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 #include <linux/lustre_acl.h>
31 #include <linux/lustre_sec.h>
32 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
33 #include <linux/lustre_compat25.h>
35 #include <linux/obd_lov.h>
36 #include <linux/lustre_audit.h>
37 #include <linux/lustre_net.h>
38 #include "llite_internal.h"
40 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
41 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
42 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
44 int ll_validate_size(struct inode *inode, __u64 *size, __u64 *blocks)
46 ldlm_policy_data_t extent = { .l_extent = { 0, OBD_OBJECT_EOF } };
47 struct obd_export *exp = ll_i2sbi(inode)->ll_dt_exp;
48 struct ll_inode_info *lli = ll_i2info(inode);
49 struct lustre_handle match_lockh = {0};
53 if (lli->lli_smd == NULL)
56 LASSERT(size != NULL && blocks != NULL);
58 flags = LDLM_FL_TEST_LOCK | LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
59 rc = obd_match(exp, lli->lli_smd, LDLM_EXTENT, &extent,
60 LCK_PR | LCK_PW, &flags, inode, &match_lockh);
62 /* we have no all needed locks,
63 * so we don't know actual size */
67 /* we know actual size! */
68 down(&lli->lli_size_sem);
69 *size = lov_merge_size(lli->lli_smd, 0);
70 *blocks = lov_merge_blocks(lli->lli_smd);
71 up(&lli->lli_size_sem);
77 int ll_md_och_close(struct obd_export *md_exp, struct inode *inode,
78 struct obd_client_handle *och, int dirty,
81 struct ptlrpc_request *req = NULL;
82 struct mdc_op_data *op_data;
83 struct obd_device *obd;
88 obd = class_exp2obd(md_exp);
90 CERROR("Invalid MDC connection handle "LPX64"\n",
91 md_exp->exp_handle.h_cookie);
95 /* here we check if this is forced umount. If so this is called on
96 * canceling "open lock" and we do not call md_close() in this case , as
97 * it will not successful, as import is already deactivated. */
98 if (obd->obd_no_recov)
101 /* prepare @op_data for close request */
102 OBD_ALLOC(op_data, sizeof(*op_data));
106 memset(op_data, 0, sizeof(*op_data));
108 valid = (OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
109 OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLEPOCH |
112 ll_inode2mdc_data(op_data, inode, valid);
114 if (0 /* ll_is_inode_dirty(inode) */) {
115 op_data->flags = MDS_BFLAG_UNCOMMITTED_WRITES;
116 op_data->valid |= OBD_MD_FLFLAGS;
120 /* we modified data through this handle */
121 op_data->io_epoch = epoch;
122 op_data->flags |= MDS_BFLAG_DIRTY_EPOCH;
123 op_data->valid |= OBD_MD_FLFLAGS | OBD_MD_FLEPOCH;
124 if (ll_validate_size(inode, &op_data->size, &op_data->blocks))
125 op_data->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
128 rc = md_close(md_exp, op_data, och, &req);
129 OBD_FREE(op_data, sizeof(*op_data));
132 /* we are the last writer, so the MDS has instructed us to get
133 * the file size and any write cookies, then close again. */
135 //ll_queue_done_writing(inode);
138 CERROR("inode %lu mdc close failed: rc = %d\n",
139 (unsigned long)inode->i_ino, rc);
142 ptlrpc_req_finished(req);
145 mdc_clear_open_replay_data(md_exp, och);
146 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
147 OBD_FREE(och, sizeof *och);
151 int ll_md_real_close(struct obd_export *md_exp,
152 struct inode *inode, int flags)
154 struct ll_inode_info *lli = ll_i2info(inode);
155 int freeing = inode->i_state & I_FREEING;
156 struct obd_client_handle **och_p;
157 struct obd_client_handle *och;
158 __u64 *och_usecount, epoch = 0;
159 int rc = 0, dirty = 0;
162 if (flags & FMODE_WRITE) {
163 och_p = &lli->lli_mds_write_och;
164 och_usecount = &lli->lli_open_fd_write_count;
165 } else if (flags & FMODE_EXEC) {
166 och_p = &lli->lli_mds_exec_och;
167 och_usecount = &lli->lli_open_fd_exec_count;
169 och_p = &lli->lli_mds_read_och;
170 och_usecount = &lli->lli_open_fd_read_count;
173 down(&lli->lli_och_sem);
174 if (*och_usecount) { /* There are still users of this handle, so
176 up(&lli->lli_och_sem);
179 if (ll_is_inode_dirty(inode)) {
180 /* the inode still has dirty pages, let's close later */
181 CDEBUG(D_INODE, "inode %lu/%u still has dirty pages\n",
182 inode->i_ino, inode->i_generation);
183 LASSERT(freeing == 0);
184 ll_queue_done_writing(inode);
185 up(&lli->lli_och_sem);
189 if (LLI_DIRTY_HANDLE(inode) && (flags & FMODE_WRITE)) {
190 clear_bit(LLI_F_DIRTY_HANDLE, &lli->lli_flags);
192 } else if (0 && !(flags & FMODE_SYNC) && !freeing) {
193 /* in order to speed up creation rate we pass
194 * closing to dedicated thread so we don't need
195 * to wait for close reply here -bzzz */
196 ll_queue_done_writing(inode);
197 up(&lli->lli_och_sem);
203 if (flags & FMODE_WRITE) {
204 epoch = lli->lli_io_epoch;
205 lli->lli_io_epoch = 0;
208 up(&lli->lli_och_sem);
211 * there might be a race and somebody have freed this och
212 * already. Another way to have this twice called is if file closing
213 * will fail due to netwok problems and on umount lock will be canceled
214 * and this will be called from block_ast callack.
216 if (och && och->och_fh.cookie != DEAD_HANDLE_MAGIC)
217 rc = ll_md_och_close(md_exp, inode, och, dirty, epoch);
222 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
225 struct ll_file_data *fd = file->private_data;
226 struct ll_inode_info *lli = ll_i2info(inode);
230 /* clear group lock, if present */
231 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
232 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
233 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
234 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
238 /* Let's see if we have good enough OPEN lock on the file and if
239 we can skip talking to MDS */
240 if (file->f_dentry->d_inode) {
242 struct obd_device *obddev;
243 struct lustre_handle lockh;
244 int flags = LDLM_FL_BLOCK_GRANTED;
245 struct ldlm_res_id file_res_id = {.name = {id_fid(&lli->lli_id),
246 id_group(&lli->lli_id)}};
247 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
249 down(&lli->lli_och_sem);
250 if (fd->fd_omode & FMODE_WRITE) {
252 LASSERT(lli->lli_open_fd_write_count);
253 lli->lli_open_fd_write_count--;
254 } else if (fd->fd_omode & FMODE_EXEC) {
256 LASSERT(lli->lli_open_fd_exec_count);
257 lli->lli_open_fd_exec_count--;
260 LASSERT(lli->lli_open_fd_read_count);
261 lli->lli_open_fd_read_count--;
263 up(&lli->lli_och_sem);
265 obddev = md_get_real_obd(md_exp, &lli->lli_id);
266 if (!ldlm_lock_match(obddev->obd_namespace, flags, &file_res_id,
267 LDLM_IBITS, &policy, lockmode, &lockh))
269 rc = ll_md_real_close(md_exp, file->f_dentry->d_inode,
272 ldlm_lock_decref(&lockh, lockmode);
276 file->private_data = NULL;
277 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof(*fd));
281 /* While this returns an error code, fput() the caller does not, so we need
282 * to make every effort to clean up all of our state here. Also, applications
283 * rarely check close errors and even if an error is returned they will not
284 * re-try the close call.
