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;
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,
56 memset(&obdo, 0, sizeof(obdo));
57 obdo.o_id = inode->i_ino;
58 obdo.o_mode = inode->i_mode;
59 obdo.o_size = inode->i_size;
60 obdo.o_blocks = inode->i_blocks;
61 if (0 /* ll_is_inode_dirty(inode) */) {
62 obdo.o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
63 valid |= OBD_MD_FLFLAGS;
66 rc = mdc_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(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=%lu/%u(%p)\n", inode->i_ino,
105 inode->i_generation, inode);
107 /* don't do anything for / */
108 if (inode->i_sb->s_root == file->f_dentry)
111 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
112 fd = (struct ll_file_data *)file->private_data;
115 rc = ll_mdc_close(sbi->ll_mdc_exp, inode, file);
119 static int ll_intent_file_open(struct file *file, void *lmm,
120 int lmmsize, struct lookup_intent *itp)
122 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
123 struct lustre_handle lockh;
124 struct mdc_op_data data;
125 struct dentry *parent = file->f_dentry->d_parent;
126 const char *name = file->f_dentry->d_name.name;
127 const int len = file->f_dentry->d_name.len;
133 ll_prepare_mdc_op_data(&data, parent->d_inode, NULL, name, len, O_RDWR);
135 rc = mdc_enqueue(sbi->ll_mdc_exp, LDLM_PLAIN, itp, LCK_PR, &data,
136 &lockh, lmm, lmmsize, ldlm_completion_ast,
137 ll_mdc_blocking_ast, parent->d_inode);
139 CERROR("lock enqueue: err: %d\n", rc);
143 static int ll_local_open(struct file *file, struct lookup_intent *it)
145 struct ptlrpc_request *req = it->d.lustre.it_data;
146 struct ll_inode_info *lli = ll_i2info(file->f_dentry->d_inode);
147 struct ll_file_data *fd;
148 struct mds_body *body;
151 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
152 LASSERT (body != NULL); /* reply already checked out */
153 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
155 LASSERT(!file->private_data);
157 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
158 /* We can't handle this well without reorganizing ll_file_open and
159 * ll_mdc_close, so don't even try right now. */
162 memcpy(&fd->fd_mds_och.och_fh, &body->handle, sizeof(body->handle));
163 fd->fd_mds_och.och_magic = OBD_CLIENT_HANDLE_MAGIC;
164 file->private_data = fd;
165 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
167 lli->lli_io_epoch = body->io_epoch;
169 mdc_set_open_replay_data(&fd->fd_mds_och, it->d.lustre.it_data);
174 /* Open a file, and (for the very first open) create objects on the OSTs at
175 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
176 * creation or open until ll_lov_setstripe() ioctl is called. We grab
177 * lli_open_sem to ensure no other process will create objects, send the
178 * stripe MD to the MDS, or try to destroy the objects if that fails.
180 * If we already have the stripe MD locally then we don't request it in
181 * mdc_open(), by passing a lmm_size = 0.
183 * It is up to the application to ensure no other processes open this file
184 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
185 * used. We might be able to avoid races of that sort by getting lli_open_sem
186 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
187 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
189 int ll_file_open(struct inode *inode, struct file *file)
191 struct ll_inode_info *lli = ll_i2info(inode);
192 struct lookup_intent *it;
193 struct lov_stripe_md *lsm;
194 struct ptlrpc_request *req;
198 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
199 inode->i_generation, inode);
201 /* don't do anything for / */
202 if (inode->i_sb->s_root == file->f_dentry)
207 if (!it->d.lustre.it_disposition) {
208 struct lookup_intent oit = { .it_op = IT_OPEN,
209 .it_flags = file->f_flags };
211 rc = ll_intent_file_open(file, NULL, 0, it);
216 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
217 rc = it_open_error(DISP_OPEN_OPEN, it);
221 rc = ll_local_open(file, it);
225 if (!S_ISREG(inode->i_mode))
230 if (file->f_flags & O_LOV_DELAY_CREATE ||
231 !(file->f_mode & FMODE_WRITE)) {
232 CDEBUG(D_INODE, "object creation was delayed\n");
236 file->f_flags &= ~O_LOV_DELAY_CREATE;
239 req = it->d.lustre.it_data;
240 ptlrpc_req_finished(req);
242 ll_open_complete(inode);
246 /* Fills the obdo with the attributes for the inode defined by lsm */
247 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
250 struct ptlrpc_request_set *set;
254 LASSERT(lsm != NULL);
256 memset(oa, 0, sizeof *oa);
257 oa->o_id = lsm->lsm_object_id;
258 oa->o_mode = S_IFREG;
259 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
260 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
263 set = ptlrpc_prep_set();
265 CERROR ("ENOMEM allocing request set\n");
