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 rc = mdc_close(mdc_exp, &obdo, och, &req);
65 /* We are the last writer, so the MDS has instructed us to get
66 * the file size and any write cookies, then close again. */
67 //ll_queue_done_writing(inode);
70 CERROR("inode %lu mdc close failed: rc = %d\n",
74 rc = ll_objects_destroy(req, file->f_dentry->d_inode);
76 CERROR("inode %lu ll_objects destroy: rc = %d\n",
80 mdc_clear_open_replay_data(och);
81 ptlrpc_req_finished(req);
82 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
83 file->private_data = NULL;
84 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof *fd);
89 /* While this returns an error code, fput() the caller does not, so we need
90 * to make every effort to clean up all of our state here. Also, applications
91 * rarely check close errors and even if an error is returned they will not
92 * re-try the close call.
94 int ll_file_release(struct inode *inode, struct file *file)
96 struct ll_file_data *fd;
97 struct ll_sb_info *sbi = ll_i2sbi(inode);
101 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
102 inode->i_generation, inode);
104 /* don't do anything for / */
105 if (inode->i_sb->s_root == file->f_dentry)
108 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
109 fd = (struct ll_file_data *)file->private_data;
112 rc = ll_mdc_close(sbi->ll_mdc_exp, inode, file);
116 static int ll_intent_file_open(struct file *file, void *lmm,
117 int lmmsize, struct lookup_intent *itp)
119 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
120 struct lustre_handle lockh;
121 struct mdc_op_data data;
122 struct dentry *parent = file->f_dentry->d_parent;
123 const char *name = file->f_dentry->d_name.name;
124 const int len = file->f_dentry->d_name.len;
130 ll_prepare_mdc_op_data(&data, parent->d_inode, NULL, name, len, O_RDWR);
132 rc = mdc_enqueue(sbi->ll_mdc_exp, LDLM_IBITS, itp, LCK_PW, &data,
133 &lockh, lmm, lmmsize, ldlm_completion_ast,
134 ll_mdc_blocking_ast, parent->d_inode);
136 CERROR("lock enqueue: err: %d\n", rc);
140 int ll_local_open(struct file *file, struct lookup_intent *it)
142 struct ptlrpc_request *req = it->d.lustre.it_data;
143 struct ll_inode_info *lli = ll_i2info(file->f_dentry->d_inode);
144 struct ll_file_data *fd;
145 struct mds_body *body;
148 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
149 LASSERT (body != NULL); /* reply already checked out */
150 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
152 LASSERT(!file->private_data);
154 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
155 /* We can't handle this well without reorganizing ll_file_open and
156 * ll_mdc_close, so don't even try right now. */
159 memcpy(&fd->fd_mds_och.och_fh, &body->handle, sizeof(body->handle));
160 fd->fd_mds_och.och_magic = OBD_CLIENT_HANDLE_MAGIC;
161 file->private_data = fd;
162 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
164 lli->lli_io_epoch = body->io_epoch;
166 mdc_set_open_replay_data(&fd->fd_mds_och, it->d.lustre.it_data);
171 /* Open a file, and (for the very first open) create objects on the OSTs at
172 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
173 * creation or open until ll_lov_setstripe() ioctl is called. We grab
174 * lli_open_sem to ensure no other process will create objects, send the
175 * stripe MD to the MDS, or try to destroy the objects if that fails.
177 * If we already have the stripe MD locally then we don't request it in
178 * mdc_open(), by passing a lmm_size = 0.
