1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Phil Schwan <phil@clusterfs.com>
7 * Author: Andreas Dilger <adilger@clusterfs.com>
9 * This file is part of Lustre, http://www.lustre.org.
11 * Lustre is free software; you can redistribute it and/or
12 * modify it under the terms of version 2 of the GNU General Public
13 * License as published by the Free Software Foundation.
15 * Lustre is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with Lustre; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #define DEBUG_SUBSYSTEM S_LLITE
26 #include <lustre_dlm.h>
27 #include <lustre_lite.h>
28 #include <lustre_mdc.h>
29 #include <linux/pagemap.h>
30 #include <linux/file.h>
31 #include "llite_internal.h"
33 /* also used by llite/special.c:ll_special_open() */
34 struct ll_file_data *ll_file_data_get(void)
36 struct ll_file_data *fd;
38 OBD_SLAB_ALLOC_PTR(fd, ll_file_data_slab);
42 static void ll_file_data_put(struct ll_file_data *fd)
45 OBD_SLAB_FREE_PTR(fd, ll_file_data_slab);
48 void ll_pack_inode2opdata(struct inode *inode, struct md_op_data *op_data,
49 struct lustre_handle *fh)
51 op_data->op_fid1 = ll_i2info(inode)->lli_fid;
52 op_data->op_attr.ia_mode = inode->i_mode;
53 op_data->op_attr.ia_atime = inode->i_atime;
54 op_data->op_attr.ia_mtime = inode->i_mtime;
55 op_data->op_attr.ia_ctime = inode->i_ctime;
56 op_data->op_attr.ia_size = i_size_read(inode);
57 op_data->op_attr_blocks = inode->i_blocks;
58 ((struct ll_iattr *)&op_data->op_attr)->ia_attr_flags = inode->i_flags;
59 op_data->op_ioepoch = ll_i2info(inode)->lli_ioepoch;
60 memcpy(&op_data->op_handle, fh, sizeof(op_data->op_handle));
61 op_data->op_capa1 = ll_mdscapa_get(inode);
64 static void ll_prepare_close(struct inode *inode, struct md_op_data *op_data,
65 struct obd_client_handle *och)
69 op_data->op_attr.ia_valid = ATTR_MODE | ATTR_ATIME_SET |
70 ATTR_MTIME_SET | ATTR_CTIME_SET;
72 if (!(och->och_flags & FMODE_WRITE))
75 if (!(ll_i2mdexp(inode)->exp_connect_flags & OBD_CONNECT_SOM) ||
76 !S_ISREG(inode->i_mode))
77 op_data->op_attr.ia_valid |= ATTR_SIZE | ATTR_BLOCKS;
79 ll_epoch_close(inode, op_data, &och, 0);
82 ll_pack_inode2opdata(inode, op_data, &och->och_fh);
86 static int ll_close_inode_openhandle(struct obd_export *md_exp,
88 struct obd_client_handle *och)
90 struct obd_export *exp = ll_i2mdexp(inode);
91 struct md_op_data *op_data;
92 struct ptlrpc_request *req = NULL;
93 struct obd_device *obd = class_exp2obd(exp);
100 * XXX: in case of LMV, is this correct to access
103 CERROR("Invalid MDC connection handle "LPX64"\n",
104 ll_i2mdexp(inode)->exp_handle.h_cookie);
109 * here we check if this is forced umount. If so this is called on
110 * canceling "open lock" and we do not call md_close() in this case, as
111 * it will not be successful, as import is already deactivated.
116 OBD_ALLOC_PTR(op_data);
118 GOTO(out, rc = -ENOMEM); // XXX We leak openhandle and request here.
120 ll_prepare_close(inode, op_data, och);
121 epoch_close = (op_data->op_flags & MF_EPOCH_CLOSE);
122 rc = md_close(md_exp, op_data, och->och_mod, &req);
127 /* This close must have the epoch closed. */
128 LASSERT(exp->exp_connect_flags & OBD_CONNECT_SOM);
129 LASSERT(epoch_close);
130 /* MDS has instructed us to obtain Size-on-MDS attribute from
131 * OSTs and send setattr to back to MDS. */
132 rc = ll_sizeonmds_update(inode, och->och_mod,
133 &och->och_fh, op_data->op_ioepoch);
135 CERROR("inode %lu mdc Size-on-MDS update failed: "
136 "rc = %d\n", inode->i_ino, rc);
140 CERROR("inode %lu mdc close failed: rc = %d\n",
143 ll_finish_md_op_data(op_data);
146 rc = ll_objects_destroy(req, inode);
148 CERROR("inode %lu ll_objects destroy: rc = %d\n",
155 if ((exp->exp_connect_flags & OBD_CONNECT_SOM) && !epoch_close &&
156 S_ISREG(inode->i_mode) && (och->och_flags & FMODE_WRITE)) {
157 ll_queue_done_writing(inode, LLIF_DONE_WRITING);
160 ptlrpc_close_replay_seq(req);
161 md_clear_open_replay_data(md_exp, och);
162 /* Free @och if it is not waiting for DONE_WRITING. */
163 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
166 if (req) /* This is close request */
167 ptlrpc_req_finished(req);
171 int ll_md_real_close(struct inode *inode, int flags)
173 struct ll_inode_info *lli = ll_i2info(inode);
174 struct obd_client_handle **och_p;
175 struct obd_client_handle *och;
180 if (flags & FMODE_WRITE) {
181 och_p = &lli->lli_mds_write_och;
182 och_usecount = &lli->lli_open_fd_write_count;
183 } else if (flags & FMODE_EXEC) {
184 och_p = &lli->lli_mds_exec_och;
185 och_usecount = &lli->lli_open_fd_exec_count;
187 LASSERT(flags & FMODE_READ);
188 och_p = &lli->lli_mds_read_och;
189 och_usecount = &lli->lli_open_fd_read_count;
192 down(&lli->lli_och_sem);
193 if (*och_usecount) { /* There are still users of this handle, so
195 up(&lli->lli_och_sem);
200 up(&lli->lli_och_sem);
202 if (och) { /* There might be a race and somebody have freed this och
204 rc = ll_close_inode_openhandle(ll_i2sbi(inode)->ll_md_exp,
211 int ll_md_close(struct obd_export *md_exp, struct inode *inode,
214 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
215 struct ll_inode_info *lli = ll_i2info(inode);
219 /* clear group lock, if present */
220 if (unlikely(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
221 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
222 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
223 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP,
227 /* Let's see if we have good enough OPEN lock on the file and if
228 we can skip talking to MDS */
229 if (file->f_dentry->d_inode) { /* Can this ever be false? */
231 int flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
232 struct lustre_handle lockh;
233 struct inode *inode = file->f_dentry->d_inode;
234 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_OPEN}};
236 down(&lli->lli_och_sem);
237 if (fd->fd_omode & FMODE_WRITE) {
239 LASSERT(lli->lli_open_fd_write_count);
240 lli->lli_open_fd_write_count--;
241 } else if (fd->fd_omode & FMODE_EXEC) {
243 LASSERT(lli->lli_open_fd_exec_count);
244 lli->lli_open_fd_exec_count--;
247 LASSERT(lli->lli_open_fd_read_count);
248 lli->lli_open_fd_read_count--;
250 up(&lli->lli_och_sem);
252 if (!md_lock_match(md_exp, flags, ll_inode2fid(inode),
253 LDLM_IBITS, &policy, lockmode,
255 rc = ll_md_real_close(file->f_dentry->d_inode,
259 CERROR("Releasing a file %p with negative dentry %p. Name %s",
260 file, file->f_dentry, file->f_dentry->d_name.name);
263 LUSTRE_FPRIVATE(file) = NULL;
264 ll_file_data_put(fd);
265 ll_capa_close(inode);
270 int lov_test_and_clear_async_rc(struct lov_stripe_md *lsm);
272 /* While this returns an error code, fput() the caller does not, so we need
273 * to make every effort to clean up all of our state here. Also, applications
274 * rarely check close errors and even if an error is returned they will not
275 * re-try the close call.
277 int ll_file_release(struct inode *inode, struct file *file)
279 struct ll_file_data *fd;
280 struct ll_sb_info *sbi = ll_i2sbi(inode);
281 struct ll_inode_info *lli = ll_i2info(inode);
282 struct lov_stripe_md *lsm = lli->lli_smd;
286 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
287 inode->i_generation, inode);
289 #ifdef CONFIG_FS_POSIX_ACL
290 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
291 inode == inode->i_sb->s_root->d_inode) {
292 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
295 if (unlikely(fd->fd_flags & LL_FILE_RMTACL)) {
296 fd->fd_flags &= ~LL_FILE_RMTACL;
297 rct_del(&sbi->ll_rct, cfs_curproc_pid());
298 et_search_free(&sbi->ll_et, cfs_curproc_pid());
303 if (inode->i_sb->s_root != file->f_dentry)
304 ll_stats_ops_tally(sbi, LPROC_LL_RELEASE, 1);
305 fd = LUSTRE_FPRIVATE(file);
308 /* The last ref on @file, maybe not the the owner pid of statahead.
309 * Different processes can open the same dir, "ll_opendir_key" means:
310 * it is me that should stop the statahead thread. */
311 if (lli->lli_opendir_key == fd)
312 ll_stop_statahead(inode, fd);
314 if (inode->i_sb->s_root == file->f_dentry) {
315 LUSTRE_FPRIVATE(file) = NULL;
316 ll_file_data_put(fd);
321 lov_test_and_clear_async_rc(lsm);
322 lli->lli_async_rc = 0;
324 rc = ll_md_close(sbi->ll_md_exp, inode, file);
328 static int ll_intent_file_open(struct file *file, void *lmm,
329 int lmmsize, struct lookup_intent *itp)
331 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
332 struct dentry *parent = file->f_dentry->d_parent;
333 const char *name = file->f_dentry->d_name.name;
334 const int len = file->f_dentry->d_name.len;
335 struct md_op_data *op_data;
336 struct ptlrpc_request *req;
343 /* Usually we come here only for NFSD, and we want open lock.
