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,
47 ldlm_policy_data_t extent = { .l_extent = { 0, OBD_OBJECT_EOF } };
48 struct obd_export *exp = ll_i2sbi(inode)->ll_dt_exp;
49 struct ll_inode_info *lli = ll_i2info(inode);
50 struct lustre_handle match_lockh = {0};
54 if (lli->lli_smd == NULL)
57 LASSERT(size != NULL && blocks != NULL);
59 flags = LDLM_FL_TEST_LOCK | LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
60 rc = obd_match(exp, lli->lli_smd, LDLM_EXTENT, &extent,
61 LCK_PR | LCK_PW, &flags, inode, &match_lockh);
63 /* we have no all needed locks,
64 * so we don't know actual size */
68 /* we know actual size! */
69 down(&lli->lli_size_sem);
70 *size = lov_merge_size(lli->lli_smd, 0);
71 *blocks = lov_merge_blocks(lli->lli_smd);
72 *mtime = lov_merge_mtime(lli->lli_smd, LTIME_S(inode->i_mtime));
74 up(&lli->lli_size_sem);
80 int ll_md_och_close(struct obd_export *md_exp, struct inode *inode,
81 struct obd_client_handle *och, int dirty,
84 struct ptlrpc_request *req = NULL;
85 struct mdc_op_data *op_data;
86 struct obd_device *obd;
91 obd = class_exp2obd(md_exp);
93 CERROR("Invalid MDC connection handle "LPX64"\n",
94 md_exp->exp_handle.h_cookie);
98 /* here we check if this is forced umount. If so this is called on
99 * canceling "open lock" and we do not call md_close() in this case , as
100 * it will not successful, as import is already deactivated. */
101 if (obd->obd_no_recov)
104 /* prepare @op_data for close request */
105 OBD_ALLOC(op_data, sizeof(*op_data));
109 memset(op_data, 0, sizeof(*op_data));
111 valid = (OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
112 OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLEPOCH |
115 ll_inode2mdc_data(op_data, inode, valid);
117 if (0 /* ll_is_inode_dirty(inode) */) {
118 op_data->flags = MDS_BFLAG_UNCOMMITTED_WRITES;
119 op_data->valid |= OBD_MD_FLFLAGS;
123 /* we modified data through this handle */
124 op_data->io_epoch = epoch;
125 op_data->flags |= MDS_BFLAG_DIRTY_EPOCH;
126 op_data->valid |= OBD_MD_FLFLAGS | OBD_MD_FLEPOCH;
127 if (ll_validate_size(inode, &op_data->size, &op_data->blocks,
129 op_data->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLMTIME;
132 rc = md_close(md_exp, op_data, och, &req);
133 OBD_FREE(op_data, sizeof(*op_data));
136 /* we are the last writer, so the MDS has instructed us to get
137 * the file size and any write cookies, then close again. */
139 //ll_queue_done_writing(inode);
142 CERROR("inode %lu mdc close failed: rc = %d\n",
143 (unsigned long)inode->i_ino, rc);
146 ptlrpc_req_finished(req);
149 mdc_clear_open_replay_data(md_exp, och);
150 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
151 OBD_FREE(och, sizeof *och);
155 int ll_md_real_close(struct obd_export *md_exp,
156 struct inode *inode, int flags)
158 struct ll_inode_info *lli = ll_i2info(inode);
159 int freeing = inode->i_state & I_FREEING;
160 struct obd_client_handle **och_p;
161 struct obd_client_handle *och;
162 __u64 *och_usecount, epoch = 0;
163 int rc = 0, dirty = 0;
166 if (flags & FMODE_WRITE) {
167 och_p = &lli->lli_mds_write_och;
168 och_usecount = &lli->lli_open_fd_write_count;
169 } else if (flags & FMODE_EXEC) {
170 och_p = &lli->lli_mds_exec_och;
171 och_usecount = &lli->lli_open_fd_exec_count;
173 och_p = &lli->lli_mds_read_och;
174 och_usecount = &lli->lli_open_fd_read_count;
177 down(&lli->lli_och_sem);
178 if (*och_usecount) { /* There are still users of this handle, so
180 up(&lli->lli_och_sem);
183 if (ll_is_inode_dirty(inode)) {
184 /* the inode still has dirty pages, let's close later */
185 CDEBUG(D_INODE, "inode %lu/%u still has dirty pages\n",
186 inode->i_ino, inode->i_generation);
187 LASSERT(freeing == 0);
188 ll_queue_done_writing(inode);
189 up(&lli->lli_och_sem);
193 if (LLI_DIRTY_HANDLE(inode) && (flags & FMODE_WRITE)) {
194 clear_bit(LLI_F_DIRTY_HANDLE, &lli->lli_flags);
196 } else if (0 && !(flags & FMODE_SYNC) && !freeing) {
197 /* in order to speed up creation rate we pass
198 * closing to dedicated thread so we don't need
199 * to wait for close reply here -bzzz */
200 ll_queue_done_writing(inode);
201 up(&lli->lli_och_sem);
207 if (flags & FMODE_WRITE) {
208 epoch = lli->lli_io_epoch;
209 lli->lli_io_epoch = 0;
212 up(&lli->lli_och_sem);
215 * there might be a race and somebody have freed this och
216 * already. Another way to have this twice called is if file closing
217 * will fail due to netwok problems and on umount lock will be canceled
218 * and this will be called from block_ast callack.
220 if (och && och->och_fh.cookie != DEAD_HANDLE_MAGIC)
221 rc = ll_md_och_close(md_exp, inode, och, dirty, epoch);
226 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
229 struct ll_file_data *fd = file->private_data;
230 struct ll_inode_info *lli = ll_i2info(inode);
234 /* clear group lock, if present */
235 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
236 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
237 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
238 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
242 /* Let's see if we have good enough OPEN lock on the file and if
243 we can skip talking to MDS */
244 if (file->f_dentry->d_inode) {
246 struct obd_device *obddev;
247 struct lustre_handle lockh;
248 int flags = LDLM_FL_BLOCK_GRANTED;
249 struct ldlm_res_id file_res_id = {.name = {id_fid(&lli->lli_id),
250 id_group(&lli->lli_id)}};
251 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
253 down(&lli->lli_och_sem);
254 if (fd->fd_omode & FMODE_WRITE) {
256 LASSERT(lli->lli_open_fd_write_count);
257 lli->lli_open_fd_write_count--;
258 } else if (fd->fd_omode & FMODE_EXEC) {
260 LASSERT(lli->lli_open_fd_exec_count);
261 lli->lli_open_fd_exec_count--;
264 LASSERT(lli->lli_open_fd_read_count);
265 lli->lli_open_fd_read_count--;
267 up(&lli->lli_och_sem);
269 obddev = md_get_real_obd(md_exp, &lli->lli_id);
270 if (!ldlm_lock_match(obddev->obd_namespace, flags, &file_res_id,
271 LDLM_IBITS, &policy, lockmode, &lockh))
273 rc = ll_md_real_close(md_exp, file->f_dentry->d_inode,
276 ldlm_lock_decref(&lockh, lockmode);
280 file->private_data = NULL;
281 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof(*fd));
285 /* While this returns an error code, fput() the caller does not, so we need
286 * to make every effort to clean up all of our state here. Also, applications
287 * rarely check close errors and even if an error is returned they will not
288 * re-try the close call.
