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"
50 #include <lustre/ll_fiemap.h>
52 /* also used by llite/special.c:ll_special_open() */
53 struct ll_file_data *ll_file_data_get(void)
55 struct ll_file_data *fd;
57 OBD_SLAB_ALLOC_PTR(fd, ll_file_data_slab);
61 static void ll_file_data_put(struct ll_file_data *fd)
64 OBD_SLAB_FREE_PTR(fd, ll_file_data_slab);
67 void ll_pack_inode2opdata(struct inode *inode, struct md_op_data *op_data,
68 struct lustre_handle *fh)
70 op_data->op_fid1 = ll_i2info(inode)->lli_fid;
71 op_data->op_attr.ia_mode = inode->i_mode;
72 op_data->op_attr.ia_atime = inode->i_atime;
73 op_data->op_attr.ia_mtime = inode->i_mtime;
74 op_data->op_attr.ia_ctime = inode->i_ctime;
75 op_data->op_attr.ia_size = i_size_read(inode);
76 op_data->op_attr_blocks = inode->i_blocks;
77 ((struct ll_iattr *)&op_data->op_attr)->ia_attr_flags = inode->i_flags;
78 op_data->op_ioepoch = ll_i2info(inode)->lli_ioepoch;
79 memcpy(&op_data->op_handle, fh, sizeof(op_data->op_handle));
80 op_data->op_capa1 = ll_mdscapa_get(inode);
83 static void ll_prepare_close(struct inode *inode, struct md_op_data *op_data,
84 struct obd_client_handle *och)
88 op_data->op_attr.ia_valid = ATTR_MODE | ATTR_ATIME_SET |
89 ATTR_MTIME_SET | ATTR_CTIME_SET;
91 if (!(och->och_flags & FMODE_WRITE))
94 if (!(ll_i2mdexp(inode)->exp_connect_flags & OBD_CONNECT_SOM) ||
95 !S_ISREG(inode->i_mode))
96 op_data->op_attr.ia_valid |= ATTR_SIZE | ATTR_BLOCKS;
98 ll_epoch_close(inode, op_data, &och, 0);
101 ll_pack_inode2opdata(inode, op_data, &och->och_fh);
105 static int ll_close_inode_openhandle(struct obd_export *md_exp,
107 struct obd_client_handle *och)
109 struct obd_export *exp = ll_i2mdexp(inode);
110 struct md_op_data *op_data;
111 struct ptlrpc_request *req = NULL;
112 struct obd_device *obd = class_exp2obd(exp);
119 * XXX: in case of LMV, is this correct to access
122 CERROR("Invalid MDC connection handle "LPX64"\n",
123 ll_i2mdexp(inode)->exp_handle.h_cookie);
128 * here we check if this is forced umount. If so this is called on
129 * canceling "open lock" and we do not call md_close() in this case, as
130 * it will not be successful, as import is already deactivated.
135 OBD_ALLOC_PTR(op_data);
137 GOTO(out, rc = -ENOMEM); // XXX We leak openhandle and request here.
139 ll_prepare_close(inode, op_data, och);
140 epoch_close = (op_data->op_flags & MF_EPOCH_CLOSE);
141 rc = md_close(md_exp, op_data, och->och_mod, &req);
146 /* This close must have the epoch closed. */
147 LASSERT(exp->exp_connect_flags & OBD_CONNECT_SOM);
148 LASSERT(epoch_close);
149 /* MDS has instructed us to obtain Size-on-MDS attribute from
150 * OSTs and send setattr to back to MDS. */
151 rc = ll_sizeonmds_update(inode, och->och_mod,
152 &och->och_fh, op_data->op_ioepoch);
154 CERROR("inode %lu mdc Size-on-MDS update failed: "
155 "rc = %d\n", inode->i_ino, rc);
159 CERROR("inode %lu mdc close failed: rc = %d\n",
162 ll_finish_md_op_data(op_data);
165 rc = ll_objects_destroy(req, inode);
167 CERROR("inode %lu ll_objects destroy: rc = %d\n",
174 if ((exp->exp_connect_flags & OBD_CONNECT_SOM) && !epoch_close &&
175 S_ISREG(inode->i_mode) && (och->och_flags & FMODE_WRITE)) {
176 ll_queue_done_writing(inode, LLIF_DONE_WRITING);
179 ptlrpc_close_replay_seq(req);
180 md_clear_open_replay_data(md_exp, och);
181 /* Free @och if it is not waiting for DONE_WRITING. */
182 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
185 if (req) /* This is close request */
186 ptlrpc_req_finished(req);
190 int ll_md_real_close(struct inode *inode, int flags)
192 struct ll_inode_info *lli = ll_i2info(inode);
193 struct obd_client_handle **och_p;
194 struct obd_client_handle *och;
199 if (flags & FMODE_WRITE) {
200 och_p = &lli->lli_mds_write_och;
201 och_usecount = &lli->lli_open_fd_write_count;
202 } else if (flags & FMODE_EXEC) {
203 och_p = &lli->lli_mds_exec_och;
204 och_usecount = &lli->lli_open_fd_exec_count;
206 LASSERT(flags & FMODE_READ);
207 och_p = &lli->lli_mds_read_och;
208 och_usecount = &lli->lli_open_fd_read_count;
211 down(&lli->lli_och_sem);
212 if (*och_usecount) { /* There are still users of this handle, so
214 up(&lli->lli_och_sem);
219 up(&lli->lli_och_sem);
221 if (och) { /* There might be a race and somebody have freed this och
223 rc = ll_close_inode_openhandle(ll_i2sbi(inode)->ll_md_exp,
230 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
233 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
234 struct ll_inode_info *lli = ll_i2info(inode);
238 /* clear group lock, if present */
239 if (unlikely(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
240 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
241 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
242 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
246 /* Let's see if we have good enough OPEN lock on the file and if
247 we can skip talking to MDS */
248 if (file->f_dentry->d_inode) { /* Can this ever be false? */
250 int flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
251 struct lustre_handle lockh;
252 struct inode *inode = file->f_dentry->d_inode;
253 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
255 down(&lli->lli_och_sem);
256 if (fd->fd_omode & FMODE_WRITE) {
258 LASSERT(lli->lli_open_fd_write_count);
259 lli->lli_open_fd_write_count--;
260 } else if (fd->fd_omode & FMODE_EXEC) {
262 LASSERT(lli->lli_open_fd_exec_count);
263 lli->lli_open_fd_exec_count--;
266 LASSERT(lli->lli_open_fd_read_count);
267 lli->lli_open_fd_read_count--;
269 up(&lli->lli_och_sem);
271 if (!md_lock_match(md_exp, flags, ll_inode2fid(inode),
272 LDLM_IBITS, &policy, lockmode,
274 rc = ll_md_real_close(file->f_dentry->d_inode,
278 CERROR("Releasing a file %p with negative dentry %p. Name %s",
279 file, file->f_dentry, file->f_dentry->d_name.name);
282 LUSTRE_FPRIVATE(file) = NULL;
283 ll_file_data_put(fd);
284 ll_capa_close(inode);
289 int lov_test_and_clear_async_rc(struct lov_stripe_md *lsm);
291 /* While this returns an error code, fput() the caller does not, so we need
292 * to make every effort to clean up all of our state here. Also, applications
293 * rarely check close errors and even if an error is returned they will not
294 * re-try the close call.
