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>
33 #include "llite_internal.h"
35 int ll_mdc_close(struct obd_export *mdc_exp, struct inode *inode,
38 struct ll_file_data *fd = file->private_data;
39 struct ptlrpc_request *req = NULL;
40 struct obd_client_handle *och = &fd->fd_mds_och;
45 /* clear group lock, if present */
46 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
47 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
48 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
49 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
53 obdo.o_id = inode->i_ino;
54 obdo.o_valid = OBD_MD_FLID;
55 obdo_from_inode(&obdo, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
56 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
57 OBD_MD_FLATIME | OBD_MD_FLMTIME |
59 if (0 /* ll_is_inode_dirty(inode) */) {
60 obdo.o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
61 obdo.o_valid |= OBD_MD_FLFLAGS;
63 obdo.o_mds = ll_i2info(inode)->lli_mds;
64 rc = md_close(mdc_exp, &obdo, och, &req);
67 /* We are the last writer, so the MDS has instructed us to get
68 * the file size and any write cookies, then close again. */
69 //ll_queue_done_writing(inode);
72 CERROR("inode %lu mdc close failed: rc = %d\n",
76 rc = ll_objects_destroy(req, file->f_dentry->d_inode);
78 CERROR("inode %lu ll_objects destroy: rc = %d\n",
82 mdc_clear_open_replay_data(mdc_exp, och);
83 ptlrpc_req_finished(req);
84 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
85 file->private_data = NULL;
86 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof *fd);
91 /* While this returns an error code, fput() the caller does not, so we need
92 * to make every effort to clean up all of our state here. Also, applications
93 * rarely check close errors and even if an error is returned they will not
94 * re-try the close call.
96 int ll_file_release(struct inode *inode, struct file *file)
98 struct ll_file_data *fd;
99 struct ll_sb_info *sbi = ll_i2sbi(inode);
103 CDEBUG(D_VFSTRACE, "VFS Op:inode=%u/%lu/%u(%p)\n",
104 ll_i2info(inode)->lli_mds, inode->i_ino,
105 inode->i_generation, inode);
107 /* don't do anything for / */
108 if (inode->i_sb->s_root == file->f_dentry)
111 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
112 fd = (struct ll_file_data *)file->private_data;
115 rc = ll_mdc_close(sbi->ll_mdc_exp, inode, file);
119 static int ll_intent_file_open(struct file *file, void *lmm,
120 int lmmsize, struct lookup_intent *itp)
122 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
123 struct lustre_handle lockh;
124 struct mdc_op_data data;
125 struct dentry *parent = file->f_dentry->d_parent;
126 const char *name = file->f_dentry->d_name.name;
127 const int len = file->f_dentry->d_name.len;
133 ll_prepare_mdc_op_data(&data, parent->d_inode, NULL, name, len, O_RDWR);
135 rc = md_enqueue(sbi->ll_mdc_exp, LDLM_IBITS, itp, LCK_PR, &data,
136 &lockh, lmm, lmmsize, ldlm_completion_ast,
137 ll_mdc_blocking_ast, NULL);
139 if (itp->d.lustre.it_lock_mode)
140 memcpy(&itp->d.lustre.it_lock_handle,
141 &lockh, sizeof(lockh));
143 CERROR("lock enqueue: err: %d\n", rc);
149 int ll_local_open(struct file *file, struct lookup_intent *it)
151 struct ptlrpc_request *req = it->d.lustre.it_data;
152 struct ll_inode_info *lli = ll_i2info(file->f_dentry->d_inode);
153 struct obd_export *mdc_exp = ll_i2mdcexp(file->f_dentry->d_inode);
154 struct ll_file_data *fd;
155 struct mds_body *body;
158 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
159 LASSERT (body != NULL); /* reply already checked out */
160 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
162 LASSERT(!file->private_data);
164 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
165 /* We can't handle this well without reorganizing ll_file_open and
166 * ll_mdc_close, so don't even try right now. */
169 memcpy(&fd->fd_mds_och.och_fh, &body->handle, sizeof(body->handle));
170 fd->fd_mds_och.och_magic = OBD_CLIENT_HANDLE_MAGIC;
171 file->private_data = fd;
172 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
174 lli->lli_io_epoch = body->io_epoch;
176 mdc_set_open_replay_data(mdc_exp, &fd->fd_mds_och, it->d.lustre.it_data);
181 /* Open a file, and (for the very first open) create objects on the OSTs at
182 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
183 * creation or open until ll_lov_setstripe() ioctl is called. We grab
184 * lli_open_sem to ensure no other process will create objects, send the
185 * stripe MD to the MDS, or try to destroy the objects if that fails.
