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 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
31 #include <linux/lustre_compat25.h>
34 #include "llite_internal.h"
36 static int ll_mdc_close(struct obd_export *mdc_exp, struct inode *inode,
39 struct ll_file_data *fd = file->private_data;
40 struct ptlrpc_request *req = NULL;
41 struct obd_client_handle *och = &fd->fd_mds_och;
48 memset(&obdo, 0, sizeof(obdo));
49 obdo.o_id = inode->i_ino;
50 obdo.o_mode = inode->i_mode;
51 obdo.o_size = inode->i_size;
52 obdo.o_blocks = inode->i_blocks;
53 if (0 /* ll_is_inode_dirty(inode) */) {
54 obdo.o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
55 valid |= OBD_MD_FLFLAGS;
58 rc = mdc_close(mdc_exp, &obdo, och, &req);
60 /* We are the last writer, so the MDS has instructed us to get
61 * the file size and any write cookies, then close again. */
62 //ll_queue_done_writing(inode);
65 CERROR("inode %lu mdc close failed: rc = %d\n",
69 rc = ll_objects_destroy(req, file->f_dentry->d_inode);
71 CERROR("inode %lu ll_objects destroy: rc = %d\n",
75 mdc_clear_open_replay_data(och);
76 ptlrpc_req_finished(req);
77 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
78 file->private_data = NULL;
79 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof *fd);
84 /* While this returns an error code, fput() the caller does not, so we need
85 * to make every effort to clean up all of our state here. Also, applications
86 * rarely check close errors and even if an error is returned they will not
87 * re-try the close call.
89 int ll_file_release(struct inode *inode, struct file *file)
91 struct ll_file_data *fd;
92 struct ll_sb_info *sbi = ll_i2sbi(inode);
96 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
97 inode->i_generation, inode);
99 /* don't do anything for / */
100 if (inode->i_sb->s_root == file->f_dentry)
103 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
104 fd = (struct ll_file_data *)file->private_data;
107 rc = ll_mdc_close(sbi->ll_mdc_exp, inode, file);
111 static int ll_intent_file_open(struct file *file, void *lmm,
112 int lmmsize, struct lookup_intent *itp)
114 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
115 struct lustre_handle lockh;
116 struct mdc_op_data data;
117 struct dentry *parent = file->f_dentry->d_parent;
118 const char *name = file->f_dentry->d_name.name;
119 const int len = file->f_dentry->d_name.len;
125 ll_prepare_mdc_op_data(&data, parent->d_inode, NULL, name, len, O_RDWR);
127 rc = mdc_enqueue(sbi->ll_mdc_exp, LDLM_PLAIN, itp, LCK_PR, &data,
128 &lockh, lmm, lmmsize, ldlm_completion_ast,
129 ll_mdc_blocking_ast, parent->d_inode);
131 CERROR("lock enqueue: err: %d\n", rc);
135 static int ll_local_open(struct file *file, struct lookup_intent *it)
137 struct ptlrpc_request *req = it->d.lustre.it_data;
138 struct ll_inode_info *lli = ll_i2info(file->f_dentry->d_inode);
139 struct ll_file_data *fd;
140 struct mds_body *body;
143 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
144 LASSERT (body != NULL); /* reply already checked out */
145 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
147 LASSERT(!file->private_data);
149 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
150 /* We can't handle this well without reorganizing ll_file_open and
151 * ll_mdc_close, so don't even try right now. */
154 memcpy(&fd->fd_mds_och.och_fh, &body->handle, sizeof(body->handle));
155 fd->fd_mds_och.och_magic = OBD_CLIENT_HANDLE_MAGIC;
156 file->private_data = fd;
157 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
159 lli->lli_io_epoch = body->io_epoch;
161 mdc_set_open_replay_data(&fd->fd_mds_och, it->d.lustre.it_data);
166 /* Open a file, and (for the very first open) create objects on the OSTs at
167 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
168 * creation or open until ll_lov_setstripe() ioctl is called. We grab
169 * lli_open_sem to ensure no other process will create objects, send the
170 * stripe MD to the MDS, or try to destroy the objects if that fails.
