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) {
373 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
374 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
375 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
378 ll_pgcache_lock(inode->i_mapping);
379 if (list_empty(&inode->i_mapping->dirty_pages) &&
380 list_empty(&inode->i_mapping->clean_pages) &&
381 list_empty(&inode->i_mapping->locked_pages)) {
382 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
383 ll_pgcache_unlock(inode->i_mapping);
386 ll_pgcache_unlock(inode->i_mapping);
388 conditional_schedule();
390 page = find_get_page(inode->i_mapping, i);
393 LL_CDEBUG_PAGE(D_PAGE, page, "locking page\n");
396 /* page->mapping to check with racing against teardown */
397 if (page->mapping && PageDirty(page) && !discard) {
398 ClearPageDirty(page);
399 LL_CDEBUG_PAGE(D_PAGE, page, "found dirty\n");
400 ll_pgcache_lock(inode->i_mapping);
401 list_del(&page->list);
402 list_add(&page->list, &inode->i_mapping->locked_pages);
403 ll_pgcache_unlock(inode->i_mapping);
405 rc = ll_call_writepage(inode, page);
407 CERROR("writepage of page %p failed: %d\n",
409 /* either waiting for io to complete or reacquiring
410 * the lock that the failed writepage released */
414 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
415 /* check to see if another DLM lock covers this page */
416 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
417 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
419 &lock->l_resource->lr_name, LDLM_EXTENT,
420 &tmpex, LCK_PR | LCK_PW, &lockh);
421 if (rc2 == 0 && page->mapping != NULL) {
422 // checking again to account for writeback's lock_page()
423 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
424 ll_truncate_complete_page(page);
427 page_cache_release(page);
438 static int ll_extent_lock_callback(struct ldlm_lock *lock,
439 struct ldlm_lock_desc *new, void *data,
442 struct lustre_handle lockh = { 0 };
446 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
447 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
452 case LDLM_CB_BLOCKING:
453 ldlm_lock2handle(lock, &lockh);
454 rc = ldlm_cli_cancel(&lockh);
456 CERROR("ldlm_cli_cancel failed: %d\n", rc);
458 case LDLM_CB_CANCELING: {
460 struct ll_inode_info *lli;
461 struct lov_stripe_md *lsm;
465 /* This lock wasn't granted, don't try to evict pages */
466 if (lock->l_req_mode != lock->l_granted_mode)
469 inode = ll_inode_from_lock(lock);
472 lli = ll_i2info(inode);
475 if (lli->lli_smd == NULL)
479 stripe = ll_lock_to_stripe_offset(inode, lock);
480 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
483 kms = ldlm_extent_shift_kms(lock,
484 lsm->lsm_oinfo[stripe].loi_kms);
485 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
486 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
487 lsm->lsm_oinfo[stripe].loi_kms, kms);
488 lsm->lsm_oinfo[stripe].loi_kms = kms;
490 //ll_try_done_writing(inode);
503 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
505 /* XXX ALLOCATE - 160 bytes */
506 struct inode *inode = ll_inode_from_lock(lock);
507 struct ll_inode_info *lli = ll_i2info(inode);
508 struct lustre_handle lockh = { 0 };
513 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
514 LDLM_FL_BLOCK_CONV)) {
515 LBUG(); /* not expecting any blocked async locks yet */
516 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
518 ldlm_lock_dump(D_OTHER, lock, 0);
519 ldlm_reprocess_all(lock->l_resource);
523 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
525 stripe = ll_lock_to_stripe_offset(inode, lock);
527 if (lock->l_lvb_len) {
528 struct lov_stripe_md *lsm = lli->lli_smd;
530 lvb = lock->l_lvb_data;
531 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
534 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
535 kms = ldlm_extent_shift_kms(NULL, kms);
536 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
537 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
538 lsm->lsm_oinfo[stripe].loi_kms, kms);
539 lsm->lsm_oinfo[stripe].loi_kms = kms;
544 wake_up(&lock->l_waitq);
546 ldlm_lock2handle(lock, &lockh);
547 ldlm_lock_decref(&lockh, LCK_PR);
