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 "llite_internal.h"
39 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
40 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
41 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
43 int ll_validate_size(struct inode *inode, __u64 *size, __u64 *blocks)
45 ldlm_policy_data_t extent = { .l_extent = { 0, OBD_OBJECT_EOF } };
46 struct obd_export *exp = ll_i2sbi(inode)->ll_dt_exp;
47 struct ll_inode_info *lli = ll_i2info(inode);
48 struct lustre_handle match_lockh = {0};
52 if (lli->lli_smd == NULL)
55 LASSERT(size != NULL && blocks != NULL);
57 flags = LDLM_FL_TEST_LOCK | LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
58 rc = obd_match(exp, lli->lli_smd, LDLM_EXTENT, &extent,
59 LCK_PR | LCK_PW, &flags, inode, &match_lockh);
61 /* we have no all needed locks,
62 * so we don't know actual size */
66 /* we know actual size! */
67 down(&lli->lli_size_sem);
68 *size = lov_merge_size(lli->lli_smd, 0);
69 *blocks = lov_merge_blocks(lli->lli_smd);
70 up(&lli->lli_size_sem);
76 int ll_md_och_close(struct obd_export *md_exp, struct inode *inode,
77 struct obd_client_handle *och, int dirty,
80 struct ptlrpc_request *req = NULL;
81 struct mdc_op_data *op_data;
82 struct obd_device *obd;
87 obd = class_exp2obd(md_exp);
89 CERROR("Invalid MDC connection handle "LPX64"\n",
90 md_exp->exp_handle.h_cookie);
94 /* here we check if this is forced umount. If so this is called on
95 * canceling "open lock" and we do not call md_close() in this case , as
96 * it will not successful, as import is already deactivated. */
97 if (obd->obd_no_recov)
100 /* prepare @op_data for close request */
101 OBD_ALLOC(op_data, sizeof(*op_data));
105 memset(op_data, 0, sizeof(*op_data));
107 valid = (OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
108 OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLEPOCH |
111 ll_inode2mdc_data(op_data, inode, valid);
113 if (0 /* ll_is_inode_dirty(inode) */) {
114 op_data->flags = MDS_BFLAG_UNCOMMITTED_WRITES;
115 op_data->valid |= OBD_MD_FLFLAGS;
119 /* we modified data through this handle */
120 op_data->io_epoch = epoch;
121 op_data->flags |= MDS_BFLAG_DIRTY_EPOCH;
122 op_data->valid |= OBD_MD_FLFLAGS | OBD_MD_FLEPOCH;
123 if (ll_validate_size(inode, &op_data->size, &op_data->blocks))
124 op_data->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
127 rc = md_close(md_exp, op_data, och, &req);
128 OBD_FREE(op_data, sizeof(*op_data));
131 /* we are the last writer, so the MDS has instructed us to get
132 * the file size and any write cookies, then close again. */
134 //ll_queue_done_writing(inode);
137 CERROR("inode %lu mdc close failed: rc = %d\n",
138 (unsigned long)inode->i_ino, rc);
141 ptlrpc_req_finished(req);
144 mdc_clear_open_replay_data(md_exp, och);
145 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
146 OBD_FREE(och, sizeof *och);
150 int ll_md_real_close(struct obd_export *md_exp,
151 struct inode *inode, int flags)
153 struct ll_inode_info *lli = ll_i2info(inode);
154 int freeing = inode->i_state & I_FREEING;
155 struct obd_client_handle **och_p;
156 struct obd_client_handle *och;
157 __u64 *och_usecount, epoch = 0;
158 int rc = 0, dirty = 0;
161 if (flags & FMODE_WRITE) {
162 och_p = &lli->lli_mds_write_och;
163 och_usecount = &lli->lli_open_fd_write_count;
164 } else if (flags & FMODE_EXEC) {
165 och_p = &lli->lli_mds_exec_och;
166 och_usecount = &lli->lli_open_fd_exec_count;
168 och_p = &lli->lli_mds_read_och;
169 och_usecount = &lli->lli_open_fd_read_count;
172 down(&lli->lli_och_sem);
173 if (*och_usecount) { /* There are still users of this handle, so
175 up(&lli->lli_och_sem);
178 if (ll_is_inode_dirty(inode)) {
179 /* the inode still has dirty pages, let's close later */
180 CDEBUG(D_INODE, "inode %lu/%u still has dirty pages\n",
181 inode->i_ino, inode->i_generation);
182 LASSERT(freeing == 0);
183 ll_queue_done_writing(inode);
184 up(&lli->lli_och_sem);
188 if (LLI_DIRTY_HANDLE(inode) && (flags & FMODE_WRITE)) {
189 clear_bit(LLI_F_DIRTY_HANDLE, &lli->lli_flags);
191 } else if (0 && !(flags & FMODE_SYNC) && !freeing) {
192 /* in order to speed up creation rate we pass
193 * closing to dedicated thread so we don't need
194 * to wait for close reply here -bzzz */
195 ll_queue_done_writing(inode);
196 up(&lli->lli_och_sem);
202 if (flags & FMODE_WRITE) {
203 epoch = lli->lli_io_epoch;
204 lli->lli_io_epoch = 0;
207 up(&lli->lli_och_sem);
210 * there might be a race and somebody have freed this och
211 * already. Another way to have this twice called is if file closing
212 * will fail due to netwok problems and on umount lock will be canceled
213 * and this will be called from block_ast callack.
215 if (och && och->och_fh.cookie != DEAD_HANDLE_MAGIC)
216 rc = ll_md_och_close(md_exp, inode, och, dirty, epoch);
221 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
224 struct ll_file_data *fd = file->private_data;
225 struct ll_inode_info *lli = ll_i2info(inode);
229 /* clear group lock, if present */
230 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
231 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
232 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
233 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
237 /* Let's see if we have good enough OPEN lock on the file and if
238 we can skip talking to MDS */
239 if (file->f_dentry->d_inode) {
241 struct obd_device *obddev;
242 struct lustre_handle lockh;
243 int flags = LDLM_FL_BLOCK_GRANTED;
244 struct ldlm_res_id file_res_id = {.name = {id_fid(&lli->lli_id),
245 id_group(&lli->lli_id)}};
246 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
248 down(&lli->lli_och_sem);
249 if (fd->fd_omode & FMODE_WRITE) {
251 LASSERT(lli->lli_open_fd_write_count);
252 lli->lli_open_fd_write_count--;
253 } else if (fd->fd_omode & FMODE_EXEC) {
255 LASSERT(lli->lli_open_fd_exec_count);
256 lli->lli_open_fd_exec_count--;
259 LASSERT(lli->lli_open_fd_read_count);
260 lli->lli_open_fd_read_count--;
262 up(&lli->lli_och_sem);
264 obddev = md_get_real_obd(md_exp, &lli->lli_id);
265 if (!ldlm_lock_match(obddev->obd_namespace, flags, &file_res_id,
266 LDLM_IBITS, &policy, lockmode, &lockh))
268 rc = ll_md_real_close(md_exp, file->f_dentry->d_inode,
271 ldlm_lock_decref(&lockh, lockmode);
275 file->private_data = NULL;
276 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof(*fd));
280 /* While this returns an error code, fput() the caller does not, so we need
281 * to make every effort to clean up all of our state here. Also, applications
282 * rarely check close errors and even if an error is returned they will not
283 * re-try the close call.
