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"
34 #include <linux/obd_lov.h>
36 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;
46 /* clear group lock, if present */
47 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
48 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
49 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
50 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
54 obdo.o_id = inode->i_ino;
55 obdo.o_valid = OBD_MD_FLID;
56 obdo_from_inode(&obdo, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
57 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
58 OBD_MD_FLATIME | OBD_MD_FLMTIME |
60 if (0 /* ll_is_inode_dirty(inode) */) {
61 obdo.o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
62 obdo.o_valid |= OBD_MD_FLFLAGS;
64 obdo.o_mds = ll_i2info(inode)->lli_mds;
65 rc = md_close(mdc_exp, &obdo, och, &req);
68 /* We are the last writer, so the MDS has instructed us to get
69 * the file size and any write cookies, then close again. */
70 //ll_queue_done_writing(inode);
73 CERROR("inode %lu mdc close failed: rc = %d\n",
77 rc = ll_objects_destroy(req, file->f_dentry->d_inode);
79 CERROR("inode %lu ll_objects destroy: rc = %d\n",
83 mdc_clear_open_replay_data(mdc_exp, och);
84 ptlrpc_req_finished(req);
85 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
86 file->private_data = NULL;
87 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof *fd);
92 /* While this returns an error code, fput() the caller does not, so we need
93 * to make every effort to clean up all of our state here. Also, applications
94 * rarely check close errors and even if an error is returned they will not
95 * re-try the close call.
97 int ll_file_release(struct inode *inode, struct file *file)
99 struct ll_file_data *fd;
100 struct ll_sb_info *sbi = ll_i2sbi(inode);
104 CDEBUG(D_VFSTRACE, "VFS Op:inode=%u/%lu/%u(%p)\n",
105 ll_i2info(inode)->lli_mds, inode->i_ino,
106 inode->i_generation, inode);
108 /* don't do anything for / */
109 if (inode->i_sb->s_root == file->f_dentry)
112 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
113 fd = (struct ll_file_data *)file->private_data;
116 rc = ll_mdc_close(sbi->ll_mdc_exp, inode, file);
120 static int ll_intent_file_open(struct file *file, void *lmm,
121 int lmmsize, struct lookup_intent *itp)
123 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
124 struct lustre_handle lockh;
125 struct mdc_op_data data;
126 struct dentry *parent = file->f_dentry->d_parent;
127 const char *name = file->f_dentry->d_name.name;
128 const int len = file->f_dentry->d_name.len;
134 ll_prepare_mdc_op_data(&data, parent->d_inode, NULL, name, len, O_RDWR);
136 rc = md_enqueue(sbi->ll_mdc_exp, LDLM_IBITS, itp, LCK_PR, &data,
137 &lockh, lmm, lmmsize, ldlm_completion_ast,
138 ll_mdc_blocking_ast, NULL);
140 if (itp->d.lustre.it_lock_mode)
141 memcpy(&itp->d.lustre.it_lock_handle,
142 &lockh, sizeof(lockh));
144 CERROR("lock enqueue: err: %d\n", rc);
150 int ll_local_open(struct file *file, struct lookup_intent *it)
152 struct ptlrpc_request *req = it->d.lustre.it_data;
153 struct ll_inode_info *lli = ll_i2info(file->f_dentry->d_inode);
154 struct obd_export *mdc_exp = ll_i2mdcexp(file->f_dentry->d_inode);
155 struct ll_file_data *fd;
156 struct mds_body *body;
159 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
160 LASSERT (body != NULL); /* reply already checked out */
161 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
163 LASSERT(!file->private_data);
165 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
166 /* We can't handle this well without reorganizing ll_file_open and
167 * ll_mdc_close, so don't even try right now. */
170 memcpy(&fd->fd_mds_och.och_fh, &body->handle, sizeof(body->handle));
171 fd->fd_mds_och.och_magic = OBD_CLIENT_HANDLE_MAGIC;
172 file->private_data = fd;
173 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
175 lli->lli_io_epoch = body->io_epoch;
177 mdc_set_open_replay_data(mdc_exp, &fd->fd_mds_och, it->d.lustre.it_data);
182 /* Open a file, and (for the very first open) create objects on the OSTs at
183 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
184 * creation or open until ll_lov_setstripe() ioctl is called. We grab
185 * lli_open_sem to ensure no other process will create objects, send the
186 * stripe MD to the MDS, or try to destroy the objects if that fails.
