(&lli->lli_och_sem);
191 obddev = md_get_real_obd(md_exp, &lli->lli_id);
192 if (!ldlm_lock_match(obddev->obd_namespace, flags, &file_res_id,
193 LDLM_IBITS, &policy, lockmode, &lockh))
195 rc = ll_md_real_close(md_exp, file->f_dentry->d_inode,
198 ldlm_lock_decref(&lockh, lockmode);
202 file->private_data = NULL;
203 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof(*fd));
207 /* While this returns an error code, fput() the caller does not, so we need
208 * to make every effort to clean up all of our state here. Also, applications
209 * rarely check close errors and even if an error is returned they will not
210 * re-try the close call.
212 int ll_file_release(struct inode *inode, struct file *file)
214 struct ll_file_data *fd;
215 struct ll_sb_info *sbi = ll_i2sbi(inode);
219 CDEBUG(D_VFSTRACE, "VFS Op:inode="DLID4"(%p)\n",
220 OLID4(&ll_i2info(inode)->lli_id), inode);
222 /* don't do anything for / */
223 if (inode->i_sb->s_root == file->f_dentry)
226 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
227 fd = (struct ll_file_data *)file->private_data;
230 rc = ll_md_close(sbi->ll_md_exp, inode, file);
234 static int ll_intent_file_open(struct file *file, void *lmm,
235 int lmmsize, struct lookup_intent *itp)
237 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
238 const char *name = (char *)file->f_dentry->d_name.name;
239 struct dentry *parent = file->f_dentry->d_parent;
240 const int len = file->f_dentry->d_name.len;
241 struct lustre_handle lockh;
242 struct mdc_op_data *op_data;
248 OBD_ALLOC(op_data, sizeof(*op_data));
252 ll_prepare_mdc_data(op_data, parent->d_inode, NULL,
255 rc = md_enqueue(sbi->ll_md_exp, LDLM_IBITS, itp, LCK_PR, op_data,
256 &lockh, lmm, lmmsize, ldlm_completion_ast,
257 ll_mdc_blocking_ast, NULL);
258 OBD_FREE(op_data, sizeof(*op_data));
260 if (LUSTRE_IT(itp)->it_lock_mode)
261 memcpy(&LUSTRE_IT(itp)->it_lock_handle,
262 &lockh, sizeof(lockh));
265 CERROR("lock enqueue: err: %d\n", rc);
270 void ll_och_fill(struct inode *inode, struct lookup_intent *it,
271 struct obd_client_handle *och)
273 struct ptlrpc_request *req = LUSTRE_IT(it)->it_data;
274 struct ll_inode_info *lli = ll_i2info(inode);
275 struct mds_body *body;
278 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
279 LASSERT (body != NULL); /* reply already checked out */
280 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
282 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
283 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
284 lli->lli_io_epoch = body->io_epoch;
285 mdc_set_open_replay_data(ll_i2mdexp(inode), och,
286 LUSTRE_IT(it)->it_data);
289 int ll_local_open(struct file *file, struct lookup_intent *it,
290 struct obd_client_handle *och)
292 struct ll_file_data *fd;
296 ll_och_fill(file->f_dentry->d_inode, it, och);
298 LASSERTF(file->private_data == NULL, "file %.*s/%.*s ino %lu/%u (%o)\n",
299 file->f_dentry->d_name.len, file->f_dentry->d_name.name,
300 file->f_dentry->d_parent->d_name.len,
301 file->f_dentry->d_parent->d_name.name,
302 file->f_dentry->d_inode->i_ino,
303 file->f_dentry->d_inode->i_generation,
304 file->f_dentry->d_inode->i_mode);
306 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
308 /* We can't handle this well without reorganizing ll_file_open and
309 * ll_md_close(), so don't even try right now. */
312 file->private_data = fd;
313 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
314 fd->fd_omode = it->it_flags;
318 /* Open a file, and (for the very first open) create objects on the OSTs at
319 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
320 * creation or open until ll_lov_setstripe() ioctl is called. We grab
321 * lli_open_sem to ensure no other process will create objects, send the
322 * stripe MD to the MDS, or try to destroy the objects if that fails.
324 * If we already have the stripe MD locally then we don't request it in
325 * mdc_open(), by passing a lmm_size = 0.
327 * It is up to the application to ensure no other processes open this file
328 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
329 * used. We might be able to avoid races of that sort by getting lli_open_sem
330 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
331 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
333 int ll_file_open(struct inode *inode, struct file *file)
335 struct ll_inode_info *lli = ll_i2info(inode);
336 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
337 .it_flags = file->f_flags };
338 struct lov_stripe_md *lsm;
339 struct ptlrpc_request *req;
341 struct obd_client_handle **och_p;
345 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n",
346 inode->i_ino, inode->i_generation, inode, file->f_flags);
348 /* don't do anything for / */
349 if (inode->i_sb->s_root == file->f_dentry)
352 if ((file->f_flags+1) & O_ACCMODE)
354 if (file->f_flags & O_TRUNC)
360 * sometimes LUSTRE_IT(it) may not be allocated like opening file by
361 * dentry_open() from GNS stuff.