285 int ll_file_release(struct inode *inode, struct file *file)
287 struct ll_file_data *fd;
288 struct ll_sb_info *sbi = ll_i2sbi(inode);
292 CDEBUG(D_VFSTRACE, "VFS Op:inode="DLID4"(%p)\n",
293 OLID4(&ll_i2info(inode)->lli_id), inode);
295 /* don't do anything for / */
296 if (inode->i_sb->s_root == file->f_dentry)
299 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
300 fd = (struct ll_file_data *)file->private_data;
303 rc = ll_md_close(sbi->ll_md_exp, inode, file);
307 static int ll_intent_file_open(struct file *file, void *lmm,
308 int lmmsize, struct lookup_intent *itp)
310 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
311 const char *name = (char *)file->f_dentry->d_name.name;
312 struct dentry *parent = file->f_dentry->d_parent;
313 const int len = file->f_dentry->d_name.len;
314 struct lustre_handle lockh;
315 struct mdc_op_data *op_data;
321 OBD_ALLOC(op_data, sizeof(*op_data));
325 ll_prepare_mdc_data(op_data, parent->d_inode, NULL,
328 rc = md_enqueue(sbi->ll_md_exp, LDLM_IBITS, itp, LCK_PR, op_data,
329 &lockh, lmm, lmmsize, ldlm_completion_ast,
330 ll_mdc_blocking_ast, NULL);
331 OBD_FREE(op_data, sizeof(*op_data));
333 if (LUSTRE_IT(itp)->it_lock_mode)
334 memcpy(&LUSTRE_IT(itp)->it_lock_handle,
335 &lockh, sizeof(lockh));
338 CERROR("lock enqueue: err: %d\n", rc);
343 int ll_och_fill(struct inode *inode, struct lookup_intent *it,
344 struct obd_client_handle *och)
346 struct ptlrpc_request *req = LUSTRE_IT(it)->it_data;
347 struct ll_inode_info *lli = ll_i2info(inode);
348 struct mds_body *body;
353 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
354 LASSERT (body != NULL); /* reply already checked out */
355 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
357 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
358 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
359 if (it->it_flags & FMODE_WRITE) {
360 if (lli->lli_io_epoch && lli->lli_io_epoch != body->io_epoch)
361 CDEBUG(D_ERROR, "new epoch?! "LPD64" != "LPD64"\n",
362 lli->lli_io_epoch, body->io_epoch);
363 lli->lli_io_epoch = body->io_epoch;
365 mdc_set_open_replay_data(ll_i2mdexp(inode), och,
366 LUSTRE_IT(it)->it_data);
371 int ll_local_open(struct file *file, struct lookup_intent *it,
372 struct obd_client_handle *och)
374 struct ll_file_data *fd;
375 struct inode *inode = file->f_dentry->d_inode;
380 rc = ll_och_fill(inode, it, och);
385 LASSERTF(file->private_data == NULL, "file %.*s/%.*s ino %lu/%u (%o)\n",
386 file->f_dentry->d_name.len, file->f_dentry->d_name.name,
387 file->f_dentry->d_parent->d_name.len,
388 file->f_dentry->d_parent->d_name.name,
389 file->f_dentry->d_inode->i_ino,
390 file->f_dentry->d_inode->i_generation,
391 file->f_dentry->d_inode->i_mode);
393 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
395 /* We can't handle this well without reorganizing ll_file_open and
396 * ll_md_close(), so don't even try right now. */
399 file->private_data = fd;
400 ll_readahead_init(inode, &fd->fd_ras);
401 fd->fd_omode = it->it_flags;
406 /* Open a file, and (for the very first open) create objects on the OSTs at
407 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
408 * creation or open until ll_lov_setstripe() ioctl is called. We grab
409 * lli_open_sem to ensure no other process will create objects, send the
410 * stripe MD to the MDS, or try to destroy the objects if that fails.
412 * If we already have the stripe MD locally then we don't request it in
413 * mdc_open(), by passing a lmm_size = 0.
415 * It is up to the application to ensure no other processes open this file
416 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
417 * used. We might be able to avoid races of that sort by getting lli_open_sem
418 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
419 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
421 int ll_file_open(struct inode *inode, struct file *file)
423 struct ll_inode_info *lli = ll_i2info(inode);
424 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
425 .it_flags = file->f_flags };
426 struct lov_stripe_md *lsm;
427 struct ptlrpc_request *req;
429 struct obd_client_handle **och_p = NULL;
430 __u64 *och_usecount = NULL;
433 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n",
434 inode->i_ino, inode->i_generation, inode, file->f_flags);
436 /* don't do anything for / */
437 if (inode->i_sb->s_root == file->f_dentry)
440 if ((file->f_flags+1) & O_ACCMODE)
442 if (file->f_flags & O_TRUNC)
447 /* sometimes LUSTRE_IT(it) may not be allocated like opening file by
448 * dentry_open() from GNS stuff. */
449 if (!it || !LUSTRE_IT(it)) {
451 rc = ll_intent_alloc(it);
456 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
458 /* mdc_intent_lock() didn't get a request ref if there was an open
459 * error, so don't do cleanup on the * request here (bug 3430) */
460 if (LUSTRE_IT(it)->it_disposition) {
461 rc = it_open_error(DISP_OPEN_OPEN, it);
466 /* Let's see if we have file open on MDS already. */
467 if (it->it_flags & FMODE_WRITE) {
468 och_p = &lli->lli_mds_write_och;
469 och_usecount = &lli->lli_open_fd_write_count;
470 } else if (it->it_flags & FMODE_EXEC) {
471 och_p = &lli->lli_mds_exec_och;
472 och_usecount = &lli->lli_open_fd_exec_count;
