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)
79 struct ptlrpc_request *req = NULL;
80 struct mdc_op_data *op_data;
81 struct obd_device *obd;
86 obd = class_exp2obd(md_exp);
88 CERROR("Invalid MDC connection handle "LPX64"\n",
89 md_exp->exp_handle.h_cookie);
93 /* here we check if this is forced umount. If so this is called on
94 * canceling "open lock" and we do not call md_close() in this case , as
95 * it will not successful, as import is already deactivated. */
96 if (obd->obd_no_recov)
99 /* prepare @op_data for close request */
100 OBD_ALLOC(op_data, sizeof(*op_data));
104 memset(op_data, 0, sizeof(*op_data));
106 valid = (OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
107 OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLEPOCH |
110 ll_inode2mdc_data(op_data, inode, valid);
112 if (0 /* ll_is_inode_dirty(inode) */) {
113 op_data->flags = MDS_BFLAG_UNCOMMITTED_WRITES;
114 op_data->valid |= OBD_MD_FLFLAGS;
118 /* we modified data through this handle */
119 op_data->flags |= MDS_BFLAG_DIRTY_EPOCH;
120 op_data->valid |= OBD_MD_FLFLAGS;
121 if (ll_validate_size(inode, &op_data->size, &op_data->blocks))
122 op_data->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
125 rc = md_close(md_exp, op_data, och, &req);
126 OBD_FREE(op_data, sizeof(*op_data));
129 /* we are the last writer, so the MDS has instructed us to get
130 * the file size and any write cookies, then close again. */
132 //ll_queue_done_writing(inode);
135 CERROR("inode %lu mdc close failed: rc = %d\n",
136 (unsigned long)inode->i_ino, rc);
139 ptlrpc_req_finished(req);
142 mdc_clear_open_replay_data(md_exp, och);
144 capa_put(och->och_capa, CLIENT_CAPA);
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;
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);
204 up(&lli->lli_och_sem);
207 * there might be a race and somebody have freed this och
208 * already. Another way to have this twice called is if file closing
209 * will fail due to netwok problems and on umount lock will be canceled
210 * and this will be called from block_ast callack.
212 if (och && och->och_fh.cookie != DEAD_HANDLE_MAGIC)
213 rc = ll_md_och_close(md_exp, inode, och, dirty);
218 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
221 struct ll_file_data *fd = file->private_data;
222 struct ll_inode_info *lli = ll_i2info(inode);
226 /* clear group lock, if present */
227 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
228 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
229 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
230 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
234 /* Let's see if we have good enough OPEN lock on the file and if
235 we can skip talking to MDS */
236 if (file->f_dentry->d_inode) {
238 struct obd_device *obddev;
239 struct lustre_handle lockh;
240 int flags = LDLM_FL_BLOCK_GRANTED;
241 struct ldlm_res_id file_res_id = {.name = {id_fid(&lli->lli_id),
242 id_group(&lli->lli_id)}};
243 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
245 down(&lli->lli_och_sem);
246 if (fd->fd_omode & FMODE_WRITE) {
248 LASSERT(lli->lli_open_fd_write_count);
249 lli->lli_open_fd_write_count--;
250 } else if (fd->fd_omode & FMODE_EXEC) {
252 LASSERT(lli->lli_open_fd_exec_count);
253 lli->lli_open_fd_exec_count--;
256 LASSERT(lli->lli_open_fd_read_count);
257 lli->lli_open_fd_read_count--;
259 up(&lli->lli_och_sem);
261 obddev = md_get_real_obd(md_exp, &lli->lli_id);
262 if (!ldlm_lock_match(obddev->obd_namespace, flags, &file_res_id,
263 LDLM_IBITS, &policy, lockmode, &lockh))
265 rc = ll_md_real_close(md_exp, file->f_dentry->d_inode,
268 ldlm_lock_decref(&lockh, lockmode);
272 file->private_data = NULL;
273 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof(*fd));
277 /* While this returns an error code, fput() the caller does not, so we need
278 * to make every effort to clean up all of our state here. Also, applications
279 * rarely check close errors and even if an error is returned they will not
280 * re-try the close call.
282 int ll_file_release(struct inode *inode, struct file *file)
284 struct ll_file_data *fd;
285 struct ll_sb_info *sbi = ll_i2sbi(inode);
289 CDEBUG(D_VFSTRACE, "VFS Op:inode="DLID4"(%p)\n",
290 OLID4(&ll_i2info(inode)->lli_id), inode);
292 /* don't do anything for / */
293 if (inode->i_sb->s_root == file->f_dentry)
296 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
297 fd = (struct ll_file_data *)file->private_data;
300 rc = ll_md_close(sbi->ll_md_exp, inode, file);
304 static int ll_intent_file_open(struct file *file, void *lmm,
305 int lmmsize, struct lookup_intent *itp)
307 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
308 const char *name = (char *)file->f_dentry->d_name.name;
309 struct dentry *parent = file->f_dentry->d_parent;
310 const int len = file->f_dentry->d_name.len;
311 struct lustre_handle lockh;
312 struct mdc_op_data *op_data;
318 OBD_ALLOC(op_data, sizeof(*op_data));
322 ll_prepare_mdc_data(op_data, parent->d_inode, NULL,
325 rc = md_enqueue(sbi->ll_md_exp, LDLM_IBITS, itp, LCK_PR, op_data,
326 &lockh, lmm, lmmsize, ldlm_completion_ast,
327 ll_mdc_blocking_ast, NULL);
328 OBD_FREE(op_data, sizeof(*op_data));
330 if (LUSTRE_IT(itp)->it_lock_mode)
331 memcpy(&LUSTRE_IT(itp)->it_lock_handle,
332 &lockh, sizeof(lockh));
335 CERROR("lock enqueue: err: %d\n", rc);
340 int ll_och_fill(struct inode *inode, struct lookup_intent *it,
341 struct obd_client_handle *och)
343 struct ptlrpc_request *req = LUSTRE_IT(it)->it_data;
344 struct ll_inode_info *lli = ll_i2info(inode);
345 struct mds_body *body;
350 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
351 LASSERT (body != NULL); /* reply already checked out */
352 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
354 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
355 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
356 lli->lli_io_epoch = body->io_epoch;
357 mdc_set_open_replay_data(ll_i2mdexp(inode), och,
358 LUSTRE_IT(it)->it_data);
360 if (S_ISREG(inode->i_mode) && (body->valid & OBD_MD_CAPA))
361 rc = ll_set_och_capa(inode, it, och);
365 int ll_local_open(struct file *file, struct lookup_intent *it,
366 struct obd_client_handle *och)
368 struct ll_file_data *fd;
373 rc = ll_och_fill(file->f_dentry->d_inode, it, och);
378 LASSERTF(file->private_data == NULL, "file %.*s/%.*s ino %lu/%u (%o)\n",
379 file->f_dentry->d_name.len, file->f_dentry->d_name.name,
380 file->f_dentry->d_parent->d_name.len,
381 file->f_dentry->d_parent->d_name.name,
382 file->f_dentry->d_inode->i_ino,
383 file->f_dentry->d_inode->i_generation,
384 file->f_dentry->d_inode->i_mode);
386 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
388 /* We can't handle this well without reorganizing ll_file_open and
389 * ll_md_close(), so don't even try right now. */
392 file->private_data = fd;
393 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
394 fd->fd_omode = it->it_flags;
399 /* Open a file, and (for the very first open) create objects on the OSTs at
400 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
401 * creation or open until ll_lov_setstripe() ioctl is called. We grab
402 * lli_open_sem to ensure no other process will create objects, send the
403 * stripe MD to the MDS, or try to destroy the objects if that fails.