180 * It is up to the application to ensure no other processes open this file
181 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
182 * used. We might be able to avoid races of that sort by getting lli_open_sem
183 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
184 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
186 int ll_file_open(struct inode *inode, struct file *file)
188 struct ll_inode_info *lli = ll_i2info(inode);
189 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
190 .it_flags = file->f_flags };
191 struct lov_stripe_md *lsm;
192 struct ptlrpc_request *req;
196 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
197 inode->i_generation, inode);
199 /* don't do anything for / */
200 if (inode->i_sb->s_root == file->f_dentry)
205 if (!it || !it->d.lustre.it_disposition) {
207 rc = ll_intent_file_open(file, NULL, 0, it);
212 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
213 rc = it_open_error(DISP_OPEN_OPEN, it);
217 rc = ll_local_open(file, it);
221 if (!S_ISREG(inode->i_mode))
226 if (file->f_flags & O_LOV_DELAY_CREATE ||
227 !(file->f_mode & FMODE_WRITE)) {
228 CDEBUG(D_INODE, "object creation was delayed\n");
232 file->f_flags &= ~O_LOV_DELAY_CREATE;
235 req = it->d.lustre.it_data;
236 ptlrpc_req_finished(req);
238 ll_open_complete(inode);
242 /* Fills the obdo with the attributes for the inode defined by lsm */
243 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
246 struct ptlrpc_request_set *set;
250 LASSERT(lsm != NULL);
252 memset(oa, 0, sizeof *oa);
253 oa->o_id = lsm->lsm_object_id;
254 oa->o_mode = S_IFREG;
255 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
256 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
259 set = ptlrpc_prep_set();
261 CERROR ("ENOMEM allocing request set\n");
264 rc = obd_getattr_async(exp, oa, lsm, set);
266 rc = ptlrpc_set_wait(set);
267 ptlrpc_set_destroy(set);
272 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
273 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
277 static inline void ll_remove_suid(struct inode *inode)
281 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
282 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
284 /* was any of the uid bits set? */
285 mode &= inode->i_mode;
286 if (mode && !capable(CAP_FSETID)) {
287 inode->i_mode &= ~mode;
288 // XXX careful here - we cannot change the size
292 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
294 struct ll_inode_info *lli = ll_i2info(inode);
295 struct lov_stripe_md *lsm = lli->lli_smd;
296 struct obd_export *exp = ll_i2obdexp(inode);
299 struct ldlm_lock *lock;
300 struct lov_stripe_md *lsm;
301 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
302 __u32 stripe, vallen = sizeof(stripe);
306 if (lsm->lsm_stripe_count == 1)
309 /* get our offset in the lov */
310 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
312 CERROR("obd_get_info: rc = %d\n", rc);
315 LASSERT(stripe < lsm->lsm_stripe_count);
319 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
320 * we get a lock cancellation for each stripe, so we have to map the obd's
321 * region back onto the stripes in the file that it held.
323 * No one can dirty the extent until we've finished our work and they can
324 * enqueue another lock. The DLM protects us from ll_file_read/write here,
325 * but other kernel actors could have pages locked.
327 * Called with the DLM lock held. */
328 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
329 struct ldlm_lock *lock, __u32 stripe)
331 ldlm_policy_data_t tmpex;
332 unsigned long start, end, count, skip, i, j;
334 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
335 struct lustre_handle lockh;
338 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
339 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
340 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
343 /* our locks are page granular thanks to osc_enqueue, we invalidate the
345 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
346 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
350 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
351 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
352 if (lsm->lsm_stripe_count > 1) {
353 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
354 skip = (lsm->lsm_stripe_count - 1) * count;
355 start += start/count * skip + stripe * count;
357 end += end/count * skip + stripe * count;
359 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
362 i = (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
366 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
367 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
368 count, skip, end, discard ? " (DISCARDING)" : "");
370 /* this is the simplistic implementation of page eviction at
371 * cancelation. It is careful to get races with other page
372 * lockers handled correctly. fixes from bug 20 will make it
373 * more efficient by associating locks with pages and with
374 * batching writeback under the lock explicitly. */
375 for (i = start, j = start % count; i <= end;
376 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
378 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
384 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
385 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
386 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
389 if (!mapping_has_pages(inode->i_mapping)) {
390 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
394 conditional_schedule();
396 page = find_get_page(inode->i_mapping, i);
399 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
400 i, tmpex.l_extent.start);
403 /* page->mapping to check with racing against teardown */
404 if (!discard && clear_page_dirty_for_io(page)) {
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 ldlm_lock2handle(lock, &lockh);
417 rc2 = ldlm_lock_match(NULL,
418 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
420 NULL, 0, &tmpex, 0, &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);
