1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
7 * Author: Andreas Dilger <adilger@clusterfs.com>
9 * This file is part of Lustre, http://www.lustre.org.
11 * Lustre is free software; you can redistribute it and/or
12 * modify it under the terms of version 2 of the GNU General Public
13 * License as published by the Free Software Foundation.
15 * Lustre is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with Lustre; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #define DEBUG_SUBSYSTEM S_LLITE
26 #include <linux/lustre_dlm.h>
27 #include <linux/lustre_lite.h>
28 #include <linux/pagemap.h>
29 #include <linux/file.h>
30 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
31 #include <linux/lustre_compat25.h>
34 #include "llite_internal.h"
36 static int ll_mdc_close(struct obd_export *mdc_exp, struct inode *inode,
39 struct ll_file_data *fd = file->private_data;
40 struct ptlrpc_request *req = NULL;
41 struct obd_client_handle *och = &fd->fd_mds_och;
48 memset(&obdo, 0, sizeof(obdo));
49 obdo.o_id = inode->i_ino;
50 obdo.o_mode = inode->i_mode;
51 obdo.o_size = inode->i_size;
52 obdo.o_blocks = inode->i_blocks;
53 if (0 /* ll_is_inode_dirty(inode) */) {
54 obdo.o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
55 valid |= OBD_MD_FLFLAGS;
58 rc = mdc_close(mdc_exp, &obdo, och, &req);
60 /* We are the last writer, so the MDS has instructed us to get
61 * the file size and any write cookies, then close again. */
62 //ll_queue_done_writing(inode);
65 CERROR("inode %lu mdc close failed: rc = %d\n",
69 rc = ll_objects_destroy(req, file->f_dentry->d_inode);
71 CERROR("inode %lu ll_objects destroy: rc = %d\n",
75 mdc_clear_open_replay_data(och);
76 ptlrpc_req_finished(req);
77 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
78 file->private_data = NULL;
79 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof *fd);
84 /* While this returns an error code, fput() the caller does not, so we need
85 * to make every effort to clean up all of our state here. Also, applications
86 * rarely check close errors and even if an error is returned they will not
87 * re-try the close call.
89 int ll_file_release(struct inode *inode, struct file *file)
91 struct ll_file_data *fd;
92 struct ll_sb_info *sbi = ll_i2sbi(inode);
96 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
97 inode->i_generation, inode);
99 /* don't do anything for / */
100 if (inode->i_sb->s_root == file->f_dentry)
103 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
104 fd = (struct ll_file_data *)file->private_data;
107 rc = ll_mdc_close(sbi->ll_mdc_exp, inode, file);
111 static int ll_intent_file_open(struct file *file, void *lmm,
112 int lmmsize, struct lookup_intent *itp)
114 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
115 struct lustre_handle lockh;
116 struct mdc_op_data data;
117 struct dentry *parent = file->f_dentry->d_parent;
118 const char *name = file->f_dentry->d_name.name;
119 const int len = file->f_dentry->d_name.len;
125 ll_prepare_mdc_op_data(&data, parent->d_inode, NULL, name, len, O_RDWR);
127 rc = mdc_enqueue(sbi->ll_mdc_exp, LDLM_PLAIN, itp, LCK_PR, &data,
128 &lockh, lmm, lmmsize, ldlm_completion_ast,
129 ll_mdc_blocking_ast, parent->d_inode);
131 CERROR("lock enqueue: err: %d\n", rc);
135 static int ll_local_open(struct file *file, struct lookup_intent *it)
137 struct ptlrpc_request *req = it->d.lustre.it_data;
138 struct ll_inode_info *lli = ll_i2info(file->f_dentry->d_inode);
139 struct ll_file_data *fd;
140 struct mds_body *body;
143 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
144 LASSERT (body != NULL); /* reply already checked out */
145 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
147 LASSERT(!file->private_data);
149 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
150 /* We can't handle this well without reorganizing ll_file_open and
151 * ll_mdc_close, so don't even try right now. */
154 memcpy(&fd->fd_mds_och.och_fh, &body->handle, sizeof(body->handle));
155 fd->fd_mds_och.och_magic = OBD_CLIENT_HANDLE_MAGIC;
156 file->private_data = fd;
157 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
159 lli->lli_io_epoch = body->io_epoch;
161 mdc_set_open_replay_data(&fd->fd_mds_och, it->d.lustre.it_data);
166 /* Open a file, and (for the very first open) create objects on the OSTs at
167 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
168 * creation or open until ll_lov_setstripe() ioctl is called. We grab
169 * lli_open_sem to ensure no other process will create objects, send the
170 * stripe MD to the MDS, or try to destroy the objects if that fails.
