1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
40 * Author: Andreas Dilger <adilger@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_LLITE
44 #include <lustre_dlm.h>
45 #include <lustre_lite.h>
46 #include <lustre_mdc.h>
47 #include <linux/pagemap.h>
48 #include <linux/file.h>
49 #include "llite_internal.h"
51 /* also used by llite/special.c:ll_special_open() */
52 struct ll_file_data *ll_file_data_get(void)
54 struct ll_file_data *fd;
56 OBD_SLAB_ALLOC_PTR(fd, ll_file_data_slab);
60 static void ll_file_data_put(struct ll_file_data *fd)
63 OBD_SLAB_FREE_PTR(fd, ll_file_data_slab);
66 void ll_pack_inode2opdata(struct inode *inode, struct md_op_data *op_data,
67 struct lustre_handle *fh)
69 op_data->op_fid1 = ll_i2info(inode)->lli_fid;
70 op_data->op_attr.ia_mode = inode->i_mode;
71 op_data->op_attr.ia_atime = inode->i_atime;
72 op_data->op_attr.ia_mtime = inode->i_mtime;
73 op_data->op_attr.ia_ctime = inode->i_ctime;
74 op_data->op_attr.ia_size = i_size_read(inode);
75 op_data->op_attr_blocks = inode->i_blocks;
76 ((struct ll_iattr *)&op_data->op_attr)->ia_attr_flags = inode->i_flags;
77 op_data->op_ioepoch = ll_i2info(inode)->lli_ioepoch;
78 memcpy(&op_data->op_handle, fh, sizeof(op_data->op_handle));
79 op_data->op_capa1 = ll_mdscapa_get(inode);
82 static void ll_prepare_close(struct inode *inode, struct md_op_data *op_data,
83 struct obd_client_handle *och)
87 op_data->op_attr.ia_valid = ATTR_MODE | ATTR_ATIME_SET |
88 ATTR_MTIME_SET | ATTR_CTIME_SET;
90 if (!(och->och_flags & FMODE_WRITE))
93 if (!(ll_i2mdexp(inode)->exp_connect_flags & OBD_CONNECT_SOM) ||
94 !S_ISREG(inode->i_mode))
95 op_data->op_attr.ia_valid |= ATTR_SIZE | ATTR_BLOCKS;
97 ll_epoch_close(inode, op_data, &och, 0);
100 ll_pack_inode2opdata(inode, op_data, &och->och_fh);
104 static int ll_close_inode_openhandle(struct obd_export *md_exp,
106 struct obd_client_handle *och)
108 struct obd_export *exp = ll_i2mdexp(inode);
109 struct md_op_data *op_data;
110 struct ptlrpc_request *req = NULL;
111 struct obd_device *obd = class_exp2obd(exp);
118 * XXX: in case of LMV, is this correct to access
121 CERROR("Invalid MDC connection handle "LPX64"\n",
122 ll_i2mdexp(inode)->exp_handle.h_cookie);
127 * here we check if this is forced umount. If so this is called on
128 * canceling "open lock" and we do not call md_close() in this case, as
129 * it will not be successful, as import is already deactivated.
134 OBD_ALLOC_PTR(op_data);
136 GOTO(out, rc = -ENOMEM); // XXX We leak openhandle and request here.
138 ll_prepare_close(inode, op_data, och);
139 epoch_close = (op_data->op_flags & MF_EPOCH_CLOSE);
140 rc = md_close(md_exp, op_data, och->och_mod, &req);
145 /* This close must have the epoch closed. */
146 LASSERT(exp->exp_connect_flags & OBD_CONNECT_SOM);
147 LASSERT(epoch_close);
148 /* MDS has instructed us to obtain Size-on-MDS attribute from
149 * OSTs and send setattr to back to MDS. */
150 rc = ll_sizeonmds_update(inode, och->och_mod,
151 &och->och_fh, op_data->op_ioepoch);
153 CERROR("inode %lu mdc Size-on-MDS update failed: "
154 "rc = %d\n", inode->i_ino, rc);
158 CERROR("inode %lu mdc close failed: rc = %d\n",
161 ll_finish_md_op_data(op_data);
164 rc = ll_objects_destroy(req, inode);
166 CERROR("inode %lu ll_objects destroy: rc = %d\n",
173 if ((exp->exp_connect_flags & OBD_CONNECT_SOM) && !epoch_close &&
174 S_ISREG(inode->i_mode) && (och->och_flags & FMODE_WRITE)) {
175 ll_queue_done_writing(inode, LLIF_DONE_WRITING);
178 ptlrpc_close_replay_seq(req);
179 md_clear_open_replay_data(md_exp, och);
180 /* Free @och if it is not waiting for DONE_WRITING. */
181 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
184 if (req) /* This is close request */
185 ptlrpc_req_finished(req);
189 int ll_md_real_close(struct inode *inode, int flags)
191 struct ll_inode_info *lli = ll_i2info(inode);
192 struct obd_client_handle **och_p;
193 struct obd_client_handle *och;
198 if (flags & FMODE_WRITE) {
199 och_p = &lli->lli_mds_write_och;
200 och_usecount = &lli->lli_open_fd_write_count;
201 } else if (flags & FMODE_EXEC) {
202 och_p = &lli->lli_mds_exec_och;
203 och_usecount = &lli->lli_open_fd_exec_count;
205 LASSERT(flags & FMODE_READ);
206 och_p = &lli->lli_mds_read_och;
207 och_usecount = &lli->lli_open_fd_read_count;
210 down(&lli->lli_och_sem);
211 if (*och_usecount) { /* There are still users of this handle, so
213 up(&lli->lli_och_sem);
218 up(&lli->lli_och_sem);
220 if (och) { /* There might be a race and somebody have freed this och
222 rc = ll_close_inode_openhandle(ll_i2sbi(inode)->ll_md_exp,
229 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
232 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
233 struct ll_inode_info *lli = ll_i2info(inode);
237 /* clear group lock, if present */
238 if (unlikely(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
239 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
240 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
241 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
245 /* Let's see if we have good enough OPEN lock on the file and if
246 we can skip talking to MDS */
247 if (file->f_dentry->d_inode) { /* Can this ever be false? */
249 int flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
250 struct lustre_handle lockh;
251 struct inode *inode = file->f_dentry->d_inode;
252 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
254 down(&lli->lli_och_sem);
255 if (fd->fd_omode & FMODE_WRITE) {
257 LASSERT(lli->lli_open_fd_write_count);
258 lli->lli_open_fd_write_count--;
259 } else if (fd->fd_omode & FMODE_EXEC) {
261 LASSERT(lli->lli_open_fd_exec_count);
262 lli->lli_open_fd_exec_count--;
265 LASSERT(lli->lli_open_fd_read_count);
266 lli->lli_open_fd_read_count--;
268 up(&lli->lli_och_sem);
270 if (!md_lock_match(md_exp, flags, ll_inode2fid(inode),
271 LDLM_IBITS, &policy, lockmode,
273 rc = ll_md_real_close(file->f_dentry->d_inode,
277 CERROR("Releasing a file %p with negative dentry %p. Name %s",
278 file, file->f_dentry, file->f_dentry->d_name.name);
281 LUSTRE_FPRIVATE(file) = NULL;
282 ll_file_data_put(fd);
283 ll_capa_close(inode);
288 int lov_test_and_clear_async_rc(struct lov_stripe_md *lsm);
290 /* While this returns an error code, fput() the caller does not, so we need
291 * to make every effort to clean up all of our state here. Also, applications
292 * rarely check close errors and even if an error is returned they will not
293 * re-try the close call.
295 int ll_file_release(struct inode *inode, struct file *file)
297 struct ll_file_data *fd;
298 struct ll_sb_info *sbi = ll_i2sbi(inode);
299 struct ll_inode_info *lli = ll_i2info(inode);
300 struct lov_stripe_md *lsm = lli->lli_smd;
304 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
305 inode->i_generation, inode);
307 #ifdef CONFIG_FS_POSIX_ACL
308 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
309 inode == inode->i_sb->s_root->d_inode) {
310 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
313 if (unlikely(fd->fd_flags & LL_FILE_RMTACL)) {
314 fd->fd_flags &= ~LL_FILE_RMTACL;
315 rct_del(&sbi->ll_rct, cfs_curproc_pid());
316 et_search_free(&sbi->ll_et, cfs_curproc_pid());
321 if (inode->i_sb->s_root != file->f_dentry)
322 ll_stats_ops_tally(sbi, LPROC_LL_RELEASE, 1);
323 fd = LUSTRE_FPRIVATE(file);
326 /* The last ref on @file, maybe not the the owner pid of statahead.
327 * Different processes can open the same dir, "ll_opendir_key" means:
328 * it is me that should stop the statahead thread. */
329 if (lli->lli_opendir_key == fd)
330 ll_stop_statahead(inode, fd);
332 if (inode->i_sb->s_root == file->f_dentry) {
333 LUSTRE_FPRIVATE(file) = NULL;
334 ll_file_data_put(fd);
339 lov_test_and_clear_async_rc(lsm);
340 lli->lli_async_rc = 0;
342 rc = ll_md_close(sbi->ll_md_exp, inode, file);
346 static int ll_intent_file_open(struct file *file, void *lmm,
347 int lmmsize, struct lookup_intent *itp)
349 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
350 struct dentry *parent = file->f_dentry->d_parent;
351 const char *name = file->f_dentry->d_name.name;
352 const int len = file->f_dentry->d_name.len;
353 struct md_op_data *op_data;
354 struct ptlrpc_request *req;
361 /* Usually we come here only for NFSD, and we want open lock.
362 But we can also get here with pre 2.6.15 patchless kernels, and in
363 that case that lock is also ok */
364 /* We can also get here if there was cached open handle in revalidate_it
365 * but it disappeared while we were getting from there to ll_file_open.
366 * But this means this file was closed and immediatelly opened which
367 * makes a good candidate for using OPEN lock */
368 /* If lmmsize & lmm are not 0, we are just setting stripe info
369 * parameters. No need for the open lock */
370 if (!lmm && !lmmsize)
371 itp->it_flags |= MDS_OPEN_LOCK;
373 op_data = ll_prep_md_op_data(NULL, parent->d_inode,
374 file->f_dentry->d_inode, name, len,
375 O_RDWR, LUSTRE_OPC_ANY, NULL);
377 RETURN(PTR_ERR(op_data));
379 rc = md_intent_lock(sbi->ll_md_exp, op_data, lmm, lmmsize, itp,
380 0 /*unused */, &req, ll_md_blocking_ast, 0);
381 ll_finish_md_op_data(op_data);
383 /* reason for keep own exit path - don`t flood log
384 * with messages with -ESTALE errors.