405 * If we already have the stripe MD locally then we don't request it in
406 * mdc_open(), by passing a lmm_size = 0.
408 * It is up to the application to ensure no other processes open this file
409 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
410 * used. We might be able to avoid races of that sort by getting lli_open_sem
411 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
412 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
414 int ll_file_open(struct inode *inode, struct file *file)
416 struct ll_inode_info *lli = ll_i2info(inode);
417 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
418 .it_flags = file->f_flags };
419 struct lov_stripe_md *lsm;
420 struct ptlrpc_request *req;
422 struct obd_client_handle **och_p = NULL;
423 __u64 *och_usecount = NULL;
426 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n",
427 inode->i_ino, inode->i_generation, inode, file->f_flags);
429 /* don't do anything for / */
430 if (inode->i_sb->s_root == file->f_dentry)
433 if ((file->f_flags+1) & O_ACCMODE)
435 if (file->f_flags & O_TRUNC)
440 * sometimes LUSTRE_IT(it) may not be allocated like opening file by
441 * dentry_open() from GNS stuff.
443 if (!it || !LUSTRE_IT(it)) {
445 rc = ll_intent_alloc(it);
450 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
453 * mdc_intent_lock() didn't get a request ref if there was an open
454 * error, so don't do cleanup on the * request here (bug 3430)
456 if (LUSTRE_IT(it)->it_disposition) {
457 rc = it_open_error(DISP_OPEN_OPEN, it);
462 /* Let's see if we have file open on MDS already. */
463 if (it->it_flags & FMODE_WRITE) {
464 och_p = &lli->lli_mds_write_och;
465 och_usecount = &lli->lli_open_fd_write_count;
466 } else if (it->it_flags & FMODE_EXEC) {
467 och_p = &lli->lli_mds_exec_och;
468 och_usecount = &lli->lli_open_fd_exec_count;
470 och_p = &lli->lli_mds_read_och;
471 och_usecount = &lli->lli_open_fd_read_count;
474 rc = ll_crypto_decrypt_key(inode, it);
478 down(&lli->lli_och_sem);
479 if (*och_p) { /* Open handle is present */
480 if (it_disposition(it, DISP_LOOKUP_POS) && /* Positive lookup */
481 it_disposition(it, DISP_OPEN_OPEN)) { /* & OPEN happened */
482 struct obd_client_handle *och;
483 /* Well, there's extra open request that we do not need,
484 let's close it somehow*/
485 OBD_ALLOC(och, sizeof (struct obd_client_handle));
487 up(&lli->lli_och_sem);
491 rc = ll_och_fill(inode, it, och);
493 up(&lli->lli_och_sem);
497 /* ll_md_och_close() will free och */
498 ll_md_och_close(ll_i2mdexp(inode), inode, och, 0);
502 rc = ll_local_open(file, it, NULL);
506 LASSERT(*och_usecount == 0);
507 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
509 GOTO(out, rc = -ENOMEM);
512 if (!it || !LUSTRE_IT(it) || !LUSTRE_IT(it)->it_disposition) {
514 * we are going to replace intent here, and that may
515 * possibly change access mode (FMODE_EXEC can only be
516 * set in intent), but I hope it never happens (I was
517 * not able to trigger it yet at least) -- green
520 /* FIXME: FMODE_EXEC is not covered by O_ACCMODE! */
521 LASSERT(!(it->it_flags & FMODE_EXEC));
522 LASSERTF((it->it_flags & O_ACCMODE) ==
523 (oit.it_flags & O_ACCMODE), "Changing intent "
524 "flags %x to incompatible %x\n", it->it_flags,
527 rc = ll_intent_file_open(file, NULL, 0, it);
530 rc = it_open_error(DISP_OPEN_OPEN, it);
532 GOTO(out_och_free, rc);
534 mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle,
535 file->f_dentry->d_inode);
537 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
538 rc = ll_local_open(file, it, *och_p);
539 LASSERTF(rc == 0, "rc = %d\n", rc);
541 up(&lli->lli_och_sem);
544 * must do this outside lli_och_sem lock to prevent deadlock where
545 * different kind of OPEN lock for this same inode gets cancelled by
548 if (!S_ISREG(inode->i_mode))
553 if (file->f_flags & O_LOV_DELAY_CREATE ||
554 !(file->f_mode & FMODE_WRITE)) {
555 CDEBUG(D_INODE, "object creation was delayed\n");
559 file->f_flags &= ~O_LOV_DELAY_CREATE;
562 /* audit stuff if there was no RPC */
563 if (LUSTRE_IT(it)->it_data == 0) {
564 ll_audit_log(inode, AUDIT_OPEN, rc);
567 req = LUSTRE_IT(it)->it_data;
568 ll_intent_drop_lock(it);
569 ll_intent_release(it);
570 ptlrpc_req_finished(req);
572 ll_open_complete(inode);
576 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
577 *och_p = NULL; /* OBD_FREE writes some magic there */
580 up(&lli->lli_och_sem);
586 /* Fills the obdo with the attributes for the inode defined by lsm */
587 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
590 struct ptlrpc_request_set *set;
594 LASSERT(lsm != NULL);
596 memset(oa, 0, sizeof *oa);
597 oa->o_id = lsm->lsm_object_id;
598 oa->o_gr = lsm->lsm_object_gr;
599 oa->o_mode = S_IFREG;
600 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
601 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
602 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
604 set = ptlrpc_prep_set();
608 rc = obd_getattr_async(exp, oa, lsm, set);
610 rc = ptlrpc_set_wait(set);
611 ptlrpc_set_destroy(set);
616 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
617 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
621 static inline void ll_remove_suid(struct inode *inode)
625 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
626 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
628 /* was any of the uid bits set? */
629 mode &= inode->i_mode;
630 if (mode && !capable(CAP_FSETID)) {
631 inode->i_mode &= ~mode;
632 // XXX careful here - we cannot change the size
636 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
638 struct ll_inode_info *lli = ll_i2info(inode);
639 struct lov_stripe_md *lsm = lli->lli_smd;
640 struct obd_export *exp = ll_i2dtexp(inode);
643 struct ldlm_lock *lock;
644 struct lov_stripe_md *lsm;
645 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
646 __u32 stripe, vallen = sizeof(stripe);
650 if (lsm->lsm_stripe_count == 1)
651 GOTO(check, stripe = 0);
653 /* get our offset in the lov */
654 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
656 CERROR("obd_get_info: rc = %d\n", rc);
659 LASSERT(stripe < lsm->lsm_stripe_count);
662 if (lsm->lsm_oinfo[stripe].loi_id != lock->l_resource->lr_name.name[0]||
663 lsm->lsm_oinfo[stripe].loi_gr != lock->l_resource->lr_name.name[2]){