286 int ll_file_release(struct inode *inode, struct file *file)
288 struct ll_file_data *fd;
289 struct ll_sb_info *sbi = ll_i2sbi(inode);
293 CDEBUG(D_VFSTRACE, "VFS Op:inode="DLID4"(%p)\n",
294 OLID4(&ll_i2info(inode)->lli_id), inode);
296 /* don't do anything for / */
297 if (inode->i_sb->s_root == file->f_dentry)
300 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
301 fd = (struct ll_file_data *)file->private_data;
304 rc = ll_md_close(sbi->ll_md_exp, inode, file);
308 static int ll_intent_file_open(struct file *file, void *lmm,
309 int lmmsize, struct lookup_intent *itp)
311 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
312 const char *name = (char *)file->f_dentry->d_name.name;
313 struct dentry *parent = file->f_dentry->d_parent;
314 const int len = file->f_dentry->d_name.len;
315 struct lustre_handle lockh;
316 struct mdc_op_data *op_data;
322 OBD_ALLOC(op_data, sizeof(*op_data));
326 ll_prepare_mdc_data(op_data, parent->d_inode, NULL,
329 rc = md_enqueue(sbi->ll_md_exp, LDLM_IBITS, itp, LCK_PR, op_data,
330 &lockh, lmm, lmmsize, ldlm_completion_ast,
331 ll_mdc_blocking_ast, NULL);
332 OBD_FREE(op_data, sizeof(*op_data));
334 if (LUSTRE_IT(itp)->it_lock_mode)
335 memcpy(&LUSTRE_IT(itp)->it_lock_handle,
336 &lockh, sizeof(lockh));
339 CERROR("lock enqueue: err: %d\n", rc);
344 int ll_och_fill(struct inode *inode, struct lookup_intent *it,
345 struct obd_client_handle *och)
347 struct ptlrpc_request *req = LUSTRE_IT(it)->it_data;
348 struct ll_inode_info *lli = ll_i2info(inode);
349 struct mds_body *body;
354 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
355 LASSERT (body != NULL); /* reply already checked out */
356 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
358 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
359 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
360 if (it->it_flags & FMODE_WRITE) {
361 if (lli->lli_io_epoch && lli->lli_io_epoch != body->io_epoch)
362 CDEBUG(D_ERROR, "new epoch?! "LPD64" != "LPD64"\n",
363 lli->lli_io_epoch, body->io_epoch);
364 lli->lli_io_epoch = body->io_epoch;
366 mdc_set_open_replay_data(ll_i2mdexp(inode), och,
367 LUSTRE_IT(it)->it_data);
372 int ll_local_open(struct file *file, struct lookup_intent *it,
373 struct obd_client_handle *och)
375 struct ll_file_data *fd;
376 struct inode *inode = file->f_dentry->d_inode;
381 rc = ll_och_fill(inode, it, och);
386 LASSERTF(file->private_data == NULL, "file %.*s/%.*s ino %lu/%u (%o)\n",
387 file->f_dentry->d_name.len, file->f_dentry->d_name.name,
388 file->f_dentry->d_parent->d_name.len,
389 file->f_dentry->d_parent->d_name.name,
390 file->f_dentry->d_inode->i_ino,
391 file->f_dentry->d_inode->i_generation,
392 file->f_dentry->d_inode->i_mode);
394 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
396 /* We can't handle this well without reorganizing ll_file_open and
397 * ll_md_close(), so don't even try right now. */
400 file->private_data = fd;
401 ll_readahead_init(inode, &fd->fd_ras);
402 fd->fd_omode = it->it_flags;
407 /* Open a file, and (for the very first open) create objects on the OSTs at
408 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
409 * creation or open until ll_lov_setstripe() ioctl is called. We grab
410 * lli_open_sem to ensure no other process will create objects, send the
411 * stripe MD to the MDS, or try to destroy the objects if that fails.
413 * If we already have the stripe MD locally then we don't request it in
414 * mdc_open(), by passing a lmm_size = 0.
416 * It is up to the application to ensure no other processes open this file
417 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
418 * used. We might be able to avoid races of that sort by getting lli_open_sem
419 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
420 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
422 int ll_file_open(struct inode *inode, struct file *file)
424 struct ll_inode_info *lli = ll_i2info(inode);
425 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
426 .it_flags = file->f_flags };
427 struct lov_stripe_md *lsm;
428 struct ptlrpc_request *req;
430 struct obd_client_handle **och_p = NULL;
431 __u64 *och_usecount = NULL;
434 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n",
435 inode->i_ino, inode->i_generation, inode, file->f_flags);
437 /* don't do anything for / */
438 if (inode->i_sb->s_root == file->f_dentry)
441 if ((file->f_flags+1) & O_ACCMODE)
443 if (file->f_flags & O_TRUNC)
448 /* sometimes LUSTRE_IT(it) may not be allocated like opening file by
449 * dentry_open() from GNS stuff. */
450 if (!it || !LUSTRE_IT(it)) {
452 rc = ll_intent_alloc(it);
457 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
459 /* mdc_intent_lock() didn't get a request ref if there was an open
460 * error, so don't do cleanup on the * request here (bug 3430) */
461 if (LUSTRE_IT(it)->it_disposition) {
462 rc = it_open_error(DISP_OPEN_OPEN, it);
467 /* Let's see if we have file open on MDS already. */
468 if (it->it_flags & FMODE_WRITE) {
469 och_p = &lli->lli_mds_write_och;
470 och_usecount = &lli->lli_open_fd_write_count;
471 } else if (it->it_flags & FMODE_EXEC) {
472 och_p = &lli->lli_mds_exec_och;
473 och_usecount = &lli->lli_open_fd_exec_count;
475 och_p = &lli->lli_mds_read_och;
476 och_usecount = &lli->lli_open_fd_read_count;
479 rc = ll_crypto_decrypt_key(inode, it);
483 down(&lli->lli_och_sem);
484 if (*och_p) { /* Open handle is present */
485 if (it_disposition(it, DISP_LOOKUP_POS) && /* Positive lookup */
486 it_disposition(it, DISP_OPEN_OPEN)) { /* & OPEN happened */
487 struct obd_client_handle *och;
488 /* Well, there's extra open request that we do not need,
489 let's close it somehow*/
490 OBD_ALLOC(och, sizeof (struct obd_client_handle));
492 up(&lli->lli_och_sem);
496 rc = ll_och_fill(inode, it, och);
498 up(&lli->lli_och_sem);
502 /* ll_md_och_close() will free och */
503 ll_md_och_close(ll_i2mdexp(inode), inode, och, 0, 0);
507 rc = ll_local_open(file, it, NULL);
511 rc = ll_set_capa(inode, it, *och_p);
513 up(&lli->lli_och_sem);
517 LASSERT(*och_usecount == 0);
518 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
520 GOTO(out, rc = -ENOMEM);
523 if (!it || !LUSTRE_IT(it) || !LUSTRE_IT(it)->it_disposition) {
525 * we are going to replace intent here, and that may
526 * possibly change access mode (FMODE_EXEC can only be
527 * set in intent), but I hope it never happens (I was
528 * not able to trigger it yet at least) -- green
531 /* FIXME: FMODE_EXEC is not covered by O_ACCMODE! */
532 LASSERT(!(it->it_flags & FMODE_EXEC));
533 LASSERTF((it->it_flags & O_ACCMODE) ==
534 (oit.it_flags & O_ACCMODE), "Changing intent "
535 "flags %x to incompatible %x\n", it->it_flags,
538 rc = ll_intent_file_open(file, NULL, 0, it);
541 rc = it_open_error(DISP_OPEN_OPEN, it);
543 GOTO(out_och_free, rc);
545 mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle,
546 file->f_dentry->d_inode);
548 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
549 rc = ll_local_open(file, it, *och_p);
550 LASSERTF(rc == 0, "rc = %d\n", rc);
552 rc = ll_set_capa(inode, it, *och_p);
554 up(&lli->lli_och_sem);
558 up(&lli->lli_och_sem);
561 * must do this outside lli_och_sem lock to prevent deadlock where
562 * different kind of OPEN lock for this same inode gets cancelled by
565 if (!S_ISREG(inode->i_mode))
570 if (file->f_flags & O_LOV_DELAY_CREATE ||
571 !(file->f_mode & FMODE_WRITE)) {
572 CDEBUG(D_INODE, "object creation was delayed\n");
576 file->f_flags &= ~O_LOV_DELAY_CREATE;