363 if (!it || !LUSTRE_IT(it)) {
365 rc = ll_intent_alloc(it);
370 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
373 * mdc_intent_lock() didn't get a request ref if there was an open
374 * error, so don't do cleanup on the * request here (bug 3430)
376 if (LUSTRE_IT(it)->it_disposition) {
377 rc = it_open_error(DISP_OPEN_OPEN, it);
382 /* Let's see if we have file open on MDS already. */
383 if (it->it_flags & FMODE_WRITE) {
384 och_p = &lli->lli_mds_write_och;
385 och_usecount = &lli->lli_open_fd_write_count;
386 } else if (it->it_flags & FMODE_EXEC) {
387 och_p = &lli->lli_mds_exec_och;
388 och_usecount = &lli->lli_open_fd_exec_count;
390 och_p = &lli->lli_mds_read_och;
391 och_usecount = &lli->lli_open_fd_read_count;
394 down(&lli->lli_och_sem);
395 if (*och_p) { /* Open handle is present */
396 if (LUSTRE_IT(it)->it_disposition) {
397 struct obd_client_handle *och;
398 /* Well, there's extra open request that we do not need,
399 let's close it somehow*/
400 OBD_ALLOC(och, sizeof (struct obd_client_handle));
402 up(&lli->lli_och_sem);
406 ll_och_fill(inode, it, och);
407 /* ll_md_och_close() will free och */
408 ll_md_och_close(ll_i2mdexp(inode), inode, och);
412 rc = ll_local_open(file, it, NULL);
416 LASSERT(*och_usecount == 0);
417 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
419 GOTO(out, rc = -ENOMEM);
422 if (!it || !LUSTRE_IT(it) || !LUSTRE_IT(it)->it_disposition) {
424 * we are going to replace intent here, and that may
425 * possibly change access mode (FMODE_EXEC can only be
426 * set in intent), but I hope it never happens (I was
427 * not able to trigger it yet at least) -- green
430 /* FIXME: FMODE_EXEC is not covered by O_ACCMODE! */
431 LASSERT(!(it->it_flags & FMODE_EXEC));
432 LASSERTF((it->it_flags & O_ACCMODE) ==
433 (oit.it_flags & O_ACCMODE), "Changing intent "
434 "flags %x to incompatible %x\n", it->it_flags,
437 rc = ll_intent_file_open(file, NULL, 0, it);
440 rc = it_open_error(DISP_OPEN_OPEN, it);
442 GOTO(out_och_free, rc);
444 mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle,
445 file->f_dentry->d_inode);
447 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
448 rc = ll_local_open(file, it, *och_p);
449 LASSERTF(rc == 0, "rc = %d\n", rc);
451 up(&lli->lli_och_sem);
454 * must do this outside lli_och_sem lock to prevent deadlock where
455 * different kind of OPEN lock for this same inode gets cancelled by
459 if (!S_ISREG(inode->i_mode))
464 if (file->f_flags & O_LOV_DELAY_CREATE ||
465 !(file->f_mode & FMODE_WRITE)) {
466 CDEBUG(D_INODE, "object creation was delayed\n");
470 file->f_flags &= ~O_LOV_DELAY_CREATE;
473 req = LUSTRE_IT(it)->it_data;
474 ll_intent_drop_lock(it);
475 ll_intent_release(it);
476 ptlrpc_req_finished(req);
478 ll_open_complete(inode);
482 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
483 *och_p = NULL; /* OBD_FREE writes some magic there */
486 up(&lli->lli_och_sem);
492 /* Fills the obdo with the attributes for the inode defined by lsm */
493 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
496 struct ptlrpc_request_set *set;
500 LASSERT(lsm != NULL);
502 memset(oa, 0, sizeof *oa);
503 oa->o_id = lsm->lsm_object_id;
504 oa->o_gr = lsm->lsm_object_gr;
505 oa->o_mode = S_IFREG;
506 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
507 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
508 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
510 set = ptlrpc_prep_set();
514 rc = obd_getattr_async(exp, oa, lsm, set);
516 rc = ptlrpc_set_wait(set);
517 ptlrpc_set_destroy(set);
522 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
523 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
527 static inline void ll_remove_suid(struct inode *inode)
531 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
532 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
534 /* was any of the uid bits set? */
535 mode &= inode->i_mode;
536 if (mode && !capable(CAP_FSETID)) {
537 inode->i_mode &= ~mode;
538 // XXX careful here - we cannot change the size
542 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
544 struct ll_inode_info *lli = ll_i2info(inode);
545 struct lov_stripe_md *lsm = lli->lli_smd;
546 struct obd_export *exp = ll_i2dtexp(inode);
549 struct ldlm_lock *lock;
550 struct lov_stripe_md *lsm;
551 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
552 __u32 stripe, vallen = sizeof(stripe);
556 if (lsm->lsm_stripe_count == 1)
557 GOTO(check, stripe = 0);
559 /* get our offset in the lov */
560 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
562 CERROR("obd_get_info: rc = %d\n", rc);
565 LASSERT(stripe < lsm->lsm_stripe_count);
568 if (lsm->lsm_oinfo[stripe].loi_id != lock->l_resource->lr_name.name[0]||
569 lsm->lsm_oinfo[stripe].loi_gr != lock->l_resource->lr_name.name[2]){
570 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64
571 " inode=%lu/%u (%p)\n",
572 lsm->lsm_oinfo[stripe].loi_id,
573 lsm->lsm_oinfo[stripe].loi_gr,
574 inode->i_ino, inode->i_generation, inode);
575 return -ELDLM_NO_LOCK_DATA;
581 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
582 * we get a lock cancellation for each stripe, so we have to map the obd's
583 * region back onto the stripes in the file that it held.
585 * No one can dirty the extent until we've finished our work and they can
586 * enqueue another lock. The DLM protects us from ll_file_read/write here,
587 * but other kernel actors could have pages locked.
