1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
7 * Author: Andreas Dilger <adilger@clusterfs.com>
9 * This file is part of Lustre, http://www.lustre.org.
11 * Lustre is free software; you can redistribute it and/or
12 * modify it under the terms of version 2 of the GNU General Public
13 * License as published by the Free Software Foundation.
15 * Lustre is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with Lustre; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #define DEBUG_SUBSYSTEM S_LLITE
26 #include <linux/lustre_dlm.h>
27 #include <linux/lustre_lite.h>
28 #include <linux/pagemap.h>
29 #include <linux/file.h>
30 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
31 #include <linux/lustre_compat25.h>
33 #include "llite_internal.h"
34 #include <linux/obd_lov.h>
36 int ll_mdc_close(struct obd_export *mdc_exp, struct inode *inode,
39 struct ll_file_data *fd = file->private_data;
40 struct ptlrpc_request *req = NULL;
41 struct obd_client_handle *och = &fd->fd_mds_och;
46 /* clear group lock, if present */
47 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
48 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
49 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
50 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
54 obdo.o_id = inode->i_ino;
55 obdo.o_valid = OBD_MD_FLID;
56 obdo_from_inode(&obdo, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
57 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
58 OBD_MD_FLATIME | OBD_MD_FLMTIME |
60 if (0 /* ll_is_inode_dirty(inode) */) {
61 obdo.o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
62 obdo.o_valid |= OBD_MD_FLFLAGS;
64 obdo.o_mds = ll_i2info(inode)->lli_mds;
65 rc = md_close(mdc_exp, &obdo, och, &req);
68 /* We are the last writer, so the MDS has instructed us to get
69 * the file size and any write cookies, then close again. */
70 //ll_queue_done_writing(inode);
73 CERROR("inode %lu mdc close failed: rc = %d\n",
77 rc = ll_objects_destroy(req, file->f_dentry->d_inode, 1);
79 CERROR("inode %lu ll_objects destroy: rc = %d\n",
83 mdc_clear_open_replay_data(mdc_exp, och);
84 ptlrpc_req_finished(req);
85 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
86 file->private_data = NULL;
87 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof *fd);
92 /* While this returns an error code, fput() the caller does not, so we need
93 * to make every effort to clean up all of our state here. Also, applications
94 * rarely check close errors and even if an error is returned they will not
95 * re-try the close call.
97 int ll_file_release(struct inode *inode, struct file *file)
99 struct ll_file_data *fd;
100 struct ll_sb_info *sbi = ll_i2sbi(inode);
104 CDEBUG(D_VFSTRACE, "VFS Op:inode=%u/%lu/%u(%p)\n",
105 ll_i2info(inode)->lli_mds, inode->i_ino,
106 inode->i_generation, inode);
108 /* don't do anything for / */
109 if (inode->i_sb->s_root == file->f_dentry)
112 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
113 fd = (struct ll_file_data *)file->private_data;
116 rc = ll_mdc_close(sbi->ll_mdc_exp, inode, file);
120 static int ll_intent_file_open(struct file *file, void *lmm,
121 int lmmsize, struct lookup_intent *itp)
123 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
124 struct lustre_handle lockh;
125 struct mdc_op_data data;
126 struct dentry *parent = file->f_dentry->d_parent;
127 const char *name = file->f_dentry->d_name.name;
128 const int len = file->f_dentry->d_name.len;
134 ll_prepare_mdc_op_data(&data, parent->d_inode, NULL, name, len, O_RDWR);
136 rc = md_enqueue(sbi->ll_mdc_exp, LDLM_IBITS, itp, LCK_PR, &data,
137 &lockh, lmm, lmmsize, ldlm_completion_ast,
138 ll_mdc_blocking_ast, NULL);
140 if (itp->d.lustre.it_lock_mode)
141 memcpy(&itp->d.lustre.it_lock_handle,
142 &lockh, sizeof(lockh));
144 CERROR("lock enqueue: err: %d\n", rc);
150 int ll_local_open(struct file *file, struct lookup_intent *it)
152 struct ptlrpc_request *req = it->d.lustre.it_data;
153 struct ll_inode_info *lli = ll_i2info(file->f_dentry->d_inode);
154 struct obd_export *mdc_exp = ll_i2mdcexp(file->f_dentry->d_inode);
155 struct ll_file_data *fd;
156 struct mds_body *body;
159 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
160 LASSERT (body != NULL); /* reply already checked out */
161 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
163 LASSERT(!file->private_data);
165 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
166 /* We can't handle this well without reorganizing ll_file_open and