172 * If we already have the stripe MD locally then we don't request it in
173 * mdc_open(), by passing a lmm_size = 0.
175 * It is up to the application to ensure no other processes open this file
176 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
177 * used. We might be able to avoid races of that sort by getting lli_open_sem
178 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
179 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
181 int ll_file_open(struct inode *inode, struct file *file)
183 struct ll_inode_info *lli = ll_i2info(inode);
184 struct lookup_intent *it;
185 struct lov_stripe_md *lsm;
186 struct ptlrpc_request *req;
190 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
191 inode->i_generation, inode);
193 /* don't do anything for / */
194 if (inode->i_sb->s_root == file->f_dentry)
199 if (!it->d.lustre.it_disposition) {
200 struct lookup_intent oit = { .it_op = IT_OPEN,
201 .it_flags = file->f_flags };
203 rc = ll_intent_file_open(file, NULL, 0, it);
208 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
209 rc = it_open_error(DISP_OPEN_OPEN, it);
213 rc = ll_local_open(file, it);
217 if (!S_ISREG(inode->i_mode))
222 if (file->f_flags & O_LOV_DELAY_CREATE ||
223 !(file->f_mode & FMODE_WRITE)) {
224 CDEBUG(D_INODE, "object creation was delayed\n");
228 file->f_flags &= ~O_LOV_DELAY_CREATE;
231 req = it->d.lustre.it_data;
232 ptlrpc_req_finished(req);
234 ll_open_complete(inode);
238 /* Fills the obdo with the attributes for the inode defined by lsm */
239 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
242 struct ptlrpc_request_set *set;
246 LASSERT(lsm != NULL);
248 memset(oa, 0, sizeof *oa);
249 oa->o_id = lsm->lsm_object_id;
250 oa->o_mode = S_IFREG;
251 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
252 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
255 set = ptlrpc_prep_set();
257 CERROR ("ENOMEM allocing request set\n");
260 rc = obd_getattr_async(exp, oa, lsm, set);
262 rc = ptlrpc_set_wait(set);
263 ptlrpc_set_destroy(set);
268 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
269 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
273 static inline void ll_remove_suid(struct inode *inode)
277 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
278 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
280 /* was any of the uid bits set? */
281 mode &= inode->i_mode;
282 if (mode && !capable(CAP_FSETID)) {
283 inode->i_mode &= ~mode;
284 // XXX careful here - we cannot change the size
288 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
290 struct ll_inode_info *lli = ll_i2info(inode);
291 struct lov_stripe_md *lsm = lli->lli_smd;
292 struct obd_export *exp = ll_i2obdexp(inode);
295 struct ldlm_lock *lock;
296 struct lov_stripe_md *lsm;
297 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
298 __u32 stripe, vallen = sizeof(stripe);
302 if (lsm->lsm_stripe_count == 1)
305 /* get our offset in the lov */
306 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
308 CERROR("obd_get_info: rc = %d\n", rc);
311 LASSERT(stripe < lsm->lsm_stripe_count);
315 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
316 * we get a lock cancellation for each stripe, so we have to map the obd's
317 * region back onto the stripes in the file that it held.
