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 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
32 #include <linux/lustre_compat25.h>
34 #include "llite_internal.h"
35 #include <linux/obd_lov.h>
37 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
38 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
39 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
41 int ll_validate_size(struct inode *inode, __u64 *size, __u64 *blocks)
43 ldlm_policy_data_t extent = { .l_extent = { 0, OBD_OBJECT_EOF } };
44 struct obd_export *exp = ll_i2sbi(inode)->ll_dt_exp;
45 struct ll_inode_info *lli = ll_i2info(inode);
46 struct lustre_handle match_lockh = {0};
50 if (lli->lli_smd == NULL)
53 LASSERT(size != NULL && blocks != NULL);
55 flags = LDLM_FL_TEST_LOCK | LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
56 rc = obd_match(exp, lli->lli_smd, LDLM_EXTENT, &extent,
57 LCK_PR | LCK_PW, &flags, inode, &match_lockh);
59 /* we have no all needed locks,
60 * so we don't know actual size */
64 /* we know actual size! */
65 down(&lli->lli_size_sem);
66 *size = lov_merge_size(lli->lli_smd, 0);
67 *blocks = lov_merge_blocks(lli->lli_smd);
68 up(&lli->lli_size_sem);
74 int ll_md_och_close(struct obd_export *md_exp, struct inode *inode,
75 struct obd_client_handle *och, int dirty)
77 struct ptlrpc_request *req = NULL;
78 struct obdo *obdo = NULL;
79 struct obd_device *obd;
83 obd = class_exp2obd(md_exp);
85 CERROR("Invalid MDC connection handle "LPX64"\n",
86 md_exp->exp_handle.h_cookie);
92 * here we check if this is forced umount. If so this is called on
93 * canceling "open lock" and we do not call md_close() in this case , as
94 * it will not successful, as import is already deactivated.
96 if (obd->obd_no_recov)
99 /* closing opened file */
104 obdo->o_id = inode->i_ino;
105 obdo->o_generation = inode->i_generation;
106 obdo->o_valid = OBD_MD_FLID;
107 obdo_from_inode(obdo, inode, (OBD_MD_FLTYPE | OBD_MD_FLMODE |
108 OBD_MD_FLATIME | OBD_MD_FLMTIME |
110 if (0 /* ll_is_inode_dirty(inode) */) {
111 obdo->o_flags = MDS_BFLAG_UNCOMMITTED_WRITES;
112 obdo->o_valid |= OBD_MD_FLFLAGS;
114 obdo->o_fid = id_fid(&ll_i2info(inode)->lli_id);
115 obdo->o_mds = id_group(&ll_i2info(inode)->lli_id);
117 obdo->o_valid |= OBD_MD_FLEPOCH;
118 obdo->o_easize = ll_i2info(inode)->lli_io_epoch;
121 /* we modified data through this handle */
122 obdo->o_flags |= MDS_BFLAG_DIRTY_EPOCH;
123 obdo->o_valid |= OBD_MD_FLFLAGS;
124 if (ll_validate_size(inode, &obdo->o_size, &obdo->o_blocks))
125 obdo->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
128 rc = md_close(md_exp, obdo, och, &req);
133 * we are the last writer, so the MDS has instructed us to get
134 * the file size and any write cookies, then close again.
137 //ll_queue_done_writing(inode);
140 CERROR("inode %lu mdc close failed: rc = %d\n",
141 (unsigned long)inode->i_ino, rc);
144 ptlrpc_req_finished(req);
147 mdc_clear_open_replay_data(md_exp, och);
148 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
149 OBD_FREE(och, sizeof *och);
153 int ll_md_real_close(struct obd_export *md_exp,
154 struct inode *inode, int flags)
156 struct ll_inode_info *lli = ll_i2info(inode);
157 int freeing = inode->i_state & I_FREEING;
158 struct obd_client_handle **och_p;
159 struct obd_client_handle *och;
161 int rc = 0, dirty = 0;
164 if (flags & FMODE_WRITE) {
165 och_p = &lli->lli_mds_write_och;
166 och_usecount = &lli->lli_open_fd_write_count;
167 } else if (flags & FMODE_EXEC) {
168 och_p = &lli->lli_mds_exec_och;
169 och_usecount = &lli->lli_open_fd_exec_count;
171 och_p = &lli->lli_mds_read_och;
172 och_usecount = &lli->lli_open_fd_read_count;
175 down(&lli->lli_och_sem);
176 if (*och_usecount) { /* There are still users of this handle, so
178 up(&lli->lli_och_sem);
181 if (ll_is_inode_dirty(inode)) {
182 /* the inode still has dirty pages, let's close later */
183 CDEBUG(D_INODE, "inode %lu/%u still has dirty pages\n",
184 inode->i_ino, inode->i_generation);
185 LASSERT(freeing == 0);
186 ll_queue_done_writing(inode);
187 up(&lli->lli_och_sem);
191 if (LLI_DIRTY_HANDLE(inode) && (flags & FMODE_WRITE)) {
192 clear_bit(LLI_F_DIRTY_HANDLE, &lli->lli_flags);
194 } else if (0 && !(flags & FMODE_SYNC) && !freeing) {
195 /* in order to speed up creation rate we pass
196 * closing to dedicated thread so we don't need
197 * to wait for close reply here -bzzz */
198 ll_queue_done_writing(inode);
199 up(&lli->lli_och_sem);
207 up(&lli->lli_och_sem);
210 * there might be a race and somebody have freed this och
211 * already. Another way to have this twice called is if file closing
212 * will fail due to netwok problems and on umount lock will be canceled
213 * and this will be called from block_ast callack.
215 if (och && och->och_fh.cookie != DEAD_HANDLE_MAGIC)
216 rc = ll_md_och_close(md_exp, inode, och, dirty);
221 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
224 struct ll_file_data *fd = file->private_data;
225 struct ll_inode_info *lli = ll_i2info(inode);
229 /* clear group lock, if present */
230 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
231 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
232 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
233 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
237 /* Let's see if we have good enough OPEN lock on the file and if
238 we can skip talking to MDS */
239 if (file->f_dentry->d_inode) {
241 struct obd_device *obddev;
242 struct lustre_handle lockh;
243 int flags = LDLM_FL_BLOCK_GRANTED;
244 struct ldlm_res_id file_res_id = {.name = {id_fid(&lli->lli_id),
245 id_group(&lli->lli_id)}};
246 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
248 down(&lli->lli_och_sem);
249 if (fd->fd_omode & FMODE_WRITE) {
251 LASSERT(lli->lli_open_fd_write_count);
252 lli->lli_open_fd_write_count--;
253 } else if (fd->fd_omode & FMODE_EXEC) {
255 LASSERT(lli->lli_open_fd_exec_count);
256 lli->lli_open_fd_exec_count--;
259 LASSERT(lli->lli_open_fd_read_count);
260 lli->lli_open_fd_read_count--;
262 up(&lli->lli_och_sem);
264 obddev = md_get_real_obd(md_exp, &lli->lli_id);
265 if (!ldlm_lock_match(obddev->obd_namespace, flags, &file_res_id,
266 LDLM_IBITS, &policy, lockmode, &lockh))
268 rc = ll_md_real_close(md_exp, file->f_dentry->d_inode,
271 ldlm_lock_decref(&lockh, lockmode);
275 file->private_data = NULL;
276 OBD_SLAB_FREE(fd, ll_file_data_slab, sizeof(*fd));
280 /* While this returns an error code, fput() the caller does not, so we need
281 * to make every effort to clean up all of our state here. Also, applications
282 * rarely check close errors and even if an error is returned they will not
283 * re-try the close call.