386 if (!it_disposition(itp, DISP_OPEN_OPEN) ||
387 it_open_error(DISP_OPEN_OPEN, itp))
389 ll_release_openhandle(file->f_dentry, itp);
393 if (rc != 0 || it_open_error(DISP_OPEN_OPEN, itp)) {
394 rc = rc ? rc : it_open_error(DISP_OPEN_OPEN, itp);
395 CDEBUG(D_VFSTRACE, "lock enqueue: err: %d\n", rc);
399 if (itp->d.lustre.it_lock_mode)
400 md_set_lock_data(sbi->ll_md_exp,
401 &itp->d.lustre.it_lock_handle,
402 file->f_dentry->d_inode);
404 rc = ll_prep_inode(&file->f_dentry->d_inode, req, NULL);
406 ptlrpc_req_finished(itp->d.lustre.it_data);
409 it_clear_disposition(itp, DISP_ENQ_COMPLETE);
410 ll_intent_drop_lock(itp);
415 static int ll_och_fill(struct obd_export *md_exp, struct ll_inode_info *lli,
416 struct lookup_intent *it, struct obd_client_handle *och)
418 struct ptlrpc_request *req = it->d.lustre.it_data;
419 struct mdt_body *body;
423 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
424 LASSERT(body != NULL); /* reply already checked out */
426 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
427 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
428 och->och_fid = lli->lli_fid;
429 och->och_flags = it->it_flags;
430 lli->lli_ioepoch = body->ioepoch;
432 return md_set_open_replay_data(md_exp, och, req);
435 int ll_local_open(struct file *file, struct lookup_intent *it,
436 struct ll_file_data *fd, struct obd_client_handle *och)
438 struct inode *inode = file->f_dentry->d_inode;
439 struct ll_inode_info *lli = ll_i2info(inode);
442 LASSERT(!LUSTRE_FPRIVATE(file));
447 struct ptlrpc_request *req = it->d.lustre.it_data;
448 struct mdt_body *body;
451 rc = ll_och_fill(ll_i2sbi(inode)->ll_md_exp, lli, it, och);
455 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
456 if ((it->it_flags & FMODE_WRITE) &&
457 (body->valid & OBD_MD_FLSIZE))
458 CDEBUG(D_INODE, "Epoch "LPU64" opened on "DFID"\n",
459 lli->lli_ioepoch, PFID(&lli->lli_fid));
462 LUSTRE_FPRIVATE(file) = fd;
463 ll_readahead_init(inode, &fd->fd_ras);
464 fd->fd_omode = it->it_flags;
468 /* Open a file, and (for the very first open) create objects on the OSTs at
469 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
470 * creation or open until ll_lov_setstripe() ioctl is called. We grab
471 * lli_open_sem to ensure no other process will create objects, send the
472 * stripe MD to the MDS, or try to destroy the objects if that fails.
474 * If we already have the stripe MD locally then we don't request it in
475 * md_open(), by passing a lmm_size = 0.
477 * It is up to the application to ensure no other processes open this file
478 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
479 * used. We might be able to avoid races of that sort by getting lli_open_sem
480 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
481 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
483 int ll_file_open(struct inode *inode, struct file *file)
485 struct ll_inode_info *lli = ll_i2info(inode);
486 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
487 .it_flags = file->f_flags };
488 struct lov_stripe_md *lsm;
489 struct ptlrpc_request *req = NULL;
490 struct obd_client_handle **och_p;
492 struct ll_file_data *fd;
493 int rc = 0, opendir_set = 0;
496 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n", inode->i_ino,
497 inode->i_generation, inode, file->f_flags);
499 #ifdef HAVE_VFS_INTENT_PATCHES
502 it = file->private_data; /* XXX: compat macro */
503 file->private_data = NULL; /* prevent ll_local_open assertion */
506 fd = ll_file_data_get();
510 if (S_ISDIR(inode->i_mode)) {
511 spin_lock(&lli->lli_lock);
512 /* "lli->lli_opendir_pid != 0" means someone has set it.
513 * "lli->lli_sai != NULL" means the previous statahead has not
515 if (lli->lli_opendir_pid == 0 && lli->lli_sai == NULL) {
517 lli->lli_opendir_pid = cfs_curproc_pid();
518 lli->lli_opendir_key = fd;
519 } else if (unlikely(lli->lli_opendir_pid == cfs_curproc_pid())) {
520 /* Two cases for this:
521 * (1) The same process open such directory many times.
522 * (2) The old process opened the directory, and exited
523 * before its children processes. Then new process
524 * with the same pid opens such directory before the
525 * old process's children processes exit.
526 * Change the owner to the latest one. */
528 lli->lli_opendir_key = fd;
530 spin_unlock(&lli->lli_lock);
533 if (inode->i_sb->s_root == file->f_dentry) {
534 LUSTRE_FPRIVATE(file) = fd;
538 if (!it || !it->d.lustre.it_disposition) {
539 /* Convert f_flags into access mode. We cannot use file->f_mode,
540 * because everything but O_ACCMODE mask was stripped from
542 if ((oit.it_flags + 1) & O_ACCMODE)
544 if (file->f_flags & O_TRUNC)
545 oit.it_flags |= FMODE_WRITE;
547 /* kernel only call f_op->open in dentry_open. filp_open calls
548 * dentry_open after call to open_namei that checks permissions.
549 * Only nfsd_open call dentry_open directly without checking
550 * permissions and because of that this code below is safe. */
551 if (oit.it_flags & FMODE_WRITE)
552 oit.it_flags |= MDS_OPEN_OWNEROVERRIDE;
554 /* We do not want O_EXCL here, presumably we opened the file
555 * already? XXX - NFS implications? */
556 oit.it_flags &= ~O_EXCL;
562 /* Let's see if we have file open on MDS already. */
563 if (it->it_flags & FMODE_WRITE) {
564 och_p = &lli->lli_mds_write_och;
565 och_usecount = &lli->lli_open_fd_write_count;
566 } else if (it->it_flags & FMODE_EXEC) {
567 och_p = &lli->lli_mds_exec_och;
568 och_usecount = &lli->lli_open_fd_exec_count;
570 och_p = &lli->lli_mds_read_och;
571 och_usecount = &lli->lli_open_fd_read_count;
574 down(&lli->lli_och_sem);
575 if (*och_p) { /* Open handle is present */
576 if (it_disposition(it, DISP_OPEN_OPEN)) {
577 /* Well, there's extra open request that we do not need,
578 let's close it somehow. This will decref request. */
579 rc = it_open_error(DISP_OPEN_OPEN, it);
581 ll_file_data_put(fd);
582 GOTO(out_och_free, rc);
584 ll_release_openhandle(file->f_dentry, it);
585 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
590 rc = ll_local_open(file, it, fd, NULL);
592 up(&lli->lli_och_sem);
593 ll_file_data_put(fd);
597 LASSERT(*och_usecount == 0);
598 if (!it->d.lustre.it_disposition) {
599 /* We cannot just request lock handle now, new ELC code
600 means that one of other OPEN locks for this file
601 could be cancelled, and since blocking ast handler
602 would attempt to grab och_sem as well, that would
603 result in a deadlock */
604 up(&lli->lli_och_sem);
605 it->it_flags |= O_CHECK_STALE;
606 rc = ll_intent_file_open(file, NULL, 0, it);
607 it->it_flags &= ~O_CHECK_STALE;
609 ll_file_data_put(fd);
610 GOTO(out_openerr, rc);
613 /* Got some error? Release the request */
614 if (it->d.lustre.it_status < 0) {
615 req = it->d.lustre.it_data;
616 ptlrpc_req_finished(req);
618 md_set_lock_data(ll_i2sbi(inode)->ll_md_exp,
619 &it->d.lustre.it_lock_handle,
620 file->f_dentry->d_inode);
623 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
625 ll_file_data_put(fd);
626 GOTO(out_och_free, rc = -ENOMEM);
629 req = it->d.lustre.it_data;
631 /* md_intent_lock() didn't get a request ref if there was an
632 * open error, so don't do cleanup on the request here
634 /* XXX (green): Should not we bail out on any error here, not
635 * just open error? */
636 rc = it_open_error(DISP_OPEN_OPEN, it);
638 ll_file_data_put(fd);
639 GOTO(out_och_free, rc);
642 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_OPEN, 1);
643 rc = ll_local_open(file, it, fd, *och_p);
645 up(&lli->lli_och_sem);
646 ll_file_data_put(fd);
647 GOTO(out_och_free, rc);
650 up(&lli->lli_och_sem);
652 /* Must do this outside lli_och_sem lock to prevent deadlock where
653 different kind of OPEN lock for this same inode gets cancelled
654 by ldlm_cancel_lru */
655 if (!S_ISREG(inode->i_mode))
662 if (file->f_flags & O_LOV_DELAY_CREATE ||
663 !(file->f_mode & FMODE_WRITE)) {
664 CDEBUG(D_INODE, "object creation was delayed\n");
668 file->f_flags &= ~O_LOV_DELAY_CREATE;
671 ptlrpc_req_finished(req);
673 it_clear_disposition(it, DISP_ENQ_OPEN_REF);
677 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
678 *och_p = NULL; /* OBD_FREE writes some magic there */
681 up(&lli->lli_och_sem);
683 if (opendir_set == 1) {
684 lli->lli_opendir_key = NULL;
685 lli->lli_opendir_pid = 0;
686 } else if (unlikely(opendir_set == 2)) {
687 ll_stop_statahead(inode, fd);
694 /* Fills the obdo with the attributes for the inode defined by lsm */
695 int ll_inode_getattr(struct inode *inode, struct obdo *obdo)
697 struct ptlrpc_request_set *set;
698 struct ll_inode_info *lli = ll_i2info(inode);
699 struct lov_stripe_md *lsm = lli->lli_smd;
701 struct obd_info oinfo = { { { 0 } } };
705 LASSERT(lsm != NULL);
709 oinfo.oi_oa->o_id = lsm->lsm_object_id;
710 oinfo.oi_oa->o_gr = lsm->lsm_object_gr;
711 oinfo.oi_oa->o_mode = S_IFREG;
712 oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE |
713 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
714 OBD_MD_FLBLKSZ | OBD_MD_FLATIME |
715 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
717 oinfo.oi_capa = ll_mdscapa_get(inode);
719 set = ptlrpc_prep_set();
721 CERROR("can't allocate ptlrpc set\n");
724 rc = obd_getattr_async(ll_i2dtexp(inode), &oinfo, set);
726 rc = ptlrpc_set_wait(set);
727 ptlrpc_set_destroy(set);
729 capa_put(oinfo.oi_capa);
733 oinfo.oi_oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ |
734 OBD_MD_FLATIME | OBD_MD_FLMTIME |
735 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
737 obdo_refresh_inode(inode, oinfo.oi_oa, oinfo.oi_oa->o_valid);
738 CDEBUG(D_INODE, "objid "LPX64" size %Lu, blocks %llu, blksize %lu\n",
739 lli->lli_smd->lsm_object_id, i_size_read(inode),
740 (unsigned long long)inode->i_blocks,
741 (unsigned long)ll_inode_blksize(inode));
745 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
747 struct ll_inode_info *lli = ll_i2info(inode);
748 struct lov_stripe_md *lsm = lli->lli_smd;
749 struct obd_export *exp = ll_i2dtexp(inode);
752 struct ldlm_lock *lock;
753 } key = { .name = KEY_LOCK_TO_STRIPE, .lock = lock };
754 __u32 stripe, vallen = sizeof(stripe);
755 struct lov_oinfo *loinfo;
759 if (lsm->lsm_stripe_count == 1)
760 GOTO(check, stripe = 0);
762 /* get our offset in the lov */
763 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe, lsm);
765 CERROR("obd_get_info: rc = %d\n", rc);
768 LASSERT(stripe < lsm->lsm_stripe_count);
771 loinfo = lsm->lsm_oinfo[stripe];
772 if (!osc_res_name_eq(loinfo->loi_id, loinfo->loi_gr,
773 &lock->l_resource->lr_name)){
774 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64,
775 loinfo->loi_id, loinfo->loi_gr);
776 RETURN(-ELDLM_NO_LOCK_DATA);
782 /* Get extra page reference to ensure it is not going away */
783 void ll_pin_extent_cb(void *data)
785 struct page *page = data;
787 page_cache_get(page);
792 /* Flush the page from page cache for an extent as its canceled.
793 * Page to remove is delivered as @data.
795 * No one can dirty the extent until we've finished our work and they cannot
796 * enqueue another lock. The DLM protects us from ll_file_read/write here,
797 * but other kernel actors could have pages locked.
799 * If @discard is set, there is no need to write the page if it is dirty.