319 * No one can dirty the extent until we've finished our work and they can
320 * enqueue another lock. The DLM protects us from ll_file_read/write here,
321 * but other kernel actors could have pages locked. */
322 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
323 struct ldlm_lock *lock, __u32 stripe)
325 struct ldlm_extent *extent = &lock->l_policy_data.l_extent;
326 unsigned long start, end, count, skip, i, j;
328 int rc, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
331 CDEBUG(D_INODE, "obdo %lu inode %p ["LPU64"->"LPU64"] size: %llu\n",
332 inode->i_ino, inode, extent->start, extent->end, inode->i_size);
334 /* our locks are page granular thanks to osc_enqueue, we invalidate the
336 LASSERT((extent->start & ~PAGE_CACHE_MASK) == 0);
337 LASSERT(((extent->end+1) & ~PAGE_CACHE_MASK) == 0);
339 start = extent->start >> PAGE_CACHE_SHIFT;
342 end = (extent->end >> PAGE_CACHE_SHIFT) + 1;
343 if ((end << PAGE_CACHE_SHIFT) < extent->end)
345 if (lsm->lsm_stripe_count > 1) {
346 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
347 skip = (lsm->lsm_stripe_count - 1) * count;
348 start += (start/count * skip) + (stripe * count);
350 end += (end/count * skip) + (stripe * count);
353 i = (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
357 CDEBUG(D_INODE, "walking page indices start: %lu j: %lu count: %lu "
358 "skip: %lu end: %lu%s\n", start, start % count, count, skip, end,
359 discard ? " (DISCARDING)" : "");
361 /* this is the simplistic implementation of page eviction at
362 * cancelation. It is careful to get races with other page
363 * lockers handled correctly. fixes from bug 20 will make it
364 * more efficient by associating locks with pages and with
365 * batching writeback under the lock explicitly. */
366 for (i = start, j = start % count ; ; j++, i++) {
374 ll_pgcache_lock(inode->i_mapping);
375 if (list_empty(&inode->i_mapping->dirty_pages) &&
376 list_empty(&inode->i_mapping->clean_pages) &&
377 list_empty(&inode->i_mapping->locked_pages)) {
378 CDEBUG(D_INODE, "nothing left\n");
379 ll_pgcache_unlock(inode->i_mapping);
382 ll_pgcache_unlock(inode->i_mapping);
384 conditional_schedule();
386 page = find_get_page(inode->i_mapping, i);
389 LL_CDEBUG_PAGE(page, "locking page\n");
392 /* page->mapping to check with racing against teardown */
393 if (page->mapping && PageDirty(page) && !discard) {
394 ClearPageDirty(page);
395 LL_CDEBUG_PAGE(page, "found dirty\n");
396 ll_pgcache_lock(inode->i_mapping);
397 list_del(&page->list);
398 list_add(&page->list, &inode->i_mapping->locked_pages);
399 ll_pgcache_unlock(inode->i_mapping);
401 rc = ll_call_writepage(inode, page);
403 CERROR("writepage of page %p failed: %d\n",
405 /* either waiting for io to complete or reacquiring
406 * the lock that the failed writepage released */
410 /* checking again to account for writeback's lock_page() */
411 if (page->mapping != NULL) {
412 LL_CDEBUG_PAGE(page, "truncating\n");
413 ll_truncate_complete_page(page);
416 page_cache_release(page);
421 static int ll_extent_lock_callback(struct ldlm_lock *lock,
422 struct ldlm_lock_desc *new, void *data,
425 struct lustre_handle lockh = { 0 };
429 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
430 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
435 case LDLM_CB_BLOCKING:
436 ldlm_lock2handle(lock, &lockh);
437 rc = ldlm_cli_cancel(&lockh);
439 CERROR("ldlm_cli_cancel failed: %d\n", rc);
441 case LDLM_CB_CANCELING: {
442 struct inode *inode = ll_inode_from_lock(lock);
443 struct ll_inode_info *lli;
444 struct lov_stripe_md *lsm;
450 lli = ll_i2info(inode);
453 if (lli->lli_smd == NULL)
457 stripe = ll_lock_to_stripe_offset(inode, lock);
458 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
461 kms = ldlm_extent_shift_kms(lock,
462 lsm->lsm_oinfo[stripe].loi_kms);
463 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
464 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
465 lsm->lsm_oinfo[stripe].loi_kms, kms);
466 lsm->lsm_oinfo[stripe].loi_kms = kms;
468 //ll_try_done_writing(inode);
481 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
483 /* XXX ALLOCATE - 160 bytes */
484 struct inode *inode = ll_inode_from_lock(lock);
485 struct ll_inode_info *lli = ll_i2info(inode);
486 struct lustre_handle lockh = { 0 };
491 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
492 LDLM_FL_BLOCK_CONV)) {
493 LBUG(); /* not expecting any blocked async locks yet */
494 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
496 ldlm_lock_dump(D_OTHER, lock, 0);
497 ldlm_reprocess_all(lock->l_resource);
501 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