429 LASSERTF(tmpex.l_extent.start <=
430 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
431 lock->l_policy_data.l_extent.end + 1),
432 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
433 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
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);
486 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
487 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
488 lsm->lsm_oinfo[stripe].loi_kms, kms);
489 lsm->lsm_oinfo[stripe].loi_kms = kms;
491 //ll_try_done_writing(inode);
504 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
506 /* XXX ALLOCATE - 160 bytes */
507 struct inode *inode = ll_inode_from_lock(lock);
508 struct ll_inode_info *lli = ll_i2info(inode);
509 struct lustre_handle lockh = { 0 };
514 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
515 LDLM_FL_BLOCK_CONV)) {
516 LBUG(); /* not expecting any blocked async locks yet */
517 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
519 ldlm_lock_dump(D_OTHER, lock, 0);
520 ldlm_reprocess_all(lock->l_resource);
524 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
526 stripe = ll_lock_to_stripe_offset(inode, lock);
528 if (lock->l_lvb_len) {
529 struct lov_stripe_md *lsm = lli->lli_smd;
531 lvb = lock->l_lvb_data;
532 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
535 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
536 kms = ldlm_extent_shift_kms(NULL, kms);
537 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
538 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
539 lsm->lsm_oinfo[stripe].loi_kms, kms);
540 lsm->lsm_oinfo[stripe].loi_kms = kms;
545 wake_up(&lock->l_waitq);
547 ldlm_lock2handle(lock, &lockh);
548 ldlm_lock_decref(&lockh, LCK_PR);
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 ll_inode_info *lli;
559 int rc, size = sizeof(*lvb), stripe = 0;
563 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
564 lli = ll_i2info(inode);
566 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
567 if (lli->lli_smd == NULL)
568 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
570 /* First, find out which stripe index this lock corresponds to. */
571 if (lli->lli_smd->lsm_stripe_count > 1)
572 stripe = ll_lock_to_stripe_offset(inode, lock);
574 rc = lustre_pack_reply(req, 1, &size, NULL);
576 CERROR("lustre_pack_reply: %d\n", rc);
580 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
581 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
583 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
584 inode->i_size, stripe, lvb->lvb_size);
590 /* These errors are normal races, so we don't want to fill the console
591 * with messages by calling ptlrpc_error() */
592 if (rc == -ELDLM_NO_LOCK_DATA)
593 lustre_pack_reply(req, 0, NULL, NULL);
599 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
600 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
601 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
603 /* NB: lov_merge_size will prefer locally cached writes if they extend the
604 * file (because it prefers KMS over RSS when larger) */
605 int ll_glimpse_size(struct inode *inode, struct ost_lvb *lvb)
607 struct ll_inode_info *lli = ll_i2info(inode);
608 struct ll_sb_info *sbi = ll_i2sbi(inode);
609 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
610 struct lustre_handle lockh = { 0 };
611 int rc, flags = LDLM_FL_HAS_INTENT;
614 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
616 rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
617 LCK_PR, &flags, ll_extent_lock_callback,
618 ldlm_completion_ast, ll_glimpse_callback, inode,
619 sizeof(*lvb), lustre_swab_ost_lvb, &lockh);
621 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
622 RETURN(rc > 0 ? -EIO : rc);
625 lvb->lvb_size = lov_merge_size(lli->lli_smd, 0);
626 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
627 //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
629 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
630 lvb->lvb_size, lvb->lvb_blocks);
632 obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
637 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
638 struct lov_stripe_md *lsm, int mode,
639 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
642 struct ll_sb_info *sbi = ll_i2sbi(inode);
646 LASSERT(lockh->cookie == 0);
648 /* XXX phil: can we do this? won't it screw the file size up? */
649 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
650 (sbi->ll_flags & LL_SBI_NOLCK))
653 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
654 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
656 rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
657 &ast_flags, ll_extent_lock_callback,
658 ldlm_completion_ast, ll_glimpse_callback, inode,
659 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
663 if (policy->l_extent.start == 0 &&
664 policy->l_extent.end == OBD_OBJECT_EOF)
665 inode->i_size = lov_merge_size(lsm, 1);
667 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
672 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
673 struct lov_stripe_md *lsm, int mode,
674 struct lustre_handle *lockh)
676 struct ll_sb_info *sbi = ll_i2sbi(inode);
680 /* XXX phil: can we do this? won't it screw the file size up? */
681 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
682 (sbi->ll_flags & LL_SBI_NOLCK))
685 rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
690 static ssize_t ll_file_read(struct file *filp, char *buf, size_t count,
693 struct ll_file_data *fd = filp->private_data;
694 struct inode *inode = filp->f_dentry->d_inode;
695 struct ll_inode_info *lli = ll_i2info(inode);
696 struct lov_stripe_md *lsm = lli->lli_smd;
697 struct lustre_handle lockh = { 0 };
698 ldlm_policy_data_t policy;
703 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
704 inode->i_ino, inode->i_generation, inode, count, *ppos);
706 /* "If nbyte is 0, read() will return 0 and have no other results."