579 /* audit stuff if there was no RPC */
580 if (LUSTRE_IT(it)->it_data == 0)
581 ll_audit_log(inode, AUDIT_OPEN, rc);
583 req = LUSTRE_IT(it)->it_data;
584 ll_intent_drop_lock(it);
585 ll_intent_release(it);
588 ptlrpc_req_finished(req);
590 ll_open_complete(inode);
594 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
595 *och_p = NULL; /* OBD_FREE writes some magic there */
598 up(&lli->lli_och_sem);
604 /* Fills the obdo with the attributes for the inode defined by lsm */
605 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
608 struct ptlrpc_request_set *set;
612 LASSERT(lsm != NULL);
614 memset(oa, 0, sizeof *oa);
615 oa->o_id = lsm->lsm_object_id;
616 oa->o_gr = lsm->lsm_object_gr;
617 oa->o_mode = S_IFREG;
618 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
619 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
620 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
622 set = ptlrpc_prep_set();
626 rc = obd_getattr_async(exp, oa, lsm, set);
628 rc = ptlrpc_set_wait(set);
629 ptlrpc_set_destroy(set);
634 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
635 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
639 static inline void ll_remove_suid(struct inode *inode)
643 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
644 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
646 /* was any of the uid bits set? */
647 mode &= inode->i_mode;
648 if (mode && !capable(CAP_FSETID)) {
649 inode->i_mode &= ~mode;
650 // XXX careful here - we cannot change the size
654 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
656 struct ll_inode_info *lli = ll_i2info(inode);
657 struct lov_stripe_md *lsm = lli->lli_smd;
658 struct obd_export *exp = ll_i2dtexp(inode);
661 struct ldlm_lock *lock;
662 struct lov_stripe_md *lsm;
663 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
664 __u32 stripe, vallen = sizeof(stripe);
668 if (lsm->lsm_stripe_count == 1)
669 GOTO(check, stripe = 0);
671 /* get our offset in the lov */
672 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
674 CERROR("obd_get_info: rc = %d\n", rc);
677 LASSERT(stripe < lsm->lsm_stripe_count);
680 if (lsm->lsm_oinfo[stripe].loi_id != lock->l_resource->lr_name.name[0]||
681 lsm->lsm_oinfo[stripe].loi_gr != lock->l_resource->lr_name.name[2]){
682 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64
683 " inode=%lu/%u (%p)\n",
684 lsm->lsm_oinfo[stripe].loi_id,
685 lsm->lsm_oinfo[stripe].loi_gr,
686 inode->i_ino, inode->i_generation, inode);
687 return -ELDLM_NO_LOCK_DATA;
693 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
694 * we get a lock cancellation for each stripe, so we have to map the obd's
695 * region back onto the stripes in the file that it held.
697 * No one can dirty the extent until we've finished our work and they can
698 * enqueue another lock. The DLM protects us from ll_file_read/write here,
699 * but other kernel actors could have pages locked.
701 * Called with the DLM lock held. */
702 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
703 struct ldlm_lock *lock, __u32 stripe)
705 ldlm_policy_data_t tmpex;
706 unsigned long start, end, count, skip, i, j;
708 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
709 struct lustre_handle lockh;
712 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
713 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
714 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
717 /* our locks are page granular thanks to osc_enqueue, we invalidate the
719 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
720 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
724 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
725 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
726 if (lsm->lsm_stripe_count > 1) {
727 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
728 skip = (lsm->lsm_stripe_count - 1) * count;
729 start += start/count * skip + stripe * count;
731 end += end/count * skip + stripe * count;
733 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
736 i = inode->i_size ? (inode->i_size - 1) >> PAGE_CACHE_SHIFT : 0;
740 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
741 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
742 count, skip, end, discard ? " (DISCARDING)" : "");
744 /* walk through the vmas on the inode and tear down mmaped pages that
745 * intersect with the lock. this stops immediately if there are no
746 * mmap()ed regions of the file. This is not efficient at all and
747 * should be short lived. We'll associate mmap()ed pages with the lock
748 * and will be able to find them directly */
750 for (i = start; i <= end; i += (j + skip)) {
751 j = min(count - (i % count), end - i + 1);
753 LASSERT(inode->i_mapping);
754 if (ll_teardown_mmaps(inode->i_mapping,
755 (__u64)i << PAGE_CACHE_SHIFT,
756 ((__u64)(i+j) << PAGE_CACHE_SHIFT) - 1) )
760 /* this is the simplistic implementation of page eviction at
761 * cancelation. It is careful to get races with other page
762 * lockers handled correctly. fixes from bug 20 will make it
763 * more efficient by associating locks with pages and with
764 * batching writeback under the lock explicitly. */
765 for (i = start, j = start % count; i <= end;
766 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
768 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
774 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
775 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
776 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
779 if (!mapping_has_pages(inode->i_mapping)) {
780 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
786 page = find_get_page(inode->i_mapping, i);
789 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
790 i, tmpex.l_extent.start);
793 /* page->mapping to check with racing against teardown */
794 if (!discard && clear_page_dirty_for_io(page)) {
795 rc = ll_call_writepage(inode, page);
797 CERROR("writepage of page %p failed: %d\n",
799 /* either waiting for io to complete or reacquiring
800 * the lock that the failed writepage released */
804 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
805 /* check to see if another DLM lock covers this page */
806 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
807 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
809 &lock->l_resource->lr_name, LDLM_EXTENT,
810 &tmpex, LCK_PR | LCK_PW, &lockh);
811 if (rc2 == 0 && page->mapping != NULL) {
812 // checking again to account for writeback's lock_page()
813 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
814 ll_ra_accounting(page, inode->i_mapping);
815 ll_truncate_complete_page(page);
818 page_cache_release(page);
820 LASSERTF(tmpex.l_extent.start <=
821 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
822 lock->l_policy_data.l_extent.end + 1),
823 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
824 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
829 static int ll_extent_lock_callback(struct ldlm_lock *lock,
830 struct ldlm_lock_desc *new, void *data,
833 struct lustre_handle lockh = { 0 };
837 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
838 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
843 case LDLM_CB_BLOCKING:
844 ldlm_lock2handle(lock, &lockh);
845 rc = ldlm_cli_cancel(&lockh);
847 CERROR("ldlm_cli_cancel failed: %d\n", rc);
849 case LDLM_CB_CANCELING: {
851 struct ll_inode_info *lli;
852 struct lov_stripe_md *lsm;
856 /* This lock wasn't granted, don't try to evict pages */
857 if (lock->l_req_mode != lock->l_granted_mode)