344 But we can also get here with pre 2.6.15 patchless kernels, and in
345 that case that lock is also ok */
346 /* We can also get here if there was cached open handle in revalidate_it
347 * but it disappeared while we were getting from there to ll_file_open.
348 * But this means this file was closed and immediatelly opened which
349 * makes a good candidate for using OPEN lock */
350 /* If lmmsize & lmm are not 0, we are just setting stripe info
351 * parameters. No need for the open lock */
352 if (!lmm && !lmmsize)
353 itp->it_flags |= MDS_OPEN_LOCK;
355 op_data = ll_prep_md_op_data(NULL, parent->d_inode,
356 file->f_dentry->d_inode, name, len,
357 O_RDWR, LUSTRE_OPC_ANY, NULL);
359 RETURN(PTR_ERR(op_data));
361 rc = md_intent_lock(sbi->ll_md_exp, op_data, lmm, lmmsize, itp,
362 0 /*unused */, &req, ll_md_blocking_ast, 0);
363 ll_finish_md_op_data(op_data);
365 /* reason for keep own exit path - don`t flood log
366 * with messages with -ESTALE errors.
368 if (!it_disposition(itp, DISP_OPEN_OPEN) ||
369 it_open_error(DISP_OPEN_OPEN, itp))
371 ll_release_openhandle(file->f_dentry, itp);
375 if (rc != 0 || it_open_error(DISP_OPEN_OPEN, itp)) {
376 rc = rc ? rc : it_open_error(DISP_OPEN_OPEN, itp);
377 CDEBUG(D_VFSTRACE, "lock enqueue: err: %d\n", rc);
381 if (itp->d.lustre.it_lock_mode)
382 md_set_lock_data(sbi->ll_md_exp,
383 &itp->d.lustre.it_lock_handle,
384 file->f_dentry->d_inode);
386 rc = ll_prep_inode(&file->f_dentry->d_inode, req, NULL);
388 ptlrpc_req_finished(itp->d.lustre.it_data);
391 it_clear_disposition(itp, DISP_ENQ_COMPLETE);
392 ll_intent_drop_lock(itp);
397 static int ll_och_fill(struct obd_export *md_exp, struct ll_inode_info *lli,
398 struct lookup_intent *it, struct obd_client_handle *och)
400 struct ptlrpc_request *req = it->d.lustre.it_data;
401 struct mdt_body *body;
405 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
406 LASSERT(body != NULL); /* reply already checked out */
408 memcpy(&och->och_fh, &body->handle, sizeof(body->handle));
409 och->och_magic = OBD_CLIENT_HANDLE_MAGIC;
410 och->och_fid = lli->lli_fid;
411 och->och_flags = it->it_flags;
412 lli->lli_ioepoch = body->ioepoch;
414 return md_set_open_replay_data(md_exp, och, req);
417 int ll_local_open(struct file *file, struct lookup_intent *it,
418 struct ll_file_data *fd, struct obd_client_handle *och)
420 struct inode *inode = file->f_dentry->d_inode;
421 struct ll_inode_info *lli = ll_i2info(inode);
424 LASSERT(!LUSTRE_FPRIVATE(file));
429 struct ptlrpc_request *req = it->d.lustre.it_data;
430 struct mdt_body *body;
433 rc = ll_och_fill(ll_i2sbi(inode)->ll_md_exp, lli, it, och);
437 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
438 if ((it->it_flags & FMODE_WRITE) &&
439 (body->valid & OBD_MD_FLSIZE))
440 CDEBUG(D_INODE, "Epoch "LPU64" opened on "DFID"\n",
441 lli->lli_ioepoch, PFID(&lli->lli_fid));
444 LUSTRE_FPRIVATE(file) = fd;
445 ll_readahead_init(inode, &fd->fd_ras);
446 fd->fd_omode = it->it_flags;
450 /* Open a file, and (for the very first open) create objects on the OSTs at
451 * this time. If opened with O_LOV_DELAY_CREATE, then we don't do the object
452 * creation or open until ll_lov_setstripe() ioctl is called. We grab
453 * lli_open_sem to ensure no other process will create objects, send the
454 * stripe MD to the MDS, or try to destroy the objects if that fails.
456 * If we already have the stripe MD locally then we don't request it in
457 * md_open(), by passing a lmm_size = 0.
459 * It is up to the application to ensure no other processes open this file
460 * in the O_LOV_DELAY_CREATE case, or the default striping pattern will be
461 * used. We might be able to avoid races of that sort by getting lli_open_sem
462 * before returning in the O_LOV_DELAY_CREATE case and dropping it here
463 * or in ll_file_release(), but I'm not sure that is desirable/necessary.
465 int ll_file_open(struct inode *inode, struct file *file)
467 struct ll_inode_info *lli = ll_i2info(inode);
468 struct lookup_intent *it, oit = { .it_op = IT_OPEN,
469 .it_flags = file->f_flags };
470 struct lov_stripe_md *lsm;
471 struct ptlrpc_request *req = NULL;
472 struct obd_client_handle **och_p;
474 struct ll_file_data *fd;
475 int rc = 0, opendir_set = 0;
478 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), flags %o\n", inode->i_ino,
479 inode->i_generation, inode, file->f_flags);
481 #ifdef HAVE_VFS_INTENT_PATCHES
484 it = file->private_data; /* XXX: compat macro */
485 file->private_data = NULL; /* prevent ll_local_open assertion */
488 fd = ll_file_data_get();
492 if (S_ISDIR(inode->i_mode)) {
493 spin_lock(&lli->lli_lock);
494 /* "lli->lli_opendir_pid != 0" means someone has set it.
495 * "lli->lli_sai != NULL" means the previous statahead has not
497 if (lli->lli_opendir_pid == 0 && lli->lli_sai == NULL) {
499 lli->lli_opendir_pid = cfs_curproc_pid();
500 lli->lli_opendir_key = fd;
501 } else if (unlikely(lli->lli_opendir_pid == cfs_curproc_pid())) {
502 /* Two cases for this:
503 * (1) The same process open such directory many times.
504 * (2) The old process opened the directory, and exited
505 * before its children processes. Then new process
506 * with the same pid opens such directory before the
507 * old process's children processes exit.
508 * Change the owner to the latest one. */
510 lli->lli_opendir_key = fd;
512 spin_unlock(&lli->lli_lock);
515 if (inode->i_sb->s_root == file->f_dentry) {
516 LUSTRE_FPRIVATE(file) = fd;
520 if (!it || !it->d.lustre.it_disposition) {
521 /* Convert f_flags into access mode. We cannot use file->f_mode,
522 * because everything but O_ACCMODE mask was stripped from
524 if ((oit.it_flags + 1) & O_ACCMODE)
526 if (file->f_flags & O_TRUNC)
527 oit.it_flags |= FMODE_WRITE;
529 /* kernel only call f_op->open in dentry_open. filp_open calls
530 * dentry_open after call to open_namei that checks permissions.