290 int ll_file_release(struct inode *inode, struct file *file)
292 struct ll_file_data *fd;
293 struct ll_sb_info *sbi = ll_i2sbi(inode);
297 CDEBUG(D_VFSTRACE, "VFS Op:inode="DLID4"(%p)\n",
298 OLID4(&ll_i2info(inode)->lli_id), inode);
300 /* don't do anything for / */
301 if (inode->i_sb->s_root == file->f_dentry)
304 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
305 fd = (struct ll_file_data *)file->private_data;
308 rc = ll_md_close(sbi->ll_md_exp, inode, file);
312 static int ll_intent_file_open(struct file *file, void *lmm,
313 int lmmsize, struct lookup_intent *itp)
315 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
316 const char *name = (char *)file->f_dentry->d_name.name;
317 struct dentry *parent = file->f_dentry->d_parent;
318 const int len = file->f_dentry->d_name.len;
319 struct lustre_handle lockh;
320 struct mdc_op_data *op_data;
326 OBD_ALLOC(op_data, sizeof(*op_data));
330 ll_prepare_mdc_data(op_data, parent->d_inode, NULL,
333 rc = md_enqueue(sbi->ll_md_exp, LDLM_IBITS, itp, LCK_PR, op_data,
334 &lockh, lmm, lmmsize, ldlm_completion_ast,
335 ll_mdc_blocking_ast, NULL);
336 OBD_FREE(op_data, sizeof(*op_data));
338 if (LUSTRE_IT(itp)->it_lock_mode)
339 memcpy(&LUSTRE_IT(itp)->it_lock_handle,
340 &lockh, sizeof(lockh));
343 CERROR("lock enqueue: err: %d\n", rc);
348 int ll_och_fill(struct inode *inode, struct lookup_intent *it,
349 struct obd_client_handle *och)
351 struct ptlrpc_request *req = LUSTRE_IT(it)->it_data;
352 struct ll_inode_info *lli = ll_i2info(inode);
353 struct mds_body *body;
358 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
359 LASSERT (body != NULL); /* reply already checked out */
360 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
362 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
363 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
364 if (it->it_flags & FMODE_WRITE) {
365 if (lli->lli_io_epoch && lli->lli_io_epoch != body->io_epoch)
366 CDEBUG(D_ERROR, "new epoch?! "LPD64" != "LPD64"\n",
367 lli->lli_io_epoch, body->io_epoch);
368 lli->lli_io_epoch = body->io_epoch;
370 mdc_set_open_replay_data(ll_i2mdexp(inode), och,
371 LUSTRE_IT(it)->it_data);
376 int ll_local_open(struct file *file, struct lookup_intent *it,
377 struct obd_client_handle *och)
379 struct ll_file_data *fd;
380 struct inode *inode = file->f_dentry->d_inode;
385 rc = ll_och_fill(inode, it, och);
390 LASSERTF(file->private_data == NULL, "file %.*s/%.*s ino %lu/%u (%o)\n",
391 file->f_dentry->d_name.len, file->f_dentry->d_name.name,
392 file->f_dentry->d_parent->d_name.len,
393 file->f_dentry->d_parent->d_name.name,
394 file->f_dentry->d_inode->i_ino,
395 file->f_dentry->d_inode->i_generation,
396 file->f_dentry->d_inode->i_mode);
398 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
400 /* We can't handle this well without reorganizing ll_file_open and
401 * ll_md_close(), so don't even try right now. */
404 file->private_data = fd;
405 ll_readahead_init(inode, &fd->fd_ras);
406 fd->fd_omode = it->it_flags;
411 /* Open a file, and (for the very first open) create objects on the OSTs at
412 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
413 * creation or open until ll_lov_setstripe() ioctl is called. We grab
414 * lli_open_sem to ensure no other process will create objects, send the
415 * stripe MD to the MDS, or try to destroy the objects if that fails.
417 * If we already have the stripe MD locally then we don't request it in
418 * mdc_open(), by passing a lmm_size = 0.
420 * It is up to the application to ensure no other processes open this file
421 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
422 * used. We might be able to avoid races of that sort by getting lli_open_sem
423 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
424 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
426 int ll_file_open(struct inode *inode, struct file *file)
428 struct ll_inode_info *lli = ll_i2info(inode);
429 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
430 .it_flags = file->f_flags };
431 struct lov_stripe_md *lsm;
432 struct ptlrpc_request *req;
434 struct obd_client_handle **och_p = NULL;
435 __u64 *och_usecount = NULL;
438 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n",
439 inode->i_ino, inode->i_generation, inode, file->f_flags);
441 /* don't do anything for / */
442 if (inode->i_sb->s_root == file->f_dentry)
445 if ((file->f_flags+1) & O_ACCMODE)
447 if (file->f_flags & O_TRUNC)
452 /* sometimes LUSTRE_IT(it) may not be allocated like opening file by
453 * dentry_open() from GNS stuff. */
454 if (!it || !LUSTRE_IT(it)) {
456 rc = ll_intent_alloc(it);
461 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
463 /* mdc_intent_lock() didn't get a request ref if there was an open
464 * error, so don't do cleanup on the * request here (bug 3430) */
465 if (LUSTRE_IT(it)->it_disposition) {
466 rc = it_open_error(DISP_OPEN_OPEN, it);
471 /* Let's see if we have file open on MDS already. */
472 if (it->it_flags & FMODE_WRITE) {
473 och_p = &lli->lli_mds_write_och;
474 och_usecount = &lli->lli_open_fd_write_count;
475 } else if (it->it_flags & FMODE_EXEC) {
476 och_p = &lli->lli_mds_exec_och;
477 och_usecount = &lli->lli_open_fd_exec_count;
479 och_p = &lli->lli_mds_read_och;
480 och_usecount = &lli->lli_open_fd_read_count;
483 rc = ll_crypto_decrypt_key(inode, it);
487 down(&lli->lli_och_sem);
488 if (*och_p) { /* Open handle is present */
489 if (it_disposition(it, DISP_LOOKUP_POS) && /* Positive lookup */
490 it_disposition(it, DISP_OPEN_OPEN)) { /* & OPEN happened */
491 struct obd_client_handle *och;
492 /* Well, there's extra open request that we do not need,
493 let's close it somehow*/
494 OBD_ALLOC(och, sizeof (struct obd_client_handle));
496 up(&lli->lli_och_sem);
500 rc = ll_och_fill(inode, it, och);
502 up(&lli->lli_och_sem);
506 /* ll_md_och_close() will free och */
507 ll_md_och_close(ll_i2mdexp(inode), inode, och, 0, 0);
511 rc = ll_local_open(file, it, NULL);
515 rc = ll_set_capa(inode, it, *och_p);
517 up(&lli->lli_och_sem);
521 LASSERT(*och_usecount == 0);
522 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
524 GOTO(out, rc = -ENOMEM);
527 if (!it || !LUSTRE_IT(it) || !LUSTRE_IT(it)->it_disposition) {
529 * we are going to replace intent here, and that may
530 * possibly change access mode (FMODE_EXEC can only be
531 * set in intent), but I hope it never happens (I was
532 * not able to trigger it yet at least) -- green
535 /* FIXME: FMODE_EXEC is not covered by O_ACCMODE! */
536 LASSERT(!(it->it_flags & FMODE_EXEC));
537 LASSERTF((it->it_flags & O_ACCMODE) ==
538 (oit.it_flags & O_ACCMODE), "Changing intent "
539 "flags %x to incompatible %x\n", it->it_flags,
542 rc = ll_intent_file_open(file, NULL, 0, it);
545 rc = it_open_error(DISP_OPEN_OPEN, it);
547 GOTO(out_och_free, rc);
549 mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle,
550 file->f_dentry->d_inode);
552 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
553 rc = ll_local_open(file, it, *och_p);
554 LASSERTF(rc == 0, "rc = %d\n", rc);
556 rc = ll_set_capa(inode, it, *och_p);
558 up(&lli->lli_och_sem);
562 up(&lli->lli_och_sem);
565 * must do this outside lli_och_sem lock to prevent deadlock where
566 * different kind of OPEN lock for this same inode gets cancelled by
569 if (!S_ISREG(inode->i_mode))
574 if (file->f_flags & O_LOV_DELAY_CREATE ||
575 !(file->f_mode & FMODE_WRITE)) {
576 CDEBUG(D_INODE, "object creation was delayed\n");
580 file->f_flags &= ~O_LOV_DELAY_CREATE;