589 * Called with the DLM lock held. */
590 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
591 struct ldlm_lock *lock, __u32 stripe)
593 ldlm_policy_data_t tmpex;
594 unsigned long start, end, count, skip, i, j;
596 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
597 struct lustre_handle lockh;
600 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
601 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
602 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
605 /* our locks are page granular thanks to osc_enqueue, we invalidate the
607 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
608 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
612 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
613 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
614 if (lsm->lsm_stripe_count > 1) {
615 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
616 skip = (lsm->lsm_stripe_count - 1) * count;
617 start += start/count * skip + stripe * count;
619 end += end/count * skip + stripe * count;
621 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
624 i = inode->i_size ? (inode->i_size - 1) >> PAGE_CACHE_SHIFT : 0;
628 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
629 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
630 count, skip, end, discard ? " (DISCARDING)" : "");
632 /* walk through the vmas on the inode and tear down mmaped pages that
633 * intersect with the lock. this stops immediately if there are no
634 * mmap()ed regions of the file. This is not efficient at all and
635 * should be short lived. We'll associate mmap()ed pages with the lock
636 * and will be able to find them directly */
638 for (i = start; i <= end; i += (j + skip)) {
639 j = min(count - (i % count), end - i + 1);
641 LASSERT(inode->i_mapping);
642 if (ll_teardown_mmaps(inode->i_mapping, i << PAGE_CACHE_SHIFT,
643 ((i+j) << PAGE_CACHE_SHIFT) - 1) )
647 /* this is the simplistic implementation of page eviction at
648 * cancelation. It is careful to get races with other page
649 * lockers handled correctly. fixes from bug 20 will make it
650 * more efficient by associating locks with pages and with
651 * batching writeback under the lock explicitly. */
652 for (i = start, j = start % count; i <= end;
653 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
655 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
661 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
662 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
663 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
666 if (!mapping_has_pages(inode->i_mapping)) {
667 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
673 page = find_get_page(inode->i_mapping, i);
676 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
677 i, tmpex.l_extent.start);
680 /* page->mapping to check with racing against teardown */
681 if (!discard && clear_page_dirty_for_io(page)) {
682 rc = ll_call_writepage(inode, page);
684 CERROR("writepage of page %p failed: %d\n",
686 /* either waiting for io to complete or reacquiring
687 * the lock that the failed writepage released */
691 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
692 /* check to see if another DLM lock covers this page */
693 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
694 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
696 &lock->l_resource->lr_name, LDLM_EXTENT,
697 &tmpex, LCK_PR | LCK_PW, &lockh);
698 if (rc2 == 0 && page->mapping != NULL) {
699 // checking again to account for writeback's lock_page()
700 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
701 ll_ra_accounting(page, inode->i_mapping);
702 ll_truncate_complete_page(page);
705 page_cache_release(page);
707 LASSERTF(tmpex.l_extent.start <=
708 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
709 lock->l_policy_data.l_extent.end + 1),
710 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
711 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
716 static int ll_extent_lock_callback(struct ldlm_lock *lock,
717 struct ldlm_lock_desc *new, void *data,
720 struct lustre_handle lockh = { 0 };
724 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
725 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
730 case LDLM_CB_BLOCKING:
731 ldlm_lock2handle(lock, &lockh);
732 rc = ldlm_cli_cancel(&lockh);
734 CERROR("ldlm_cli_cancel failed: %d\n", rc);
736 case LDLM_CB_CANCELING: {
738 struct ll_inode_info *lli;
739 struct lov_stripe_md *lsm;
743 /* This lock wasn't granted, don't try to evict pages */
744 if (lock->l_req_mode != lock->l_granted_mode)
747 inode = ll_inode_from_lock(lock);
750 lli = ll_i2info(inode);
753 if (lli->lli_smd == NULL)
757 stripe = ll_lock_to_stripe_offset(inode, lock);
760 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
762 l_lock(&lock->l_resource->lr_namespace->ns_lock);
763 down(&lli->lli_size_sem);
764 kms = ldlm_extent_shift_kms(lock,
765 lsm->lsm_oinfo[stripe].loi_kms);
767 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
768 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
769 lsm->lsm_oinfo[stripe].loi_kms, kms);
770 lsm->lsm_oinfo[stripe].loi_kms = kms;
771 up(&lli->lli_size_sem);
772 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
773 //ll_try_done_writing(inode);
786 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
788 /* XXX ALLOCATE - 160 bytes */
789 struct inode *inode = ll_inode_from_lock(lock);
790 struct ll_inode_info *lli = ll_i2info(inode);
791 struct lustre_handle lockh = { 0 };
796 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
797 LDLM_FL_BLOCK_CONV)) {
798 LBUG(); /* not expecting any blocked async locks yet */
799 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
801 ldlm_lock_dump(D_OTHER, lock, 0);
802 ldlm_reprocess_all(lock->l_resource);
806 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
808 stripe = ll_lock_to_stripe_offset(inode, lock);
812 if (lock->l_lvb_len) {
813 struct lov_stripe_md *lsm = lli->lli_smd;
815 lvb = lock->l_lvb_data;
816 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
818 l_lock(&lock->l_resource->lr_namespace->ns_lock);
820 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
821 kms = ldlm_extent_shift_kms(NULL, kms);
822 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
823 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
824 lsm->lsm_oinfo[stripe].loi_kms, kms);
825 lsm->lsm_oinfo[stripe].loi_kms = kms;
827 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
832 wake_up(&lock->l_waitq);
834 ldlm_lock2handle(lock, &lockh);
835 ldlm_lock_decref(&lockh, LCK_PR);