188 * If we already have the stripe MD locally then we don't request it in
189 * mdc_open(), by passing a lmm_size = 0.
191 * It is up to the application to ensure no other processes open this file
192 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
193 * used. We might be able to avoid races of that sort by getting lli_open_sem
194 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
195 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
197 int ll_file_open(struct inode *inode, struct file *file)
199 struct ll_inode_info *lli = ll_i2info(inode);
200 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
201 .it_flags = file->f_flags };
202 struct lov_stripe_md *lsm;
203 struct ptlrpc_request *req;
207 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
208 inode->i_generation, inode);
210 /* don't do anything for / */
211 if (inode->i_sb->s_root == file->f_dentry)
216 if (!it || !it->d.lustre.it_disposition) {
218 rc = ll_intent_file_open(file, NULL, 0, it);
223 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
224 rc = it_open_error(DISP_OPEN_OPEN, it);
228 rc = ll_local_open(file, it);
232 ll_intent_drop_lock(it);
234 if (!S_ISREG(inode->i_mode))
239 if (file->f_flags & O_LOV_DELAY_CREATE ||
240 !(file->f_mode & FMODE_WRITE)) {
241 CDEBUG(D_INODE, "object creation was delayed\n");
245 file->f_flags &= ~O_LOV_DELAY_CREATE;
248 req = it->d.lustre.it_data;
249 ptlrpc_req_finished(req);
251 ll_open_complete(inode);
255 /* Fills the obdo with the attributes for the inode defined by lsm */
256 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
259 struct ptlrpc_request_set *set;
263 LASSERT(lsm != NULL);
265 memset(oa, 0, sizeof *oa);
266 oa->o_id = lsm->lsm_object_id;
267 oa->o_gr = lsm->lsm_object_gr;
268 oa->o_mode = S_IFREG;
269 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
270 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
271 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
273 set = ptlrpc_prep_set();
275 CERROR ("ENOMEM allocing request set\n");
278 rc = obd_getattr_async(exp, oa, lsm, set);
280 rc = ptlrpc_set_wait(set);
281 ptlrpc_set_destroy(set);
286 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
287 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
291 static inline void ll_remove_suid(struct inode *inode)
295 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
296 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
298 /* was any of the uid bits set? */
299 mode &= inode->i_mode;
300 if (mode && !capable(CAP_FSETID)) {
301 inode->i_mode &= ~mode;
302 // XXX careful here - we cannot change the size
306 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
308 struct ll_inode_info *lli = ll_i2info(inode);
309 struct lov_stripe_md *lsm = lli->lli_smd;
310 struct obd_export *exp = ll_i2obdexp(inode);
313 struct ldlm_lock *lock;
314 struct lov_stripe_md *lsm;
315 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
316 __u32 stripe, vallen = sizeof(stripe);
320 if (lsm->lsm_stripe_count == 1)
323 /* get our offset in the lov */
324 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
326 CERROR("obd_get_info: rc = %d\n", rc);
329 LASSERT(stripe < lsm->lsm_stripe_count);
333 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
334 * we get a lock cancellation for each stripe, so we have to map the obd's
335 * region back onto the stripes in the file that it held.
337 * No one can dirty the extent until we've finished our work and they can
338 * enqueue another lock. The DLM protects us from ll_file_read/write here,
339 * but other kernel actors could have pages locked.