187 * If we already have the stripe MD locally then we don't request it in
188 * mdc_open(), by passing a lmm_size = 0.
190 * It is up to the application to ensure no other processes open this file
191 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
192 * used. We might be able to avoid races of that sort by getting lli_open_sem
193 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
194 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
196 int ll_file_open(struct inode *inode, struct file *file)
198 struct ll_inode_info *lli = ll_i2info(inode);
199 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
200 .it_flags = file->f_flags };
201 struct lov_stripe_md *lsm;
202 struct ptlrpc_request *req;
206 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
207 inode->i_generation, inode);
209 /* don't do anything for / */
210 if (inode->i_sb->s_root == file->f_dentry)
215 if (!it || !it->d.lustre.it_disposition) {
217 rc = ll_intent_file_open(file, NULL, 0, it);
222 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
223 rc = it_open_error(DISP_OPEN_OPEN, it);
227 rc = ll_local_open(file, it);
231 ll_intent_drop_lock(it);
233 if (!S_ISREG(inode->i_mode))
238 if (file->f_flags & O_LOV_DELAY_CREATE ||
239 !(file->f_mode & FMODE_WRITE)) {
240 CDEBUG(D_INODE, "object creation was delayed\n");
244 file->f_flags &= ~O_LOV_DELAY_CREATE;
247 req = it->d.lustre.it_data;
248 ptlrpc_req_finished(req);
250 ll_open_complete(inode);
254 /* Fills the obdo with the attributes for the inode defined by lsm */
255 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
258 struct ptlrpc_request_set *set;
262 LASSERT(lsm != NULL);
264 memset(oa, 0, sizeof *oa);
265 oa->o_id = lsm->lsm_object_id;
266 oa->o_gr = lsm->lsm_object_gr;
267 oa->o_mode = S_IFREG;
268 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
269 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
270 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
272 set = ptlrpc_prep_set();
274 CERROR ("ENOMEM allocing request set\n");
277 rc = obd_getattr_async(exp, oa, lsm, set);
279 rc = ptlrpc_set_wait(set);
280 ptlrpc_set_destroy(set);
285 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
286 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
290 static inline void ll_remove_suid(struct inode *inode)
294 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
295 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
297 /* was any of the uid bits set? */
298 mode &= inode->i_mode;
299 if (mode && !capable(CAP_FSETID)) {
300 inode->i_mode &= ~mode;
301 // XXX careful here - we cannot change the size
305 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
307 struct ll_inode_info *lli = ll_i2info(inode);
308 struct lov_stripe_md *lsm = lli->lli_smd;
309 struct obd_export *exp = ll_i2obdexp(inode);
312 struct ldlm_lock *lock;
313 struct lov_stripe_md *lsm;
314 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
315 __u32 stripe, vallen = sizeof(stripe);
319 if (lsm->lsm_stripe_count == 1)
322 /* get our offset in the lov */
323 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
325 CERROR("obd_get_info: rc = %d\n", rc);
328 LASSERT(stripe < lsm->lsm_stripe_count);
332 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
333 * we get a lock cancellation for each stripe, so we have to map the obd's
334 * region back onto the stripes in the file that it held.
336 * No one can dirty the extent until we've finished our work and they can
337 * enqueue another lock. The DLM protects us from ll_file_read/write here,
338 * but other kernel actors could have pages locked.