167 * ll_mdc_close, so don't even try right now. */
170 memcpy(&fd->fd_mds_och.och_fh, &body->handle, sizeof(body->handle));
171 fd->fd_mds_och.och_magic = OBD_CLIENT_HANDLE_MAGIC;
172 file->private_data = fd;
173 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
175 lli->lli_io_epoch = body->io_epoch;
177 mdc_set_open_replay_data(mdc_exp, &fd->fd_mds_och, it->d.lustre.it_data);
182 /* Open a file, and (for the very first open) create objects on the OSTs at
183 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
184 * creation or open until ll_lov_setstripe() ioctl is called. We grab
185 * lli_open_sem to ensure no other process will create objects, send the
186 * stripe MD to the MDS, or try to destroy the objects if that fails.
188 * If we already have the stripe MD locally then we don't request it in
189 * mdc_open(), by passing a lmm_size = 0.
191 * It is up to the application to ensure no other processes open this file
192 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
193 * used. We might be able to avoid races of that sort by getting lli_open_sem
194 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
195 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
197 int ll_file_open(struct inode *inode, struct file *file)
199 struct ll_inode_info *lli = ll_i2info(inode);
200 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
201 .it_flags = file->f_flags };
202 struct lov_stripe_md *lsm;
203 struct ptlrpc_request *req;
207 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
208 inode->i_generation, inode);
210 /* don't do anything for / */
211 if (inode->i_sb->s_root == file->f_dentry)
216 if (!it || !it->d.lustre.it_disposition) {
218 rc = ll_intent_file_open(file, NULL, 0, it);
223 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
224 rc = it_open_error(DISP_OPEN_OPEN, it);
228 rc = ll_local_open(file, it);
232 if (!S_ISREG(inode->i_mode))
237 if (file->f_flags & O_LOV_DELAY_CREATE ||
238 !(file->f_mode & FMODE_WRITE)) {
239 CDEBUG(D_INODE, "object creation was delayed\n");
243 file->f_flags &= ~O_LOV_DELAY_CREATE;
246 req = it->d.lustre.it_data;
247 ptlrpc_req_finished(req);
249 ll_open_complete(inode);
253 /* Fills the obdo with the attributes for the inode defined by lsm */
254 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
257 struct ptlrpc_request_set *set;
261 LASSERT(lsm != NULL);
263 memset(oa, 0, sizeof *oa);
264 oa->o_id = lsm->lsm_object_id;
265 oa->o_gr = lsm->lsm_object_gr;
266 oa->o_mode = S_IFREG;
267 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
268 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
269 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
271 set = ptlrpc_prep_set();
273 CERROR ("ENOMEM allocing request set\n");
276 rc = obd_getattr_async(exp, oa, lsm, set);
278 rc = ptlrpc_set_wait(set);
279 ptlrpc_set_destroy(set);
284 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
285 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
289 static inline void ll_remove_suid(struct inode *inode)
293 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
294 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
296 /* was any of the uid bits set? */
297 mode &= inode->i_mode;
298 if (mode && !capable(CAP_FSETID)) {
299 inode->i_mode &= ~mode;
300 // XXX careful here - we cannot change the size
304 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
306 struct ll_inode_info *lli = ll_i2info(inode);
307 struct lov_stripe_md *lsm = lli->lli_smd;
308 struct obd_export *exp = ll_i2obdexp(inode);
311 struct ldlm_lock *lock;
312 struct lov_stripe_md *lsm;
313 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
314 __u32 stripe, vallen = sizeof(stripe);
318 if (lsm->lsm_stripe_count == 1)
319 GOTO(check, stripe = 0);
321 /* get our offset in the lov */
322 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
324 CERROR("obd_get_info: rc = %d\n", rc);
327 LASSERT(stripe < lsm->lsm_stripe_count);
330 if (lsm->lsm_oinfo[stripe].loi_id != lock->l_resource->lr_name.name[0]||
331 lsm->lsm_oinfo[stripe].loi_gr != lock->l_resource->lr_name.name[2]){
332 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64,
333 lsm->lsm_oinfo[stripe].loi_id,
334 lsm->lsm_oinfo[stripe].loi_gr);
335 RETURN(-ELDLM_NO_LOCK_DATA);
341 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
342 * we get a lock cancellation for each stripe, so we have to map the obd's
343 * region back onto the stripes in the file that it held.