341 * Called with the DLM lock held. */
342 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
343 struct ldlm_lock *lock, __u32 stripe)
345 ldlm_policy_data_t tmpex;
346 unsigned long start, end, count, skip, i, j;
348 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
349 struct lustre_handle lockh;
352 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
353 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
354 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
357 /* our locks are page granular thanks to osc_enqueue, we invalidate the
359 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
360 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
364 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
365 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
366 if (lsm->lsm_stripe_count > 1) {
367 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
368 skip = (lsm->lsm_stripe_count - 1) * count;
369 start += start/count * skip + stripe * count;
371 end += end/count * skip + stripe * count;
373 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
376 i = (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
380 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
381 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
382 count, skip, end, discard ? " (DISCARDING)" : "");
384 /* this is the simplistic implementation of page eviction at
385 * cancelation. It is careful to get races with other page
386 * lockers handled correctly. fixes from bug 20 will make it
387 * more efficient by associating locks with pages and with
388 * batching writeback under the lock explicitly. */
389 for (i = start, j = start % count; i <= end;
390 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
392 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
398 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
399 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
400 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
403 if (!mapping_has_pages(inode->i_mapping)) {
404 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
408 conditional_schedule();
410 page = find_get_page(inode->i_mapping, i);
413 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
414 i, tmpex.l_extent.start);
417 /* page->mapping to check with racing against teardown */
418 if (!discard && clear_page_dirty_for_io(page)) {
419 rc = ll_call_writepage(inode, page);
421 CERROR("writepage of page %p failed: %d\n",
423 /* either waiting for io to complete or reacquiring
424 * the lock that the failed writepage released */
428 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
429 /* check to see if another DLM lock covers this page */
430 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
431 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
433 &lock->l_resource->lr_name, LDLM_EXTENT,
434 &tmpex, LCK_PR | LCK_PW, &lockh);
435 if (rc2 == 0 && page->mapping != NULL) {
436 // checking again to account for writeback's lock_page()
437 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
438 ll_truncate_complete_page(page);
441 page_cache_release(page);
443 LASSERTF(tmpex.l_extent.start <=
444 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
445 lock->l_policy_data.l_extent.end + 1),
446 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
447 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
452 static int ll_extent_lock_callback(struct ldlm_lock *lock,
453 struct ldlm_lock_desc *new, void *data,
456 struct lustre_handle lockh = { 0 };
460 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
461 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
466 case LDLM_CB_BLOCKING:
467 ldlm_lock2handle(lock, &lockh);
468 rc = ldlm_cli_cancel(&lockh);
470 CERROR("ldlm_cli_cancel failed: %d\n", rc);
472 case LDLM_CB_CANCELING: {
474 struct ll_inode_info *lli;
475 struct lov_stripe_md *lsm;
479 /* This lock wasn't granted, don't try to evict pages */
480 if (lock->l_req_mode != lock->l_granted_mode)
483 inode = ll_inode_from_lock(lock);
486 lli = ll_i2info(inode);
489 if (lli->lli_smd == NULL)
493 stripe = ll_lock_to_stripe_offset(inode, lock);
494 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
497 kms = ldlm_extent_shift_kms(lock,
498 lsm->lsm_oinfo[stripe].loi_kms);
500 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
501 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
502 lsm->lsm_oinfo[stripe].loi_kms, kms);
503 lsm->lsm_oinfo[stripe].loi_kms = kms;
505 //ll_try_done_writing(inode);
518 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
520 /* XXX ALLOCATE - 160 bytes */
521 struct inode *inode = ll_inode_from_lock(lock);
522 struct ll_inode_info *lli = ll_i2info(inode);
523 struct lustre_handle lockh = { 0 };
528 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
529 LDLM_FL_BLOCK_CONV)) {
530 LBUG(); /* not expecting any blocked async locks yet */
531 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
533 ldlm_lock_dump(D_OTHER, lock, 0);
534 ldlm_reprocess_all(lock->l_resource);
538 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
540 stripe = ll_lock_to_stripe_offset(inode, lock);