268 rc = obd_getattr_async(exp, oa, lsm, set);
270 rc = ptlrpc_set_wait(set);
271 ptlrpc_set_destroy(set);
276 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
277 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
281 static inline void ll_remove_suid(struct inode *inode)
285 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
286 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
288 /* was any of the uid bits set? */
289 mode &= inode->i_mode;
290 if (mode && !capable(CAP_FSETID)) {
291 inode->i_mode &= ~mode;
292 // XXX careful here - we cannot change the size
296 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
298 struct ll_inode_info *lli = ll_i2info(inode);
299 struct lov_stripe_md *lsm = lli->lli_smd;
300 struct obd_export *exp = ll_i2obdexp(inode);
303 struct ldlm_lock *lock;
304 struct lov_stripe_md *lsm;
305 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
306 __u32 stripe, vallen = sizeof(stripe);
310 if (lsm->lsm_stripe_count == 1)
313 /* get our offset in the lov */
314 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
316 CERROR("obd_get_info: rc = %d\n", rc);
319 LASSERT(stripe < lsm->lsm_stripe_count);
323 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
324 * we get a lock cancellation for each stripe, so we have to map the obd's
325 * region back onto the stripes in the file that it held.
327 * No one can dirty the extent until we've finished our work and they can
328 * enqueue another lock. The DLM protects us from ll_file_read/write here,
329 * but other kernel actors could have pages locked.
331 * Called with the DLM lock held. */
332 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
333 struct ldlm_lock *lock, __u32 stripe)
335 ldlm_policy_data_t tmpex;
336 unsigned long start, end, count, skip, i, j;
338 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
339 struct lustre_handle lockh;
342 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
343 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
344 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
347 /* our locks are page granular thanks to osc_enqueue, we invalidate the
349 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
350 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
354 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
355 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
356 if (lsm->lsm_stripe_count > 1) {
357 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
358 skip = (lsm->lsm_stripe_count - 1) * count;
359 start += start/count * skip + stripe * count;
361 end += end/count * skip + stripe * count;
363 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
366 i = (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
370 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
371 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
372 count, skip, end, discard ? " (DISCARDING)" : "");
374 /* this is the simplistic implementation of page eviction at
375 * cancelation. It is careful to get races with other page
376 * lockers handled correctly. fixes from bug 20 will make it
377 * more efficient by associating locks with pages and with
378 * batching writeback under the lock explicitly. */
379 for (i = start, j = start % count; i <= end;
380 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
382 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
388 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
389 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
390 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
393 ll_pgcache_lock(inode->i_mapping);
394 if (list_empty(&inode->i_mapping->dirty_pages) &&
395 list_empty(&inode->i_mapping->clean_pages) &&
396 list_empty(&inode->i_mapping->locked_pages)) {
397 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
398 ll_pgcache_unlock(inode->i_mapping);
401 ll_pgcache_unlock(inode->i_mapping);
403 conditional_schedule();
405 page = find_get_page(inode->i_mapping, i);
408 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
409 i, tmpex.l_extent.start);
412 /* page->mapping to check with racing against teardown */
413 if (page->mapping && PageDirty(page) && !discard) {
414 ClearPageDirty(page);
415 LL_CDEBUG_PAGE(D_PAGE, page, "found dirty\n");
416 ll_pgcache_lock(inode->i_mapping);
417 list_del(&page->list);
418 list_add(&page->list, &inode->i_mapping->locked_pages);
419 ll_pgcache_unlock(inode->i_mapping);
421 rc = ll_call_writepage(inode, page);
423 CERROR("writepage of page %p failed: %d\n",
425 /* either waiting for io to complete or reacquiring
426 * the lock that the failed writepage released */
430 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
431 /* check to see if another DLM lock covers this page */