860 inode = ll_inode_from_lock(lock);
863 lli = ll_i2info(inode);
866 if (lli->lli_smd == NULL)
870 stripe = ll_lock_to_stripe_offset(inode, lock);
872 CERROR("ll_lock_to_stripe_offset failed: %d\n", stripe);
876 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
878 down(&lli->lli_size_sem);
879 lock_res_and_lock(lock);
880 kms = ldlm_extent_shift_kms(lock,
881 lsm->lsm_oinfo[stripe].loi_kms);
883 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
884 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
885 lsm->lsm_oinfo[stripe].loi_kms, kms);
886 lsm->lsm_oinfo[stripe].loi_kms = kms;
887 unlock_res_and_lock(lock);
888 up(&lli->lli_size_sem);
889 //ll_try_done_writing(inode);
902 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
904 /* XXX ALLOCATE - 160 bytes */
905 struct inode *inode = ll_inode_from_lock(lock);
906 struct ll_inode_info *lli = ll_i2info(inode);
907 struct lustre_handle lockh = { 0 };
912 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
913 LDLM_FL_BLOCK_CONV)) {
914 LBUG(); /* not expecting any blocked async locks yet */
915 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
917 ldlm_lock_dump(D_OTHER, lock, 0);
918 ldlm_reprocess_all(lock->l_resource);
922 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
924 stripe = ll_lock_to_stripe_offset(inode, lock);
926 CERROR("ll_lock_to_stripe_offset failed: %d\n", stripe);
930 if (lock->l_lvb_len) {
931 struct lov_stripe_md *lsm = lli->lli_smd;
933 lvb = lock->l_lvb_data;
934 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
937 lock_res_and_lock(lock);
938 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
939 kms = ldlm_extent_shift_kms(NULL, kms);
940 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
941 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
942 lsm->lsm_oinfo[stripe].loi_kms, kms);
943 lsm->lsm_oinfo[stripe].loi_kms = kms;
944 unlock_res_and_lock(lock);
950 wake_up(&lock->l_waitq);
952 ldlm_lock2handle(lock, &lockh);
953 ldlm_lock_decref(&lockh, LCK_PR);
958 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
960 struct ptlrpc_request *req = reqp;
961 struct inode *inode = ll_inode_from_lock(lock);
962 struct ll_inode_info *lli;
964 struct lov_stripe_md *lsm;
965 int rc, size = sizeof(*lvb), stripe;
969 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
970 lli = ll_i2info(inode);
972 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
976 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
978 /* First, find out which stripe index this lock corresponds to. */
979 stripe = ll_lock_to_stripe_offset(inode, lock);
981 CERROR("ll_lock_to_stripe_offset failed: %d\n", stripe);
982 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
985 rc = lustre_pack_reply(req, 1, &size, NULL);
987 CERROR("lustre_pack_reply: %d\n", rc);
991 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
992 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
993 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
994 lvb->lvb_atime = LTIME_S(inode->i_atime);
995 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
997 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
998 inode->i_size, stripe, lvb->lvb_size);
1004 /* These errors are normal races, so we don't want to fill the console
1005 * with messages by calling ptlrpc_error() */
1006 if (rc == -ELDLM_NO_LOCK_DATA)
1007 lustre_pack_reply(req, 0, NULL, NULL);
1009 req->rq_status = rc;
1013 /* NB: lov_merge_size will prefer locally cached writes if they extend the
1014 * file (because it prefers KMS over RSS when larger) */
1015 int ll_glimpse_size(struct inode *inode)
1017 struct ll_inode_info *lli = ll_i2info(inode);
1018 struct ll_sb_info *sbi = ll_i2sbi(inode);
1019 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1020 struct lustre_handle lockh = { 0 };
1021 int rc, flags = LDLM_FL_HAS_INTENT;
1024 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1026 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
1027 LCK_PR, &flags, ll_extent_lock_callback,
1028 ldlm_completion_ast, ll_glimpse_callback, inode,
1029 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1034 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1035 RETURN(rc > 0 ? -EIO : rc);
1038 down(&lli->lli_size_sem);
1039 inode->i_size = lov_merge_size(lli->lli_smd, 0);
1040 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
1041 up(&lli->lli_size_sem);
1043 LTIME_S(inode->i_mtime) = lov_merge_mtime(lli->lli_smd,
1044 LTIME_S(inode->i_mtime));
1046 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
1047 (__u64)inode->i_size, (__u64)inode->i_blocks);
1049 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
1053 void ll_stime_record(struct ll_sb_info *sbi, struct timeval *start,
1054 struct obd_service_time *stime)
1056 struct timeval stop;
1057 do_gettimeofday(&stop);
1059 spin_lock(&sbi->ll_lock);
1060 lprocfs_stime_record(stime, &stop, start);
1061 spin_unlock(&sbi->ll_lock);
1064 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
1065 struct lov_stripe_md *lsm, int mode,
1066 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
1067 int ast_flags, struct obd_service_time *stime)
1069 struct ll_inode_info *lli = ll_i2info(inode);
1070 struct ll_sb_info *sbi = ll_i2sbi(inode);
1071 struct timeval start;
1075 LASSERT(lockh->cookie == 0);
1077 /* XXX phil: can we do this? won't it screw the file size up? */
1078 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1079 (sbi->ll_flags & LL_SBI_NOLCK))
1082 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
1083 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
1085 do_gettimeofday(&start);
1086 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
1087 &ast_flags, ll_extent_lock_callback,
1088 ldlm_completion_ast, ll_glimpse_callback, inode,
1089 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
1093 ll_stime_record(sbi, &start, stime);
1095 if (policy->l_extent.start == 0 &&
1096 policy->l_extent.end == OBD_OBJECT_EOF) {
1097 /* vmtruncate()->ll_truncate() first sets the i_size and then
1098 * the kms under both a DLM lock and the i_sem. If we don't
1099 * get the i_sem here we can match the DLM lock and reset
1100 * i_size from the kms before the truncating path has updated
1101 * the kms. generic_file_write can then trust the stale i_size
1102 * when doing appending writes and effectively cancel the
1103 * result of the truncate. Getting the i_sem after the enqueue
1104 * maintains the DLM -> i_sem acquiry order. */
1105 down(&lli->lli_size_sem);
1106 inode->i_size = lov_merge_size(lsm, 1);
1107 up(&lli->lli_size_sem);
1111 LTIME_S(inode->i_mtime) =
1112 lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1118 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1119 struct lov_stripe_md *lsm, int mode,
1120 struct lustre_handle *lockh)
1122 struct ll_sb_info *sbi = ll_i2sbi(inode);
1126 /* XXX phil: can we do this? won't it screw the file size up? */
1127 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1128 (sbi->ll_flags & LL_SBI_NOLCK))
1131 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1136 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1139 struct inode *inode = file->f_dentry->d_inode;
1140 struct ll_inode_info *lli = ll_i2info(inode);
1141 struct lov_stripe_md *lsm = lli->lli_smd;
1142 struct ll_lock_tree tree;
1143 struct ll_lock_tree_node *node;
1148 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1149 inode->i_ino, inode->i_generation, inode, count, *ppos);
1151 /* "If nbyte is 0, read() will return 0 and have no other results."