340 * Called with the DLM lock held. */
341 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
342 struct ldlm_lock *lock, __u32 stripe)
344 ldlm_policy_data_t tmpex;
345 unsigned long start, end, count, skip, i, j;
347 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
348 struct lustre_handle lockh;
351 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
352 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
353 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
356 /* our locks are page granular thanks to osc_enqueue, we invalidate the
358 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
359 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
363 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
364 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
365 if (lsm->lsm_stripe_count > 1) {
366 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
367 skip = (lsm->lsm_stripe_count - 1) * count;
368 start += start/count * skip + stripe * count;
370 end += end/count * skip + stripe * count;
372 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
375 i = (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
379 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
380 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
381 count, skip, end, discard ? " (DISCARDING)" : "");
383 /* this is the simplistic implementation of page eviction at
384 * cancelation. It is careful to get races with other page
385 * lockers handled correctly. fixes from bug 20 will make it
386 * more efficient by associating locks with pages and with
387 * batching writeback under the lock explicitly. */
388 for (i = start, j = start % count; i <= end;
389 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
391 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
397 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
398 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
399 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
402 if (!mapping_has_pages(inode->i_mapping)) {
403 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
407 conditional_schedule();
409 page = find_get_page(inode->i_mapping, i);
412 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
413 i, tmpex.l_extent.start);
416 /* page->mapping to check with racing against teardown */
417 if (!discard && clear_page_dirty_for_io(page)) {
418 rc = ll_call_writepage(inode, page);
420 CERROR("writepage of page %p failed: %d\n",
422 /* either waiting for io to complete or reacquiring
423 * the lock that the failed writepage released */
427 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
428 /* check to see if another DLM lock covers this page */
429 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
430 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
432 &lock->l_resource->lr_name, LDLM_EXTENT,
433 &tmpex, LCK_PR | LCK_PW, &lockh);
434 if (rc2 == 0 && page->mapping != NULL) {
435 // checking again to account for writeback's lock_page()
436 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
437 ll_truncate_complete_page(page);
440 page_cache_release(page);
442 LASSERTF(tmpex.l_extent.start <=
443 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
444 lock->l_policy_data.l_extent.end + 1),
445 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
446 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
451 static int ll_extent_lock_callback(struct ldlm_lock *lock,
452 struct ldlm_lock_desc *new, void *data,
455 struct lustre_handle lockh = { 0 };
459 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
460 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
465 case LDLM_CB_BLOCKING:
466 ldlm_lock2handle(lock, &lockh);
467 rc = ldlm_cli_cancel(&lockh);
469 CERROR("ldlm_cli_cancel failed: %d\n", rc);
471 case LDLM_CB_CANCELING: {
473 struct ll_inode_info *lli;
474 struct lov_stripe_md *lsm;
478 /* This lock wasn't granted, don't try to evict pages */
479 if (lock->l_req_mode != lock->l_granted_mode)
482 inode = ll_inode_from_lock(lock);
485 lli = ll_i2info(inode);
488 if (lli->lli_smd == NULL)
492 stripe = ll_lock_to_stripe_offset(inode, lock);
493 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
496 kms = ldlm_extent_shift_kms(lock,
497 lsm->lsm_oinfo[stripe].loi_kms);
499 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
500 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
501 lsm->lsm_oinfo[stripe].loi_kms, kms);
502 lsm->lsm_oinfo[stripe].loi_kms = kms;
504 //ll_try_done_writing(inode);
517 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
519 /* XXX ALLOCATE - 160 bytes */
520 struct inode *inode = ll_inode_from_lock(lock);
521 struct ll_inode_info *lli = ll_i2info(inode);
522 struct lustre_handle lockh = { 0 };
527 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
528 LDLM_FL_BLOCK_CONV)) {
529 LBUG(); /* not expecting any blocked async locks yet */
530 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
532 ldlm_lock_dump(D_OTHER, lock, 0);