432 ldlm_lock2handle(lock, &lockh);
433 rc2 = ldlm_lock_match(NULL,
434 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
436 NULL, 0, &tmpex, 0, &lockh);
437 if (rc2 == 0 && page->mapping != NULL) {
438 // checking again to account for writeback's lock_page()
439 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
440 ll_truncate_complete_page(page);
443 page_cache_release(page);
445 LASSERTF(tmpex.l_extent.start <=
446 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
447 lock->l_policy_data.l_extent.end + 1),
448 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
449 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
454 static int ll_extent_lock_callback(struct ldlm_lock *lock,
455 struct ldlm_lock_desc *new, void *data,
458 struct lustre_handle lockh = { 0 };
462 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
463 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
468 case LDLM_CB_BLOCKING:
469 ldlm_lock2handle(lock, &lockh);
470 rc = ldlm_cli_cancel(&lockh);
472 CERROR("ldlm_cli_cancel failed: %d\n", rc);
474 case LDLM_CB_CANCELING: {
476 struct ll_inode_info *lli;
477 struct lov_stripe_md *lsm;
481 /* This lock wasn't granted, don't try to evict pages */
482 if (lock->l_req_mode != lock->l_granted_mode)
485 inode = ll_inode_from_lock(lock);
488 lli = ll_i2info(inode);
491 if (lli->lli_smd == NULL)
495 stripe = ll_lock_to_stripe_offset(inode, lock);
496 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
499 kms = ldlm_extent_shift_kms(lock,
500 lsm->lsm_oinfo[stripe].loi_kms);
501 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
502 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
503 lsm->lsm_oinfo[stripe].loi_kms, kms);
504 lsm->lsm_oinfo[stripe].loi_kms = kms;
506 //ll_try_done_writing(inode);
519 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
521 /* XXX ALLOCATE - 160 bytes */
522 struct inode *inode = ll_inode_from_lock(lock);
523 struct ll_inode_info *lli = ll_i2info(inode);
524 struct lustre_handle lockh = { 0 };
529 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
530 LDLM_FL_BLOCK_CONV)) {
531 LBUG(); /* not expecting any blocked async locks yet */
532 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
534 ldlm_lock_dump(D_OTHER, lock, 0);
535 ldlm_reprocess_all(lock->l_resource);
539 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
541 stripe = ll_lock_to_stripe_offset(inode, lock);
543 if (lock->l_lvb_len) {
544 struct lov_stripe_md *lsm = lli->lli_smd;
546 lvb = lock->l_lvb_data;
547 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
550 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
551 kms = ldlm_extent_shift_kms(NULL, kms);
552 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
553 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
554 lsm->lsm_oinfo[stripe].loi_kms, kms);
555 lsm->lsm_oinfo[stripe].loi_kms = kms;
560 wake_up(&lock->l_waitq);
562 ldlm_lock2handle(lock, &lockh);
563 ldlm_lock_decref(&lockh, LCK_PR);
568 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
570 struct ptlrpc_request *req = reqp;
571 struct inode *inode = ll_inode_from_lock(lock);
572 struct ll_inode_info *lli;
574 int rc, size = sizeof(*lvb), stripe = 0;
578 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
579 lli = ll_i2info(inode);
581 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
582 if (lli->lli_smd == NULL)
583 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
585 /* First, find out which stripe index this lock corresponds to. */
586 if (lli->lli_smd->lsm_stripe_count > 1)
587 stripe = ll_lock_to_stripe_offset(inode, lock);
589 rc = lustre_pack_reply(req, 1, &size, NULL);
591 CERROR("lustre_pack_reply: %d\n", rc);
595 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
596 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
598 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
599 inode->i_size, stripe, lvb->lvb_size);
605 /* These errors are normal races, so we don't want to fill the console
606 * with messages by calling ptlrpc_error() */
607 if (rc == -ELDLM_NO_LOCK_DATA)
608 lustre_pack_reply(req, 0, NULL, NULL);
614 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
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;
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_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
642 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64"\n", lvb->lvb_size);
644 obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
649 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
650 struct lov_stripe_md *lsm, int mode,
651 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
654 struct ll_sb_info *sbi = ll_i2sbi(inode);
658 LASSERT(lockh->cookie == 0);
660 /* XXX phil: can we do this? won't it screw the file size up? */