296 int ll_file_release(struct inode *inode, struct file *file)
298 struct ll_file_data *fd;
299 struct ll_sb_info *sbi = ll_i2sbi(inode);
300 struct ll_inode_info *lli = ll_i2info(inode);
301 struct lov_stripe_md *lsm = lli->lli_smd;
305 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
306 inode->i_generation, inode);
308 #ifdef CONFIG_FS_POSIX_ACL
309 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
310 inode == inode->i_sb->s_root->d_inode) {
311 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
314 if (unlikely(fd->fd_flags & LL_FILE_RMTACL)) {
315 fd->fd_flags &= ~LL_FILE_RMTACL;
316 rct_del(&sbi->ll_rct, cfs_curproc_pid());
317 et_search_free(&sbi->ll_et, cfs_curproc_pid());
322 if (inode->i_sb->s_root != file->f_dentry)
323 ll_stats_ops_tally(sbi, LPROC_LL_RELEASE, 1);
324 fd = LUSTRE_FPRIVATE(file);
327 /* The last ref on @file, maybe not the the owner pid of statahead.
328 * Different processes can open the same dir, "ll_opendir_key" means:
329 * it is me that should stop the statahead thread. */
330 if (lli->lli_opendir_key == fd)
331 ll_stop_statahead(inode, fd);
333 if (inode->i_sb->s_root == file->f_dentry) {
334 LUSTRE_FPRIVATE(file) = NULL;
335 ll_file_data_put(fd);
340 lov_test_and_clear_async_rc(lsm);
341 lli->lli_async_rc = 0;
343 rc = ll_md_close(sbi->ll_md_exp, inode, file);
347 static int ll_intent_file_open(struct file *file, void *lmm,
348 int lmmsize, struct lookup_intent *itp)
350 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
351 struct dentry *parent = file->f_dentry->d_parent;
352 const char *name = file->f_dentry->d_name.name;
353 const int len = file->f_dentry->d_name.len;
354 struct md_op_data *op_data;
355 struct ptlrpc_request *req;
362 /* Usually we come here only for NFSD, and we want open lock.
363 But we can also get here with pre 2.6.15 patchless kernels, and in
364 that case that lock is also ok */
365 /* We can also get here if there was cached open handle in revalidate_it
366 * but it disappeared while we were getting from there to ll_file_open.
367 * But this means this file was closed and immediatelly opened which
368 * makes a good candidate for using OPEN lock */
369 /* If lmmsize & lmm are not 0, we are just setting stripe info
370 * parameters. No need for the open lock */
371 if (!lmm && !lmmsize)
372 itp->it_flags |= MDS_OPEN_LOCK;
374 op_data = ll_prep_md_op_data(NULL, parent->d_inode,
375 file->f_dentry->d_inode, name, len,
376 O_RDWR, LUSTRE_OPC_ANY, NULL);
378 RETURN(PTR_ERR(op_data));
380 rc = md_intent_lock(sbi->ll_md_exp, op_data, lmm, lmmsize, itp,
381 0 /*unused */, &req, ll_md_blocking_ast, 0);
382 ll_finish_md_op_data(op_data);
384 /* reason for keep own exit path - don`t flood log
385 * with messages with -ESTALE errors.
387 if (!it_disposition(itp, DISP_OPEN_OPEN) ||
388 it_open_error(DISP_OPEN_OPEN, itp))
390 ll_release_openhandle(file->f_dentry, itp);
394 if (rc != 0 || it_open_error(DISP_OPEN_OPEN, itp)) {
395 rc = rc ? rc : it_open_error(DISP_OPEN_OPEN, itp);
396 CDEBUG(D_VFSTRACE, "lock enqueue: err: %d\n", rc);
400 if (itp->d.lustre.it_lock_mode)
401 md_set_lock_data(sbi->ll_md_exp,
402 &itp->d.lustre.it_lock_handle,
403 file->f_dentry->d_inode);
405 rc = ll_prep_inode(&file->f_dentry->d_inode, req, NULL);
407 ptlrpc_req_finished(itp->d.lustre.it_data);
408 it_clear_disposition(itp, DISP_ENQ_COMPLETE);
409 ll_intent_drop_lock(itp);
414 static int ll_och_fill(struct obd_export *md_exp, struct ll_inode_info *lli,
415 struct lookup_intent *it, struct obd_client_handle *och)
417 struct ptlrpc_request *req = it->d.lustre.it_data;
418 struct mdt_body *body;
422 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
423 LASSERT(body != NULL); /* reply already checked out */
425 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
426 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
427 och->och_fid = lli->lli_fid;
428 och->och_flags = it->it_flags;
429 lli->lli_ioepoch = body->ioepoch;
431 return md_set_open_replay_data(md_exp, och, req);
434 int ll_local_open(struct file *file, struct lookup_intent *it,
435 struct ll_file_data *fd, struct obd_client_handle *och)
437 struct inode *inode = file->f_dentry->d_inode;
438 struct ll_inode_info *lli = ll_i2info(inode);
441 LASSERT(!LUSTRE_FPRIVATE(file));
446 struct ptlrpc_request *req = it->d.lustre.it_data;
447 struct mdt_body *body;
450 rc = ll_och_fill(ll_i2sbi(inode)->ll_md_exp, lli, it, och);
454 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
455 if ((it->it_flags & FMODE_WRITE) &&
456 (body->valid & OBD_MD_FLSIZE))
457 CDEBUG(D_INODE, "Epoch "LPU64" opened on "DFID"\n",
458 lli->lli_ioepoch, PFID(&lli->lli_fid));
461 LUSTRE_FPRIVATE(file) = fd;
462 ll_readahead_init(inode, &fd->fd_ras);
463 fd->fd_omode = it->it_flags;
467 /* Open a file, and (for the very first open) create objects on the OSTs at
468 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
469 * creation or open until ll_lov_setstripe() ioctl is called. We grab
470 * lli_open_sem to ensure no other process will create objects, send the
471 * stripe MD to the MDS, or try to destroy the objects if that fails.
473 * If we already have the stripe MD locally then we don't request it in
474 * md_open(), by passing a lmm_size = 0.
476 * It is up to the application to ensure no other processes open this file
477 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
478 * used. We might be able to avoid races of that sort by getting lli_open_sem
479 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
480 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
482 int ll_file_open(struct inode *inode, struct file *file)
484 struct ll_inode_info *lli = ll_i2info(inode);
485 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
486 .it_flags = file->f_flags };
487 struct lov_stripe_md *lsm;
488 struct ptlrpc_request *req = NULL;
489 struct obd_client_handle **och_p;
491 struct ll_file_data *fd;
492 int rc = 0, opendir_set = 0;
495 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n", inode->i_ino,
496 inode->i_generation, inode, file->f_flags);
498 #ifdef HAVE_VFS_INTENT_PATCHES
501 it = file->private_data; /* XXX: compat macro */
502 file->private_data = NULL; /* prevent ll_local_open assertion */
505 fd = ll_file_data_get();
509 if (S_ISDIR(inode->i_mode)) {
510 spin_lock(&lli->lli_lock);
511 /* "lli->lli_opendir_pid != 0" means someone has set it.