503 stripe = ll_lock_to_stripe_offset(inode, lock);
505 if (lock->l_lvb_len) {
506 struct lov_stripe_md *lsm = lli->lli_smd;
508 lvb = lock->l_lvb_data;
509 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
512 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
513 kms = ldlm_extent_shift_kms(NULL, kms);
514 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
515 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
516 lsm->lsm_oinfo[stripe].loi_kms, kms);
517 lsm->lsm_oinfo[stripe].loi_kms = kms;
522 wake_up(&lock->l_waitq);
524 ldlm_lock2handle(lock, &lockh);
525 ldlm_lock_decref(&lockh, LCK_PR);
530 /* This function is a disaster. I hate the LOV. */
531 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
533 struct ptlrpc_request *req = reqp;
534 struct inode *inode = ll_inode_from_lock(lock);
535 struct obd_export *exp;
536 struct ll_inode_info *lli;
541 struct lov_stripe_md *lsm;
543 __u32 vallen = sizeof(data);
544 int rc, size = sizeof(*lvb);
549 lli = ll_i2info(inode);
552 if (lli->lli_smd == NULL)
554 exp = ll_i2obdexp(inode);
556 /* First, find out which stripe index this lock corresponds to. */
557 if (lli->lli_smd->lsm_stripe_count > 1)
558 data.stripe_number = ll_lock_to_stripe_offset(inode, lock);
560 data.stripe_number = 0;
562 data.size = inode->i_size;
563 data.lsm = lli->lli_smd;
565 rc = obd_get_info(exp, strlen("size_to_stripe"), "size_to_stripe",
568 CERROR("obd_get_info: rc = %d\n", rc);
572 LDLM_DEBUG(lock, "i_size: %Lu -> stripe number %d -> size %Lu",
573 inode->i_size, data.stripe_number, data.size);
575 rc = lustre_pack_reply(req, 1, &size, NULL);
577 CERROR("lustre_pack_reply: %d\n", rc);
581 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
582 lvb->lvb_size = data.size;
590 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
591 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
593 /* NB: lov_merge_size will prefer locally cached writes if they extend the
594 * file (because it prefers KMS over RSS when larger) */
595 int ll_glimpse_size(struct inode *inode, struct ost_lvb *lvb)
597 struct ll_inode_info *lli = ll_i2info(inode);
598 struct ll_sb_info *sbi = ll_i2sbi(inode);
599 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
600 struct lustre_handle lockh;
601 int rc, flags = LDLM_FL_HAS_INTENT;
604 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
606 rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
607 LCK_PR, &flags, ll_extent_lock_callback,
608 ldlm_completion_ast, ll_glimpse_callback, inode,
609 sizeof(*lvb), lustre_swab_ost_lvb, &lockh);
613 lvb->lvb_size = lov_merge_size(lli->lli_smd, 0);
614 //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
616 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64"\n", lvb->lvb_size);
618 obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
623 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
624 struct lov_stripe_md *lsm, int mode,
625 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
628 struct ll_sb_info *sbi = ll_i2sbi(inode);
632 LASSERT(lockh->cookie == 0);
634 /* XXX phil: can we do this? won't it screw the file size up? */
635 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
636 (sbi->ll_flags & LL_SBI_NOLCK))
639 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
640 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
642 rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
643 &ast_flags, ll_extent_lock_callback,
644 ldlm_completion_ast, ll_glimpse_callback, inode,
645 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
649 if (policy->l_extent.start == 0 &&
650 policy->l_extent.end == OBD_OBJECT_EOF)
651 inode->i_size = lov_merge_size(lsm, 1);
653 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
658 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
659 struct lov_stripe_md *lsm, int mode,
660 struct lustre_handle *lockh)
662 struct ll_sb_info *sbi = ll_i2sbi(inode);
666 /* XXX phil: can we do this? won't it screw the file size up? */
667 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
668 (sbi->ll_flags & LL_SBI_NOLCK))
671 rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
676 static ssize_t ll_file_read(struct file *filp, char *buf, size_t count,
679 struct ll_file_data *fd = filp->private_data;
680 struct inode *inode = filp->f_dentry->d_inode;
681 struct ll_inode_info *lli = ll_i2info(inode);
682 struct lov_stripe_md *lsm = lli->lli_smd;
683 struct lustre_handle lockh = { 0 };
684 ldlm_policy_data_t policy;
689 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
690 inode->i_ino, inode->i_generation, inode, count, *ppos);
692 /* "If nbyte is 0, read() will return 0 and have no other results."