664 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64
665 " inode=%lu/%u (%p)\n",
666 lsm->lsm_oinfo[stripe].loi_id,
667 lsm->lsm_oinfo[stripe].loi_gr,
668 inode->i_ino, inode->i_generation, inode);
669 return -ELDLM_NO_LOCK_DATA;
675 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
676 * we get a lock cancellation for each stripe, so we have to map the obd's
677 * region back onto the stripes in the file that it held.
679 * No one can dirty the extent until we've finished our work and they can
680 * enqueue another lock. The DLM protects us from ll_file_read/write here,
681 * but other kernel actors could have pages locked.
683 * Called with the DLM lock held. */
684 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
685 struct ldlm_lock *lock, __u32 stripe)
687 ldlm_policy_data_t tmpex;
688 unsigned long start, end, count, skip, i, j;
690 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
691 struct lustre_handle lockh;
694 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
695 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
696 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
699 /* our locks are page granular thanks to osc_enqueue, we invalidate the
701 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
702 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
706 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
707 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
708 if (lsm->lsm_stripe_count > 1) {
709 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
710 skip = (lsm->lsm_stripe_count - 1) * count;
711 start += start/count * skip + stripe * count;
713 end += end/count * skip + stripe * count;
715 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
718 i = inode->i_size ? (inode->i_size - 1) >> PAGE_CACHE_SHIFT : 0;
722 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
723 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
724 count, skip, end, discard ? " (DISCARDING)" : "");
726 /* walk through the vmas on the inode and tear down mmaped pages that
727 * intersect with the lock. this stops immediately if there are no
728 * mmap()ed regions of the file. This is not efficient at all and
729 * should be short lived. We'll associate mmap()ed pages with the lock
730 * and will be able to find them directly */
732 for (i = start; i <= end; i += (j + skip)) {
733 j = min(count - (i % count), end - i + 1);
735 LASSERT(inode->i_mapping);
736 if (ll_teardown_mmaps(inode->i_mapping,
737 (__u64)i << PAGE_CACHE_SHIFT,
738 ((__u64)(i+j) << PAGE_CACHE_SHIFT) - 1) )
742 /* this is the simplistic implementation of page eviction at
743 * cancelation. It is careful to get races with other page
744 * lockers handled correctly. fixes from bug 20 will make it
745 * more efficient by associating locks with pages and with
746 * batching writeback under the lock explicitly. */
747 for (i = start, j = start % count; i <= end;
748 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
750 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
756 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
757 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
758 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
761 if (!mapping_has_pages(inode->i_mapping)) {
762 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
768 page = find_get_page(inode->i_mapping, i);
771 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
772 i, tmpex.l_extent.start);
775 /* page->mapping to check with racing against teardown */
776 if (!discard && clear_page_dirty_for_io(page)) {
777 rc = ll_call_writepage(inode, page);
779 CERROR("writepage of page %p failed: %d\n",
781 /* either waiting for io to complete or reacquiring
782 * the lock that the failed writepage released */
786 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
787 /* check to see if another DLM lock covers this page */
788 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
789 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
791 &lock->l_resource->lr_name, LDLM_EXTENT,
792 &tmpex, LCK_PR | LCK_PW, &lockh);
793 if (rc2 == 0 && page->mapping != NULL) {
794 // checking again to account for writeback's lock_page()
795 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
796 ll_ra_accounting(page, inode->i_mapping);
797 ll_truncate_complete_page(page);
800 page_cache_release(page);
802 LASSERTF(tmpex.l_extent.start <=
803 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
804 lock->l_policy_data.l_extent.end + 1),
805 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
806 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
811 static int ll_extent_lock_callback(struct ldlm_lock *lock,
812 struct ldlm_lock_desc *new, void *data,
815 struct lustre_handle lockh = { 0 };
819 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
820 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
825 case LDLM_CB_BLOCKING:
826 ldlm_lock2handle(lock, &lockh);
827 rc = ldlm_cli_cancel(&lockh);
829 CERROR("ldlm_cli_cancel failed: %d\n", rc);
831 case LDLM_CB_CANCELING: {
833 struct ll_inode_info *lli;
834 struct lov_stripe_md *lsm;
838 /* This lock wasn't granted, don't try to evict pages */
839 if (lock->l_req_mode != lock->l_granted_mode)
842 inode = ll_inode_from_lock(lock);
845 lli = ll_i2info(inode);
848 if (lli->lli_smd == NULL)
852 stripe = ll_lock_to_stripe_offset(inode, lock);
855 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
857 down(&lli->lli_size_sem);
858 lock_res_and_lock(lock);
859 kms = ldlm_extent_shift_kms(lock,
860 lsm->lsm_oinfo[stripe].loi_kms);
862 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
863 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
864 lsm->lsm_oinfo[stripe].loi_kms, kms);
865 lsm->lsm_oinfo[stripe].loi_kms = kms;
866 unlock_res_and_lock(lock);
867 up(&lli->lli_size_sem);
868 //ll_try_done_writing(inode);
881 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
883 /* XXX ALLOCATE - 160 bytes */
884 struct inode *inode = ll_inode_from_lock(lock);
885 struct ll_inode_info *lli = ll_i2info(inode);
886 struct lustre_handle lockh = { 0 };
891 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
892 LDLM_FL_BLOCK_CONV)) {
893 LBUG(); /* not expecting any blocked async locks yet */