533 ldlm_reprocess_all(lock->l_resource);
537 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
539 stripe = ll_lock_to_stripe_offset(inode, lock);
541 if (lock->l_lvb_len) {
542 struct lov_stripe_md *lsm = lli->lli_smd;
544 lvb = lock->l_lvb_data;
545 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
548 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
549 kms = ldlm_extent_shift_kms(NULL, kms);
550 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
551 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
552 lsm->lsm_oinfo[stripe].loi_kms, kms);
553 lsm->lsm_oinfo[stripe].loi_kms = kms;
558 wake_up(&lock->l_waitq);
560 ldlm_lock2handle(lock, &lockh);
561 ldlm_lock_decref(&lockh, LCK_PR);
566 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
568 struct ptlrpc_request *req = reqp;
569 struct inode *inode = ll_inode_from_lock(lock);
570 struct ll_inode_info *lli;
572 int rc, size = sizeof(*lvb), stripe = 0;
576 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
577 lli = ll_i2info(inode);
579 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
580 if (lli->lli_smd == NULL)
581 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
583 /* First, find out which stripe index this lock corresponds to. */
584 if (lli->lli_smd->lsm_stripe_count > 1)
585 stripe = ll_lock_to_stripe_offset(inode, lock);
587 rc = lustre_pack_reply(req, 1, &size, NULL);
589 CERROR("lustre_pack_reply: %d\n", rc);
593 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
594 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
596 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
597 inode->i_size, stripe, lvb->lvb_size);
603 /* These errors are normal races, so we don't want to fill the console
604 * with messages by calling ptlrpc_error() */
605 if (rc == -ELDLM_NO_LOCK_DATA)
606 lustre_pack_reply(req, 0, NULL, NULL);
612 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
613 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
614 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
616 /* NB: lov_merge_size will prefer locally cached writes if they extend the
617 * file (because it prefers KMS over RSS when larger) */
618 int ll_glimpse_size(struct inode *inode, struct ost_lvb *lvb)
620 struct ll_inode_info *lli = ll_i2info(inode);
621 struct ll_sb_info *sbi = ll_i2sbi(inode);
622 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
623 struct lustre_handle lockh = { 0 };
624 int rc, flags = LDLM_FL_HAS_INTENT;
627 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
629 rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
630 LCK_PR, &flags, ll_extent_lock_callback,
631 ldlm_completion_ast, ll_glimpse_callback, inode,
632 sizeof(*lvb), lustre_swab_ost_lvb, &lockh);
634 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
635 RETURN(rc > 0 ? -EIO : rc);
638 lvb->lvb_size = lov_merge_size(lli->lli_smd, 0);
639 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
640 //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
642 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
643 lvb->lvb_size, lvb->lvb_blocks);
645 obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
650 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
651 struct lov_stripe_md *lsm, int mode,
652 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
655 struct ll_sb_info *sbi = ll_i2sbi(inode);
659 LASSERT(lockh->cookie == 0);
661 /* XXX phil: can we do this? won't it screw the file size up? */
662 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
663 (sbi->ll_flags & LL_SBI_NOLCK))
666 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
667 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
669 rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
670 &ast_flags, ll_extent_lock_callback,
671 ldlm_completion_ast, ll_glimpse_callback, inode,
672 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
676 if (policy->l_extent.start == 0 &&
677 policy->l_extent.end == OBD_OBJECT_EOF)
678 inode->i_size = lov_merge_size(lsm, 1);
680 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
685 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
686 struct lov_stripe_md *lsm, int mode,
687 struct lustre_handle *lockh)
689 struct ll_sb_info *sbi = ll_i2sbi(inode);
693 /* XXX phil: can we do this? won't it screw the file size up? */
694 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
695 (sbi->ll_flags & LL_SBI_NOLCK))
698 rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
703 static ssize_t ll_file_read(struct file *filp, char *buf, size_t count,
706 struct ll_file_data *fd = filp->private_data;
707 struct inode *inode = filp->f_dentry->d_inode;
708 struct ll_inode_info *lli = ll_i2info(inode);
709 struct lov_stripe_md *lsm = lli->lli_smd;
710 struct lustre_handle lockh = { 0 };
711 ldlm_policy_data_t policy;
716 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
717 inode->i_ino, inode->i_generation, inode, count, *ppos);
719 /* "If nbyte is 0, read() will return 0 and have no other results."