285 int ll_file_release(struct inode *inode, struct file *file)
287 struct ll_file_data *fd;
288 struct ll_sb_info *sbi = ll_i2sbi(inode);
292 CDEBUG(D_VFSTRACE, "VFS Op:inode="DLID4"(%p)\n",
293 OLID4(&ll_i2info(inode)->lli_id), inode);
295 /* don't do anything for / */
296 if (inode->i_sb->s_root == file->f_dentry)
299 lprocfs_counter_incr(sbi->ll_stats, LPROC_LL_RELEASE);
300 fd = (struct ll_file_data *)file->private_data;
303 rc = ll_md_close(sbi->ll_md_exp, inode, file);
307 static int ll_intent_file_open(struct file *file, void *lmm,
308 int lmmsize, struct lookup_intent *itp)
310 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
311 const char *name = (char *)file->f_dentry->d_name.name;
312 struct dentry *parent = file->f_dentry->d_parent;
313 const int len = file->f_dentry->d_name.len;
314 struct lustre_handle lockh;
315 struct mdc_op_data *op_data;
321 OBD_ALLOC(op_data, sizeof(*op_data));
325 ll_prepare_mdc_data(op_data, parent->d_inode, NULL,
328 rc = md_enqueue(sbi->ll_md_exp, LDLM_IBITS, itp, LCK_PR, op_data,
329 &lockh, lmm, lmmsize, ldlm_completion_ast,
330 ll_mdc_blocking_ast, NULL);
331 OBD_FREE(op_data, sizeof(*op_data));
333 if (LUSTRE_IT(itp)->it_lock_mode)
334 memcpy(&LUSTRE_IT(itp)->it_lock_handle,
335 &lockh, sizeof(lockh));
338 CERROR("lock enqueue: err: %d\n", rc);
343 void ll_och_fill(struct inode *inode, struct lookup_intent *it,
344 struct obd_client_handle *och)
346 struct ptlrpc_request *req = LUSTRE_IT(it)->it_data;
347 struct ll_inode_info *lli = ll_i2info(inode);
348 struct mds_body *body;
351 body = lustre_msg_buf (req->rq_repmsg, 1, sizeof (*body));
352 LASSERT (body != NULL); /* reply already checked out */
353 LASSERT_REPSWABBED (req, 1); /* and swabbed down */
355 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
356 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
357 lli->lli_io_epoch = body->io_epoch;
358 mdc_set_open_replay_data(ll_i2mdexp(inode), och,
359 LUSTRE_IT(it)->it_data);
362 int ll_local_open(struct file *file, struct lookup_intent *it,
363 struct obd_client_handle *och)
365 struct ll_file_data *fd;
369 ll_och_fill(file->f_dentry->d_inode, it, och);
371 LASSERTF(file->private_data == NULL, "file %.*s/%.*s ino %lu/%u (%o)\n",
372 file->f_dentry->d_name.len, file->f_dentry->d_name.name,
373 file->f_dentry->d_parent->d_name.len,
374 file->f_dentry->d_parent->d_name.name,
375 file->f_dentry->d_inode->i_ino,
376 file->f_dentry->d_inode->i_generation,
377 file->f_dentry->d_inode->i_mode);
379 OBD_SLAB_ALLOC(fd, ll_file_data_slab, SLAB_KERNEL, sizeof *fd);
381 /* We can't handle this well without reorganizing ll_file_open and
382 * ll_md_close(), so don't even try right now. */
385 file->private_data = fd;
386 ll_readahead_init(file->f_dentry->d_inode, &fd->fd_ras);
387 fd->fd_omode = it->it_flags;
391 /* Open a file, and (for the very first open) create objects on the OSTs at
392 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
393 * creation or open until ll_lov_setstripe() ioctl is called. We grab
394 * lli_open_sem to ensure no other process will create objects, send the
395 * stripe MD to the MDS, or try to destroy the objects if that fails.
397 * If we already have the stripe MD locally then we don't request it in
398 * mdc_open(), by passing a lmm_size = 0.
400 * It is up to the application to ensure no other processes open this file
401 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
402 * used. We might be able to avoid races of that sort by getting lli_open_sem
403 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
404 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
406 int ll_file_open(struct inode *inode, struct file *file)
408 struct ll_inode_info *lli = ll_i2info(inode);
409 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
410 .it_flags = file->f_flags };
411 struct lov_stripe_md *lsm;
412 struct ptlrpc_request *req;
414 struct obd_client_handle **och_p = NULL;
415 __u64 *och_usecount = NULL;
418 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n",
419 inode->i_ino, inode->i_generation, inode, file->f_flags);
421 /* don't do anything for / */
422 if (inode->i_sb->s_root == file->f_dentry)
425 if ((file->f_flags+1) & O_ACCMODE)
427 if (file->f_flags & O_TRUNC)
432 * sometimes LUSTRE_IT(it) may not be allocated like opening file by
433 * dentry_open() from GNS stuff.