531 * Only nfsd_open call dentry_open directly without checking
532 * permissions and because of that this code below is safe. */
533 if (oit.it_flags & FMODE_WRITE)
534 oit.it_flags |= MDS_OPEN_OWNEROVERRIDE;
536 /* We do not want O_EXCL here, presumably we opened the file
537 * already? XXX - NFS implications? */
538 oit.it_flags &= ~O_EXCL;
544 /* Let's see if we have file open on MDS already. */
545 if (it->it_flags & FMODE_WRITE) {
546 och_p = &lli->lli_mds_write_och;
547 och_usecount = &lli->lli_open_fd_write_count;
548 } else if (it->it_flags & FMODE_EXEC) {
549 och_p = &lli->lli_mds_exec_och;
550 och_usecount = &lli->lli_open_fd_exec_count;
552 och_p = &lli->lli_mds_read_och;
553 och_usecount = &lli->lli_open_fd_read_count;
556 down(&lli->lli_och_sem);
557 if (*och_p) { /* Open handle is present */
558 if (it_disposition(it, DISP_OPEN_OPEN)) {
559 /* Well, there's extra open request that we do not need,
560 let's close it somehow. This will decref request. */
561 rc = it_open_error(DISP_OPEN_OPEN, it);
563 ll_file_data_put(fd);
564 GOTO(out_och_free, rc);
566 ll_release_openhandle(file->f_dentry, it);
567 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
572 rc = ll_local_open(file, it, fd, NULL);
574 up(&lli->lli_och_sem);
575 ll_file_data_put(fd);
579 LASSERT(*och_usecount == 0);
580 if (!it->d.lustre.it_disposition) {
581 /* We cannot just request lock handle now, new ELC code
582 means that one of other OPEN locks for this file
583 could be cancelled, and since blocking ast handler
584 would attempt to grab och_sem as well, that would
585 result in a deadlock */
586 up(&lli->lli_och_sem);
587 it->it_flags |= O_CHECK_STALE;
588 rc = ll_intent_file_open(file, NULL, 0, it);
589 it->it_flags &= ~O_CHECK_STALE;
591 ll_file_data_put(fd);
592 GOTO(out_openerr, rc);
595 /* Got some error? Release the request */
596 if (it->d.lustre.it_status < 0) {
597 req = it->d.lustre.it_data;
598 ptlrpc_req_finished(req);
600 md_set_lock_data(ll_i2sbi(inode)->ll_md_exp,
601 &it->d.lustre.it_lock_handle,
602 file->f_dentry->d_inode);
605 OBD_ALLOC(*och_p, sizeof (struct obd_client_handle));
607 ll_file_data_put(fd);
608 GOTO(out_och_free, rc = -ENOMEM);
611 req = it->d.lustre.it_data;
613 /* md_intent_lock() didn't get a request ref if there was an
614 * open error, so don't do cleanup on the request here
616 /* XXX (green): Should not we bail out on any error here, not
617 * just open error? */
618 rc = it_open_error(DISP_OPEN_OPEN, it);
620 ll_file_data_put(fd);
621 GOTO(out_och_free, rc);
624 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_OPEN, 1);
625 rc = ll_local_open(file, it, fd, *och_p);
627 up(&lli->lli_och_sem);
628 ll_file_data_put(fd);
629 GOTO(out_och_free, rc);
632 up(&lli->lli_och_sem);
634 /* Must do this outside lli_och_sem lock to prevent deadlock where
635 different kind of OPEN lock for this same inode gets cancelled
636 by ldlm_cancel_lru */
637 if (!S_ISREG(inode->i_mode))
644 if (file->f_flags & O_LOV_DELAY_CREATE ||
645 !(file->f_mode & FMODE_WRITE)) {
646 CDEBUG(D_INODE, "object creation was delayed\n");
650 file->f_flags &= ~O_LOV_DELAY_CREATE;
653 ptlrpc_req_finished(req);
655 it_clear_disposition(it, DISP_ENQ_OPEN_REF);
659 OBD_FREE(*och_p, sizeof (struct obd_client_handle));
660 *och_p = NULL; /* OBD_FREE writes some magic there */
663 up(&lli->lli_och_sem);
665 if (opendir_set == 1) {
666 lli->lli_opendir_key = NULL;
667 lli->lli_opendir_pid = 0;
668 } else if (unlikely(opendir_set == 2)) {
669 ll_stop_statahead(inode, fd);
676 /* Fills the obdo with the attributes for the inode defined by lsm */
677 int ll_inode_getattr(struct inode *inode, struct obdo *obdo)
679 struct ptlrpc_request_set *set;
680 struct ll_inode_info *lli = ll_i2info(inode);
681 struct lov_stripe_md *lsm = lli->lli_smd;
683 struct obd_info oinfo = { { { 0 } } };
687 LASSERT(lsm != NULL);
691 oinfo.oi_oa->o_id = lsm->lsm_object_id;
692 oinfo.oi_oa->o_gr = lsm->lsm_object_gr;
693 oinfo.oi_oa->o_mode = S_IFREG;
694 oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE |
695 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
696 OBD_MD_FLBLKSZ | OBD_MD_FLATIME |
697 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
699 oinfo.oi_capa = ll_mdscapa_get(inode);
701 set = ptlrpc_prep_set();
703 CERROR("can't allocate ptlrpc set\n");
706 rc = obd_getattr_async(ll_i2dtexp(inode), &oinfo, set);
708 rc = ptlrpc_set_wait(set);
709 ptlrpc_set_destroy(set);
711 capa_put(oinfo.oi_capa);
715 oinfo.oi_oa->o_valid &= (OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ |
716 OBD_MD_FLATIME | OBD_MD_FLMTIME |
717 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
719 obdo_refresh_inode(inode, oinfo.oi_oa, oinfo.oi_oa->o_valid);
720 CDEBUG(D_INODE, "objid "LPX64" size %Lu, blocks %llu, blksize %lu\n",
721 lli->lli_smd->lsm_object_id, i_size_read(inode),
722 (unsigned long long)inode->i_blocks,
723 (unsigned long)ll_inode_blksize(inode));
727 static inline void ll_remove_suid(struct inode *inode)
731 /* set S_IGID if S_IXGRP is set, and always set S_ISUID */
732 mode = (inode->i_mode & S_IXGRP)*(S_ISGID/S_IXGRP) | S_ISUID;
734 /* was any of the uid bits set? */
735 mode &= inode->i_mode;
736 if (mode && !capable(CAP_FSETID)) {
737 inode->i_mode &= ~mode;
738 // XXX careful here - we cannot change the size
742 static int ll_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock)
744 struct ll_inode_info *lli = ll_i2info(inode);
745 struct lov_stripe_md *lsm = lli->lli_smd;
746 struct obd_export *exp = ll_i2dtexp(inode);
749 struct ldlm_lock *lock;
750 struct lov_stripe_md *lsm;
751 } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm };
752 __u32 stripe, vallen = sizeof(stripe);
756 if (lsm->lsm_stripe_count == 1)
757 GOTO(check, stripe = 0);
759 /* get our offset in the lov */
760 rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe);
762 CERROR("obd_get_info: rc = %d\n", rc);
765 LASSERT(stripe < lsm->lsm_stripe_count);
768 if (lsm->lsm_oinfo[stripe]->loi_id != lock->l_resource->lr_name.name[0]||
769 lsm->lsm_oinfo[stripe]->loi_gr != lock->l_resource->lr_name.name[2]){
770 LDLM_ERROR(lock, "resource doesn't match object "LPU64"/"LPU64,
771 lsm->lsm_oinfo[stripe]->loi_id,
772 lsm->lsm_oinfo[stripe]->loi_gr);
773 RETURN(-ELDLM_NO_LOCK_DATA);
779 /* Get extra page reference to ensure it is not going away */
780 void ll_pin_extent_cb(void *data)
782 struct page *page = data;
784 page_cache_get(page);
789 /* Flush the page from page cache for an extent as its canceled.
790 * Page to remove is delivered as @data.
792 * No one can dirty the extent until we've finished our work and they cannot
793 * enqueue another lock. The DLM protects us from ll_file_read/write here,
794 * but other kernel actors could have pages locked.
796 * If @discard is set, there is no need to write the page if it is dirty.
798 * Called with the DLM lock held. */
799 int ll_page_removal_cb(void *data, int discard)
802 struct page *page = data;
803 struct address_space *mapping;
807 /* We have page reference already from ll_pin_page */
810 /* Already truncated by somebody */
813 mapping = page->mapping;
815 ll_teardown_mmaps(mapping,
816 (__u64)page->index << PAGE_CACHE_SHIFT,
817 ((__u64)page->index<<PAGE_CACHE_SHIFT)|
819 LL_CDEBUG_PAGE(D_PAGE, page, "removing page\n");
821 if (!discard && clear_page_dirty_for_io(page)) {
822 LASSERT(page->mapping);
823 rc = ll_call_writepage(page->mapping->host, page);
824 /* either waiting for io to complete or reacquiring
825 * the lock that the failed writepage released */
827 wait_on_page_writeback(page);
829 CERROR("writepage inode %lu(%p) of page %p "
830 "failed: %d\n", mapping->host->i_ino,
831 mapping->host, page, rc);
833 set_bit(AS_ENOSPC, &mapping->flags);
835 set_bit(AS_EIO, &mapping->flags);
837 set_bit(AS_EIO, &mapping->flags);
839 if (page->mapping != NULL) {
840 struct ll_async_page *llap = llap_cast_private(page);
841 /* checking again to account for writeback's lock_page() */
842 LL_CDEBUG_PAGE(D_PAGE, page, "truncating\n");
844 ll_ra_accounting(llap, page->mapping);
845 ll_truncate_complete_page(page);
849 LASSERT(!PageWriteback(page));
851 page_cache_release(page);
856 int ll_extent_lock_cancel_cb(struct ldlm_lock *lock, struct ldlm_lock_desc *new,
857 void *data, int flag)
860 struct ll_inode_info *lli;
861 struct lov_stripe_md *lsm;