583 /* audit stuff if there was no RPC */
584 if (LUSTRE_IT(it)->it_data == 0)
585 ll_audit_log(inode, AUDIT_OPEN, rc);
587 req = LUSTRE_IT(it)->it_data;
588 ll_intent_drop_lock(it);
589 ll_intent_release(it);
592 ptlrpc_req_finished(req);
594 ll_open_complete(inode);
598 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
599 *och_p = NULL; /* OBD_FREE writes some magic there */
602 up(&lli->lli_och_sem);
608 /* Fills the obdo with the attributes for the inode defined by lsm */
609 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
612 struct ptlrpc_request_set *set;
616 LASSERT(lsm != NULL);
618 memset(oa, 0, sizeof *oa);
619 oa->o_id = lsm->lsm_object_id;
620 oa->o_gr = lsm->lsm_object_gr;
621 oa->o_mode = S_IFREG;
622 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
623 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
624 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
626 set = ptlrpc_prep_set();
630 rc = obd_getattr_async(exp, oa, lsm, set);
632 rc = ptlrpc_set_wait(set);
633 ptlrpc_set_destroy(set);
638 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
639 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
643 static inline void ll_remove_suid(struct inode *inode)
647 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
648 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
650 /* was any of the uid bits set? */
651 mode &= inode->i_mode;
652 if (mode && !capable(CAP_FSETID)) {
653 inode->i_mode &= ~mode;
654 // XXX careful here - we cannot change the size
658 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
660 struct ll_inode_info *lli = ll_i2info(inode);
661 struct lov_stripe_md *lsm = lli->lli_smd;
662 struct obd_export *exp = ll_i2dtexp(inode);
665 struct ldlm_lock *lock;
666 struct lov_stripe_md *lsm;
667 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
668 __u32 stripe, vallen = sizeof(stripe);
672 if (lsm->lsm_stripe_count == 1)
673 GOTO(check, stripe = 0);
675 /* get our offset in the lov */
676 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
678 CERROR("obd_get_info: rc = %d\n", rc);
681 LASSERT(stripe < lsm->lsm_stripe_count);
684 if (lsm->lsm_oinfo[stripe].loi_id != lock->l_resource->lr_name.name[0]||
685 lsm->lsm_oinfo[stripe].loi_gr != lock->l_resource->lr_name.name[2]){
686 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64
687 " inode=%lu/%u (%p)\n",
688 lsm->lsm_oinfo[stripe].loi_id,
689 lsm->lsm_oinfo[stripe].loi_gr,
690 inode->i_ino, inode->i_generation, inode);
691 return -ELDLM_NO_LOCK_DATA;
697 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
698 * we get a lock cancellation for each stripe, so we have to map the obd's
699 * region back onto the stripes in the file that it held.
701 * No one can dirty the extent until we've finished our work and they can
702 * enqueue another lock. The DLM protects us from ll_file_read/write here,
703 * but other kernel actors could have pages locked.
705 * Called with the DLM lock held. */
706 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
707 struct ldlm_lock *lock, __u32 stripe)
709 ldlm_policy_data_t tmpex;
710 unsigned long start, end, count, skip, i, j;
712 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
713 struct lustre_handle lockh;
716 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
717 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
718 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
721 /* our locks are page granular thanks to osc_enqueue, we invalidate the
723 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
724 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
728 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
729 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
730 if (lsm->lsm_stripe_count > 1) {
731 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
732 skip = (lsm->lsm_stripe_count - 1) * count;
733 start += start/count * skip + stripe * count;
735 end += end/count * skip + stripe * count;
737 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
740 i = inode->i_size ? (inode->i_size - 1) >> PAGE_CACHE_SHIFT : 0;
744 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
745 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
746 count, skip, end, discard ? " (DISCARDING)" : "");
748 /* walk through the vmas on the inode and tear down mmaped pages that
749 * intersect with the lock. this stops immediately if there are no
750 * mmap()ed regions of the file. This is not efficient at all and
751 * should be short lived. We'll associate mmap()ed pages with the lock
752 * and will be able to find them directly */
754 for (i = start; i <= end; i += (j + skip)) {
755 j = min(count - (i % count), end - i + 1);
757 LASSERT(inode->i_mapping);
758 if (ll_teardown_mmaps(inode->i_mapping,
759 (__u64)i << PAGE_CACHE_SHIFT,
760 ((__u64)(i+j) << PAGE_CACHE_SHIFT) - 1) )
764 /* this is the simplistic implementation of page eviction at
765 * cancelation. It is careful to get races with other page
766 * lockers handled correctly. fixes from bug 20 will make it
767 * more efficient by associating locks with pages and with
768 * batching writeback under the lock explicitly. */
769 for (i = start, j = start % count; i <= end;
770 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
772 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
778 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
779 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
780 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
783 if (!mapping_has_pages(inode->i_mapping)) {
784 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
790 page = find_get_page(inode->i_mapping, i);
793 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
794 i, tmpex.l_extent.start);
797 /* page->mapping to check with racing against teardown */
798 if (!discard && clear_page_dirty_for_io(page)) {
799 rc = ll_call_writepage(inode, page);
801 CERROR("writepage of page %p failed: %d\n",
803 /* either waiting for io to complete or reacquiring
804 * the lock that the failed writepage released */
808 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
809 /* check to see if another DLM lock covers this page */
810 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
811 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
813 &lock->l_resource->lr_name, LDLM_EXTENT,
814 &tmpex, LCK_PR | LCK_PW, &lockh);
815 if (rc2 == 0 && page->mapping != NULL) {
816 // checking again to account for writeback's lock_page()
817 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
818 ll_ra_accounting(page, inode->i_mapping);
819 ll_truncate_complete_page(page);
822 page_cache_release(page);
824 LASSERTF(tmpex.l_extent.start <=
825 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
826 lock->l_policy_data.l_extent.end + 1),
827 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
828 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
833 static int ll_extent_lock_callback(struct ldlm_lock *lock,
834 struct ldlm_lock_desc *new, void *data,
837 struct lustre_handle lockh = { 0 };
841 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
842 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
847 case LDLM_CB_BLOCKING:
848 ldlm_lock2handle(lock, &lockh);
849 rc = ldlm_cli_cancel(&lockh);
851 CERROR("ldlm_cli_cancel failed: %d\n", rc);
853 case LDLM_CB_CANCELING: {
855 struct ll_inode_info *lli;
856 struct lov_stripe_md *lsm;
860 /* This lock wasn't granted, don't try to evict pages */
861 if (lock->l_req_mode != lock->l_granted_mode)