894 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
896 ldlm_lock_dump(D_OTHER, lock, 0);
897 ldlm_reprocess_all(lock->l_resource);
901 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
903 stripe = ll_lock_to_stripe_offset(inode, lock);
907 if (lock->l_lvb_len) {
908 struct lov_stripe_md *lsm = lli->lli_smd;
910 lvb = lock->l_lvb_data;
911 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
914 lock_res_and_lock(lock);
915 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
916 kms = ldlm_extent_shift_kms(NULL, kms);
917 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
918 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
919 lsm->lsm_oinfo[stripe].loi_kms, kms);
920 lsm->lsm_oinfo[stripe].loi_kms = kms;
921 unlock_res_and_lock(lock);
927 wake_up(&lock->l_waitq);
929 ldlm_lock2handle(lock, &lockh);
930 ldlm_lock_decref(&lockh, LCK_PR);
935 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
937 struct ptlrpc_request *req = reqp;
938 struct inode *inode = ll_inode_from_lock(lock);
939 struct ll_inode_info *lli;
941 struct lov_stripe_md *lsm;
942 int rc, size = sizeof(*lvb), stripe;
946 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
947 lli = ll_i2info(inode);
949 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
953 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
955 /* First, find out which stripe index this lock corresponds to. */
956 stripe = ll_lock_to_stripe_offset(inode, lock);
958 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
960 rc = lustre_pack_reply(req, 1, &size, NULL);
962 CERROR("lustre_pack_reply: %d\n", rc);
966 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
967 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
968 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
969 lvb->lvb_atime = LTIME_S(inode->i_atime);
970 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
972 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
973 inode->i_size, stripe, lvb->lvb_size);
979 /* These errors are normal races, so we don't want to fill the console
980 * with messages by calling ptlrpc_error() */
981 if (rc == -ELDLM_NO_LOCK_DATA)
982 lustre_pack_reply(req, 0, NULL, NULL);
988 /* NB: lov_merge_size will prefer locally cached writes if they extend the
989 * file (because it prefers KMS over RSS when larger) */
990 int ll_glimpse_size(struct inode *inode)
992 struct ll_inode_info *lli = ll_i2info(inode);
993 struct ll_sb_info *sbi = ll_i2sbi(inode);
994 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
995 struct lustre_handle lockh = { 0 };
996 int rc, flags = LDLM_FL_HAS_INTENT;
999 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1001 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
1002 LCK_PR, &flags, ll_extent_lock_callback,
1003 ldlm_completion_ast, ll_glimpse_callback, inode,
1004 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1009 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1010 RETURN(rc > 0 ? -EIO : rc);
1013 down(&lli->lli_size_sem);
1014 inode->i_size = lov_merge_size(lli->lli_smd, 0);
1015 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
1016 up(&lli->lli_size_sem);
1018 LTIME_S(inode->i_mtime) = lov_merge_mtime(lli->lli_smd,
1019 LTIME_S(inode->i_mtime));
1021 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
1022 (__u64)inode->i_size, (__u64)inode->i_blocks);
1024 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
1028 void ll_stime_record(struct ll_sb_info *sbi, struct timeval *start,
1029 struct obd_service_time *stime)
1031 struct timeval stop;
1032 do_gettimeofday(&stop);
1034 spin_lock(&sbi->ll_lock);
1035 lprocfs_stime_record(stime, &stop, start);
1036 spin_unlock(&sbi->ll_lock);
1039 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
1040 struct lov_stripe_md *lsm, int mode,
1041 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
1042 int ast_flags, struct obd_service_time *stime)
1044 struct ll_inode_info *lli = ll_i2info(inode);
1045 struct ll_sb_info *sbi = ll_i2sbi(inode);
1046 struct timeval start;
1050 LASSERT(lockh->cookie == 0);
1052 /* XXX phil: can we do this? won't it screw the file size up? */
1053 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1054 (sbi->ll_flags & LL_SBI_NOLCK))
1057 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
1058 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
1060 do_gettimeofday(&start);
1061 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
1062 &ast_flags, ll_extent_lock_callback,
1063 ldlm_completion_ast, ll_glimpse_callback, inode,
1064 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
1068 ll_stime_record(sbi, &start, stime);
1070 if (policy->l_extent.start == 0 &&
1071 policy->l_extent.end == OBD_OBJECT_EOF) {
1072 /* vmtruncate()->ll_truncate() first sets the i_size and then
1073 * the kms under both a DLM lock and the i_sem. If we don't
1074 * get the i_sem here we can match the DLM lock and reset
1075 * i_size from the kms before the truncating path has updated
1076 * the kms. generic_file_write can then trust the stale i_size
1077 * when doing appending writes and effectively cancel the
1078 * result of the truncate. Getting the i_sem after the enqueue
1079 * maintains the DLM -> i_sem acquiry order. */
1080 down(&lli->lli_size_sem);
1081 inode->i_size = lov_merge_size(lsm, 1);
1082 up(&lli->lli_size_sem);
1086 LTIME_S(inode->i_mtime) =
1087 lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1093 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1094 struct lov_stripe_md *lsm, int mode,
1095 struct lustre_handle *lockh)
1097 struct ll_sb_info *sbi = ll_i2sbi(inode);
1101 /* XXX phil: can we do this? won't it screw the file size up? */
1102 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1103 (sbi->ll_flags & LL_SBI_NOLCK))
1106 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1111 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1114 struct inode *inode = file->f_dentry->d_inode;
1115 struct ll_inode_info *lli = ll_i2info(inode);
1116 struct lov_stripe_md *lsm = lli->lli_smd;
1117 struct ll_lock_tree tree;
1118 struct ll_lock_tree_node *node;
1123 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1124 inode->i_ino, inode->i_generation, inode, count, *ppos);
1126 /* "If nbyte is 0, read() will return 0 and have no other results."