661 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
662 (sbi->ll_flags & LL_SBI_NOLCK))
665 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
666 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
668 rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
669 &ast_flags, ll_extent_lock_callback,
670 ldlm_completion_ast, ll_glimpse_callback, inode,
671 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
675 if (policy->l_extent.start == 0 &&
676 policy->l_extent.end == OBD_OBJECT_EOF)
677 inode->i_size = lov_merge_size(lsm, 1);
679 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
684 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
685 struct lov_stripe_md *lsm, int mode,
686 struct lustre_handle *lockh)
688 struct ll_sb_info *sbi = ll_i2sbi(inode);
692 /* XXX phil: can we do this? won't it screw the file size up? */
693 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
694 (sbi->ll_flags & LL_SBI_NOLCK))
697 rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
702 static ssize_t ll_file_read(struct file *filp, char *buf, size_t count,
705 struct ll_file_data *fd = filp->private_data;
706 struct inode *inode = filp->f_dentry->d_inode;
707 struct ll_inode_info *lli = ll_i2info(inode);
708 struct lov_stripe_md *lsm = lli->lli_smd;
709 struct lustre_handle lockh = { 0 };
710 ldlm_policy_data_t policy;
715 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
716 inode->i_ino, inode->i_generation, inode, count, *ppos);
718 /* "If nbyte is 0, read() will return 0 and have no other results."
719 * -- Single Unix Spec */
723 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
729 policy.l_extent.start = *ppos;
730 policy.l_extent.end = *ppos + count - 1;
732 err = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
733 (filp->f_flags & O_NONBLOCK)?LDLM_FL_BLOCK_NOWAIT:
738 kms = lov_merge_size(lsm, 1);
739 if (policy.l_extent.end > kms) {
740 /* A glimpse is necessary to determine whether we return a short
741 * read or some zeroes at the end of the buffer */
743 retval = ll_glimpse_size(inode, &lvb);
746 inode->i_size = lvb.lvb_size;
751 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
752 inode->i_ino, count, *ppos, inode->i_size);
754 /* turn off the kernel's read-ahead */
755 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
758 filp->f_ra.ra_pages = 0;
760 retval = generic_file_read(filp, buf, count, ppos);
763 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
768 * Write to a file (through the page cache).
770 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
773 struct ll_file_data *fd = file->private_data;
774 struct inode *inode = file->f_dentry->d_inode;
775 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
776 struct lustre_handle lockh = { 0 };
777 ldlm_policy_data_t policy;
778 loff_t maxbytes = ll_file_maxbytes(inode);
783 if (file->f_flags & O_NONBLOCK)
784 nonblock = LDLM_FL_BLOCK_NOWAIT;
785 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
786 inode->i_ino, inode->i_generation, inode, count, *ppos);
788 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
790 /* POSIX, but surprised the VFS doesn't check this already */
794 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
795 * called on the file, don't fail the below assertion (bug 2388). */
796 if (file->f_flags & O_LOV_DELAY_CREATE && lsm == NULL)
801 if (file->f_flags & O_APPEND) {
802 policy.l_extent.start = 0;
803 policy.l_extent.end = OBD_OBJECT_EOF;
805 policy.l_extent.start = *ppos;
806 policy.l_extent.end = *ppos + count - 1;
809 err = ll_extent_lock(fd, inode, lsm, LCK_PW, &policy, &lockh, nonblock);
813 /* this is ok, g_f_w will overwrite this under i_sem if it races
814 * with a local truncate, it just makes our maxbyte checking easier */
815 if (file->f_flags & O_APPEND)
816 *ppos = inode->i_size;
818 if (*ppos >= maxbytes) {
819 if (count || *ppos > maxbytes) {
820 send_sig(SIGXFSZ, current, 0);
821 GOTO(out, retval = -EFBIG);
824 if (*ppos + count > maxbytes)
825 count = maxbytes - *ppos;
827 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
828 inode->i_ino, count, *ppos);
830 /* generic_file_write handles O_APPEND after getting i_sem */
831 retval = generic_file_write(file, buf, count, ppos);
834 ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
835 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
836 retval > 0 ? retval : 0);
840 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
843 struct ll_inode_info *lli = ll_i2info(inode);
844 struct obd_export *exp = ll_i2obdexp(inode);
845 struct ll_recreate_obj ucreatp;
846 struct obd_trans_info oti = { 0 };
847 struct obdo *oa = NULL;
850 struct lov_stripe_md *lsm, *lsm2;
853 if (!capable (CAP_SYS_ADMIN))