864 inode = ll_inode_from_lock(lock);
867 lli = ll_i2info(inode);
870 if (lli->lli_smd == NULL)
874 stripe = ll_lock_to_stripe_offset(inode, lock);
876 CERROR("ll_lock_to_stripe_offset failed: %d\n", stripe);
880 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
882 down(&lli->lli_size_sem);
883 lock_res_and_lock(lock);
884 kms = ldlm_extent_shift_kms(lock,
885 lsm->lsm_oinfo[stripe].loi_kms);
887 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
888 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
889 lsm->lsm_oinfo[stripe].loi_kms, kms);
890 lsm->lsm_oinfo[stripe].loi_kms = kms;
891 unlock_res_and_lock(lock);
892 up(&lli->lli_size_sem);
893 //ll_try_done_writing(inode);
906 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
908 /* XXX ALLOCATE - 160 bytes */
909 struct inode *inode = ll_inode_from_lock(lock);
910 struct ll_inode_info *lli = ll_i2info(inode);
911 struct lustre_handle lockh = { 0 };
916 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
917 LDLM_FL_BLOCK_CONV)) {
918 LBUG(); /* not expecting any blocked async locks yet */
919 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
921 ldlm_lock_dump(D_OTHER, lock, 0);
922 ldlm_reprocess_all(lock->l_resource);
926 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
928 stripe = ll_lock_to_stripe_offset(inode, lock);
930 CERROR("ll_lock_to_stripe_offset failed: %d\n", stripe);
934 if (lock->l_lvb_len) {
935 struct lov_stripe_md *lsm = lli->lli_smd;
937 lvb = lock->l_lvb_data;
938 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
941 lock_res_and_lock(lock);
942 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
943 kms = ldlm_extent_shift_kms(NULL, kms);
944 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
945 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
946 lsm->lsm_oinfo[stripe].loi_kms, kms);
947 lsm->lsm_oinfo[stripe].loi_kms = kms;
948 unlock_res_and_lock(lock);
954 wake_up(&lock->l_waitq);
956 ldlm_lock2handle(lock, &lockh);
957 ldlm_lock_decref(&lockh, LCK_PR);
962 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
964 struct ptlrpc_request *req = reqp;
965 struct inode *inode = ll_inode_from_lock(lock);
966 struct ll_inode_info *lli;
968 struct lov_stripe_md *lsm;
969 int rc, size = sizeof(*lvb), stripe;
973 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
974 lli = ll_i2info(inode);
976 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
980 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
982 /* First, find out which stripe index this lock corresponds to. */
983 stripe = ll_lock_to_stripe_offset(inode, lock);
985 CERROR("ll_lock_to_stripe_offset failed: %d\n", stripe);
986 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
989 rc = lustre_pack_reply(req, 1, &size, NULL);
991 CERROR("lustre_pack_reply: %d\n", rc);
995 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
996 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
997 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
998 lvb->lvb_atime = LTIME_S(inode->i_atime);
999 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
1001 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
1002 inode->i_size, stripe, lvb->lvb_size);
1008 /* These errors are normal races, so we don't want to fill the console
1009 * with messages by calling ptlrpc_error() */
1010 if (rc == -ELDLM_NO_LOCK_DATA)
1011 lustre_pack_reply(req, 0, NULL, NULL);
1013 req->rq_status = rc;
1017 /* NB: lov_merge_size will prefer locally cached writes if they extend the
1018 * file (because it prefers KMS over RSS when larger) */
1019 int ll_glimpse_size(struct inode *inode)
1021 struct ll_inode_info *lli = ll_i2info(inode);
1022 struct ll_sb_info *sbi = ll_i2sbi(inode);
1023 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1024 struct lustre_handle lockh = { 0 };
1025 int rc, flags = LDLM_FL_HAS_INTENT;
1028 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1030 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
1031 LCK_PR, &flags, ll_extent_lock_callback,
1032 ldlm_completion_ast, ll_glimpse_callback, inode,
1033 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1038 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1039 RETURN(rc > 0 ? -EIO : rc);
1042 down(&lli->lli_size_sem);
1043 inode->i_size = lov_merge_size(lli->lli_smd, 0);
1044 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
1045 LTIME_S(inode->i_mtime) = lov_merge_mtime(lli->lli_smd,
1046 LTIME_S(inode->i_mtime));
1047 up(&lli->lli_size_sem);
1049 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
1050 (__u64)inode->i_size, (__u64)inode->i_blocks);
1052 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
1056 void ll_stime_record(struct ll_sb_info *sbi, struct timeval *start,
1057 struct obd_service_time *stime)
1059 struct timeval stop;
1060 do_gettimeofday(&stop);
1062 spin_lock(&sbi->ll_lock);
1063 lprocfs_stime_record(stime, &stop, start);
1064 spin_unlock(&sbi->ll_lock);
1067 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
1068 struct lov_stripe_md *lsm, int mode,
1069 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
1070 int ast_flags, struct obd_service_time *stime)
1072 struct ll_inode_info *lli = ll_i2info(inode);
1073 struct ll_sb_info *sbi = ll_i2sbi(inode);
1074 struct timeval start;
1078 LASSERT(lockh->cookie == 0);
1080 /* XXX phil: can we do this? won't it screw the file size up? */
1081 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1082 (sbi->ll_flags & LL_SBI_NOLCK))
1085 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
1086 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
1088 do_gettimeofday(&start);
1089 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
1090 &ast_flags, ll_extent_lock_callback,
1091 ldlm_completion_ast, ll_glimpse_callback, inode,
1092 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
1096 ll_stime_record(sbi, &start, stime);
1098 if (policy->l_extent.start == 0 &&
1099 policy->l_extent.end == OBD_OBJECT_EOF) {
1100 /* vmtruncate()->ll_truncate() first sets the i_size and then
1101 * the kms under both a DLM lock and the i_sem. If we don't
1102 * get the i_sem here we can match the DLM lock and reset
1103 * i_size from the kms before the truncating path has updated
1104 * the kms. generic_file_write can then trust the stale i_size
1105 * when doing appending writes and effectively cancel the
1106 * result of the truncate. Getting the i_sem after the enqueue
1107 * maintains the DLM -> i_sem acquiry order. */
1108 down(&lli->lli_size_sem);
1109 inode->i_size = lov_merge_size(lsm, 1);
1110 up(&lli->lli_size_sem);
1114 LTIME_S(inode->i_mtime) =
1115 lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1121 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1122 struct lov_stripe_md *lsm, int mode,
1123 struct lustre_handle *lockh)
1125 struct ll_sb_info *sbi = ll_i2sbi(inode);
1129 /* XXX phil: can we do this? won't it screw the file size up? */
1130 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1131 (sbi->ll_flags & LL_SBI_NOLCK))
1134 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1139 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1142 struct inode *inode = file->f_dentry->d_inode;
1143 struct ll_inode_info *lli = ll_i2info(inode);
1144 struct lov_stripe_md *lsm = lli->lli_smd;
1145 struct ll_lock_tree tree;
1146 struct ll_lock_tree_node *node;
1151 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1152 inode->i_ino, inode->i_generation, inode, count, *ppos);
1154 /* "If nbyte is 0, read() will return 0 and have no other results."