435 if (!it || !LUSTRE_IT(it)) {
437 rc = ll_intent_alloc(it);
442 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
445 * mdc_intent_lock() didn't get a request ref if there was an open
446 * error, so don't do cleanup on the * request here (bug 3430)
448 if (LUSTRE_IT(it)->it_disposition) {
449 rc = it_open_error(DISP_OPEN_OPEN, it);
454 /* Let's see if we have file open on MDS already. */
455 if (it->it_flags & FMODE_WRITE) {
456 och_p = &lli->lli_mds_write_och;
457 och_usecount = &lli->lli_open_fd_write_count;
458 } else if (it->it_flags & FMODE_EXEC) {
459 och_p = &lli->lli_mds_exec_och;
460 och_usecount = &lli->lli_open_fd_exec_count;
462 och_p = &lli->lli_mds_read_och;
463 och_usecount = &lli->lli_open_fd_read_count;
466 rc = ll_crypto_decrypt_key(inode, it);
470 down(&lli->lli_och_sem);
471 if (*och_p) { /* Open handle is present */
472 if (it_disposition(it, DISP_LOOKUP_POS) && /* Positive lookup */
473 it_disposition(it, DISP_OPEN_OPEN)) { /* & OPEN happened */
474 struct obd_client_handle *och;
475 /* Well, there's extra open request that we do not need,
476 let's close it somehow*/
477 OBD_ALLOC(och, sizeof (struct obd_client_handle));
479 up(&lli->lli_och_sem);
483 ll_och_fill(inode, it, och);
484 /* ll_md_och_close() will free och */
485 ll_md_och_close(ll_i2mdexp(inode), inode, och, 0);
489 rc = ll_local_open(file, it, NULL);
493 LASSERT(*och_usecount == 0);
494 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
496 GOTO(out, rc = -ENOMEM);
499 if (!it || !LUSTRE_IT(it) || !LUSTRE_IT(it)->it_disposition) {
501 * we are going to replace intent here, and that may
502 * possibly change access mode (FMODE_EXEC can only be
503 * set in intent), but I hope it never happens (I was
504 * not able to trigger it yet at least) -- green
507 /* FIXME: FMODE_EXEC is not covered by O_ACCMODE! */
508 LASSERT(!(it->it_flags & FMODE_EXEC));
509 LASSERTF((it->it_flags & O_ACCMODE) ==
510 (oit.it_flags & O_ACCMODE), "Changing intent "
511 "flags %x to incompatible %x\n", it->it_flags,
514 rc = ll_intent_file_open(file, NULL, 0, it);
517 rc = it_open_error(DISP_OPEN_OPEN, it);
519 GOTO(out_och_free, rc);
521 mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle,
522 file->f_dentry->d_inode);
524 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_OPEN);
525 rc = ll_local_open(file, it, *och_p);
526 LASSERTF(rc == 0, "rc = %d\n", rc);
528 up(&lli->lli_och_sem);
531 * must do this outside lli_och_sem lock to prevent deadlock where
532 * different kind of OPEN lock for this same inode gets cancelled by
536 if (!S_ISREG(inode->i_mode))
541 if (file->f_flags & O_LOV_DELAY_CREATE ||
542 !(file->f_mode & FMODE_WRITE)) {
543 CDEBUG(D_INODE, "object creation was delayed\n");
547 file->f_flags &= ~O_LOV_DELAY_CREATE;
550 req = LUSTRE_IT(it)->it_data;
551 ll_intent_drop_lock(it);
552 ll_intent_release(it);
553 ptlrpc_req_finished(req);
555 ll_open_complete(inode);
559 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
560 *och_p = NULL; /* OBD_FREE writes some magic there */
563 up(&lli->lli_och_sem);
569 /* Fills the obdo with the attributes for the inode defined by lsm */
570 int ll_lsm_getattr(struct obd_export *exp, struct lov_stripe_md *lsm,
573 struct ptlrpc_request_set *set;
577 LASSERT(lsm != NULL);
579 memset(oa, 0, sizeof *oa);
580 oa->o_id = lsm->lsm_object_id;
581 oa->o_gr = lsm->lsm_object_gr;
582 oa->o_mode = S_IFREG;
583 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLSIZE |
584 OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
585 OBD_MD_FLCTIME | OBD_MD_FLGROUP;
587 set = ptlrpc_prep_set();
591 rc = obd_getattr_async(exp, oa, lsm, set);
593 rc = ptlrpc_set_wait(set);
594 ptlrpc_set_destroy(set);
599 oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ | OBD_MD_FLMTIME |
600 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
604 static inline void ll_remove_suid(struct inode *inode)
608 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
609 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
611 /* was any of the uid bits set? */
612 mode &= inode->i_mode;
613 if (mode && !capable(CAP_FSETID)) {
614 inode->i_mode &= ~mode;
615 // XXX careful here - we cannot change the size
619 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
621 struct ll_inode_info *lli = ll_i2info(inode);
622 struct lov_stripe_md *lsm = lli->lli_smd;
623 struct obd_export *exp = ll_i2dtexp(inode);
626 struct ldlm_lock *lock;
627 struct lov_stripe_md *lsm;
628 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
629 __u32 stripe, vallen = sizeof(stripe);
633 if (lsm->lsm_stripe_count == 1)
634 GOTO(check, stripe = 0);
636 /* get our offset in the lov */
637 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
639 CERROR("obd_get_info: rc = %d\n", rc);
642 LASSERT(stripe < lsm->lsm_stripe_count);
645 if (lsm->lsm_oinfo[stripe].loi_id != lock->l_resource->lr_name.name[0]||
646 lsm->lsm_oinfo[stripe].loi_gr != lock->l_resource->lr_name.name[2]){
647 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64
648 " inode=%lu/%u (%p)\n",
649 lsm->lsm_oinfo[stripe].loi_id,
650 lsm->lsm_oinfo[stripe].loi_gr,
651 inode->i_ino, inode->i_generation, inode);
652 return -ELDLM_NO_LOCK_DATA;
658 /* Flush the page cache for an extent as its canceled. When we're on an LOV,
659 * we get a lock cancellation for each stripe, so we have to map the obd's
660 * region back onto the stripes in the file that it held.
662 * No one can dirty the extent until we've finished our work and they can
663 * enqueue another lock. The DLM protects us from ll_file_read/write here,
664 * but other kernel actors could have pages locked.
666 * Called with the DLM lock held. */
667 void ll_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm,
668 struct ldlm_lock *lock, __u32 stripe)
670 ldlm_policy_data_t tmpex;
671 unsigned long start, end, count, skip, i, j;
673 int rc, rc2, discard = lock->l_flags & LDLM_FL_DISCARD_DATA;
674 struct lustre_handle lockh;
677 memcpy(&tmpex, &lock->l_policy_data, sizeof(tmpex));
678 CDEBUG(D_INODE|D_PAGE, "inode %lu(%p) ["LPU64"->"LPU64"] size: %llu\n",
679 inode->i_ino, inode, tmpex.l_extent.start, tmpex.l_extent.end,
682 /* our locks are page granular thanks to osc_enqueue, we invalidate the
684 LASSERT((tmpex.l_extent.start & ~PAGE_CACHE_MASK) == 0);
685 LASSERT(((tmpex.l_extent.end + 1) & ~PAGE_CACHE_MASK) == 0);
689 start = tmpex.l_extent.start >> PAGE_CACHE_SHIFT;
690 end = tmpex.l_extent.end >> PAGE_CACHE_SHIFT;
691 if (lsm->lsm_stripe_count > 1) {
692 count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT;
693 skip = (lsm->lsm_stripe_count - 1) * count;
694 start += start/count * skip + stripe * count;
696 end += end/count * skip + stripe * count;
698 if (end < tmpex.l_extent.end >> PAGE_CACHE_SHIFT)
701 i = inode->i_size ? (inode->i_size - 1) >> PAGE_CACHE_SHIFT : 0;
705 CDEBUG(D_INODE|D_PAGE, "walking page indices start: %lu j: %lu "