542 if (lock->l_lvb_len) {
543 struct lov_stripe_md *lsm = lli->lli_smd;
545 lvb = lock->l_lvb_data;
546 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
549 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
550 kms = ldlm_extent_shift_kms(NULL, kms);
551 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
552 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
553 lsm->lsm_oinfo[stripe].loi_kms, kms);
554 lsm->lsm_oinfo[stripe].loi_kms = kms;
559 wake_up(&lock->l_waitq);
561 ldlm_lock2handle(lock, &lockh);
562 ldlm_lock_decref(&lockh, LCK_PR);
567 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
569 struct ptlrpc_request *req = reqp;
570 struct inode *inode = ll_inode_from_lock(lock);
571 struct ll_inode_info *lli;
573 int rc, size = sizeof(*lvb), stripe = 0;
577 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
578 lli = ll_i2info(inode);
580 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
581 if (lli->lli_smd == NULL)
582 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
584 /* First, find out which stripe index this lock corresponds to. */
585 if (lli->lli_smd->lsm_stripe_count > 1)
586 stripe = ll_lock_to_stripe_offset(inode, lock);
588 rc = lustre_pack_reply(req, 1, &size, NULL);
590 CERROR("lustre_pack_reply: %d\n", rc);
594 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
595 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
597 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
598 inode->i_size, stripe, lvb->lvb_size);
604 /* These errors are normal races, so we don't want to fill the console
605 * with messages by calling ptlrpc_error() */
606 if (rc == -ELDLM_NO_LOCK_DATA)
607 lustre_pack_reply(req, 0, NULL, NULL);
613 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
614 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
615 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
617 /* NB: lov_merge_size will prefer locally cached writes if they extend the
618 * file (because it prefers KMS over RSS when larger) */
619 int ll_glimpse_size(struct inode *inode, struct ost_lvb *lvb)
621 struct ll_inode_info *lli = ll_i2info(inode);
622 struct ll_sb_info *sbi = ll_i2sbi(inode);
623 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
624 struct lustre_handle lockh = { 0 };
625 int rc, flags = LDLM_FL_HAS_INTENT;
628 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
630 rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
631 LCK_PR, &flags, ll_extent_lock_callback,
632 ldlm_completion_ast, ll_glimpse_callback, inode,
633 sizeof(*lvb), lustre_swab_ost_lvb, &lockh);
635 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
636 RETURN(rc > 0 ? -EIO : rc);
639 lvb->lvb_size = lov_merge_size(lli->lli_smd, 0);
640 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
641 //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
643 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
644 lvb->lvb_size, lvb->lvb_blocks);
646 obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
651 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
652 struct lov_stripe_md *lsm, int mode,
653 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
656 struct ll_sb_info *sbi = ll_i2sbi(inode);
660 LASSERT(lockh->cookie == 0);
662 /* XXX phil: can we do this? won't it screw the file size up? */
663 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
664 (sbi->ll_flags & LL_SBI_NOLCK))
667 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
668 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
670 rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
671 &ast_flags, ll_extent_lock_callback,
672 ldlm_completion_ast, ll_glimpse_callback, inode,
673 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
677 if (policy->l_extent.start == 0 &&
678 policy->l_extent.end == OBD_OBJECT_EOF)
679 inode->i_size = lov_merge_size(lsm, 1);
681 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
686 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
687 struct lov_stripe_md *lsm, int mode,
688 struct lustre_handle *lockh)
690 struct ll_sb_info *sbi = ll_i2sbi(inode);
694 /* XXX phil: can we do this? won't it screw the file size up? */
695 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
696 (sbi->ll_flags & LL_SBI_NOLCK))
699 rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
704 static ssize_t ll_file_read(struct file *filp, char *buf, size_t count,
707 struct ll_file_data *fd = filp->private_data;
708 struct inode *inode = filp->f_dentry->d_inode;
709 struct ll_inode_info *lli = ll_i2info(inode);
710 struct lov_stripe_md *lsm = lli->lli_smd;
711 struct lustre_handle lockh = { 0 };
712 ldlm_policy_data_t policy;
717 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
718 inode->i_ino, inode->i_generation, inode, count, *ppos);
720 /* "If nbyte is 0, read() will return 0 and have no other results."
721 * -- Single Unix Spec */
725 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
731 policy.l_extent.start = *ppos;
732 policy.l_extent.end = *ppos + count - 1;
734 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
735 (filp->f_flags & O_NONBLOCK) ?