345 * No one can dirty the extent until we've finished our work and they can
346 * enqueue another lock. The DLM protects us from ll_file_read/write here,
347 * but other kernel actors could have pages locked.
349 * Called with the DLM lock held. */
350 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
351 struct ldlm_lock *lock, __u32 stripe)
353 ldlm_policy_data_t tmpex;
354 unsigned long start, end, count, skip, i, j;
356 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
357 struct lustre_handle lockh;
360 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
361 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
362 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
365 /* our locks are page granular thanks to osc_enqueue, we invalidate the
367 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
368 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
372 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
373 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
374 if (lsm->lsm_stripe_count > 1) {
375 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
376 skip = (lsm->lsm_stripe_count - 1) * count;
377 start += start/count * skip + stripe * count;
379 end += end/count * skip + stripe * count;
381 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
384 i = (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
388 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
389 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
390 count, skip, end, discard ? " (DISCARDING)" : "");
392 /* this is the simplistic implementation of page eviction at
393 * cancelation. It is careful to get races with other page
394 * lockers handled correctly. fixes from bug 20 will make it
395 * more efficient by associating locks with pages and with
396 * batching writeback under the lock explicitly. */
397 for (i = start, j = start % count; i <= end;
398 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
400 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
406 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
407 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
408 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
411 if (!mapping_has_pages(inode->i_mapping)) {
412 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
416 conditional_schedule();
418 page = find_get_page(inode->i_mapping, i);
421 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
422 i, tmpex.l_extent.start);
425 /* page->mapping to check with racing against teardown */
426 if (!discard && clear_page_dirty_for_io(page)) {
427 rc = ll_call_writepage(inode, page);
429 CERROR("writepage of page %p failed: %d\n",
431 /* either waiting for io to complete or reacquiring
432 * the lock that the failed writepage released */
436 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
437 /* check to see if another DLM lock covers this page */
438 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
439 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
441 &lock->l_resource->lr_name, LDLM_EXTENT,
442 &tmpex, LCK_PR | LCK_PW, &lockh);
443 if (rc2 == 0 && page->mapping != NULL) {
444 // checking again to account for writeback's lock_page()
445 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
446 ll_truncate_complete_page(page);
449 page_cache_release(page);
451 LASSERTF(tmpex.l_extent.start <=
452 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
453 lock->l_policy_data.l_extent.end + 1),
454 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
455 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
460 static int ll_extent_lock_callback(struct ldlm_lock *lock,
461 struct ldlm_lock_desc *new, void *data,
464 struct lustre_handle lockh = { 0 };
468 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
469 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
474 case LDLM_CB_BLOCKING:
475 ldlm_lock2handle(lock, &lockh);
476 rc = ldlm_cli_cancel(&lockh);
478 CERROR("ldlm_cli_cancel failed: %d\n", rc);
480 case LDLM_CB_CANCELING: {
482 struct ll_inode_info *lli;
483 struct lov_stripe_md *lsm;