856 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
857 sizeof(struct ll_recreate_obj));
866 down(&lli->lli_open_sem);
869 up(&lli->lli_open_sem);
873 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
874 (lsm->lsm_stripe_count));
876 OBD_ALLOC(lsm2, lsm_size);
878 up(&lli->lli_open_sem);
883 oa->o_id = ucreatp.lrc_id;
884 oa->o_nlink = ucreatp.lrc_ost_idx;
885 oa->o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS;
886 oa->o_flags |= OBD_FL_RECREATE_OBJS;
887 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
888 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
890 oti.oti_objid = NULL;
891 memcpy(lsm2, lsm, lsm_size);
892 rc = obd_create(exp, oa, &lsm2, &oti);
894 up(&lli->lli_open_sem);
895 OBD_FREE(lsm2, lsm_size);
900 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
901 int flags, struct lov_user_md *lum,
904 struct ll_inode_info *lli = ll_i2info(inode);
906 struct obd_export *exp = ll_i2obdexp(inode);
907 struct lov_stripe_md *lsm;
908 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
909 struct ptlrpc_request *req = NULL;
914 down(&lli->lli_open_sem);
917 up(&lli->lli_open_sem);
918 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
923 f = get_empty_filp();
927 f->f_dentry = file->f_dentry;
928 f->f_vfsmnt = file->f_vfsmnt;
930 rc = ll_intent_file_open(f, lum, lum_size, &oit);
933 if (it_disposition(&oit, DISP_LOOKUP_NEG))
935 req = oit.d.lustre.it_data;
936 rc = oit.d.lustre.it_status;
941 rc = mdc_req2lustre_md(req, 1, exp, &md);
944 ll_update_inode(f->f_dentry->d_inode, md.body, md.lsm);
946 rc = ll_local_open(f, &oit);
949 ll_intent_release(&oit);
951 rc = ll_file_release(f->f_dentry->d_inode, f);
956 up(&lli->lli_open_sem);
958 ptlrpc_req_finished(req);
962 static int ll_lov_setea(struct inode *inode, struct file *file,
965 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
966 struct lov_user_md *lump;
967 int lum_size = sizeof(struct lov_user_md) +
968 sizeof(struct lov_user_ost_data);
972 if (!capable (CAP_SYS_ADMIN))
975 OBD_ALLOC(lump, lum_size);
979 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
981 OBD_FREE(lump, lum_size);
985 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
987 OBD_FREE(lump, lum_size);
991 static int ll_lov_setstripe(struct inode *inode, struct file *file,
994 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
996 int flags = FMODE_WRITE;
999 /* Bug 1152: copy properly when this is no longer true */
1000 LASSERT(sizeof(lum) == sizeof(*lump));
1001 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1002 rc = copy_from_user(&lum, lump, sizeof(lum));
1006 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1010 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1012 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1017 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2obdexp(inode), 0, lsm,
1021 static int ll_get_grouplock(struct inode *inode, struct file *file,
1024 struct ll_file_data *fd = file->private_data;
1025 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1026 .end = OBD_OBJECT_EOF}};
1027 struct lustre_handle lockh = { 0 };
1028 struct ll_inode_info *lli = ll_i2info(inode);
1029 struct lov_stripe_md *lsm = lli->lli_smd;
1034 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1038 policy.l_extent.gid = arg;
1039 if (file->f_flags & O_NONBLOCK)
1040 flags = LDLM_FL_BLOCK_NOWAIT;
1042 err = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags);
1046 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1048 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1053 static int ll_put_grouplock(struct inode *inode, struct file *file,
1056 struct ll_file_data *fd = file->private_data;
1057 struct ll_inode_info *lli = ll_i2info(inode);
1058 struct lov_stripe_md *lsm = lli->lli_smd;
1062 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1063 /* Ugh, it's already unlocked. */
1067 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1070 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1072 err = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1077 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1082 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1085 struct ll_file_data *fd = file->private_data;
1089 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1090 inode->i_generation, inode, cmd);
1092 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1095 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1097 case LL_IOC_GETFLAGS:
1098 /* Get the current value of the file flags */
1099 return put_user(fd->fd_flags, (int *)arg);
1100 case LL_IOC_SETFLAGS:
1101 case LL_IOC_CLRFLAGS:
1102 /* Set or clear specific file flags */
1103 /* XXX This probably needs checks to ensure the flags are
1104 * not abused, and to handle any flag side effects.