736 LDLM_FL_BLOCK_NOWAIT: 0);
740 kms = lov_merge_size(lsm, 1);
741 if (*ppos + count - 1 > kms) {
742 /* A glimpse is necessary to determine whether we return a short
743 * read or some zeroes at the end of the buffer */
745 retval = ll_glimpse_size(inode, &lvb);
748 inode->i_size = lvb.lvb_size;
753 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
754 inode->i_ino, count, *ppos, inode->i_size);
756 /* turn off the kernel's read-ahead */
757 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
760 filp->f_ra.ra_pages = 0;
762 retval = generic_file_read(filp, buf, count, ppos);
765 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
770 * Write to a file (through the page cache).
772 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
775 struct ll_file_data *fd = file->private_data;
776 struct inode *inode = file->f_dentry->d_inode;
777 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
778 struct lustre_handle lockh = { 0 };
779 ldlm_policy_data_t policy;
780 loff_t maxbytes = ll_file_maxbytes(inode);
782 int nonblock = 0, rc;
784 if (file->f_flags & O_NONBLOCK)
785 nonblock = LDLM_FL_BLOCK_NOWAIT;
786 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
787 inode->i_ino, inode->i_generation, inode, count, *ppos);
789 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
791 /* POSIX, but surprised the VFS doesn't check this already */
795 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
796 * called on the file, don't fail the below assertion (bug 2388). */
797 if (file->f_flags & O_LOV_DELAY_CREATE && lsm == NULL)
802 if (file->f_flags & O_APPEND) {
803 policy.l_extent.start = 0;
804 policy.l_extent.end = OBD_OBJECT_EOF;
806 policy.l_extent.start = *ppos;
807 policy.l_extent.end = *ppos + count - 1;
810 rc = ll_extent_lock(fd, inode, lsm, LCK_PW, &policy, &lockh, nonblock);
814 /* this is ok, g_f_w will overwrite this under i_sem if it races
815 * with a local truncate, it just makes our maxbyte checking easier */
816 if (file->f_flags & O_APPEND)
817 *ppos = inode->i_size;
819 if (*ppos >= maxbytes) {
820 if (count || *ppos > maxbytes) {
821 send_sig(SIGXFSZ, current, 0);
822 GOTO(out, retval = -EFBIG);
825 if (*ppos + count > maxbytes)
826 count = maxbytes - *ppos;
828 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
829 inode->i_ino, count, *ppos);
831 /* generic_file_write handles O_APPEND after getting i_sem */
832 retval = generic_file_write(file, buf, count, ppos);
835 ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
836 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
837 retval > 0 ? retval : 0);
841 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
844 struct ll_inode_info *lli = ll_i2info(inode);
845 struct obd_export *exp = ll_i2obdexp(inode);
846 struct ll_recreate_obj ucreatp;
847 struct obd_trans_info oti = { 0 };
848 struct obdo *oa = NULL;
851 struct lov_stripe_md *lsm, *lsm2;
854 if (!capable (CAP_SYS_ADMIN))
857 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
858 sizeof(struct ll_recreate_obj));
867 down(&lli->lli_open_sem);
870 up(&lli->lli_open_sem);
874 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
875 (lsm->lsm_stripe_count));
877 OBD_ALLOC(lsm2, lsm_size);
879 up(&lli->lli_open_sem);
884 oa->o_id = ucreatp.lrc_id;
885 oa->o_nlink = ucreatp.lrc_ost_idx;
886 oa->o_gr = ucreatp.lrc_group;
887 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLFLAGS;
888 oa->o_flags |= OBD_FL_RECREATE_OBJS;
889 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
890 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
892 oti.oti_objid = NULL;
893 memcpy(lsm2, lsm, lsm_size);
894 rc = obd_create(exp, oa, &lsm2, &oti);
896 up(&lli->lli_open_sem);
897 OBD_FREE(lsm2, lsm_size);
902 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
903 int flags, struct lov_user_md *lum,
906 struct ll_inode_info *lli = ll_i2info(inode);