840 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
842 struct ptlrpc_request *req = reqp;
843 struct inode *inode = ll_inode_from_lock(lock);
844 struct ll_inode_info *lli;
846 struct lov_stripe_md *lsm;
847 int rc, size = sizeof(*lvb), stripe;
851 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
852 lli = ll_i2info(inode);
854 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
858 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
860 /* First, find out which stripe index this lock corresponds to. */
861 stripe = ll_lock_to_stripe_offset(inode, lock);
863 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
865 rc = lustre_pack_reply(req, 1, &size, NULL);
867 CERROR("lustre_pack_reply: %d\n", rc);
871 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
872 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
873 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
874 lvb->lvb_atime = LTIME_S(inode->i_atime);
875 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
877 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
878 inode->i_size, stripe, lvb->lvb_size);
884 /* These errors are normal races, so we don't want to fill the console
885 * with messages by calling ptlrpc_error() */
886 if (rc == -ELDLM_NO_LOCK_DATA)
887 lustre_pack_reply(req, 0, NULL, NULL);
893 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
894 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
895 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
897 /* NB: lov_merge_size will prefer locally cached writes if they extend the
898 * file (because it prefers KMS over RSS when larger) */
899 int ll_glimpse_size(struct inode *inode)
901 struct ll_inode_info *lli = ll_i2info(inode);
902 struct ll_sb_info *sbi = ll_i2sbi(inode);
903 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
904 struct lustre_handle lockh = { 0 };
905 int rc, flags = LDLM_FL_HAS_INTENT;
908 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
910 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
911 LCK_PR, &flags, ll_extent_lock_callback,
912 ldlm_completion_ast, ll_glimpse_callback, inode,
913 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
918 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
919 RETURN(rc > 0 ? -EIO : rc);
922 down(&lli->lli_size_sem);
923 inode->i_size = lov_merge_size(lli->lli_smd, 0);
924 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
925 up(&lli->lli_size_sem);
927 LTIME_S(inode->i_mtime) = lov_merge_mtime(lli->lli_smd,
928 LTIME_S(inode->i_mtime));
930 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
931 (__u64)inode->i_size, (__u64)inode->i_blocks);
933 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
937 void ll_stime_record(struct ll_sb_info *sbi, struct timeval *start,
938 struct obd_service_time *stime)
941 do_gettimeofday(&stop);
943 spin_lock(&sbi->ll_lock);
944 lprocfs_stime_record(stime, &stop, start);
945 spin_unlock(&sbi->ll_lock);
948 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
949 struct lov_stripe_md *lsm, int mode,
950 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
951 int ast_flags, struct obd_service_time *stime)
953 struct ll_inode_info *lli = ll_i2info(inode);
954 struct ll_sb_info *sbi = ll_i2sbi(inode);
955 struct timeval start;
959 LASSERT(lockh->cookie == 0);
961 /* XXX phil: can we do this? won't it screw the file size up? */
962 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
963 (sbi->ll_flags & LL_SBI_NOLCK))
966 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
967 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
969 do_gettimeofday(&start);
970 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
971 &ast_flags, ll_extent_lock_callback,
972 ldlm_completion_ast, ll_glimpse_callback, inode,
973 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
977 ll_stime_record(sbi, &start, stime);
979 if (policy->l_extent.start == 0 &&
980 policy->l_extent.end == OBD_OBJECT_EOF) {
981 /* vmtruncate()->ll_truncate() first sets the i_size and then
982 * the kms under both a DLM lock and the i_sem. If we don't
983 * get the i_sem here we can match the DLM lock and reset
984 * i_size from the kms before the truncating path has updated
985 * the kms. generic_file_write can then trust the stale i_size
986 * when doing appending writes and effectively cancel the
987 * result of the truncate. Getting the i_sem after the enqueue
988 * maintains the DLM -> i_sem acquiry order. */
989 down(&lli->lli_size_sem);
990 inode->i_size = lov_merge_size(lsm, 1);
991 up(&lli->lli_size_sem);
995 LTIME_S(inode->i_mtime) =
996 lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1002 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1003 struct lov_stripe_md *lsm, int mode,
1004 struct lustre_handle *lockh)
1006 struct ll_sb_info *sbi = ll_i2sbi(inode);
1010 /* XXX phil: can we do this? won't it screw the file size up? */
1011 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1012 (sbi->ll_flags & LL_SBI_NOLCK))
1015 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1020 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1023 struct inode *inode = file->f_dentry->d_inode;
1024 struct ll_inode_info *lli = ll_i2info(inode);
1025 struct lov_stripe_md *lsm = lli->lli_smd;
1026 struct ll_lock_tree tree;
1027 struct ll_lock_tree_node *node;
1032 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1033 inode->i_ino, inode->i_generation, inode, count, *ppos);
1035 /* "If nbyte is 0, read() will return 0 and have no other results."
1036 * -- Single Unix Spec */
1040 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
1046 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1049 tree.lt_fd = file->private_data;
1051 rc = ll_tree_lock(&tree, node, inode, buf, count,
1052 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1056 down(&lli->lli_size_sem);
1057 kms = lov_merge_size(lsm, 1);
1058 if (*ppos + count - 1 > kms) {
1059 /* A glimpse is necessary to determine whether we return a short
1060 * read or some zeroes at the end of the buffer */
1061 up(&lli->lli_size_sem);
1062 retval = ll_glimpse_size(inode);
1066 inode->i_size = kms;
1067 up(&lli->lli_size_sem);
1070 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1071 inode->i_ino, count, *ppos, inode->i_size);
1073 /* turn off the kernel's read-ahead */
1074 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1077 file->f_ra.ra_pages = 0;
1079 retval = generic_file_read(file, buf, count, ppos);
1082 ll_tree_unlock(&tree, inode);
1087 * Write to a file (through the page cache).