706 "count: %lu skip: %lu end: %lu%s\n", start, start % count,
707 count, skip, end, discard ? " (DISCARDING)" : "");
709 /* walk through the vmas on the inode and tear down mmaped pages that
710 * intersect with the lock. this stops immediately if there are no
711 * mmap()ed regions of the file. This is not efficient at all and
712 * should be short lived. We'll associate mmap()ed pages with the lock
713 * and will be able to find them directly */
715 for (i = start; i <= end; i += (j + skip)) {
716 j = min(count - (i % count), end - i + 1);
718 LASSERT(inode->i_mapping);
719 if (ll_teardown_mmaps(inode->i_mapping,
720 (__u64)i << PAGE_CACHE_SHIFT,
721 ((__u64)(i+j) << PAGE_CACHE_SHIFT) - 1) )
725 /* this is the simplistic implementation of page eviction at
726 * cancelation. It is careful to get races with other page
727 * lockers handled correctly. fixes from bug 20 will make it
728 * more efficient by associating locks with pages and with
729 * batching writeback under the lock explicitly. */
730 for (i = start, j = start % count; i <= end;
731 j++, i++, tmpex.l_extent.start += PAGE_CACHE_SIZE) {
733 CDEBUG(D_PAGE, "skip index %lu to %lu\n", i, i + skip);
739 LASSERTF(tmpex.l_extent.start< lock->l_policy_data.l_extent.end,
740 LPU64" >= "LPU64" start %lu i %lu end %lu\n",
741 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
744 if (!mapping_has_pages(inode->i_mapping)) {
745 CDEBUG(D_INODE|D_PAGE, "nothing left\n");
751 page = find_get_page(inode->i_mapping, i);
754 LL_CDEBUG_PAGE(D_PAGE, page, "lock page idx %lu ext "LPU64"\n",
755 i, tmpex.l_extent.start);
758 /* page->mapping to check with racing against teardown */
759 if (!discard && clear_page_dirty_for_io(page)) {
760 rc = ll_call_writepage(inode, page);
762 CERROR("writepage of page %p failed: %d\n",
764 /* either waiting for io to complete or reacquiring
765 * the lock that the failed writepage released */
769 tmpex.l_extent.end = tmpex.l_extent.start + PAGE_CACHE_SIZE - 1;
770 /* check to see if another DLM lock covers this page */
771 rc2 = ldlm_lock_match(lock->l_resource->lr_namespace,
772 LDLM_FL_BLOCK_GRANTED|LDLM_FL_CBPENDING |
774 &lock->l_resource->lr_name, LDLM_EXTENT,
775 &tmpex, LCK_PR | LCK_PW, &lockh);
776 if (rc2 == 0 && page->mapping != NULL) {
777 // checking again to account for writeback's lock_page()
778 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
779 ll_ra_accounting(page, inode->i_mapping);
780 ll_truncate_complete_page(page);
783 page_cache_release(page);
785 LASSERTF(tmpex.l_extent.start <=
786 (lock->l_policy_data.l_extent.end == ~0ULL ? ~0ULL :
787 lock->l_policy_data.l_extent.end + 1),
788 "loop too long "LPU64" > "LPU64" start %lu i %lu end %lu\n",
789 tmpex.l_extent.start, lock->l_policy_data.l_extent.end,
794 static int ll_extent_lock_callback(struct ldlm_lock *lock,
795 struct ldlm_lock_desc *new, void *data,
798 struct lustre_handle lockh = { 0 };
802 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
803 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
808 case LDLM_CB_BLOCKING:
809 ldlm_lock2handle(lock, &lockh);
810 rc = ldlm_cli_cancel(&lockh);
812 CERROR("ldlm_cli_cancel failed: %d\n", rc);
814 case LDLM_CB_CANCELING: {
816 struct ll_inode_info *lli;
817 struct lov_stripe_md *lsm;
821 /* This lock wasn't granted, don't try to evict pages */
822 if (lock->l_req_mode != lock->l_granted_mode)
825 inode = ll_inode_from_lock(lock);
828 lli = ll_i2info(inode);
831 if (lli->lli_smd == NULL)
835 stripe = ll_lock_to_stripe_offset(inode, lock);
838 ll_pgcache_remove_extent(inode, lsm, lock, stripe);
840 down(&lli->lli_size_sem);
841 lock_res_and_lock(lock);
842 kms = ldlm_extent_shift_kms(lock,
843 lsm->lsm_oinfo[stripe].loi_kms);
845 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
846 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
847 lsm->lsm_oinfo[stripe].loi_kms, kms);
848 lsm->lsm_oinfo[stripe].loi_kms = kms;
849 unlock_res_and_lock(lock);
850 up(&lli->lli_size_sem);
851 //ll_try_done_writing(inode);
864 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
866 /* XXX ALLOCATE - 160 bytes */
867 struct inode *inode = ll_inode_from_lock(lock);
868 struct ll_inode_info *lli = ll_i2info(inode);
869 struct lustre_handle lockh = { 0 };
874 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
875 LDLM_FL_BLOCK_CONV)) {
876 LBUG(); /* not expecting any blocked async locks yet */
877 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
879 ldlm_lock_dump(D_OTHER, lock, 0);
880 ldlm_reprocess_all(lock->l_resource);
884 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
886 stripe = ll_lock_to_stripe_offset(inode, lock);
890 if (lock->l_lvb_len) {
891 struct lov_stripe_md *lsm = lli->lli_smd;
893 lvb = lock->l_lvb_data;
894 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
897 lock_res_and_lock(lock);
898 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
899 kms = ldlm_extent_shift_kms(NULL, kms);
900 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
901 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
902 lsm->lsm_oinfo[stripe].loi_kms, kms);
903 lsm->lsm_oinfo[stripe].loi_kms = kms;
904 unlock_res_and_lock(lock);
910 wake_up(&lock->l_waitq);
912 ldlm_lock2handle(lock, &lockh);
913 ldlm_lock_decref(&lockh, LCK_PR);
918 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
920 struct ptlrpc_request *req = reqp;
921 struct inode *inode = ll_inode_from_lock(lock);
922 struct ll_inode_info *lli;
924 struct lov_stripe_md *lsm;
925 int rc, size = sizeof(*lvb), stripe;
929 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
930 lli = ll_i2info(inode);
932 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
936 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
938 /* First, find out which stripe index this lock corresponds to. */
939 stripe = ll_lock_to_stripe_offset(inode, lock);
941 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
943 rc = lustre_pack_reply(req, 1, &size, NULL);
945 CERROR("lustre_pack_reply: %d\n", rc);
949 lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb));
950 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe].loi_kms;
951 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
952 lvb->lvb_atime = LTIME_S(inode->i_atime);
953 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
955 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64,
956 inode->i_size, stripe, lvb->lvb_size);
962 /* These errors are normal races, so we don't want to fill the console
963 * with messages by calling ptlrpc_error() */
964 if (rc == -ELDLM_NO_LOCK_DATA)
965 lustre_pack_reply(req, 0, NULL, NULL);
971 /* NB: lov_merge_size will prefer locally cached writes if they extend the
972 * file (because it prefers KMS over RSS when larger) */
973 int ll_glimpse_size(struct inode *inode)
975 struct ll_inode_info *lli = ll_i2info(inode);
976 struct ll_sb_info *sbi = ll_i2sbi(inode);
977 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