474 och_p = &lli->lli_mds_read_och;
475 och_usecount = &lli->lli_open_fd_read_count;
478 rc = ll_crypto_decrypt_key(inode, it);
482 down(&lli->lli_och_sem);
483 if (*och_p) { /* Open handle is present */
484 if (it_disposition(it, DISP_LOOKUP_POS) && /* Positive lookup */
485 it_disposition(it, DISP_OPEN_OPEN)) { /* & OPEN happened */
486 struct obd_client_handle *och;
487 /* Well, there's extra open request that we do not need,
488 let's close it somehow*/
489 OBD_ALLOC(och, sizeof (struct obd_client_handle));
491 up(&lli->lli_och_sem);
495 rc = ll_och_fill(inode, it, och);
497 up(&lli->lli_och_sem);
501 /* ll_md_och_close() will free och */
502 ll_md_och_close(ll_i2mdexp(inode), inode, och, 0, 0);
506 rc = ll_local_open(file, it, NULL);
510 rc = ll_set_capa(inode, it, *och_p);
512 up(&lli->lli_och_sem);
516 LASSERT(*och_usecount == 0);
517 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
519 GOTO(out, rc = -ENOMEM);
522 if (!it || !LUSTRE_IT(it) || !LUSTRE_IT(it)->it_disposition) {
524 * we are going to replace intent here, and that may
525 * possibly change access mode (FMODE_EXEC can only be
526 * set in intent), but I hope it never happens (I was
527 * not able to trigger it yet at least) -- green
530 /* FIXME: FMODE_EXEC is not covered by O_ACCMODE! */
531 LASSERT(!(it->it_flags & FMODE_EXEC));
532 LASSERTF((it->it_flags & O_ACCMODE) ==
533 (oit.it_flags & O_ACCMODE), "Changing intent "
534 "flags %x to incompatible %x\n", it->it_flags,
537 rc = ll_intent_file_open(file, NULL, 0, it);
540 rc = it_open_error(DISP_OPEN_OPEN, it);
542 GOTO(out_och_free, rc);
544 mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle,
545 file->f_dentry->d_inode);
547 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
548 rc = ll_local_open(file, it, *och_p);
549 LASSERTF(rc == 0, "rc = %d\n", rc);
551 rc = ll_set_capa(inode, it, *och_p);
553 up(&lli->lli_och_sem);
557 up(&lli->lli_och_sem);
560 * must do this outside lli_och_sem lock to prevent deadlock where
561 * different kind of OPEN lock for this same inode gets cancelled by
564 if (!S_ISREG(inode->i_mode))
569 if (file->f_flags & O_LOV_DELAY_CREATE ||
570 !(file->f_mode & FMODE_WRITE)) {
571 CDEBUG(D_INODE, "object creation was delayed\n");
575 file->f_flags &= ~O_LOV_DELAY_CREATE;
578 /* audit stuff if there was no RPC */
579 if (LUSTRE_IT(it)->it_data == 0)
580 ll_audit_log(inode, AUDIT_OPEN, rc);
582 req = LUSTRE_IT(it)->it_data;
583 ll_intent_drop_lock(it);
584 ll_intent_release(it);
587 ptlrpc_req_finished(req);
589 ll_open_complete(inode);
593 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
594 *och_p = NULL; /* OBD_FREE writes some magic there */
597 up(&lli->lli_och_sem);
603 /* Fills the obdo with the attributes for the inode defined by lsm */
604 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
607 struct ptlrpc_request_set *set;
611 LASSERT(lsm != NULL);
613 memset(oa, 0, sizeof *oa);
614 oa->o_id = lsm->lsm_object_id;
615 oa->o_gr = lsm->lsm_object_gr;
616 oa->o_mode = S_IFREG;
617 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
618 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
619 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
621 set = ptlrpc_prep_set();
625 rc = obd_getattr_async(exp, oa, lsm, set);
627 rc = ptlrpc_set_wait(set);
628 ptlrpc_set_destroy(set);
633 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
634 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
638 static inline void ll_remove_suid(struct inode *inode)
642 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
643 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
645 /* was any of the uid bits set? */
646 mode &= inode->i_mode;
647 if (mode && !capable(CAP_FSETID)) {
648 inode->i_mode &= ~mode;
649 // XXX careful here - we cannot change the size
653 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
655 struct ll_inode_info *lli = ll_i2info(inode);
656 struct lov_stripe_md *lsm = lli->lli_smd;
657 struct obd_export *exp = ll_i2dtexp(inode);
660 struct ldlm_lock *lock;
661 struct lov_stripe_md *lsm;
662 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
663 __u32 stripe, vallen = sizeof(stripe);
667 if (lsm->lsm_stripe_count == 1)
668 GOTO(check, stripe = 0);
670 /* get our offset in the lov */
671 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
673 CERROR("obd_get_info: rc = %d\n", rc);
676 LASSERT(stripe < lsm->lsm_stripe_count);
679 if (lsm->lsm_oinfo[stripe].loi_id != lock->l_resource->lr_name.name[0]||
680 lsm->lsm_oinfo[stripe].loi_gr != lock->l_resource->lr_name.name[2]){
681 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64
682 " inode=%lu/%u (%p)\n",
683 lsm->lsm_oinfo[stripe].loi_id,
684 lsm->lsm_oinfo[stripe].loi_gr,
685 inode->i_ino, inode->i_generation, inode);
686 return -ELDLM_NO_LOCK_DATA;
692 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
693 * we get a lock cancellation for each stripe, so we have to map the obd's
694 * region back onto the stripes in the file that it held.
696 * No one can dirty the extent until we've finished our work and they can
697 * enqueue another lock. The DLM protects us from ll_file_read/write here,
698 * but other kernel actors could have pages locked.