172 * If we already have the stripe MD locally then we don't request it in
173 * mdc_open(), by passing a lmm_size = 0.
175 * It is up to the application to ensure no other processes open this file
176 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
177 * used. We might be able to avoid races of that sort by getting lli_open_sem
178 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
179 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
181 int ll_file_open(struct inode *inode, struct file *file)
183 struct ll_inode_info *lli = ll_i2info(inode);
184 struct lookup_intent *it;
185 struct lov_stripe_md *lsm;
186 struct ptlrpc_request *req;
190 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
191 inode->i_generation, inode);
193 /* don't do anything for / */
194 if (inode->i_sb->s_root == file->f_dentry)
199 if (!it->d.lustre.it_disposition) {
200 struct lookup_intent oit = { .it_op = IT_OPEN,
201 .it_flags = file->f_flags };
203 rc = ll_intent_file_open(file, NULL, 0, it);
208 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
209 rc = it_open_error(DISP_OPEN_OPEN, it);
213 rc = ll_local_open(file, it);
217 if (!S_ISREG(inode->i_mode))
222 if (file->f_flags & O_LOV_DELAY_CREATE ||
223 !(file->f_mode & FMODE_WRITE)) {
224 CDEBUG(D_INODE, "object creation was delayed\n");
228 file->f_flags &= ~O_LOV_DELAY_CREATE;
231 req = it->d.lustre.it_data;
232 ptlrpc_req_finished(req);
234 ll_open_complete(inode);
238 /* Fills the obdo with the attributes for the inode defined by lsm */
239 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
242 struct ptlrpc_request_set *set;
246 LASSERT(lsm != NULL);
248 memset(oa, 0, sizeof *oa);
249 oa->o_id = lsm->lsm_object_id;
250 oa->o_mode = S_IFREG;
251 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
252 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
255 set = ptlrpc_prep_set();
257 CERROR ("ENOMEM allocing request set\n");
260 rc = obd_getattr_async(exp, oa, lsm, set);
262 rc = ptlrpc_set_wait(set);
263 ptlrpc_set_destroy(set);
268 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
269 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
273 static inline void ll_remove_suid(struct inode *inode)
277 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
278 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
280 /* was any of the uid bits set? */
281 mode &= inode->i_mode;
282 if (mode && !capable(CAP_FSETID)) {
283 inode->i_mode &= ~mode;
284 // XXX careful here - we cannot change the size
288 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
290 struct ll_inode_info *lli = ll_i2info(inode);
291 struct lov_stripe_md *lsm = lli->lli_smd;
292 struct obd_export *exp = ll_i2obdexp(inode);
295 struct ldlm_lock *lock;
296 struct lov_stripe_md *lsm;
297 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
298 __u32 stripe, vallen = sizeof(stripe);
302 if (lsm->lsm_stripe_count == 1)
305 /* get our offset in the lov */
306 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
308 CERROR("obd_get_info: rc = %d\n", rc);
311 LASSERT(stripe < lsm->lsm_stripe_count);
315 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
316 * we get a lock cancellation for each stripe, so we have to map the obd's
317 * region back onto the stripes in the file that it held.
319 * No one can dirty the extent until we've finished our work and they can
320 * enqueue another lock. The DLM protects us from ll_file_read/write here,
321 * but other kernel actors could have pages locked.