693 * -- Single Unix Spec */
697 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
703 policy.l_extent.start = *ppos;
704 policy.l_extent.end = *ppos + count - 1;
706 err = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh, 0);
710 kms = lov_merge_size(lsm, 1);
711 if (policy.l_extent.end > kms) {
712 /* A glimpse is necessary to determine whether we return a short
713 * read or some zeroes at the end of the buffer */
715 retval = ll_glimpse_size(inode, &lvb);
718 inode->i_size = lvb.lvb_size;
723 CDEBUG(D_INFO, "Reading inode %lu, "LPSZ" bytes, offset %Ld, i_size "
724 LPU64"\n", inode->i_ino, count, *ppos, inode->i_size);
726 /* turn off the kernel's read-ahead */
727 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
730 filp->f_ra.ra_pages = 0;
732 retval = generic_file_read(filp, buf, count, ppos);
735 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
740 * Write to a file (through the page cache).
742 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
745 struct ll_file_data *fd = file->private_data;
746 struct inode *inode = file->f_dentry->d_inode;
747 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
748 struct lustre_handle lockh = { 0 };
749 ldlm_policy_data_t policy;
750 loff_t maxbytes = ll_file_maxbytes(inode);
754 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
755 inode->i_ino, inode->i_generation, inode, count, *ppos);
757 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
759 /* POSIX, but surprised the VFS doesn't check this already */
763 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
764 * called on the file, don't fail the below assertion (bug 2388). */
765 if (file->f_flags & O_LOV_DELAY_CREATE && lsm == NULL)
770 if (file->f_flags & O_APPEND) {
771 policy.l_extent.start = 0;
772 policy.l_extent.end = OBD_OBJECT_EOF;
774 policy.l_extent.start = *ppos;
775 policy.l_extent.end = *ppos + count - 1;
778 err = ll_extent_lock(fd, inode, lsm, LCK_PW, &policy, &lockh, 0);
782 /* this is ok, g_f_w will overwrite this under i_sem if it races
783 * with a local truncate, it just makes our maxbyte checking easier */
784 if (file->f_flags & O_APPEND)
785 *ppos = inode->i_size;
787 if (*ppos >= maxbytes) {
788 if (count || *ppos > maxbytes) {
789 send_sig(SIGXFSZ, current, 0);
790 GOTO(out, retval = -EFBIG);
793 if (*ppos + count > maxbytes)
794 count = maxbytes - *ppos;
796 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
797 inode->i_ino, count, *ppos);
799 /* generic_file_write handles O_APPEND after getting i_sem */
800 retval = generic_file_write(file, buf, count, ppos);
804 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
806 ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
810 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
813 struct ll_inode_info *lli = ll_i2info(inode);
814 struct obd_export *exp = ll_i2obdexp(inode);
815 struct ll_recreate_obj ucreatp;
816 struct obd_trans_info oti = { 0 };
817 struct obdo *oa = NULL;
820 struct lov_stripe_md *lsm, *lsm2;
823 if (!capable (CAP_SYS_ADMIN))
826 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
827 sizeof(struct ll_recreate_obj));
836 down(&lli->lli_open_sem);
839 up(&lli->lli_open_sem);
843 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
844 (lsm->lsm_stripe_count));
846 OBD_ALLOC(lsm2, lsm_size);
848 up(&lli->lli_open_sem);
853 oa->o_id = ucreatp.lrc_id;
854 oa->o_nlink = ucreatp.lrc_ost_idx;
855 oa->o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS;
856 oa->o_flags |= OBD_FL_RECREATE_OBJS;
857 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
858 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
860 oti.oti_objid = NULL;