978 struct lustre_handle lockh = { 0 };
979 int rc, flags = LDLM_FL_HAS_INTENT;
982 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
984 rc = obd_enqueue(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, &policy,
985 LCK_PR, &flags, ll_extent_lock_callback,
986 ldlm_completion_ast, ll_glimpse_callback, inode,
987 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
992 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
993 RETURN(rc > 0 ? -EIO : rc);
996 down(&lli->lli_size_sem);
997 inode->i_size = lov_merge_size(lli->lli_smd, 0);
998 inode->i_blocks = lov_merge_blocks(lli->lli_smd);
999 up(&lli->lli_size_sem);
1001 LTIME_S(inode->i_mtime) = lov_merge_mtime(lli->lli_smd,
1002 LTIME_S(inode->i_mtime));
1004 CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64", blocks: "LPU64"\n",
1005 (__u64)inode->i_size, (__u64)inode->i_blocks);
1007 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
1011 void ll_stime_record(struct ll_sb_info *sbi, struct timeval *start,
1012 struct obd_service_time *stime)
1014 struct timeval stop;
1015 do_gettimeofday(&stop);
1017 spin_lock(&sbi->ll_lock);
1018 lprocfs_stime_record(stime, &stop, start);
1019 spin_unlock(&sbi->ll_lock);
1022 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
1023 struct lov_stripe_md *lsm, int mode,
1024 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
1025 int ast_flags, struct obd_service_time *stime)
1027 struct ll_inode_info *lli = ll_i2info(inode);
1028 struct ll_sb_info *sbi = ll_i2sbi(inode);
1029 struct timeval start;
1033 LASSERT(lockh->cookie == 0);
1035 /* XXX phil: can we do this? won't it screw the file size up? */
1036 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1037 (sbi->ll_flags & LL_SBI_NOLCK))
1040 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
1041 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
1043 do_gettimeofday(&start);
1044 rc = obd_enqueue(sbi->ll_dt_exp, lsm, LDLM_EXTENT, policy, mode,
1045 &ast_flags, ll_extent_lock_callback,
1046 ldlm_completion_ast, ll_glimpse_callback, inode,
1047 sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh);
1051 ll_stime_record(sbi, &start, stime);
1053 if (policy->l_extent.start == 0 &&
1054 policy->l_extent.end == OBD_OBJECT_EOF) {
1055 /* vmtruncate()->ll_truncate() first sets the i_size and then
1056 * the kms under both a DLM lock and the i_sem. If we don't
1057 * get the i_sem here we can match the DLM lock and reset
1058 * i_size from the kms before the truncating path has updated
1059 * the kms. generic_file_write can then trust the stale i_size
1060 * when doing appending writes and effectively cancel the
1061 * result of the truncate. Getting the i_sem after the enqueue
1062 * maintains the DLM -> i_sem acquiry order. */
1063 down(&lli->lli_size_sem);
1064 inode->i_size = lov_merge_size(lsm, 1);
1065 up(&lli->lli_size_sem);
1069 LTIME_S(inode->i_mtime) =
1070 lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1076 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1077 struct lov_stripe_md *lsm, int mode,
1078 struct lustre_handle *lockh)
1080 struct ll_sb_info *sbi = ll_i2sbi(inode);
1084 /* XXX phil: can we do this? won't it screw the file size up? */
1085 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1086 (sbi->ll_flags & LL_SBI_NOLCK))
1089 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1094 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1097 struct inode *inode = file->f_dentry->d_inode;
1098 struct ll_inode_info *lli = ll_i2info(inode);
1099 struct lov_stripe_md *lsm = lli->lli_smd;
1100 struct ll_lock_tree tree;
1101 struct ll_lock_tree_node *node;
1106 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1107 inode->i_ino, inode->i_generation, inode, count, *ppos);
1109 /* "If nbyte is 0, read() will return 0 and have no other results."
1110 * -- Single Unix Spec */
1114 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_READ_BYTES,
1120 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1123 tree.lt_fd = file->private_data;
1125 rc = ll_tree_lock(&tree, node, inode, buf, count,
1126 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1130 down(&lli->lli_size_sem);
1131 kms = lov_merge_size(lsm, 1);
1132 if (*ppos + count - 1 > kms) {
1133 /* A glimpse is necessary to determine whether we return a short
1134 * read or some zeroes at the end of the buffer */
1135 up(&lli->lli_size_sem);
1136 retval = ll_glimpse_size(inode);
1140 inode->i_size = kms;
1141 up(&lli->lli_size_sem);
1144 CDEBUG(D_INFO, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1145 inode->i_ino, count, *ppos, inode->i_size);
1147 /* turn off the kernel's read-ahead */
1148 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1151 file->f_ra.ra_pages = 0;
1153 retval = generic_file_read(file, buf, count, ppos);
1156 ll_tree_unlock(&tree, inode);
1161 * Write to a file (through the page cache).
1163 static ssize_t ll_file_write(struct file *file, const char *buf,
1164 size_t count, loff_t *ppos)
1166 struct inode *inode = file->f_dentry->d_inode;
1167 loff_t maxbytes = ll_file_maxbytes(inode);
1168 struct ll_lock_tree tree;
1169 struct ll_lock_tree_node *node;
1174 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1175 inode->i_ino, inode->i_generation, inode, count, *ppos);
1177 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1179 /* POSIX, but surprised the VFS doesn't check this already */
1183 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1184 * called on the file, don't fail the below assertion (bug 2388). */
1185 if (file->f_flags & O_LOV_DELAY_CREATE &&
1186 ll_i2info(inode)->lli_smd == NULL)
1189 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1191 if (file->f_flags & O_APPEND)
1192 node = ll_node_from_inode(inode, 0, OBD_OBJECT_EOF, LCK_PW);
1194 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1,
1198 RETURN(PTR_ERR(node));
1200 tree.lt_fd = file->private_data;
1202 rc = ll_tree_lock(&tree, node, inode, buf, count,
1203 file->f_flags & O_NONBLOCK ? LDLM_FL_BLOCK_NOWAIT :0);
1207 /* this is ok, g_f_w will overwrite this under i_sem if it races
1208 * with a local truncate, it just makes our maxbyte checking easier */
1209 if (file->f_flags & O_APPEND)
1210 *ppos = inode->i_size;
1212 if (*ppos >= maxbytes) {
1213 if (count || *ppos > maxbytes) {
1214 send_sig(SIGXFSZ, current, 0);
1215 GOTO(out, retval = -EFBIG);
1218 if (*ppos + count > maxbytes)
1219 count = maxbytes - *ppos;
1221 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1222 inode->i_ino, count, *ppos);
1224 /* mark open handle dirty */
1225 set_bit(LLI_F_DIRTY_HANDLE, &(ll_i2info(inode)->lli_flags));
1227 /* generic_file_write handles O_APPEND after getting i_sem */
1228 retval = generic_file_write(file, buf, count, ppos);
1231 ll_tree_unlock(&tree, inode);
1232 /* serialize with mmap/munmap/mremap */
1233 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats, LPROC_LL_WRITE_BYTES,