552 /* This function is a disaster. I hate the LOV. */
553 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
555 struct ptlrpc_request *req = reqp;
556 struct inode *inode = ll_inode_from_lock(lock);
557 struct obd_export *exp;
558 struct ll_inode_info *lli;
563 struct lov_stripe_md *lsm;
565 __u32 vallen = sizeof(data);
566 int rc, size = sizeof(*lvb);
571 lli = ll_i2info(inode);
574 if (lli->lli_smd == NULL)
576 exp = ll_i2obdexp(inode);
578 /* First, find out which stripe index this lock corresponds to. */
579 if (lli->lli_smd->lsm_stripe_count > 1)
580 data.stripe_number = ll_lock_to_stripe_offset(inode, lock);
582 data.stripe_number = 0;
584 data.size = inode->i_size;
585 data.lsm = lli->lli_smd;
587 rc = obd_get_info(exp, strlen("size_to_stripe"), "size_to_stripe",
590 CERROR("obd_get_info: rc = %d\n", rc);
594 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> size "LPU64,
595 inode->i_size, data.stripe_number, data.size);
597 rc = lustre_pack_reply(req, 1, &size, NULL);
599 CERROR("lustre_pack_reply: %d\n", rc);
603 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
604 lvb->lvb_size = data.size;
612 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
613 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
615 /* NB: lov_merge_size will prefer locally cached writes if they extend the
616 * file (because it prefers KMS over RSS when larger) */
617 int ll_glimpse_size(struct inode *inode, struct ost_lvb *lvb)
619 struct ll_inode_info *lli = ll_i2info(inode);
620 struct ll_sb_info *sbi = ll_i2sbi(inode);
621 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
622 struct lustre_handle lockh;
623 int rc, flags = LDLM_FL_HAS_INTENT;
626 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
628 rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
629 LCK_PR, &flags, ll_extent_lock_callback,
630 ldlm_completion_ast, ll_glimpse_callback, inode,
631 sizeof(*lvb), lustre_swab_ost_lvb, &lockh);
633 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
637 lvb->lvb_size = lov_merge_size(lli->lli_smd, 0);
638 //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
640 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64"\n", lvb->lvb_size);
642 obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
647 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
648 struct lov_stripe_md *lsm, int mode,
649 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
652 struct ll_sb_info *sbi = ll_i2sbi(inode);
656 LASSERT(lockh->cookie == 0);
658 /* XXX phil: can we do this? won't it screw the file size up? */
659 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
660 (sbi->ll_flags & LL_SBI_NOLCK))
663 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
664 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
666 rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
667 &ast_flags, ll_extent_lock_callback,
668 ldlm_completion_ast, ll_glimpse_callback, inode,
669 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
673 if (policy->l_extent.start == 0 &&
674 policy->l_extent.end == OBD_OBJECT_EOF)
675 inode->i_size = lov_merge_size(lsm, 1);
677 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
682 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
683 struct lov_stripe_md *lsm, int mode,
684 struct lustre_handle *lockh)
686 struct ll_sb_info *sbi = ll_i2sbi(inode);
690 /* XXX phil: can we do this? won't it screw the file size up? */
691 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
692 (sbi->ll_flags & LL_SBI_NOLCK))
695 rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
700 static ssize_t ll_file_read(struct file *filp, char *buf, size_t count,
703 struct ll_file_data *fd = filp->private_data;
704 struct inode *inode = filp->f_dentry->d_inode;
705 struct ll_inode_info *lli = ll_i2info(inode);
706 struct lov_stripe_md *lsm = lli->lli_smd;
707 struct lustre_handle lockh = { 0 };
708 ldlm_policy_data_t policy;
713 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
714 inode->i_ino, inode->i_generation, inode, count, *ppos);
716 /* "If nbyte is 0, read() will return 0 and have no other results."