512 * "lli->lli_sai != NULL" means the previous statahead has not
514 if (lli->lli_opendir_pid == 0 && lli->lli_sai == NULL) {
516 lli->lli_opendir_pid = cfs_curproc_pid();
517 lli->lli_opendir_key = fd;
518 } else if (unlikely(lli->lli_opendir_pid == cfs_curproc_pid())) {
519 /* Two cases for this:
520 * (1) The same process open such directory many times.
521 * (2) The old process opened the directory, and exited
522 * before its children processes. Then new process
523 * with the same pid opens such directory before the
524 * old process's children processes exit.
525 * Change the owner to the latest one. */
527 lli->lli_opendir_key = fd;
529 spin_unlock(&lli->lli_lock);
532 if (inode->i_sb->s_root == file->f_dentry) {
533 LUSTRE_FPRIVATE(file) = fd;
537 if (!it || !it->d.lustre.it_disposition) {
538 /* Convert f_flags into access mode. We cannot use file->f_mode,
539 * because everything but O_ACCMODE mask was stripped from
541 if ((oit.it_flags + 1) & O_ACCMODE)
543 if (file->f_flags & O_TRUNC)
544 oit.it_flags |= FMODE_WRITE;
546 /* kernel only call f_op->open in dentry_open. filp_open calls
547 * dentry_open after call to open_namei that checks permissions.
548 * Only nfsd_open call dentry_open directly without checking
549 * permissions and because of that this code below is safe. */
550 if (oit.it_flags & FMODE_WRITE)
551 oit.it_flags |= MDS_OPEN_OWNEROVERRIDE;
553 /* We do not want O_EXCL here, presumably we opened the file
554 * already? XXX - NFS implications? */
555 oit.it_flags &= ~O_EXCL;
561 /* Let's see if we have file open on MDS already. */
562 if (it->it_flags & FMODE_WRITE) {
563 och_p = &lli->lli_mds_write_och;
564 och_usecount = &lli->lli_open_fd_write_count;
565 } else if (it->it_flags & FMODE_EXEC) {
566 och_p = &lli->lli_mds_exec_och;
567 och_usecount = &lli->lli_open_fd_exec_count;
569 och_p = &lli->lli_mds_read_och;
570 och_usecount = &lli->lli_open_fd_read_count;
573 down(&lli->lli_och_sem);
574 if (*och_p) { /* Open handle is present */
575 if (it_disposition(it, DISP_OPEN_OPEN)) {
576 /* Well, there's extra open request that we do not need,
577 let's close it somehow. This will decref request. */
578 rc = it_open_error(DISP_OPEN_OPEN, it);
580 up(&lli->lli_och_sem);
581 ll_file_data_put(fd);
582 GOTO(out_openerr, 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);
593 up(&lli->lli_och_sem);
594 ll_file_data_put(fd);
595 GOTO(out_openerr, rc);
598 LASSERT(*och_usecount == 0);
599 if (!it->d.lustre.it_disposition) {
600 /* We cannot just request lock handle now, new ELC code
601 means that one of other OPEN locks for this file
602 could be cancelled, and since blocking ast handler
603 would attempt to grab och_sem as well, that would
604 result in a deadlock */
605 up(&lli->lli_och_sem);
606 it->it_flags |= O_CHECK_STALE;
607 rc = ll_intent_file_open(file, NULL, 0, it);
608 it->it_flags &= ~O_CHECK_STALE;
610 ll_file_data_put(fd);
611 GOTO(out_openerr, rc);
614 /* Got some error? Release the request */
615 if (it->d.lustre.it_status < 0) {
616 req = it->d.lustre.it_data;
617 ptlrpc_req_finished(req);
619 md_set_lock_data(ll_i2sbi(inode)->ll_md_exp,
620 &it->d.lustre.it_lock_handle,
621 file->f_dentry->d_inode);
624 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
626 ll_file_data_put(fd);
627 GOTO(out_och_free, rc = -ENOMEM);
630 req = it->d.lustre.it_data;
632 /* md_intent_lock() didn't get a request ref if there was an
633 * open error, so don't do cleanup on the request here
635 /* XXX (green): Should not we bail out on any error here, not
636 * just open error? */
637 rc = it_open_error(DISP_OPEN_OPEN, it);
639 ll_file_data_put(fd);
640 GOTO(out_och_free, rc);
643 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_OPEN, 1);
644 rc = ll_local_open(file, it, fd, *och_p);
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 (!cfs_capable(CFS_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);
2094 if ((lmm->lmm_magic != cpu_to_le32(LOV_MAGIC_V1)) &&
2095 (lmm->lmm_magic != cpu_to_le32(LOV_MAGIC_V3)) &&
2096 (lmm->lmm_magic != cpu_to_le32(LOV_MAGIC_JOIN))) {
2097 GOTO(out, rc = -EPROTO);
2101 * This is coming from the MDS, so is probably in
2102 * little endian. We convert it to host endian before
2103 * passing it to userspace.
2105 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC)) {
2106 /* if function called for directory - we should
2107 * avoid swab not existent lsm objects */
2108 if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) {
2109 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
2110 if (S_ISREG(body->mode))
2111 lustre_swab_lov_user_md_objects(
2112 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
2113 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
2114 } else if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)) {
2115 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
2116 if (S_ISREG(body->mode))
2117 lustre_swab_lov_user_md_objects(
2118 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
2119 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
2120 } else if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_JOIN)) {
2121 lustre_swab_lov_user_md_join((struct lov_user_md_join *)lmm);
2125 if (lmm->lmm_magic == LOV_MAGIC_JOIN) {
2126 struct lov_stripe_md *lsm;
2127 struct lov_user_md_join *lmj;
2128 int lmj_size, i, aindex = 0;
2130 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
2132 GOTO(out, rc = -ENOMEM);
2133 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
2135 GOTO(out_free_memmd, rc);
2137 lmj_size = sizeof(struct lov_user_md_join) +
2138 lsm->lsm_stripe_count *
2139 sizeof(struct lov_user_ost_data_join);
2140 OBD_ALLOC(lmj, lmj_size);
2142 GOTO(out_free_memmd, rc = -ENOMEM);
2144 memcpy(lmj, lmm, sizeof(struct lov_user_md_join));
2145 for (i = 0; i < lsm->lsm_stripe_count; i++) {
2146 struct lov_extent *lex =
2147 &lsm->lsm_array->lai_ext_array[aindex];
2149 if (lex->le_loi_idx + lex->le_stripe_count <= i)
2151 CDEBUG(D_INFO, "aindex %d i %d l_extent_start "
2152 LPU64" len %d\n", aindex, i,
2153 lex->le_start, (int)lex->le_len);
2154 lmj->lmm_objects[i].l_extent_start =
2157 if ((int)lex->le_len == -1)
2158 lmj->lmm_objects[i].l_extent_end = -1;
2160 lmj->lmm_objects[i].l_extent_end =
2161 lex->le_start + lex->le_len;
2162 lmj->lmm_objects[i].l_object_id =
2163 lsm->lsm_oinfo[i]->loi_id;
2164 lmj->lmm_objects[i].l_object_gr =
2165 lsm->lsm_oinfo[i]->loi_gr;
2166 lmj->lmm_objects[i].l_ost_gen =
2167 lsm->lsm_oinfo[i]->loi_ost_gen;
2168 lmj->lmm_objects[i].l_ost_idx =
2169 lsm->lsm_oinfo[i]->loi_ost_idx;