801 * Called with the DLM lock held. */
802 int ll_page_removal_cb(void *data, int discard)
805 struct page *page = data;
806 struct address_space *mapping;
810 /* We have page reference already from ll_pin_page */
813 /* Already truncated by somebody */
816 mapping = page->mapping;
818 ll_teardown_mmaps(mapping,
819 (__u64)page->index << PAGE_CACHE_SHIFT,
820 ((__u64)page->index<<PAGE_CACHE_SHIFT)|
822 LL_CDEBUG_PAGE(D_PAGE, page, "removing page\n");
824 if (!discard && clear_page_dirty_for_io(page)) {
825 LASSERT(page->mapping);
826 rc = ll_call_writepage(page->mapping->host, page);
827 /* either waiting for io to complete or reacquiring
828 * the lock that the failed writepage released */
830 wait_on_page_writeback(page);
832 CERROR("writepage inode %lu(%p) of page %p "
833 "failed: %d\n", mapping->host->i_ino,
834 mapping->host, page, rc);
836 set_bit(AS_ENOSPC, &mapping->flags);
838 set_bit(AS_EIO, &mapping->flags);
840 set_bit(AS_EIO, &mapping->flags);
842 if (page->mapping != NULL) {
843 struct ll_async_page *llap = llap_cast_private(page);
844 /* checking again to account for writeback's lock_page() */
845 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
847 ll_ra_accounting(llap, page->mapping);
848 ll_truncate_complete_page(page);
852 LASSERT(!PageWriteback(page));
854 page_cache_release(page);
859 int ll_extent_lock_cancel_cb(struct ldlm_lock *lock, struct ldlm_lock_desc *new,
860 void *data, int flag)
863 struct ll_inode_info *lli;
864 struct lov_stripe_md *lsm;
870 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
871 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
875 inode = ll_inode_from_lock(lock);
878 lli = ll_i2info(inode);
881 if (lli->lli_smd == NULL)
885 stripe = ll_lock_to_stripe_offset(inode, lock);
889 lov_stripe_lock(lsm);
890 lock_res_and_lock(lock);
891 kms = ldlm_extent_shift_kms(lock,
892 lsm->lsm_oinfo[stripe]->loi_kms);
894 if (lsm->lsm_oinfo[stripe]->loi_kms != kms)
895 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
896 lsm->lsm_oinfo[stripe]->loi_kms, kms);
897 lsm->lsm_oinfo[stripe]->loi_kms = kms;
898 unlock_res_and_lock(lock);
899 lov_stripe_unlock(lsm);
900 ll_queue_done_writing(inode, 0);
909 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
911 /* XXX ALLOCATE - 160 bytes */
912 struct inode *inode = ll_inode_from_lock(lock);
913 struct ll_inode_info *lli = ll_i2info(inode);
914 struct lustre_handle lockh = { 0 };
919 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
920 LDLM_FL_BLOCK_CONV)) {
921 LBUG(); /* not expecting any blocked async locks yet */
922 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
924 ldlm_lock_dump(D_OTHER, lock, 0);
925 ldlm_reprocess_all(lock->l_resource);
929 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
931 stripe = ll_lock_to_stripe_offset(inode, lock);
935 if (lock->l_lvb_len) {
936 struct lov_stripe_md *lsm = lli->lli_smd;
938 lvb = lock->l_lvb_data;
939 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
941 lock_res_and_lock(lock);
942 ll_inode_size_lock(inode, 1);
943 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
944 kms = ldlm_extent_shift_kms(NULL, kms);
945 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
946 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
947 lsm->lsm_oinfo[stripe].loi_kms, kms);
948 lsm->lsm_oinfo[stripe].loi_kms = kms;
949 ll_inode_size_unlock(inode, 1);
950 unlock_res_and_lock(lock);
955 wake_up(&lock->l_waitq);
957 ldlm_lock2handle(lock, &lockh);
958 ldlm_lock_decref(&lockh, LCK_PR);
963 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
965 struct ptlrpc_request *req = reqp;
966 struct inode *inode = ll_inode_from_lock(lock);
967 struct ll_inode_info *lli;
968 struct lov_stripe_md *lsm;
974 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
975 lli = ll_i2info(inode);
977 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
980 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
982 /* First, find out which stripe index this lock corresponds to. */
983 stripe = ll_lock_to_stripe_offset(inode, lock);
985 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
987 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
988 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
990 rc = req_capsule_server_pack(&req->rq_pill);
992 CERROR("lustre_pack_reply: %d\n", rc);
996 lvb = req_capsule_server_get(&req->rq_pill, &RMF_DLM_LVB);
997 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe]->loi_kms;
998 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
999 lvb->lvb_atime = LTIME_S(inode->i_atime);
1000 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
1002 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64
1003 " atime "LPU64", mtime "LPU64", ctime "LPU64,
1004 i_size_read(inode), stripe, lvb->lvb_size, lvb->lvb_mtime,
1005 lvb->lvb_atime, lvb->lvb_ctime);
1010 /* These errors are normal races, so we don't want to fill the console
1011 * with messages by calling ptlrpc_error() */
1012 if (rc == -ELDLM_NO_LOCK_DATA)
1013 lustre_pack_reply(req, 1, NULL, NULL);
1015 req->rq_status = rc;
1019 static int ll_merge_lvb(struct inode *inode)
1021 struct ll_inode_info *lli = ll_i2info(inode);
1022 struct ll_sb_info *sbi = ll_i2sbi(inode);
1028 ll_inode_size_lock(inode, 1);
1029 inode_init_lvb(inode, &lvb);
1030 rc = obd_merge_lvb(sbi->ll_dt_exp, lli->lli_smd, &lvb, 0);
1031 i_size_write(inode, lvb.lvb_size);
1032 inode->i_blocks = lvb.lvb_blocks;
1034 LTIME_S(inode->i_mtime) = lvb.lvb_mtime;
1035 LTIME_S(inode->i_atime) = lvb.lvb_atime;
1036 LTIME_S(inode->i_ctime) = lvb.lvb_ctime;
1037 ll_inode_size_unlock(inode, 1);
1042 int ll_local_size(struct inode *inode)
1044 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1045 struct ll_inode_info *lli = ll_i2info(inode);
1046 struct ll_sb_info *sbi = ll_i2sbi(inode);
1047 struct lustre_handle lockh = { 0 };
1052 if (lli->lli_smd->lsm_stripe_count == 0)
1055 rc = obd_match(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT,
1056 &policy, LCK_PR, &flags, inode, &lockh);
1062 rc = ll_merge_lvb(inode);
1063 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
1067 int ll_glimpse_ioctl(struct ll_sb_info *sbi, struct lov_stripe_md *lsm,
1070 struct lustre_handle lockh = { 0 };
1071 struct ldlm_enqueue_info einfo = { 0 };
1072 struct obd_info oinfo = { { { 0 } } };
1078 einfo.ei_type = LDLM_EXTENT;
1079 einfo.ei_mode = LCK_PR;
1080 einfo.ei_cb_bl = osc_extent_blocking_cb;
1081 einfo.ei_cb_cp = ldlm_completion_ast;
1082 einfo.ei_cb_gl = ll_glimpse_callback;
1083 einfo.ei_cbdata = NULL;
1085 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
1086 oinfo.oi_lockh = &lockh;
1088 oinfo.oi_flags = LDLM_FL_HAS_INTENT;
1090 rc = obd_enqueue_rqset(sbi->ll_dt_exp, &oinfo, &einfo);
1094 CERROR("obd_enqueue returned rc %d, "
1095 "returning -EIO\n", rc);
1096 RETURN(rc > 0 ? -EIO : rc);
1099 lov_stripe_lock(lsm);
1100 memset(&lvb, 0, sizeof(lvb));
1101 obd_merge_lvb(sbi->ll_dt_exp, lsm, &lvb, 0);
1102 st->st_size = lvb.lvb_size;
1103 st->st_blocks = lvb.lvb_blocks;
1104 st->st_mtime = lvb.lvb_mtime;
1105 st->st_atime = lvb.lvb_atime;
1106 st->st_ctime = lvb.lvb_ctime;
1107 lov_stripe_unlock(lsm);
1112 /* NB: obd_merge_lvb will prefer locally cached writes if they extend the
1113 * file (because it prefers KMS over RSS when larger) */
1114 int ll_glimpse_size(struct inode *inode, int ast_flags)
1116 struct ll_inode_info *lli = ll_i2info(inode);
1117 struct ll_sb_info *sbi = ll_i2sbi(inode);
1118 struct lustre_handle lockh = { 0 };
1119 struct ldlm_enqueue_info einfo = { 0 };
1120 struct obd_info oinfo = { { { 0 } } };
1124 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
1127 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1129 if (!lli->lli_smd) {
1130 CDEBUG(D_DLMTRACE, "No objects for inode %lu\n", inode->i_ino);
1134 /* NOTE: this looks like DLM lock request, but it may not be one. Due
1135 * to LDLM_FL_HAS_INTENT flag, this is glimpse request, that
1136 * won't revoke any conflicting DLM locks held. Instead,
1137 * ll_glimpse_callback() will be called on each client
1138 * holding a DLM lock against this file, and resulting size
1139 * will be returned for each stripe. DLM lock on [0, EOF] is
1140 * acquired only if there were no conflicting locks. */
1141 einfo.ei_type = LDLM_EXTENT;
1142 einfo.ei_mode = LCK_PR;
1143 einfo.ei_cb_bl = osc_extent_blocking_cb;
1144 einfo.ei_cb_cp = ldlm_completion_ast;
1145 einfo.ei_cb_gl = ll_glimpse_callback;
1146 einfo.ei_cbdata = inode;
1148 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
1149 oinfo.oi_lockh = &lockh;
1150 oinfo.oi_md = lli->lli_smd;
1151 oinfo.oi_flags = ast_flags | LDLM_FL_HAS_INTENT;
1153 rc = obd_enqueue_rqset(sbi->ll_dt_exp, &oinfo, &einfo);
1157 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1158 RETURN(rc > 0 ? -EIO : rc);
1161 rc = ll_merge_lvb(inode);
1163 CDEBUG(D_DLMTRACE, "glimpse: size: %llu, blocks: %llu\n",
1164 i_size_read(inode), (unsigned long long)inode->i_blocks);
1169 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
1170 struct lov_stripe_md *lsm, int mode,
1171 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
1174 struct ll_sb_info *sbi = ll_i2sbi(inode);
1176 struct ldlm_enqueue_info einfo = { 0 };
1177 struct obd_info oinfo = { { { 0 } } };
1181 LASSERT(!lustre_handle_is_used(lockh));
1182 LASSERT(lsm != NULL);
1184 /* don't drop the mmapped file to LRU */
1185 if (mapping_mapped(inode->i_mapping))
1186 ast_flags |= LDLM_FL_NO_LRU;
1188 /* XXX phil: can we do this? won't it screw the file size up? */
1189 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1190 (sbi->ll_flags & LL_SBI_NOLCK))
1193 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
1194 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
1196 einfo.ei_type = LDLM_EXTENT;
1197 einfo.ei_mode = mode;
1198 einfo.ei_cb_bl = osc_extent_blocking_cb;
1199 einfo.ei_cb_cp = ldlm_completion_ast;
1200 einfo.ei_cb_gl = ll_glimpse_callback;
1201 einfo.ei_cbdata = inode;
1203 oinfo.oi_policy = *policy;
1204 oinfo.oi_lockh = lockh;
1206 oinfo.oi_flags = ast_flags;
1208 rc = obd_enqueue(sbi->ll_dt_exp, &oinfo, &einfo, NULL);
1209 *policy = oinfo.oi_policy;
1213 ll_inode_size_lock(inode, 1);
1214 inode_init_lvb(inode, &lvb);
1215 obd_merge_lvb(sbi->ll_dt_exp, lsm, &lvb, 1);
1217 if (policy->l_extent.start == 0 &&
1218 policy->l_extent.end == OBD_OBJECT_EOF) {
1219 /* vmtruncate()->ll_truncate() first sets the i_size and then