1089 static ssize_t ll_file_write(struct file *file, const char *buf,
1090 size_t count, loff_t *ppos)
1092 struct inode *inode = file->f_dentry->d_inode;
1093 loff_t maxbytes = ll_file_maxbytes(inode);
1094 struct ll_lock_tree tree;
1095 struct ll_lock_tree_node *node;
1100 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1101 inode->i_ino, inode->i_generation, inode, count, *ppos);
1103 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1105 /* POSIX, but surprised the VFS doesn't check this already */
1109 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1110 * called on the file, don't fail the below assertion (bug 2388). */
1111 if (file->f_flags & O_LOV_DELAY_CREATE &&
1112 ll_i2info(inode)->lli_smd == NULL)
1115 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1117 if (file->f_flags & O_APPEND)
1118 node = ll_node_from_inode(inode, 0, OBD_OBJECT_EOF, LCK_PW);
1120 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1124 RETURN(PTR_ERR(node));
1126 tree.lt_fd = file->private_data;
1128 rc = ll_tree_lock(&tree, node, inode, buf, count,
1129 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1133 /* this is ok, g_f_w will overwrite this under i_sem if it races
1134 * with a local truncate, it just makes our maxbyte checking easier */
1135 if (file->f_flags & O_APPEND)
1136 *ppos = inode->i_size;
1138 if (*ppos >= maxbytes) {
1139 if (count || *ppos > maxbytes) {
1140 send_sig(SIGXFSZ, current, 0);
1141 GOTO(out, retval = -EFBIG);
1144 if (*ppos + count > maxbytes)
1145 count = maxbytes - *ppos;
1147 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1148 inode->i_ino, count, *ppos);
1150 /* generic_file_write handles O_APPEND after getting i_sem */
1151 retval = generic_file_write(file, buf, count, ppos);
1154 ll_tree_unlock(&tree, inode);
1155 /* serialize with mmap/munmap/mremap */
1156 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
1157 retval > 0 ? retval : 0);
1161 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
1162 int flags, struct lov_user_md *lum,
1165 struct ll_inode_info *lli = ll_i2info(inode);
1167 struct obd_export *exp = ll_i2dtexp(inode);
1168 struct lov_stripe_md *lsm;
1169 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
1170 struct ptlrpc_request *req = NULL;
1172 struct lustre_md md;
1173 struct obd_client_handle *och;
1177 if ((file->f_flags+1) & O_ACCMODE)
1179 if (file->f_flags & O_TRUNC)
1182 down(&lli->lli_open_sem);
1185 up(&lli->lli_open_sem);
1186 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
1191 f = get_empty_filp();
1195 f->f_dentry = file->f_dentry;
1196 f->f_vfsmnt = file->f_vfsmnt;
1199 rc = ll_intent_alloc(&oit);
1203 rc = ll_intent_file_open(f, lum, lum_size, &oit);
1206 if (it_disposition(&oit, DISP_LOOKUP_NEG))
1209 req = LUSTRE_IT(&oit)->it_data;
1210 rc = LUSTRE_IT(&oit)->it_status;
1215 rc = mdc_req2lustre_md(ll_i2mdexp(inode), req, 1, exp, &md);
1218 ll_update_inode(f->f_dentry->d_inode, &md);
1220 OBD_ALLOC(och, sizeof(struct obd_client_handle));
1221 rc = ll_local_open(f, &oit, och);
1222 if (rc) { /* Actually ll_local_open cannot fail! */
1225 if (LUSTRE_IT(&oit)->it_lock_mode) {
1226 ldlm_lock_decref_and_cancel((struct lustre_handle *)
1227 &LUSTRE_IT(&oit)->it_lock_handle,
1228 LUSTRE_IT(&oit)->it_lock_mode);
1229 LUSTRE_IT(&oit)->it_lock_mode = 0;
1232 ll_intent_release(&oit);
1234 /* ll_file_release will decrease the count, but won't free anything
1235 because we have at least one more reference coming from actual open
1237 down(&lli->lli_och_sem);
1238 lli->lli_open_fd_write_count++;
1239 up(&lli->lli_och_sem);
1240 rc = ll_file_release(f->f_dentry->d_inode, f);
1242 /* Now also destroy our supplemental och */
1243 ll_md_och_close(ll_i2mdexp(inode), f->f_dentry->d_inode, och);
1246 ll_intent_release(&oit);
1249 up(&lli->lli_open_sem);
1251 ptlrpc_req_finished(req);
1255 static int ll_lov_setea(struct inode *inode, struct file *file,
1258 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
1259 struct lov_user_md *lump;
1260 int lum_size = sizeof(struct lov_user_md) +
1261 sizeof(struct lov_user_ost_data);
1265 if (!capable (CAP_SYS_ADMIN))
1268 OBD_ALLOC(lump, lum_size);
1272 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
1274 OBD_FREE(lump, lum_size);
1278 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1280 OBD_FREE(lump, lum_size);
1284 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1287 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1289 int flags = FMODE_WRITE;
1292 /* Bug 1152: copy properly when this is no longer true */
1293 LASSERT(sizeof(lum) == sizeof(*lump));
1294 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1295 rc = copy_from_user(&lum, lump, sizeof(lum));
1299 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1301 put_user(0, &lump->lmm_stripe_count);
1302 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
1303 0, ll_i2info(inode)->lli_smd, lump);
1308 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1310 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1315 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