323 * Called with the DLM lock held. */
324 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
325 struct ldlm_lock *lock, __u32 stripe)
327 ldlm_policy_data_t tmpex;
328 unsigned long start, end, count, skip, i, j;
330 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
331 struct lustre_handle lockh;
334 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
335 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
336 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
339 /* our locks are page granular thanks to osc_enqueue, we invalidate the
341 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
342 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
346 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
347 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
348 if (lsm->lsm_stripe_count > 1) {
349 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
350 skip = (lsm->lsm_stripe_count - 1) * count;
351 start += start/count * skip + stripe * count;
353 end += end/count * skip + stripe * count;
355 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
358 i = (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
362 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
363 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
364 count, skip, end, discard ? " (DISCARDING)" : "");
366 /* this is the simplistic implementation of page eviction at
367 * cancelation. It is careful to get races with other page
368 * lockers handled correctly. fixes from bug 20 will make it
369 * more efficient by associating locks with pages and with
370 * batching writeback under the lock explicitly. */
371 for (i = start, j = start % count; i <= end;
372 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
374 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
380 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
381 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
382 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
385 ll_pgcache_lock(inode->i_mapping);
386 if (list_empty(&inode->i_mapping->dirty_pages) &&
387 list_empty(&inode->i_mapping->clean_pages) &&
388 list_empty(&inode->i_mapping->locked_pages)) {
389 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
390 ll_pgcache_unlock(inode->i_mapping);
393 ll_pgcache_unlock(inode->i_mapping);
395 conditional_schedule();
397 page = find_get_page(inode->i_mapping, i);
400 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
401 i, tmpex.l_extent.start);
404 /* page->mapping to check with racing against teardown */
405 if (page->mapping && PageDirty(page) && !discard) {
406 ClearPageDirty(page);
407 LL_CDEBUG_PAGE(D_PAGE, page, "found dirty\n");
408 ll_pgcache_lock(inode->i_mapping);
409 list_del(&page->list);
410 list_add(&page->list, &inode->i_mapping->locked_pages);
411 ll_pgcache_unlock(inode->i_mapping);
413 rc = ll_call_writepage(inode, page);
415 CERROR("writepage of page %p failed: %d\n",
417 /* either waiting for io to complete or reacquiring
418 * the lock that the failed writepage released */
422 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
423 /* check to see if another DLM lock covers this page */
424 ldlm_lock2handle(lock, &lockh);
425 rc2 = ldlm_lock_match(NULL,
426 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
428 NULL, 0, &tmpex, 0, &lockh);
429 if (rc2 == 0 && page->mapping != NULL) {
430 // checking again to account for writeback's lock_page()
431 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
432 ll_truncate_complete_page(page);
435 page_cache_release(page);
437 LASSERTF(tmpex.l_extent.start <=
438 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
439 lock->l_policy_data.l_extent.end + 1),
440 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
441 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
446 static int ll_extent_lock_callback(struct ldlm_lock *lock,
447 struct ldlm_lock_desc *new, void *data,
450 struct lustre_handle lockh = { 0 };
454 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
455 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
460 case LDLM_CB_BLOCKING:
461 ldlm_lock2handle(lock, &lockh);
462 rc = ldlm_cli_cancel(&lockh);
464 CERROR("ldlm_cli_cancel failed: %d\n", rc);
466 case LDLM_CB_CANCELING: {
468 struct ll_inode_info *lli;
469 struct lov_stripe_md *lsm;
473 /* This lock wasn't granted, don't try to evict pages */
474 if (lock->l_req_mode != lock->l_granted_mode)
477 inode = ll_inode_from_lock(lock);
480 lli = ll_i2info(inode);
483 if (lli->lli_smd == NULL)
487 stripe = ll_lock_to_stripe_offset(inode, lock);
488 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
491 kms = ldlm_extent_shift_kms(lock,
492 lsm->lsm_oinfo[stripe].loi_kms);
493 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
494 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
495 lsm->lsm_oinfo[stripe].loi_kms, kms);
496 lsm->lsm_oinfo[stripe].loi_kms = kms;
498 //ll_try_done_writing(inode);
511 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
513 /* XXX ALLOCATE - 160 bytes */
514 struct inode *inode = ll_inode_from_lock(lock);
515 struct ll_inode_info *lli = ll_i2info(inode);