717 * -- Single Unix Spec */
721 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
727 policy.l_extent.start = *ppos;
728 policy.l_extent.end = *ppos + count - 1;
730 err = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh, 0);
734 kms = lov_merge_size(lsm, 1);
735 if (policy.l_extent.end > kms) {
736 /* A glimpse is necessary to determine whether we return a short
737 * read or some zeroes at the end of the buffer */
739 retval = ll_glimpse_size(inode, &lvb);
742 inode->i_size = lvb.lvb_size;
747 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
748 inode->i_ino, count, *ppos, inode->i_size);
750 /* turn off the kernel's read-ahead */
751 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
754 filp->f_ra.ra_pages = 0;
756 retval = generic_file_read(filp, buf, count, ppos);
759 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
764 * Write to a file (through the page cache).
766 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
769 struct ll_file_data *fd = file->private_data;
770 struct inode *inode = file->f_dentry->d_inode;
771 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
772 struct lustre_handle lockh = { 0 };
773 ldlm_policy_data_t policy;
774 loff_t maxbytes = ll_file_maxbytes(inode);
778 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
779 inode->i_ino, inode->i_generation, inode, count, *ppos);
781 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
783 /* POSIX, but surprised the VFS doesn't check this already */
787 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
788 * called on the file, don't fail the below assertion (bug 2388). */
789 if (file->f_flags & O_LOV_DELAY_CREATE && lsm == NULL)
794 if (file->f_flags & O_APPEND) {
795 policy.l_extent.start = 0;
796 policy.l_extent.end = OBD_OBJECT_EOF;
798 policy.l_extent.start = *ppos;
799 policy.l_extent.end = *ppos + count - 1;
802 err = ll_extent_lock(fd, inode, lsm, LCK_PW, &policy, &lockh, 0);
806 /* this is ok, g_f_w will overwrite this under i_sem if it races
807 * with a local truncate, it just makes our maxbyte checking easier */
808 if (file->f_flags & O_APPEND)
809 *ppos = inode->i_size;
811 if (*ppos >= maxbytes) {
812 if (count || *ppos > maxbytes) {
813 send_sig(SIGXFSZ, current, 0);
814 GOTO(out, retval = -EFBIG);
817 if (*ppos + count > maxbytes)
818 count = maxbytes - *ppos;
820 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
821 inode->i_ino, count, *ppos);
823 /* generic_file_write handles O_APPEND after getting i_sem */
824 retval = generic_file_write(file, buf, count, ppos);
828 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
830 ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
834 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
837 struct ll_inode_info *lli = ll_i2info(inode);
838 struct obd_export *exp = ll_i2obdexp(inode);
839 struct ll_recreate_obj ucreatp;
840 struct obd_trans_info oti = { 0 };
841 struct obdo *oa = NULL;
844 struct lov_stripe_md *lsm, *lsm2;
847 if (!capable (CAP_SYS_ADMIN))
850 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
851 sizeof(struct ll_recreate_obj));
860 down(&lli->lli_open_sem);
863 up(&lli->lli_open_sem);
867 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
868 (lsm->lsm_stripe_count));
870 OBD_ALLOC(lsm2, lsm_size);
872 up(&lli->lli_open_sem);
877 oa->o_id = ucreatp.lrc_id;
878 oa->o_nlink = ucreatp.lrc_ost_idx;
879 oa->o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS;
880 oa->o_flags |= OBD_FL_RECREATE_OBJS;
881 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
882 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
884 oti.oti_objid = NULL;