487 /* This lock wasn't granted, don't try to evict pages */
488 if (lock->l_req_mode != lock->l_granted_mode)
491 inode = ll_inode_from_lock(lock);
494 lli = ll_i2info(inode);
497 if (lli->lli_smd == NULL)
501 stripe = ll_lock_to_stripe_offset(inode, lock);
504 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
507 kms = ldlm_extent_shift_kms(lock,
508 lsm->lsm_oinfo[stripe].loi_kms);
510 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
511 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
512 lsm->lsm_oinfo[stripe].loi_kms, kms);
513 lsm->lsm_oinfo[stripe].loi_kms = kms;
515 //ll_try_done_writing(inode);
528 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
530 /* XXX ALLOCATE - 160 bytes */
531 struct inode *inode = ll_inode_from_lock(lock);
532 struct ll_inode_info *lli = ll_i2info(inode);
533 struct lustre_handle lockh = { 0 };
538 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
539 LDLM_FL_BLOCK_CONV)) {
540 LBUG(); /* not expecting any blocked async locks yet */
541 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
543 ldlm_lock_dump(D_OTHER, lock, 0);
544 ldlm_reprocess_all(lock->l_resource);
548 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
550 stripe = ll_lock_to_stripe_offset(inode, lock);
554 if (lock->l_lvb_len) {
555 struct lov_stripe_md *lsm = lli->lli_smd;
557 lvb = lock->l_lvb_data;
558 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
561 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
562 kms = ldlm_extent_shift_kms(NULL, kms);
563 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
564 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
565 lsm->lsm_oinfo[stripe].loi_kms, kms);
566 lsm->lsm_oinfo[stripe].loi_kms = kms;
572 wake_up(&lock->l_waitq);
574 ldlm_lock2handle(lock, &lockh);
575 ldlm_lock_decref(&lockh, LCK_PR);
580 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
582 struct ptlrpc_request *req = reqp;
583 struct inode *inode = ll_inode_from_lock(lock);
584 struct ll_inode_info *lli;
586 int rc, size = sizeof(*lvb), stripe;
590 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
591 lli = ll_i2info(inode);
593 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
594 if (lli->lli_smd == NULL)
595 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
597 /* First, find out which stripe index this lock corresponds to. */
598 stripe = ll_lock_to_stripe_offset(inode, lock);
600 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
602 rc = lustre_pack_reply(req, 1, &size, NULL);
604 CERROR("lustre_pack_reply: %d\n", rc);
608 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
609 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
611 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
612 inode->i_size, stripe, lvb->lvb_size);
618 /* These errors are normal races, so we don't want to fill the console
619 * with messages by calling ptlrpc_error() */
620 if (rc == -ELDLM_NO_LOCK_DATA)
621 lustre_pack_reply(req, 0, NULL, NULL);
627 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
628 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
629 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
631 /* NB: lov_merge_size will prefer locally cached writes if they extend the
632 * file (because it prefers KMS over RSS when larger) */
633 int ll_glimpse_size(struct inode *inode, struct ost_lvb *lvb)
635 struct ll_inode_info *lli = ll_i2info(inode);
636 struct ll_sb_info *sbi = ll_i2sbi(inode);
637 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
638 struct lustre_handle lockh = { 0 };
639 int rc, flags = LDLM_FL_HAS_INTENT;
642 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
644 rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy,
645 LCK_PR, &flags, ll_extent_lock_callback,
646 ldlm_completion_ast, ll_glimpse_callback, inode,
647 sizeof(*lvb), lustre_swab_ost_lvb, &lockh);
649 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
650 RETURN(rc > 0 ? -EIO : rc);
653 lvb->lvb_size = lov_merge_size(lli->lli_smd, 0);
654 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