1220 * the kms under both a DLM lock and the
1221 * ll_inode_size_lock(). If we don't get the
1222 * ll_inode_size_lock() here we can match the DLM lock and
1223 * reset i_size from the kms before the truncating path has
1224 * updated the kms. generic_file_write can then trust the
1225 * stale i_size when doing appending writes and effectively
1226 * cancel the result of the truncate. Getting the
1227 * ll_inode_size_lock() after the enqueue maintains the DLM
1228 * -> ll_inode_size_lock() acquiring order. */
1229 i_size_write(inode, lvb.lvb_size);
1230 CDEBUG(D_INODE, "inode=%lu, updating i_size %llu\n",
1231 inode->i_ino, i_size_read(inode));
1235 LTIME_S(inode->i_mtime) = lvb.lvb_mtime;
1236 LTIME_S(inode->i_atime) = lvb.lvb_atime;
1237 LTIME_S(inode->i_ctime) = lvb.lvb_ctime;
1239 ll_inode_size_unlock(inode, 1);
1244 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1245 struct lov_stripe_md *lsm, int mode,
1246 struct lustre_handle *lockh)
1248 struct ll_sb_info *sbi = ll_i2sbi(inode);
1252 /* XXX phil: can we do this? won't it screw the file size up? */
1253 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1254 (sbi->ll_flags & LL_SBI_NOLCK))
1257 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1262 static void ll_set_file_contended(struct inode *inode)
1264 struct ll_inode_info *lli = ll_i2info(inode);
1265 cfs_time_t now = cfs_time_current();
1267 spin_lock(&lli->lli_lock);
1268 lli->lli_contention_time = now;
1269 lli->lli_flags |= LLIF_CONTENDED;
1270 spin_unlock(&lli->lli_lock);
1273 void ll_clear_file_contended(struct inode *inode)
1275 struct ll_inode_info *lli = ll_i2info(inode);
1277 spin_lock(&lli->lli_lock);
1278 lli->lli_flags &= ~LLIF_CONTENDED;
1279 spin_unlock(&lli->lli_lock);
1282 static int ll_is_file_contended(struct file *file)
1284 struct inode *inode = file->f_dentry->d_inode;
1285 struct ll_inode_info *lli = ll_i2info(inode);
1286 struct ll_sb_info *sbi = ll_i2sbi(inode);
1287 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1290 if (!(sbi->ll_lco.lco_flags & OBD_CONNECT_SRVLOCK)) {
1291 CDEBUG(D_INFO, "the server does not support SRVLOCK feature,"
1292 " osc connect flags = 0x"LPX64"\n",
1293 sbi->ll_lco.lco_flags);
1296 if (fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK))
1298 if (lli->lli_flags & LLIF_CONTENDED) {
1299 cfs_time_t cur_time = cfs_time_current();
1300 cfs_time_t retry_time;
1302 retry_time = cfs_time_add(
1303 lli->lli_contention_time,
1304 cfs_time_seconds(sbi->ll_contention_time));
1305 if (cfs_time_after(cur_time, retry_time)) {
1306 ll_clear_file_contended(inode);
1314 static int ll_file_get_tree_lock(struct ll_lock_tree *tree, struct file *file,
1315 const char *buf, size_t count,
1316 loff_t start, loff_t end, int rw)
1319 int tree_locked = 0;
1321 struct inode * inode = file->f_dentry->d_inode;
1324 append = (rw == OBD_BRW_WRITE) && (file->f_flags & O_APPEND);
1326 if (append || !ll_is_file_contended(file)) {
1327 struct ll_lock_tree_node *node;
1330 ast_flags = append ? 0 : LDLM_FL_DENY_ON_CONTENTION;
1331 if (file->f_flags & O_NONBLOCK)
1332 ast_flags |= LDLM_FL_BLOCK_NOWAIT;
1333 node = ll_node_from_inode(inode, start, end,
1334 (rw == OBD_BRW_WRITE) ? LCK_PW : LCK_PR);
1339 tree->lt_fd = LUSTRE_FPRIVATE(file);
1340 rc = ll_tree_lock(tree, node, buf, count, ast_flags);
1343 else if (rc == -EUSERS)
1344 ll_set_file_contended(inode);
1348 RETURN(tree_locked);
1354 * Checks if requested extent lock is compatible with a lock under a page.
1356 * Checks if the lock under \a page is compatible with a read or write lock
1357 * (specified by \a rw) for an extent [\a start , \a end].
1359 * \param page the page under which lock is considered
1360 * \param rw OBD_BRW_READ if requested for reading,
1361 * OBD_BRW_WRITE if requested for writing
1362 * \param start start of the requested extent
1363 * \param end end of the requested extent
1364 * \param cookie transparent parameter for passing locking context
1366 * \post result == 1, *cookie == context, appropriate lock is referenced or
1369 * \retval 1 owned lock is reused for the request
1370 * \retval 0 no lock reused for the request
1372 * \see ll_release_short_lock
1374 static int ll_reget_short_lock(struct page *page, int rw,
1375 obd_off start, obd_off end,
1378 struct ll_async_page *llap;
1379 struct obd_export *exp;
1380 struct inode *inode = page->mapping->host;
1384 exp = ll_i2dtexp(inode);
1388 llap = llap_cast_private(page);
1392 RETURN(obd_reget_short_lock(exp, ll_i2info(inode)->lli_smd,
1393 &llap->llap_cookie, rw, start, end,
1398 * Releases a reference to a lock taken in a "fast" way.
1400 * Releases a read or a write (specified by \a rw) lock
1401 * referenced by \a cookie.
1403 * \param inode inode to which data belong
1404 * \param end end of the locked extent
1405 * \param rw OBD_BRW_READ if requested for reading,
1406 * OBD_BRW_WRITE if requested for writing
1407 * \param cookie transparent parameter for passing locking context
1409 * \post appropriate lock is dereferenced
1411 * \see ll_reget_short_lock
1413 static void ll_release_short_lock(struct inode *inode, obd_off end,
1414 void *cookie, int rw)
1416 struct obd_export *exp;
1419 exp = ll_i2dtexp(inode);
1423 rc = obd_release_short_lock(exp, ll_i2info(inode)->lli_smd, end,
1426 CERROR("unlock failed (%d)\n", rc);
1430 * Checks if requested extent lock is compatible
1431 * with a lock under a page in page cache.
1433 * Checks if a lock under some \a page is compatible with a read or write lock
1434 * (specified by \a rw) for an extent [\a start , \a end].
1436 * \param file the file under which lock is considered
1437 * \param rw OBD_BRW_READ if requested for reading,
1438 * OBD_BRW_WRITE if requested for writing
1439 * \param ppos start of the requested extent
1440 * \param end end of the requested extent
1441 * \param cookie transparent parameter for passing locking context
1442 * \param buf userspace buffer for the data
1444 * \post result == 1, *cookie == context, appropriate lock is referenced
1447 * \retval 1 owned lock is reused for the request
1448 * \retval 0 no lock reused for the request
1450 * \see ll_file_put_fast_lock
1452 static inline int ll_file_get_fast_lock(struct file *file,
1453 obd_off ppos, obd_off end,
1454 char *buf, void **cookie, int rw)
1461 if (!ll_region_mapped((unsigned long)buf, end - ppos)) {
1462 page = find_lock_page(file->f_dentry->d_inode->i_mapping,
1463 ppos >> CFS_PAGE_SHIFT);
1465 if (ll_reget_short_lock(page, rw, ppos, end, cookie))
1469 page_cache_release(page);
1477 * Releases a reference to a lock taken in a "fast" way.
1479 * Releases a read or a write (specified by \a rw) lock
1480 * referenced by \a cookie.
1482 * \param inode inode to which data belong
1483 * \param end end of the locked extent
1484 * \param rw OBD_BRW_READ if requested for reading,
1485 * OBD_BRW_WRITE if requested for writing
1486 * \param cookie transparent parameter for passing locking context
1488 * \post appropriate lock is dereferenced
1490 * \see ll_file_get_fast_lock
1492 static inline void ll_file_put_fast_lock(struct inode *inode, obd_off end,
1493 void *cookie, int rw)
1495 ll_release_short_lock(inode, end, cookie, rw);
1498 enum ll_lock_style {
1499 LL_LOCK_STYLE_NOLOCK = 0,
1500 LL_LOCK_STYLE_FASTLOCK = 1,
1501 LL_LOCK_STYLE_TREELOCK = 2
1505 * Checks if requested extent lock is compatible with a lock
1506 * under a page cache page.
1508 * Checks if the lock under \a page is compatible with a read or write lock
1509 * (specified by \a rw) for an extent [\a start , \a end].
1511 * \param file file under which I/O is processed
1512 * \param rw OBD_BRW_READ if requested for reading,
1513 * OBD_BRW_WRITE if requested for writing
1514 * \param ppos start of the requested extent
1515 * \param end end of the requested extent
1516 * \param cookie transparent parameter for passing locking context
1517 * (only used with LL_LOCK_STYLE_FASTLOCK)
1518 * \param tree lock tree (only used with LL_LOCK_STYLE_TREELOCK)
1519 * \param buf userspace buffer for the data
1521 * \retval LL_LOCK_STYLE_FASTLOCK owned lock is reused through fast lock
1522 * \retval LL_LOCK_STYLE_TREELOCK got a lock through tree lock
1523 * \retval LL_LOCK_STYLE_NOLOCK got no lock
1525 * \see ll_file_put_lock
1527 static inline int ll_file_get_lock(struct file *file, obd_off ppos,
1528 obd_off end, char *buf, void **cookie,
1529 struct ll_lock_tree *tree, int rw)
1535 if (ll_file_get_fast_lock(file, ppos, end, buf, cookie, rw))
1536 RETURN(LL_LOCK_STYLE_FASTLOCK);
1538 rc = ll_file_get_tree_lock(tree, file, buf, ppos - end, ppos, end, rw);
1539 /* rc: 1 for tree lock, 0 for no lock, <0 for error */
1542 RETURN(LL_LOCK_STYLE_TREELOCK);
1544 RETURN(LL_LOCK_STYLE_NOLOCK);
1547 /* an error happened if we reached this point, rc = -errno here */
1552 * Drops the lock taken by ll_file_get_lock.
1554 * Releases a read or a write (specified by \a rw) lock
1555 * referenced by \a tree or \a cookie.
1557 * \param inode inode to which data belong
1558 * \param end end of the locked extent
1559 * \param lockstyle facility through which the lock was taken
1560 * \param rw OBD_BRW_READ if requested for reading,
1561 * OBD_BRW_WRITE if requested for writing
1562 * \param cookie transparent parameter for passing locking context
1563 * (only used with LL_LOCK_STYLE_FASTLOCK)
1564 * \param tree lock tree (only used with LL_LOCK_STYLE_TREELOCK)
1566 * \post appropriate lock is dereferenced
1568 * \see ll_file_get_lock
1570 static inline void ll_file_put_lock(struct inode *inode, obd_off end,
1571 enum ll_lock_style lock_style,
1572 void *cookie, struct ll_lock_tree *tree,
1576 switch (lock_style) {
1577 case LL_LOCK_STYLE_TREELOCK:
1578 ll_tree_unlock(tree);
1580 case LL_LOCK_STYLE_FASTLOCK:
1581 ll_file_put_fast_lock(inode, end, cookie, rw);
1584 CERROR("invalid locking style (%d)\n", lock_style);
1588 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1591 struct inode *inode = file->f_dentry->d_inode;
1592 struct ll_inode_info *lli = ll_i2info(inode);
1593 struct lov_stripe_md *lsm = lli->lli_smd;
1594 struct ll_sb_info *sbi = ll_i2sbi(inode);
1595 struct ll_lock_tree tree;
1597 struct ll_ra_read bead;
1600 ssize_t retval, chunk, sum = 0;
1606 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1607 inode->i_ino, inode->i_generation, inode, count, *ppos);
1608 /* "If nbyte is 0, read() will return 0 and have no other results."