2171 lmm = (struct lov_mds_md *)lmj;
2174 obd_free_memmd(sbi->ll_dt_exp, &lsm);
2178 *lmm_size = lmmsize;
2183 static int ll_lov_setea(struct inode *inode, struct file *file,
2186 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
2187 struct lov_user_md *lump;
2188 int lum_size = sizeof(struct lov_user_md) +
2189 sizeof(struct lov_user_ost_data);
2193 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
2196 OBD_ALLOC(lump, lum_size);
2200 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
2202 OBD_FREE(lump, lum_size);
2206 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
2208 OBD_FREE(lump, lum_size);
2212 static int ll_lov_setstripe(struct inode *inode, struct file *file,
2215 struct lov_user_md_v3 lumv3;
2216 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
2217 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
2218 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
2221 int flags = FMODE_WRITE;
2224 /* first try with v1 which is smaller than v3 */
2225 lum_size = sizeof(struct lov_user_md_v1);
2226 rc = copy_from_user(lumv1, lumv1p, lum_size);
2230 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
2231 lum_size = sizeof(struct lov_user_md_v3);
2232 rc = copy_from_user(&lumv3, lumv3p, lum_size);
2237 rc = ll_lov_setstripe_ea_info(inode, file, flags, lumv1, lum_size);
2239 put_user(0, &lumv1p->lmm_stripe_count);
2240 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
2241 0, ll_i2info(inode)->lli_smd,
2247 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
2249 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
2254 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
2258 static int ll_get_grouplock(struct inode *inode, struct file *file,
2261 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
2262 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
2263 .end = OBD_OBJECT_EOF}};
2264 struct lustre_handle lockh = { 0 };
2265 struct ll_inode_info *lli = ll_i2info(inode);
2266 struct lov_stripe_md *lsm = lli->lli_smd;
2270 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
2274 policy.l_extent.gid = arg;
2275 if (file->f_flags & O_NONBLOCK)
2276 flags = LDLM_FL_BLOCK_NOWAIT;
2278 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags);
2282 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
2284 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
2289 static int ll_put_grouplock(struct inode *inode, struct file *file,
2292 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
2293 struct ll_inode_info *lli = ll_i2info(inode);
2294 struct lov_stripe_md *lsm = lli->lli_smd;
2298 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
2299 /* Ugh, it's already unlocked. */
2303 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
2306 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
2308 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
2313 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
2318 static int join_sanity_check(struct inode *head, struct inode *tail)
2321 if ((ll_i2sbi(head)->ll_flags & LL_SBI_JOIN) == 0) {
2322 CERROR("server do not support join \n");
2325 if (!S_ISREG(tail->i_mode) || !S_ISREG(head->i_mode)) {
2326 CERROR("tail ino %lu and ino head %lu must be regular\n",
2327 head->i_ino, tail->i_ino);
2330 if (head->i_ino == tail->i_ino) {
2331 CERROR("file %lu can not be joined to itself \n", head->i_ino);
2334 if (i_size_read(head) % JOIN_FILE_ALIGN) {
2335 CERROR("hsize %llu must be times of 64K\n", i_size_read(head));
2341 static int join_file(struct inode *head_inode, struct file *head_filp,
2342 struct file *tail_filp)
2344 struct dentry *tail_dentry = tail_filp->f_dentry;
2345 struct lookup_intent oit = {.it_op = IT_OPEN,
2346 .it_flags = head_filp->f_flags|O_JOIN_FILE};
2347 struct ldlm_enqueue_info einfo = { LDLM_IBITS, LCK_CW,
2348 ll_md_blocking_ast, ldlm_completion_ast, NULL, NULL };
2350 struct lustre_handle lockh;
2351 struct md_op_data *op_data;
2356 tail_dentry = tail_filp->f_dentry;
2358 data = i_size_read(head_inode);
2359 op_data = ll_prep_md_op_data(NULL, head_inode,
2360 tail_dentry->d_parent->d_inode,
2361 tail_dentry->d_name.name,
2362 tail_dentry->d_name.len, 0,
2363 LUSTRE_OPC_ANY, &data);
2364 if (IS_ERR(op_data))
2365 RETURN(PTR_ERR(op_data));
2367 rc = md_enqueue(ll_i2mdexp(head_inode), &einfo, &oit,
2368 op_data, &lockh, NULL, 0, NULL, 0);
2370 ll_finish_md_op_data(op_data);
2374 rc = oit.d.lustre.it_status;
2376 if (rc < 0 || it_open_error(DISP_OPEN_OPEN, &oit)) {
2377 rc = rc ? rc : it_open_error(DISP_OPEN_OPEN, &oit);
2378 ptlrpc_req_finished((struct ptlrpc_request *)
2379 oit.d.lustre.it_data);
2383 if (oit.d.lustre.it_lock_mode) { /* If we got lock - release it right
2385 ldlm_lock_decref(&lockh, oit.d.lustre.it_lock_mode);
2386 oit.d.lustre.it_lock_mode = 0;
2388 ptlrpc_req_finished((struct ptlrpc_request *) oit.d.lustre.it_data);
2389 it_clear_disposition(&oit, DISP_ENQ_COMPLETE);
2390 ll_release_openhandle(head_filp->f_dentry, &oit);
2392 ll_intent_release(&oit);
2396 static int ll_file_join(struct inode *head, struct file *filp,
2397 char *filename_tail)
2399 struct inode *tail = NULL, *first = NULL, *second = NULL;
2400 struct dentry *tail_dentry;
2401 struct file *tail_filp, *first_filp, *second_filp;
2402 struct ll_lock_tree first_tree, second_tree;
2403 struct ll_lock_tree_node *first_node, *second_node;
2404 struct ll_inode_info *hlli = ll_i2info(head), *tlli;
2405 int rc = 0, cleanup_phase = 0;
2408 CDEBUG(D_VFSTRACE, "VFS Op:head=%lu/%u(%p) tail %s\n",
2409 head->i_ino, head->i_generation, head, filename_tail);
2411 tail_filp = filp_open(filename_tail, O_WRONLY, 0644);
2412 if (IS_ERR(tail_filp)) {
2413 CERROR("Can not open tail file %s", filename_tail);
2414 rc = PTR_ERR(tail_filp);
2417 tail = igrab(tail_filp->f_dentry->d_inode);
2419 tlli = ll_i2info(tail);
2420 tail_dentry = tail_filp->f_dentry;
2421 LASSERT(tail_dentry);
2424 /*reorder the inode for lock sequence*/
2425 first = head->i_ino > tail->i_ino ? head : tail;
2426 second = head->i_ino > tail->i_ino ? tail : head;
2427 first_filp = head->i_ino > tail->i_ino ? filp : tail_filp;
2428 second_filp = head->i_ino > tail->i_ino ? tail_filp : filp;
2430 CDEBUG(D_INFO, "reorder object from %lu:%lu to %lu:%lu \n",
2431 head->i_ino, tail->i_ino, first->i_ino, second->i_ino);
2432 first_node = ll_node_from_inode(first, 0, OBD_OBJECT_EOF, LCK_EX);
2433 if (IS_ERR(first_node)){
2434 rc = PTR_ERR(first_node);
2437 first_tree.lt_fd = first_filp->private_data;
2438 rc = ll_tree_lock(&first_tree, first_node, NULL, 0, 0);