861 memcpy(lsm2, lsm, lsm_size);
862 rc = obd_create(exp, oa, &lsm2, &oti);
864 up(&lli->lli_open_sem);
865 OBD_FREE(lsm2, lsm_size);
870 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
871 int flags, struct lov_user_md *lum, int lum_size)
873 struct ll_inode_info *lli = ll_i2info(inode);
875 struct obd_export *exp = ll_i2obdexp(inode);
876 struct lov_stripe_md *lsm;
877 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
878 struct ptlrpc_request *req = NULL;
883 down(&lli->lli_open_sem);
886 up(&lli->lli_open_sem);
887 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
892 f = get_empty_filp();
896 f->f_dentry = file->f_dentry;
897 f->f_vfsmnt = file->f_vfsmnt;
899 rc = ll_intent_file_open(f, lum, lum_size, &oit);
902 if (it_disposition(&oit, DISP_LOOKUP_NEG))
904 req = oit.d.lustre.it_data;
905 rc = oit.d.lustre.it_status;
910 rc = mdc_req2lustre_md(req, 1, exp, &md);
913 ll_update_inode(f->f_dentry->d_inode, md.body, md.lsm);
915 rc = ll_local_open(f, &oit);
918 ll_intent_release(&oit);
920 rc = ll_file_release(f->f_dentry->d_inode, f);
925 up(&lli->lli_open_sem);
927 ptlrpc_req_finished(req);
931 static int ll_lov_setea(struct inode *inode, struct file *file,
934 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
935 struct lov_user_md *lump;
936 int lum_size = sizeof(struct lov_user_md) +
937 sizeof(struct lov_user_ost_data);
941 if (!capable (CAP_SYS_ADMIN))
944 OBD_ALLOC(lump, lum_size);
948 rc = copy_from_user(lump, (struct lov_user_md *)arg,
951 OBD_FREE(lump, lum_size);
955 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
957 OBD_FREE(lump, lum_size);
961 static int ll_lov_setstripe(struct inode *inode, struct file *file,
964 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
966 int flags = FMODE_WRITE;
969 /* Bug 1152: copy properly when this is no longer true */
970 LASSERT(sizeof(lum) == sizeof(*lump));
971 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
972 rc = copy_from_user(&lum, lump, sizeof(lum));
976 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
980 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
982 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
987 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2obdexp(inode), 0, lsm,
991 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
994 struct ll_file_data *fd = file->private_data;
998 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
999 inode->i_generation, inode, cmd);
1001 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1004 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1006 case LL_IOC_GETFLAGS:
1007 /* Get the current value of the file flags */
1008 return put_user(fd->fd_flags, (int *)arg);
1009 case LL_IOC_SETFLAGS:
1010 case LL_IOC_CLRFLAGS:
1011 /* Set or clear specific file flags */
1012 /* XXX This probably needs checks to ensure the flags are
1013 * not abused, and to handle any flag side effects.
1015 if (get_user(flags, (int *) arg))
1018 if (cmd == LL_IOC_SETFLAGS)
1019 fd->fd_flags |= flags;
1021 fd->fd_flags &= ~flags;
1023 case LL_IOC_LOV_SETSTRIPE:
1024 RETURN(ll_lov_setstripe(inode, file, arg));
1025 case LL_IOC_LOV_SETEA:
1026 RETURN( ll_lov_setea(inode, file, arg) );
1027 case LL_IOC_LOV_GETSTRIPE:
1028 RETURN(ll_lov_getstripe(inode, arg));
1029 case LL_IOC_RECREATE_OBJ:
1030 RETURN(ll_lov_recreate_obj(inode, file, arg));
1031 case EXT3_IOC_GETFLAGS:
1032 case EXT3_IOC_SETFLAGS:
1033 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1034 /* We need to special case any other ioctls we want to handle,
1035 * to send them to the MDS/OST as appropriate and to properly
1036 * network encode the arg field.