1609 * -- Single Unix Spec */
1613 ll_stats_ops_tally(sbi, LPROC_LL_READ_BYTES, count);
1616 /* Read on file with no objects should return zero-filled
1617 * buffers up to file size (we can get non-zero sizes with
1618 * mknod + truncate, then opening file for read. This is a
1619 * common pattern in NFS case, it seems). Bug 6243 */
1621 /* Since there are no objects on OSTs, we have nothing to get
1622 * lock on and so we are forced to access inode->i_size
1625 /* Read beyond end of file */
1626 if (*ppos >= i_size_read(inode))
1629 if (count > i_size_read(inode) - *ppos)
1630 count = i_size_read(inode) - *ppos;
1631 /* Make sure to correctly adjust the file pos pointer for
1633 notzeroed = clear_user(buf, count);
1641 if (sbi->ll_max_rw_chunk != 0) {
1642 /* first, let's know the end of the current stripe */
1644 obd_extent_calc(sbi->ll_dt_exp, lsm, OBD_CALC_STRIPE_END, &end);
1646 /* correct, the end is beyond the request */
1647 if (end > *ppos + count - 1)
1648 end = *ppos + count - 1;
1650 /* and chunk shouldn't be too large even if striping is wide */
1651 if (end - *ppos > sbi->ll_max_rw_chunk)
1652 end = *ppos + sbi->ll_max_rw_chunk - 1;
1654 end = *ppos + count - 1;
1657 lock_style = ll_file_get_lock(file, (obd_off)(*ppos), end,
1658 buf, &cookie, &tree, OBD_BRW_READ);
1660 GOTO(out, retval = lock_style);
1662 ll_inode_size_lock(inode, 1);
1664 * Consistency guarantees: following possibilities exist for the
1665 * relation between region being read and real file size at this
1668 * (A): the region is completely inside of the file;
1670 * (B-x): x bytes of region are inside of the file, the rest is
1673 * (C): the region is completely outside of the file.
1675 * This classification is stable under DLM lock acquired by
1676 * ll_tree_lock() above, because to change class, other client has to
1677 * take DLM lock conflicting with our lock. Also, any updates to
1678 * ->i_size by other threads on this client are serialized by
1679 * ll_inode_size_lock(). This guarantees that short reads are handled
1680 * correctly in the face of concurrent writes and truncates.
1682 inode_init_lvb(inode, &lvb);
1683 obd_merge_lvb(sbi->ll_dt_exp, lsm, &lvb, 1);
1685 if (*ppos + count - 1 > kms) {
1686 /* A glimpse is necessary to determine whether we return a
1687 * short read (B) or some zeroes at the end of the buffer (C) */
1688 ll_inode_size_unlock(inode, 1);
1689 retval = ll_glimpse_size(inode, LDLM_FL_BLOCK_GRANTED);
1691 if (lock_style != LL_LOCK_STYLE_NOLOCK)
1692 ll_file_put_lock(inode, end, lock_style,
1693 cookie, &tree, OBD_BRW_READ);
1697 /* region is within kms and, hence, within real file size (A).
1698 * We need to increase i_size to cover the read region so that
1699 * generic_file_read() will do its job, but that doesn't mean
1700 * the kms size is _correct_, it is only the _minimum_ size.
1701 * If someone does a stat they will get the correct size which
1702 * will always be >= the kms value here. b=11081 */
1703 if (i_size_read(inode) < kms)
1704 i_size_write(inode, kms);
1705 ll_inode_size_unlock(inode, 1);
1708 chunk = end - *ppos + 1;
1709 CDEBUG(D_INODE, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1710 inode->i_ino, chunk, *ppos, i_size_read(inode));
1712 if (lock_style != LL_LOCK_STYLE_NOLOCK) {
1713 /* turn off the kernel's read-ahead */
1714 file->f_ra.ra_pages = 0;
1716 /* initialize read-ahead window once per syscall */
1719 bead.lrr_start = *ppos >> CFS_PAGE_SHIFT;
1720 bead.lrr_count = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1721 ll_ra_read_in(file, &bead);
1725 file_accessed(file);
1726 retval = generic_file_read(file, buf, chunk, ppos);
1727 ll_file_put_lock(inode, end, lock_style, cookie, &tree,
1730 retval = ll_file_lockless_io(file, buf, chunk, ppos, READ);
1733 ll_rw_stats_tally(sbi, current->pid, file, chunk, 0);
1739 if (retval == chunk && count > 0)
1745 ll_ra_read_ex(file, &bead);
1746 retval = (sum > 0) ? sum : retval;
1751 * Write to a file (through the page cache).
1753 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
1756 struct inode *inode = file->f_dentry->d_inode;
1757 struct ll_sb_info *sbi = ll_i2sbi(inode);
1758 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1759 struct ll_lock_tree tree;
1760 loff_t maxbytes = ll_file_maxbytes(inode);
1761 loff_t lock_start, lock_end, end;
1762 ssize_t retval, chunk, sum = 0;
1766 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1767 inode->i_ino, inode->i_generation, inode, count, *ppos);
1769 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1771 /* POSIX, but surprised the VFS doesn't check this already */
1775 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1776 * called on the file, don't fail the below assertion (bug 2388). */
1777 if (file->f_flags & O_LOV_DELAY_CREATE &&
1778 ll_i2info(inode)->lli_smd == NULL)
1781 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1783 down(&ll_i2info(inode)->lli_write_sem);
1786 chunk = 0; /* just to fix gcc's warning */
1787 end = *ppos + count - 1;
1789 if (file->f_flags & O_APPEND) {
1791 lock_end = OBD_OBJECT_EOF;
1792 } else if (sbi->ll_max_rw_chunk != 0) {
1793 /* first, let's know the end of the current stripe */
1795 obd_extent_calc(sbi->ll_dt_exp, lsm, OBD_CALC_STRIPE_END,
1798 /* correct, the end is beyond the request */
1799 if (end > *ppos + count - 1)
1800 end = *ppos + count - 1;
1802 /* and chunk shouldn't be too large even if striping is wide */
1803 if (end - *ppos > sbi->ll_max_rw_chunk)
1804 end = *ppos + sbi->ll_max_rw_chunk - 1;
1809 lock_end = *ppos + count - 1;
1812 tree_locked = ll_file_get_tree_lock(&tree, file, buf, count,
1813 lock_start, lock_end, OBD_BRW_WRITE);
1814 if (tree_locked < 0)
1815 GOTO(out, retval = tree_locked);
1817 /* This is ok, g_f_w will overwrite this under i_sem if it races
1818 * with a local truncate, it just makes our maxbyte checking easier.
1819 * The i_size value gets updated in ll_extent_lock() as a consequence
1820 * of the [0,EOF] extent lock we requested above. */
1821 if (file->f_flags & O_APPEND) {
1822 *ppos = i_size_read(inode);
1823 end = *ppos + count - 1;
1826 if (*ppos >= maxbytes) {
1827 send_sig(SIGXFSZ, current, 0);
1828 GOTO(out_unlock, retval = -EFBIG);
1830 if (end > maxbytes - 1)
1833 /* generic_file_write handles O_APPEND after getting i_mutex */
1834 chunk = end - *ppos + 1;
1835 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1836 inode->i_ino, chunk, *ppos);
1838 retval = generic_file_write(file, buf, chunk, ppos);
1840 retval = ll_file_lockless_io(file, (char*)buf, chunk,
1842 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, file, chunk, 1);
1846 ll_tree_unlock(&tree);
1853 if (retval == chunk && count > 0)
1857 up(&ll_i2info(inode)->lli_write_sem);
1859 retval = (sum > 0) ? sum : retval;
1860 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_WRITE_BYTES,
1861 retval > 0 ? retval : 0);
1866 * Send file content (through pagecache) somewhere with helper
1868 static ssize_t ll_file_sendfile(struct file *in_file, loff_t *ppos,size_t count,
1869 read_actor_t actor, void *target)
1871 struct inode *inode = in_file->f_dentry->d_inode;
1872 struct ll_inode_info *lli = ll_i2info(inode);
1873 struct lov_stripe_md *lsm = lli->lli_smd;
1874 struct ll_lock_tree tree;
1875 struct ll_lock_tree_node *node;
1877 struct ll_ra_read bead;
1882 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1883 inode->i_ino, inode->i_generation, inode, count, *ppos);
1885 /* "If nbyte is 0, read() will return 0 and have no other results."
1886 * -- Single Unix Spec */
1890 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_READ_BYTES, count);
1891 /* turn off the kernel's read-ahead */
1892 in_file->f_ra.ra_pages = 0;
1894 /* File with no objects, nothing to lock */
1896 RETURN(generic_file_sendfile(in_file, ppos, count, actor, target));
1898 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1, LCK_PR);
1900 RETURN(PTR_ERR(node));
1902 tree.lt_fd = LUSTRE_FPRIVATE(in_file);
1903 rc = ll_tree_lock(&tree, node, NULL, count,
1904 in_file->f_flags & O_NONBLOCK?LDLM_FL_BLOCK_NOWAIT:0);
1908 ll_clear_file_contended(inode);
1909 ll_inode_size_lock(inode, 1);
1911 * Consistency guarantees: following possibilities exist for the
1912 * relation between region being read and real file size at this
1915 * (A): the region is completely inside of the file;
1917 * (B-x): x bytes of region are inside of the file, the rest is
1920 * (C): the region is completely outside of the file.
1922 * This classification is stable under DLM lock acquired by
1923 * ll_tree_lock() above, because to change class, other client has to
1924 * take DLM lock conflicting with our lock. Also, any updates to
1925 * ->i_size by other threads on this client are serialized by
1926 * ll_inode_size_lock(). This guarantees that short reads are handled
1927 * correctly in the face of concurrent writes and truncates.