1234 retval > 0 ? retval : 0);
1238 static int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
1239 int flags, struct lov_user_md *lum,
1242 struct ll_inode_info *lli = ll_i2info(inode);
1244 struct obd_export *exp = ll_i2dtexp(inode);
1245 struct lov_stripe_md *lsm;
1246 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
1247 struct ptlrpc_request *req = NULL;
1249 struct lustre_md md;
1250 struct obd_client_handle *och;
1254 if ((file->f_flags+1) & O_ACCMODE)
1256 if (file->f_flags & O_TRUNC)
1259 down(&lli->lli_open_sem);
1262 up(&lli->lli_open_sem);
1263 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
1268 f = get_empty_filp();
1272 f->f_dentry = file->f_dentry;
1273 f->f_vfsmnt = file->f_vfsmnt;
1276 rc = ll_intent_alloc(&oit);
1280 rc = ll_intent_file_open(f, lum, lum_size, &oit);
1283 if (it_disposition(&oit, DISP_LOOKUP_NEG))
1286 req = LUSTRE_IT(&oit)->it_data;
1287 rc = LUSTRE_IT(&oit)->it_status;
1292 rc = mdc_req2lustre_md(ll_i2mdexp(inode), req, 1, exp, &md);
1295 ll_update_inode(f->f_dentry->d_inode, &md);
1297 OBD_ALLOC(och, sizeof(struct obd_client_handle));
1298 rc = ll_local_open(f, &oit, och);
1299 if (rc) { /* Actually ll_local_open cannot fail! */
1302 if (LUSTRE_IT(&oit)->it_lock_mode) {
1303 ldlm_lock_decref_and_cancel((struct lustre_handle *)
1304 &LUSTRE_IT(&oit)->it_lock_handle,
1305 LUSTRE_IT(&oit)->it_lock_mode);
1306 LUSTRE_IT(&oit)->it_lock_mode = 0;
1309 ll_intent_release(&oit);
1311 /* ll_file_release will decrease the count, but won't free anything
1312 because we have at least one more reference coming from actual open
1314 down(&lli->lli_och_sem);
1315 lli->lli_open_fd_write_count++;
1316 up(&lli->lli_och_sem);
1317 rc = ll_file_release(f->f_dentry->d_inode, f);
1319 /* Now also destroy our supplemental och */
1320 ll_md_och_close(ll_i2mdexp(inode), f->f_dentry->d_inode, och, 0);
1323 ll_intent_release(&oit);
1326 up(&lli->lli_open_sem);
1328 ptlrpc_req_finished(req);
1332 static int ll_lov_setea(struct inode *inode, struct file *file,
1335 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
1336 struct lov_user_md *lump;
1337 int lum_size = sizeof(struct lov_user_md) +
1338 sizeof(struct lov_user_ost_data);
1342 if (!capable (CAP_SYS_ADMIN))
1345 OBD_ALLOC(lump, lum_size);
1349 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
1351 OBD_FREE(lump, lum_size);
1355 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1357 OBD_FREE(lump, lum_size);
1361 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1364 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1366 int flags = FMODE_WRITE;
1369 /* Bug 1152: copy properly when this is no longer true */
1370 LASSERT(sizeof(lum) == sizeof(*lump));
1371 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1372 rc = copy_from_user(&lum, lump, sizeof(lum));
1376 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1378 put_user(0, &lump->lmm_stripe_count);
1379 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
1380 0, ll_i2info(inode)->lli_smd, lump);
1385 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1387 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1392 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
1396 static int ll_get_grouplock(struct inode *inode, struct file *file,
1399 struct ll_file_data *fd = file->private_data;
1400 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1401 .end = OBD_OBJECT_EOF}};
1402 struct lustre_handle lockh = { 0 };
1403 struct ll_inode_info *lli = ll_i2info(inode);
1404 struct lov_stripe_md *lsm = lli->lli_smd;
1408 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
1412 policy.l_extent.gid = arg;
1413 if (file->f_flags & O_NONBLOCK)
1414 flags = LDLM_FL_BLOCK_NOWAIT;
1416 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags,
1417 &ll_i2sbi(inode)->ll_grouplock_stime);
1421 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
1423 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
1428 static int ll_put_grouplock(struct inode *inode, struct file *file,
1431 struct ll_file_data *fd = file->private_data;
1432 struct ll_inode_info *lli = ll_i2info(inode);
1433 struct lov_stripe_md *lsm = lli->lli_smd;
1437 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1438 /* Ugh, it's already unlocked. */
1442 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
1445 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
1447 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
1452 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
1457 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
1460 struct ll_file_data *fd = file->private_data;
1461 struct ll_sb_info *sbi = ll_i2sbi(inode);
1465 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
1466 inode->i_generation, inode, cmd);
1468 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
1471 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
1473 case LL_IOC_GETFLAGS:
1474 /* Get the current value of the file flags */
1475 return put_user(fd->fd_flags, (int *)arg);
1476 case LL_IOC_SETFLAGS:
1477 case LL_IOC_CLRFLAGS:
1478 /* Set or clear specific file flags */
1479 /* XXX This probably needs checks to ensure the flags are
1480 * not abused, and to handle any flag side effects.
1482 if (get_user(flags, (int *) arg))
1485 if (cmd == LL_IOC_SETFLAGS)
1486 fd->fd_flags |= flags;
1488 fd->fd_flags &= ~flags;
1490 case LL_IOC_LOV_SETSTRIPE:
1491 RETURN(ll_lov_setstripe(inode, file, arg));
1492 case LL_IOC_LOV_SETEA:
1493 RETURN(ll_lov_setea(inode, file, arg));
1494 case IOC_MDC_SHOWFID: {
1495 struct lustre_id *idp = (struct lustre_id *)arg;
1496 struct lustre_id id;
1500 filename = getname((const char *)arg);
1501 if (IS_ERR(filename))
1502 RETURN(PTR_ERR(filename));
1504 ll_inode2id(&id, inode);
1506 rc = ll_get_fid(sbi->ll_md_exp, &id, filename, &id);
1508 GOTO(out_filename, rc);
1510 rc = copy_to_user(idp, &id, sizeof(*idp));
1512 GOTO(out_filename, rc = -EFAULT);
1519 case LL_IOC_KEY_TYPE: {
1520 struct obd_ioctl_data *data;
1523 int typelen, rc, len = 0;
1525 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1530 type = data->ioc_inlbuf1;
1531 typelen = data->ioc_inllen1;
1534 CDEBUG(D_INFO, "LL_IOC_KEY_TYPE missing filename\n");
1535 GOTO(out, rc = -EINVAL);
1537 ll_set_sb_gksinfo(inode->i_sb, type);
1540 obd_ioctl_freedata(buf, len);
1544 case LL_IOC_LOV_GETSTRIPE:
1545 RETURN(ll_lov_getstripe(inode, arg));
1546 case EXT3_IOC_GETFLAGS:
1547 case EXT3_IOC_SETFLAGS:
1548 RETURN( ll_iocontrol(inode, file, cmd, arg) );
1549 case LL_IOC_GROUP_LOCK:
1550 RETURN(ll_get_grouplock(inode, file, arg));
1551 case LL_IOC_GROUP_UNLOCK:
1552 RETURN(ll_put_grouplock(inode, file, arg));
1553 case EXT3_IOC_GETVERSION_OLD:
1554 case EXT3_IOC_GETVERSION:
1555 return put_user(inode->i_generation, (int *) arg);
1556 /* We need to special case any other ioctls we want to handle,
1557 * to send them to the MDS/OST as appropriate and to properly
1558 * network encode the arg field.