867 if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) {
868 LDLM_ERROR(lock, "cancelling lock with bad data %p", data);
872 inode = ll_inode_from_lock(lock);
875 lli = ll_i2info(inode);
878 if (lli->lli_smd == NULL)
882 stripe = ll_lock_to_stripe_offset(inode, lock);
886 lov_stripe_lock(lsm);
887 lock_res_and_lock(lock);
888 kms = ldlm_extent_shift_kms(lock,
889 lsm->lsm_oinfo[stripe]->loi_kms);
891 if (lsm->lsm_oinfo[stripe]->loi_kms != kms)
892 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
893 lsm->lsm_oinfo[stripe]->loi_kms, kms);
894 lsm->lsm_oinfo[stripe]->loi_kms = kms;
895 unlock_res_and_lock(lock);
896 lov_stripe_unlock(lsm);
897 ll_queue_done_writing(inode, 0);
906 int ll_async_completion_ast(struct ldlm_lock *lock, int flags, void *data)
908 /* XXX ALLOCATE - 160 bytes */
909 struct inode *inode = ll_inode_from_lock(lock);
910 struct ll_inode_info *lli = ll_i2info(inode);
911 struct lustre_handle lockh = { 0 };
916 if (flags & (LDLM_FL_BLOCK_WAIT | LDLM_FL_BLOCK_GRANTED |
917 LDLM_FL_BLOCK_CONV)) {
918 LBUG(); /* not expecting any blocked async locks yet */
919 LDLM_DEBUG(lock, "client-side async enqueue returned a blocked "
921 ldlm_lock_dump(D_OTHER, lock, 0);
922 ldlm_reprocess_all(lock->l_resource);
926 LDLM_DEBUG(lock, "client-side async enqueue: granted/glimpsed");
928 stripe = ll_lock_to_stripe_offset(inode, lock);
932 if (lock->l_lvb_len) {
933 struct lov_stripe_md *lsm = lli->lli_smd;
935 lvb = lock->l_lvb_data;
936 lsm->lsm_oinfo[stripe].loi_rss = lvb->lvb_size;
938 lock_res_and_lock(lock);
939 ll_inode_size_lock(inode, 1);
940 kms = MAX(lsm->lsm_oinfo[stripe].loi_kms, lvb->lvb_size);
941 kms = ldlm_extent_shift_kms(NULL, kms);
942 if (lsm->lsm_oinfo[stripe].loi_kms != kms)
943 LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64,
944 lsm->lsm_oinfo[stripe].loi_kms, kms);
945 lsm->lsm_oinfo[stripe].loi_kms = kms;
946 ll_inode_size_unlock(inode, 1);
947 unlock_res_and_lock(lock);
952 wake_up(&lock->l_waitq);
954 ldlm_lock2handle(lock, &lockh);
955 ldlm_lock_decref(&lockh, LCK_PR);
960 static int ll_glimpse_callback(struct ldlm_lock *lock, void *reqp)
962 struct ptlrpc_request *req = reqp;
963 struct inode *inode = ll_inode_from_lock(lock);
964 struct ll_inode_info *lli;
965 struct lov_stripe_md *lsm;
971 GOTO(out, rc = -ELDLM_NO_LOCK_DATA);
972 lli = ll_i2info(inode);
974 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
977 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
979 /* First, find out which stripe index this lock corresponds to. */
980 stripe = ll_lock_to_stripe_offset(inode, lock);
982 GOTO(iput, rc = -ELDLM_NO_LOCK_DATA);
984 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
985 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
987 rc = req_capsule_server_pack(&req->rq_pill);
989 CERROR("lustre_pack_reply: %d\n", rc);
993 lvb = req_capsule_server_get(&req->rq_pill, &RMF_DLM_LVB);
994 lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe]->loi_kms;
995 lvb->lvb_mtime = LTIME_S(inode->i_mtime);
996 lvb->lvb_atime = LTIME_S(inode->i_atime);
997 lvb->lvb_ctime = LTIME_S(inode->i_ctime);
999 LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64
1000 " atime "LPU64", mtime "LPU64", ctime "LPU64,
1001 i_size_read(inode), stripe, lvb->lvb_size, lvb->lvb_mtime,
1002 lvb->lvb_atime, lvb->lvb_ctime);
1007 /* These errors are normal races, so we don't want to fill the console
1008 * with messages by calling ptlrpc_error() */
1009 if (rc == -ELDLM_NO_LOCK_DATA)
1010 lustre_pack_reply(req, 1, NULL, NULL);
1012 req->rq_status = rc;
1016 static int ll_merge_lvb(struct inode *inode)
1018 struct ll_inode_info *lli = ll_i2info(inode);
1019 struct ll_sb_info *sbi = ll_i2sbi(inode);
1025 ll_inode_size_lock(inode, 1);
1026 inode_init_lvb(inode, &lvb);
1027 rc = obd_merge_lvb(sbi->ll_dt_exp, lli->lli_smd, &lvb, 0);
1028 i_size_write(inode, lvb.lvb_size);
1029 inode->i_blocks = lvb.lvb_blocks;
1031 LTIME_S(inode->i_mtime) = lvb.lvb_mtime;
1032 LTIME_S(inode->i_atime) = lvb.lvb_atime;
1033 LTIME_S(inode->i_ctime) = lvb.lvb_ctime;
1034 ll_inode_size_unlock(inode, 1);
1039 int ll_local_size(struct inode *inode)
1041 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1042 struct ll_inode_info *lli = ll_i2info(inode);
1043 struct ll_sb_info *sbi = ll_i2sbi(inode);
1044 struct lustre_handle lockh = { 0 };
1049 if (lli->lli_smd->lsm_stripe_count == 0)
1052 rc = obd_match(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT,
1053 &policy, LCK_PR, &flags, inode, &lockh);
1059 rc = ll_merge_lvb(inode);
1060 obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh);
1064 int ll_glimpse_ioctl(struct ll_sb_info *sbi, struct lov_stripe_md *lsm,
1067 struct lustre_handle lockh = { 0 };
1068 struct ldlm_enqueue_info einfo = { 0 };
1069 struct obd_info oinfo = { { { 0 } } };
1075 einfo.ei_type = LDLM_EXTENT;
1076 einfo.ei_mode = LCK_PR;
1077 einfo.ei_cb_bl = osc_extent_blocking_cb;
1078 einfo.ei_cb_cp = ldlm_completion_ast;
1079 einfo.ei_cb_gl = ll_glimpse_callback;
1080 einfo.ei_cbdata = NULL;
1082 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
1083 oinfo.oi_lockh = &lockh;
1085 oinfo.oi_flags = LDLM_FL_HAS_INTENT;
1087 rc = obd_enqueue_rqset(sbi->ll_dt_exp, &oinfo, &einfo);
1091 CERROR("obd_enqueue returned rc %d, "
1092 "returning -EIO\n", rc);
1093 RETURN(rc > 0 ? -EIO : rc);
1096 lov_stripe_lock(lsm);
1097 memset(&lvb, 0, sizeof(lvb));
1098 obd_merge_lvb(sbi->ll_dt_exp, lsm, &lvb, 0);
1099 st->st_size = lvb.lvb_size;
1100 st->st_blocks = lvb.lvb_blocks;
1101 st->st_mtime = lvb.lvb_mtime;
1102 st->st_atime = lvb.lvb_atime;
1103 st->st_ctime = lvb.lvb_ctime;
1104 lov_stripe_unlock(lsm);
1109 /* NB: obd_merge_lvb will prefer locally cached writes if they extend the
1110 * file (because it prefers KMS over RSS when larger) */
1111 int ll_glimpse_size(struct inode *inode, int ast_flags)
1113 struct ll_inode_info *lli = ll_i2info(inode);
1114 struct ll_sb_info *sbi = ll_i2sbi(inode);
1115 struct lustre_handle lockh = { 0 };
1116 struct ldlm_enqueue_info einfo = { 0 };
1117 struct obd_info oinfo = { { { 0 } } };
1121 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
1124 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1126 if (!lli->lli_smd) {
1127 CDEBUG(D_DLMTRACE, "No objects for inode %lu\n", inode->i_ino);
1131 /* NOTE: this looks like DLM lock request, but it may not be one. Due
1132 * to LDLM_FL_HAS_INTENT flag, this is glimpse request, that
1133 * won't revoke any conflicting DLM locks held. Instead,
1134 * ll_glimpse_callback() will be called on each client
1135 * holding a DLM lock against this file, and resulting size
1136 * will be returned for each stripe. DLM lock on [0, EOF] is
1137 * acquired only if there were no conflicting locks. */
1138 einfo.ei_type = LDLM_EXTENT;
1139 einfo.ei_mode = LCK_PR;
1140 einfo.ei_cb_bl = osc_extent_blocking_cb;
1141 einfo.ei_cb_cp = ldlm_completion_ast;
1142 einfo.ei_cb_gl = ll_glimpse_callback;
1143 einfo.ei_cbdata = inode;
1145 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
1146 oinfo.oi_lockh = &lockh;
1147 oinfo.oi_md = lli->lli_smd;
1148 oinfo.oi_flags = ast_flags | LDLM_FL_HAS_INTENT;
1150 rc = obd_enqueue_rqset(sbi->ll_dt_exp, &oinfo, &einfo);
1154 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1155 RETURN(rc > 0 ? -EIO : rc);
1158 rc = ll_merge_lvb(inode);
1160 CDEBUG(D_DLMTRACE, "glimpse: size: %llu, blocks: %llu\n",
1161 i_size_read(inode), (unsigned long long)inode->i_blocks);
1166 int ll_extent_lock(struct ll_file_data *fd, struct inode *inode,
1167 struct lov_stripe_md *lsm, int mode,
1168 ldlm_policy_data_t *policy, struct lustre_handle *lockh,
1171 struct ll_sb_info *sbi = ll_i2sbi(inode);
1173 struct ldlm_enqueue_info einfo = { 0 };
1174 struct obd_info oinfo = { { { 0 } } };
1178 LASSERT(!lustre_handle_is_used(lockh));
1179 LASSERT(lsm != NULL);
1181 /* don't drop the mmapped file to LRU */
1182 if (mapping_mapped(inode->i_mapping))
1183 ast_flags |= LDLM_FL_NO_LRU;
1185 /* XXX phil: can we do this? won't it screw the file size up? */
1186 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1187 (sbi->ll_flags & LL_SBI_NOLCK))
1190 CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n",
1191 inode->i_ino, policy->l_extent.start, policy->l_extent.end);
1193 einfo.ei_type = LDLM_EXTENT;
1194 einfo.ei_mode = mode;
1195 einfo.ei_cb_bl = osc_extent_blocking_cb;
1196 einfo.ei_cb_cp = ldlm_completion_ast;