2443 second_node = ll_node_from_inode(second, 0, OBD_OBJECT_EOF, LCK_EX);
2444 if (IS_ERR(second_node)){
2445 rc = PTR_ERR(second_node);
2448 second_tree.lt_fd = second_filp->private_data;
2449 rc = ll_tree_lock(&second_tree, second_node, NULL, 0, 0);
2454 rc = join_sanity_check(head, tail);
2458 rc = join_file(head, filp, tail_filp);
2462 switch (cleanup_phase) {
2464 ll_tree_unlock(&second_tree);
2465 obd_cancel_unused(ll_i2dtexp(second),
2466 ll_i2info(second)->lli_smd, 0, NULL);
2468 ll_tree_unlock(&first_tree);
2469 obd_cancel_unused(ll_i2dtexp(first),
2470 ll_i2info(first)->lli_smd, 0, NULL);
2472 filp_close(tail_filp, 0);
2475 if (head && rc == 0) {
2476 obd_free_memmd(ll_i2sbi(head)->ll_dt_exp,
2478 hlli->lli_smd = NULL;
2483 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2490 * Close inode open handle
2492 * \param dentry [in] dentry which contains the inode
2493 * \param it [in,out] intent which contains open info and result
2496 * \retval <0 failure
2498 int ll_release_openhandle(struct dentry *dentry, struct lookup_intent *it)
2500 struct inode *inode = dentry->d_inode;
2501 struct obd_client_handle *och;
2507 /* Root ? Do nothing. */
2508 if (dentry->d_inode->i_sb->s_root == dentry)
2511 /* No open handle to close? Move away */
2512 if (!it_disposition(it, DISP_OPEN_OPEN))
2515 LASSERT(it_open_error(DISP_OPEN_OPEN, it) == 0);
2517 OBD_ALLOC(och, sizeof(*och));
2519 GOTO(out, rc = -ENOMEM);
2521 ll_och_fill(ll_i2sbi(inode)->ll_md_exp,
2522 ll_i2info(inode), it, och);
2524 rc = ll_close_inode_openhandle(ll_i2sbi(inode)->ll_md_exp,
2527 /* this one is in place of ll_file_open */
2528 if (it_disposition(it, DISP_ENQ_OPEN_REF))
2529 ptlrpc_req_finished(it->d.lustre.it_data);
2530 it_clear_disposition(it, DISP_ENQ_OPEN_REF);
2535 * Get size for inode for which FIEMAP mapping is requested.
2536 * Make the FIEMAP get_info call and returns the result.
2538 int ll_fiemap(struct inode *inode, struct ll_user_fiemap *fiemap,
2541 struct obd_export *exp = ll_i2dtexp(inode);
2542 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
2543 struct ll_fiemap_info_key fm_key = { .name = KEY_FIEMAP, };
2544 int vallen = num_bytes;
2548 /* If the stripe_count > 1 and the application does not understand
2549 * DEVICE_ORDER flag, then it cannot interpret the extents correctly.
2551 if (lsm->lsm_stripe_count > 1 &&
2552 !(fiemap->fm_flags & FIEMAP_FLAG_DEVICE_ORDER))
2555 fm_key.oa.o_id = lsm->lsm_object_id;
2556 fm_key.oa.o_gr = lsm->lsm_object_gr;
2557 fm_key.oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2559 obdo_from_inode(&fm_key.oa, inode, OBD_MD_FLFID | OBD_MD_FLGROUP |
2562 /* If filesize is 0, then there would be no objects for mapping */
2563 if (fm_key.oa.o_size == 0) {
2564 fiemap->fm_mapped_extents = 0;
2568 memcpy(&fm_key.fiemap, fiemap, sizeof(*fiemap));
2570 rc = obd_get_info(exp, sizeof(fm_key), &fm_key, &vallen, fiemap, lsm);
2572 CERROR("obd_get_info failed: rc = %d\n", rc);
2577 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
2580 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
2584 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
2585 inode->i_generation, inode, cmd);
2586 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
2588 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
2589 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
2593 case LL_IOC_GETFLAGS:
2594 /* Get the current value of the file flags */
2595 return put_user(fd->fd_flags, (int *)arg);
2596 case LL_IOC_SETFLAGS:
2597 case LL_IOC_CLRFLAGS:
2598 /* Set or clear specific file flags */
2599 /* XXX This probably needs checks to ensure the flags are
2600 * not abused, and to handle any flag side effects.
2602 if (get_user(flags, (int *) arg))
2605 if (cmd == LL_IOC_SETFLAGS) {
2606 if ((flags & LL_FILE_IGNORE_LOCK) &&
2607 !(file->f_flags & O_DIRECT)) {
2608 CERROR("%s: unable to disable locking on "
2609 "non-O_DIRECT file\n", current->comm);
2613 fd->fd_flags |= flags;
2615 fd->fd_flags &= ~flags;
2618 case LL_IOC_LOV_SETSTRIPE:
2619 RETURN(ll_lov_setstripe(inode, file, arg));
2620 case LL_IOC_LOV_SETEA:
2621 RETURN(ll_lov_setea(inode, file, arg));
2622 case LL_IOC_LOV_GETSTRIPE:
2623 RETURN(ll_lov_getstripe(inode, arg));
2624 case LL_IOC_RECREATE_OBJ:
2625 RETURN(ll_lov_recreate_obj(inode, file, arg));
2626 case EXT3_IOC_FIEMAP: {
2627 struct ll_user_fiemap *fiemap_s;
2628 size_t num_bytes, ret_bytes;
2629 unsigned int extent_count;
2632 /* Get the extent count so we can calculate the size of
2633 * required fiemap buffer */
2634 if (get_user(extent_count,
2635 &((struct ll_user_fiemap __user *)arg)->fm_extent_count))
2637 num_bytes = sizeof(*fiemap_s) + (extent_count *
2638 sizeof(struct ll_fiemap_extent));
2639 OBD_VMALLOC(fiemap_s, num_bytes);
2640 if (fiemap_s == NULL)
2643 if (copy_from_user(fiemap_s,(struct ll_user_fiemap __user *)arg,
2645 GOTO(error, rc = -EFAULT);
2647 if (fiemap_s->fm_flags & ~LUSTRE_FIEMAP_FLAGS_COMPAT) {
2648 fiemap_s->fm_flags = fiemap_s->fm_flags &
2649 ~LUSTRE_FIEMAP_FLAGS_COMPAT;
2650 if (copy_to_user((char *)arg, fiemap_s,
2652 GOTO(error, rc = -EFAULT);
2654 GOTO(error, rc = -EBADR);
2657 /* If fm_extent_count is non-zero, read the first extent since
2658 * it is used to calculate end_offset and device from previous
2661 if (copy_from_user(&fiemap_s->fm_extents[0],
2662 (char __user *)arg + sizeof(*fiemap_s),
2663 sizeof(struct ll_fiemap_extent)))
2664 GOTO(error, rc = -EFAULT);
2667 if (fiemap_s->fm_flags & FIEMAP_FLAG_SYNC) {
2670 rc = filemap_fdatawrite(inode->i_mapping);
2675 rc = ll_fiemap(inode, fiemap_s, num_bytes);
2679 ret_bytes = sizeof(struct ll_user_fiemap);
2681 if (extent_count != 0)
2682 ret_bytes += (fiemap_s->fm_mapped_extents *
2683 sizeof(struct ll_fiemap_extent));
2685 if (copy_to_user((void *)arg, fiemap_s, ret_bytes))
2689 OBD_VFREE(fiemap_s, num_bytes);
2692 case EXT3_IOC_GETFLAGS:
2693 case EXT3_IOC_SETFLAGS:
2694 RETURN(ll_iocontrol(inode, file, cmd, arg));
2695 case EXT3_IOC_GETVERSION_OLD:
2696 case EXT3_IOC_GETVERSION:
2697 RETURN(put_user(inode->i_generation, (int *)arg));
2702 ftail = getname((const char *)arg);
2704 RETURN(PTR_ERR(ftail));
2705 rc = ll_file_join(inode, file, ftail);
2709 case LL_IOC_GROUP_LOCK:
2710 RETURN(ll_get_grouplock(inode, file, arg));
2711 case LL_IOC_GROUP_UNLOCK:
2712 RETURN(ll_put_grouplock(inode, file, arg));
2713 case IOC_OBD_STATFS:
2714 RETURN(ll_obd_statfs(inode, (void *)arg));
2716 /* We need to special case any other ioctls we want to handle,
2717 * to send them to the MDS/OST as appropriate and to properly
2718 * network encode the arg field.