1559 case EXT2_IOC_GETVERSION_OLD:
1560 case EXT2_IOC_GETVERSION_NEW:
1561 case EXT2_IOC_SETVERSION_OLD:
1562 case EXT2_IOC_SETVERSION_NEW:
1563 case EXT3_IOC_SETVERSION_OLD:
1564 case EXT3_IOC_SETVERSION:
1566 case LL_IOC_FLUSH_CRED:
1567 RETURN(ll_flush_cred(inode));
1569 RETURN( obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
1574 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
1576 struct inode *inode = file->f_dentry->d_inode;
1577 struct ll_file_data *fd = file->private_data;
1578 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1579 struct lustre_handle lockh = {0};
1582 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),to=%llu\n", inode->i_ino,
1583 inode->i_generation, inode,
1584 offset + ((origin==2) ? inode->i_size : file->f_pos));
1586 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_LLSEEK);
1587 if (origin == 2) { /* SEEK_END */
1588 ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF }};
1589 struct ll_inode_info *lli = ll_i2info(inode);
1590 int nonblock = 0, rc;
1592 if (file->f_flags & O_NONBLOCK)
1593 nonblock = LDLM_FL_BLOCK_NOWAIT;
1595 rc = ll_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh,
1596 nonblock, &ll_i2sbi(inode)->ll_seek_stime);
1600 down(&lli->lli_size_sem);
1601 offset += inode->i_size;
1602 up(&lli->lli_size_sem);
1603 } else if (origin == 1) { /* SEEK_CUR */
1604 offset += file->f_pos;
1608 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
1609 if (offset != file->f_pos) {
1610 file->f_pos = offset;
1611 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1613 file->f_version = ++event;
1620 ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh);
1624 int ll_fsync(struct file *file, struct dentry *dentry, int data)
1626 struct inode *inode = dentry->d_inode;
1627 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1628 struct lustre_id id;
1629 struct ptlrpc_request *req;
1632 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
1633 inode->i_generation, inode);
1635 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_FSYNC);
1637 /* fsync's caller has already called _fdata{sync,write}, we want
1638 * that IO to finish before calling the osc and mdc sync methods */
1639 rc = filemap_fdatawait(inode->i_mapping);
1641 ll_inode2id(&id, inode);
1642 err = md_sync(ll_i2sbi(inode)->ll_md_exp, &id, &req);
1646 ptlrpc_req_finished(req);
1649 struct obdo *oa = obdo_alloc();
1652 RETURN(rc ? rc : -ENOMEM);
1654 oa->o_id = lsm->lsm_object_id;
1655 oa->o_gr = lsm->lsm_object_gr;
1656 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
1658 obdo_from_inode(oa, inode, (OBD_MD_FLTYPE | OBD_MD_FLATIME |
1659 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1662 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
1672 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
1674 struct inode *inode = file->f_dentry->d_inode;
1675 struct ll_inode_info *li = ll_i2info(inode);
1676 struct ll_sb_info *sbi = ll_i2sbi(inode);
1677 struct obd_device *obddev;
1678 struct ldlm_res_id res_id =
1679 { .name = {id_fid(&li->lli_id), id_group(&li->lli_id), LDLM_FLOCK} };
1680 struct lustre_handle lockh = {0};
1681 ldlm_policy_data_t flock;
1682 ldlm_mode_t mode = 0;
1687 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
1688 inode->i_ino, file_lock);
1690 flock.l_flock.pid = file_lock->fl_pid;
1691 flock.l_flock.start = file_lock->fl_start;
1692 flock.l_flock.end = file_lock->fl_end;
1694 switch (file_lock->fl_type) {
1699 /* An unlock request may or may not have any relation to
1700 * existing locks so we may not be able to pass a lock handle
1701 * via a normal ldlm_lock_cancel() request. The request may even
1702 * unlock a byte range in the middle of an existing lock. In
1703 * order to process an unlock request we need all of the same
1704 * information that is given with a normal read or write record
1705 * lock request. To avoid creating another ldlm unlock (cancel)
1706 * message we'll treat a LCK_NL flock request as an unlock. */
1713 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
1728 flags = LDLM_FL_BLOCK_NOWAIT;
1734 flags = LDLM_FL_TEST_LOCK;
1735 /* Save the old mode so that if the mode in the lock changes we
1736 * can decrement the appropriate reader or writer refcount. */
1737 file_lock->fl_type = mode;
1740 CERROR("unknown fcntl lock command: %d\n", cmd);
1744 CDEBUG(D_DLMTRACE, "inode=%lu, pid="LPU64", flags=%#x, mode=%u, "
1745 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
1746 flags, mode, flock.l_flock.start, flock.l_flock.end);
1748 obddev = md_get_real_obd(sbi->ll_md_exp, &li->lli_id);
1749 rc = ldlm_cli_enqueue(obddev->obd_self_export, NULL,
1750 obddev->obd_namespace,
1751 res_id, LDLM_FLOCK, &flock, mode, &flags,
1752 NULL, ldlm_flock_completion_ast, NULL, file_lock,
1753 NULL, 0, NULL, &lockh);
1757 int ll_inode_revalidate_it(struct dentry *dentry)
1759 struct lookup_intent oit = { .it_op = IT_GETATTR };
1760 struct inode *inode = dentry->d_inode;
1761 struct ptlrpc_request *req = NULL;
1762 struct ll_inode_info *lli;
1763 struct lov_stripe_md *lsm;
1764 struct ll_sb_info *sbi;
1765 struct lustre_id id;
1770 CERROR("REPORT THIS LINE TO PETER\n");
1774 sbi = ll_i2sbi(inode);
1776 ll_inode2id(&id, inode);
1777 lli = ll_i2info(inode);
1778 LASSERT(id_fid(&id) != 0);
1780 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), name=%s(%p)\n",
1781 inode->i_ino, inode->i_generation, inode, dentry->d_name.name,
1784 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
1785 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_REVALIDATE);
1788 rc = ll_intent_alloc(&oit);
1792 rc = ll_crypto_init_it_key(inode, &oit);
1796 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
1797 &oit, 0, &req, ll_mdc_blocking_ast);
1801 rc = revalidate_it_finish(req, 1, &oit, dentry);
1806 ll_lookup_finish_locks(&oit, dentry);
1808 if (!LLI_HAVE_FLSIZE(inode)) {
1809 /* if object not yet allocated, don't validate size */
1812 /* ll_glimpse_size() will prefer locally cached
1813 * writes if they extend the file */
1814 rc = ll_glimpse_size(inode);
1819 ll_intent_release(&oit);
1821 ptlrpc_req_finished(req);
1825 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