720 * -- Single Unix Spec */
724 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
730 policy.l_extent.start = *ppos;
731 policy.l_extent.end = *ppos + count - 1;
733 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
734 (filp->f_flags & O_NONBLOCK) ?
735 LDLM_FL_BLOCK_NOWAIT: 0);
739 kms = lov_merge_size(lsm, 1);
740 if (*ppos + count - 1 > kms) {
741 /* A glimpse is necessary to determine whether we return a short
742 * read or some zeroes at the end of the buffer */
744 retval = ll_glimpse_size(inode, &lvb);
747 inode->i_size = lvb.lvb_size;
752 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
753 inode->i_ino, count, *ppos, inode->i_size);
755 /* turn off the kernel's read-ahead */
756 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
759 filp->f_ra.ra_pages = 0;
761 retval = generic_file_read(filp, buf, count, ppos);
764 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
769 * Write to a file (through the page cache).
771 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
774 struct ll_file_data *fd = file->private_data;
775 struct inode *inode = file->f_dentry->d_inode;
776 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
777 struct lustre_handle lockh = { 0 };
778 ldlm_policy_data_t policy;
779 loff_t maxbytes = ll_file_maxbytes(inode);
781 int nonblock = 0, rc;
783 if (file->f_flags & O_NONBLOCK)
784 nonblock = LDLM_FL_BLOCK_NOWAIT;
785 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
786 inode->i_ino, inode->i_generation, inode, count, *ppos);
788 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
790 /* POSIX, but surprised the VFS doesn't check this already */
794 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
795 * called on the file, don't fail the below assertion (bug 2388). */
796 if (file->f_flags & O_LOV_DELAY_CREATE && lsm == NULL)
801 if (file->f_flags & O_APPEND) {
802 policy.l_extent.start = 0;
803 policy.l_extent.end = OBD_OBJECT_EOF;
805 policy.l_extent.start = *ppos;
806 policy.l_extent.end = *ppos + count - 1;
809 rc = ll_extent_lock(fd, inode, lsm, LCK_PW, &policy, &lockh, nonblock);
813 /* this is ok, g_f_w will overwrite this under i_sem if it races
814 * with a local truncate, it just makes our maxbyte checking easier */
815 if (file->f_flags & O_APPEND)
816 *ppos = inode->i_size;
818 if (*ppos >= maxbytes) {
819 if (count || *ppos > maxbytes) {
820 send_sig(SIGXFSZ, current, 0);
821 GOTO(out, retval = -EFBIG);
824 if (*ppos + count > maxbytes)
825 count = maxbytes - *ppos;
827 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
828 inode->i_ino, count, *ppos);
830 /* generic_file_write handles O_APPEND after getting i_sem */
831 retval = generic_file_write(file, buf, count, ppos);
834 ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
835 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
836 retval > 0 ? retval : 0);
840 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
843 struct ll_inode_info *lli = ll_i2info(inode);
844 struct obd_export *exp = ll_i2obdexp(inode);
845 struct ll_recreate_obj ucreatp;
846 struct obd_trans_info oti = { 0 };
847 struct obdo *oa = NULL;
850 struct lov_stripe_md *lsm, *lsm2;
853 if (!capable (CAP_SYS_ADMIN))
856 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
857 sizeof(struct ll_recreate_obj));
866 down(&lli->lli_open_sem);
869 up(&lli->lli_open_sem);
873 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
874 (lsm->lsm_stripe_count));
876 OBD_ALLOC(lsm2, lsm_size);
878 up(&lli->lli_open_sem);
883 oa->o_id = ucreatp.lrc_id;
884 oa->o_nlink = ucreatp.lrc_ost_idx;
885 oa->o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS;
886 oa->o_flags |= OBD_FL_RECREATE_OBJS;
887 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
888 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
890 oti.oti_objid = NULL;
891 memcpy(lsm2, lsm, lsm_size);
892 rc = obd_create(exp, oa, &lsm2, &oti);
894 up(&lli->lli_open_sem);
895 OBD_FREE(lsm2, lsm_size);
900 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