700 * Called with the DLM lock held. */
701 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
702 struct ldlm_lock *lock, __u32 stripe)
704 ldlm_policy_data_t tmpex;
705 unsigned long start, end, count, skip, i, j;
707 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
708 struct lustre_handle lockh;
711 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
712 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
713 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
716 /* our locks are page granular thanks to osc_enqueue, we invalidate the
718 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
719 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
723 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
724 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
725 if (lsm->lsm_stripe_count > 1) {
726 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
727 skip = (lsm->lsm_stripe_count - 1) * count;
728 start += start/count * skip + stripe * count;
730 end += end/count * skip + stripe * count;
732 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
735 i = inode->i_size ? (inode->i_size - 1) >> PAGE_CACHE_SHIFT : 0;
739 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
740 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
741 count, skip, end, discard ? " (DISCARDING)" : "");
743 /* walk through the vmas on the inode and tear down mmaped pages that
744 * intersect with the lock. this stops immediately if there are no
745 * mmap()ed regions of the file. This is not efficient at all and
746 * should be short lived. We'll associate mmap()ed pages with the lock
747 * and will be able to find them directly */
749 for (i = start; i <= end; i += (j + skip)) {
750 j = min(count - (i % count), end - i + 1);
752 LASSERT(inode->i_mapping);
753 if (ll_teardown_mmaps(inode->i_mapping,
754 (__u64)i << PAGE_CACHE_SHIFT,
755 ((__u64)(i+j) << PAGE_CACHE_SHIFT) - 1) )
759 /* this is the simplistic implementation of page eviction at
760 * cancelation. It is careful to get races with other page
761 * lockers handled correctly. fixes from bug 20 will make it
762 * more efficient by associating locks with pages and with
763 * batching writeback under the lock explicitly. */
764 for (i = start, j = start % count; i <= end;
765 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
767 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
773 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
774 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
775 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
778 if (!mapping_has_pages(inode->i_mapping)) {
779 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
785 page = find_get_page(inode->i_mapping, i);
788 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
789 i, tmpex.l_extent.start);
792 /* page->mapping to check with racing against teardown */
793 if (!discard && clear_page_dirty_for_io(page)) {
794 rc = ll_call_writepage(inode, page);
796 CERROR("writepage of page %p failed: %d\n",
798 /* either waiting for io to complete or reacquiring
799 * the lock that the failed writepage released */
803 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
804 /* check to see if another DLM lock covers this page */
805 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
806 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
808 &lock->l_resource->lr_name, LDLM_EXTENT,
809 &tmpex, LCK_PR | LCK_PW, &lockh);
810 if (rc2 == 0 && page->mapping != NULL) {
811 // checking again to account for writeback's lock_page()
812 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
813 ll_ra_accounting(page, inode->i_mapping);
814 ll_truncate_complete_page(page);
817 page_cache_release(page);
819 LASSERTF(tmpex.l_extent.start <=
820 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
821 lock->l_policy_data.l_extent.end + 1),
822 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
823 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
828 static int ll_extent_lock_callback(struct ldlm_lock *lock,
829 struct ldlm_lock_desc *new, void *data,
832 struct lustre_handle lockh = { 0 };
836 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
837 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
842 case LDLM_CB_BLOCKING:
843 ldlm_lock2handle(lock, &lockh);
844 rc = ldlm_cli_cancel(&lockh);
846 CERROR("ldlm_cli_cancel failed: %d\n", rc);
848 case LDLM_CB_CANCELING: {
850 struct ll_inode_info *lli;
851 struct lov_stripe_md *lsm;
855 /* This lock wasn't granted, don't try to evict pages */
856 if (lock->l_req_mode != lock->l_granted_mode)
859 inode = ll_inode_from_lock(lock);
862 lli = ll_i2info(inode);
865 if (lli->lli_smd == NULL)
869 stripe = ll_lock_to_stripe_offset(inode, lock);
871 CERROR("ll_lock_to_stripe_offset failed: %d\n", stripe);
875 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
877 down(&lli->lli_size_sem);
878 lock_res_and_lock(lock);
879 kms = ldlm_extent_shift_kms(lock,
880 lsm->lsm_oinfo[stripe].loi_kms);
882 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
883 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
884 lsm->lsm_oinfo[stripe].loi_kms, kms);
885 lsm->lsm_oinfo[stripe].loi_kms = kms;
886 unlock_res_and_lock(lock);
887 up(&lli->lli_size_sem);
888 //ll_try_done_writing(inode);
901 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
903 /* XXX ALLOCATE - 160 bytes */
904 struct inode *inode = ll_inode_from_lock(lock);
905 struct ll_inode_info *lli = ll_i2info(inode);
906 struct lustre_handle lockh = { 0 };
911 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
912 LDLM_FL_BLOCK_CONV)) {
913 LBUG(); /* not expecting any blocked async locks yet */
914 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
916 ldlm_lock_dump(D_OTHER, lock, 0);
917 ldlm_reprocess_all(lock->l_resource);
921 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
923 stripe = ll_lock_to_stripe_offset(inode, lock);
925 CERROR("ll_lock_to_stripe_offset failed: %d\n", stripe);
929 if (lock->l_lvb_len) {
930 struct lov_stripe_md *lsm = lli->lli_smd;
932 lvb = lock->l_lvb_data;
933 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
936 lock_res_and_lock(lock);
937 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
938 kms = ldlm_extent_shift_kms(NULL, kms);
939 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
940 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
941 lsm->lsm_oinfo[stripe].loi_kms, kms);
942 lsm->lsm_oinfo[stripe].loi_kms = kms;
943 unlock_res_and_lock(lock);
949 wake_up(&lock->l_waitq);
951 ldlm_lock2handle(lock, &lockh);
952 ldlm_lock_decref(&lockh, LCK_PR);
957 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
959 struct ptlrpc_request *req = reqp;
960 struct inode *inode = ll_inode_from_lock(lock);
961 struct ll_inode_info *lli;
963 struct lov_stripe_md *lsm;
964 int rc, size = sizeof(*lvb), stripe;
968 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
969 lli = ll_i2info(inode);
971 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
975 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
977 /* First, find out which stripe index this lock corresponds to. */
978 stripe = ll_lock_to_stripe_offset(inode, lock);
980 CERROR("ll_lock_to_stripe_offset failed: %d\n", stripe);
981 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
984 rc = lustre_pack_reply(req, 1, &size, NULL);
986 CERROR("lustre_pack_reply: %d\n", rc);