1152 * -- Single Unix Spec */
1156 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
1162 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1165 tree.lt_fd = file->private_data;
1167 rc = ll_tree_lock(&tree, node, inode, buf, count,
1168 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1172 down(&lli->lli_size_sem);
1173 kms = lov_merge_size(lsm, 1);
1174 if (*ppos + count - 1 > kms) {
1175 /* A glimpse is necessary to determine whether we return a short
1176 * read or some zeroes at the end of the buffer */
1177 up(&lli->lli_size_sem);
1178 retval = ll_glimpse_size(inode);
1182 inode->i_size = kms;
1183 up(&lli->lli_size_sem);
1186 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1187 inode->i_ino, count, *ppos, inode->i_size);
1189 /* turn off the kernel's read-ahead */
1190 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1193 file->f_ra.ra_pages = 0;
1195 retval = generic_file_read(file, buf, count, ppos);
1196 ll_audit_log(inode, AUDIT_READ, retval);
1199 ll_tree_unlock(&tree, inode);
1204 * Write to a file (through the page cache).
1206 static ssize_t ll_file_write(struct file *file, const char *buf,
1207 size_t count, loff_t *ppos)
1209 struct inode *inode = file->f_dentry->d_inode;
1210 loff_t maxbytes = ll_file_maxbytes(inode);
1211 struct ll_lock_tree tree;
1212 struct ll_lock_tree_node *node;
1217 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1218 inode->i_ino, inode->i_generation, inode, count, *ppos);
1220 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1222 /* POSIX, but surprised the VFS doesn't check this already */
1226 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1227 * called on the file, don't fail the below assertion (bug 2388). */
1228 if (file->f_flags & O_LOV_DELAY_CREATE &&
1229 ll_i2info(inode)->lli_smd == NULL)
1232 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1234 if (file->f_flags & O_APPEND)
1235 node = ll_node_from_inode(inode, 0, OBD_OBJECT_EOF, LCK_PW);
1237 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1241 RETURN(PTR_ERR(node));
1243 tree.lt_fd = file->private_data;
1245 rc = ll_tree_lock(&tree, node, inode, buf, count,
1246 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1250 /* this is ok, g_f_w will overwrite this under i_sem if it races
1251 * with a local truncate, it just makes our maxbyte checking easier */
1252 if (file->f_flags & O_APPEND)
1253 *ppos = inode->i_size;
1255 if (*ppos >= maxbytes) {
1256 if (count || *ppos > maxbytes) {
1257 send_sig(SIGXFSZ, current, 0);
1258 GOTO(out, retval = -EFBIG);
1261 if (*ppos + count > maxbytes)
1262 count = maxbytes - *ppos;
1264 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1265 inode->i_ino, count, *ppos);
1267 /* mark open handle dirty */
1268 set_bit(LLI_F_DIRTY_HANDLE, &(ll_i2info(inode)->lli_flags));
1270 /* generic_file_write handles O_APPEND after getting i_sem */
1271 retval = generic_file_write(file, buf, count, ppos);
1275 /* ll_audit_log(inode, AUDIT_WRITE, retval); */
1277 ll_tree_unlock(&tree, inode);
1278 /* serialize with mmap/munmap/mremap */
1279 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
1280 retval > 0 ? retval : 0);
1284 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
1285 int flags, struct lov_user_md *lum,
1288 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
1289 struct ll_inode_info *lli = ll_i2info(inode);
1290 struct ptlrpc_request *req = NULL;
1291 struct obd_client_handle *och;
1292 struct lov_stripe_md *lsm;
1293 struct lustre_md md;
1299 if ((file->f_flags+1) & O_ACCMODE)
1301 if (file->f_flags & O_TRUNC)
1304 down(&lli->lli_open_sem);
1307 up(&lli->lli_open_sem);
1308 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
1313 f = get_empty_filp();
1317 f->f_dentry = file->f_dentry;
1318 f->f_vfsmnt = file->f_vfsmnt;
1321 rc = ll_intent_alloc(&oit);
1325 rc = ll_intent_file_open(f, lum, lum_size, &oit);
1328 if (it_disposition(&oit, DISP_LOOKUP_NEG))
1331 req = LUSTRE_IT(&oit)->it_data;
1332 rc = LUSTRE_IT(&oit)->it_status;
1337 rc = mdc_req2lustre_md(ll_i2mdexp(inode), req, 1,
1338 ll_i2dtexp(inode), &md);
1341 ll_update_inode(f->f_dentry->d_inode, &md);
1343 OBD_ALLOC(och, sizeof(struct obd_client_handle));
1345 GOTO(out, rc = -ENOMEM);
1347 /* actually ll_local_open() cannot fail! */
1348 rc = ll_local_open(f, &oit, och);
1352 if (LUSTRE_IT(&oit)->it_lock_mode) {
1353 ldlm_lock_decref_and_cancel((struct lustre_handle *)
1354 &LUSTRE_IT(&oit)->it_lock_handle,
1355 LUSTRE_IT(&oit)->it_lock_mode);
1356 LUSTRE_IT(&oit)->it_lock_mode = 0;
1359 ll_intent_release(&oit);
1361 /* ll_file_release() will decrease the count, but won't free anything
1362 * because we have at least one more reference coming from actual
1364 down(&lli->lli_och_sem);
1365 lli->lli_open_fd_write_count++;
1366 up(&lli->lli_och_sem);
1367 rc = ll_file_release(f->f_dentry->d_inode, f);
1369 /* Now also destroy our supplemental och */
1370 ll_md_och_close(ll_i2mdexp(inode), f->f_dentry->d_inode, och, 0, 0);
1373 ll_intent_release(&oit);
1376 up(&lli->lli_open_sem);
1378 ptlrpc_req_finished(req);
1382 static int ll_lov_setea(struct inode *inode, struct file *file,
1385 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
1386 int lum_size = sizeof(struct lov_user_md) +
1387 sizeof(struct lov_user_ost_data);
1388 struct lov_user_md *lump;
1392 if (!capable(CAP_SYS_ADMIN))
1395 OBD_ALLOC(lump, lum_size);
1399 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
1401 GOTO(out_free_lump, rc = -EFAULT);
1403 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1406 OBD_FREE(lump, lum_size);
1410 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1413 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1414 int rc, flags = FMODE_WRITE;
1417 /* Bug 1152: copy properly when this is no longer true */
1418 LASSERT(sizeof(lum) == sizeof(*lump));
1419 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1420 rc = copy_from_user(&lum, lump, sizeof(lum));
1424 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1426 put_user(0, &lump->lmm_stripe_count);
1427 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
1428 0, ll_i2info(inode)->lli_smd, lump);