908 struct obd_export *exp = ll_i2obdexp(inode);
909 struct lov_stripe_md *lsm;
910 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
911 struct ptlrpc_request *req = NULL;
916 down(&lli->lli_open_sem);
919 up(&lli->lli_open_sem);
920 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
925 f = get_empty_filp();
929 f->f_dentry = file->f_dentry;
930 f->f_vfsmnt = file->f_vfsmnt;
932 rc = ll_intent_file_open(f, lum, lum_size, &oit);
935 if (it_disposition(&oit, DISP_LOOKUP_NEG))
937 req = oit.d.lustre.it_data;
938 rc = oit.d.lustre.it_status;
943 rc = mdc_req2lustre_md(ll_i2mdcexp(inode), req, 1, exp, &md);
946 ll_update_inode(f->f_dentry->d_inode, &md);
948 rc = ll_local_open(f, &oit);
951 ll_intent_release(&oit);
953 rc = ll_file_release(f->f_dentry->d_inode, f);
958 up(&lli->lli_open_sem);
960 ptlrpc_req_finished(req);
964 static int ll_lov_setea(struct inode *inode, struct file *file,
967 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
968 struct lov_user_md *lump;
969 int lum_size = sizeof(struct lov_user_md) +
970 sizeof(struct lov_user_ost_data);
974 if (!capable (CAP_SYS_ADMIN))
977 OBD_ALLOC(lump, lum_size);
981 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
983 OBD_FREE(lump, lum_size);
987 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
989 OBD_FREE(lump, lum_size);
993 static int ll_lov_setstripe(struct inode *inode, struct file *file,
996 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
998 int flags = FMODE_WRITE;
1001 /* Bug 1152: copy properly when this is no longer true */
1002 LASSERT(sizeof(lum) == sizeof(*lump));
1003 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1004 rc = copy_from_user(&lum, lump, sizeof(lum));
1008 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1012 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1014 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1019 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2obdexp(inode), 0, lsm,
1023 static int ll_get_grouplock(struct inode *inode, struct file *file,
1026 struct ll_file_data *fd = file->private_data;
1027 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1028 .end = OBD_OBJECT_EOF}};
1029 struct lustre_handle lockh = { 0 };
1030 struct ll_inode_info *lli = ll_i2info(inode);
1031 struct lov_stripe_md *lsm = lli->lli_smd;
1035 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1039 policy.l_extent.gid = arg;
1040 if (file->f_flags & O_NONBLOCK)
1041 flags = LDLM_FL_BLOCK_NOWAIT;
1043 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags);
1047 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1049 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1054 static int ll_put_grouplock(struct inode *inode, struct file *file,
1057 struct ll_file_data *fd = file->private_data;
1058 struct ll_inode_info *lli = ll_i2info(inode);
1059 struct lov_stripe_md *lsm = lli->lli_smd;
1063 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1064 /* Ugh, it's already unlocked. */
1068 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1071 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1073 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1078 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1083 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1086 struct ll_file_data *fd = file->private_data;
1090 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1091 inode->i_generation, inode, cmd);
1093 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1096 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1098 case LL_IOC_GETFLAGS:
1099 /* Get the current value of the file flags */
1100 return put_user(fd->fd_flags, (int *)arg);
1101 case LL_IOC_SETFLAGS:
1102 case LL_IOC_CLRFLAGS:
1103 /* Set or clear specific file flags */
1104 /* XXX This probably needs checks to ensure the flags are
1105 * not abused, and to handle any flag side effects.