1319 static int ll_get_grouplock(struct inode *inode, struct file *file,
1322 struct ll_file_data *fd = file->private_data;
1323 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1324 .end = OBD_OBJECT_EOF}};
1325 struct lustre_handle lockh = { 0 };
1326 struct ll_inode_info *lli = ll_i2info(inode);
1327 struct lov_stripe_md *lsm = lli->lli_smd;
1331 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1335 policy.l_extent.gid = arg;
1336 if (file->f_flags & O_NONBLOCK)
1337 flags = LDLM_FL_BLOCK_NOWAIT;
1339 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags,
1340 &ll_i2sbi(inode)->ll_grouplock_stime);
1344 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1346 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1351 static int ll_put_grouplock(struct inode *inode, struct file *file,
1354 struct ll_file_data *fd = file->private_data;
1355 struct ll_inode_info *lli = ll_i2info(inode);
1356 struct lov_stripe_md *lsm = lli->lli_smd;
1360 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1361 /* Ugh, it's already unlocked. */
1365 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1368 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1370 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1375 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1380 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1383 struct ll_file_data *fd = file->private_data;
1384 struct ll_sb_info *sbi = ll_i2sbi(inode);
1388 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1389 inode->i_generation, inode, cmd);
1391 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1394 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1396 case LL_IOC_GETFLAGS:
1397 /* Get the current value of the file flags */
1398 return put_user(fd->fd_flags, (int *)arg);
1399 case LL_IOC_SETFLAGS:
1400 case LL_IOC_CLRFLAGS:
1401 /* Set or clear specific file flags */
1402 /* XXX This probably needs checks to ensure the flags are
1403 * not abused, and to handle any flag side effects.
1405 if (get_user(flags, (int *) arg))
1408 if (cmd == LL_IOC_SETFLAGS)
1409 fd->fd_flags |= flags;
1411 fd->fd_flags &= ~flags;
1413 case LL_IOC_LOV_SETSTRIPE:
1414 RETURN(ll_lov_setstripe(inode, file, arg));
1415 case LL_IOC_LOV_SETEA:
1416 RETURN(ll_lov_setea(inode, file, arg));
1417 case IOC_MDC_SHOWFID: {
1418 struct lustre_id *idp = (struct lustre_id *)arg;
1419 struct lustre_id id;
1423 filename = getname((const char *)arg);
1424 if (IS_ERR(filename))
1425 RETURN(PTR_ERR(filename));
1427 ll_inode2id(&id, inode);
1429 rc = ll_get_fid(sbi->ll_md_exp, &id, filename, &id);
1431 GOTO(out_filename, rc);
1433 rc = copy_to_user(idp, &id, sizeof(*idp));
1435 GOTO(out_filename, rc = -EFAULT);
1442 case LL_IOC_LOV_GETSTRIPE:
1443 RETURN(ll_lov_getstripe(inode, arg));
1444 case EXT3_IOC_GETFLAGS:
1445 case EXT3_IOC_SETFLAGS:
1446 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1447 case LL_IOC_GROUP_LOCK:
1448 RETURN(ll_get_grouplock(inode, file, arg));
1449 case LL_IOC_GROUP_UNLOCK:
1450 RETURN(ll_put_grouplock(inode, file, arg));
1451 case EXT3_IOC_GETVERSION_OLD:
1452 case EXT3_IOC_GETVERSION:
1453 return put_user(inode->i_generation, (int *) arg);
1454 /* We need to special case any other ioctls we want to handle,
1455 * to send them to the MDS/OST as appropriate and to properly
1456 * network encode the arg field.
1457 case EXT2_IOC_GETVERSION_OLD:
1458 case EXT2_IOC_GETVERSION_NEW:
1459 case EXT2_IOC_SETVERSION_OLD:
1460 case EXT2_IOC_SETVERSION_NEW:
1461 case EXT3_IOC_SETVERSION_OLD:
1462 case EXT3_IOC_SETVERSION:
1465 RETURN( obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
1470 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1472 struct inode *inode = file->f_dentry->d_inode;
1473 struct ll_file_data *fd = file->private_data;
1474 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1475 struct lustre_handle lockh = {0};
1478 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1479 inode->i_generation, inode,
1480 offset + ((origin==2) ? inode->i_size : file->f_pos));
1482 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1483 if (origin == 2) { /* SEEK_END */
1484 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1485 struct ll_inode_info *lli = ll_i2info(inode);
1486 int nonblock = 0, rc;
1488 if (file->f_flags & O_NONBLOCK)
1489 nonblock = LDLM_FL_BLOCK_NOWAIT;
1491 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1492 nonblock, &ll_i2sbi(inode)->ll_seek_stime);
1496 down(&lli->lli_size_sem);
1497 offset += inode->i_size;
1498 up(&lli->lli_size_sem);
1499 } else if (origin == 1) { /* SEEK_CUR */
1500 offset += file->f_pos;
1504 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1505 if (offset != file->f_pos) {
1506 file->f_pos = offset;
1507 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1509 file->f_version = ++event;
1516 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1520 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1522 struct inode *inode = dentry->d_inode;