885 memcpy(lsm2, lsm, lsm_size);
886 rc = obd_create(exp, oa, &lsm2, &oti);
888 up(&lli->lli_open_sem);
889 OBD_FREE(lsm2, lsm_size);
894 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
895 int flags, struct lov_user_md *lum, int lum_size)
897 struct ll_inode_info *lli = ll_i2info(inode);
899 struct obd_export *exp = ll_i2obdexp(inode);
900 struct lov_stripe_md *lsm;
901 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
902 struct ptlrpc_request *req = NULL;
907 down(&lli->lli_open_sem);
910 up(&lli->lli_open_sem);
911 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
916 f = get_empty_filp();
920 f->f_dentry = file->f_dentry;
921 f->f_vfsmnt = file->f_vfsmnt;
923 rc = ll_intent_file_open(f, lum, lum_size, &oit);
926 if (it_disposition(&oit, DISP_LOOKUP_NEG))
928 req = oit.d.lustre.it_data;
929 rc = oit.d.lustre.it_status;
934 rc = mdc_req2lustre_md(req, 1, exp, &md);
937 ll_update_inode(f->f_dentry->d_inode, md.body, md.lsm);
939 rc = ll_local_open(f, &oit);
942 ll_intent_release(&oit);
944 rc = ll_file_release(f->f_dentry->d_inode, f);
949 up(&lli->lli_open_sem);
951 ptlrpc_req_finished(req);
955 static int ll_lov_setea(struct inode *inode, struct file *file,
958 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
959 struct lov_user_md *lump;
960 int lum_size = sizeof(struct lov_user_md) +
961 sizeof(struct lov_user_ost_data);
965 if (!capable (CAP_SYS_ADMIN))
968 OBD_ALLOC(lump, lum_size);
972 rc = copy_from_user(lump, (struct lov_user_md *)arg,
975 OBD_FREE(lump, lum_size);
979 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
981 OBD_FREE(lump, lum_size);
985 static int ll_lov_setstripe(struct inode *inode, struct file *file,
988 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
990 int flags = FMODE_WRITE;
993 /* Bug 1152: copy properly when this is no longer true */
994 LASSERT(sizeof(lum) == sizeof(*lump));
995 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
996 rc = copy_from_user(&lum, lump, sizeof(lum));
1000 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1004 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1006 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1011 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2obdexp(inode), 0, lsm,
1015 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1018 struct ll_file_data *fd = file->private_data;
1022 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1023 inode->i_generation, inode, cmd);
1025 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1028 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1030 case LL_IOC_GETFLAGS:
1031 /* Get the current value of the file flags */
1032 return put_user(fd->fd_flags, (int *)arg);
1033 case LL_IOC_SETFLAGS:
1034 case LL_IOC_CLRFLAGS:
1035 /* Set or clear specific file flags */
1036 /* XXX This probably needs checks to ensure the flags are
1037 * not abused, and to handle any flag side effects.
1039 if (get_user(flags, (int *) arg))
1042 if (cmd == LL_IOC_SETFLAGS)
1043 fd->fd_flags |= flags;
1045 fd->fd_flags &= ~flags;
1047 case LL_IOC_LOV_SETSTRIPE:
1048 RETURN(ll_lov_setstripe(inode, file, arg));
1049 case LL_IOC_LOV_SETEA:
1050 RETURN( ll_lov_setea(inode, file, arg) );
1051 case LL_IOC_LOV_GETSTRIPE:
1052 RETURN(ll_lov_getstripe(inode, arg));
1053 case LL_IOC_RECREATE_OBJ:
1054 RETURN(ll_lov_recreate_obj(inode, file, arg));
1055 case EXT3_IOC_GETFLAGS:
1056 case EXT3_IOC_SETFLAGS:
1057 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1058 /* We need to special case any other ioctls we want to handle,
1059 * to send them to the MDS/OST as appropriate and to properly
1060 * network encode the arg field.