707 * -- Single Unix Spec */
711 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
717 policy.l_extent.start = *ppos;
718 policy.l_extent.end = *ppos + count - 1;
720 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
721 (filp->f_flags & O_NONBLOCK) ?
722 LDLM_FL_BLOCK_NOWAIT: 0);
726 kms = lov_merge_size(lsm, 1);
727 if (*ppos + count - 1 > kms) {
728 /* A glimpse is necessary to determine whether we return a short
729 * read or some zeroes at the end of the buffer */
731 retval = ll_glimpse_size(inode, &lvb);
734 inode->i_size = lvb.lvb_size;
739 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
740 inode->i_ino, count, *ppos, inode->i_size);
742 /* turn off the kernel's read-ahead */
743 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
746 filp->f_ra.ra_pages = 0;
748 retval = generic_file_read(filp, buf, count, ppos);
751 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
756 * Write to a file (through the page cache).
758 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
761 struct ll_file_data *fd = file->private_data;
762 struct inode *inode = file->f_dentry->d_inode;
763 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
764 struct lustre_handle lockh = { 0 };
765 ldlm_policy_data_t policy;
766 loff_t maxbytes = ll_file_maxbytes(inode);
768 int nonblock = 0, rc;
770 if (file->f_flags & O_NONBLOCK)
771 nonblock = LDLM_FL_BLOCK_NOWAIT;
772 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
773 inode->i_ino, inode->i_generation, inode, count, *ppos);
775 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
777 /* POSIX, but surprised the VFS doesn't check this already */
781 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
782 * called on the file, don't fail the below assertion (bug 2388). */
783 if (file->f_flags & O_LOV_DELAY_CREATE && lsm == NULL)
788 if (file->f_flags & O_APPEND) {
789 policy.l_extent.start = 0;
790 policy.l_extent.end = OBD_OBJECT_EOF;
792 policy.l_extent.start = *ppos;
793 policy.l_extent.end = *ppos + count - 1;
796 rc = ll_extent_lock(fd, inode, lsm, LCK_PW, &policy, &lockh, nonblock);
800 /* this is ok, g_f_w will overwrite this under i_sem if it races
801 * with a local truncate, it just makes our maxbyte checking easier */
802 if (file->f_flags & O_APPEND)
803 *ppos = inode->i_size;
805 if (*ppos >= maxbytes) {
806 if (count || *ppos > maxbytes) {
807 send_sig(SIGXFSZ, current, 0);
808 GOTO(out, retval = -EFBIG);
811 if (*ppos + count > maxbytes)
812 count = maxbytes - *ppos;
814 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
815 inode->i_ino, count, *ppos);
817 /* generic_file_write handles O_APPEND after getting i_sem */
818 retval = generic_file_write(file, buf, count, ppos);
821 ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
822 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
823 retval > 0 ? retval : 0);
827 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
830 struct ll_inode_info *lli = ll_i2info(inode);
831 struct obd_export *exp = ll_i2obdexp(inode);
832 struct ll_recreate_obj ucreatp;
833 struct obd_trans_info oti = { 0 };
834 struct obdo *oa = NULL;
837 struct lov_stripe_md *lsm, *lsm2;
840 if (!capable (CAP_SYS_ADMIN))
843 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
844 sizeof(struct ll_recreate_obj));
853 down(&lli->lli_open_sem);
856 up(&lli->lli_open_sem);
860 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
861 (lsm->lsm_stripe_count));
863 OBD_ALLOC(lsm2, lsm_size);
865 up(&lli->lli_open_sem);
870 oa->o_id = ucreatp.lrc_id;
871 oa->o_nlink = ucreatp.lrc_ost_idx;
872 oa->o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS;
873 oa->o_flags |= OBD_FL_RECREATE_OBJS;
874 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
875 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
877 oti.oti_objid = NULL;
878 memcpy(lsm2, lsm, lsm_size);
879 rc = obd_create(exp, oa, &lsm2, &oti);
881 up(&lli->lli_open_sem);
882 OBD_FREE(lsm2, lsm_size);
887 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
888 int flags, struct lov_user_md *lum,
891 struct ll_inode_info *lli = ll_i2info(inode);
893 struct obd_export *exp = ll_i2obdexp(inode);
894 struct lov_stripe_md *lsm;
895 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
896 struct ptlrpc_request *req = NULL;
901 down(&lli->lli_open_sem);
904 up(&lli->lli_open_sem);
905 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
910 f = get_empty_filp();