1929 inode_init_lvb(inode, &lvb);
1930 obd_merge_lvb(ll_i2sbi(inode)->ll_dt_exp, lsm, &lvb, 1);
1932 if (*ppos + count - 1 > kms) {
1933 /* A glimpse is necessary to determine whether we return a
1934 * short read (B) or some zeroes at the end of the buffer (C) */
1935 ll_inode_size_unlock(inode, 1);
1936 retval = ll_glimpse_size(inode, LDLM_FL_BLOCK_GRANTED);
1940 /* region is within kms and, hence, within real file size (A) */
1941 i_size_write(inode, kms);
1942 ll_inode_size_unlock(inode, 1);
1945 CDEBUG(D_INFO, "Send ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1946 inode->i_ino, count, *ppos, i_size_read(inode));
1948 bead.lrr_start = *ppos >> CFS_PAGE_SHIFT;
1949 bead.lrr_count = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1950 ll_ra_read_in(in_file, &bead);
1952 file_accessed(in_file);
1953 retval = generic_file_sendfile(in_file, ppos, count, actor, target);
1954 ll_ra_read_ex(in_file, &bead);
1957 ll_tree_unlock(&tree);
1961 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
1964 struct ll_inode_info *lli = ll_i2info(inode);
1965 struct obd_export *exp = ll_i2dtexp(inode);
1966 struct ll_recreate_obj ucreatp;
1967 struct obd_trans_info oti = { 0 };
1968 struct obdo *oa = NULL;
1971 struct lov_stripe_md *lsm, *lsm2;
1974 if (!capable (CAP_SYS_ADMIN))
1977 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
1978 sizeof(struct ll_recreate_obj));
1986 down(&lli->lli_size_sem);
1989 GOTO(out, rc = -ENOENT);
1990 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
1991 (lsm->lsm_stripe_count));
1993 OBD_ALLOC(lsm2, lsm_size);
1995 GOTO(out, rc = -ENOMEM);
1997 oa->o_id = ucreatp.lrc_id;
1998 oa->o_gr = ucreatp.lrc_group;
1999 oa->o_nlink = ucreatp.lrc_ost_idx;
2000 oa->o_flags |= OBD_FL_RECREATE_OBJS;
2001 oa->o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
2002 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
2003 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
2005 memcpy(lsm2, lsm, lsm_size);
2006 rc = obd_create(exp, oa, &lsm2, &oti);
2008 OBD_FREE(lsm2, lsm_size);
2011 up(&lli->lli_size_sem);
2016 int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
2017 int flags, struct lov_user_md *lum, int lum_size)
2019 struct ll_inode_info *lli = ll_i2info(inode);
2020 struct lov_stripe_md *lsm;
2021 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
2025 down(&lli->lli_size_sem);
2028 up(&lli->lli_size_sem);
2029 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
2034 rc = ll_intent_file_open(file, lum, lum_size, &oit);
2037 if (it_disposition(&oit, DISP_LOOKUP_NEG))
2038 GOTO(out_req_free, rc = -ENOENT);
2039 rc = oit.d.lustre.it_status;
2041 GOTO(out_req_free, rc);
2043 ll_release_openhandle(file->f_dentry, &oit);
2046 up(&lli->lli_size_sem);
2047 ll_intent_release(&oit);
2050 ptlrpc_req_finished((struct ptlrpc_request *) oit.d.lustre.it_data);
2054 int ll_lov_getstripe_ea_info(struct inode *inode, const char *filename,
2055 struct lov_mds_md **lmmp, int *lmm_size,
2056 struct ptlrpc_request **request)
2058 struct ll_sb_info *sbi = ll_i2sbi(inode);
2059 struct mdt_body *body;
2060 struct lov_mds_md *lmm = NULL;
2061 struct ptlrpc_request *req = NULL;
2062 struct obd_capa *oc;
2065 rc = ll_get_max_mdsize(sbi, &lmmsize);
2069 oc = ll_mdscapa_get(inode);
2070 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode),
2071 oc, filename, strlen(filename) + 1,
2072 OBD_MD_FLEASIZE | OBD_MD_FLDIREA, lmmsize,
2073 ll_i2suppgid(inode), &req);
2076 CDEBUG(D_INFO, "md_getattr_name failed "
2077 "on %s: rc %d\n", filename, rc);
2081 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
2082 LASSERT(body != NULL); /* checked by mdc_getattr_name */
2084 lmmsize = body->eadatasize;
2086 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
2088 GOTO(out, rc = -ENODATA);
2091 lmm = req_capsule_server_sized_get(&req->rq_pill, &RMF_MDT_MD, lmmsize);
2092 LASSERT(lmm != NULL);
2095 * This is coming from the MDS, so is probably in
2096 * little endian. We convert it to host endian before
2097 * passing it to userspace.
2099 if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
2100 lustre_swab_lov_user_md((struct lov_user_md *)lmm);
2101 lustre_swab_lov_user_md_objects((struct lov_user_md *)lmm);
2102 } else if (lmm->lmm_magic == __swab32(LOV_MAGIC_JOIN)) {
2103 lustre_swab_lov_user_md_join((struct lov_user_md_join *)lmm);
2106 if (lmm->lmm_magic == LOV_MAGIC_JOIN) {
2107 struct lov_stripe_md *lsm;
2108 struct lov_user_md_join *lmj;
2109 int lmj_size, i, aindex = 0;
2111 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
2113 GOTO(out, rc = -ENOMEM);
2114 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
2116 GOTO(out_free_memmd, rc);
2118 lmj_size = sizeof(struct lov_user_md_join) +
2119 lsm->lsm_stripe_count *
2120 sizeof(struct lov_user_ost_data_join);
2121 OBD_ALLOC(lmj, lmj_size);
2123 GOTO(out_free_memmd, rc = -ENOMEM);
2125 memcpy(lmj, lmm, sizeof(struct lov_user_md_join));
2126 for (i = 0; i < lsm->lsm_stripe_count; i++) {
2127 struct lov_extent *lex =
2128 &lsm->lsm_array->lai_ext_array[aindex];
2130 if (lex->le_loi_idx + lex->le_stripe_count <= i)
2132 CDEBUG(D_INFO, "aindex %d i %d l_extent_start "
2133 LPU64" len %d\n", aindex, i,
2134 lex->le_start, (int)lex->le_len);
2135 lmj->lmm_objects[i].l_extent_start =
2138 if ((int)lex->le_len == -1)
2139 lmj->lmm_objects[i].l_extent_end = -1;
2141 lmj->lmm_objects[i].l_extent_end =
2142 lex->le_start + lex->le_len;
2143 lmj->lmm_objects[i].l_object_id =
2144 lsm->lsm_oinfo[i]->loi_id;
2145 lmj->lmm_objects[i].l_object_gr =
2146 lsm->lsm_oinfo[i]->loi_gr;
2147 lmj->lmm_objects[i].l_ost_gen =
2148 lsm->lsm_oinfo[i]->loi_ost_gen;
2149 lmj->lmm_objects[i].l_ost_idx =
2150 lsm->lsm_oinfo[i]->loi_ost_idx;
2152 lmm = (struct lov_mds_md *)lmj;
2155 obd_free_memmd(sbi->ll_dt_exp, &lsm);
2159 *lmm_size = lmmsize;
2164 static int ll_lov_setea(struct inode *inode, struct file *file,
2167 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
2168 struct lov_user_md *lump;
2169 int lum_size = sizeof(struct lov_user_md) +
2170 sizeof(struct lov_user_ost_data);
2174 if (!capable (CAP_SYS_ADMIN))
2177 OBD_ALLOC(lump, lum_size);
2181 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
2183 OBD_FREE(lump, lum_size);
2187 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
2189 OBD_FREE(lump, lum_size);
2193 static int ll_lov_setstripe(struct inode *inode, struct file *file,
2196 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
2198 int flags = FMODE_WRITE;
2201 /* Bug 1152: copy properly when this is no longer true */
2202 LASSERT(sizeof(lum) == sizeof(*lump));
2203 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
2204 rc = copy_from_user(&lum, lump, sizeof(lum));
2208 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
2210 put_user(0, &lump->lmm_stripe_count);
2211 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
2212 0, ll_i2info(inode)->lli_smd, lump);
2217 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
2219 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
2224 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
2228 static int ll_get_grouplock(struct inode *inode, struct file *file,
2231 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
2232 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
2233 .end = OBD_OBJECT_EOF}};
2234 struct lustre_handle lockh = { 0 };
2235 struct ll_inode_info *lli = ll_i2info(inode);
2236 struct lov_stripe_md *lsm = lli->lli_smd;
2240 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
2244 policy.l_extent.gid = arg;
2245 if (file->f_flags & O_NONBLOCK)
2246 flags = LDLM_FL_BLOCK_NOWAIT;
2248 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags);
2252 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
2254 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
2259 static int ll_put_grouplock(struct inode *inode, struct file *file,
2262 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
2263 struct ll_inode_info *lli = ll_i2info(inode);
2264 struct lov_stripe_md *lsm = lli->lli_smd;
2268 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
2269 /* Ugh, it's already unlocked. */
2273 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
2276 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
2278 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
2283 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
2288 static int join_sanity_check(struct inode *head, struct inode *tail)
2291 if ((ll_i2sbi(head)->ll_flags & LL_SBI_JOIN) == 0) {
2292 CERROR("server do not support join \n");
2295 if (!S_ISREG(tail->i_mode) || !S_ISREG(head->i_mode)) {
2296 CERROR("tail ino %lu and ino head %lu must be regular\n",
2297 head->i_ino, tail->i_ino);
2300 if (head->i_ino == tail->i_ino) {
2301 CERROR("file %lu can not be joined to itself \n", head->i_ino);
2304 if (i_size_read(head) % JOIN_FILE_ALIGN) {
2305 CERROR("hsize %llu must be times of 64K\n", i_size_read(head));
2311 static int join_file(struct inode *head_inode, struct file *head_filp,
2312 struct file *tail_filp)
2314 struct dentry *tail_dentry = tail_filp->f_dentry;
2315 struct lookup_intent oit = {.it_op = IT_OPEN,
2316 .it_flags = head_filp->f_flags|O_JOIN_FILE};
2317 struct ldlm_enqueue_info einfo = { LDLM_IBITS, LCK_CW,
2318 ll_md_blocking_ast, ldlm_completion_ast, NULL, NULL };
2320 struct lustre_handle lockh;
2321 struct md_op_data *op_data;
2326 tail_dentry = tail_filp->f_dentry;
2328 data = i_size_read(head_inode);
2329 op_data = ll_prep_md_op_data(NULL, head_inode,
2330 tail_dentry->d_parent->d_inode,
2331 tail_dentry->d_name.name,
2332 tail_dentry->d_name.len, 0,
2333 LUSTRE_OPC_ANY, &data);
2334 if (IS_ERR(op_data))
2335 RETURN(PTR_ERR(op_data));
2337 rc = md_enqueue(ll_i2mdexp(head_inode), &einfo, &oit,
2338 op_data, &lockh, NULL, 0, 0);
2340 ll_finish_md_op_data(op_data);
2344 rc = oit.d.lustre.it_status;
2346 if (rc < 0 || it_open_error(DISP_OPEN_OPEN, &oit)) {
2347 rc = rc ? rc : it_open_error(DISP_OPEN_OPEN, &oit);
2348 ptlrpc_req_finished((struct ptlrpc_request *)
2349 oit.d.lustre.it_data);
2353 if (oit.d.lustre.it_lock_mode) { /* If we got lock - release it right
2355 ldlm_lock_decref(&lockh, oit.d.lustre.it_lock_mode);
2356 oit.d.lustre.it_lock_mode = 0;
2358 ll_release_openhandle(head_filp->f_dentry, &oit);
2360 ll_intent_release(&oit);
2364 static int ll_file_join(struct inode *head, struct file *filp,
2365 char *filename_tail)
2367 struct inode *tail = NULL, *first = NULL, *second = NULL;
2368 struct dentry *tail_dentry;
2369 struct file *tail_filp, *first_filp, *second_filp;
2370 struct ll_lock_tree first_tree, second_tree;
2371 struct ll_lock_tree_node *first_node, *second_node;
2372 struct ll_inode_info *hlli = ll_i2info(head), *tlli;
2373 int rc = 0, cleanup_phase = 0;
2376 CDEBUG(D_VFSTRACE, "VFS Op:head=%lu/%u(%p) tail %s\n",
2377 head->i_ino, head->i_generation, head, filename_tail);
2379 tail_filp = filp_open(filename_tail, O_WRONLY, 0644);
2380 if (IS_ERR(tail_filp)) {
2381 CERROR("Can not open tail file %s", filename_tail);
2382 rc = PTR_ERR(tail_filp);
2385 tail = igrab(tail_filp->f_dentry->d_inode);
2387 tlli = ll_i2info(tail);
2388 tail_dentry = tail_filp->f_dentry;
2389 LASSERT(tail_dentry);
2392 /*reorder the inode for lock sequence*/
2393 first = head->i_ino > tail->i_ino ? head : tail;
2394 second = head->i_ino > tail->i_ino ? tail : head;
2395 first_filp = head->i_ino > tail->i_ino ? filp : tail_filp;
2396 second_filp = head->i_ino > tail->i_ino ? tail_filp : filp;
2398 CDEBUG(D_INFO, "reorder object from %lu:%lu to %lu:%lu \n",
2399 head->i_ino, tail->i_ino, first->i_ino, second->i_ino);
2400 first_node = ll_node_from_inode(first, 0, OBD_OBJECT_EOF, LCK_EX);
2401 if (IS_ERR(first_node)){
2402 rc = PTR_ERR(first_node);
2405 first_tree.lt_fd = first_filp->private_data;
2406 rc = ll_tree_lock(&first_tree, first_node, NULL, 0, 0);
2411 second_node = ll_node_from_inode(second, 0, OBD_OBJECT_EOF, LCK_EX);
2412 if (IS_ERR(second_node)){
2413 rc = PTR_ERR(second_node);
2416 second_tree.lt_fd = second_filp->private_data;
2417 rc = ll_tree_lock(&second_tree, second_node, NULL, 0, 0);
2422 rc = join_sanity_check(head, tail);
2426 rc = join_file(head, filp, tail_filp);
2430 switch (cleanup_phase) {
2432 ll_tree_unlock(&second_tree);
2433 obd_cancel_unused(ll_i2dtexp(second),
2434 ll_i2info(second)->lli_smd, 0, NULL);
2436 ll_tree_unlock(&first_tree);
2437 obd_cancel_unused(ll_i2dtexp(first),
2438 ll_i2info(first)->lli_smd, 0, NULL);
2440 filp_close(tail_filp, 0);
2443 if (head && rc == 0) {
2444 obd_free_memmd(ll_i2sbi(head)->ll_dt_exp,
2446 hlli->lli_smd = NULL;
2451 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2457 int ll_release_openhandle(struct dentry *dentry, struct lookup_intent *it)
2459 struct inode *inode = dentry->d_inode;
2460 struct obd_client_handle *och;
2466 /* Root ? Do nothing. */
2467 if (dentry->d_inode->i_sb->s_root == dentry)
2470 /* No open handle to close? Move away */
2471 if (!it_disposition(it, DISP_OPEN_OPEN))
2474 LASSERT(it_open_error(DISP_OPEN_OPEN, it) == 0);
2476 OBD_ALLOC(och, sizeof(*och));
2478 GOTO(out, rc = -ENOMEM);
2480 ll_och_fill(ll_i2sbi(inode)->ll_md_exp,
2481 ll_i2info(inode), it, och);
2483 rc = ll_close_inode_openhandle(ll_i2sbi(inode)->ll_md_exp,
2486 /* this one is in place of ll_file_open */
2487 ptlrpc_req_finished(it->d.lustre.it_data);
2488 it_clear_disposition(it, DISP_ENQ_OPEN_REF);
2492 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
2495 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
2499 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
2500 inode->i_generation, inode, cmd);
2501 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
2503 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
2504 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
2508 case LL_IOC_GETFLAGS:
2509 /* Get the current value of the file flags */
2510 return put_user(fd->fd_flags, (int *)arg);
2511 case LL_IOC_SETFLAGS:
2512 case LL_IOC_CLRFLAGS:
2513 /* Set or clear specific file flags */
2514 /* XXX This probably needs checks to ensure the flags are
2515 * not abused, and to handle any flag side effects.