655 //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime);
657 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
658 lvb->lvb_size, lvb->lvb_blocks);
660 obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh);
665 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
666 struct lov_stripe_md *lsm, int mode,
667 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
670 struct ll_sb_info *sbi = ll_i2sbi(inode);
674 LASSERT(lockh->cookie == 0);
676 /* XXX phil: can we do this? won't it screw the file size up? */
677 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
678 (sbi->ll_flags & LL_SBI_NOLCK))
681 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
682 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
684 rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode,
685 &ast_flags, ll_extent_lock_callback,
686 ldlm_completion_ast, ll_glimpse_callback, inode,
687 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
691 if (policy->l_extent.start == 0 &&
692 policy->l_extent.end == OBD_OBJECT_EOF)
693 inode->i_size = lov_merge_size(lsm, 1);
695 //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime);
700 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
701 struct lov_stripe_md *lsm, int mode,
702 struct lustre_handle *lockh)
704 struct ll_sb_info *sbi = ll_i2sbi(inode);
708 /* XXX phil: can we do this? won't it screw the file size up? */
709 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
710 (sbi->ll_flags & LL_SBI_NOLCK))
713 rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh);
718 static ssize_t ll_file_read(struct file *filp, char *buf, size_t count,
721 struct ll_file_data *fd = filp->private_data;
722 struct inode *inode = filp->f_dentry->d_inode;
723 struct ll_inode_info *lli = ll_i2info(inode);
724 struct lov_stripe_md *lsm = lli->lli_smd;
725 struct lustre_handle lockh = { 0 };
726 ldlm_policy_data_t policy;
731 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
732 inode->i_ino, inode->i_generation, inode, count, *ppos);
734 /* "If nbyte is 0, read() will return 0 and have no other results."
735 * -- Single Unix Spec */
739 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
745 policy.l_extent.start = *ppos;
746 policy.l_extent.end = *ppos + count - 1;
748 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
749 (filp->f_flags & O_NONBLOCK) ?
750 LDLM_FL_BLOCK_NOWAIT: 0);
754 kms = lov_merge_size(lsm, 1);
755 if (*ppos + count - 1 > kms) {
756 /* A glimpse is necessary to determine whether we return a short
757 * read or some zeroes at the end of the buffer */
759 retval = ll_glimpse_size(inode, &lvb);
762 inode->i_size = lvb.lvb_size;
767 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
768 inode->i_ino, count, *ppos, inode->i_size);
770 /* turn off the kernel's read-ahead */
771 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
774 filp->f_ra.ra_pages = 0;
776 retval = generic_file_read(filp, buf, count, ppos);
779 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
784 * Write to a file (through the page cache).
786 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
789 struct ll_file_data *fd = file->private_data;
790 struct inode *inode = file->f_dentry->d_inode;
791 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
792 struct lustre_handle lockh = { 0 };
793 ldlm_policy_data_t policy;
794 loff_t maxbytes = ll_file_maxbytes(inode);
796 int nonblock = 0, rc;
798 if (file->f_flags & O_NONBLOCK)
799 nonblock = LDLM_FL_BLOCK_NOWAIT;
800 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
801 inode->i_ino, inode->i_generation, inode, count, *ppos);
803 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
805 /* POSIX, but surprised the VFS doesn't check this already */
809 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
810 * called on the file, don't fail the below assertion (bug 2388). */
811 if (file->f_flags & O_LOV_DELAY_CREATE && lsm == NULL)
816 if (file->f_flags & O_APPEND) {
817 policy.l_extent.start = 0;
818 policy.l_extent.end = OBD_OBJECT_EOF;
820 policy.l_extent.start = *ppos;
821 policy.l_extent.end = *ppos + count - 1;