914 f->f_dentry = file->f_dentry;
915 f->f_vfsmnt = file->f_vfsmnt;
917 rc = ll_intent_file_open(f, lum, lum_size, &oit);
920 if (it_disposition(&oit, DISP_LOOKUP_NEG))
922 req = oit.d.lustre.it_data;
923 rc = oit.d.lustre.it_status;
928 rc = mdc_req2lustre_md(req, 1, exp, &md);
931 ll_update_inode(f->f_dentry->d_inode, md.body, md.lsm);
933 rc = ll_local_open(f, &oit);
936 ll_intent_release(&oit);
938 rc = ll_file_release(f->f_dentry->d_inode, f);
943 up(&lli->lli_open_sem);
945 ptlrpc_req_finished(req);
949 static int ll_lov_setea(struct inode *inode, struct file *file,
952 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
953 struct lov_user_md *lump;
954 int lum_size = sizeof(struct lov_user_md) +
955 sizeof(struct lov_user_ost_data);
959 if (!capable (CAP_SYS_ADMIN))
962 OBD_ALLOC(lump, lum_size);
966 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
968 OBD_FREE(lump, lum_size);
972 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
974 OBD_FREE(lump, lum_size);
978 static int ll_lov_setstripe(struct inode *inode, struct file *file,
981 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
983 int flags = FMODE_WRITE;
986 /* Bug 1152: copy properly when this is no longer true */
987 LASSERT(sizeof(lum) == sizeof(*lump));
988 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
989 rc = copy_from_user(&lum, lump, sizeof(lum));
993 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
997 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
999 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1004 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2obdexp(inode), 0, lsm,
1008 static int ll_get_grouplock(struct inode *inode, struct file *file,
1011 struct ll_file_data *fd = file->private_data;
1012 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1013 .end = OBD_OBJECT_EOF}};
1014 struct lustre_handle lockh = { 0 };
1015 struct ll_inode_info *lli = ll_i2info(inode);
1016 struct lov_stripe_md *lsm = lli->lli_smd;
1020 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1024 policy.l_extent.gid = arg;
1025 if (file->f_flags & O_NONBLOCK)
1026 flags = LDLM_FL_BLOCK_NOWAIT;
1028 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags);
1032 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1034 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1039 static int ll_put_grouplock(struct inode *inode, struct file *file,
1042 struct ll_file_data *fd = file->private_data;
1043 struct ll_inode_info *lli = ll_i2info(inode);
1044 struct lov_stripe_md *lsm = lli->lli_smd;
1048 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1049 /* Ugh, it's already unlocked. */
1053 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1056 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1058 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1063 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1068 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1071 struct ll_file_data *fd = file->private_data;
1075 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1076 inode->i_generation, inode, cmd);
1078 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1081 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1083 case LL_IOC_GETFLAGS:
1084 /* Get the current value of the file flags */
1085 return put_user(fd->fd_flags, (int *)arg);
1086 case LL_IOC_SETFLAGS:
1087 case LL_IOC_CLRFLAGS:
1088 /* Set or clear specific file flags */
1089 /* XXX This probably needs checks to ensure the flags are
1090 * not abused, and to handle any flag side effects.
1092 if (get_user(flags, (int *) arg))
1095 if (cmd == LL_IOC_SETFLAGS)
1096 fd->fd_flags |= flags;
1098 fd->fd_flags &= ~flags;
1100 case LL_IOC_LOV_SETSTRIPE:
1101 RETURN(ll_lov_setstripe(inode, file, arg));
1102 case LL_IOC_LOV_SETEA:
1103 RETURN(ll_lov_setea(inode, file, arg));
1104 case LL_IOC_LOV_GETSTRIPE:
1105 RETURN(ll_lov_getstripe(inode, arg));
1106 case LL_IOC_RECREATE_OBJ:
1107 RETURN(ll_lov_recreate_obj(inode, file, arg));
1108 case EXT3_IOC_GETFLAGS:
1109 case EXT3_IOC_SETFLAGS:
1110 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1111 case LL_IOC_GROUP_LOCK:
1112 RETURN(ll_get_grouplock(inode, file, arg));
1113 case LL_IOC_GROUP_UNLOCK:
1114 RETURN(ll_put_grouplock(inode, file, arg));
1115 /* We need to special case any other ioctls we want to handle,
1116 * to send them to the MDS/OST as appropriate and to properly
1117 * network encode the arg field.