516 struct lustre_handle lockh = { 0 };
521 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
522 LDLM_FL_BLOCK_CONV)) {
523 LBUG(); /* not expecting any blocked async locks yet */
524 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
526 ldlm_lock_dump(D_OTHER, lock, 0);
527 ldlm_reprocess_all(lock->l_resource);
531 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
533 stripe = ll_lock_to_stripe_offset(inode, lock);
535 if (lock->l_lvb_len) {
536 struct lov_stripe_md *lsm = lli->lli_smd;
538 lvb = lock->l_lvb_data;
539 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
542 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
543 kms = ldlm_extent_shift_kms(NULL, kms);
544 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
545 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
546 lsm->lsm_oinfo[stripe].loi_kms, kms);
547 lsm->lsm_oinfo[stripe].loi_kms = kms;
552 wake_up(&lock->l_waitq);
554 ldlm_lock2handle(lock, &lockh);
555 ldlm_lock_decref(&lockh, LCK_PR);
560 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
562 struct ptlrpc_request *req = reqp;
563 struct inode *inode = ll_inode_from_lock(lock);
564 struct ll_inode_info *lli;
566 int rc, size = sizeof(*lvb), stripe = 0;
570 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
571 lli = ll_i2info(inode);
573 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
574 if (lli->lli_smd == NULL)
575 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
577 /* First, find out which stripe index this lock corresponds to. */
578 if (lli->lli_smd->lsm_stripe_count > 1)
579 stripe = ll_lock_to_stripe_offset(inode, lock);
581 rc = lustre_pack_reply(req, 1, &size, NULL);
583 CERROR("lustre_pack_reply: %d\n", rc);
587 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
588 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
590 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
591 inode->i_size, stripe, lvb->lvb_size);
597 /* These errors are normal races, so we don't want to fill the console
598 * with messages by calling ptlrpc_error() */
599 if (rc == -ELDLM_NO_LOCK_DATA)
600 lustre_pack_reply(req, 0, NULL, NULL);
606 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
607 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
609 /* NB: lov_merge_size will prefer locally cached writes if they extend the
610 * file (because it prefers KMS over RSS when larger) */
611 int ll_glimpse_size(struct inode *inode, struct ost_lvb *lvb)
613 struct ll_inode_info *lli = ll_i2info(inode);
614 struct ll_sb_info *sbi = ll_i2sbi(inode);
615 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
616 struct lustre_handle lockh;
617 int rc, flags = LDLM_FL_HAS_INTENT;
620 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
622 rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
623 LCK_PR, &flags, ll_extent_lock_callback,
624 ldlm_completion_ast, ll_glimpse_callback, inode,
625 sizeof(*lvb), lustre_swab_ost_lvb, &lockh);
627 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
628 RETURN(rc > 0 ? -EIO : rc);
631 lvb->lvb_size = lov_merge_size(lli->lli_smd, 0);
632 //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
634 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64"\n", lvb->lvb_size);
636 obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
641 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
642 struct lov_stripe_md *lsm, int mode,
643 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
646 struct ll_sb_info *sbi = ll_i2sbi(inode);
650 LASSERT(lockh->cookie == 0);
652 /* XXX phil: can we do this? won't it screw the file size up? */
653 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
654 (sbi->ll_flags & LL_SBI_NOLCK))
657 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
658 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
660 rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
661 &ast_flags, ll_extent_lock_callback,
662 ldlm_completion_ast, ll_glimpse_callback, inode,
663 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
667 if (policy->l_extent.start == 0 &&
668 policy->l_extent.end == OBD_OBJECT_EOF)
669 inode->i_size = lov_merge_size(lsm, 1);
671 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
676 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
677 struct lov_stripe_md *lsm, int mode,
678 struct lustre_handle *lockh)
680 struct ll_sb_info *sbi = ll_i2sbi(inode);
684 /* XXX phil: can we do this? won't it screw the file size up? */
685 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
686 (sbi->ll_flags & LL_SBI_NOLCK))
689 rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
694 static ssize_t ll_file_read(struct file *filp, char *buf, size_t count,
697 struct ll_file_data *fd = filp->private_data;
698 struct inode *inode = filp->f_dentry->d_inode;
699 struct ll_inode_info *lli = ll_i2info(inode);
700 struct lov_stripe_md *lsm = lli->lli_smd;
701 struct lustre_handle lockh = { 0 };
702 ldlm_policy_data_t policy;
707 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
708 inode->i_ino, inode->i_generation, inode, count, *ppos);
710 /* "If nbyte is 0, read() will return 0 and have no other results."