1155 * -- Single Unix Spec */
1159 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
1165 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1168 tree.lt_fd = file->private_data;
1170 rc = ll_tree_lock(&tree, node, inode, buf, count,
1171 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1175 down(&lli->lli_size_sem);
1176 kms = lov_merge_size(lsm, 1);
1177 if (*ppos + count - 1 > kms) {
1178 /* A glimpse is necessary to determine whether we return a short
1179 * read or some zeroes at the end of the buffer */
1180 up(&lli->lli_size_sem);
1181 retval = ll_glimpse_size(inode);
1185 inode->i_size = kms;
1186 up(&lli->lli_size_sem);
1189 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1190 inode->i_ino, count, *ppos, inode->i_size);
1192 /* turn off the kernel's read-ahead */
1193 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1196 file->f_ra.ra_pages = 0;
1198 retval = generic_file_read(file, buf, count, ppos);
1199 ll_audit_log(inode, AUDIT_READ, retval);
1202 ll_tree_unlock(&tree, inode);
1207 * Write to a file (through the page cache).
1209 static ssize_t ll_file_write(struct file *file, const char *buf,
1210 size_t count, loff_t *ppos)
1212 struct inode *inode = file->f_dentry->d_inode;
1213 loff_t maxbytes = ll_file_maxbytes(inode);
1214 struct ll_lock_tree tree;
1215 struct ll_lock_tree_node *node;
1220 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1221 inode->i_ino, inode->i_generation, inode, count, *ppos);
1223 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1225 /* POSIX, but surprised the VFS doesn't check this already */
1229 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1230 * called on the file, don't fail the below assertion (bug 2388). */
1231 if (file->f_flags & O_LOV_DELAY_CREATE &&
1232 ll_i2info(inode)->lli_smd == NULL)
1235 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1237 if (file->f_flags & O_APPEND)
1238 node = ll_node_from_inode(inode, 0, OBD_OBJECT_EOF, LCK_PW);
1240 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1244 RETURN(PTR_ERR(node));
1246 tree.lt_fd = file->private_data;
1248 rc = ll_tree_lock(&tree, node, inode, buf, count,
1249 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1253 /* this is ok, g_f_w will overwrite this under i_sem if it races
1254 * with a local truncate, it just makes our maxbyte checking easier */
1255 if (file->f_flags & O_APPEND)
1256 *ppos = inode->i_size;
1258 if (*ppos >= maxbytes) {
1259 if (count || *ppos > maxbytes) {
1260 send_sig(SIGXFSZ, current, 0);
1261 GOTO(out, retval = -EFBIG);
1264 if (*ppos + count > maxbytes)
1265 count = maxbytes - *ppos;
1267 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1268 inode->i_ino, count, *ppos);
1270 /* mark open handle dirty */
1271 set_bit(LLI_F_DIRTY_HANDLE, &(ll_i2info(inode)->lli_flags));
1273 /* generic_file_write handles O_APPEND after getting i_sem */
1274 retval = generic_file_write(file, buf, count, ppos);
1278 /* ll_audit_log(inode, AUDIT_WRITE, retval); */
1280 ll_tree_unlock(&tree, inode);
1281 /* serialize with mmap/munmap/mremap */
1282 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
1283 retval > 0 ? retval : 0);
1287 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
1288 int flags, struct lov_user_md *lum,
1291 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
1292 struct ll_inode_info *lli = ll_i2info(inode);
1293 struct ptlrpc_request *req = NULL;
1294 struct obd_client_handle *och;
1295 struct lov_stripe_md *lsm;
1296 struct lustre_md md;
1302 if ((file->f_flags+1) & O_ACCMODE)
1304 if (file->f_flags & O_TRUNC)
1307 down(&lli->lli_open_sem);
1310 up(&lli->lli_open_sem);
1311 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
1316 f = get_empty_filp();
1320 f->f_dentry = file->f_dentry;
1321 f->f_vfsmnt = file->f_vfsmnt;
1324 rc = ll_intent_alloc(&oit);
1328 rc = ll_intent_file_open(f, lum, lum_size, &oit);
1331 if (it_disposition(&oit, DISP_LOOKUP_NEG))
1334 req = LUSTRE_IT(&oit)->it_data;
1335 rc = LUSTRE_IT(&oit)->it_status;
1340 rc = mdc_req2lustre_md(ll_i2mdexp(inode), req, 1,
1341 ll_i2dtexp(inode), &md);
1344 ll_update_inode(f->f_dentry->d_inode, &md);
1346 OBD_ALLOC(och, sizeof(struct obd_client_handle));
1348 GOTO(out, rc = -ENOMEM);
1350 /* actually ll_local_open() cannot fail! */
1351 rc = ll_local_open(f, &oit, och);
1355 if (LUSTRE_IT(&oit)->it_lock_mode) {
1356 ldlm_lock_decref_and_cancel((struct lustre_handle *)
1357 &LUSTRE_IT(&oit)->it_lock_handle,
1358 LUSTRE_IT(&oit)->it_lock_mode);
1359 LUSTRE_IT(&oit)->it_lock_mode = 0;
1362 ll_intent_release(&oit);
1364 /* ll_file_release() will decrease the count, but won't free anything
1365 * because we have at least one more reference coming from actual
1367 down(&lli->lli_och_sem);
1368 lli->lli_open_fd_write_count++;
1369 up(&lli->lli_och_sem);
1370 rc = ll_file_release(f->f_dentry->d_inode, f);
1372 /* Now also destroy our supplemental och */
1373 ll_md_och_close(ll_i2mdexp(inode), f->f_dentry->d_inode, och, 0, 0);
1376 ll_intent_release(&oit);
1379 up(&lli->lli_open_sem);
1381 ptlrpc_req_finished(req);
1385 static int ll_lov_setea(struct inode *inode, struct file *file,
1388 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
1389 int lum_size = sizeof(struct lov_user_md) +
1390 sizeof(struct lov_user_ost_data);
1391 struct lov_user_md *lump;
1395 if (!capable(CAP_SYS_ADMIN))
1398 OBD_ALLOC(lump, lum_size);
1402 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
1404 GOTO(out_free_lump, rc = -EFAULT);
1406 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1409 OBD_FREE(lump, lum_size);
1413 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1416 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1417 int rc, flags = FMODE_WRITE;
1420 /* Bug 1152: copy properly when this is no longer true */
1421 LASSERT(sizeof(lum) == sizeof(*lump));
1422 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1423 rc = copy_from_user(&lum, lump, sizeof(lum));
1427 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1429 put_user(0, &lump->lmm_stripe_count);
1430 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
1431 0, ll_i2info(inode)->lli_smd, lump);
1436 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1438 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1443 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
1447 static int ll_get_grouplock(struct inode *inode, struct file *file,
1450 struct ll_file_data *fd = file->private_data;
1451 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1452 .end = OBD_OBJECT_EOF}};
1453 struct lustre_handle lockh = { 0 };
1454 struct ll_inode_info *lli = ll_i2info(inode);
1455 struct lov_stripe_md *lsm = lli->lli_smd;
1459 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1463 policy.l_extent.gid = arg;
1464 if (file->f_flags & O_NONBLOCK)
1465 flags = LDLM_FL_BLOCK_NOWAIT;
1467 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags,
1468 &ll_i2sbi(inode)->ll_grouplock_stime);
1472 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1474 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1479 static int ll_put_grouplock(struct inode *inode, struct file *file,
1482 struct ll_file_data *fd = file->private_data;
1483 struct ll_inode_info *lli = ll_i2info(inode);
1484 struct lov_stripe_md *lsm = lli->lli_smd;
1488 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1489 /* Ugh, it's already unlocked. */
1493 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1496 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1498 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1503 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1508 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1511 struct ll_file_data *fd = file->private_data;
1512 struct ll_sb_info *sbi = ll_i2sbi(inode);
1516 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1517 inode->i_generation, inode, cmd);
1519 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1522 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1524 case LL_IOC_GETFLAGS:
1525 /* Get the current value of the file flags */
1526 return put_user(fd->fd_flags, (int *)arg);
1527 case LL_IOC_SETFLAGS:
1528 case LL_IOC_CLRFLAGS:
1529 /* Set or clear specific file flags */
1530 /* XXX This probably needs checks to ensure the flags are
1531 * not abused, and to handle any flag side effects.