2719 case EXT3_IOC_SETVERSION_OLD:
2720 case EXT3_IOC_SETVERSION:
2722 case LL_IOC_FLUSHCTX:
2723 RETURN(ll_flush_ctx(inode));
2728 ll_iocontrol_call(inode, file, cmd, arg, &err))
2731 RETURN(obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
2737 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
2739 struct inode *inode = file->f_dentry->d_inode;
2740 struct ll_inode_info *lli = ll_i2info(inode);
2741 struct lov_stripe_md *lsm = lli->lli_smd;
2744 retval = offset + ((origin == 2) ? i_size_read(inode) :
2745 (origin == 1) ? file->f_pos : 0);
2746 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), to=%Lu=%#Lx(%s)\n",
2747 inode->i_ino, inode->i_generation, inode, retval, retval,
2748 origin == 2 ? "SEEK_END": origin == 1 ? "SEEK_CUR" : "SEEK_SET");
2749 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_LLSEEK, 1);
2751 if (origin == 2) { /* SEEK_END */
2752 int nonblock = 0, rc;
2754 if (file->f_flags & O_NONBLOCK)
2755 nonblock = LDLM_FL_BLOCK_NOWAIT;
2758 rc = ll_glimpse_size(inode, nonblock);
2763 ll_inode_size_lock(inode, 0);
2764 offset += i_size_read(inode);
2765 ll_inode_size_unlock(inode, 0);
2766 } else if (origin == 1) { /* SEEK_CUR */
2767 offset += file->f_pos;
2771 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
2772 if (offset != file->f_pos) {
2773 file->f_pos = offset;
2774 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2776 file->f_version = ++event;
2785 int ll_fsync(struct file *file, struct dentry *dentry, int data)
2787 struct inode *inode = dentry->d_inode;
2788 struct ll_inode_info *lli = ll_i2info(inode);
2789 struct lov_stripe_md *lsm = lli->lli_smd;
2790 struct ptlrpc_request *req;
2791 struct obd_capa *oc;
2794 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
2795 inode->i_generation, inode);
2796 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_FSYNC, 1);
2798 /* fsync's caller has already called _fdata{sync,write}, we want
2799 * that IO to finish before calling the osc and mdc sync methods */
2800 rc = filemap_fdatawait(inode->i_mapping);
2802 /* catch async errors that were recorded back when async writeback
2803 * failed for pages in this mapping. */
2804 err = lli->lli_async_rc;
2805 lli->lli_async_rc = 0;
2809 err = lov_test_and_clear_async_rc(lsm);
2814 oc = ll_mdscapa_get(inode);
2815 err = md_sync(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode), oc,
2821 ptlrpc_req_finished(req);
2828 RETURN(rc ? rc : -ENOMEM);
2830 oa->o_id = lsm->lsm_object_id;
2831 oa->o_gr = lsm->lsm_object_gr;
2832 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2833 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
2834 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
2837 oc = ll_osscapa_get(inode, CAPA_OPC_OSS_WRITE);
2838 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
2839 0, OBD_OBJECT_EOF, oc);
2849 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
2851 struct inode *inode = file->f_dentry->d_inode;
2852 struct ll_sb_info *sbi = ll_i2sbi(inode);
2853 struct ldlm_enqueue_info einfo = { .ei_type = LDLM_FLOCK,
2854 .ei_cb_cp =ldlm_flock_completion_ast,
2855 .ei_cbdata = file_lock };
2856 struct md_op_data *op_data;
2857 struct lustre_handle lockh = {0};
2858 ldlm_policy_data_t flock;
2863 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
2864 inode->i_ino, file_lock);
2866 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_FLOCK, 1);
2868 if (file_lock->fl_flags & FL_FLOCK) {
2869 LASSERT((cmd == F_SETLKW) || (cmd == F_SETLK));
2870 /* set missing params for flock() calls */
2871 file_lock->fl_end = OFFSET_MAX;
2872 file_lock->fl_pid = current->tgid;
2874 flock.l_flock.pid = file_lock->fl_pid;
2875 flock.l_flock.start = file_lock->fl_start;
2876 flock.l_flock.end = file_lock->fl_end;
2878 switch (file_lock->fl_type) {
2880 einfo.ei_mode = LCK_PR;
2883 /* An unlock request may or may not have any relation to
2884 * existing locks so we may not be able to pass a lock handle
2885 * via a normal ldlm_lock_cancel() request. The request may even
2886 * unlock a byte range in the middle of an existing lock. In
2887 * order to process an unlock request we need all of the same
2888 * information that is given with a normal read or write record
2889 * lock request. To avoid creating another ldlm unlock (cancel)
2890 * message we'll treat a LCK_NL flock request as an unlock. */
2891 einfo.ei_mode = LCK_NL;
2894 einfo.ei_mode = LCK_PW;
2897 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
2912 flags = LDLM_FL_BLOCK_NOWAIT;
2918 flags = LDLM_FL_TEST_LOCK;
2919 /* Save the old mode so that if the mode in the lock changes we
2920 * can decrement the appropriate reader or writer refcount. */
2921 file_lock->fl_type = einfo.ei_mode;
2924 CERROR("unknown fcntl lock command: %d\n", cmd);
2928 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
2929 LUSTRE_OPC_ANY, NULL);
2930 if (IS_ERR(op_data))
2931 RETURN(PTR_ERR(op_data));
2933 CDEBUG(D_DLMTRACE, "inode=%lu, pid=%u, flags=%#x, mode=%u, "
2934 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
2935 flags, einfo.ei_mode, flock.l_flock.start, flock.l_flock.end);
2937 rc = md_enqueue(sbi->ll_md_exp, &einfo, NULL,
2938 op_data, &lockh, &flock, 0, NULL /* req */, flags);
2940 ll_finish_md_op_data(op_data);
2942 if ((file_lock->fl_flags & FL_FLOCK) &&
2943 (rc == 0 || file_lock->fl_type == F_UNLCK))
2944 ll_flock_lock_file_wait(file, file_lock, (cmd == F_SETLKW));