990 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
991 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
992 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
993 lvb->lvb_atime = LTIME_S(inode->i_atime);
994 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
996 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
997 inode->i_size, stripe, lvb->lvb_size);
1003 /* These errors are normal races, so we don't want to fill the console
1004 * with messages by calling ptlrpc_error() */
1005 if (rc == -ELDLM_NO_LOCK_DATA)
1006 lustre_pack_reply(req, 0, NULL, NULL);
1008 req->rq_status = rc;
1012 /* NB: lov_merge_size will prefer locally cached writes if they extend the
1013 * file (because it prefers KMS over RSS when larger) */
1014 int ll_glimpse_size(struct inode *inode)
1016 struct ll_inode_info *lli = ll_i2info(inode);
1017 struct ll_sb_info *sbi = ll_i2sbi(inode);
1018 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1019 struct lustre_handle lockh = { 0 };
1020 int rc, flags = LDLM_FL_HAS_INTENT;
1023 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1025 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
1026 LCK_PR, &flags, ll_extent_lock_callback,
1027 ldlm_completion_ast, ll_glimpse_callback, inode,
1028 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1033 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1034 RETURN(rc > 0 ? -EIO : rc);
1037 down(&lli->lli_size_sem);
1038 inode->i_size = lov_merge_size(lli->lli_smd, 0);
1039 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
1040 up(&lli->lli_size_sem);
1042 LTIME_S(inode->i_mtime) = lov_merge_mtime(lli->lli_smd,
1043 LTIME_S(inode->i_mtime));
1045 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
1046 (__u64)inode->i_size, (__u64)inode->i_blocks);
1048 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
1052 void ll_stime_record(struct ll_sb_info *sbi, struct timeval *start,
1053 struct obd_service_time *stime)
1055 struct timeval stop;
1056 do_gettimeofday(&stop);
1058 spin_lock(&sbi->ll_lock);
1059 lprocfs_stime_record(stime, &stop, start);
1060 spin_unlock(&sbi->ll_lock);
1063 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
1064 struct lov_stripe_md *lsm, int mode,
1065 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
1066 int ast_flags, struct obd_service_time *stime)
1068 struct ll_inode_info *lli = ll_i2info(inode);
1069 struct ll_sb_info *sbi = ll_i2sbi(inode);
1070 struct timeval start;
1074 LASSERT(lockh->cookie == 0);
1076 /* XXX phil: can we do this? won't it screw the file size up? */
1077 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1078 (sbi->ll_flags & LL_SBI_NOLCK))
1081 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
1082 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
1084 do_gettimeofday(&start);
1085 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
1086 &ast_flags, ll_extent_lock_callback,
1087 ldlm_completion_ast, ll_glimpse_callback, inode,
1088 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
1092 ll_stime_record(sbi, &start, stime);
1094 if (policy->l_extent.start == 0 &&
1095 policy->l_extent.end == OBD_OBJECT_EOF) {
1096 /* vmtruncate()->ll_truncate() first sets the i_size and then
1097 * the kms under both a DLM lock and the i_sem. If we don't
1098 * get the i_sem here we can match the DLM lock and reset
1099 * i_size from the kms before the truncating path has updated
1100 * the kms. generic_file_write can then trust the stale i_size
1101 * when doing appending writes and effectively cancel the
1102 * result of the truncate. Getting the i_sem after the enqueue
1103 * maintains the DLM -> i_sem acquiry order. */
1104 down(&lli->lli_size_sem);
1105 inode->i_size = lov_merge_size(lsm, 1);
1106 up(&lli->lli_size_sem);
1110 LTIME_S(inode->i_mtime) =
1111 lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1117 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1118 struct lov_stripe_md *lsm, int mode,
1119 struct lustre_handle *lockh)
1121 struct ll_sb_info *sbi = ll_i2sbi(inode);
1125 /* XXX phil: can we do this? won't it screw the file size up? */
1126 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1127 (sbi->ll_flags & LL_SBI_NOLCK))
1130 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1135 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1138 struct inode *inode = file->f_dentry->d_inode;
1139 struct ll_inode_info *lli = ll_i2info(inode);
1140 struct lov_stripe_md *lsm = lli->lli_smd;
1141 struct ll_lock_tree tree;
1142 struct ll_lock_tree_node *node;
1147 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1148 inode->i_ino, inode->i_generation, inode, count, *ppos);
1150 /* "If nbyte is 0, read() will return 0 and have no other results."
1151 * -- Single Unix Spec */
1155 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
1161 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1164 tree.lt_fd = file->private_data;
1166 rc = ll_tree_lock(&tree, node, inode, buf, count,
1167 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1171 down(&lli->lli_size_sem);
1172 kms = lov_merge_size(lsm, 1);
1173 if (*ppos + count - 1 > kms) {
1174 /* A glimpse is necessary to determine whether we return a short
1175 * read or some zeroes at the end of the buffer */
1176 up(&lli->lli_size_sem);
1177 retval = ll_glimpse_size(inode);
1181 inode->i_size = kms;
1182 up(&lli->lli_size_sem);
1185 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1186 inode->i_ino, count, *ppos, inode->i_size);
1188 /* turn off the kernel's read-ahead */
1189 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1192 file->f_ra.ra_pages = 0;
1194 retval = generic_file_read(file, buf, count, ppos);
1195 ll_audit_log(inode, AUDIT_READ, retval);
1198 ll_tree_unlock(&tree, inode);
1203 * Write to a file (through the page cache).
1205 static ssize_t ll_file_write(struct file *file, const char *buf,
1206 size_t count, loff_t *ppos)
1208 struct inode *inode = file->f_dentry->d_inode;
1209 loff_t maxbytes = ll_file_maxbytes(inode);
1210 struct ll_lock_tree tree;
1211 struct ll_lock_tree_node *node;
1216 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1217 inode->i_ino, inode->i_generation, inode, count, *ppos);
1219 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1221 /* POSIX, but surprised the VFS doesn't check this already */
1225 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1226 * called on the file, don't fail the below assertion (bug 2388). */
1227 if (file->f_flags & O_LOV_DELAY_CREATE &&
1228 ll_i2info(inode)->lli_smd == NULL)
1231 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1233 if (file->f_flags & O_APPEND)
1234 node = ll_node_from_inode(inode, 0, OBD_OBJECT_EOF, LCK_PW);
1236 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1240 RETURN(PTR_ERR(node));
1242 tree.lt_fd = file->private_data;
1244 rc = ll_tree_lock(&tree, node, inode, buf, count,
1245 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1249 /* this is ok, g_f_w will overwrite this under i_sem if it races
1250 * with a local truncate, it just makes our maxbyte checking easier */
1251 if (file->f_flags & O_APPEND)
1252 *ppos = inode->i_size;
1254 if (*ppos >= maxbytes) {
1255 if (count || *ppos > maxbytes) {
1256 send_sig(SIGXFSZ, current, 0);
1257 GOTO(out, retval = -EFBIG);
1260 if (*ppos + count > maxbytes)
1261 count = maxbytes - *ppos;