1037 case EXT2_IOC_GETVERSION_OLD:
1038 case EXT2_IOC_GETVERSION_NEW:
1039 case EXT2_IOC_SETVERSION_OLD:
1040 case EXT2_IOC_SETVERSION_NEW:
1043 RETURN( obd_iocontrol(cmd, ll_i2obdexp(inode), 0, NULL,
1048 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1050 struct inode *inode = file->f_dentry->d_inode;
1051 struct ll_file_data *fd = file->private_data;
1052 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1053 struct lustre_handle lockh = {0};
1056 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1057 inode->i_generation, inode,
1058 offset + ((origin==2) ? inode->i_size : file->f_pos));
1060 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1061 if (origin == 2) { /* SEEK_END */
1063 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1064 err = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,0);
1065 if (err != ELDLM_OK)
1068 offset += inode->i_size;
1069 } else if (origin == 1) { /* SEEK_CUR */
1070 offset += file->f_pos;
1074 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1075 if (offset != file->f_pos) {
1076 file->f_pos = offset;
1077 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1079 file->f_version = ++event;
1086 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1090 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1092 struct inode *inode = dentry->d_inode;
1093 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1095 struct ptlrpc_request *req;
1098 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1099 inode->i_generation, inode);
1101 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1103 /* fsync's caller has already called _fdata{sync,write}, we want
1104 * that IO to finish before calling the osc and mdc sync methods */
1105 rc = filemap_fdatawait(inode->i_mapping);
1107 ll_inode2fid(&fid, inode);
1108 err = mdc_sync(ll_i2sbi(inode)->ll_mdc_exp, &fid, &req);
1112 ptlrpc_req_finished(req);
1115 struct obdo *oa = obdo_alloc();
1118 RETURN(rc ? rc : -ENOMEM);
1120 oa->o_id = lsm->lsm_object_id;
1121 oa->o_valid = OBD_MD_FLID;
1122 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
1123 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
1125 err = obd_sync(ll_i2sbi(inode)->ll_osc_exp, oa, lsm,
1135 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1137 struct inode *inode = file->f_dentry->d_inode;
1138 struct ll_sb_info *sbi = ll_i2sbi(inode);
1139 struct obd_device *obddev;
1140 struct ldlm_res_id res_id =
1141 { .name = {inode->i_ino, inode->i_generation, LDLM_FLOCK} };
1142 struct lustre_handle lockh = {0};
1143 ldlm_policy_data_t flock;
1144 ldlm_mode_t mode = 0;
1149 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1150 inode->i_ino, file_lock);
1152 flock.l_flock.pid = file_lock->fl_pid;
1153 flock.l_flock.start = file_lock->fl_start;
1154 flock.l_flock.end = file_lock->fl_end;
1156 switch (file_lock->fl_type) {
1161 /* An unlock request may or may not have any relation to
1162 * existing locks so we may not be able to pass a lock handle
1163 * via a normal ldlm_lock_cancel() request. The request may even
1164 * unlock a byte range in the middle of an existing lock. In
1165 * order to process an unlock request we need all of the same
1166 * information that is given with a normal read or write record
1167 * lock request. To avoid creating another ldlm unlock (cancel)
1168 * message we'll treat a LCK_NL flock request as an unlock. */
1175 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1184 flags = LDLM_FL_BLOCK_NOWAIT;
1187 flags = LDLM_FL_TEST_LOCK;
1188 /* Save the old mode so that if the mode in the lock changes we
1189 * can decrement the appropriate reader or writer refcount. */
1190 file_lock->fl_type = mode;
1193 CERROR("unknown fcntl lock command: %d\n", cmd);
1197 CDEBUG(D_DLMTRACE, "inode=%lu, pid=%u, flags=%#x, mode=%u, "