1533 if (get_user(flags, (int *) arg))
1536 if (cmd == LL_IOC_SETFLAGS)
1537 fd->fd_flags |= flags;
1539 fd->fd_flags &= ~flags;
1541 case LL_IOC_LOV_SETSTRIPE:
1542 RETURN(ll_lov_setstripe(inode, file, arg));
1543 case LL_IOC_LOV_SETEA:
1544 RETURN(ll_lov_setea(inode, file, arg));
1545 case IOC_MDC_SHOWFID: {
1546 struct lustre_id *idp = (struct lustre_id *)arg;
1547 struct lustre_id id;
1551 filename = getname((const char *)arg);
1552 if (IS_ERR(filename))
1553 RETURN(PTR_ERR(filename));
1555 ll_inode2id(&id, inode);
1557 rc = ll_get_fid(sbi->ll_md_exp, &id, filename, &id);
1559 GOTO(out_filename, rc);
1561 rc = copy_to_user(idp, &id, sizeof(*idp));
1563 GOTO(out_filename, rc = -EFAULT);
1570 case LL_IOC_KEY_TYPE: {
1571 struct obd_ioctl_data *data;
1574 int typelen, rc, len = 0;
1576 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1581 type = data->ioc_inlbuf1;
1582 typelen = data->ioc_inllen1;
1585 CDEBUG(D_INFO, "LL_IOC_KEY_TYPE missing filename\n");
1586 GOTO(out, rc = -EINVAL);
1588 ll_set_sb_gksinfo(inode->i_sb, type);
1591 obd_ioctl_freedata(buf, len);
1595 case LL_IOC_LOV_GETSTRIPE:
1596 RETURN(ll_lov_getstripe(inode, arg));
1597 case EXT3_IOC_GETFLAGS:
1598 case EXT3_IOC_SETFLAGS:
1599 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1600 case LL_IOC_GROUP_LOCK:
1601 RETURN(ll_get_grouplock(inode, file, arg));
1602 case LL_IOC_GROUP_UNLOCK:
1603 RETURN(ll_put_grouplock(inode, file, arg));
1604 case EXT3_IOC_GETVERSION_OLD:
1605 case EXT3_IOC_GETVERSION:
1606 return put_user(inode->i_generation, (int *) arg);
1607 /* We need to special case any other ioctls we want to handle,
1608 * to send them to the MDS/OST as appropriate and to properly
1609 * network encode the arg field.
1610 case EXT2_IOC_GETVERSION_OLD:
1611 case EXT2_IOC_GETVERSION_NEW:
1612 case EXT2_IOC_SETVERSION_OLD:
1613 case EXT2_IOC_SETVERSION_NEW:
1614 case EXT3_IOC_SETVERSION_OLD:
1615 case EXT3_IOC_SETVERSION:
1617 case LL_IOC_FLUSH_CRED:
1618 RETURN(ll_flush_cred(inode));
1620 RETURN(ll_set_audit(inode, arg));
1622 RETURN( obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
1627 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1629 struct inode *inode = file->f_dentry->d_inode;
1630 struct ll_file_data *fd = file->private_data;
1631 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1632 struct lustre_handle lockh = {0};
1635 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1636 inode->i_generation, inode,
1637 offset + ((origin==2) ? inode->i_size : file->f_pos));
1639 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1640 if (origin == 2) { /* SEEK_END */
1641 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1642 struct ll_inode_info *lli = ll_i2info(inode);
1643 int nonblock = 0, rc;
1645 if (file->f_flags & O_NONBLOCK)
1646 nonblock = LDLM_FL_BLOCK_NOWAIT;
1648 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1649 nonblock, &ll_i2sbi(inode)->ll_seek_stime);
1653 down(&lli->lli_size_sem);
1654 offset += inode->i_size;
1655 up(&lli->lli_size_sem);
1656 } else if (origin == 1) { /* SEEK_CUR */
1657 offset += file->f_pos;
1661 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1662 if (offset != file->f_pos) {
1663 file->f_pos = offset;
1664 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1666 file->f_version = ++event;
1673 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1677 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1679 struct inode *inode = dentry->d_inode;
1680 struct ll_inode_info *lli = ll_i2info(inode);
1681 struct address_space *mapping = inode->i_mapping;
1682 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1683 struct lustre_id id;
1684 struct ptlrpc_request *req;
1687 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1688 inode->i_generation, inode);
1690 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1692 /* fsync's caller has already called _fdata{sync,write}, we want
1693 * that IO to finish before calling the osc and mdc sync methods */
1694 rc = filemap_fdatawait(inode->i_mapping);
1696 /* 2.6 implements filemap_fdatawait() using PG_writeback which we
1697 * don't support now. so, wait until all submited llaps are gone */
1698 wait_event(lli->lli_dirty_wait, !ll_is_inode_dirty(mapping->host));
1700 ll_inode2id(&id, inode);
1701 err = md_sync(ll_i2sbi(inode)->ll_md_exp, &id, &req);
1705 ptlrpc_req_finished(req);
1708 struct obdo *oa = obdo_alloc();
1711 RETURN(rc ? rc : -ENOMEM);
1713 oa->o_id = lsm->lsm_object_id;
1714 oa->o_gr = lsm->lsm_object_gr;
1715 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1717 obdo_from_inode(oa, inode, (OBD_MD_FLTYPE | OBD_MD_FLATIME |
1718 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1721 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
1731 int ll_file_flock(struct file *file, int cmd, struct file_lock *fl)
1733 struct inode *inode = file->f_dentry->d_inode;
1734 struct ll_inode_info *li = ll_i2info(inode);
1735 struct ll_sb_info *sbi = ll_i2sbi(inode);
1736 struct obd_device *obd = md_get_real_obd(sbi->ll_md_exp, &li->lli_id);
1737 struct ldlm_res_id res_id = { .name = {id_fid(&li->lli_id),
1738 id_group(&li->lli_id), LDLM_FLOCK} };
1739 struct lustre_handle lockh = {0};
1740 struct ptlrpc_connection *conn;
1741 ptl_process_id_t ptlpid;
1742 ldlm_policy_data_t flock;
1743 ldlm_mode_t mode = 0;
1748 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu cmd=%d file_lock=%p\n",
1749 inode->i_ino, cmd, fl);
1751 if (!(fl->fl_flags & FL_POSIX))
1754 switch (fl->fl_type) {