2945 #ifdef HAVE_F_OP_FLOCK
2946 if ((file_lock->fl_flags & FL_POSIX) &&
2947 (rc == 0 || file_lock->fl_type == F_UNLCK) &&
2948 !(flags & LDLM_FL_TEST_LOCK))
2949 posix_lock_file_wait(file, file_lock);
2955 int ll_file_noflock(struct file *file, int cmd, struct file_lock *file_lock)
2962 int ll_have_md_lock(struct inode *inode, __u64 bits)
2964 struct lustre_handle lockh;
2965 ldlm_policy_data_t policy = { .l_inodebits = {bits}};
2973 fid = &ll_i2info(inode)->lli_fid;
2974 CDEBUG(D_INFO, "trying to match res "DFID"\n", PFID(fid));
2976 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING | LDLM_FL_TEST_LOCK;
2977 if (md_lock_match(ll_i2mdexp(inode), flags, fid, LDLM_IBITS, &policy,
2978 LCK_CR|LCK_CW|LCK_PR|LCK_PW, &lockh)) {
2984 ldlm_mode_t ll_take_md_lock(struct inode *inode, __u64 bits,
2985 struct lustre_handle *lockh)
2987 ldlm_policy_data_t policy = { .l_inodebits = {bits}};
2993 fid = &ll_i2info(inode)->lli_fid;
2994 CDEBUG(D_INFO, "trying to match res "DFID"\n", PFID(fid));
2996 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
2997 rc = md_lock_match(ll_i2mdexp(inode), flags, fid, LDLM_IBITS, &policy,
2998 LCK_CR|LCK_CW|LCK_PR|LCK_PW, lockh);
3002 static int ll_inode_revalidate_fini(struct inode *inode, int rc) {
3003 if (rc == -ENOENT) { /* Already unlinked. Just update nlink
3004 * and return success */
3006 /* This path cannot be hit for regular files unless in
3007 * case of obscure races, so no need to to validate
3009 if (!S_ISREG(inode->i_mode) &&
3010 !S_ISDIR(inode->i_mode))
3015 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
3023 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
3025 struct inode *inode = dentry->d_inode;
3026 struct ptlrpc_request *req = NULL;
3027 struct ll_sb_info *sbi;
3028 struct obd_export *exp;
3033 CERROR("REPORT THIS LINE TO PETER\n");
3036 sbi = ll_i2sbi(inode);
3038 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s\n",
3039 inode->i_ino, inode->i_generation, inode, dentry->d_name.name);
3041 exp = ll_i2mdexp(inode);
3043 if (exp->exp_connect_flags & OBD_CONNECT_ATTRFID) {
3044 struct lookup_intent oit = { .it_op = IT_GETATTR };
3045 struct md_op_data *op_data;
3047 /* Call getattr by fid, so do not provide name at all. */
3048 op_data = ll_prep_md_op_data(NULL, dentry->d_parent->d_inode,
3049 dentry->d_inode, NULL, 0, 0,
3050 LUSTRE_OPC_ANY, NULL);
3051 if (IS_ERR(op_data))
3052 RETURN(PTR_ERR(op_data));
3054 oit.it_flags |= O_CHECK_STALE;
3055 rc = md_intent_lock(exp, op_data, NULL, 0,
3056 /* we are not interested in name
3059 ll_md_blocking_ast, 0);
3060 ll_finish_md_op_data(op_data);
3061 oit.it_flags &= ~O_CHECK_STALE;
3063 rc = ll_inode_revalidate_fini(inode, rc);
3067 rc = ll_revalidate_it_finish(req, &oit, dentry);
3069 ll_intent_release(&oit);
3073 /* Unlinked? Unhash dentry, so it is not picked up later by
3074 do_lookup() -> ll_revalidate_it(). We cannot use d_drop
3075 here to preserve get_cwd functionality on 2.6.
3077 if (!dentry->d_inode->i_nlink) {
3078 spin_lock(&ll_lookup_lock);
3079 spin_lock(&dcache_lock);
3080 ll_drop_dentry(dentry);
3081 spin_unlock(&dcache_lock);
3082 spin_unlock(&ll_lookup_lock);
3085 ll_lookup_finish_locks(&oit, dentry);
3086 } else if (!ll_have_md_lock(dentry->d_inode, MDS_INODELOCK_UPDATE |
3087 MDS_INODELOCK_LOOKUP)) {
3088 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
3089 obd_valid valid = OBD_MD_FLGETATTR;
3090 struct obd_capa *oc;
3093 if (S_ISREG(inode->i_mode)) {
3094 rc = ll_get_max_mdsize(sbi, &ealen);
3097 valid |= OBD_MD_FLEASIZE | OBD_MD_FLMODEASIZE;
3099 /* Once OBD_CONNECT_ATTRFID is not supported, we can't find one
3100 * capa for this inode. Because we only keep capas of dirs
3102 oc = ll_mdscapa_get(inode);
3103 rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode), oc, valid,
3107 rc = ll_inode_revalidate_fini(inode, rc);
3111 rc = ll_prep_inode(&inode, req, NULL);
3116 /* if object not yet allocated, don't validate size */
3117 if (ll_i2info(inode)->lli_smd == NULL)
3120 /* ll_glimpse_size will prefer locally cached writes if they extend
3122 rc = ll_glimpse_size(inode, 0);
3125 ptlrpc_req_finished(req);
3129 int ll_getattr_it(struct vfsmount *mnt, struct dentry *de,
3130 struct lookup_intent *it, struct kstat *stat)
3132 struct inode *inode = de->d_inode;
3135 res = ll_inode_revalidate_it(de, it);
3136 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_GETATTR, 1);
3141 stat->dev = inode->i_sb->s_dev;
3142 stat->ino = inode->i_ino;
3143 stat->mode = inode->i_mode;
3144 stat->nlink = inode->i_nlink;
3145 stat->uid = inode->i_uid;
3146 stat->gid = inode->i_gid;
3147 stat->rdev = kdev_t_to_nr(inode->i_rdev);
3148 stat->atime = inode->i_atime;
3149 stat->mtime = inode->i_mtime;
3150 stat->ctime = inode->i_ctime;
3151 #ifdef HAVE_INODE_BLKSIZE
3152 stat->blksize = inode->i_blksize;
3154 stat->blksize = 1 << inode->i_blkbits;
3157 ll_inode_size_lock(inode, 0);
3158 stat->size = i_size_read(inode);
3159 stat->blocks = inode->i_blocks;
3160 ll_inode_size_unlock(inode, 0);
3164 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
3166 struct lookup_intent it = { .it_op = IT_GETATTR };
3168 return ll_getattr_it(mnt, de, &it, stat);
3172 int lustre_check_acl(struct inode *inode, int mask)
3174 #ifdef CONFIG_FS_POSIX_ACL
3175 struct ll_inode_info *lli = ll_i2info(inode);
3176 struct posix_acl *acl;