711 * -- Single Unix Spec */
715 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
721 policy.l_extent.start = *ppos;
722 policy.l_extent.end = *ppos + count - 1;
724 err = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh, 0);
728 kms = lov_merge_size(lsm, 1);
729 if (policy.l_extent.end > kms) {
730 /* A glimpse is necessary to determine whether we return a short
731 * read or some zeroes at the end of the buffer */
733 retval = ll_glimpse_size(inode, &lvb);
736 inode->i_size = lvb.lvb_size;
741 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
742 inode->i_ino, count, *ppos, inode->i_size);
744 /* turn off the kernel's read-ahead */
745 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
748 filp->f_ra.ra_pages = 0;
750 retval = generic_file_read(filp, buf, count, ppos);
753 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
758 * Write to a file (through the page cache).
760 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
763 struct ll_file_data *fd = file->private_data;
764 struct inode *inode = file->f_dentry->d_inode;
765 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
766 struct lustre_handle lockh = { 0 };
767 ldlm_policy_data_t policy;
768 loff_t maxbytes = ll_file_maxbytes(inode);
772 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
773 inode->i_ino, inode->i_generation, inode, count, *ppos);
775 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
777 /* POSIX, but surprised the VFS doesn't check this already */
781 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
782 * called on the file, don't fail the below assertion (bug 2388). */
783 if (file->f_flags & O_LOV_DELAY_CREATE && lsm == NULL)
788 if (file->f_flags & O_APPEND) {
789 policy.l_extent.start = 0;
790 policy.l_extent.end = OBD_OBJECT_EOF;
792 policy.l_extent.start = *ppos;
793 policy.l_extent.end = *ppos + count - 1;
796 err = ll_extent_lock(fd, inode, lsm, LCK_PW, &policy, &lockh, 0);
800 /* this is ok, g_f_w will overwrite this under i_sem if it races
801 * with a local truncate, it just makes our maxbyte checking easier */
802 if (file->f_flags & O_APPEND)
803 *ppos = inode->i_size;
805 if (*ppos >= maxbytes) {
806 if (count || *ppos > maxbytes) {
807 send_sig(SIGXFSZ, current, 0);
808 GOTO(out, retval = -EFBIG);
811 if (*ppos + count > maxbytes)
812 count = maxbytes - *ppos;
814 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
815 inode->i_ino, count, *ppos);
817 /* generic_file_write handles O_APPEND after getting i_sem */
818 retval = generic_file_write(file, buf, count, ppos);
821 ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
822 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
823 retval > 0 ? retval : 0);
827 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
830 struct ll_inode_info *lli = ll_i2info(inode);
831 struct obd_export *exp = ll_i2obdexp(inode);
832 struct ll_recreate_obj ucreatp;
833 struct obd_trans_info oti = { 0 };
834 struct obdo *oa = NULL;
837 struct lov_stripe_md *lsm, *lsm2;
840 if (!capable (CAP_SYS_ADMIN))
843 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
844 sizeof(struct ll_recreate_obj));
853 down(&lli->lli_open_sem);
856 up(&lli->lli_open_sem);
860 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
861 (lsm->lsm_stripe_count));
863 OBD_ALLOC(lsm2, lsm_size);
865 up(&lli->lli_open_sem);
870 oa->o_id = ucreatp.lrc_id;
871 oa->o_nlink = ucreatp.lrc_ost_idx;
872 oa->o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS;
873 oa->o_flags |= OBD_FL_RECREATE_OBJS;
874 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
875 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