1760 /* An unlock request may or may not have any relation to
1761 * existing locks so we may not be able to pass a lock handle
1762 * via a normal ldlm_lock_cancel() request. The request may even
1763 * unlock a byte range in the middle of an existing lock. In
1764 * order to process an unlock request we need all of the same
1765 * information that is given with a normal read or write record
1766 * lock request. To avoid creating another ldlm unlock (cancel)
1767 * message we'll treat a LCK_NL flock request as an unlock. */
1774 CERROR("unknown fcntl lock type: %d\n", fl->fl_type);
1783 flags |= LDLM_FL_BLOCK_NOWAIT;
1794 flags = LDLM_FL_TEST_LOCK;
1795 /* Save the old mode so that if the mode in the lock changes we
1796 * can decrement the appropriate reader or writer refcount. */
1800 CERROR("unknown fcntl lock command: %d\n", cmd);
1804 /* Since we're called on every close to remove any oustanding Posix
1805 * flocks owned by the process it's worth a little effort to avoid
1806 * the RPCs if there are no flocks on this file from this node. */
1807 if (mode == LCK_NL && fl->fl_start == 0 && fl->fl_end >= OFFSET_MAX) {
1808 struct ldlm_resource *res;
1810 res = ldlm_resource_get(obd->obd_namespace, NULL,
1811 res_id, LDLM_FLOCK, 0);
1815 ldlm_resource_putref(res);
1818 conn = class_exp2cliimp(obd->obd_self_export)->imp_connection;
1819 if (!conn || !conn->c_peer.peer_ni)
1822 rc = PtlGetId(conn->c_peer.peer_ni->pni_ni_h, &ptlpid);
1826 flock.l_flock.start = fl->fl_start;
1827 flock.l_flock.end = fl->fl_end;
1828 /* XXX - ptlpid.pid is currently coming back a constant; i.e. 12345. */
1829 flock.l_flock.pid = fl->fl_pid;
1830 flock.l_flock.nid = ptlpid.nid;
1831 flock.l_flock.blocking_pid = 0;
1832 flock.l_flock.blocking_nid = 0;
1834 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1835 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1836 flags, mode, flock.l_flock.start, flock.l_flock.end);
1838 rc = ldlm_cli_enqueue(obd->obd_self_export, NULL, obd->obd_namespace,
1839 res_id, LDLM_FLOCK, &flock, mode, &flags,
1840 NULL, ldlm_flock_completion_ast, NULL,
1841 md_get_real_obd(sbi->ll_md_exp, &sbi->ll_rootid),
1842 NULL, 0, NULL, &lockh);
1844 if (flags & LDLM_FL_TEST_LOCK) {
1845 struct ldlm_lock *lock = ldlm_handle2lock(&lockh);
1847 fl->fl_start = lock->l_policy_data.l_flock.start;
1848 fl->fl_end = lock->l_policy_data.l_flock.end;
1849 fl->fl_pid = lock->l_policy_data.l_flock.pid;
1851 switch (lock->l_granted_mode) {
1853 fl->fl_type = F_RDLCK;
1856 fl->fl_type = F_WRLCK;
1859 fl->fl_type = F_UNLCK;
1862 CERROR("unexpected lock type: %d returned from server."
1863 "\n", lock->l_granted_mode);
1868 /* offset the addref() done by ldlm_handle2lock() above. */
1869 LDLM_LOCK_PUT(lock);
1870 ldlm_lock_decref(&lockh, mode);
1871 ldlm_cli_cancel(&lockh);
1873 /* the LDLM_CBPENDING flag was set in the lock by the
1874 * completion AST so the ldlm_lock_decref() call above
1875 * scheduled a blocking AST which will do the final
1876 * lock put on the lock. */
1882 int ll_inode_revalidate_it(struct dentry *dentry)
1884 struct lookup_intent oit = { .it_op = IT_GETATTR };
1885 struct inode *inode = dentry->d_inode;
1886 struct ptlrpc_request *req = NULL;
1887 struct ll_inode_info *lli;
1888 struct lov_stripe_md *lsm;
1889 struct ll_sb_info *sbi;
1890 struct lustre_id id;
1895 CERROR("REPORT THIS LINE TO PETER\n");
1899 sbi = ll_i2sbi(inode);
1901 ll_inode2id(&id, inode);
1902 lli = ll_i2info(inode);
1903 LASSERT(id_fid(&id) != 0);
1905 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), name=%s(%p)\n",
1906 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1909 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1910 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1913 rc = ll_intent_alloc(&oit);
1917 rc = ll_crypto_init_it_key(inode, &oit);
1921 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1922 &oit, 0, &req, ll_mdc_blocking_ast);
1926 rc = revalidate_it_finish(req, 1, &oit, dentry);
1931 ll_lookup_finish_locks(&oit, dentry);
1934 if (!req && (oit.it_op & IT_GETATTR))
1935 ll_audit_log(inode, AUDIT_STAT, 0);
1938 if (!LLI_HAVE_FLSIZE(inode)) {
1939 /* if object not yet allocated, don't validate size */
1942 /* ll_glimpse_size() will prefer locally cached
1943 * writes if they extend the file */
1944 rc = ll_glimpse_size(inode);
1950 ll_intent_release(&oit);
1952 ptlrpc_req_finished(req);
1957 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1958 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
1961 struct inode *inode = de->d_inode;
1962 struct ll_inode_info *lli = ll_i2info(inode);
1964 res = ll_inode_revalidate_it(de);
1965 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1970 stat->ino = inode->i_ino;
1971 stat->mode = inode->i_mode;
1972 stat->nlink = inode->i_nlink;
1973 stat->uid = inode->i_uid;
1974 stat->gid = inode->i_gid;
1975 stat->atime = inode->i_atime;
1976 stat->mtime = inode->i_mtime;
1977 stat->ctime = inode->i_ctime;
1978 stat->blksize = inode->i_blksize;
1980 down(&lli->lli_size_sem);
1981 stat->size = inode->i_size;
1982 stat->blocks = inode->i_blocks;
1983 up(&lli->lli_size_sem);
1985 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1986 stat->dev = id_group(&ll_i2info(inode)->lli_id);
1992 int ll_setxattr_internal(struct inode *inode, const char *name,
1993 const void *value, size_t size, int flags,
1996 struct ll_sb_info *sbi = ll_i2sbi(inode);
1997 struct ptlrpc_request *request = NULL;
1998 struct mdc_op_data *op_data;
2001 int rc = 0, key_size = 0;