1106 if (get_user(flags, (int *) arg))
1109 if (cmd == LL_IOC_SETFLAGS)
1110 fd->fd_flags |= flags;
1112 fd->fd_flags &= ~flags;
1114 case LL_IOC_LOV_SETSTRIPE:
1115 RETURN(ll_lov_setstripe(inode, file, arg));
1116 case LL_IOC_LOV_SETEA:
1117 RETURN(ll_lov_setea(inode, file, arg));
1118 case LL_IOC_LOV_GETSTRIPE:
1119 RETURN(ll_lov_getstripe(inode, arg));
1120 case LL_IOC_RECREATE_OBJ:
1121 RETURN(ll_lov_recreate_obj(inode, file, arg));
1122 case EXT3_IOC_GETFLAGS:
1123 case EXT3_IOC_SETFLAGS:
1124 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1125 case LL_IOC_GROUP_LOCK:
1126 RETURN(ll_get_grouplock(inode, file, arg));
1127 case LL_IOC_GROUP_UNLOCK:
1128 RETURN(ll_put_grouplock(inode, file, arg));
1129 /* We need to special case any other ioctls we want to handle,
1130 * to send them to the MDS/OST as appropriate and to properly
1131 * network encode the arg field.
1132 case EXT2_IOC_GETVERSION_OLD:
1133 case EXT2_IOC_GETVERSION_NEW:
1134 case EXT2_IOC_SETVERSION_OLD:
1135 case EXT2_IOC_SETVERSION_NEW:
1138 RETURN( obd_iocontrol(cmd, ll_i2obdexp(inode), 0, NULL,
1143 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1145 struct inode *inode = file->f_dentry->d_inode;
1146 struct ll_file_data *fd = file->private_data;
1147 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1148 struct lustre_handle lockh = {0};
1151 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1152 inode->i_generation, inode,
1153 offset + ((origin==2) ? inode->i_size : file->f_pos));
1155 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1156 if (origin == 2) { /* SEEK_END */
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 err = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1166 if (err != ELDLM_OK)
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 = mdc_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_valid = OBD_MD_FLID;
1223 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
1224 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
1226 err = obd_sync(ll_i2sbi(inode)->ll_osc_exp, oa, lsm,
1236 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1238 struct inode *inode = file->f_dentry->d_inode;
1239 struct ll_sb_info *sbi = ll_i2sbi(inode);
1240 struct obd_device *obddev;
1241 struct ldlm_res_id res_id =
1242 { .name = {inode->i_ino, inode->i_generation, LDLM_FLOCK} };
1243 struct lustre_handle lockh = {0};
1244 ldlm_policy_data_t flock;
1245 ldlm_mode_t mode = 0;
1250 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1251 inode->i_ino, file_lock);
1253 flock.l_flock.pid = file_lock->fl_pid;
1254 flock.l_flock.start = file_lock->fl_start;
1255 flock.l_flock.end = file_lock->fl_end;
1257 switch (file_lock->fl_type) {
1262 /* An unlock request may or may not have any relation to
1263 * existing locks so we may not be able to pass a lock handle
1264 * via a normal ldlm_lock_cancel() request. The request may even
1265 * unlock a byte range in the middle of an existing lock. In
1266 * order to process an unlock request we need all of the same
1267 * information that is given with a normal read or write record
1268 * lock request. To avoid creating another ldlm unlock (cancel)
1269 * message we'll treat a LCK_NL flock request as an unlock. */
1276 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1285 flags = LDLM_FL_BLOCK_NOWAIT;
1288 flags = LDLM_FL_TEST_LOCK;
1289 /* Save the old mode so that if the mode in the lock changes we
1290 * can decrement the appropriate reader or writer refcount. */
1291 file_lock->fl_type = mode;
1294 CERROR("unknown fcntl lock command: %d\n", cmd);
1298 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1299 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1300 flags, mode, flock.l_flock.start, flock.l_flock.end);
1302 obddev = sbi->ll_mdc_exp->exp_obd;