2517 if (get_user(flags, (int *) arg))
2520 if (cmd == LL_IOC_SETFLAGS) {
2521 if ((flags & LL_FILE_IGNORE_LOCK) &&
2522 !(file->f_flags & O_DIRECT)) {
2523 CERROR("%s: unable to disable locking on "
2524 "non-O_DIRECT file\n", current->comm);
2528 fd->fd_flags |= flags;
2530 fd->fd_flags &= ~flags;
2533 case LL_IOC_LOV_SETSTRIPE:
2534 RETURN(ll_lov_setstripe(inode, file, arg));
2535 case LL_IOC_LOV_SETEA:
2536 RETURN(ll_lov_setea(inode, file, arg));
2537 case LL_IOC_LOV_GETSTRIPE:
2538 RETURN(ll_lov_getstripe(inode, arg));
2539 case LL_IOC_RECREATE_OBJ:
2540 RETURN(ll_lov_recreate_obj(inode, file, arg));
2541 case EXT3_IOC_GETFLAGS:
2542 case EXT3_IOC_SETFLAGS:
2543 RETURN(ll_iocontrol(inode, file, cmd, arg));
2544 case EXT3_IOC_GETVERSION_OLD:
2545 case EXT3_IOC_GETVERSION:
2546 RETURN(put_user(inode->i_generation, (int *)arg));
2551 ftail = getname((const char *)arg);
2553 RETURN(PTR_ERR(ftail));
2554 rc = ll_file_join(inode, file, ftail);
2558 case LL_IOC_GROUP_LOCK:
2559 RETURN(ll_get_grouplock(inode, file, arg));
2560 case LL_IOC_GROUP_UNLOCK:
2561 RETURN(ll_put_grouplock(inode, file, arg));
2562 case IOC_OBD_STATFS:
2563 RETURN(ll_obd_statfs(inode, (void *)arg));
2565 /* We need to special case any other ioctls we want to handle,
2566 * to send them to the MDS/OST as appropriate and to properly
2567 * network encode the arg field.
2568 case EXT3_IOC_SETVERSION_OLD:
2569 case EXT3_IOC_SETVERSION:
2571 case LL_IOC_FLUSHCTX:
2572 RETURN(ll_flush_ctx(inode));
2577 ll_iocontrol_call(inode, file, cmd, arg, &err))
2580 RETURN(obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
2586 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
2588 struct inode *inode = file->f_dentry->d_inode;
2589 struct ll_inode_info *lli = ll_i2info(inode);
2590 struct lov_stripe_md *lsm = lli->lli_smd;
2593 retval = offset + ((origin == 2) ? i_size_read(inode) :
2594 (origin == 1) ? file->f_pos : 0);
2595 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), to=%Lu=%#Lx(%s)\n",
2596 inode->i_ino, inode->i_generation, inode, retval, retval,
2597 origin == 2 ? "SEEK_END": origin == 1 ? "SEEK_CUR" : "SEEK_SET");
2598 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_LLSEEK, 1);
2600 if (origin == 2) { /* SEEK_END */
2601 int nonblock = 0, rc;
2603 if (file->f_flags & O_NONBLOCK)
2604 nonblock = LDLM_FL_BLOCK_NOWAIT;
2607 rc = ll_glimpse_size(inode, nonblock);
2612 ll_inode_size_lock(inode, 0);
2613 offset += i_size_read(inode);
2614 ll_inode_size_unlock(inode, 0);
2615 } else if (origin == 1) { /* SEEK_CUR */
2616 offset += file->f_pos;
2620 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
2621 if (offset != file->f_pos) {
2622 file->f_pos = offset;
2623 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2625 file->f_version = ++event;
2634 int ll_fsync(struct file *file, struct dentry *dentry, int data)
2636 struct inode *inode = dentry->d_inode;
2637 struct ll_inode_info *lli = ll_i2info(inode);
2638 struct lov_stripe_md *lsm = lli->lli_smd;
2639 struct ptlrpc_request *req;
2640 struct obd_capa *oc;
2643 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
2644 inode->i_generation, inode);
2645 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_FSYNC, 1);
2647 /* fsync's caller has already called _fdata{sync,write}, we want
2648 * that IO to finish before calling the osc and mdc sync methods */
2649 rc = filemap_fdatawait(inode->i_mapping);
2651 /* catch async errors that were recorded back when async writeback
2652 * failed for pages in this mapping. */
2653 err = lli->lli_async_rc;
2654 lli->lli_async_rc = 0;
2658 err = lov_test_and_clear_async_rc(lsm);
2663 oc = ll_mdscapa_get(inode);
2664 err = md_sync(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode), oc,
2670 ptlrpc_req_finished(req);
2677 RETURN(rc ? rc : -ENOMEM);
2679 oa->o_id = lsm->lsm_object_id;
2680 oa->o_gr = lsm->lsm_object_gr;
2681 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2682 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
2683 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
2686 oc = ll_osscapa_get(inode, CAPA_OPC_OSS_WRITE);
2687 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
2688 0, OBD_OBJECT_EOF, oc);
2698 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
2700 struct inode *inode = file->f_dentry->d_inode;
2701 struct ll_sb_info *sbi = ll_i2sbi(inode);
2702 struct ldlm_res_id res_id =
2703 { .name = { fid_seq(ll_inode2fid(inode)),
2704 fid_oid(ll_inode2fid(inode)),
2705 fid_ver(ll_inode2fid(inode)),
2707 struct ldlm_enqueue_info einfo = { LDLM_FLOCK, 0, NULL,
2708 ldlm_flock_completion_ast, NULL, file_lock };
2709 struct lustre_handle lockh = {0};
2710 ldlm_policy_data_t flock;
2715 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
2716 inode->i_ino, file_lock);
2718 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_FLOCK, 1);
2720 if (file_lock->fl_flags & FL_FLOCK) {
2721 LASSERT((cmd == F_SETLKW) || (cmd == F_SETLK));
2722 /* set missing params for flock() calls */
2723 file_lock->fl_end = OFFSET_MAX;
2724 file_lock->fl_pid = current->tgid;
2726 flock.l_flock.pid = file_lock->fl_pid;
2727 flock.l_flock.start = file_lock->fl_start;
2728 flock.l_flock.end = file_lock->fl_end;
2730 switch (file_lock->fl_type) {
2732 einfo.ei_mode = LCK_PR;
2735 /* An unlock request may or may not have any relation to
2736 * existing locks so we may not be able to pass a lock handle
2737 * via a normal ldlm_lock_cancel() request. The request may even
2738 * unlock a byte range in the middle of an existing lock. In
2739 * order to process an unlock request we need all of the same
2740 * information that is given with a normal read or write record
2741 * lock request. To avoid creating another ldlm unlock (cancel)
2742 * message we'll treat a LCK_NL flock request as an unlock. */
2743 einfo.ei_mode = LCK_NL;
2746 einfo.ei_mode = LCK_PW;
2749 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
2764 flags = LDLM_FL_BLOCK_NOWAIT;
2770 flags = LDLM_FL_TEST_LOCK;
2771 /* Save the old mode so that if the mode in the lock changes we
2772 * can decrement the appropriate reader or writer refcount. */
2773 file_lock->fl_type = einfo.ei_mode;
2776 CERROR("unknown fcntl lock command: %d\n", cmd);
2780 CDEBUG(D_DLMTRACE, "inode=%lu, pid=%u, flags=%#x, mode=%u, "
2781 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
2782 flags, einfo.ei_mode, flock.l_flock.start, flock.l_flock.end);
2784 rc = ldlm_cli_enqueue(sbi->ll_md_exp, NULL, &einfo, &res_id,
2785 &flock, &flags, NULL, 0, NULL, &lockh, 0);
2786 if ((file_lock->fl_flags & FL_FLOCK) &&
2787 (rc == 0 || file_lock->fl_type == F_UNLCK))
2788 ll_flock_lock_file_wait(file, file_lock, (cmd == F_SETLKW));
2789 #ifdef HAVE_F_OP_FLOCK
2790 if ((file_lock->fl_flags & FL_POSIX) &&
2791 (rc == 0 || file_lock->fl_type == F_UNLCK) &&
2792 !(flags & LDLM_FL_TEST_LOCK))
2793 posix_lock_file_wait(file, file_lock);
2799 int ll_file_noflock(struct file *file, int cmd, struct file_lock *file_lock)
2806 int ll_have_md_lock(struct inode *inode, __u64 bits)
2808 struct lustre_handle lockh;
2809 ldlm_policy_data_t policy = { .l_inodebits = {bits}};
2817 fid = &ll_i2info(inode)->lli_fid;
2818 CDEBUG(D_INFO, "trying to match res "DFID"\n", PFID(fid));
2820 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING | LDLM_FL_TEST_LOCK;
2821 if (md_lock_match(ll_i2mdexp(inode), flags, fid, LDLM_IBITS, &policy,
2822 LCK_CR|LCK_CW|LCK_PR|LCK_PW, &lockh)) {
2828 ldlm_mode_t ll_take_md_lock(struct inode *inode, __u64 bits,
2829 struct lustre_handle *lockh)
2831 ldlm_policy_data_t policy = { .l_inodebits = {bits}};
2837 fid = &ll_i2info(inode)->lli_fid;
2838 CDEBUG(D_INFO, "trying to match res "DFID"\n", PFID(fid));
2840 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
2841 rc = md_lock_match(ll_i2mdexp(inode), flags, fid, LDLM_IBITS, &policy,
2842 LCK_CR|LCK_CW|LCK_PR|LCK_PW, lockh);
2846 static int ll_inode_revalidate_fini(struct inode *inode, int rc) {
2847 if (rc == -ENOENT) { /* Already unlinked. Just update nlink
2848 * and return success */
2850 /* This path cannot be hit for regular files unless in
2851 * case of obscure races, so no need to to validate
2853 if (!S_ISREG(inode->i_mode) &&
2854 !S_ISDIR(inode->i_mode))
2859 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
2867 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
2869 struct inode *inode = dentry->d_inode;
2870 struct ptlrpc_request *req = NULL;
2871 struct ll_sb_info *sbi;
2872 struct obd_export *exp;
2877 CERROR("REPORT THIS LINE TO PETER\n");
2880 sbi = ll_i2sbi(inode);
2882 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s\n",
2883 inode->i_ino, inode->i_generation, inode, dentry->d_name.name);
2885 exp = ll_i2mdexp(inode);
2887 if (exp->exp_connect_flags & OBD_CONNECT_ATTRFID) {
2888 struct lookup_intent oit = { .it_op = IT_GETATTR };
2889 struct md_op_data *op_data;
2891 /* Call getattr by fid, so do not provide name at all. */
2892 op_data = ll_prep_md_op_data(NULL, dentry->d_parent->d_inode,
2893 dentry->d_inode, NULL, 0, 0,
2894 LUSTRE_OPC_ANY, NULL);
2895 if (IS_ERR(op_data))
2896 RETURN(PTR_ERR(op_data));
2898 oit.it_flags |= O_CHECK_STALE;
2899 rc = md_intent_lock(exp, op_data, NULL, 0,
2900 /* we are not interested in name
2903 ll_md_blocking_ast, 0);
2904 ll_finish_md_op_data(op_data);
2905 oit.it_flags &= ~O_CHECK_STALE;
2907 rc = ll_inode_revalidate_fini(inode, rc);
2911 rc = ll_revalidate_it_finish(req, &oit, dentry);
2913 ll_intent_release(&oit);
2917 /* Unlinked? Unhash dentry, so it is not picked up later by
2918 do_lookup() -> ll_revalidate_it(). We cannot use d_drop
2919 here to preserve get_cwd functionality on 2.6.