1107 if (get_user(flags, (int *) arg))
1110 if (cmd == LL_IOC_SETFLAGS)
1111 fd->fd_flags |= flags;
1113 fd->fd_flags &= ~flags;
1115 case LL_IOC_LOV_SETSTRIPE:
1116 RETURN(ll_lov_setstripe(inode, file, arg));
1117 case LL_IOC_LOV_SETEA:
1118 RETURN(ll_lov_setea(inode, file, arg));
1119 case LL_IOC_LOV_GETSTRIPE:
1120 RETURN(ll_lov_getstripe(inode, arg));
1121 case LL_IOC_RECREATE_OBJ:
1122 RETURN(ll_lov_recreate_obj(inode, file, arg));
1123 case EXT3_IOC_GETFLAGS:
1124 case EXT3_IOC_SETFLAGS:
1125 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1126 case LL_IOC_GROUP_LOCK:
1127 RETURN(ll_get_grouplock(inode, file, arg));
1128 case LL_IOC_GROUP_UNLOCK:
1129 RETURN(ll_put_grouplock(inode, file, arg));
1130 /* We need to special case any other ioctls we want to handle,
1131 * to send them to the MDS/OST as appropriate and to properly
1132 * network encode the arg field.
1133 case EXT2_IOC_GETVERSION_OLD:
1134 case EXT2_IOC_GETVERSION_NEW:
1135 case EXT2_IOC_SETVERSION_OLD:
1136 case EXT2_IOC_SETVERSION_NEW:
1139 RETURN( obd_iocontrol(cmd, ll_i2obdexp(inode), 0, NULL,
1144 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1146 struct inode *inode = file->f_dentry->d_inode;
1147 struct ll_file_data *fd = file->private_data;
1148 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1149 struct lustre_handle lockh = {0};
1152 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1153 inode->i_generation, inode,
1154 offset + ((origin==2) ? inode->i_size : file->f_pos));
1156 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1157 if (origin == 2) { /* SEEK_END */
1158 int nonblock = 0, rc;
1159 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1161 if (file->f_flags & O_NONBLOCK)
1162 nonblock = LDLM_FL_BLOCK_NOWAIT;
1164 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1169 offset += inode->i_size;
1170 } else if (origin == 1) { /* SEEK_CUR */
1171 offset += file->f_pos;
1175 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1176 if (offset != file->f_pos) {
1177 file->f_pos = offset;
1178 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1180 file->f_version = ++event;
1187 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1191 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1193 struct inode *inode = dentry->d_inode;
1194 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1196 struct ptlrpc_request *req;
1199 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1200 inode->i_generation, inode);
1202 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1204 /* fsync's caller has already called _fdata{sync,write}, we want
1205 * that IO to finish before calling the osc and mdc sync methods */
1206 rc = filemap_fdatawait(inode->i_mapping);
1208 ll_inode2fid(&fid, inode);
1209 err = md_sync(ll_i2sbi(inode)->ll_mdc_exp, &fid, &req);
1213 ptlrpc_req_finished(req);
1216 struct obdo *oa = obdo_alloc();
1219 RETURN(rc ? rc : -ENOMEM);
1221 oa->o_id = lsm->lsm_object_id;
1222 oa->o_gr = lsm->lsm_object_gr;
1223 oa->o_valid = OBD_MD_FLID;
1224 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
1225 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1228 err = obd_sync(ll_i2sbi(inode)->ll_osc_exp, oa, lsm,
1238 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1240 struct inode *inode = file->f_dentry->d_inode;
1241 struct ll_sb_info *sbi = ll_i2sbi(inode);
1242 struct obd_device *obddev;
1243 struct ldlm_res_id res_id =
1244 { .name = {inode->i_ino, inode->i_generation, LDLM_FLOCK} };
1245 struct lustre_handle lockh = {0};
1246 ldlm_policy_data_t flock;
1247 ldlm_mode_t mode = 0;
1252 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1253 inode->i_ino, file_lock);
1255 flock.l_flock.pid = file_lock->fl_pid;
1256 flock.l_flock.start = file_lock->fl_start;
1257 flock.l_flock.end = file_lock->fl_end;
1259 switch (file_lock->fl_type) {
1264 /* An unlock request may or may not have any relation to
1265 * existing locks so we may not be able to pass a lock handle