1127 * -- Single Unix Spec */
1131 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
1137 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1140 tree.lt_fd = file->private_data;
1142 rc = ll_tree_lock(&tree, node, inode, buf, count,
1143 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1147 down(&lli->lli_size_sem);
1148 kms = lov_merge_size(lsm, 1);
1149 if (*ppos + count - 1 > kms) {
1150 /* A glimpse is necessary to determine whether we return a short
1151 * read or some zeroes at the end of the buffer */
1152 up(&lli->lli_size_sem);
1153 retval = ll_glimpse_size(inode);
1157 inode->i_size = kms;
1158 up(&lli->lli_size_sem);
1161 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1162 inode->i_ino, count, *ppos, inode->i_size);
1164 /* turn off the kernel's read-ahead */
1165 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1168 file->f_ra.ra_pages = 0;
1170 retval = generic_file_read(file, buf, count, ppos);
1171 ll_audit_log(inode, AUDIT_READ, retval);
1174 ll_tree_unlock(&tree, inode);
1179 * Write to a file (through the page cache).
1181 static ssize_t ll_file_write(struct file *file, const char *buf,
1182 size_t count, loff_t *ppos)
1184 struct inode *inode = file->f_dentry->d_inode;
1185 loff_t maxbytes = ll_file_maxbytes(inode);
1186 struct ll_lock_tree tree;
1187 struct ll_lock_tree_node *node;
1192 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1193 inode->i_ino, inode->i_generation, inode, count, *ppos);
1195 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1197 /* POSIX, but surprised the VFS doesn't check this already */
1201 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1202 * called on the file, don't fail the below assertion (bug 2388). */
1203 if (file->f_flags & O_LOV_DELAY_CREATE &&
1204 ll_i2info(inode)->lli_smd == NULL)
1207 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1209 if (file->f_flags & O_APPEND)
1210 node = ll_node_from_inode(inode, 0, OBD_OBJECT_EOF, LCK_PW);
1212 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1216 RETURN(PTR_ERR(node));
1218 tree.lt_fd = file->private_data;
1220 rc = ll_tree_lock(&tree, node, inode, buf, count,
1221 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1225 /* this is ok, g_f_w will overwrite this under i_sem if it races
1226 * with a local truncate, it just makes our maxbyte checking easier */
1227 if (file->f_flags & O_APPEND)
1228 *ppos = inode->i_size;
1230 if (*ppos >= maxbytes) {
1231 if (count || *ppos > maxbytes) {
1232 send_sig(SIGXFSZ, current, 0);
1233 GOTO(out, retval = -EFBIG);
1236 if (*ppos + count > maxbytes)
1237 count = maxbytes - *ppos;
1239 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1240 inode->i_ino, count, *ppos);
1242 /* mark open handle dirty */
1243 set_bit(LLI_F_DIRTY_HANDLE, &(ll_i2info(inode)->lli_flags));
1245 /* generic_file_write handles O_APPEND after getting i_sem */
1246 retval = generic_file_write(file, buf, count, ppos);
1250 ll_audit_log(inode, AUDIT_WRITE, retval);
1252 ll_tree_unlock(&tree, inode);
1253 /* serialize with mmap/munmap/mremap */
1254 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
1255 retval > 0 ? retval : 0);
1259 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
1260 int flags, struct lov_user_md *lum,
1263 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
1264 struct ll_inode_info *lli = ll_i2info(inode);
1265 struct ptlrpc_request *req = NULL;
1266 struct obd_client_handle *och;
1267 struct lov_stripe_md *lsm;
1268 struct lustre_md md;
1274 if ((file->f_flags+1) & O_ACCMODE)
1276 if (file->f_flags & O_TRUNC)
1279 down(&lli->lli_open_sem);
1282 up(&lli->lli_open_sem);
1283 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
1288 f = get_empty_filp();
1292 f->f_dentry = file->f_dentry;
1293 f->f_vfsmnt = file->f_vfsmnt;
1296 rc = ll_intent_alloc(&oit);
1300 rc = ll_intent_file_open(f, lum, lum_size, &oit);
1303 if (it_disposition(&oit, DISP_LOOKUP_NEG))
1306 req = LUSTRE_IT(&oit)->it_data;
1307 rc = LUSTRE_IT(&oit)->it_status;
1312 rc = mdc_req2lustre_md(ll_i2mdexp(inode), req, 1,
1313 ll_i2dtexp(inode), &md);
1316 ll_update_inode(f->f_dentry->d_inode, &md);
1318 OBD_ALLOC(och, sizeof(struct obd_client_handle));
1320 GOTO(out, rc = -ENOMEM);
1322 /* actually ll_local_open() cannot fail! */
1323 rc = ll_local_open(f, &oit, och);
1327 if (LUSTRE_IT(&oit)->it_lock_mode) {
1328 ldlm_lock_decref_and_cancel((struct lustre_handle *)
1329 &LUSTRE_IT(&oit)->it_lock_handle,
1330 LUSTRE_IT(&oit)->it_lock_mode);
1331 LUSTRE_IT(&oit)->it_lock_mode = 0;
1334 ll_intent_release(&oit);
1336 /* ll_file_release() will decrease the count, but won't free anything
1337 * because we have at least one more reference coming from actual
1339 down(&lli->lli_och_sem);
1340 lli->lli_open_fd_write_count++;
1341 up(&lli->lli_och_sem);
1342 rc = ll_file_release(f->f_dentry->d_inode, f);
1344 /* Now also destroy our supplemental och */
1345 ll_md_och_close(ll_i2mdexp(inode), f->f_dentry->d_inode, och, 0);
1348 ll_intent_release(&oit);
1351 up(&lli->lli_open_sem);