2921 if (!dentry->d_inode->i_nlink) {
2922 spin_lock(&dcache_lock);
2923 ll_drop_dentry(dentry);
2924 spin_unlock(&dcache_lock);
2927 ll_lookup_finish_locks(&oit, dentry);
2928 } else if (!ll_have_md_lock(dentry->d_inode, MDS_INODELOCK_UPDATE |
2929 MDS_INODELOCK_LOOKUP)) {
2930 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
2931 obd_valid valid = OBD_MD_FLGETATTR;
2932 struct obd_capa *oc;
2935 if (S_ISREG(inode->i_mode)) {
2936 rc = ll_get_max_mdsize(sbi, &ealen);
2939 valid |= OBD_MD_FLEASIZE | OBD_MD_FLMODEASIZE;
2941 /* Once OBD_CONNECT_ATTRFID is not supported, we can't find one
2942 * capa for this inode. Because we only keep capas of dirs
2944 oc = ll_mdscapa_get(inode);
2945 rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode), oc, valid,
2949 rc = ll_inode_revalidate_fini(inode, rc);
2953 rc = ll_prep_inode(&inode, req, NULL);
2958 /* if object not yet allocated, don't validate size */
2959 if (ll_i2info(inode)->lli_smd == NULL)
2962 /* ll_glimpse_size will prefer locally cached writes if they extend
2964 rc = ll_glimpse_size(inode, 0);
2967 ptlrpc_req_finished(req);
2971 int ll_getattr_it(struct vfsmount *mnt, struct dentry *de,
2972 struct lookup_intent *it, struct kstat *stat)
2974 struct inode *inode = de->d_inode;
2977 res = ll_inode_revalidate_it(de, it);
2978 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_GETATTR, 1);
2983 stat->dev = inode->i_sb->s_dev;
2984 stat->ino = inode->i_ino;
2985 stat->mode = inode->i_mode;
2986 stat->nlink = inode->i_nlink;
2987 stat->uid = inode->i_uid;
2988 stat->gid = inode->i_gid;
2989 stat->rdev = kdev_t_to_nr(inode->i_rdev);
2990 stat->atime = inode->i_atime;
2991 stat->mtime = inode->i_mtime;
2992 stat->ctime = inode->i_ctime;
2993 #ifdef HAVE_INODE_BLKSIZE
2994 stat->blksize = inode->i_blksize;
2996 stat->blksize = 1 << inode->i_blkbits;
2999 ll_inode_size_lock(inode, 0);
3000 stat->size = i_size_read(inode);
3001 stat->blocks = inode->i_blocks;
3002 ll_inode_size_unlock(inode, 0);
3006 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
3008 struct lookup_intent it = { .it_op = IT_GETATTR };
3010 return ll_getattr_it(mnt, de, &it, stat);
3014 int lustre_check_acl(struct inode *inode, int mask)
3016 #ifdef CONFIG_FS_POSIX_ACL
3017 struct ll_inode_info *lli = ll_i2info(inode);
3018 struct posix_acl *acl;
3022 spin_lock(&lli->lli_lock);
3023 acl = posix_acl_dup(lli->lli_posix_acl);
3024 spin_unlock(&lli->lli_lock);
3029 rc = posix_acl_permission(inode, acl, mask);
3030 posix_acl_release(acl);
3038 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,10))
3039 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
3041 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), mask %o\n",
3042 inode->i_ino, inode->i_generation, inode, mask);
3043 if (ll_i2sbi(inode)->ll_flags & LL_SBI_RMT_CLIENT)
3044 return lustre_check_remote_perm(inode, mask);
3046 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_INODE_PERM, 1);
3047 return generic_permission(inode, mask, lustre_check_acl);
3050 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
3052 int mode = inode->i_mode;
3055 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), mask %o\n",
3056 inode->i_ino, inode->i_generation, inode, mask);
3058 if (ll_i2sbi(inode)->ll_flags & LL_SBI_RMT_CLIENT)
3059 return lustre_check_remote_perm(inode, mask);
3061 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_INODE_PERM, 1);
3063 if ((mask & MAY_WRITE) && IS_RDONLY(inode) &&
3064 (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
3066 if ((mask & MAY_WRITE) && IS_IMMUTABLE(inode))
3068 if (current->fsuid == inode->i_uid) {
3071 if (((mode >> 3) & mask & S_IRWXO) != mask)
3073 rc = lustre_check_acl(inode, mask);
3077 goto check_capabilities;
3081 if (in_group_p(inode->i_gid))
3084 if ((mode & mask & S_IRWXO) == mask)
3088 if (!(mask & MAY_EXEC) ||
3089 (inode->i_mode & S_IXUGO) || S_ISDIR(inode->i_mode))
3090 if (capable(CAP_DAC_OVERRIDE))
3093 if (capable(CAP_DAC_READ_SEARCH) && ((mask == MAY_READ) ||
3094 (S_ISDIR(inode->i_mode) && !(mask & MAY_WRITE))))
3101 /* -o localflock - only provides locally consistent flock locks */
3102 struct file_operations ll_file_operations = {
3103 .read = ll_file_read,
3104 .write = ll_file_write,
3105 .ioctl = ll_file_ioctl,
3106 .open = ll_file_open,
3107 .release = ll_file_release,
3108 .mmap = ll_file_mmap,
3109 .llseek = ll_file_seek,
3110 .sendfile = ll_file_sendfile,
3114 struct file_operations ll_file_operations_flock = {
3115 .read = ll_file_read,
3116 .write = ll_file_write,
3117 .ioctl = ll_file_ioctl,
3118 .open = ll_file_open,
3119 .release = ll_file_release,
3120 .mmap = ll_file_mmap,
3121 .llseek = ll_file_seek,
3122 .sendfile = ll_file_sendfile,
3124 #ifdef HAVE_F_OP_FLOCK
3125 .flock = ll_file_flock,
3127 .lock = ll_file_flock
3130 /* These are for -o noflock - to return ENOSYS on flock calls */
3131 struct file_operations ll_file_operations_noflock = {
3132 .read = ll_file_read,
3133 .write = ll_file_write,
3134 .ioctl = ll_file_ioctl,
3135 .open = ll_file_open,
3136 .release = ll_file_release,
3137 .mmap = ll_file_mmap,
3138 .llseek = ll_file_seek,
3139 .sendfile = ll_file_sendfile,
3141 #ifdef HAVE_F_OP_FLOCK
3142 .flock = ll_file_noflock,
3144 .lock = ll_file_noflock
3147 struct inode_operations ll_file_inode_operations = {
3148 #ifdef HAVE_VFS_INTENT_PATCHES
3149 .setattr_raw = ll_setattr_raw,
3151 .setattr = ll_setattr,
3152 .truncate = ll_truncate,
3153 .getattr = ll_getattr,
3154 .permission = ll_inode_permission,
3155 .setxattr = ll_setxattr,
3156 .getxattr = ll_getxattr,
3157 .listxattr = ll_listxattr,
3158 .removexattr = ll_removexattr,
3161 /* dynamic ioctl number support routins */
3162 static struct llioc_ctl_data {
3163 struct rw_semaphore ioc_sem;
3164 struct list_head ioc_head;
3166 __RWSEM_INITIALIZER(llioc.ioc_sem),
3167 CFS_LIST_HEAD_INIT(llioc.ioc_head)
3172 struct list_head iocd_list;
3173 unsigned int iocd_size;
3174 llioc_callback_t iocd_cb;
3175 unsigned int iocd_count;
3176 unsigned int iocd_cmd[0];
3179 void *ll_iocontrol_register(llioc_callback_t cb, int count, unsigned int *cmd)
3182 struct llioc_data *in_data = NULL;
3185 if (cb == NULL || cmd == NULL ||
3186 count > LLIOC_MAX_CMD || count < 0)
3189 size = sizeof(*in_data) + count * sizeof(unsigned int);
3190 OBD_ALLOC(in_data, size);
3191 if (in_data == NULL)
3194 memset(in_data, 0, sizeof(*in_data));
3195 in_data->iocd_size = size;
3196 in_data->iocd_cb = cb;
3197 in_data->iocd_count = count;
3198 memcpy(in_data->iocd_cmd, cmd, sizeof(unsigned int) * count);
3200 down_write(&llioc.ioc_sem);
3201 list_add_tail(&in_data->iocd_list, &llioc.ioc_head);
3202 up_write(&llioc.ioc_sem);
3207 void ll_iocontrol_unregister(void *magic)
3209 struct llioc_data *tmp;
3214 down_write(&llioc.ioc_sem);
3215 list_for_each_entry(tmp, &llioc.ioc_head, iocd_list) {
3217 unsigned int size = tmp->iocd_size;
3219 list_del(&tmp->iocd_list);
3220 up_write(&llioc.ioc_sem);
3222 OBD_FREE(tmp, size);
3226 up_write(&llioc.ioc_sem);
3228 CWARN("didn't find iocontrol register block with magic: %p\n", magic);
3231 EXPORT_SYMBOL(ll_iocontrol_register);
3232 EXPORT_SYMBOL(ll_iocontrol_unregister);
3234 enum llioc_iter ll_iocontrol_call(struct inode *inode, struct file *file,
3235 unsigned int cmd, unsigned long arg, int *rcp)
3237 enum llioc_iter ret = LLIOC_CONT;
3238 struct llioc_data *data;
3239 int rc = -EINVAL, i;
3241 down_read(&llioc.ioc_sem);
3242 list_for_each_entry(data, &llioc.ioc_head, iocd_list) {
3243 for (i = 0; i < data->iocd_count; i++) {
3244 if (cmd != data->iocd_cmd[i])
3247 ret = data->iocd_cb(inode, file, cmd, arg, data, &rc);
3251 if (ret == LLIOC_STOP)
3254 up_read(&llioc.ioc_sem);