1061 case EXT2_IOC_GETVERSION_OLD:
1062 case EXT2_IOC_GETVERSION_NEW:
1063 case EXT2_IOC_SETVERSION_OLD:
1064 case EXT2_IOC_SETVERSION_NEW:
1067 RETURN( obd_iocontrol(cmd, ll_i2obdexp(inode), 0, NULL,
1072 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1074 struct inode *inode = file->f_dentry->d_inode;
1075 struct ll_file_data *fd = file->private_data;
1076 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1077 struct lustre_handle lockh = {0};
1080 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1081 inode->i_generation, inode,
1082 offset + ((origin==2) ? inode->i_size : file->f_pos));
1084 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1085 if (origin == 2) { /* SEEK_END */
1087 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1088 err = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,0);
1089 if (err != ELDLM_OK)
1092 offset += inode->i_size;
1093 } else if (origin == 1) { /* SEEK_CUR */
1094 offset += file->f_pos;
1098 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1099 if (offset != file->f_pos) {
1100 file->f_pos = offset;
1101 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1103 file->f_version = ++event;
1110 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1114 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1116 struct inode *inode = dentry->d_inode;
1117 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1119 struct ptlrpc_request *req;
1122 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1123 inode->i_generation, inode);
1125 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1127 /* fsync's caller has already called _fdata{sync,write}, we want
1128 * that IO to finish before calling the osc and mdc sync methods */
1129 rc = filemap_fdatawait(inode->i_mapping);
1131 ll_inode2fid(&fid, inode);
1132 err = mdc_sync(ll_i2sbi(inode)->ll_mdc_exp, &fid, &req);
1136 ptlrpc_req_finished(req);
1139 struct obdo *oa = obdo_alloc();
1142 RETURN(rc ? rc : -ENOMEM);
1144 oa->o_id = lsm->lsm_object_id;
1145 oa->o_valid = OBD_MD_FLID;
1146 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
1147 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
1149 err = obd_sync(ll_i2sbi(inode)->ll_osc_exp, oa, lsm,
1159 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1161 struct inode *inode = file->f_dentry->d_inode;
1162 struct ll_sb_info *sbi = ll_i2sbi(inode);
1163 struct obd_device *obddev;
1164 struct ldlm_res_id res_id =
1165 { .name = {inode->i_ino, inode->i_generation, LDLM_FLOCK} };
1166 struct lustre_handle lockh = {0};
1167 ldlm_policy_data_t flock;
1168 ldlm_mode_t mode = 0;
1173 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1174 inode->i_ino, file_lock);
1176 flock.l_flock.pid = file_lock->fl_pid;
1177 flock.l_flock.start = file_lock->fl_start;
1178 flock.l_flock.end = file_lock->fl_end;
1180 switch (file_lock->fl_type) {
1185 /* An unlock request may or may not have any relation to
1186 * existing locks so we may not be able to pass a lock handle
1187 * via a normal ldlm_lock_cancel() request. The request may even
1188 * unlock a byte range in the middle of an existing lock. In
1189 * order to process an unlock request we need all of the same
1190 * information that is given with a normal read or write record
1191 * lock request. To avoid creating another ldlm unlock (cancel)
1192 * message we'll treat a LCK_NL flock request as an unlock. */
1199 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1208 flags = LDLM_FL_BLOCK_NOWAIT;
1211 flags = LDLM_FL_TEST_LOCK;
1212 /* Save the old mode so that if the mode in the lock changes we
1213 * can decrement the appropriate reader or writer refcount. */
1214 file_lock->fl_type = mode;
1217 CERROR("unknown fcntl lock command: %d\n", cmd);
1221 CDEBUG(D_DLMTRACE, "inode=%lu, pid=%u, flags=%#x, mode=%u, "
1222 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1223 flags, mode, flock.l_flock.start, flock.l_flock.end);