1197 einfo.ei_cb_gl = ll_glimpse_callback;
1198 einfo.ei_cbdata = inode;
1200 oinfo.oi_policy = *policy;
1201 oinfo.oi_lockh = lockh;
1203 oinfo.oi_flags = ast_flags;
1205 rc = obd_enqueue(sbi->ll_dt_exp, &oinfo, &einfo, NULL);
1206 *policy = oinfo.oi_policy;
1210 ll_inode_size_lock(inode, 1);
1211 inode_init_lvb(inode, &lvb);
1212 obd_merge_lvb(sbi->ll_dt_exp, lsm, &lvb, 1);
1214 if (policy->l_extent.start == 0 &&
1215 policy->l_extent.end == OBD_OBJECT_EOF) {
1216 /* vmtruncate()->ll_truncate() first sets the i_size and then
1217 * the kms under both a DLM lock and the
1218 * ll_inode_size_lock(). If we don't get the
1219 * ll_inode_size_lock() here we can match the DLM lock and
1220 * reset i_size from the kms before the truncating path has
1221 * updated the kms. generic_file_write can then trust the
1222 * stale i_size when doing appending writes and effectively
1223 * cancel the result of the truncate. Getting the
1224 * ll_inode_size_lock() after the enqueue maintains the DLM
1225 * -> ll_inode_size_lock() acquiring order. */
1226 i_size_write(inode, lvb.lvb_size);
1227 CDEBUG(D_INODE, "inode=%lu, updating i_size %llu\n",
1228 inode->i_ino, i_size_read(inode));
1232 LTIME_S(inode->i_mtime) = lvb.lvb_mtime;
1233 LTIME_S(inode->i_atime) = lvb.lvb_atime;
1234 LTIME_S(inode->i_ctime) = lvb.lvb_ctime;
1236 ll_inode_size_unlock(inode, 1);
1241 int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode,
1242 struct lov_stripe_md *lsm, int mode,
1243 struct lustre_handle *lockh)
1245 struct ll_sb_info *sbi = ll_i2sbi(inode);
1249 /* XXX phil: can we do this? won't it screw the file size up? */
1250 if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) ||
1251 (sbi->ll_flags & LL_SBI_NOLCK))
1254 rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh);
1259 static void ll_set_file_contended(struct inode *inode)
1261 struct ll_inode_info *lli = ll_i2info(inode);
1262 cfs_time_t now = cfs_time_current();
1264 spin_lock(&lli->lli_lock);
1265 lli->lli_contention_time = now;
1266 lli->lli_flags |= LLIF_CONTENDED;
1267 spin_unlock(&lli->lli_lock);
1270 void ll_clear_file_contended(struct inode *inode)
1272 struct ll_inode_info *lli = ll_i2info(inode);
1274 spin_lock(&lli->lli_lock);
1275 lli->lli_flags &= ~LLIF_CONTENDED;
1276 spin_unlock(&lli->lli_lock);
1279 static int ll_is_file_contended(struct file *file)
1281 struct inode *inode = file->f_dentry->d_inode;
1282 struct ll_inode_info *lli = ll_i2info(inode);
1283 struct ll_sb_info *sbi = ll_i2sbi(inode);
1284 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1287 if (!(sbi->ll_lco.lco_flags & OBD_CONNECT_SRVLOCK)) {
1288 CDEBUG(D_INFO, "the server does not support SRVLOCK feature,"
1289 " osc connect flags = 0x"LPX64"\n",
1290 sbi->ll_lco.lco_flags);
1293 if (fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK))
1295 if (lli->lli_flags & LLIF_CONTENDED) {
1296 cfs_time_t cur_time = cfs_time_current();
1297 cfs_time_t retry_time;
1299 retry_time = cfs_time_add(
1300 lli->lli_contention_time,
1301 cfs_time_seconds(sbi->ll_contention_time));
1302 if (cfs_time_after(cur_time, retry_time)) {
1303 ll_clear_file_contended(inode);
1311 static int ll_file_get_tree_lock(struct ll_lock_tree *tree, struct file *file,
1312 const char *buf, size_t count,
1313 loff_t start, loff_t end, int rw)
1316 int tree_locked = 0;
1318 struct inode * inode = file->f_dentry->d_inode;
1321 append = (rw == WRITE) && (file->f_flags & O_APPEND);
1323 if (append || !ll_is_file_contended(file)) {
1324 struct ll_lock_tree_node *node;
1327 ast_flags = append ? 0 : LDLM_FL_DENY_ON_CONTENTION;
1328 if (file->f_flags & O_NONBLOCK)
1329 ast_flags |= LDLM_FL_BLOCK_NOWAIT;
1330 node = ll_node_from_inode(inode, start, end,
1331 (rw == WRITE) ? LCK_PW : LCK_PR);
1336 tree->lt_fd = LUSTRE_FPRIVATE(file);
1337 rc = ll_tree_lock(tree, node, buf, count, ast_flags);
1340 else if (rc == -EUSERS)
1341 ll_set_file_contended(inode);
1345 RETURN(tree_locked);
1350 static ssize_t ll_file_read(struct file *file, char *buf, size_t count,
1353 struct inode *inode = file->f_dentry->d_inode;
1354 struct ll_inode_info *lli = ll_i2info(inode);
1355 struct lov_stripe_md *lsm = lli->lli_smd;
1356 struct ll_sb_info *sbi = ll_i2sbi(inode);
1357 struct ll_lock_tree tree;
1359 struct ll_ra_read bead;
1362 ssize_t retval, chunk, sum = 0;
1367 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1368 inode->i_ino, inode->i_generation, inode, count, *ppos);
1369 /* "If nbyte is 0, read() will return 0 and have no other results."
1370 * -- Single Unix Spec */
1374 ll_stats_ops_tally(sbi, LPROC_LL_READ_BYTES, count);
1377 /* Read on file with no objects should return zero-filled
1378 * buffers up to file size (we can get non-zero sizes with
1379 * mknod + truncate, then opening file for read. This is a
1380 * common pattern in NFS case, it seems). Bug 6243 */
1382 /* Since there are no objects on OSTs, we have nothing to get
1383 * lock on and so we are forced to access inode->i_size
1386 /* Read beyond end of file */
1387 if (*ppos >= i_size_read(inode))
1390 if (count > i_size_read(inode) - *ppos)
1391 count = i_size_read(inode) - *ppos;
1392 /* Make sure to correctly adjust the file pos pointer for
1394 notzeroed = clear_user(buf, count);
1402 if (sbi->ll_max_rw_chunk != 0) {
1403 /* first, let's know the end of the current stripe */
1405 obd_extent_calc(sbi->ll_dt_exp, lsm, OBD_CALC_STRIPE_END,
1408 /* correct, the end is beyond the request */
1409 if (end > *ppos + count - 1)
1410 end = *ppos + count - 1;
1412 /* and chunk shouldn't be too large even if striping is wide */
1413 if (end - *ppos > sbi->ll_max_rw_chunk)
1414 end = *ppos + sbi->ll_max_rw_chunk - 1;
1416 end = *ppos + count - 1;
1419 tree_locked = ll_file_get_tree_lock(&tree, file, buf,
1420 count, *ppos, end, READ);
1421 if (tree_locked < 0)
1422 GOTO(out, retval = tree_locked);
1424 ll_inode_size_lock(inode, 1);
1426 * Consistency guarantees: following possibilities exist for the
1427 * relation between region being read and real file size at this
1430 * (A): the region is completely inside of the file;
1432 * (B-x): x bytes of region are inside of the file, the rest is
1435 * (C): the region is completely outside of the file.
1437 * This classification is stable under DLM lock acquired by
1438 * ll_tree_lock() above, because to change class, other client has to
1439 * take DLM lock conflicting with our lock. Also, any updates to
1440 * ->i_size by other threads on this client are serialized by
1441 * ll_inode_size_lock(). This guarantees that short reads are handled
1442 * correctly in the face of concurrent writes and truncates.
1444 inode_init_lvb(inode, &lvb);
1445 obd_merge_lvb(sbi->ll_dt_exp, lsm, &lvb, 1);
1447 if (*ppos + count - 1 > kms) {
1448 /* A glimpse is necessary to determine whether we return a
1449 * short read (B) or some zeroes at the end of the buffer (C) */
1450 ll_inode_size_unlock(inode, 1);
1451 retval = ll_glimpse_size(inode, LDLM_FL_BLOCK_GRANTED);
1454 ll_tree_unlock(&tree);
1458 /* region is within kms and, hence, within real file size (A).
1459 * We need to increase i_size to cover the read region so that
1460 * generic_file_read() will do its job, but that doesn't mean
1461 * the kms size is _correct_, it is only the _minimum_ size.
1462 * If someone does a stat they will get the correct size which
1463 * will always be >= the kms value here. b=11081 */
1464 if (i_size_read(inode) < kms)
1465 i_size_write(inode, kms);
1466 ll_inode_size_unlock(inode, 1);
1469 chunk = end - *ppos + 1;
1470 CDEBUG(D_INODE, "Read ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1471 inode->i_ino, chunk, *ppos, i_size_read(inode));
1474 /* turn off the kernel's read-ahead */
1475 file->f_ra.ra_pages = 0;
1477 /* initialize read-ahead window once per syscall */
1480 bead.lrr_start = *ppos >> CFS_PAGE_SHIFT;
1481 bead.lrr_count = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1482 ll_ra_read_in(file, &bead);
1486 file_accessed(file);
1487 retval = generic_file_read(file, buf, chunk, ppos);
1488 ll_tree_unlock(&tree);
1490 retval = ll_file_lockless_io(file, buf, chunk, ppos, READ);
1493 ll_rw_stats_tally(sbi, current->pid, file, chunk, 0);
1499 if (retval == chunk && count > 0)
1505 ll_ra_read_ex(file, &bead);
1506 retval = (sum > 0) ? sum : retval;
1511 * Write to a file (through the page cache).