1523 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1524 struct lustre_id id;
1525 struct ptlrpc_request *req;
1528 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1529 inode->i_generation, inode);
1531 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1533 /* fsync's caller has already called _fdata{sync,write}, we want
1534 * that IO to finish before calling the osc and mdc sync methods */
1535 rc = filemap_fdatawait(inode->i_mapping);
1537 ll_inode2id(&id, inode);
1538 err = md_sync(ll_i2sbi(inode)->ll_md_exp, &id, &req);
1542 ptlrpc_req_finished(req);
1545 struct obdo *oa = obdo_alloc();
1548 RETURN(rc ? rc : -ENOMEM);
1550 oa->o_id = lsm->lsm_object_id;
1551 oa->o_gr = lsm->lsm_object_gr;
1552 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1554 obdo_from_inode(oa, inode, (OBD_MD_FLTYPE | OBD_MD_FLATIME |
1555 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1558 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
1568 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1570 struct inode *inode = file->f_dentry->d_inode;
1571 struct ll_inode_info *li = ll_i2info(inode);
1572 struct ll_sb_info *sbi = ll_i2sbi(inode);
1573 struct obd_device *obddev;
1574 struct ldlm_res_id res_id =
1575 { .name = {id_fid(&li->lli_id), id_group(&li->lli_id), LDLM_FLOCK} };
1576 struct lustre_handle lockh = {0};
1577 ldlm_policy_data_t flock;
1578 ldlm_mode_t mode = 0;
1583 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1584 inode->i_ino, file_lock);
1586 flock.l_flock.pid = file_lock->fl_pid;
1587 flock.l_flock.start = file_lock->fl_start;
1588 flock.l_flock.end = file_lock->fl_end;
1590 switch (file_lock->fl_type) {
1595 /* An unlock request may or may not have any relation to
1596 * existing locks so we may not be able to pass a lock handle
1597 * via a normal ldlm_lock_cancel() request. The request may even
1598 * unlock a byte range in the middle of an existing lock. In
1599 * order to process an unlock request we need all of the same
1600 * information that is given with a normal read or write record
1601 * lock request. To avoid creating another ldlm unlock (cancel)
1602 * message we'll treat a LCK_NL flock request as an unlock. */
1609 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1624 flags = LDLM_FL_BLOCK_NOWAIT;
1630 flags = LDLM_FL_TEST_LOCK;
1631 /* Save the old mode so that if the mode in the lock changes we
1632 * can decrement the appropriate reader or writer refcount. */
1633 file_lock->fl_type = mode;
1636 CERROR("unknown fcntl lock command: %d\n", cmd);
1640 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1641 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1642 flags, mode, flock.l_flock.start, flock.l_flock.end);
1644 obddev = md_get_real_obd(sbi->ll_md_exp, &li->lli_id);
1645 rc = ldlm_cli_enqueue(obddev->obd_self_export, NULL,
1646 obddev->obd_namespace,
1647 res_id, LDLM_FLOCK, &flock, mode, &flags,
1648 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1649 NULL, 0, NULL, &lockh);
1653 int ll_inode_revalidate_it(struct dentry *dentry)
1655 struct lookup_intent oit = { .it_op = IT_GETATTR };
1656 struct inode *inode = dentry->d_inode;
1657 struct ptlrpc_request *req = NULL;
1658 struct ll_inode_info *lli;
1659 struct lov_stripe_md *lsm;
1660 struct ll_sb_info *sbi;
1661 struct lustre_id id;
1666 CERROR("REPORT THIS LINE TO PETER\n");
1670 sbi = ll_i2sbi(inode);
1672 ll_inode2id(&id, inode);
1673 lli = ll_i2info(inode);
1674 LASSERT(id_fid(&id) != 0);
1676 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), name=%s(%p)\n",
1677 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1680 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1681 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1684 rc = ll_intent_alloc(&oit);
1688 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1689 &oit, 0, &req, ll_mdc_blocking_ast);
1693 rc = revalidate_it_finish(req, 1, &oit, dentry);
1698 ll_lookup_finish_locks(&oit, dentry);
1701 if (lsm == NULL) /* object not yet allocated, don't validate size */
1705 * ll_glimpse_size() will prefer locally cached writes if they extend
1708 rc = ll_glimpse_size(inode);
1711 ll_intent_release(&oit);
1713 ptlrpc_req_finished(req);
1717 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1718 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
1721 struct inode *inode = de->d_inode;
1722 struct ll_inode_info *lli = ll_i2info(inode);
1724 res = ll_inode_revalidate_it(de);
1725 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1730 stat->ino = inode->i_ino;
1731 stat->mode = inode->i_mode;
1732 stat->nlink = inode->i_nlink;
1733 stat->uid = inode->i_uid;
1734 stat->gid = inode->i_gid;
1735 stat->atime = inode->i_atime;
1736 stat->mtime = inode->i_mtime;
1737 stat->ctime = inode->i_ctime;
1738 stat->blksize = inode->i_blksize;
1740 down(&lli->lli_size_sem);
1741 stat->size = inode->i_size;
1742 stat->blocks = inode->i_blocks;
1743 up(&lli->lli_size_sem);
1745 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1746 stat->dev = id_group(&ll_i2info(inode)->lli_id);