1353 ptlrpc_req_finished(req);
1357 static int ll_lov_setea(struct inode *inode, struct file *file,
1360 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
1361 int lum_size = sizeof(struct lov_user_md) +
1362 sizeof(struct lov_user_ost_data);
1363 struct lov_user_md *lump;
1367 if (!capable(CAP_SYS_ADMIN))
1370 OBD_ALLOC(lump, lum_size);
1374 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
1376 GOTO(out_free_lump, rc = -EFAULT);
1378 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1381 OBD_FREE(lump, lum_size);
1385 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1388 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1389 int rc, flags = FMODE_WRITE;
1392 /* Bug 1152: copy properly when this is no longer true */
1393 LASSERT(sizeof(lum) == sizeof(*lump));
1394 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1395 rc = copy_from_user(&lum, lump, sizeof(lum));
1399 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1401 put_user(0, &lump->lmm_stripe_count);
1402 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
1403 0, ll_i2info(inode)->lli_smd, lump);
1408 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1410 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1415 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
1419 static int ll_get_grouplock(struct inode *inode, struct file *file,
1422 struct ll_file_data *fd = file->private_data;
1423 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1424 .end = OBD_OBJECT_EOF}};
1425 struct lustre_handle lockh = { 0 };
1426 struct ll_inode_info *lli = ll_i2info(inode);
1427 struct lov_stripe_md *lsm = lli->lli_smd;
1431 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1435 policy.l_extent.gid = arg;
1436 if (file->f_flags & O_NONBLOCK)
1437 flags = LDLM_FL_BLOCK_NOWAIT;
1439 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags,
1440 &ll_i2sbi(inode)->ll_grouplock_stime);
1444 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1446 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1451 static int ll_put_grouplock(struct inode *inode, struct file *file,
1454 struct ll_file_data *fd = file->private_data;
1455 struct ll_inode_info *lli = ll_i2info(inode);
1456 struct lov_stripe_md *lsm = lli->lli_smd;
1460 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1461 /* Ugh, it's already unlocked. */
1465 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1468 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1470 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1475 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1480 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1483 struct ll_file_data *fd = file->private_data;
1484 struct ll_sb_info *sbi = ll_i2sbi(inode);
1488 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1489 inode->i_generation, inode, cmd);
1491 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1494 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1496 case LL_IOC_GETFLAGS:
1497 /* Get the current value of the file flags */
1498 return put_user(fd->fd_flags, (int *)arg);
1499 case LL_IOC_SETFLAGS:
1500 case LL_IOC_CLRFLAGS:
1501 /* Set or clear specific file flags */
1502 /* XXX This probably needs checks to ensure the flags are
1503 * not abused, and to handle any flag side effects.
1505 if (get_user(flags, (int *) arg))
1508 if (cmd == LL_IOC_SETFLAGS)
1509 fd->fd_flags |= flags;
1511 fd->fd_flags &= ~flags;
1513 case LL_IOC_LOV_SETSTRIPE:
1514 RETURN(ll_lov_setstripe(inode, file, arg));
1515 case LL_IOC_LOV_SETEA:
1516 RETURN(ll_lov_setea(inode, file, arg));
1517 case IOC_MDC_SHOWFID: {
1518 struct lustre_id *idp = (struct lustre_id *)arg;
1519 struct lustre_id id;
1523 filename = getname((const char *)arg);
1524 if (IS_ERR(filename))
1525 RETURN(PTR_ERR(filename));
1527 ll_inode2id(&id, inode);
1529 rc = ll_get_fid(sbi->ll_md_exp, &id, filename, &id);
1531 GOTO(out_filename, rc);
1533 rc = copy_to_user(idp, &id, sizeof(*idp));
1535 GOTO(out_filename, rc = -EFAULT);
1542 case LL_IOC_KEY_TYPE: {
1543 struct obd_ioctl_data *data;
1546 int typelen, rc, len = 0;
1548 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1553 type = data->ioc_inlbuf1;
1554 typelen = data->ioc_inllen1;
1557 CDEBUG(D_INFO, "LL_IOC_KEY_TYPE missing filename\n");
1558 GOTO(out, rc = -EINVAL);
1560 ll_set_sb_gksinfo(inode->i_sb, type);
1563 obd_ioctl_freedata(buf, len);
1567 case LL_IOC_LOV_GETSTRIPE:
1568 RETURN(ll_lov_getstripe(inode, arg));
1569 case EXT3_IOC_GETFLAGS:
1570 case EXT3_IOC_SETFLAGS:
1571 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1572 case LL_IOC_GROUP_LOCK:
1573 RETURN(ll_get_grouplock(inode, file, arg));
1574 case LL_IOC_GROUP_UNLOCK:
1575 RETURN(ll_put_grouplock(inode, file, arg));
1576 case EXT3_IOC_GETVERSION_OLD:
1577 case EXT3_IOC_GETVERSION:
1578 return put_user(inode->i_generation, (int *) arg);
1579 /* We need to special case any other ioctls we want to handle,
1580 * to send them to the MDS/OST as appropriate and to properly
1581 * network encode the arg field.