1513 static ssize_t ll_file_write(struct file *file, const char *buf, size_t count,
1516 struct inode *inode = file->f_dentry->d_inode;
1517 struct ll_sb_info *sbi = ll_i2sbi(inode);
1518 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1519 struct ll_lock_tree tree;
1520 loff_t maxbytes = ll_file_maxbytes(inode);
1521 loff_t lock_start, lock_end, end;
1522 ssize_t retval, chunk, sum = 0;
1526 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1527 inode->i_ino, inode->i_generation, inode, count, *ppos);
1529 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1531 /* POSIX, but surprised the VFS doesn't check this already */
1535 /* If file was opened for LL_IOC_LOV_SETSTRIPE but the ioctl wasn't
1536 * called on the file, don't fail the below assertion (bug 2388). */
1537 if (file->f_flags & O_LOV_DELAY_CREATE &&
1538 ll_i2info(inode)->lli_smd == NULL)
1541 LASSERT(ll_i2info(inode)->lli_smd != NULL);
1543 down(&ll_i2info(inode)->lli_write_sem);
1546 chunk = 0; /* just to fix gcc's warning */
1547 end = *ppos + count - 1;
1549 if (file->f_flags & O_APPEND) {
1551 lock_end = OBD_OBJECT_EOF;
1552 } else if (sbi->ll_max_rw_chunk != 0) {
1553 /* first, let's know the end of the current stripe */
1555 obd_extent_calc(sbi->ll_dt_exp, lsm, OBD_CALC_STRIPE_END,
1558 /* correct, the end is beyond the request */
1559 if (end > *ppos + count - 1)
1560 end = *ppos + count - 1;
1562 /* and chunk shouldn't be too large even if striping is wide */
1563 if (end - *ppos > sbi->ll_max_rw_chunk)
1564 end = *ppos + sbi->ll_max_rw_chunk - 1;
1569 lock_end = *ppos + count - 1;
1572 tree_locked = ll_file_get_tree_lock(&tree, file, buf, count,
1573 lock_start, lock_end, WRITE);
1574 if (tree_locked < 0)
1575 GOTO(out, retval = tree_locked);
1577 /* This is ok, g_f_w will overwrite this under i_sem if it races
1578 * with a local truncate, it just makes our maxbyte checking easier.
1579 * The i_size value gets updated in ll_extent_lock() as a consequence
1580 * of the [0,EOF] extent lock we requested above. */
1581 if (file->f_flags & O_APPEND) {
1582 *ppos = i_size_read(inode);
1583 end = *ppos + count - 1;
1586 if (*ppos >= maxbytes) {
1587 send_sig(SIGXFSZ, current, 0);
1588 GOTO(out_unlock, retval = -EFBIG);
1590 if (end > maxbytes - 1)
1593 /* generic_file_write handles O_APPEND after getting i_mutex */
1594 chunk = end - *ppos + 1;
1595 CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n",
1596 inode->i_ino, chunk, *ppos);
1598 retval = generic_file_write(file, buf, chunk, ppos);
1600 retval = ll_file_lockless_io(file, (char*)buf, chunk,
1602 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, file, chunk, 1);
1606 ll_tree_unlock(&tree);
1613 if (retval == chunk && count > 0)
1617 up(&ll_i2info(inode)->lli_write_sem);
1619 retval = (sum > 0) ? sum : retval;
1620 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_WRITE_BYTES,
1621 retval > 0 ? retval : 0);
1626 * Send file content (through pagecache) somewhere with helper
1628 static ssize_t ll_file_sendfile(struct file *in_file, loff_t *ppos,size_t count,
1629 read_actor_t actor, void *target)
1631 struct inode *inode = in_file->f_dentry->d_inode;
1632 struct ll_inode_info *lli = ll_i2info(inode);
1633 struct lov_stripe_md *lsm = lli->lli_smd;
1634 struct ll_lock_tree tree;
1635 struct ll_lock_tree_node *node;
1637 struct ll_ra_read bead;
1642 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),size="LPSZ",offset=%Ld\n",
1643 inode->i_ino, inode->i_generation, inode, count, *ppos);
1645 /* "If nbyte is 0, read() will return 0 and have no other results."
1646 * -- Single Unix Spec */
1650 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_READ_BYTES, count);
1651 /* turn off the kernel's read-ahead */
1652 in_file->f_ra.ra_pages = 0;
1654 /* File with no objects, nothing to lock */
1656 RETURN(generic_file_sendfile(in_file, ppos, count, actor, target));
1658 node = ll_node_from_inode(inode, *ppos, *ppos + count - 1, LCK_PR);
1660 RETURN(PTR_ERR(node));
1662 tree.lt_fd = LUSTRE_FPRIVATE(in_file);
1663 rc = ll_tree_lock(&tree, node, NULL, count,
1664 in_file->f_flags & O_NONBLOCK?LDLM_FL_BLOCK_NOWAIT:0);
1668 ll_clear_file_contended(inode);
1669 ll_inode_size_lock(inode, 1);
1671 * Consistency guarantees: following possibilities exist for the
1672 * relation between region being read and real file size at this
1675 * (A): the region is completely inside of the file;
1677 * (B-x): x bytes of region are inside of the file, the rest is
1680 * (C): the region is completely outside of the file.
1682 * This classification is stable under DLM lock acquired by
1683 * ll_tree_lock() above, because to change class, other client has to
1684 * take DLM lock conflicting with our lock. Also, any updates to
1685 * ->i_size by other threads on this client are serialized by
1686 * ll_inode_size_lock(). This guarantees that short reads are handled
1687 * correctly in the face of concurrent writes and truncates.
1689 inode_init_lvb(inode, &lvb);
1690 obd_merge_lvb(ll_i2sbi(inode)->ll_dt_exp, lsm, &lvb, 1);
1692 if (*ppos + count - 1 > kms) {
1693 /* A glimpse is necessary to determine whether we return a
1694 * short read (B) or some zeroes at the end of the buffer (C) */
1695 ll_inode_size_unlock(inode, 1);
1696 retval = ll_glimpse_size(inode, LDLM_FL_BLOCK_GRANTED);
1700 /* region is within kms and, hence, within real file size (A) */
1701 i_size_write(inode, kms);
1702 ll_inode_size_unlock(inode, 1);
1705 CDEBUG(D_INFO, "Send ino %lu, "LPSZ" bytes, offset %lld, i_size %llu\n",
1706 inode->i_ino, count, *ppos, i_size_read(inode));
1708 bead.lrr_start = *ppos >> CFS_PAGE_SHIFT;
1709 bead.lrr_count = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1710 ll_ra_read_in(in_file, &bead);
1712 file_accessed(in_file);
1713 retval = generic_file_sendfile(in_file, ppos, count, actor, target);
1714 ll_ra_read_ex(in_file, &bead);
1717 ll_tree_unlock(&tree);
1721 static int ll_lov_recreate_obj(struct inode *inode, struct file *file,
1724 struct ll_inode_info *lli = ll_i2info(inode);
1725 struct obd_export *exp = ll_i2dtexp(inode);
1726 struct ll_recreate_obj ucreatp;
1727 struct obd_trans_info oti = { 0 };
1728 struct obdo *oa = NULL;
1731 struct lov_stripe_md *lsm, *lsm2;
1734 if (!capable (CAP_SYS_ADMIN))
1737 rc = copy_from_user(&ucreatp, (struct ll_recreate_obj *)arg,
1738 sizeof(struct ll_recreate_obj));
1746 down(&lli->lli_size_sem);
1749 GOTO(out, rc = -ENOENT);
1750 lsm_size = sizeof(*lsm) + (sizeof(struct lov_oinfo) *
1751 (lsm->lsm_stripe_count));
1753 OBD_ALLOC(lsm2, lsm_size);
1755 GOTO(out, rc = -ENOMEM);
1757 oa->o_id = ucreatp.lrc_id;
1758 oa->o_gr = ucreatp.lrc_group;
1759 oa->o_nlink = ucreatp.lrc_ost_idx;
1760 oa->o_flags |= OBD_FL_RECREATE_OBJS;
1761 oa->o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
1762 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
1763 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
1765 memcpy(lsm2, lsm, lsm_size);
1766 rc = obd_create(exp, oa, &lsm2, &oti);
1768 OBD_FREE(lsm2, lsm_size);
1771 up(&lli->lli_size_sem);
1776 int ll_lov_setstripe_ea_info(struct inode *inode, struct file *file,
1777 int flags, struct lov_user_md *lum, int lum_size)
1779 struct ll_inode_info *lli = ll_i2info(inode);
1780 struct lov_stripe_md *lsm;
1781 struct lookup_intent oit = {.it_op = IT_OPEN, .it_flags = flags};
1785 down(&lli->lli_size_sem);
1788 up(&lli->lli_size_sem);
1789 CDEBUG(D_IOCTL, "stripe already exists for ino %lu\n",
1794 rc = ll_intent_file_open(file, lum, lum_size, &oit);
1797 if (it_disposition(&oit, DISP_LOOKUP_NEG))
1798 GOTO(out_req_free, rc = -ENOENT);
1799 rc = oit.d.lustre.it_status;
1801 GOTO(out_req_free, rc);
1803 ll_release_openhandle(file->f_dentry, &oit);
1806 up(&lli->lli_size_sem);
1807 ll_intent_release(&oit);
1810 ptlrpc_req_finished((struct ptlrpc_request *) oit.d.lustre.it_data);
1814 int ll_lov_getstripe_ea_info(struct inode *inode, const char *filename,
1815 struct lov_mds_md **lmmp, int *lmm_size,
1816 struct ptlrpc_request **request)
1818 struct ll_sb_info *sbi = ll_i2sbi(inode);
1819 struct mdt_body *body;
1820 struct lov_mds_md *lmm = NULL;
1821 struct ptlrpc_request *req = NULL;
1822 struct obd_capa *oc;
1825 rc = ll_get_max_mdsize(sbi, &lmmsize);
1829 oc = ll_mdscapa_get(inode);
1830 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode),
1831 oc, filename, strlen(filename) + 1,
1832 OBD_MD_FLEASIZE | OBD_MD_FLDIREA, lmmsize,
1833 ll_i2suppgid(inode), &req);
1836 CDEBUG(D_INFO, "md_getattr_name failed "
1837 "on %s: rc %d\n", filename, rc);
1841 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1842 LASSERT(body != NULL); /* checked by mdc_getattr_name */
1844 lmmsize = body->eadatasize;
1846 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
1848 GOTO(out, rc = -ENODATA);
1851 lmm = req_capsule_server_sized_get(&req->rq_pill, &RMF_MDT_MD, lmmsize);
1852 LASSERT(lmm != NULL);
1855 * This is coming from the MDS, so is probably in
1856 * little endian. We convert it to host endian before
1857 * passing it to userspace.