1433 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1435 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1440 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
1444 static int ll_get_grouplock(struct inode *inode, struct file *file,
1447 struct ll_file_data *fd = file->private_data;
1448 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1449 .end = OBD_OBJECT_EOF}};
1450 struct lustre_handle lockh = { 0 };
1451 struct ll_inode_info *lli = ll_i2info(inode);
1452 struct lov_stripe_md *lsm = lli->lli_smd;
1456 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1460 policy.l_extent.gid = arg;
1461 if (file->f_flags & O_NONBLOCK)
1462 flags = LDLM_FL_BLOCK_NOWAIT;
1464 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags,
1465 &ll_i2sbi(inode)->ll_grouplock_stime);
1469 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1471 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1476 static int ll_put_grouplock(struct inode *inode, struct file *file,
1479 struct ll_file_data *fd = file->private_data;
1480 struct ll_inode_info *lli = ll_i2info(inode);
1481 struct lov_stripe_md *lsm = lli->lli_smd;
1485 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1486 /* Ugh, it's already unlocked. */
1490 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1493 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1495 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1500 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1505 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1508 struct ll_file_data *fd = file->private_data;
1509 struct ll_sb_info *sbi = ll_i2sbi(inode);
1513 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1514 inode->i_generation, inode, cmd);
1516 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1519 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1521 case LL_IOC_GETFLAGS:
1522 /* Get the current value of the file flags */
1523 return put_user(fd->fd_flags, (int *)arg);
1524 case LL_IOC_SETFLAGS:
1525 case LL_IOC_CLRFLAGS:
1526 /* Set or clear specific file flags */
1527 /* XXX This probably needs checks to ensure the flags are
1528 * not abused, and to handle any flag side effects.
1530 if (get_user(flags, (int *) arg))
1533 if (cmd == LL_IOC_SETFLAGS)
1534 fd->fd_flags |= flags;
1536 fd->fd_flags &= ~flags;
1538 case LL_IOC_LOV_SETSTRIPE:
1539 RETURN(ll_lov_setstripe(inode, file, arg));
1540 case LL_IOC_LOV_SETEA:
1541 RETURN(ll_lov_setea(inode, file, arg));
1542 case IOC_MDC_SHOWFID: {
1543 struct lustre_id *idp = (struct lustre_id *)arg;
1544 struct lustre_id id;
1548 filename = getname((const char *)arg);
1549 if (IS_ERR(filename))
1550 RETURN(PTR_ERR(filename));
1552 ll_inode2id(&id, inode);
1554 rc = ll_get_fid(sbi->ll_md_exp, &id, filename, &id);
1556 GOTO(out_filename, rc);
1558 rc = copy_to_user(idp, &id, sizeof(*idp));
1560 GOTO(out_filename, rc = -EFAULT);
1567 case LL_IOC_KEY_TYPE: {
1568 struct obd_ioctl_data *data;
1571 int typelen, rc, len = 0;
1573 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1578 type = data->ioc_inlbuf1;
1579 typelen = data->ioc_inllen1;
1582 CDEBUG(D_INFO, "LL_IOC_KEY_TYPE missing filename\n");
1583 GOTO(out, rc = -EINVAL);
1585 ll_set_sb_gksinfo(inode->i_sb, type);
1588 obd_ioctl_freedata(buf, len);
1592 case LL_IOC_LOV_GETSTRIPE:
1593 RETURN(ll_lov_getstripe(inode, arg));
1594 case EXT3_IOC_GETFLAGS:
1595 case EXT3_IOC_SETFLAGS:
1596 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1597 case LL_IOC_GROUP_LOCK:
1598 RETURN(ll_get_grouplock(inode, file, arg));
1599 case LL_IOC_GROUP_UNLOCK:
1600 RETURN(ll_put_grouplock(inode, file, arg));
1601 case EXT3_IOC_GETVERSION_OLD:
1602 case EXT3_IOC_GETVERSION:
1603 return put_user(inode->i_generation, (int *) arg);
1604 /* We need to special case any other ioctls we want to handle,
1605 * to send them to the MDS/OST as appropriate and to properly
1606 * network encode the arg field.
1607 case EXT2_IOC_GETVERSION_OLD:
1608 case EXT2_IOC_GETVERSION_NEW:
1609 case EXT2_IOC_SETVERSION_OLD:
1610 case EXT2_IOC_SETVERSION_NEW:
1611 case EXT3_IOC_SETVERSION_OLD:
1612 case EXT3_IOC_SETVERSION:
1614 case LL_IOC_FLUSH_CRED:
1615 RETURN(ll_flush_cred(inode));
1617 RETURN(ll_set_audit(inode, arg));
1619 RETURN( obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
1624 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1626 struct inode *inode = file->f_dentry->d_inode;
1627 struct ll_file_data *fd = file->private_data;
1628 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1629 struct lustre_handle lockh = {0};
1632 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1633 inode->i_generation, inode,
1634 offset + ((origin==2) ? inode->i_size : file->f_pos));
1636 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1637 if (origin == 2) { /* SEEK_END */
1638 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1639 struct ll_inode_info *lli = ll_i2info(inode);
1640 int nonblock = 0, rc;
1642 if (file->f_flags & O_NONBLOCK)
1643 nonblock = LDLM_FL_BLOCK_NOWAIT;
1645 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1646 nonblock, &ll_i2sbi(inode)->ll_seek_stime);
1650 down(&lli->lli_size_sem);
1651 offset += inode->i_size;
1652 up(&lli->lli_size_sem);
1653 } else if (origin == 1) { /* SEEK_CUR */
1654 offset += file->f_pos;
1658 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1659 if (offset != file->f_pos) {
1660 file->f_pos = offset;
1661 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1663 file->f_version = ++event;
1670 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1674 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1676 struct inode *inode = dentry->d_inode;
1677 struct ll_inode_info *lli = ll_i2info(inode);
1678 struct address_space *mapping = inode->i_mapping;
1679 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1680 struct lustre_id id;
1681 struct ptlrpc_request *req;
1684 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1685 inode->i_generation, inode);
1687 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1689 /* fsync's caller has already called _fdata{sync,write}, we want
1690 * that IO to finish before calling the osc and mdc sync methods */
1691 rc = filemap_fdatawait(inode->i_mapping);
1693 /* 2.6 implements filemap_fdatawait() using PG_writeback which we