3180 spin_lock(&lli->lli_lock);
3181 acl = posix_acl_dup(lli->lli_posix_acl);
3182 spin_unlock(&lli->lli_lock);
3187 rc = posix_acl_permission(inode, acl, mask);
3188 posix_acl_release(acl);
3196 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,10))
3197 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
3199 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), mask %o\n",
3200 inode->i_ino, inode->i_generation, inode, mask);
3201 if (ll_i2sbi(inode)->ll_flags & LL_SBI_RMT_CLIENT)
3202 return lustre_check_remote_perm(inode, mask);
3204 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_INODE_PERM, 1);
3205 return generic_permission(inode, mask, lustre_check_acl);
3208 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
3210 int mode = inode->i_mode;
3213 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), mask %o\n",
3214 inode->i_ino, inode->i_generation, inode, mask);
3216 if (ll_i2sbi(inode)->ll_flags & LL_SBI_RMT_CLIENT)
3217 return lustre_check_remote_perm(inode, mask);
3219 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_INODE_PERM, 1);
3221 if ((mask & MAY_WRITE) && IS_RDONLY(inode) &&
3222 (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
3224 if ((mask & MAY_WRITE) && IS_IMMUTABLE(inode))
3226 if (current->fsuid == inode->i_uid) {
3229 if (((mode >> 3) & mask & S_IRWXO) != mask)
3231 rc = lustre_check_acl(inode, mask);
3235 goto check_capabilities;
3239 if (in_group_p(inode->i_gid))
3242 if ((mode & mask & S_IRWXO) == mask)
3246 if (!(mask & MAY_EXEC) ||
3247 (inode->i_mode & S_IXUGO) || S_ISDIR(inode->i_mode))
3248 if (cfs_capable(CFS_CAP_DAC_OVERRIDE))
3251 if (cfs_capable(CFS_CAP_DAC_READ_SEARCH) && ((mask == MAY_READ) ||
3252 (S_ISDIR(inode->i_mode) && !(mask & MAY_WRITE))))
3259 /* -o localflock - only provides locally consistent flock locks */
3260 struct file_operations ll_file_operations = {
3261 .read = ll_file_read,
3262 .write = ll_file_write,
3263 .ioctl = ll_file_ioctl,
3264 .open = ll_file_open,
3265 .release = ll_file_release,
3266 .mmap = ll_file_mmap,
3267 .llseek = ll_file_seek,
3268 .sendfile = ll_file_sendfile,
3272 struct file_operations ll_file_operations_flock = {
3273 .read = ll_file_read,
3274 .write = ll_file_write,
3275 .ioctl = ll_file_ioctl,
3276 .open = ll_file_open,
3277 .release = ll_file_release,
3278 .mmap = ll_file_mmap,
3279 .llseek = ll_file_seek,
3280 .sendfile = ll_file_sendfile,
3282 #ifdef HAVE_F_OP_FLOCK
3283 .flock = ll_file_flock,
3285 .lock = ll_file_flock
3288 /* These are for -o noflock - to return ENOSYS on flock calls */
3289 struct file_operations ll_file_operations_noflock = {
3290 .read = ll_file_read,
3291 .write = ll_file_write,
3292 .ioctl = ll_file_ioctl,
3293 .open = ll_file_open,
3294 .release = ll_file_release,
3295 .mmap = ll_file_mmap,
3296 .llseek = ll_file_seek,
3297 .sendfile = ll_file_sendfile,
3299 #ifdef HAVE_F_OP_FLOCK
3300 .flock = ll_file_noflock,
3302 .lock = ll_file_noflock
3305 struct inode_operations ll_file_inode_operations = {
3306 #ifdef HAVE_VFS_INTENT_PATCHES
3307 .setattr_raw = ll_setattr_raw,
3309 .setattr = ll_setattr,
3310 .truncate = ll_truncate,
3311 .getattr = ll_getattr,
3312 .permission = ll_inode_permission,
3313 .setxattr = ll_setxattr,
3314 .getxattr = ll_getxattr,
3315 .listxattr = ll_listxattr,
3316 .removexattr = ll_removexattr,
3319 /* dynamic ioctl number support routins */
3320 static struct llioc_ctl_data {
3321 struct rw_semaphore ioc_sem;
3322 struct list_head ioc_head;
3324 __RWSEM_INITIALIZER(llioc.ioc_sem),
3325 CFS_LIST_HEAD_INIT(llioc.ioc_head)
3330 struct list_head iocd_list;
3331 unsigned int iocd_size;
3332 llioc_callback_t iocd_cb;
3333 unsigned int iocd_count;
3334 unsigned int iocd_cmd[0];
3337 void *ll_iocontrol_register(llioc_callback_t cb, int count, unsigned int *cmd)
3340 struct llioc_data *in_data = NULL;
3343 if (cb == NULL || cmd == NULL ||
3344 count > LLIOC_MAX_CMD || count < 0)
3347 size = sizeof(*in_data) + count * sizeof(unsigned int);
3348 OBD_ALLOC(in_data, size);
3349 if (in_data == NULL)
3352 memset(in_data, 0, sizeof(*in_data));
3353 in_data->iocd_size = size;
3354 in_data->iocd_cb = cb;
3355 in_data->iocd_count = count;
3356 memcpy(in_data->iocd_cmd, cmd, sizeof(unsigned int) * count);
3358 down_write(&llioc.ioc_sem);
3359 list_add_tail(&in_data->iocd_list, &llioc.ioc_head);
3360 up_write(&llioc.ioc_sem);
3365 void ll_iocontrol_unregister(void *magic)
3367 struct llioc_data *tmp;
3372 down_write(&llioc.ioc_sem);
3373 list_for_each_entry(tmp, &llioc.ioc_head, iocd_list) {
3375 unsigned int size = tmp->iocd_size;
3377 list_del(&tmp->iocd_list);
3378 up_write(&llioc.ioc_sem);
3380 OBD_FREE(tmp, size);
3384 up_write(&llioc.ioc_sem);
3386 CWARN("didn't find iocontrol register block with magic: %p\n", magic);
3389 EXPORT_SYMBOL(ll_iocontrol_register);
3390 EXPORT_SYMBOL(ll_iocontrol_unregister);
3392 enum llioc_iter ll_iocontrol_call(struct inode *inode, struct file *file,
3393 unsigned int cmd, unsigned long arg, int *rcp)
3395 enum llioc_iter ret = LLIOC_CONT;
3396 struct llioc_data *data;
3397 int rc = -EINVAL, i;
3399 down_read(&llioc.ioc_sem);
3400 list_for_each_entry(data, &llioc.ioc_head, iocd_list) {
3401 for (i = 0; i < data->iocd_count; i++) {
3402 if (cmd != data->iocd_cmd[i])
3405 ret = data->iocd_cb(inode, file, cmd, arg, data, &rc);
3409 if (ret == LLIOC_STOP)
3412 up_read(&llioc.ioc_sem);