1859 if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
1860 lustre_swab_lov_user_md((struct lov_user_md *)lmm);
1861 lustre_swab_lov_user_md_objects((struct lov_user_md *)lmm);
1862 } else if (lmm->lmm_magic == __swab32(LOV_MAGIC_JOIN)) {
1863 lustre_swab_lov_user_md_join((struct lov_user_md_join *)lmm);
1866 if (lmm->lmm_magic == LOV_MAGIC_JOIN) {
1867 struct lov_stripe_md *lsm;
1868 struct lov_user_md_join *lmj;
1869 int lmj_size, i, aindex = 0;
1871 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1873 GOTO(out, rc = -ENOMEM);
1874 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
1876 GOTO(out_free_memmd, rc);
1878 lmj_size = sizeof(struct lov_user_md_join) +
1879 lsm->lsm_stripe_count *
1880 sizeof(struct lov_user_ost_data_join);
1881 OBD_ALLOC(lmj, lmj_size);
1883 GOTO(out_free_memmd, rc = -ENOMEM);
1885 memcpy(lmj, lmm, sizeof(struct lov_user_md_join));
1886 for (i = 0; i < lsm->lsm_stripe_count; i++) {
1887 struct lov_extent *lex =
1888 &lsm->lsm_array->lai_ext_array[aindex];
1890 if (lex->le_loi_idx + lex->le_stripe_count <= i)
1892 CDEBUG(D_INFO, "aindex %d i %d l_extent_start "
1893 LPU64" len %d\n", aindex, i,
1894 lex->le_start, (int)lex->le_len);
1895 lmj->lmm_objects[i].l_extent_start =
1898 if ((int)lex->le_len == -1)
1899 lmj->lmm_objects[i].l_extent_end = -1;
1901 lmj->lmm_objects[i].l_extent_end =
1902 lex->le_start + lex->le_len;
1903 lmj->lmm_objects[i].l_object_id =
1904 lsm->lsm_oinfo[i]->loi_id;
1905 lmj->lmm_objects[i].l_object_gr =
1906 lsm->lsm_oinfo[i]->loi_gr;
1907 lmj->lmm_objects[i].l_ost_gen =
1908 lsm->lsm_oinfo[i]->loi_ost_gen;
1909 lmj->lmm_objects[i].l_ost_idx =
1910 lsm->lsm_oinfo[i]->loi_ost_idx;
1912 lmm = (struct lov_mds_md *)lmj;
1915 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1919 *lmm_size = lmmsize;
1924 static int ll_lov_setea(struct inode *inode, struct file *file,
1927 int flags = MDS_OPEN_HAS_OBJS | FMODE_WRITE;
1928 struct lov_user_md *lump;
1929 int lum_size = sizeof(struct lov_user_md) +
1930 sizeof(struct lov_user_ost_data);
1934 if (!capable (CAP_SYS_ADMIN))
1937 OBD_ALLOC(lump, lum_size);
1941 rc = copy_from_user(lump, (struct lov_user_md *)arg, lum_size);
1943 OBD_FREE(lump, lum_size);
1947 rc = ll_lov_setstripe_ea_info(inode, file, flags, lump, lum_size);
1949 OBD_FREE(lump, lum_size);
1953 static int ll_lov_setstripe(struct inode *inode, struct file *file,
1956 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
1958 int flags = FMODE_WRITE;
1961 /* Bug 1152: copy properly when this is no longer true */
1962 LASSERT(sizeof(lum) == sizeof(*lump));
1963 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lump->lmm_objects[0]));
1964 rc = copy_from_user(&lum, lump, sizeof(lum));
1968 rc = ll_lov_setstripe_ea_info(inode, file, flags, &lum, sizeof(lum));
1970 put_user(0, &lump->lmm_stripe_count);
1971 rc = obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode),
1972 0, ll_i2info(inode)->lli_smd, lump);
1977 static int ll_lov_getstripe(struct inode *inode, unsigned long arg)
1979 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
1984 return obd_iocontrol(LL_IOC_LOV_GETSTRIPE, ll_i2dtexp(inode), 0, lsm,
1988 static int ll_get_grouplock(struct inode *inode, struct file *file,
1991 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1992 ldlm_policy_data_t policy = { .l_extent = { .start = 0,
1993 .end = OBD_OBJECT_EOF}};
1994 struct lustre_handle lockh = { 0 };
1995 struct ll_inode_info *lli = ll_i2info(inode);
1996 struct lov_stripe_md *lsm = lli->lli_smd;
2000 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
2004 policy.l_extent.gid = arg;
2005 if (file->f_flags & O_NONBLOCK)
2006 flags = LDLM_FL_BLOCK_NOWAIT;
2008 rc = ll_extent_lock(fd, inode, lsm, LCK_GROUP, &policy, &lockh, flags);
2012 fd->fd_flags |= LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK;
2014 memcpy(&fd->fd_cwlockh, &lockh, sizeof(lockh));
2019 static int ll_put_grouplock(struct inode *inode, struct file *file,
2022 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
2023 struct ll_inode_info *lli = ll_i2info(inode);
2024 struct lov_stripe_md *lsm = lli->lli_smd;
2028 if (!(fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
2029 /* Ugh, it's already unlocked. */
2033 if (fd->fd_gid != arg) /* Ugh? Unlocking with different gid? */
2036 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
2038 rc = ll_extent_unlock(fd, inode, lsm, LCK_GROUP, &fd->fd_cwlockh);
2043 memset(&fd->fd_cwlockh, 0, sizeof(fd->fd_cwlockh));
2048 static int join_sanity_check(struct inode *head, struct inode *tail)
2051 if ((ll_i2sbi(head)->ll_flags & LL_SBI_JOIN) == 0) {
2052 CERROR("server do not support join \n");
2055 if (!S_ISREG(tail->i_mode) || !S_ISREG(head->i_mode)) {
2056 CERROR("tail ino %lu and ino head %lu must be regular\n",
2057 head->i_ino, tail->i_ino);
2060 if (head->i_ino == tail->i_ino) {
2061 CERROR("file %lu can not be joined to itself \n", head->i_ino);
2064 if (i_size_read(head) % JOIN_FILE_ALIGN) {
2065 CERROR("hsize %llu must be times of 64K\n", i_size_read(head));
2071 static int join_file(struct inode *head_inode, struct file *head_filp,
2072 struct file *tail_filp)
2074 struct dentry *tail_dentry = tail_filp->f_dentry;
2075 struct lookup_intent oit = {.it_op = IT_OPEN,
2076 .it_flags = head_filp->f_flags|O_JOIN_FILE};
2077 struct ldlm_enqueue_info einfo = { LDLM_IBITS, LCK_CW,
2078 ll_md_blocking_ast, ldlm_completion_ast, NULL, NULL };
2080 struct lustre_handle lockh;
2081 struct md_op_data *op_data;
2086 tail_dentry = tail_filp->f_dentry;
2088 data = i_size_read(head_inode);
2089 op_data = ll_prep_md_op_data(NULL, head_inode,
2090 tail_dentry->d_parent->d_inode,
2091 tail_dentry->d_name.name,
2092 tail_dentry->d_name.len, 0,
2093 LUSTRE_OPC_ANY, &data);
2094 if (IS_ERR(op_data))
2095 RETURN(PTR_ERR(op_data));
2097 rc = md_enqueue(ll_i2mdexp(head_inode), &einfo, &oit,
2098 op_data, &lockh, NULL, 0, 0);
2100 ll_finish_md_op_data(op_data);
2104 rc = oit.d.lustre.it_status;
2106 if (rc < 0 || it_open_error(DISP_OPEN_OPEN, &oit)) {
2107 rc = rc ? rc : it_open_error(DISP_OPEN_OPEN, &oit);
2108 ptlrpc_req_finished((struct ptlrpc_request *)
2109 oit.d.lustre.it_data);
2113 if (oit.d.lustre.it_lock_mode) { /* If we got lock - release it right
2115 ldlm_lock_decref(&lockh, oit.d.lustre.it_lock_mode);
2116 oit.d.lustre.it_lock_mode = 0;
2118 ll_release_openhandle(head_filp->f_dentry, &oit);
2120 ll_intent_release(&oit);
2124 static int ll_file_join(struct inode *head, struct file *filp,
2125 char *filename_tail)
2127 struct inode *tail = NULL, *first = NULL, *second = NULL;
2128 struct dentry *tail_dentry;
2129 struct file *tail_filp, *first_filp, *second_filp;
2130 struct ll_lock_tree first_tree, second_tree;
2131 struct ll_lock_tree_node *first_node, *second_node;
2132 struct ll_inode_info *hlli = ll_i2info(head), *tlli;
2133 int rc = 0, cleanup_phase = 0;
2136 CDEBUG(D_VFSTRACE, "VFS Op:head=%lu/%u(%p) tail %s\n",
2137 head->i_ino, head->i_generation, head, filename_tail);
2139 tail_filp = filp_open(filename_tail, O_WRONLY, 0644);
2140 if (IS_ERR(tail_filp)) {
2141 CERROR("Can not open tail file %s", filename_tail);
2142 rc = PTR_ERR(tail_filp);
2145 tail = igrab(tail_filp->f_dentry->d_inode);
2147 tlli = ll_i2info(tail);
2148 tail_dentry = tail_filp->f_dentry;
2149 LASSERT(tail_dentry);