824 rc = ll_extent_lock(fd, inode, lsm, LCK_PW, &policy, &lockh, nonblock);
828 /* this is ok, g_f_w will overwrite this under i_sem if it races
829 * with a local truncate, it just makes our maxbyte checking easier */
830 if (file->f_flags & O_APPEND)
831 *ppos = inode->i_size;
833 if (*ppos >= maxbytes) {
834 if (count || *ppos > maxbytes) {
835 send_sig(SIGXFSZ, current, 0);
836 GOTO(out, retval = -EFBIG);
839 if (*ppos + count > maxbytes)
840 count = maxbytes - *ppos;
842 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
843 inode->i_ino, count, *ppos);
845 /* generic_file_write handles O_APPEND after getting i_sem */
846 retval = generic_file_write(file, buf, count, ppos);
849 ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh);
850 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
851 retval > 0 ? retval : 0);
855 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
858 struct ll_inode_info *lli = ll_i2info(inode);
859 struct obd_export *exp = ll_i2obdexp(inode);
860 struct ll_recreate_obj ucreatp;
861 struct obd_trans_info oti = { 0 };
862 struct obdo *oa = NULL;
865 struct lov_stripe_md *lsm, *lsm2;
868 if (!capable (CAP_SYS_ADMIN))
871 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
872 sizeof(struct ll_recreate_obj));
881 down(&lli->lli_open_sem);
884 up(&lli->lli_open_sem);
888 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
889 (lsm->lsm_stripe_count));
891 OBD_ALLOC(lsm2, lsm_size);
893 up(&lli->lli_open_sem);
898 oa->o_id = ucreatp.lrc_id;
899 oa->o_nlink = ucreatp.lrc_ost_idx;
900 oa->o_gr = ucreatp.lrc_group;
901 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLFLAGS;
902 oa->o_flags |= OBD_FL_RECREATE_OBJS;
903 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
904 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
906 oti.oti_objid = NULL;
907 memcpy(lsm2, lsm, lsm_size);
908 rc = obd_create(exp, oa, &lsm2, &oti);
910 up(&lli->lli_open_sem);
911 OBD_FREE(lsm2, lsm_size);
916 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
917 int flags, struct lov_user_md *lum,
920 struct ll_inode_info *lli = ll_i2info(inode);
922 struct obd_export *exp = ll_i2obdexp(inode);
923 struct lov_stripe_md *lsm;
924 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
925 struct ptlrpc_request *req = NULL;
930 down(&lli->lli_open_sem);
933 up(&lli->lli_open_sem);
934 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
939 f = get_empty_filp();
943 f->f_dentry = file->f_dentry;
944 f->f_vfsmnt = file->f_vfsmnt;
946 rc = ll_intent_file_open(f, lum, lum_size, &oit);
949 if (it_disposition(&oit, DISP_LOOKUP_NEG))
951 req = oit.d.lustre.it_data;
952 rc = oit.d.lustre.it_status;
957 rc = mdc_req2lustre_md(ll_i2mdcexp(inode), req, 1, exp, &md);
960 ll_update_inode(f->f_dentry->d_inode, &md);
962 rc = ll_local_open(f, &oit);
965 ll_intent_release(&oit);
967 rc = ll_file_release(f->f_dentry->d_inode, f);
972 up(&lli->lli_open_sem);
974 ptlrpc_req_finished(req);
978 static int ll_lov_setea(struct inode *inode, struct file *file,
981 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
982 struct lov_user_md *lump;
983 int lum_size = sizeof(struct lov_user_md) +
984 sizeof(struct lov_user_ost_data);
988 if (!capable (CAP_SYS_ADMIN))
991 OBD_ALLOC(lump, lum_size);
995 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
997 OBD_FREE(lump, lum_size);
1001 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1003 OBD_FREE(lump, lum_size);
1007 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1010 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1012 int flags = FMODE_WRITE;
1015 /* Bug 1152: copy properly when this is no longer true */
1016 LASSERT(sizeof(lum) == sizeof(*lump));
1017 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1018 rc = copy_from_user(&lum, lump, sizeof(lum));
1022 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1026 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1028 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1033 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2obdexp(inode), 0, lsm,