1263 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1264 inode->i_ino, count, *ppos);
1266 /* mark open handle dirty */
1267 set_bit(LLI_F_DIRTY_HANDLE, &(ll_i2info(inode)->lli_flags));
1269 /* generic_file_write handles O_APPEND after getting i_sem */
1270 retval = generic_file_write(file, buf, count, ppos);
1274 /* ll_audit_log(inode, AUDIT_WRITE, retval); */
1276 ll_tree_unlock(&tree, inode);
1277 /* serialize with mmap/munmap/mremap */
1278 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
1279 retval > 0 ? retval : 0);
1283 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
1284 int flags, struct lov_user_md *lum,
1287 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
1288 struct ll_inode_info *lli = ll_i2info(inode);
1289 struct ptlrpc_request *req = NULL;
1290 struct obd_client_handle *och;
1291 struct lov_stripe_md *lsm;
1292 struct lustre_md md;
1298 if ((file->f_flags+1) & O_ACCMODE)
1300 if (file->f_flags & O_TRUNC)
1303 down(&lli->lli_open_sem);
1306 up(&lli->lli_open_sem);
1307 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
1312 f = get_empty_filp();
1316 f->f_dentry = file->f_dentry;
1317 f->f_vfsmnt = file->f_vfsmnt;
1320 rc = ll_intent_alloc(&oit);
1324 rc = ll_intent_file_open(f, lum, lum_size, &oit);
1327 if (it_disposition(&oit, DISP_LOOKUP_NEG))
1330 req = LUSTRE_IT(&oit)->it_data;
1331 rc = LUSTRE_IT(&oit)->it_status;
1336 rc = mdc_req2lustre_md(ll_i2mdexp(inode), req, 1,
1337 ll_i2dtexp(inode), &md);
1340 ll_update_inode(f->f_dentry->d_inode, &md);
1342 OBD_ALLOC(och, sizeof(struct obd_client_handle));
1344 GOTO(out, rc = -ENOMEM);
1346 /* actually ll_local_open() cannot fail! */
1347 rc = ll_local_open(f, &oit, och);
1351 if (LUSTRE_IT(&oit)->it_lock_mode) {
1352 ldlm_lock_decref_and_cancel((struct lustre_handle *)
1353 &LUSTRE_IT(&oit)->it_lock_handle,
1354 LUSTRE_IT(&oit)->it_lock_mode);
1355 LUSTRE_IT(&oit)->it_lock_mode = 0;
1358 ll_intent_release(&oit);
1360 /* ll_file_release() will decrease the count, but won't free anything
1361 * because we have at least one more reference coming from actual
1363 down(&lli->lli_och_sem);
1364 lli->lli_open_fd_write_count++;
1365 up(&lli->lli_och_sem);
1366 rc = ll_file_release(f->f_dentry->d_inode, f);
1368 /* Now also destroy our supplemental och */
1369 ll_md_och_close(ll_i2mdexp(inode), f->f_dentry->d_inode, och, 0, 0);
1372 ll_intent_release(&oit);
1375 up(&lli->lli_open_sem);
1377 ptlrpc_req_finished(req);
1381 static int ll_lov_setea(struct inode *inode, struct file *file,
1384 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
1385 int lum_size = sizeof(struct lov_user_md) +
1386 sizeof(struct lov_user_ost_data);
1387 struct lov_user_md *lump;
1391 if (!capable(CAP_SYS_ADMIN))
1394 OBD_ALLOC(lump, lum_size);
1398 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
1400 GOTO(out_free_lump, rc = -EFAULT);
1402 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1405 OBD_FREE(lump, lum_size);
1409 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1412 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1413 int rc, flags = FMODE_WRITE;
1416 /* Bug 1152: copy properly when this is no longer true */
1417 LASSERT(sizeof(lum) == sizeof(*lump));
1418 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1419 rc = copy_from_user(&lum, lump, sizeof(lum));
1423 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1425 put_user(0, &lump->lmm_stripe_count);
1426 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
1427 0, ll_i2info(inode)->lli_smd, lump);
1432 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1434 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1439 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
1443 static int ll_get_grouplock(struct inode *inode, struct file *file,
1446 struct ll_file_data *fd = file->private_data;
1447 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1448 .end = OBD_OBJECT_EOF}};
1449 struct lustre_handle lockh = { 0 };
1450 struct ll_inode_info *lli = ll_i2info(inode);
1451 struct lov_stripe_md *lsm = lli->lli_smd;
1455 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1459 policy.l_extent.gid = arg;
1460 if (file->f_flags & O_NONBLOCK)
1461 flags = LDLM_FL_BLOCK_NOWAIT;
1463 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags,
1464 &ll_i2sbi(inode)->ll_grouplock_stime);
1468 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1470 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1475 static int ll_put_grouplock(struct inode *inode, struct file *file,
1478 struct ll_file_data *fd = file->private_data;
1479 struct ll_inode_info *lli = ll_i2info(inode);
1480 struct lov_stripe_md *lsm = lli->lli_smd;
1484 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1485 /* Ugh, it's already unlocked. */
1489 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1492 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1494 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1499 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1504 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1507 struct ll_file_data *fd = file->private_data;
1508 struct ll_sb_info *sbi = ll_i2sbi(inode);
1512 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1513 inode->i_generation, inode, cmd);
1515 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1518 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1520 case LL_IOC_GETFLAGS:
1521 /* Get the current value of the file flags */
1522 return put_user(fd->fd_flags, (int *)arg);
1523 case LL_IOC_SETFLAGS:
1524 case LL_IOC_CLRFLAGS:
1525 /* Set or clear specific file flags */
1526 /* XXX This probably needs checks to ensure the flags are
1527 * not abused, and to handle any flag side effects.
1529 if (get_user(flags, (int *) arg))
1532 if (cmd == LL_IOC_SETFLAGS)
1533 fd->fd_flags |= flags;
1535 fd->fd_flags &= ~flags;
1537 case LL_IOC_LOV_SETSTRIPE:
1538 RETURN(ll_lov_setstripe(inode, file, arg));
1539 case LL_IOC_LOV_SETEA:
1540 RETURN(ll_lov_setea(inode, file, arg));
1541 case IOC_MDC_SHOWFID: {
1542 struct lustre_id *idp = (struct lustre_id *)arg;
1543 struct lustre_id id;
1547 filename = getname((const char *)arg);
1548 if (IS_ERR(filename))
1549 RETURN(PTR_ERR(filename));
1551 ll_inode2id(&id, inode);
1553 rc = ll_get_fid(sbi->ll_md_exp, &id, filename, &id);
1555 GOTO(out_filename, rc);
1557 rc = copy_to_user(idp, &id, sizeof(*idp));
1559 GOTO(out_filename, rc = -EFAULT);
1566 case LL_IOC_KEY_TYPE: {
1567 struct obd_ioctl_data *data;
1570 int typelen, rc, len = 0;
1572 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1577 type = data->ioc_inlbuf1;
1578 typelen = data->ioc_inllen1;
1581 CDEBUG(D_INFO, "LL_IOC_KEY_TYPE missing filename\n");
1582 GOTO(out, rc = -EINVAL);
1584 ll_set_sb_gksinfo(inode->i_sb, type);
1587 obd_ioctl_freedata(buf, len);
1591 case LL_IOC_LOV_GETSTRIPE:
1592 RETURN(ll_lov_getstripe(inode, arg));
1593 case EXT3_IOC_GETFLAGS:
1594 case EXT3_IOC_SETFLAGS:
1595 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1596 case LL_IOC_GROUP_LOCK:
1597 RETURN(ll_get_grouplock(inode, file, arg));
1598 case LL_IOC_GROUP_UNLOCK:
1599 RETURN(ll_put_grouplock(inode, file, arg));
1600 case EXT3_IOC_GETVERSION_OLD:
1601 case EXT3_IOC_GETVERSION:
1602 return put_user(inode->i_generation, (int *) arg);
1603 /* We need to special case any other ioctls we want to handle,
1604 * to send them to the MDS/OST as appropriate and to properly
1605 * network encode the arg field.