1694 * don't support now. so, wait until all submited llaps are gone */
1695 wait_event(lli->lli_dirty_wait, !ll_is_inode_dirty(mapping->host));
1697 ll_inode2id(&id, inode);
1698 err = md_sync(ll_i2sbi(inode)->ll_md_exp, &id, &req);
1702 ptlrpc_req_finished(req);
1705 struct obdo *oa = obdo_alloc();
1708 RETURN(rc ? rc : -ENOMEM);
1710 oa->o_id = lsm->lsm_object_id;
1711 oa->o_gr = lsm->lsm_object_gr;
1712 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1714 obdo_from_inode(oa, inode, (OBD_MD_FLTYPE | OBD_MD_FLATIME |
1715 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1718 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
1728 int ll_file_flock(struct file *file, int cmd, struct file_lock *fl)
1730 struct inode *inode = file->f_dentry->d_inode;
1731 struct ll_inode_info *li = ll_i2info(inode);
1732 struct ll_sb_info *sbi = ll_i2sbi(inode);
1733 struct obd_device *obd = md_get_real_obd(sbi->ll_md_exp, &li->lli_id);
1734 struct ldlm_res_id res_id = { .name = {id_fid(&li->lli_id),
1735 id_group(&li->lli_id), LDLM_FLOCK} };
1736 struct lustre_handle lockh = {0};
1737 struct ptlrpc_connection *conn;
1738 ptl_process_id_t ptlpid;
1739 ldlm_policy_data_t flock;
1740 ldlm_mode_t mode = 0;
1745 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu cmd=%d file_lock=%p\n",
1746 inode->i_ino, cmd, fl);
1748 if (!(fl->fl_flags & FL_POSIX))
1751 switch (fl->fl_type) {
1757 /* An unlock request may or may not have any relation to
1758 * existing locks so we may not be able to pass a lock handle
1759 * via a normal ldlm_lock_cancel() request. The request may even
1760 * unlock a byte range in the middle of an existing lock. In
1761 * order to process an unlock request we need all of the same
1762 * information that is given with a normal read or write record
1763 * lock request. To avoid creating another ldlm unlock (cancel)
1764 * message we'll treat a LCK_NL flock request as an unlock. */
1771 CERROR("unknown fcntl lock type: %d\n", fl->fl_type);
1780 flags |= LDLM_FL_BLOCK_NOWAIT;
1791 flags = LDLM_FL_TEST_LOCK;
1792 /* Save the old mode so that if the mode in the lock changes we
1793 * can decrement the appropriate reader or writer refcount. */
1797 CERROR("unknown fcntl lock command: %d\n", cmd);
1801 /* Since we're called on every close to remove any oustanding Posix
1802 * flocks owned by the process it's worth a little effort to avoid
1803 * the RPCs if there are no flocks on this file from this node. */
1804 if (mode == LCK_NL && fl->fl_start == 0 && fl->fl_end >= OFFSET_MAX) {
1805 struct ldlm_resource *res;
1807 res = ldlm_resource_get(obd->obd_namespace, NULL,
1808 res_id, LDLM_FLOCK, 0);
1812 ldlm_resource_putref(res);
1815 conn = class_exp2cliimp(obd->obd_self_export)->imp_connection;
1816 if (!conn || !conn->c_peer.peer_ni)
1819 rc = PtlGetId(conn->c_peer.peer_ni->pni_ni_h, &ptlpid);
1823 flock.l_flock.start = fl->fl_start;
1824 flock.l_flock.end = fl->fl_end;
1825 /* XXX - ptlpid.pid is currently coming back a constant; i.e. 12345. */
1826 flock.l_flock.pid = fl->fl_pid;
1827 flock.l_flock.nid = ptlpid.nid;
1828 flock.l_flock.blocking_pid = 0;
1829 flock.l_flock.blocking_nid = 0;
1831 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1832 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1833 flags, mode, flock.l_flock.start, flock.l_flock.end);
1835 rc = ldlm_cli_enqueue(obd->obd_self_export, NULL, obd->obd_namespace,
1836 res_id, LDLM_FLOCK, &flock, mode, &flags,
1837 NULL, ldlm_flock_completion_ast, NULL,
1838 md_get_real_obd(sbi->ll_md_exp, &sbi->ll_rootid),
1839 NULL, 0, NULL, &lockh);
1841 if (flags & LDLM_FL_TEST_LOCK) {
1842 struct ldlm_lock *lock = ldlm_handle2lock(&lockh);
1844 fl->fl_start = lock->l_policy_data.l_flock.start;
1845 fl->fl_end = lock->l_policy_data.l_flock.end;
1846 fl->fl_pid = lock->l_policy_data.l_flock.pid;
1848 switch (lock->l_granted_mode) {
1850 fl->fl_type = F_RDLCK;
1853 fl->fl_type = F_WRLCK;
1856 fl->fl_type = F_UNLCK;
1859 CERROR("unexpected lock type: %d returned from server."
1860 "\n", lock->l_granted_mode);
1865 /* offset the addref() done by ldlm_handle2lock() above. */
1866 LDLM_LOCK_PUT(lock);
1867 ldlm_lock_decref(&lockh, mode);
1868 ldlm_cli_cancel(&lockh);
1870 /* the LDLM_CBPENDING flag was set in the lock by the
1871 * completion AST so the ldlm_lock_decref() call above
1872 * scheduled a blocking AST which will do the final
1873 * lock put on the lock. */
1879 int ll_inode_revalidate_it(struct dentry *dentry)
1881 struct lookup_intent oit = { .it_op = IT_GETATTR };
1882 struct inode *inode = dentry->d_inode;
1883 struct ptlrpc_request *req = NULL;
1884 struct ll_inode_info *lli;
1885 struct lov_stripe_md *lsm;
1886 struct ll_sb_info *sbi;
1887 struct lustre_id id;
1892 CERROR("REPORT THIS LINE TO PETER\n");
1896 sbi = ll_i2sbi(inode);
1898 ll_inode2id(&id, inode);
1899 lli = ll_i2info(inode);
1900 LASSERT(id_fid(&id) != 0);
1902 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), name=%s(%p)\n",
1903 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1906 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1907 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1910 rc = ll_intent_alloc(&oit);
1914 rc = ll_crypto_init_it_key(inode, &oit);
1918 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1919 &oit, 0, &req, ll_mdc_blocking_ast);
1923 rc = revalidate_it_finish(req, 1, &oit, dentry);
1928 ll_lookup_finish_locks(&oit, dentry);
1931 if (!req && (oit.it_op & IT_GETATTR))
1932 ll_audit_log(inode, AUDIT_STAT, 0);
1934 if (!LLI_HAVE_FLSIZE(inode)) {
1935 /* if object not yet allocated, don't validate size */
1938 /* ll_glimpse_size() will prefer locally cached
1939 * writes if they extend the file */
1940 rc = ll_glimpse_size(inode);
1945 ll_intent_release(&oit);
1947 ptlrpc_req_finished(req);
1952 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1953 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
1956 struct inode *inode = de->d_inode;
1957 struct ll_inode_info *lli = ll_i2info(inode);
1959 res = ll_inode_revalidate_it(de);
1960 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1965 stat->ino = inode->i_ino;
1966 stat->mode = inode->i_mode;
1967 stat->nlink = inode->i_nlink;
1968 stat->uid = inode->i_uid;