1037 static int ll_get_grouplock(struct inode *inode, struct file *file,
1040 struct ll_file_data *fd = file->private_data;
1041 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1042 .end = OBD_OBJECT_EOF}};
1043 struct lustre_handle lockh = { 0 };
1044 struct ll_inode_info *lli = ll_i2info(inode);
1045 struct lov_stripe_md *lsm = lli->lli_smd;
1049 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1053 policy.l_extent.gid = arg;
1054 if (file->f_flags & O_NONBLOCK)
1055 flags = LDLM_FL_BLOCK_NOWAIT;
1057 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags);
1061 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1063 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1068 static int ll_put_grouplock(struct inode *inode, struct file *file,
1071 struct ll_file_data *fd = file->private_data;
1072 struct ll_inode_info *lli = ll_i2info(inode);
1073 struct lov_stripe_md *lsm = lli->lli_smd;
1077 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1078 /* Ugh, it's already unlocked. */
1082 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1085 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1087 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1092 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1097 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1100 struct ll_file_data *fd = file->private_data;
1104 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1105 inode->i_generation, inode, cmd);
1107 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1110 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1112 case LL_IOC_GETFLAGS:
1113 /* Get the current value of the file flags */
1114 return put_user(fd->fd_flags, (int *)arg);
1115 case LL_IOC_SETFLAGS:
1116 case LL_IOC_CLRFLAGS:
1117 /* Set or clear specific file flags */
1118 /* XXX This probably needs checks to ensure the flags are
1119 * not abused, and to handle any flag side effects.
1121 if (get_user(flags, (int *) arg))
1124 if (cmd == LL_IOC_SETFLAGS)
1125 fd->fd_flags |= flags;
1127 fd->fd_flags &= ~flags;
1129 case LL_IOC_LOV_SETSTRIPE:
1130 RETURN(ll_lov_setstripe(inode, file, arg));
1131 case LL_IOC_LOV_SETEA:
1132 RETURN(ll_lov_setea(inode, file, arg));
1133 case LL_IOC_LOV_GETSTRIPE:
1134 RETURN(ll_lov_getstripe(inode, arg));
1135 case LL_IOC_RECREATE_OBJ:
1136 RETURN(ll_lov_recreate_obj(inode, file, arg));
1137 case EXT3_IOC_GETFLAGS:
1138 case EXT3_IOC_SETFLAGS:
1139 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1140 case LL_IOC_GROUP_LOCK:
1141 RETURN(ll_get_grouplock(inode, file, arg));
1142 case LL_IOC_GROUP_UNLOCK:
1143 RETURN(ll_put_grouplock(inode, file, arg));
1144 /* We need to special case any other ioctls we want to handle,
1145 * to send them to the MDS/OST as appropriate and to properly
1146 * network encode the arg field.
1147 case EXT2_IOC_GETVERSION_OLD:
1148 case EXT2_IOC_GETVERSION_NEW:
1149 case EXT2_IOC_SETVERSION_OLD:
1150 case EXT2_IOC_SETVERSION_NEW:
1153 RETURN( obd_iocontrol(cmd, ll_i2obdexp(inode), 0, NULL,
1158 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1160 struct inode *inode = file->f_dentry->d_inode;
1161 struct ll_file_data *fd = file->private_data;
1162 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1163 struct lustre_handle lockh = {0};
1166 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1167 inode->i_generation, inode,
1168 offset + ((origin==2) ? inode->i_size : file->f_pos));
1170 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1171 if (origin == 2) { /* SEEK_END */
1172 int nonblock = 0, rc;
1173 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1175 if (file->f_flags & O_NONBLOCK)
1176 nonblock = LDLM_FL_BLOCK_NOWAIT;
1178 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1183 offset += inode->i_size;
1184 } else if (origin == 1) { /* SEEK_CUR */
1185 offset += file->f_pos;
1189 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1190 if (offset != file->f_pos) {
1191 file->f_pos = offset;
1192 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1194 file->f_version = ++event;