901 int flags, struct lov_user_md *lum,
904 struct ll_inode_info *lli = ll_i2info(inode);
906 struct obd_export *exp = ll_i2obdexp(inode);
907 struct lov_stripe_md *lsm;
908 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
909 struct ptlrpc_request *req = NULL;
914 down(&lli->lli_open_sem);
917 up(&lli->lli_open_sem);
918 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
923 f = get_empty_filp();
927 f->f_dentry = file->f_dentry;
928 f->f_vfsmnt = file->f_vfsmnt;
930 rc = ll_intent_file_open(f, lum, lum_size, &oit);
933 if (it_disposition(&oit, DISP_LOOKUP_NEG))
935 req = oit.d.lustre.it_data;
936 rc = oit.d.lustre.it_status;
941 rc = mdc_req2lustre_md(ll_i2mdcexp(inode), req, 1, exp, &md);
944 ll_update_inode(f->f_dentry->d_inode, &md);
946 rc = ll_local_open(f, &oit);
949 ll_intent_release(&oit);
951 rc = ll_file_release(f->f_dentry->d_inode, f);
956 up(&lli->lli_open_sem);
958 ptlrpc_req_finished(req);
962 static int ll_lov_setea(struct inode *inode, struct file *file,
965 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
966 struct lov_user_md *lump;
967 int lum_size = sizeof(struct lov_user_md) +
968 sizeof(struct lov_user_ost_data);
972 if (!capable (CAP_SYS_ADMIN))
975 OBD_ALLOC(lump, lum_size);
979 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
981 OBD_FREE(lump, lum_size);
985 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
987 OBD_FREE(lump, lum_size);
991 static int ll_lov_setstripe(struct inode *inode, struct file *file,
994 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
996 int flags = FMODE_WRITE;
999 /* Bug 1152: copy properly when this is no longer true */
1000 LASSERT(sizeof(lum) == sizeof(*lump));
1001 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1002 rc = copy_from_user(&lum, lump, sizeof(lum));
1006 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1010 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1012 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1017 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2obdexp(inode), 0, lsm,
1021 static int ll_get_grouplock(struct inode *inode, struct file *file,
1024 struct ll_file_data *fd = file->private_data;
1025 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1026 .end = OBD_OBJECT_EOF}};
1027 struct lustre_handle lockh = { 0 };
1028 struct ll_inode_info *lli = ll_i2info(inode);
1029 struct lov_stripe_md *lsm = lli->lli_smd;
1033 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1037 policy.l_extent.gid = arg;
1038 if (file->f_flags & O_NONBLOCK)
1039 flags = LDLM_FL_BLOCK_NOWAIT;
1041 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags);
1045 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1047 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1052 static int ll_put_grouplock(struct inode *inode, struct file *file,
1055 struct ll_file_data *fd = file->private_data;
1056 struct ll_inode_info *lli = ll_i2info(inode);
1057 struct lov_stripe_md *lsm = lli->lli_smd;
1061 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1062 /* Ugh, it's already unlocked. */
1066 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1069 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1071 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1076 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1081 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1084 struct ll_file_data *fd = file->private_data;
1088 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1089 inode->i_generation, inode, cmd);
1091 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1094 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1096 case LL_IOC_GETFLAGS:
1097 /* Get the current value of the file flags */
1098 return put_user(fd->fd_flags, (int *)arg);
1099 case LL_IOC_SETFLAGS:
1100 case LL_IOC_CLRFLAGS:
1101 /* Set or clear specific file flags */
1102 /* XXX This probably needs checks to ensure the flags are
1103 * not abused, and to handle any flag side effects.