877 oti.oti_objid = NULL;
878 memcpy(lsm2, lsm, lsm_size);
879 rc = obd_create(exp, oa, &lsm2, &oti);
881 up(&lli->lli_open_sem);
882 OBD_FREE(lsm2, lsm_size);
887 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
888 int flags, struct lov_user_md *lum,
891 struct ll_inode_info *lli = ll_i2info(inode);
893 struct obd_export *exp = ll_i2obdexp(inode);
894 struct lov_stripe_md *lsm;
895 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
896 struct ptlrpc_request *req = NULL;
901 down(&lli->lli_open_sem);
904 up(&lli->lli_open_sem);
905 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
910 f = get_empty_filp();
914 f->f_dentry = file->f_dentry;
915 f->f_vfsmnt = file->f_vfsmnt;
917 rc = ll_intent_file_open(f, lum, lum_size, &oit);
920 if (it_disposition(&oit, DISP_LOOKUP_NEG))
922 req = oit.d.lustre.it_data;
923 rc = oit.d.lustre.it_status;
928 rc = mdc_req2lustre_md(req, 1, exp, &md);
931 ll_update_inode(f->f_dentry->d_inode, md.body, md.lsm);
933 rc = ll_local_open(f, &oit);
936 ll_intent_release(&oit);
938 rc = ll_file_release(f->f_dentry->d_inode, f);
943 up(&lli->lli_open_sem);
945 ptlrpc_req_finished(req);
949 static int ll_lov_setea(struct inode *inode, struct file *file,
952 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
953 struct lov_user_md *lump;
954 int lum_size = sizeof(struct lov_user_md) +
955 sizeof(struct lov_user_ost_data);
959 if (!capable (CAP_SYS_ADMIN))
962 OBD_ALLOC(lump, lum_size);
966 rc = copy_from_user(lump, (struct lov_user_md *)arg,
969 OBD_FREE(lump, lum_size);
973 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
975 OBD_FREE(lump, lum_size);
979 static int ll_lov_setstripe(struct inode *inode, struct file *file,
982 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
984 int flags = FMODE_WRITE;
987 /* Bug 1152: copy properly when this is no longer true */
988 LASSERT(sizeof(lum) == sizeof(*lump));
989 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
990 rc = copy_from_user(&lum, lump, sizeof(lum));
994 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
998 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1000 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1005 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2obdexp(inode), 0, lsm,
1009 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1012 struct ll_file_data *fd = file->private_data;
1016 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1017 inode->i_generation, inode, cmd);
1019 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1022 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1024 case LL_IOC_GETFLAGS:
1025 /* Get the current value of the file flags */
1026 return put_user(fd->fd_flags, (int *)arg);
1027 case LL_IOC_SETFLAGS:
1028 case LL_IOC_CLRFLAGS:
1029 /* Set or clear specific file flags */
1030 /* XXX This probably needs checks to ensure the flags are
1031 * not abused, and to handle any flag side effects.
1033 if (get_user(flags, (int *) arg))
1036 if (cmd == LL_IOC_SETFLAGS)
1037 fd->fd_flags |= flags;
1039 fd->fd_flags &= ~flags;
1041 case LL_IOC_LOV_SETSTRIPE:
1042 RETURN(ll_lov_setstripe(inode, file, arg));
1043 case LL_IOC_LOV_SETEA:
1044 RETURN( ll_lov_setea(inode, file, arg) );
1045 case LL_IOC_LOV_GETSTRIPE:
1046 RETURN(ll_lov_getstripe(inode, arg));
1047 case LL_IOC_RECREATE_OBJ:
1048 RETURN(ll_lov_recreate_obj(inode, file, arg));
1049 case EXT3_IOC_GETFLAGS:
1050 case EXT3_IOC_SETFLAGS:
1051 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1052 /* We need to special case any other ioctls we want to handle,
1053 * to send them to the MDS/OST as appropriate and to properly
1054 * network encode the arg field.