1201 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1205 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1207 struct inode *inode = dentry->d_inode;
1208 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1210 struct ptlrpc_request *req;
1213 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1214 inode->i_generation, inode);
1216 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1218 /* fsync's caller has already called _fdata{sync,write}, we want
1219 * that IO to finish before calling the osc and mdc sync methods */
1220 rc = filemap_fdatawait(inode->i_mapping);
1222 ll_inode2fid(&fid, inode);
1223 err = md_sync(ll_i2sbi(inode)->ll_mdc_exp, &fid, &req);
1227 ptlrpc_req_finished(req);
1230 struct obdo *oa = obdo_alloc();
1233 RETURN(rc ? rc : -ENOMEM);
1235 oa->o_id = lsm->lsm_object_id;
1236 oa->o_gr = lsm->lsm_object_gr;
1237 oa->o_valid = OBD_MD_FLID;
1238 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
1239 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1242 err = obd_sync(ll_i2sbi(inode)->ll_osc_exp, oa, lsm,
1252 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1254 struct inode *inode = file->f_dentry->d_inode;
1255 struct ll_sb_info *sbi = ll_i2sbi(inode);
1256 struct obd_device *obddev;
1257 struct ldlm_res_id res_id =
1258 { .name = {inode->i_ino, inode->i_generation, LDLM_FLOCK} };
1259 struct lustre_handle lockh = {0};
1260 ldlm_policy_data_t flock;
1261 ldlm_mode_t mode = 0;
1266 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1267 inode->i_ino, file_lock);
1269 flock.l_flock.pid = file_lock->fl_pid;
1270 flock.l_flock.start = file_lock->fl_start;
1271 flock.l_flock.end = file_lock->fl_end;
1273 switch (file_lock->fl_type) {
1278 /* An unlock request may or may not have any relation to
1279 * existing locks so we may not be able to pass a lock handle
1280 * via a normal ldlm_lock_cancel() request. The request may even
1281 * unlock a byte range in the middle of an existing lock. In
1282 * order to process an unlock request we need all of the same
1283 * information that is given with a normal read or write record
1284 * lock request. To avoid creating another ldlm unlock (cancel)
1285 * message we'll treat a LCK_NL flock request as an unlock. */
1292 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1301 flags = LDLM_FL_BLOCK_NOWAIT;
1304 flags = LDLM_FL_TEST_LOCK;
1305 /* Save the old mode so that if the mode in the lock changes we
1306 * can decrement the appropriate reader or writer refcount. */
1307 file_lock->fl_type = mode;
1310 CERROR("unknown fcntl lock command: %d\n", cmd);
1314 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1315 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1316 flags, mode, flock.l_flock.start, flock.l_flock.end);
1318 obddev = md_get_real_obd(sbi->ll_mdc_exp, NULL, 0);
1319 rc = ldlm_cli_enqueue(sbi->ll_mdc_exp, NULL, obddev->obd_namespace,
1320 res_id, LDLM_FLOCK, &flock, mode, &flags,
1321 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1322 NULL, 0, NULL, &lockh);
1326 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
1328 struct inode *inode = dentry->d_inode;
1329 struct ll_inode_info *lli;
1330 struct lov_stripe_md *lsm;
1336 CERROR("REPORT THIS LINE TO PETER\n");
1339 ll_inode2fid(&fid, inode);
1340 lli = ll_i2info(inode);
1341 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s,intent=%s\n",
1342 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1344 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1345 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1348 if (!md_valid_attrs(ll_i2mdcexp(inode), &fid)) {
1349 struct ptlrpc_request *req = NULL;
1350 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
1352 unsigned long valid = 0;
1355 if (S_ISREG(inode->i_mode)) {
1356 ealen = obd_size_diskmd(sbi->ll_osc_exp, NULL);
1357 valid |= OBD_MD_FLEASIZE;
1359 ll_inode2fid(&fid, inode);
1360 rc = md_getattr(sbi->ll_mdc_exp, &fid, valid, ealen, &req);
1362 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
1365 rc = ll_prep_inode(sbi->ll_osc_exp, sbi->ll_mdc_exp,
1366 &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,