1105 if (get_user(flags, (int *) arg))
1108 if (cmd == LL_IOC_SETFLAGS)
1109 fd->fd_flags |= flags;
1111 fd->fd_flags &= ~flags;
1113 case LL_IOC_LOV_SETSTRIPE:
1114 RETURN(ll_lov_setstripe(inode, file, arg));
1115 case LL_IOC_LOV_SETEA:
1116 RETURN(ll_lov_setea(inode, file, arg));
1117 case LL_IOC_LOV_GETSTRIPE:
1118 RETURN(ll_lov_getstripe(inode, arg));
1119 case LL_IOC_RECREATE_OBJ:
1120 RETURN(ll_lov_recreate_obj(inode, file, arg));
1121 case EXT3_IOC_GETFLAGS:
1122 case EXT3_IOC_SETFLAGS:
1123 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1124 case LL_IOC_GROUP_LOCK:
1125 RETURN(ll_get_grouplock(inode, file, arg));
1126 case LL_IOC_GROUP_UNLOCK:
1127 RETURN(ll_put_grouplock(inode, file, arg));
1128 /* We need to special case any other ioctls we want to handle,
1129 * to send them to the MDS/OST as appropriate and to properly
1130 * network encode the arg field.
1131 case EXT2_IOC_GETVERSION_OLD:
1132 case EXT2_IOC_GETVERSION_NEW:
1133 case EXT2_IOC_SETVERSION_OLD:
1134 case EXT2_IOC_SETVERSION_NEW:
1137 RETURN( obd_iocontrol(cmd, ll_i2obdexp(inode), 0, NULL,
1142 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1144 struct inode *inode = file->f_dentry->d_inode;
1145 struct ll_file_data *fd = file->private_data;
1146 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1147 struct lustre_handle lockh = {0};
1150 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1151 inode->i_generation, inode,
1152 offset + ((origin==2) ? inode->i_size : file->f_pos));
1154 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1155 if (origin == 2) { /* SEEK_END */
1156 int nonblock = 0, rc;
1157 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1159 if (file->f_flags & O_NONBLOCK)
1160 nonblock = LDLM_FL_BLOCK_NOWAIT;
1162 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1167 offset += inode->i_size;
1168 } else if (origin == 1) { /* SEEK_CUR */
1169 offset += file->f_pos;
1173 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1174 if (offset != file->f_pos) {
1175 file->f_pos = offset;
1176 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1178 file->f_version = ++event;
1185 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1189 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1191 struct inode *inode = dentry->d_inode;
1192 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1194 struct ptlrpc_request *req;
1197 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1198 inode->i_generation, inode);
1200 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1202 /* fsync's caller has already called _fdata{sync,write}, we want
1203 * that IO to finish before calling the osc and mdc sync methods */
1204 rc = filemap_fdatawait(inode->i_mapping);
1206 ll_inode2fid(&fid, inode);
1207 err = md_sync(ll_i2sbi(inode)->ll_mdc_exp, &fid, &req);
1211 ptlrpc_req_finished(req);
1214 struct obdo *oa = obdo_alloc();
1217 RETURN(rc ? rc : -ENOMEM);
1219 oa->o_id = lsm->lsm_object_id;
1220 oa->o_gr = lsm->lsm_object_gr;
1221 oa->o_valid = OBD_MD_FLID;
1222 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
1223 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1226 err = obd_sync(ll_i2sbi(inode)->ll_osc_exp, oa, lsm,
1236 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1238 struct inode *inode = file->f_dentry->d_inode;
1239 struct ll_sb_info *sbi = ll_i2sbi(inode);
1240 struct obd_device *obddev;
1241 struct ldlm_res_id res_id =
1242 { .name = {inode->i_ino, inode->i_generation, LDLM_FLOCK} };
1243 struct lustre_handle lockh = {0};
1244 ldlm_policy_data_t flock;
1245 ldlm_mode_t mode = 0;
1250 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1251 inode->i_ino, file_lock);
1253 flock.l_flock.pid = file_lock->fl_pid;
1254 flock.l_flock.start = file_lock->fl_start;
1255 flock.l_flock.end = file_lock->fl_end;
1257 switch (file_lock->fl_type) {
1262 /* An unlock request may or may not have any relation to
1263 * existing locks so we may not be able to pass a lock handle