1055 case EXT2_IOC_GETVERSION_OLD:
1056 case EXT2_IOC_GETVERSION_NEW:
1057 case EXT2_IOC_SETVERSION_OLD:
1058 case EXT2_IOC_SETVERSION_NEW:
1061 RETURN( obd_iocontrol(cmd, ll_i2obdexp(inode), 0, NULL,
1066 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1068 struct inode *inode = file->f_dentry->d_inode;
1069 struct ll_file_data *fd = file->private_data;
1070 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1071 struct lustre_handle lockh = {0};
1074 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1075 inode->i_generation, inode,
1076 offset + ((origin==2) ? inode->i_size : file->f_pos));
1078 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1079 if (origin == 2) { /* SEEK_END */
1081 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1082 err = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,0);
1083 if (err != ELDLM_OK)
1086 offset += inode->i_size;
1087 } else if (origin == 1) { /* SEEK_CUR */
1088 offset += file->f_pos;
1092 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1093 if (offset != file->f_pos) {
1094 file->f_pos = offset;
1095 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1097 file->f_version = ++event;
1104 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1108 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1110 struct inode *inode = dentry->d_inode;
1111 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1113 struct ptlrpc_request *req;
1116 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1117 inode->i_generation, inode);
1119 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1121 /* fsync's caller has already called _fdata{sync,write}, we want
1122 * that IO to finish before calling the osc and mdc sync methods */
1123 rc = filemap_fdatawait(inode->i_mapping);
1125 ll_inode2fid(&fid, inode);
1126 err = mdc_sync(ll_i2sbi(inode)->ll_mdc_exp, &fid, &req);
1130 ptlrpc_req_finished(req);
1133 struct obdo *oa = obdo_alloc();
1136 RETURN(rc ? rc : -ENOMEM);
1138 oa->o_id = lsm->lsm_object_id;
1139 oa->o_valid = OBD_MD_FLID;
1140 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
1141 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
1143 err = obd_sync(ll_i2sbi(inode)->ll_osc_exp, oa, lsm,
1153 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1155 struct inode *inode = file->f_dentry->d_inode;
1156 struct ll_sb_info *sbi = ll_i2sbi(inode);
1157 struct obd_device *obddev;
1158 struct ldlm_res_id res_id =
1159 { .name = {inode->i_ino, inode->i_generation, LDLM_FLOCK} };
1160 struct lustre_handle lockh = {0};
1161 ldlm_policy_data_t flock;
1162 ldlm_mode_t mode = 0;
1167 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1168 inode->i_ino, file_lock);
1170 flock.l_flock.pid = file_lock->fl_pid;
1171 flock.l_flock.start = file_lock->fl_start;
1172 flock.l_flock.end = file_lock->fl_end;
1174 switch (file_lock->fl_type) {
1179 /* An unlock request may or may not have any relation to
1180 * existing locks so we may not be able to pass a lock handle
1181 * via a normal ldlm_lock_cancel() request. The request may even
1182 * unlock a byte range in the middle of an existing lock. In
1183 * order to process an unlock request we need all of the same
1184 * information that is given with a normal read or write record
1185 * lock request. To avoid creating another ldlm unlock (cancel)
1186 * message we'll treat a LCK_NL flock request as an unlock. */
1193 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1202 flags = LDLM_FL_BLOCK_NOWAIT;
1205 flags = LDLM_FL_TEST_LOCK;
1206 /* Save the old mode so that if the mode in the lock changes we
1207 * can decrement the appropriate reader or writer refcount. */
1208 file_lock->fl_type = mode;
1211 CERROR("unknown fcntl lock command: %d\n", cmd);
1215 CDEBUG(D_DLMTRACE, "inode=%lu, pid=%u, flags=%#x, mode=%u, "
1216 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1217 flags, mode, flock.l_flock.start, flock.l_flock.end);