1606 case EXT2_IOC_GETVERSION_OLD:
1607 case EXT2_IOC_GETVERSION_NEW:
1608 case EXT2_IOC_SETVERSION_OLD:
1609 case EXT2_IOC_SETVERSION_NEW:
1610 case EXT3_IOC_SETVERSION_OLD:
1611 case EXT3_IOC_SETVERSION:
1613 case LL_IOC_FLUSH_CRED:
1614 RETURN(ll_flush_cred(inode));
1616 RETURN(ll_set_audit(inode, arg));
1618 RETURN( obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
1623 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1625 struct inode *inode = file->f_dentry->d_inode;
1626 struct ll_file_data *fd = file->private_data;
1627 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1628 struct lustre_handle lockh = {0};
1631 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1632 inode->i_generation, inode,
1633 offset + ((origin==2) ? inode->i_size : file->f_pos));
1635 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1636 if (origin == 2) { /* SEEK_END */
1637 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1638 struct ll_inode_info *lli = ll_i2info(inode);
1639 int nonblock = 0, rc;
1641 if (file->f_flags & O_NONBLOCK)
1642 nonblock = LDLM_FL_BLOCK_NOWAIT;
1644 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1645 nonblock, &ll_i2sbi(inode)->ll_seek_stime);
1649 down(&lli->lli_size_sem);
1650 offset += inode->i_size;
1651 up(&lli->lli_size_sem);
1652 } else if (origin == 1) { /* SEEK_CUR */
1653 offset += file->f_pos;
1657 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1658 if (offset != file->f_pos) {
1659 file->f_pos = offset;
1660 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1662 file->f_version = ++event;
1669 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1673 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1675 struct inode *inode = dentry->d_inode;
1676 struct ll_inode_info *lli = ll_i2info(inode);
1677 struct address_space *mapping = inode->i_mapping;
1678 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1679 struct lustre_id id;
1680 struct ptlrpc_request *req;
1683 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1684 inode->i_generation, inode);
1686 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1688 /* fsync's caller has already called _fdata{sync,write}, we want
1689 * that IO to finish before calling the osc and mdc sync methods */
1690 rc = filemap_fdatawait(inode->i_mapping);
1692 /* 2.6 implements filemap_fdatawait() using PG_writeback which we
1693 * don't support now. so, wait until all submited llaps are gone */
1694 wait_event(lli->lli_dirty_wait, !ll_is_inode_dirty(mapping->host));
1696 ll_inode2id(&id, inode);
1697 err = md_sync(ll_i2sbi(inode)->ll_md_exp, &id, &req);
1701 ptlrpc_req_finished(req);
1704 struct obdo *oa = obdo_alloc();
1707 RETURN(rc ? rc : -ENOMEM);
1709 oa->o_id = lsm->lsm_object_id;
1710 oa->o_gr = lsm->lsm_object_gr;
1711 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1713 obdo_from_inode(oa, inode, (OBD_MD_FLTYPE | OBD_MD_FLATIME |
1714 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1717 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
1727 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1729 struct inode *inode = file->f_dentry->d_inode;
1730 struct ll_inode_info *li = ll_i2info(inode);
1731 struct ll_sb_info *sbi = ll_i2sbi(inode);
1732 struct obd_device *obddev;
1733 struct ldlm_res_id res_id =
1734 { .name = {id_fid(&li->lli_id), id_group(&li->lli_id), LDLM_FLOCK} };
1735 struct lustre_handle lockh = {0};
1736 ldlm_policy_data_t flock;
1737 ldlm_mode_t mode = 0;
1742 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1743 inode->i_ino, file_lock);
1745 flock.l_flock.pid = file_lock->fl_pid;
1746 flock.l_flock.start = file_lock->fl_start;
1747 flock.l_flock.end = file_lock->fl_end;
1749 switch (file_lock->fl_type) {
1754 /* An unlock request may or may not have any relation to
1755 * existing locks so we may not be able to pass a lock handle
1756 * via a normal ldlm_lock_cancel() request. The request may even
1757 * unlock a byte range in the middle of an existing lock. In
1758 * order to process an unlock request we need all of the same
1759 * information that is given with a normal read or write record
1760 * lock request. To avoid creating another ldlm unlock (cancel)
1761 * message we'll treat a LCK_NL flock request as an unlock. */
1768 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1783 flags = LDLM_FL_BLOCK_NOWAIT;
1789 flags = LDLM_FL_TEST_LOCK;
1790 /* Save the old mode so that if the mode in the lock changes we
1791 * can decrement the appropriate reader or writer refcount. */
1792 file_lock->fl_type = mode;
1795 CERROR("unknown fcntl lock command: %d\n", cmd);
1799 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1800 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1801 flags, mode, flock.l_flock.start, flock.l_flock.end);
1803 obddev = md_get_real_obd(sbi->ll_md_exp, &li->lli_id);
1804 rc = ldlm_cli_enqueue(obddev->obd_self_export, NULL,
1805 obddev->obd_namespace,
1806 res_id, LDLM_FLOCK, &flock, mode, &flags,
1807 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1808 NULL, 0, NULL, &lockh);
1812 int ll_inode_revalidate_it(struct dentry *dentry)
1814 struct lookup_intent oit = { .it_op = IT_GETATTR };
1815 struct inode *inode = dentry->d_inode;
1816 struct ptlrpc_request *req = NULL;
1817 struct ll_inode_info *lli;
1818 struct lov_stripe_md *lsm;
1819 struct ll_sb_info *sbi;
1820 struct lustre_id id;
1825 CERROR("REPORT THIS LINE TO PETER\n");
1829 sbi = ll_i2sbi(inode);
1831 ll_inode2id(&id, inode);
1832 lli = ll_i2info(inode);
1833 LASSERT(id_fid(&id) != 0);
1835 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), name=%s(%p)\n",
1836 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1839 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1840 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1843 rc = ll_intent_alloc(&oit);
1847 rc = ll_crypto_init_it_key(inode, &oit);
1851 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1852 &oit, 0, &req, ll_mdc_blocking_ast);
1856 rc = revalidate_it_finish(req, 1, &oit, dentry);
1861 ll_lookup_finish_locks(&oit, dentry);
1864 if (!req && (oit.it_op & IT_GETATTR))
1865 ll_audit_log(inode, AUDIT_STAT, 0);
1867 if (!LLI_HAVE_FLSIZE(inode)) {
1868 /* if object not yet allocated, don't validate size */
1871 /* ll_glimpse_size() will prefer locally cached
1872 * writes if they extend the file */
1873 rc = ll_glimpse_size(inode);
1878 ll_intent_release(&oit);
1880 ptlrpc_req_finished(req);
1885 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1886 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