1582 case EXT2_IOC_GETVERSION_OLD:
1583 case EXT2_IOC_GETVERSION_NEW:
1584 case EXT2_IOC_SETVERSION_OLD:
1585 case EXT2_IOC_SETVERSION_NEW:
1586 case EXT3_IOC_SETVERSION_OLD:
1587 case EXT3_IOC_SETVERSION:
1589 case LL_IOC_FLUSH_CRED:
1590 RETURN(ll_flush_cred(inode));
1592 RETURN(ll_set_audit(inode, arg));
1594 RETURN( obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
1599 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1601 struct inode *inode = file->f_dentry->d_inode;
1602 struct ll_file_data *fd = file->private_data;
1603 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1604 struct lustre_handle lockh = {0};
1607 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1608 inode->i_generation, inode,
1609 offset + ((origin==2) ? inode->i_size : file->f_pos));
1611 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1612 if (origin == 2) { /* SEEK_END */
1613 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1614 struct ll_inode_info *lli = ll_i2info(inode);
1615 int nonblock = 0, rc;
1617 if (file->f_flags & O_NONBLOCK)
1618 nonblock = LDLM_FL_BLOCK_NOWAIT;
1620 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1621 nonblock, &ll_i2sbi(inode)->ll_seek_stime);
1625 down(&lli->lli_size_sem);
1626 offset += inode->i_size;
1627 up(&lli->lli_size_sem);
1628 } else if (origin == 1) { /* SEEK_CUR */
1629 offset += file->f_pos;
1633 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1634 if (offset != file->f_pos) {
1635 file->f_pos = offset;
1636 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1638 file->f_version = ++event;
1645 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1649 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1651 struct inode *inode = dentry->d_inode;
1652 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1653 struct lustre_id id;
1654 struct ptlrpc_request *req;
1657 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1658 inode->i_generation, inode);
1660 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1662 /* fsync's caller has already called _fdata{sync,write}, we want
1663 * that IO to finish before calling the osc and mdc sync methods */
1664 rc = filemap_fdatawait(inode->i_mapping);
1666 ll_inode2id(&id, inode);
1667 err = md_sync(ll_i2sbi(inode)->ll_md_exp, &id, &req);
1671 ptlrpc_req_finished(req);
1674 struct obdo *oa = obdo_alloc();
1677 RETURN(rc ? rc : -ENOMEM);
1679 oa->o_id = lsm->lsm_object_id;
1680 oa->o_gr = lsm->lsm_object_gr;
1681 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1683 obdo_from_inode(oa, inode, (OBD_MD_FLTYPE | OBD_MD_FLATIME |
1684 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1687 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
1697 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1699 struct inode *inode = file->f_dentry->d_inode;
1700 struct ll_inode_info *li = ll_i2info(inode);
1701 struct ll_sb_info *sbi = ll_i2sbi(inode);
1702 struct obd_device *obddev;
1703 struct ldlm_res_id res_id =
1704 { .name = {id_fid(&li->lli_id), id_group(&li->lli_id), LDLM_FLOCK} };
1705 struct lustre_handle lockh = {0};
1706 ldlm_policy_data_t flock;
1707 ldlm_mode_t mode = 0;
1712 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1713 inode->i_ino, file_lock);
1715 flock.l_flock.pid = file_lock->fl_pid;
1716 flock.l_flock.start = file_lock->fl_start;
1717 flock.l_flock.end = file_lock->fl_end;
1719 switch (file_lock->fl_type) {
1724 /* An unlock request may or may not have any relation to
1725 * existing locks so we may not be able to pass a lock handle
1726 * via a normal ldlm_lock_cancel() request. The request may even
1727 * unlock a byte range in the middle of an existing lock. In
1728 * order to process an unlock request we need all of the same
1729 * information that is given with a normal read or write record
1730 * lock request. To avoid creating another ldlm unlock (cancel)
1731 * message we'll treat a LCK_NL flock request as an unlock. */
1738 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1753 flags = LDLM_FL_BLOCK_NOWAIT;
1759 flags = LDLM_FL_TEST_LOCK;
1760 /* Save the old mode so that if the mode in the lock changes we
1761 * can decrement the appropriate reader or writer refcount. */
1762 file_lock->fl_type = mode;
1765 CERROR("unknown fcntl lock command: %d\n", cmd);
1769 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1770 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1771 flags, mode, flock.l_flock.start, flock.l_flock.end);
1773 obddev = md_get_real_obd(sbi->ll_md_exp, &li->lli_id);
1774 rc = ldlm_cli_enqueue(obddev->obd_self_export, NULL,
1775 obddev->obd_namespace,
1776 res_id, LDLM_FLOCK, &flock, mode, &flags,
1777 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1778 NULL, 0, NULL, &lockh);
1782 int ll_inode_revalidate_it(struct dentry *dentry)
1784 struct lookup_intent oit = { .it_op = IT_GETATTR };
1785 struct inode *inode = dentry->d_inode;
1786 struct ptlrpc_request *req = NULL;
1787 struct ll_inode_info *lli;
1788 struct lov_stripe_md *lsm;
1789 struct ll_sb_info *sbi;
1790 struct lustre_id id;
1795 CERROR("REPORT THIS LINE TO PETER\n");
1799 sbi = ll_i2sbi(inode);
1801 ll_inode2id(&id, inode);
1802 lli = ll_i2info(inode);
1803 LASSERT(id_fid(&id) != 0);
1805 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), name=%s(%p)\n",
1806 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1809 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1810 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1813 rc = ll_intent_alloc(&oit);
1817 rc = ll_crypto_init_it_key(inode, &oit);
1821 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1822 &oit, 0, &req, ll_mdc_blocking_ast);
1826 rc = revalidate_it_finish(req, 1, &oit, dentry);
1831 ll_lookup_finish_locks(&oit, dentry);
1833 if (!LLI_HAVE_FLSIZE(inode)) {
1834 /* if object not yet allocated, don't validate size */
1837 /* ll_glimpse_size() will prefer locally cached
1838 * writes if they extend the file */
1839 rc = ll_glimpse_size(inode);
1844 ll_intent_release(&oit);
1846 ptlrpc_req_finished(req);
1850 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1851 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
1854 struct inode *inode = de->d_inode;
1855 struct ll_inode_info *lli = ll_i2info(inode);
1857 res = ll_inode_revalidate_it(de);