1303 rc = ldlm_cli_enqueue(sbi->ll_mdc_exp, NULL, obddev->obd_namespace,
1304 res_id, LDLM_FLOCK, &flock, mode, &flags,
1305 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1306 NULL, 0, NULL, &lockh);
1310 static int ll_have_md_lock(struct dentry *de)
1312 struct ll_sb_info *sbi = ll_s2sbi(de->d_sb);
1313 struct lustre_handle lockh;
1314 struct ldlm_res_id res_id = { .name = {0} };
1315 struct obd_device *obddev;
1322 obddev = sbi->ll_mdc_exp->exp_obd;
1323 res_id.name[0] = de->d_inode->i_ino;
1324 res_id.name[1] = de->d_inode->i_generation;
1326 CDEBUG(D_INFO, "trying to match res "LPU64"\n", res_id.name[0]);
1328 /* FIXME use LDLM_FL_TEST_LOCK instead */
1329 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1330 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
1331 NULL, LCK_PR, &lockh)) {
1332 ldlm_lock_decref(&lockh, LCK_PR);
1336 if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN,
1337 NULL, LCK_PW, &lockh)) {
1338 ldlm_lock_decref(&lockh, LCK_PW);
1344 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
1346 struct inode *inode = dentry->d_inode;
1347 struct ll_inode_info *lli;
1348 struct lov_stripe_md *lsm;
1353 CERROR("REPORT THIS LINE TO PETER\n");
1356 lli = ll_i2info(inode);
1357 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s\n",
1358 inode->i_ino, inode->i_generation, inode, dentry->d_name.name);
1359 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1360 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1363 if (!ll_have_md_lock(dentry)) {
1364 struct ptlrpc_request *req = NULL;
1365 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
1367 unsigned long valid = 0;
1370 if (S_ISREG(inode->i_mode)) {
1371 ealen = obd_size_diskmd(sbi->ll_osc_exp, NULL);
1372 valid |= OBD_MD_FLEASIZE;
1374 ll_inode2fid(&fid, inode);
1375 rc = mdc_getattr(sbi->ll_mdc_exp, &fid, valid, ealen, &req);
1377 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
1380 rc = ll_prep_inode(sbi->ll_osc_exp, &inode, req, 0, NULL);
1382 ptlrpc_req_finished(req);
1385 ptlrpc_req_finished(req);
1389 if (lsm == NULL) /* object not yet allocated, don't validate size */
1392 /* ll_glimpse_size will prefer locally cached writes if they extend
1397 rc = ll_glimpse_size(inode, &lvb);
1398 inode->i_size = lvb.lvb_size;
1403 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1404 int ll_getattr(struct vfsmount *mnt, struct dentry *de,
1405 struct lookup_intent *it, struct kstat *stat)
1408 struct inode *inode = de->d_inode;
1410 res = ll_inode_revalidate_it(de, it);
1411 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1416 stat->dev = inode->i_sb->s_dev;
1417 stat->ino = inode->i_ino;
1418 stat->mode = inode->i_mode;
1419 stat->nlink = inode->i_nlink;
1420 stat->uid = inode->i_uid;
1421 stat->gid = inode->i_gid;
1422 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1423 stat->atime = inode->i_atime;
1424 stat->mtime = inode->i_mtime;
1425 stat->ctime = inode->i_ctime;
1426 stat->size = inode->i_size;
1427 stat->blksize = inode->i_blksize;
1428 stat->blocks = inode->i_blocks;
1433 struct file_operations ll_file_operations = {
1435 write: ll_file_write,
1436 ioctl: ll_file_ioctl,
1438 release: ll_file_release,
1439 mmap: generic_file_mmap,
1440 llseek: ll_file_seek,
1442 //lock: ll_file_flock
1445 struct inode_operations ll_file_inode_operations = {
1446 setattr_raw: ll_setattr_raw,
1447 setattr: ll_setattr,
1448 truncate: ll_truncate,
1449 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1450 getattr_it: ll_getattr,
1452 revalidate_it: ll_inode_revalidate_it,
1456 struct inode_operations ll_special_inode_operations = {
1457 setattr_raw: ll_setattr_raw,
1458 setattr: ll_setattr,
1459 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1460 getattr_it: ll_getattr,
1462 revalidate_it: ll_inode_revalidate_it,