1118 case EXT2_IOC_GETVERSION_OLD:
1119 case EXT2_IOC_GETVERSION_NEW:
1120 case EXT2_IOC_SETVERSION_OLD:
1121 case EXT2_IOC_SETVERSION_NEW:
1124 RETURN( obd_iocontrol(cmd, ll_i2obdexp(inode), 0, NULL,
1129 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1131 struct inode *inode = file->f_dentry->d_inode;
1132 struct ll_file_data *fd = file->private_data;
1133 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1134 struct lustre_handle lockh = {0};
1137 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1138 inode->i_generation, inode,
1139 offset + ((origin==2) ? inode->i_size : file->f_pos));
1141 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1142 if (origin == 2) { /* SEEK_END */
1143 int nonblock = 0, rc;
1144 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1146 if (file->f_flags & O_NONBLOCK)
1147 nonblock = LDLM_FL_BLOCK_NOWAIT;
1149 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1154 offset += inode->i_size;
1155 } else if (origin == 1) { /* SEEK_CUR */
1156 offset += file->f_pos;
1160 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1161 if (offset != file->f_pos) {
1162 file->f_pos = offset;
1163 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1165 file->f_version = ++event;
1172 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1176 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1178 struct inode *inode = dentry->d_inode;
1179 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1181 struct ptlrpc_request *req;
1184 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1185 inode->i_generation, inode);
1187 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1189 /* fsync's caller has already called _fdata{sync,write}, we want
1190 * that IO to finish before calling the osc and mdc sync methods */
1191 rc = filemap_fdatawait(inode->i_mapping);
1193 ll_inode2fid(&fid, inode);
1194 err = mdc_sync(ll_i2sbi(inode)->ll_mdc_exp, &fid, &req);
1198 ptlrpc_req_finished(req);
1201 struct obdo *oa = obdo_alloc();
1204 RETURN(rc ? rc : -ENOMEM);
1206 oa->o_id = lsm->lsm_object_id;
1207 oa->o_valid = OBD_MD_FLID;
1208 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
1209 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
1211 err = obd_sync(ll_i2sbi(inode)->ll_osc_exp, oa, lsm,
1221 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1223 struct inode *inode = file->f_dentry->d_inode;
1224 struct ll_sb_info *sbi = ll_i2sbi(inode);
1225 struct obd_device *obddev;
1226 struct ldlm_res_id res_id =
1227 { .name = {inode->i_ino, inode->i_generation, LDLM_FLOCK} };
1228 struct lustre_handle lockh = {0};
1229 ldlm_policy_data_t flock;
1230 ldlm_mode_t mode = 0;
1235 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1236 inode->i_ino, file_lock);
1238 flock.l_flock.pid = file_lock->fl_pid;
1239 flock.l_flock.start = file_lock->fl_start;
1240 flock.l_flock.end = file_lock->fl_end;
1242 switch (file_lock->fl_type) {
1247 /* An unlock request may or may not have any relation to
1248 * existing locks so we may not be able to pass a lock handle
1249 * via a normal ldlm_lock_cancel() request. The request may even
1250 * unlock a byte range in the middle of an existing lock. In
1251 * order to process an unlock request we need all of the same
1252 * information that is given with a normal read or write record
1253 * lock request. To avoid creating another ldlm unlock (cancel)
1254 * message we'll treat a LCK_NL flock request as an unlock. */
1261 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1270 flags = LDLM_FL_BLOCK_NOWAIT;
1273 flags = LDLM_FL_TEST_LOCK;
1274 /* Save the old mode so that if the mode in the lock changes we
1275 * can decrement the appropriate reader or writer refcount. */
1276 file_lock->fl_type = mode;
1279 CERROR("unknown fcntl lock command: %d\n", cmd);
1283 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1284 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1285 flags, mode, flock.l_flock.start, flock.l_flock.end);