2004 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu\n", inode->i_ino);
2005 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_SETXATTR);
2007 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
2008 RETURN(-EOPNOTSUPP);
2010 memset(&attr, 0x0, sizeof(attr));
2011 attr.ia_valid |= valid;
2012 attr.ia_attr_flags = flags;
2014 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
2015 rc = ll_crypto_get_mac(inode, &attr, (void *)value, size,
2018 CERROR("can not get right mac, rc=%d\n", rc);
2023 OBD_ALLOC(op_data, sizeof(*op_data));
2027 ll_inode2mdc_data(op_data, inode, (OBD_MD_FLID | OBD_MD_MEA));
2029 rc = md_setattr(sbi->ll_md_exp, op_data, &attr,
2030 (void *)name, strnlen(name, XATTR_NAME_MAX) + 1,
2031 (void *)value, size, key, key_size, &request);
2032 OBD_FREE(op_data, sizeof(*op_data));
2035 CDEBUG(D_SEC, "md_setattr fails: rc = %d\n", rc);
2039 if (key && key_size)
2040 OBD_FREE(key, key_size);
2041 ptlrpc_req_finished(request);
2045 int ll_setxattr(struct dentry *dentry, const char *name, const void *value,
2046 size_t size, int flags)
2049 struct posix_acl *acl;
2050 struct ll_inode_info *lli;
2053 rc = ll_setxattr_internal(dentry->d_inode, name, value, size,
2056 /* update inode's acl info */
2057 if (rc == 0 && strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
2059 acl = posix_acl_from_xattr(value, size);
2061 CERROR("convert from xattr to acl error: %ld",
2065 error = posix_acl_valid(acl);
2067 CERROR("acl valid error: %d", error);
2068 posix_acl_release(acl);
2076 lli = ll_i2info(dentry->d_inode);
2077 spin_lock(&lli->lli_lock);
2078 if (lli->lli_posix_acl != NULL)
2079 posix_acl_release(lli->lli_posix_acl);
2080 lli->lli_posix_acl = acl;
2081 spin_unlock(&lli->lli_lock);
2088 int ll_removexattr(struct dentry *dentry, const char *name)
2090 return ll_setxattr_internal(dentry->d_inode, name, NULL, 0, 0,
2094 int ll_getxattr_internal(struct inode *inode, const char *name,
2095 void *value, size_t size, __u64 valid)
2097 struct ptlrpc_request *request = NULL;
2098 struct ll_sb_info *sbi = ll_i2sbi(inode);
2099 struct lustre_id id;
2100 struct mds_body *body;
2105 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETXATTR);
2107 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
2108 RETURN(-EOPNOTSUPP);
2110 ll_inode2id(&id, inode);
2111 rc = md_getattr(sbi->ll_md_exp, &id, valid, name, NULL, 0,
2112 size, NULL, &request);
2114 if (rc != -ENODATA && rc != -EOPNOTSUPP)
2115 CERROR("rc = %d\n", rc);
2119 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
2120 LASSERT(body != NULL);
2121 LASSERT_REPSWABBED(request, 0);
2123 ea_size = body->eadatasize;
2125 GOTO(out, rc = ea_size);
2127 LASSERT(ea_size <= request->rq_repmsg->buflens[1]);
2128 ea_data = lustre_msg_buf(request->rq_repmsg, 1, ea_size);
2129 LASSERT(ea_data != NULL);
2130 LASSERT_REPSWABBED(request, 1);
2133 memcpy(value, ea_data, ea_size);
2137 ptlrpc_req_finished(request);
2141 int ll_getxattr(struct dentry *dentry, const char *name, void *value,
2144 return ll_getxattr_internal(dentry->d_inode, name,
2145 value, size, OBD_MD_FLXATTR);
2148 int ll_listxattr(struct dentry *dentry, char *list, size_t size)
2150 return ll_getxattr_internal(dentry->d_inode, NULL, list, size,
2151 OBD_MD_FLXATTRLIST);
2155 * Here we hold DLM lock across permission check, to get better
2156 * conformance especially for remote acl. will introduce certain
2157 * overhead, maybe need a profile.
2160 lustre_check_acl(struct inode *inode, int mask)
2162 struct lookup_intent it = { .it_op = IT_GETATTR };
2163 struct dentry de = { .d_inode = inode };
2164 struct ll_sb_info *sbi;
2165 struct lustre_id id;
2166 struct ptlrpc_request *req = NULL;
2167 struct ll_inode_info *lli = ll_i2info(inode);
2168 struct posix_acl *acl;
2172 sbi = ll_i2sbi(inode);
2173 ll_inode2id(&id, inode);
2175 if (ll_intent_alloc(&it))
2178 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
2179 &it, 0, &req, ll_mdc_blocking_ast);
2181 ll_intent_free(&it);
2185 rc = revalidate_it_finish(req, 1, &it, &de);
2187 ll_intent_release(&it);
2191 if (sbi->ll_remote) {
2192 rc = ll_remote_acl_permission(inode, mask);
2194 spin_lock(&lli->lli_lock);
2195 acl = posix_acl_dup(ll_i2info(inode)->lli_posix_acl);
2196 spin_unlock(&lli->lli_lock);
2201 rc = posix_acl_permission(inode, acl, mask);
2202 posix_acl_release(acl);
2206 ll_lookup_finish_locks(&it, &de);
2207 ll_intent_free(&it);
2211 ptlrpc_req_finished(req);
2216 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
2218 struct ll_sb_info *sbi = ll_i2sbi(inode);
2220 /* for remote client, permission bits in inode doesn't
2221 * play a role anymore.
2224 return lustre_check_acl(inode, mask);
2226 return generic_permission(inode, mask, lustre_check_acl);
2229 struct file_operations ll_file_operations = {
2230 .read = ll_file_read,
2231 .write = ll_file_write,
2232 .ioctl = ll_file_ioctl,
2233 .open = ll_file_open,
2234 .release = ll_file_release,
2235 .mmap = ll_file_mmap,
2236 .llseek = ll_file_seek,
2237 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
2238 .sendfile = generic_file_sendfile,
2241 .lock = ll_file_flock
2244 struct inode_operations ll_file_inode_operations = {
2245 .setattr = ll_setattr,
2246 .truncate = ll_truncate,
2247 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
2248 .getattr = ll_getattr,
2250 .revalidate_it = ll_inode_revalidate_it,
2252 .setxattr = ll_setxattr,
2253 .getxattr = ll_getxattr,
2254 .listxattr = ll_listxattr,
2255 .removexattr = ll_removexattr,
2256 .permission = ll_inode_permission,