1266 * via a normal ldlm_lock_cancel() request. The request may even
1267 * unlock a byte range in the middle of an existing lock. In
1268 * order to process an unlock request we need all of the same
1269 * information that is given with a normal read or write record
1270 * lock request. To avoid creating another ldlm unlock (cancel)
1271 * message we'll treat a LCK_NL flock request as an unlock. */
1278 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1287 flags = LDLM_FL_BLOCK_NOWAIT;
1290 flags = LDLM_FL_TEST_LOCK;
1291 /* Save the old mode so that if the mode in the lock changes we
1292 * can decrement the appropriate reader or writer refcount. */
1293 file_lock->fl_type = mode;
1296 CERROR("unknown fcntl lock command: %d\n", cmd);
1300 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1301 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1302 flags, mode, flock.l_flock.start, flock.l_flock.end);
1304 obddev = md_get_real_obd(sbi->ll_mdc_exp, NULL, 0);
1305 rc = ldlm_cli_enqueue(sbi->ll_mdc_exp, NULL, obddev->obd_namespace,
1306 res_id, LDLM_FLOCK, &flock, mode, &flags,
1307 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1308 NULL, 0, NULL, &lockh);
1312 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
1314 struct inode *inode = dentry->d_inode;
1315 struct ll_inode_info *lli;
1316 struct lov_stripe_md *lsm;
1322 CERROR("REPORT THIS LINE TO PETER\n");
1325 ll_inode2fid(&fid, inode);
1326 lli = ll_i2info(inode);
1327 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s,intent=%s\n",
1328 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1330 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1331 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1334 if (!md_valid_attrs(ll_i2mdcexp(inode), &fid)) {
1335 struct ptlrpc_request *req = NULL;
1336 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
1338 unsigned long valid = 0;
1341 if (S_ISREG(inode->i_mode)) {
1342 ealen = obd_size_diskmd(sbi->ll_osc_exp, NULL);
1343 valid |= OBD_MD_FLEASIZE;
1345 ll_inode2fid(&fid, inode);
1346 rc = md_getattr(sbi->ll_mdc_exp, &fid, valid, ealen, &req);
1348 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
1351 rc = ll_prep_inode(sbi->ll_osc_exp, sbi->ll_mdc_exp,
1352 &inode, req, 0, NULL);
1354 ptlrpc_req_finished(req);
1357 ptlrpc_req_finished(req);
1361 if (lsm == NULL) /* object not yet allocated, don't validate size */
1364 /* ll_glimpse_size will prefer locally cached writes if they extend
1369 rc = ll_glimpse_size(inode, &lvb);
1370 inode->i_size = lvb.lvb_size;
1375 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1376 int ll_getattr(struct vfsmount *mnt, struct dentry *de,
1377 struct lookup_intent *it, struct kstat *stat)
1380 struct inode *inode = de->d_inode;
1382 res = ll_inode_revalidate_it(de, it);
1383 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1388 stat->dev = inode->i_sb->s_dev;
1389 stat->ino = inode->i_ino;
1390 stat->mode = inode->i_mode;
1391 stat->nlink = inode->i_nlink;
1392 stat->uid = inode->i_uid;
1393 stat->gid = inode->i_gid;
1394 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1395 stat->atime = inode->i_atime;
1396 stat->mtime = inode->i_mtime;
1397 stat->ctime = inode->i_ctime;
1398 stat->size = inode->i_size;
1399 stat->blksize = inode->i_blksize;
1400 stat->blocks = inode->i_blocks;
1405 struct file_operations ll_file_operations = {
1406 .read = ll_file_read,
1407 .write = ll_file_write,
1408 .ioctl = ll_file_ioctl,
1409 .open = ll_file_open,
1410 .release = ll_file_release,
1411 .mmap = generic_file_mmap,
1412 .llseek = ll_file_seek,
1413 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
1414 .sendfile = generic_file_sendfile,
1417 //.lock ll_file_flock
1420 struct inode_operations ll_file_inode_operations = {
1421 .setattr_raw = ll_setattr_raw,
1422 .setattr = ll_setattr,
1423 .truncate = ll_truncate,
1424 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1425 .getattr_it = ll_getattr,
1427 .revalidate_it = ll_inode_revalidate_it,