1752 int ll_setxattr_internal(struct inode *inode, const char *name,
1753 const void *value, size_t size, int flags,
1756 struct ll_sb_info *sbi = ll_i2sbi(inode);
1757 struct ptlrpc_request *request = NULL;
1758 struct mdc_op_data op_data;
1763 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu\n", inode->i_ino);
1764 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_SETXATTR);
1766 memset(&attr, 0x0, sizeof(attr));
1767 attr.ia_valid |= valid;
1768 attr.ia_attr_flags = flags;
1770 ll_prepare_mdc_data(&op_data, inode, NULL, NULL, 0, 0);
1772 rc = md_setattr(sbi->ll_md_exp, &op_data, &attr,
1773 (void*) name, strnlen(name, XATTR_NAME_MAX)+1,
1774 (void*) value, size, &request);
1776 CERROR("md_setattr fails: rc = %d\n", rc);
1781 ptlrpc_req_finished(request);
1785 int ll_setxattr(struct dentry *dentry, const char *name, const void *value,
1786 size_t size, int flags)
1789 struct posix_acl *acl;
1790 struct ll_inode_info *lli;
1793 rc = ll_setxattr_internal(dentry->d_inode, name, value, size,
1796 /* update inode's acl info */
1797 if (rc == 0 && strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1799 acl = posix_acl_from_xattr(value, size);
1801 CERROR("convert from xattr to acl error: %ld",
1805 error = posix_acl_valid(acl);
1807 CERROR("acl valid error: %d", error);
1808 posix_acl_release(acl);
1816 lli = ll_i2info(dentry->d_inode);
1817 spin_lock(&lli->lli_lock);
1818 if (lli->lli_acl_access != NULL)
1819 posix_acl_release(lli->lli_acl_access);
1820 lli->lli_acl_access = acl;
1821 spin_unlock(&lli->lli_lock);
1828 int ll_removexattr(struct dentry *dentry, const char *name)
1830 return ll_setxattr_internal(dentry->d_inode, name, NULL, 0, 0,
1835 int ll_getxattr_internal(struct inode *inode, const char *name, int namelen,
1836 void *value, size_t size, __u64 valid)
1838 struct ptlrpc_request *request = NULL;
1839 struct ll_sb_info *sbi = ll_i2sbi(inode);
1840 struct lustre_id id;
1841 struct mds_body *body;
1846 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETXATTR);
1848 ll_inode2id(&id, inode);
1849 rc = md_getattr(sbi->ll_md_exp, &id, valid, name, namelen,
1852 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1853 CERROR("md_getattr fails: rc = %d\n", rc);
1857 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
1858 LASSERT(body != NULL);
1859 LASSERT_REPSWABBED(request, 0);
1861 ea_size = body->eadatasize;
1862 LASSERT(ea_size <= request->rq_repmsg->buflens[0]);
1865 GOTO(out, rc = ea_size);
1867 ea_data = lustre_msg_buf(request->rq_repmsg, 1, ea_size);
1868 LASSERT(ea_data != NULL);
1869 LASSERT_REPSWABBED(request, 1);
1872 memcpy(value, ea_data, ea_size);
1875 ptlrpc_req_finished(request);
1879 int ll_getxattr(struct dentry *dentry, const char *name, void *value,
1882 return ll_getxattr_internal(dentry->d_inode, name, strlen(name) + 1,
1883 value, size, OBD_MD_FLEA);
1886 int ll_listxattr(struct dentry *dentry, char *list, size_t size)
1888 return ll_getxattr_internal(dentry->d_inode, NULL, 0, list, size,
1893 * XXX We could choose not to check DLM lock. Leave the decision
1894 * to remote acl handling.
1897 lustre_check_acl(struct inode *inode, int mask)
1899 struct lookup_intent it = { .it_op = IT_GETATTR };
1900 struct dentry de = { .d_inode = inode };
1901 struct ll_sb_info *sbi;
1902 struct lustre_id id;
1903 struct ptlrpc_request *req = NULL;
1904 struct ll_inode_info *lli = ll_i2info(inode);
1905 struct posix_acl *acl;
1909 sbi = ll_i2sbi(inode);
1910 ll_inode2id(&id, inode);
1912 if (ll_intent_alloc(&it))
1915 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1916 &it, 0, &req, ll_mdc_blocking_ast);
1918 ll_intent_free(&it);
1922 rc = revalidate_it_finish(req, 1, &it, &de);
1924 ll_intent_release(&it);
1928 ll_lookup_finish_locks(&it, &de);
1929 ll_intent_free(&it);
1931 spin_lock(&lli->lli_lock);
1932 acl = posix_acl_dup(ll_i2info(inode)->lli_acl_access);
1933 spin_unlock(&lli->lli_lock);
1936 GOTO(out, rc = -EAGAIN);
1938 rc = posix_acl_permission(inode, acl, mask);
1939 posix_acl_release(acl);
1943 ptlrpc_req_finished(req);
1948 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
1950 return generic_permission(inode, mask, lustre_check_acl);
1953 struct file_operations ll_file_operations = {
1954 .read = ll_file_read,
1955 .write = ll_file_write,
1956 .ioctl = ll_file_ioctl,
1957 .open = ll_file_open,
1958 .release = ll_file_release,
1959 .mmap = ll_file_mmap,
1960 .llseek = ll_file_seek,
1961 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
1962 .sendfile = generic_file_sendfile,
1965 .lock = ll_file_flock
1968 struct inode_operations ll_file_inode_operations = {
1969 .setattr = ll_setattr,
1970 .truncate = ll_truncate,
1971 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1972 .getattr = ll_getattr,
1974 .revalidate_it = ll_inode_revalidate_it,
1976 .setxattr = ll_setxattr,
1977 .getxattr = ll_getxattr,
1978 .listxattr = ll_listxattr,
1979 .removexattr = ll_removexattr,
1980 .permission = ll_inode_permission,