1287 obddev = sbi->ll_mdc_exp->exp_obd;
1288 rc = ldlm_cli_enqueue(sbi->ll_mdc_exp, NULL, obddev->obd_namespace,
1289 res_id, LDLM_FLOCK, &flock, mode, &flags,
1290 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1291 NULL, 0, NULL, &lockh);
1295 static int ll_have_md_lock(struct dentry *de, __u64 lockpart)
1297 struct ll_sb_info *sbi = ll_s2sbi(de->d_sb);
1298 struct lustre_handle lockh;
1299 struct ldlm_res_id res_id = { .name = {0} };
1300 struct obd_device *obddev;
1302 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
1308 obddev = sbi->ll_mdc_exp->exp_obd;
1309 res_id.name[0] = de->d_inode->i_ino;
1310 res_id.name[1] = de->d_inode->i_generation;
1312 CDEBUG(D_INFO, "trying to match res "LPU64"\n", res_id.name[0]);
1314 /* FIXME use LDLM_FL_TEST_LOCK instead */
1315 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1316 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_IBITS,
1317 &policy, LCK_PR, &lockh)) {
1318 ldlm_lock_decref(&lockh, LCK_PR);
1322 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_IBITS,
1323 &policy, LCK_PW, &lockh)) {
1324 ldlm_lock_decref(&lockh, LCK_PW);
1330 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
1332 struct inode *inode = dentry->d_inode;
1333 struct ll_inode_info *lli;
1334 struct lov_stripe_md *lsm;
1339 CERROR("REPORT THIS LINE TO PETER\n");
1342 lli = ll_i2info(inode);
1343 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s\n",
1344 inode->i_ino, inode->i_generation, inode, dentry->d_name.name);
1345 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1346 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1349 if (!ll_have_md_lock(dentry, MDS_INODELOCK_UPDATE)) {
1350 struct ptlrpc_request *req = NULL;
1351 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
1353 unsigned long valid = 0;
1356 if (S_ISREG(inode->i_mode)) {
1357 ealen = obd_size_diskmd(sbi->ll_osc_exp, NULL);
1358 valid |= OBD_MD_FLEASIZE;
1360 ll_inode2fid(&fid, inode);
1361 rc = mdc_getattr(sbi->ll_mdc_exp, &fid, valid, ealen, &req);
1363 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
1366 rc = ll_prep_inode(sbi->ll_osc_exp, &inode, req, 0, NULL);
1368 ptlrpc_req_finished(req);
1371 ptlrpc_req_finished(req);
1375 if (lsm == NULL) /* object not yet allocated, don't validate size */
1378 /* ll_glimpse_size will prefer locally cached writes if they extend
1383 rc = ll_glimpse_size(inode, &lvb);
1384 inode->i_size = lvb.lvb_size;
1389 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1390 int ll_getattr(struct vfsmount *mnt, struct dentry *de,
1391 struct lookup_intent *it, struct kstat *stat)
1394 struct inode *inode = de->d_inode;
1396 res = ll_inode_revalidate_it(de, it);
1397 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1402 stat->dev = inode->i_sb->s_dev;
1403 stat->ino = inode->i_ino;
1404 stat->mode = inode->i_mode;
1405 stat->nlink = inode->i_nlink;
1406 stat->uid = inode->i_uid;
1407 stat->gid = inode->i_gid;
1408 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1409 stat->atime = inode->i_atime;
1410 stat->mtime = inode->i_mtime;
1411 stat->ctime = inode->i_ctime;
1412 stat->size = inode->i_size;
1413 stat->blksize = inode->i_blksize;
1414 stat->blocks = inode->i_blocks;
1419 struct file_operations ll_file_operations = {
1420 .read = ll_file_read,
1421 .write = ll_file_write,
1422 .ioctl = ll_file_ioctl,
1423 .open = ll_file_open,
1424 .release = ll_file_release,
1425 .mmap = generic_file_mmap,
1426 .llseek = ll_file_seek,
1427 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
1428 .sendfile = generic_file_sendfile,
1431 //.lock ll_file_flock
1434 struct inode_operations ll_file_inode_operations = {
1435 .setattr_raw = ll_setattr_raw,
1436 .setattr = ll_setattr,
1437 .truncate = ll_truncate,
1438 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1439 .getattr_it = ll_getattr,
1441 .revalidate_it = ll_inode_revalidate_it,