1219 obddev = sbi->ll_mdc_exp->exp_obd;
1220 rc = ldlm_cli_enqueue(sbi->ll_mdc_exp, NULL, obddev->obd_namespace,
1221 res_id, LDLM_FLOCK, &flock, mode, &flags,
1222 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1223 NULL, 0, NULL, &lockh);
1227 static int ll_have_md_lock(struct dentry *de)
1229 struct ll_sb_info *sbi = ll_s2sbi(de->d_sb);
1230 struct lustre_handle lockh;
1231 struct ldlm_res_id res_id = { .name = {0} };
1232 struct obd_device *obddev;
1239 obddev = sbi->ll_mdc_exp->exp_obd;
1240 res_id.name[0] = de->d_inode->i_ino;
1241 res_id.name[1] = de->d_inode->i_generation;
1243 CDEBUG(D_INFO, "trying to match res "LPU64"\n", res_id.name[0]);
1245 /* FIXME use LDLM_FL_TEST_LOCK instead */
1246 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1247 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
1248 NULL, LCK_PR, &lockh)) {
1249 ldlm_lock_decref(&lockh, LCK_PR);
1253 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
1254 NULL, LCK_PW, &lockh)) {
1255 ldlm_lock_decref(&lockh, LCK_PW);
1261 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
1263 struct inode *inode = dentry->d_inode;
1264 struct ll_inode_info *lli;
1265 struct lov_stripe_md *lsm;
1270 CERROR("REPORT THIS LINE TO PETER\n");
1273 lli = ll_i2info(inode);
1274 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s\n",
1275 inode->i_ino, inode->i_generation, inode, dentry->d_name.name);
1276 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1277 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1280 if (!ll_have_md_lock(dentry)) {
1281 struct ptlrpc_request *req = NULL;
1282 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
1284 unsigned long valid = 0;
1287 if (S_ISREG(inode->i_mode)) {
1288 ealen = obd_size_diskmd(sbi->ll_osc_exp, NULL);
1289 valid |= OBD_MD_FLEASIZE;
1291 ll_inode2fid(&fid, inode);
1292 rc = mdc_getattr(sbi->ll_mdc_exp, &fid, valid, ealen, &req);
1294 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
1297 rc = ll_prep_inode(sbi->ll_osc_exp, &inode, req, 0, NULL);
1299 ptlrpc_req_finished(req);
1302 ptlrpc_req_finished(req);
1306 if (lsm == NULL) /* object not yet allocated, don't validate size */
1309 /* ll_glimpse_size will prefer locally cached writes if they extend
1314 rc = ll_glimpse_size(inode, &lvb);
1315 inode->i_size = lvb.lvb_size;
1320 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1321 int ll_getattr(struct vfsmount *mnt, struct dentry *de,
1322 struct lookup_intent *it, struct kstat *stat)
1325 struct inode *inode = de->d_inode;
1327 res = ll_inode_revalidate_it(de, it);
1328 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1333 stat->dev = inode->i_sb->s_dev;
1334 stat->ino = inode->i_ino;
1335 stat->mode = inode->i_mode;
1336 stat->nlink = inode->i_nlink;
1337 stat->uid = inode->i_uid;
1338 stat->gid = inode->i_gid;
1339 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1340 stat->atime = inode->i_atime;
1341 stat->mtime = inode->i_mtime;
1342 stat->ctime = inode->i_ctime;
1343 stat->size = inode->i_size;
1344 stat->blksize = inode->i_blksize;
1345 stat->blocks = inode->i_blocks;
1350 struct file_operations ll_file_operations = {
1352 write: ll_file_write,
1353 ioctl: ll_file_ioctl,
1355 release: ll_file_release,
1356 mmap: generic_file_mmap,
1357 llseek: ll_file_seek,
1359 //lock: ll_file_flock
1362 struct inode_operations ll_file_inode_operations = {
1363 setattr_raw: ll_setattr_raw,
1364 setattr: ll_setattr,
1365 truncate: ll_truncate,
1366 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1367 getattr_it: ll_getattr,
1369 revalidate_it: ll_inode_revalidate_it,
1373 struct inode_operations ll_special_inode_operations = {
1374 setattr_raw: ll_setattr_raw,
1375 setattr: ll_setattr,
1376 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1377 getattr_it: ll_getattr,
1379 revalidate_it: ll_inode_revalidate_it,