1225 obddev = sbi->ll_mdc_exp->exp_obd;
1226 rc = ldlm_cli_enqueue(sbi->ll_mdc_exp, NULL, obddev->obd_namespace,
1227 res_id, LDLM_FLOCK, &flock, mode, &flags,
1228 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1229 NULL, 0, NULL, &lockh);
1233 static int ll_have_md_lock(struct dentry *de)
1235 struct ll_sb_info *sbi = ll_s2sbi(de->d_sb);
1236 struct lustre_handle lockh;
1237 struct ldlm_res_id res_id = { .name = {0} };
1238 struct obd_device *obddev;
1245 obddev = sbi->ll_mdc_exp->exp_obd;
1246 res_id.name[0] = de->d_inode->i_ino;
1247 res_id.name[1] = de->d_inode->i_generation;
1249 CDEBUG(D_INFO, "trying to match res "LPU64"\n", res_id.name[0]);
1251 /* FIXME use LDLM_FL_TEST_LOCK instead */
1252 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1253 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
1254 NULL, LCK_PR, &lockh)) {
1255 ldlm_lock_decref(&lockh, LCK_PR);
1259 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
1260 NULL, LCK_PW, &lockh)) {
1261 ldlm_lock_decref(&lockh, LCK_PW);
1267 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
1269 struct inode *inode = dentry->d_inode;
1270 struct ll_inode_info *lli;
1271 struct lov_stripe_md *lsm;
1275 CERROR("REPORT THIS LINE TO PETER\n");
1278 lli = ll_i2info(inode);
1279 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s\n",
1280 inode->i_ino, inode->i_generation, inode, dentry->d_name.name);
1281 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1282 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1285 if (!ll_have_md_lock(dentry)) {
1286 struct ptlrpc_request *req = NULL;
1287 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
1289 unsigned long valid = 0;
1292 if (S_ISREG(inode->i_mode)) {
1293 ealen = obd_size_diskmd(sbi->ll_osc_exp, NULL);
1294 valid |= OBD_MD_FLEASIZE;
1296 ll_inode2fid(&fid, inode);
1297 rc = mdc_getattr(sbi->ll_mdc_exp, &fid, valid, ealen, &req);
1299 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
1302 rc = ll_prep_inode(sbi->ll_osc_exp, &inode, req, 0, NULL);
1304 ptlrpc_req_finished(req);
1307 ptlrpc_req_finished(req);
1311 if (lsm == NULL) /* object not yet allocated, don't validate size */
1314 /* ll_glimpse_size will prefer locally cached writes if they extend
1320 err = ll_glimpse_size(inode, &lvb);
1321 inode->i_size = lvb.lvb_size;
1326 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1327 int ll_getattr(struct vfsmount *mnt, struct dentry *de,
1328 struct lookup_intent *it, struct kstat *stat)
1331 struct inode *inode = de->d_inode;
1333 res = ll_inode_revalidate_it(de, it);
1334 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1339 stat->dev = inode->i_sb->s_dev;
1340 stat->ino = inode->i_ino;
1341 stat->mode = inode->i_mode;
1342 stat->nlink = inode->i_nlink;
1343 stat->uid = inode->i_uid;
1344 stat->gid = inode->i_gid;
1345 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1346 stat->atime = inode->i_atime;
1347 stat->mtime = inode->i_mtime;
1348 stat->ctime = inode->i_ctime;
1349 stat->size = inode->i_size;
1350 stat->blksize = inode->i_blksize;
1351 stat->blocks = inode->i_blocks;
1356 struct file_operations ll_file_operations = {
1358 write: ll_file_write,
1359 ioctl: ll_file_ioctl,
1361 release: ll_file_release,
1362 mmap: generic_file_mmap,
1363 llseek: ll_file_seek,
1365 //lock: ll_file_flock
1368 struct inode_operations ll_file_inode_operations = {
1369 setattr_raw: ll_setattr_raw,
1370 setattr: ll_setattr,
1371 truncate: ll_truncate,
1372 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1373 getattr_it: ll_getattr,
1375 revalidate_it: ll_inode_revalidate_it,
1379 struct inode_operations ll_special_inode_operations = {
1380 setattr_raw: ll_setattr_raw,
1381 setattr: ll_setattr,
1382 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1383 getattr_it: ll_getattr,
1385 revalidate_it: ll_inode_revalidate_it,