1264 * via a normal ldlm_lock_cancel() request. The request may even
1265 * unlock a byte range in the middle of an existing lock. In
1266 * order to process an unlock request we need all of the same
1267 * information that is given with a normal read or write record
1268 * lock request. To avoid creating another ldlm unlock (cancel)
1269 * message we'll treat a LCK_NL flock request as an unlock. */
1276 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1285 flags = LDLM_FL_BLOCK_NOWAIT;
1288 flags = LDLM_FL_TEST_LOCK;
1289 /* Save the old mode so that if the mode in the lock changes we
1290 * can decrement the appropriate reader or writer refcount. */
1291 file_lock->fl_type = mode;
1294 CERROR("unknown fcntl lock command: %d\n", cmd);
1298 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1299 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1300 flags, mode, flock.l_flock.start, flock.l_flock.end);
1302 obddev = md_get_real_obd(sbi->ll_mdc_exp, NULL, 0);
1303 rc = ldlm_cli_enqueue(sbi->ll_mdc_exp, NULL, obddev->obd_namespace,
1304 res_id, LDLM_FLOCK, &flock, mode, &flags,
1305 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1306 NULL, 0, NULL, &lockh);
1310 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
1312 struct inode *inode = dentry->d_inode;
1313 struct ll_inode_info *lli;
1314 struct lov_stripe_md *lsm;
1320 CERROR("REPORT THIS LINE TO PETER\n");
1323 ll_inode2fid(&fid, inode);
1324 lli = ll_i2info(inode);
1325 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s,intent=%s\n",
1326 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1328 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1329 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1332 if (!md_valid_attrs(ll_i2mdcexp(inode), &fid)) {
1333 struct ptlrpc_request *req = NULL;
1334 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
1336 unsigned long valid = 0;
1339 if (S_ISREG(inode->i_mode)) {
1340 ealen = obd_size_diskmd(sbi->ll_osc_exp, NULL);
1341 valid |= OBD_MD_FLEASIZE;
1343 ll_inode2fid(&fid, inode);
1344 rc = md_getattr(sbi->ll_mdc_exp, &fid, valid, ealen, &req);
1346 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
1349 rc = ll_prep_inode(sbi->ll_osc_exp, sbi->ll_mdc_exp,
1350 &inode, req, 0, NULL);
1352 ptlrpc_req_finished(req);
1355 ptlrpc_req_finished(req);
1359 if (lsm == NULL) /* object not yet allocated, don't validate size */
1362 /* ll_glimpse_size will prefer locally cached writes if they extend
1367 rc = ll_glimpse_size(inode, &lvb);
1368 inode->i_size = lvb.lvb_size;
1373 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1374 int ll_getattr(struct vfsmount *mnt, struct dentry *de,
1375 struct lookup_intent *it, struct kstat *stat)
1378 struct inode *inode = de->d_inode;
1380 res = ll_inode_revalidate_it(de, it);
1381 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1386 stat->dev = inode->i_sb->s_dev;
1387 stat->ino = inode->i_ino;
1388 stat->mode = inode->i_mode;
1389 stat->nlink = inode->i_nlink;
1390 stat->uid = inode->i_uid;
1391 stat->gid = inode->i_gid;
1392 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1393 stat->atime = inode->i_atime;
1394 stat->mtime = inode->i_mtime;
1395 stat->ctime = inode->i_ctime;
1396 stat->size = inode->i_size;
1397 stat->blksize = inode->i_blksize;
1398 stat->blocks = inode->i_blocks;
1403 struct file_operations ll_file_operations = {
1404 .read = ll_file_read,
1405 .write = ll_file_write,
1406 .ioctl = ll_file_ioctl,
1407 .open = ll_file_open,
1408 .release = ll_file_release,
1409 .mmap = generic_file_mmap,
1410 .llseek = ll_file_seek,
1411 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
1412 .sendfile = generic_file_sendfile,
1415 //.lock ll_file_flock
1418 struct inode_operations ll_file_inode_operations = {
1419 .setattr_raw = ll_setattr_raw,
1420 .setattr = ll_setattr,
1421 .truncate = ll_truncate,
1422 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1423 .getattr_it = ll_getattr,
1425 .revalidate_it = ll_inode_revalidate_it,