1969 stat->gid = inode->i_gid;
1970 stat->atime = inode->i_atime;
1971 stat->mtime = inode->i_mtime;
1972 stat->ctime = inode->i_ctime;
1973 stat->blksize = inode->i_blksize;
1975 down(&lli->lli_size_sem);
1976 stat->size = inode->i_size;
1977 stat->blocks = inode->i_blocks;
1978 up(&lli->lli_size_sem);
1980 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1981 stat->dev = id_group(&ll_i2info(inode)->lli_id);
1987 int ll_setxattr_internal(struct inode *inode, const char *name,
1988 const void *value, size_t size, int flags,
1991 struct ll_sb_info *sbi = ll_i2sbi(inode);
1992 struct ptlrpc_request *request = NULL;
1993 struct mdc_op_data *op_data;
1996 int rc = 0, key_size = 0;
1999 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu\n", inode->i_ino);
2000 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_SETXATTR);
2002 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
2003 RETURN(-EOPNOTSUPP);
2005 memset(&attr, 0x0, sizeof(attr));
2006 attr.ia_valid |= valid;
2007 attr.ia_attr_flags = flags;
2009 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
2010 rc = ll_crypto_get_mac(inode, &attr, (void *)value, size,
2013 CERROR("can not get right mac, rc=%d\n", rc);
2018 OBD_ALLOC(op_data, sizeof(*op_data));
2022 ll_inode2mdc_data(op_data, inode, (OBD_MD_FLID | OBD_MD_MEA));
2024 rc = md_setattr(sbi->ll_md_exp, op_data, &attr,
2025 (void *)name, strnlen(name, XATTR_NAME_MAX) + 1,
2026 (void *)value, size, key, key_size, &request);
2027 OBD_FREE(op_data, sizeof(*op_data));
2030 CDEBUG(D_SEC, "md_setattr fails: rc = %d\n", rc);
2034 if (key && key_size)
2035 OBD_FREE(key, key_size);
2036 ptlrpc_req_finished(request);
2040 int ll_setxattr(struct dentry *dentry, const char *name, const void *value,
2041 size_t size, int flags)
2044 struct posix_acl *acl;
2045 struct ll_inode_info *lli;
2048 rc = ll_setxattr_internal(dentry->d_inode, name, value, size,
2051 /* update inode's acl info */
2052 if (rc == 0 && strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
2054 acl = posix_acl_from_xattr(value, size);
2056 CERROR("convert from xattr to acl error: %ld",
2060 error = posix_acl_valid(acl);
2062 CERROR("acl valid error: %d", error);
2063 posix_acl_release(acl);
2071 lli = ll_i2info(dentry->d_inode);
2072 spin_lock(&lli->lli_lock);
2073 if (lli->lli_posix_acl != NULL)
2074 posix_acl_release(lli->lli_posix_acl);
2075 lli->lli_posix_acl = acl;
2076 spin_unlock(&lli->lli_lock);
2083 int ll_removexattr(struct dentry *dentry, const char *name)
2085 return ll_setxattr_internal(dentry->d_inode, name, NULL, 0, 0,
2089 int ll_getxattr_internal(struct inode *inode, const char *name,
2090 void *value, size_t size, __u64 valid)
2092 struct ptlrpc_request *request = NULL;
2093 struct ll_sb_info *sbi = ll_i2sbi(inode);
2094 struct lustre_id id;
2095 struct mds_body *body;
2100 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETXATTR);
2102 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
2103 RETURN(-EOPNOTSUPP);
2105 ll_inode2id(&id, inode);
2106 rc = md_getattr(sbi->ll_md_exp, &id, valid, name, NULL, 0,
2107 size, NULL, &request);
2109 if (rc != -ENODATA && rc != -EOPNOTSUPP)
2110 CERROR("rc = %d\n", rc);
2114 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
2115 LASSERT(body != NULL);
2116 LASSERT_REPSWABBED(request, 0);
2118 ea_size = body->eadatasize;
2120 GOTO(out, rc = ea_size);
2122 LASSERT(ea_size <= request->rq_repmsg->buflens[1]);
2123 ea_data = lustre_msg_buf(request->rq_repmsg, 1, ea_size);
2124 LASSERT(ea_data != NULL);
2125 LASSERT_REPSWABBED(request, 1);
2128 memcpy(value, ea_data, ea_size);
2132 ptlrpc_req_finished(request);
2136 int ll_getxattr(struct dentry *dentry, const char *name, void *value,
2139 return ll_getxattr_internal(dentry->d_inode, name,
2140 value, size, OBD_MD_FLXATTR);
2143 int ll_listxattr(struct dentry *dentry, char *list, size_t size)
2145 return ll_getxattr_internal(dentry->d_inode, NULL, list, size,
2146 OBD_MD_FLXATTRLIST);
2150 * Here we hold DLM lock across permission check, to get better
2151 * conformance especially for remote acl. will introduce certain
2152 * overhead, maybe need a profile.
2155 lustre_check_acl(struct inode *inode, int mask)
2157 struct lookup_intent it = { .it_op = IT_GETATTR };
2158 struct dentry de = { .d_inode = inode };
2159 struct ll_sb_info *sbi;
2160 struct lustre_id id;
2161 struct ptlrpc_request *req = NULL;
2162 struct ll_inode_info *lli = ll_i2info(inode);
2163 struct posix_acl *acl;
2167 sbi = ll_i2sbi(inode);
2168 ll_inode2id(&id, inode);
2170 if (ll_intent_alloc(&it))
2173 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
2174 &it, 0, &req, ll_mdc_blocking_ast);
2176 ll_intent_free(&it);
2180 rc = revalidate_it_finish(req, 1, &it, &de);
2182 ll_intent_release(&it);
2186 if (sbi->ll_remote) {
2187 rc = ll_remote_acl_permission(inode, mask);
2189 spin_lock(&lli->lli_lock);
2190 acl = posix_acl_dup(ll_i2info(inode)->lli_posix_acl);
2191 spin_unlock(&lli->lli_lock);
2196 rc = posix_acl_permission(inode, acl, mask);
2197 posix_acl_release(acl);
2201 ll_lookup_finish_locks(&it, &de);
2202 ll_intent_free(&it);
2206 ptlrpc_req_finished(req);
2211 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
2213 struct ll_sb_info *sbi = ll_i2sbi(inode);
2215 /* for remote client, permission bits in inode doesn't
2216 * play a role anymore.
2219 return lustre_check_acl(inode, mask);
2221 return generic_permission(inode, mask, lustre_check_acl);
2224 struct file_operations ll_file_operations = {
2225 .read = ll_file_read,
2226 .write = ll_file_write,
2227 .ioctl = ll_file_ioctl,
2228 .open = ll_file_open,
2229 .release = ll_file_release,
2230 .mmap = ll_file_mmap,
2231 .llseek = ll_file_seek,
2232 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
2233 .sendfile = generic_file_sendfile,
2236 .lock = ll_file_flock
2239 struct inode_operations ll_file_inode_operations = {
2240 .setattr = ll_setattr,
2241 .truncate = ll_truncate,
2242 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
2243 .getattr = ll_getattr,
2245 .revalidate_it = ll_inode_revalidate_it,
2247 .setxattr = ll_setxattr,
2248 .getxattr = ll_getxattr,
2249 .listxattr = ll_listxattr,
2250 .removexattr = ll_removexattr,
2251 .permission = ll_inode_permission,