1889 struct inode *inode = de->d_inode;
1890 struct ll_inode_info *lli = ll_i2info(inode);
1892 res = ll_inode_revalidate_it(de);
1893 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1898 stat->ino = inode->i_ino;
1899 stat->mode = inode->i_mode;
1900 stat->nlink = inode->i_nlink;
1901 stat->uid = inode->i_uid;
1902 stat->gid = inode->i_gid;
1903 stat->atime = inode->i_atime;
1904 stat->mtime = inode->i_mtime;
1905 stat->ctime = inode->i_ctime;
1906 stat->blksize = inode->i_blksize;
1908 down(&lli->lli_size_sem);
1909 stat->size = inode->i_size;
1910 stat->blocks = inode->i_blocks;
1911 up(&lli->lli_size_sem);
1913 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1914 stat->dev = id_group(&ll_i2info(inode)->lli_id);
1920 int ll_setxattr_internal(struct inode *inode, const char *name,
1921 const void *value, size_t size, int flags,
1924 struct ll_sb_info *sbi = ll_i2sbi(inode);
1925 struct ptlrpc_request *request = NULL;
1926 struct mdc_op_data *op_data;
1929 int rc = 0, key_size = 0;
1932 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu\n", inode->i_ino);
1933 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_SETXATTR);
1935 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
1936 RETURN(-EOPNOTSUPP);
1938 memset(&attr, 0x0, sizeof(attr));
1939 attr.ia_valid |= valid;
1940 attr.ia_attr_flags = flags;
1942 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1943 rc = ll_crypto_get_mac(inode, &attr, (void *)value, size,
1946 CERROR("can not get right mac, rc=%d\n", rc);
1951 OBD_ALLOC(op_data, sizeof(*op_data));
1955 ll_inode2mdc_data(op_data, inode, (OBD_MD_FLID | OBD_MD_MEA));
1957 rc = md_setattr(sbi->ll_md_exp, op_data, &attr,
1958 (void *)name, strnlen(name, XATTR_NAME_MAX) + 1,
1959 (void *)value, size, key, key_size, &request);
1960 OBD_FREE(op_data, sizeof(*op_data));
1963 CDEBUG(D_SEC, "md_setattr fails: rc = %d\n", rc);
1967 if (key && key_size)
1968 OBD_FREE(key, key_size);
1969 ptlrpc_req_finished(request);
1973 int ll_setxattr(struct dentry *dentry, const char *name, const void *value,
1974 size_t size, int flags)
1977 struct posix_acl *acl;
1978 struct ll_inode_info *lli;
1981 rc = ll_setxattr_internal(dentry->d_inode, name, value, size,
1984 /* update inode's acl info */
1985 if (rc == 0 && strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1987 acl = posix_acl_from_xattr(value, size);
1989 CERROR("convert from xattr to acl error: %ld",
1993 error = posix_acl_valid(acl);
1995 CERROR("acl valid error: %d", error);
1996 posix_acl_release(acl);
2004 lli = ll_i2info(dentry->d_inode);
2005 spin_lock(&lli->lli_lock);
2006 if (lli->lli_posix_acl != NULL)
2007 posix_acl_release(lli->lli_posix_acl);
2008 lli->lli_posix_acl = acl;
2009 spin_unlock(&lli->lli_lock);
2016 int ll_removexattr(struct dentry *dentry, const char *name)
2018 return ll_setxattr_internal(dentry->d_inode, name, NULL, 0, 0,
2022 int ll_getxattr_internal(struct inode *inode, const char *name,
2023 void *value, size_t size, __u64 valid)
2025 struct ptlrpc_request *request = NULL;
2026 struct ll_sb_info *sbi = ll_i2sbi(inode);
2027 struct lustre_id id;
2028 struct mds_body *body;
2033 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETXATTR);
2035 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
2036 RETURN(-EOPNOTSUPP);
2038 ll_inode2id(&id, inode);
2039 rc = md_getattr(sbi->ll_md_exp, &id, valid, name, NULL, 0,
2040 size, NULL, &request);
2042 if (rc != -ENODATA && rc != -EOPNOTSUPP)
2043 CERROR("rc = %d\n", rc);
2047 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
2048 LASSERT(body != NULL);
2049 LASSERT_REPSWABBED(request, 0);
2051 ea_size = body->eadatasize;
2053 GOTO(out, rc = ea_size);
2055 LASSERT(ea_size <= request->rq_repmsg->buflens[1]);
2056 ea_data = lustre_msg_buf(request->rq_repmsg, 1, ea_size);
2057 LASSERT(ea_data != NULL);
2058 LASSERT_REPSWABBED(request, 1);
2061 memcpy(value, ea_data, ea_size);
2065 ptlrpc_req_finished(request);
2069 int ll_getxattr(struct dentry *dentry, const char *name, void *value,
2072 return ll_getxattr_internal(dentry->d_inode, name,
2073 value, size, OBD_MD_FLXATTR);
2076 int ll_listxattr(struct dentry *dentry, char *list, size_t size)
2078 return ll_getxattr_internal(dentry->d_inode, NULL, list, size,
2079 OBD_MD_FLXATTRLIST);
2083 * Here we hold DLM lock across permission check, to get better
2084 * conformance especially for remote acl. will introduce certain
2085 * overhead, maybe need a profile.
2088 lustre_check_acl(struct inode *inode, int mask)
2090 struct lookup_intent it = { .it_op = IT_GETATTR };
2091 struct dentry de = { .d_inode = inode };
2092 struct ll_sb_info *sbi;
2093 struct lustre_id id;
2094 struct ptlrpc_request *req = NULL;
2095 struct ll_inode_info *lli = ll_i2info(inode);
2096 struct posix_acl *acl;
2100 sbi = ll_i2sbi(inode);
2101 ll_inode2id(&id, inode);
2103 if (ll_intent_alloc(&it))
2106 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
2107 &it, 0, &req, ll_mdc_blocking_ast);
2109 ll_intent_free(&it);
2113 rc = revalidate_it_finish(req, 1, &it, &de);
2115 ll_intent_release(&it);
2119 if (sbi->ll_remote) {
2120 rc = ll_remote_acl_permission(inode, mask);
2122 spin_lock(&lli->lli_lock);
2123 acl = posix_acl_dup(ll_i2info(inode)->lli_posix_acl);
2124 spin_unlock(&lli->lli_lock);
2129 rc = posix_acl_permission(inode, acl, mask);
2130 posix_acl_release(acl);
2134 ll_lookup_finish_locks(&it, &de);
2135 ll_intent_free(&it);
2139 ptlrpc_req_finished(req);
2144 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
2146 struct ll_sb_info *sbi = ll_i2sbi(inode);
2148 /* for remote client, permission bits in inode doesn't
2149 * play a role anymore.
2152 return lustre_check_acl(inode, mask);
2154 return generic_permission(inode, mask, lustre_check_acl);
2157 struct file_operations ll_file_operations = {
2158 .read = ll_file_read,
2159 .write = ll_file_write,
2160 .ioctl = ll_file_ioctl,
2161 .open = ll_file_open,
2162 .release = ll_file_release,
2163 .mmap = ll_file_mmap,
2164 .llseek = ll_file_seek,
2165 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
2166 .sendfile = generic_file_sendfile,
2169 .lock = ll_file_flock
2172 struct inode_operations ll_file_inode_operations = {
2173 .setattr = ll_setattr,
2174 .truncate = ll_truncate,
2175 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
2176 .getattr = ll_getattr,
2178 .revalidate_it = ll_inode_revalidate_it,
2180 .setxattr = ll_setxattr,
2181 .getxattr = ll_getxattr,
2182 .listxattr = ll_listxattr,
2183 .removexattr = ll_removexattr,
2184 .permission = ll_inode_permission,