1826 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
1829 struct inode *inode = de->d_inode;
1830 struct ll_inode_info *lli = ll_i2info(inode);
1832 res = ll_inode_revalidate_it(de);
1833 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETATTR);
1838 stat->ino = inode->i_ino;
1839 stat->mode = inode->i_mode;
1840 stat->nlink = inode->i_nlink;
1841 stat->uid = inode->i_uid;
1842 stat->gid = inode->i_gid;
1843 stat->atime = inode->i_atime;
1844 stat->mtime = inode->i_mtime;
1845 stat->ctime = inode->i_ctime;
1846 stat->blksize = inode->i_blksize;
1848 down(&lli->lli_size_sem);
1849 stat->size = inode->i_size;
1850 stat->blocks = inode->i_blocks;
1851 up(&lli->lli_size_sem);
1853 stat->rdev = kdev_t_to_nr(inode->i_rdev);
1854 stat->dev = id_group(&ll_i2info(inode)->lli_id);
1860 int ll_setxattr_internal(struct inode *inode, const char *name,
1861 const void *value, size_t size, int flags,
1864 struct ll_sb_info *sbi = ll_i2sbi(inode);
1865 struct ptlrpc_request *request = NULL;
1866 struct mdc_op_data op_data;
1869 int rc = 0, key_size = 0;
1872 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu\n", inode->i_ino);
1873 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_SETXATTR);
1875 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
1876 RETURN(-EOPNOTSUPP);
1878 memset(&attr, 0x0, sizeof(attr));
1879 attr.ia_valid |= valid;
1880 attr.ia_attr_flags = flags;
1882 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1883 rc = ll_crypto_get_mac(inode, NULL, (void *)value, size,
1887 ll_prepare_mdc_data(&op_data, inode, NULL, NULL, 0, 0);
1889 rc = md_setattr(sbi->ll_md_exp, &op_data, &attr,
1890 (void*) name, strnlen(name, XATTR_NAME_MAX) + 1,
1891 (void*) value, size, key, key_size, &request);
1893 if (key && key_size)
1894 OBD_FREE(key, key_size);
1896 CERROR("md_setattr fails: rc = %d\n", rc);
1900 ptlrpc_req_finished(request);
1904 int ll_setxattr(struct dentry *dentry, const char *name, const void *value,
1905 size_t size, int flags)
1908 struct posix_acl *acl;
1909 struct ll_inode_info *lli;
1912 rc = ll_setxattr_internal(dentry->d_inode, name, value, size,
1915 /* update inode's acl info */
1916 if (rc == 0 && strcmp(name, XATTR_NAME_ACL_ACCESS) == 0) {
1918 acl = posix_acl_from_xattr(value, size);
1920 CERROR("convert from xattr to acl error: %ld",
1924 error = posix_acl_valid(acl);
1926 CERROR("acl valid error: %d", error);
1927 posix_acl_release(acl);
1935 lli = ll_i2info(dentry->d_inode);
1936 spin_lock(&lli->lli_lock);
1937 if (lli->lli_posix_acl != NULL)
1938 posix_acl_release(lli->lli_posix_acl);
1939 lli->lli_posix_acl = acl;
1940 spin_unlock(&lli->lli_lock);
1947 int ll_removexattr(struct dentry *dentry, const char *name)
1949 return ll_setxattr_internal(dentry->d_inode, name, NULL, 0, 0,
1953 int ll_getxattr_internal(struct inode *inode, const char *name,
1954 void *value, size_t size, __u64 valid)
1956 struct ptlrpc_request *request = NULL;
1957 struct ll_sb_info *sbi = ll_i2sbi(inode);
1958 struct lustre_id id;
1959 struct mds_body *body;
1964 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_GETXATTR);
1966 if (sbi->ll_remote && !strcmp(name, XATTR_NAME_ACL_ACCESS))
1967 RETURN(-EOPNOTSUPP);
1969 ll_inode2id(&id, inode);
1970 rc = md_getattr(sbi->ll_md_exp, &id, valid, name, NULL, 0,
1973 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1974 CERROR("rc = %d\n", rc);
1978 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
1979 LASSERT(body != NULL);
1980 LASSERT_REPSWABBED(request, 0);
1982 ea_size = body->eadatasize;
1984 GOTO(out, rc = ea_size);
1986 LASSERT(ea_size <= request->rq_repmsg->buflens[1]);
1987 ea_data = lustre_msg_buf(request->rq_repmsg, 1, ea_size);
1988 LASSERT(ea_data != NULL);
1989 LASSERT_REPSWABBED(request, 1);
1992 memcpy(value, ea_data, ea_size);
1996 ptlrpc_req_finished(request);
2000 int ll_getxattr(struct dentry *dentry, const char *name, void *value,
2003 return ll_getxattr_internal(dentry->d_inode, name,
2004 value, size, OBD_MD_FLXATTR);
2007 int ll_listxattr(struct dentry *dentry, char *list, size_t size)
2009 return ll_getxattr_internal(dentry->d_inode, NULL, list, size,
2010 OBD_MD_FLXATTRLIST);
2014 * XXX We could choose not to check DLM lock. Leave the decision
2015 * to remote acl handling.
2018 lustre_check_acl(struct inode *inode, int mask)
2020 struct lookup_intent it = { .it_op = IT_GETATTR };
2021 struct dentry de = { .d_inode = inode };
2022 struct ll_sb_info *sbi;
2023 struct lustre_id id;
2024 struct ptlrpc_request *req = NULL;
2025 struct ll_inode_info *lli = ll_i2info(inode);
2026 struct posix_acl *acl;
2030 sbi = ll_i2sbi(inode);
2031 ll_inode2id(&id, inode);
2033 if (ll_intent_alloc(&it))
2036 rc = md_intent_lock(sbi->ll_md_exp, &id, NULL, 0, NULL, 0, &id,
2037 &it, 0, &req, ll_mdc_blocking_ast);
2039 ll_intent_free(&it);
2043 rc = revalidate_it_finish(req, 1, &it, &de);
2045 ll_intent_release(&it);
2049 if (sbi->ll_remote) {
2050 rc = ll_remote_acl_permission(inode, mask);
2052 spin_lock(&lli->lli_lock);
2053 acl = posix_acl_dup(ll_i2info(inode)->lli_posix_acl);
2054 spin_unlock(&lli->lli_lock);
2059 rc = posix_acl_permission(inode, acl, mask);
2060 posix_acl_release(acl);
2064 ll_lookup_finish_locks(&it, &de);
2065 ll_intent_free(&it);
2069 ptlrpc_req_finished(req);
2074 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
2076 return generic_permission(inode, mask, lustre_check_acl);
2079 struct file_operations ll_file_operations = {
2080 .read = ll_file_read,
2081 .write = ll_file_write,
2082 .ioctl = ll_file_ioctl,
2083 .open = ll_file_open,
2084 .release = ll_file_release,
2085 .mmap = ll_file_mmap,
2086 .llseek = ll_file_seek,
2087 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
2088 .sendfile = generic_file_sendfile,
2091 .lock = ll_file_flock
2094 struct inode_operations ll_file_inode_operations = {
2095 .setattr = ll_setattr,
2096 .truncate = ll_truncate,
2097 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
2098 .getattr = ll_getattr,
2100 .revalidate_it = ll_inode_revalidate_it,
2102 .setxattr = ll_setxattr,
2103 .getxattr = ll_getxattr,
2104 .listxattr = ll_listxattr,
2105 .removexattr = ll_removexattr,
2106 .permission = ll_inode_permission,