1858 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1863 stat->ino = inode->i_ino;
1864 stat->mode = inode->i_mode;
1865 stat->nlink = inode->i_nlink;
1866 stat->uid = inode->i_uid;
1867 stat->gid = inode->i_gid;
1868 stat->atime = inode->i_atime;
1869 stat->mtime = inode->i_mtime;
1870 stat->ctime = inode->i_ctime;
1871 stat->blksize = inode->i_blksize;
1873 down(&lli->lli_size_sem);
1874 stat->size = inode->i_size;
1875 stat->blocks = inode->i_blocks;
1876 up(&lli->lli_size_sem);
1878 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1879 stat->dev = id_group(&ll_i2info(inode)->lli_id);
1885 int ll_setxattr_internal(struct inode *inode, const char *name,
1886 const void *value, size_t size, int flags,
1889 struct ll_sb_info *sbi = ll_i2sbi(inode);
1890 struct ptlrpc_request *request = NULL;
1891 struct mdc_op_data *op_data;
1894 int rc = 0, key_size = 0;
1897 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu\n", inode->i_ino);
1898 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_SETXATTR);
1900 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
1901 RETURN(-EOPNOTSUPP);
1903 memset(&attr, 0x0, sizeof(attr));
1904 attr.ia_valid |= valid;
1905 attr.ia_attr_flags = flags;
1907 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1908 rc = ll_crypto_get_mac(inode, &attr, (void *)value, size,
1912 OBD_ALLOC(op_data, sizeof(*op_data));
1916 ll_inode2mdc_data(op_data, inode, (OBD_MD_FLID | OBD_MD_MEA));
1918 rc = md_setattr(sbi->ll_md_exp, op_data, &attr,
1919 (void *)name, strnlen(name, XATTR_NAME_MAX) + 1,
1920 (void *)value, size, key, key_size, &request);
1921 OBD_FREE(op_data, sizeof(*op_data));
1923 if (key && key_size)
1924 OBD_FREE(key, key_size);
1926 CDEBUG(D_SEC, "md_setattr fails: rc = %d\n", rc);
1930 ptlrpc_req_finished(request);
1934 int ll_setxattr(struct dentry *dentry, const char *name, const void *value,
1935 size_t size, int flags)
1938 struct posix_acl *acl;
1939 struct ll_inode_info *lli;
1942 rc = ll_setxattr_internal(dentry->d_inode, name, value, size,
1945 /* update inode's acl info */
1946 if (rc == 0 && strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1948 acl = posix_acl_from_xattr(value, size);
1950 CERROR("convert from xattr to acl error: %ld",
1954 error = posix_acl_valid(acl);
1956 CERROR("acl valid error: %d", error);
1957 posix_acl_release(acl);
1965 lli = ll_i2info(dentry->d_inode);
1966 spin_lock(&lli->lli_lock);
1967 if (lli->lli_posix_acl != NULL)
1968 posix_acl_release(lli->lli_posix_acl);
1969 lli->lli_posix_acl = acl;
1970 spin_unlock(&lli->lli_lock);
1977 int ll_removexattr(struct dentry *dentry, const char *name)
1979 return ll_setxattr_internal(dentry->d_inode, name, NULL, 0, 0,
1983 int ll_getxattr_internal(struct inode *inode, const char *name,
1984 void *value, size_t size, __u64 valid)
1986 struct ptlrpc_request *request = NULL;
1987 struct ll_sb_info *sbi = ll_i2sbi(inode);
1988 struct lustre_id id;
1989 struct mds_body *body;
1994 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETXATTR);
1996 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
1997 RETURN(-EOPNOTSUPP);
1999 ll_inode2id(&id, inode);
2000 rc = md_getattr(sbi->ll_md_exp, &id, valid, name, NULL, 0,
2001 size, NULL, &request);
2003 if (rc != -ENODATA && rc != -EOPNOTSUPP)
2004 CERROR("rc = %d\n", rc);
2008 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
2009 LASSERT(body != NULL);
2010 LASSERT_REPSWABBED(request, 0);
2012 ea_size = body->eadatasize;
2014 GOTO(out, rc = ea_size);
2016 LASSERT(ea_size <= request->rq_repmsg->buflens[1]);
2017 ea_data = lustre_msg_buf(request->rq_repmsg, 1, ea_size);
2018 LASSERT(ea_data != NULL);
2019 LASSERT_REPSWABBED(request, 1);
2022 memcpy(value, ea_data, ea_size);
2026 ptlrpc_req_finished(request);
2030 int ll_getxattr(struct dentry *dentry, const char *name, void *value,
2033 return ll_getxattr_internal(dentry->d_inode, name,
2034 value, size, OBD_MD_FLXATTR);
2037 int ll_listxattr(struct dentry *dentry, char *list, size_t size)
2039 return ll_getxattr_internal(dentry->d_inode, NULL, list, size,
2040 OBD_MD_FLXATTRLIST);
2044 * Here we hold DLM lock across permission check, to get better
2045 * conformance especially for remote acl. will introduce certain
2046 * overhead, maybe need a profile.
2049 lustre_check_acl(struct inode *inode, int mask)
2051 struct lookup_intent it = { .it_op = IT_GETATTR };
2052 struct dentry de = { .d_inode = inode };
2053 struct ll_sb_info *sbi;
2054 struct lustre_id id;
2055 struct ptlrpc_request *req = NULL;
2056 struct ll_inode_info *lli = ll_i2info(inode);
2057 struct posix_acl *acl;
2061 sbi = ll_i2sbi(inode);
2062 ll_inode2id(&id, inode);
2064 if (ll_intent_alloc(&it))
2067 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
2068 &it, 0, &req, ll_mdc_blocking_ast);
2070 ll_intent_free(&it);
2074 rc = revalidate_it_finish(req, 1, &it, &de);
2076 ll_intent_release(&it);
2080 if (sbi->ll_remote) {
2081 rc = ll_remote_acl_permission(inode, mask);
2083 spin_lock(&lli->lli_lock);
2084 acl = posix_acl_dup(ll_i2info(inode)->lli_posix_acl);
2085 spin_unlock(&lli->lli_lock);
2090 rc = posix_acl_permission(inode, acl, mask);
2091 posix_acl_release(acl);
2095 ll_lookup_finish_locks(&it, &de);
2096 ll_intent_free(&it);
2100 ptlrpc_req_finished(req);
2105 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
2107 struct ll_sb_info *sbi = ll_i2sbi(inode);
2109 /* for remote client, permission bits in inode doesn't
2110 * play a role anymore.
2113 return lustre_check_acl(inode, mask);
2115 return generic_permission(inode, mask, lustre_check_acl);
2118 struct file_operations ll_file_operations = {
2119 .read = ll_file_read,
2120 .write = ll_file_write,
2121 .ioctl = ll_file_ioctl,
2122 .open = ll_file_open,
2123 .release = ll_file_release,
2124 .mmap = ll_file_mmap,
2125 .llseek = ll_file_seek,
2126 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
2127 .sendfile = generic_file_sendfile,
2130 .lock = ll_file_flock
2133 struct inode_operations ll_file_inode_operations = {
2134 .setattr = ll_setattr,
2135 .truncate = ll_truncate,
2136 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
2137 .getattr = ll_getattr,
2139 .revalidate_it = ll_inode_revalidate_it,
2141 .setxattr = ll_setxattr,
2142 .getxattr = ll_getxattr,
2143 .listxattr = ll_listxattr,
2144 .removexattr = ll_removexattr,
2145 .permission = ll_inode_permission,