1198 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1199 flags, mode, flock.l_flock.start, flock.l_flock.end);
1201 obddev = sbi->ll_mdc_exp->exp_obd;
1202 rc = ldlm_cli_enqueue(sbi->ll_mdc_exp, NULL, obddev->obd_namespace,
1203 res_id, LDLM_FLOCK, &flock, mode, &flags,
1204 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1205 NULL, 0, NULL, &lockh);
1209 static int ll_have_md_lock(struct dentry *de)
1211 struct ll_sb_info *sbi = ll_s2sbi(de->d_sb);
1212 struct lustre_handle lockh;
1213 struct ldlm_res_id res_id = { .name = {0} };
1214 struct obd_device *obddev;
1221 obddev = sbi->ll_mdc_exp->exp_obd;
1222 res_id.name[0] = de->d_inode->i_ino;
1223 res_id.name[1] = de->d_inode->i_generation;
1225 CDEBUG(D_INFO, "trying to match res "LPU64"\n", res_id.name[0]);
1227 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1228 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
1229 NULL, LCK_PR, &lockh)) {
1230 ldlm_lock_decref(&lockh, LCK_PR);
1234 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
1235 NULL, LCK_PW, &lockh)) {
1236 ldlm_lock_decref(&lockh, LCK_PW);
1242 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
1244 struct inode *inode = dentry->d_inode;
1245 struct ll_inode_info *lli;
1246 struct lov_stripe_md *lsm;
1250 CERROR("REPORT THIS LINE TO PETER\n");
1253 lli = ll_i2info(inode);
1254 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s\n",
1255 inode->i_ino, inode->i_generation, inode, dentry->d_name.name);
1256 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1257 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1260 if (!ll_have_md_lock(dentry)) {
1261 struct ptlrpc_request *req = NULL;
1262 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
1264 unsigned long valid = 0;
1267 if (S_ISREG(inode->i_mode)) {
1268 ealen = obd_size_diskmd(sbi->ll_osc_exp, NULL);
1269 valid |= OBD_MD_FLEASIZE;
1271 ll_inode2fid(&fid, inode);
1272 rc = mdc_getattr(sbi->ll_mdc_exp, &fid, valid, ealen, &req);
1274 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
1277 rc = ll_prep_inode(sbi->ll_osc_exp, &inode, req, 0, NULL);
1279 ptlrpc_req_finished(req);
1282 ptlrpc_req_finished(req);
1286 if (lsm == NULL) /* object not yet allocated, don't validate size */
1289 /* ll_glimpse_size will prefer locally cached writes if they extend
1295 err = ll_glimpse_size(inode, &lvb);
1296 inode->i_size = lvb.lvb_size;
1301 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1302 int ll_getattr(struct vfsmount *mnt, struct dentry *de,
1303 struct lookup_intent *it, struct kstat *stat)
1306 struct inode *inode = de->d_inode;
1308 res = ll_inode_revalidate_it(de, it);
1309 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1314 stat->dev = inode->i_sb->s_dev;
1315 stat->ino = inode->i_ino;
1316 stat->mode = inode->i_mode;
1317 stat->nlink = inode->i_nlink;
1318 stat->uid = inode->i_uid;
1319 stat->gid = inode->i_gid;
1320 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1321 stat->atime = inode->i_atime;
1322 stat->mtime = inode->i_mtime;
1323 stat->ctime = inode->i_ctime;
1324 stat->size = inode->i_size;
1325 stat->blksize = inode->i_blksize;
1326 stat->blocks = inode->i_blocks;
1331 struct file_operations ll_file_operations = {
1333 write: ll_file_write,
1334 ioctl: ll_file_ioctl,
1336 release: ll_file_release,
1337 mmap: generic_file_mmap,
1338 llseek: ll_file_seek,
1340 //lock: ll_file_flock
1343 struct inode_operations ll_file_inode_operations = {
1344 setattr_raw: ll_setattr_raw,
1345 setattr: ll_setattr,
1346 truncate: ll_truncate,
1347 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1348 getattr_it: ll_getattr,
1350 revalidate_it: ll_inode_revalidate_it,
1354 struct inode_operations ll_special_inode_operations = {
1355 setattr_raw: ll_setattr_raw,
1356 setattr: ll_setattr,
1357 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1358 getattr_it: ll_getattr,
1360 revalidate_it: ll_inode_revalidate_it,