2152 /*reorder the inode for lock sequence*/
2153 first = head->i_ino > tail->i_ino ? head : tail;
2154 second = head->i_ino > tail->i_ino ? tail : head;
2155 first_filp = head->i_ino > tail->i_ino ? filp : tail_filp;
2156 second_filp = head->i_ino > tail->i_ino ? tail_filp : filp;
2158 CDEBUG(D_INFO, "reorder object from %lu:%lu to %lu:%lu \n",
2159 head->i_ino, tail->i_ino, first->i_ino, second->i_ino);
2160 first_node = ll_node_from_inode(first, 0, OBD_OBJECT_EOF, LCK_EX);
2161 if (IS_ERR(first_node)){
2162 rc = PTR_ERR(first_node);
2165 first_tree.lt_fd = first_filp->private_data;
2166 rc = ll_tree_lock(&first_tree, first_node, NULL, 0, 0);
2171 second_node = ll_node_from_inode(second, 0, OBD_OBJECT_EOF, LCK_EX);
2172 if (IS_ERR(second_node)){
2173 rc = PTR_ERR(second_node);
2176 second_tree.lt_fd = second_filp->private_data;
2177 rc = ll_tree_lock(&second_tree, second_node, NULL, 0, 0);
2182 rc = join_sanity_check(head, tail);
2186 rc = join_file(head, filp, tail_filp);
2190 switch (cleanup_phase) {
2192 ll_tree_unlock(&second_tree);
2193 obd_cancel_unused(ll_i2dtexp(second),
2194 ll_i2info(second)->lli_smd, 0, NULL);
2196 ll_tree_unlock(&first_tree);
2197 obd_cancel_unused(ll_i2dtexp(first),
2198 ll_i2info(first)->lli_smd, 0, NULL);
2200 filp_close(tail_filp, 0);
2203 if (head && rc == 0) {
2204 obd_free_memmd(ll_i2sbi(head)->ll_dt_exp,
2206 hlli->lli_smd = NULL;
2211 CERROR("invalid cleanup_phase %d\n", cleanup_phase);
2217 int ll_release_openhandle(struct dentry *dentry, struct lookup_intent *it)
2219 struct inode *inode = dentry->d_inode;
2220 struct obd_client_handle *och;
2226 /* Root ? Do nothing. */
2227 if (dentry->d_inode->i_sb->s_root == dentry)
2230 /* No open handle to close? Move away */
2231 if (!it_disposition(it, DISP_OPEN_OPEN))
2234 LASSERT(it_open_error(DISP_OPEN_OPEN, it) == 0);
2236 OBD_ALLOC(och, sizeof(*och));
2238 GOTO(out, rc = -ENOMEM);
2240 ll_och_fill(ll_i2sbi(inode)->ll_md_exp,
2241 ll_i2info(inode), it, och);
2243 rc = ll_close_inode_openhandle(ll_i2sbi(inode)->ll_md_exp,
2246 /* this one is in place of ll_file_open */
2247 ptlrpc_req_finished(it->d.lustre.it_data);
2248 it_clear_disposition(it, DISP_ENQ_OPEN_REF);
2252 int ll_file_ioctl(struct inode *inode, struct file *file, unsigned int cmd,
2255 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
2259 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),cmd=%x\n", inode->i_ino,
2260 inode->i_generation, inode, cmd);
2261 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
2263 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
2264 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
2268 case LL_IOC_GETFLAGS:
2269 /* Get the current value of the file flags */
2270 return put_user(fd->fd_flags, (int *)arg);
2271 case LL_IOC_SETFLAGS:
2272 case LL_IOC_CLRFLAGS:
2273 /* Set or clear specific file flags */
2274 /* XXX This probably needs checks to ensure the flags are
2275 * not abused, and to handle any flag side effects.
2277 if (get_user(flags, (int *) arg))
2280 if (cmd == LL_IOC_SETFLAGS) {
2281 if ((flags & LL_FILE_IGNORE_LOCK) &&
2282 !(file->f_flags & O_DIRECT)) {
2283 CERROR("%s: unable to disable locking on "
2284 "non-O_DIRECT file\n", current->comm);
2288 fd->fd_flags |= flags;
2290 fd->fd_flags &= ~flags;
2293 case LL_IOC_LOV_SETSTRIPE:
2294 RETURN(ll_lov_setstripe(inode, file, arg));
2295 case LL_IOC_LOV_SETEA:
2296 RETURN(ll_lov_setea(inode, file, arg));
2297 case LL_IOC_LOV_GETSTRIPE:
2298 RETURN(ll_lov_getstripe(inode, arg));
2299 case LL_IOC_RECREATE_OBJ:
2300 RETURN(ll_lov_recreate_obj(inode, file, arg));
2301 case EXT3_IOC_GETFLAGS:
2302 case EXT3_IOC_SETFLAGS:
2303 RETURN(ll_iocontrol(inode, file, cmd, arg));
2304 case EXT3_IOC_GETVERSION_OLD:
2305 case EXT3_IOC_GETVERSION:
2306 RETURN(put_user(inode->i_generation, (int *)arg));
2311 ftail = getname((const char *)arg);
2313 RETURN(PTR_ERR(ftail));
2314 rc = ll_file_join(inode, file, ftail);
2318 case LL_IOC_GROUP_LOCK:
2319 RETURN(ll_get_grouplock(inode, file, arg));
2320 case LL_IOC_GROUP_UNLOCK:
2321 RETURN(ll_put_grouplock(inode, file, arg));
2322 case IOC_OBD_STATFS:
2323 RETURN(ll_obd_statfs(inode, (void *)arg));
2325 /* We need to special case any other ioctls we want to handle,
2326 * to send them to the MDS/OST as appropriate and to properly
2327 * network encode the arg field.
2328 case EXT3_IOC_SETVERSION_OLD:
2329 case EXT3_IOC_SETVERSION:
2331 case LL_IOC_FLUSHCTX:
2332 RETURN(ll_flush_ctx(inode));
2337 ll_iocontrol_call(inode, file, cmd, arg, &err))
2340 RETURN(obd_iocontrol(cmd, ll_i2dtexp(inode), 0, NULL,
2346 loff_t ll_file_seek(struct file *file, loff_t offset, int origin)
2348 struct inode *inode = file->f_dentry->d_inode;
2349 struct ll_inode_info *lli = ll_i2info(inode);
2350 struct lov_stripe_md *lsm = lli->lli_smd;
2353 retval = offset + ((origin == 2) ? i_size_read(inode) :
2354 (origin == 1) ? file->f_pos : 0);
2355 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), to=%Lu=%#Lx(%s)\n",
2356 inode->i_ino, inode->i_generation, inode, retval, retval,
2357 origin == 2 ? "SEEK_END": origin == 1 ? "SEEK_CUR" : "SEEK_SET");
2358 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_LLSEEK, 1);
2360 if (origin == 2) { /* SEEK_END */
2361 int nonblock = 0, rc;
2363 if (file->f_flags & O_NONBLOCK)
2364 nonblock = LDLM_FL_BLOCK_NOWAIT;
2367 rc = ll_glimpse_size(inode, nonblock);
2372 ll_inode_size_lock(inode, 0);
2373 offset += i_size_read(inode);
2374 ll_inode_size_unlock(inode, 0);
2375 } else if (origin == 1) { /* SEEK_CUR */
2376 offset += file->f_pos;
2380 if (offset >= 0 && offset <= ll_file_maxbytes(inode)) {
2381 if (offset != file->f_pos) {
2382 file->f_pos = offset;
2383 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2385 file->f_version = ++event;
2394 int ll_fsync(struct file *file, struct dentry *dentry, int data)
2396 struct inode *inode = dentry->d_inode;
2397 struct ll_inode_info *lli = ll_i2info(inode);
2398 struct lov_stripe_md *lsm = lli->lli_smd;
2399 struct ptlrpc_request *req;
2400 struct obd_capa *oc;
2403 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
2404 inode->i_generation, inode);
2405 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_FSYNC, 1);
2407 /* fsync's caller has already called _fdata{sync,write}, we want
2408 * that IO to finish before calling the osc and mdc sync methods */
2409 rc = filemap_fdatawait(inode->i_mapping);
2411 /* catch async errors that were recorded back when async writeback
2412 * failed for pages in this mapping. */
2413 err = lli->lli_async_rc;
2414 lli->lli_async_rc = 0;
2418 err = lov_test_and_clear_async_rc(lsm);
2423 oc = ll_mdscapa_get(inode);
2424 err = md_sync(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode), oc,
2430 ptlrpc_req_finished(req);
2437 RETURN(rc ? rc : -ENOMEM);
2439 oa->o_id = lsm->lsm_object_id;
2440 oa->o_gr = lsm->lsm_object_gr;
2441 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2442 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
2443 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
2446 oc = ll_osscapa_get(inode, CAPA_OPC_OSS_WRITE);
2447 err = obd_sync(ll_i2sbi(inode)->ll_dt_exp, oa, lsm,
2448 0, OBD_OBJECT_EOF, oc);
2458 int ll_file_flock(struct file *file, int cmd, struct file_lock *file_lock)
2460 struct inode *inode = file->f_dentry->d_inode;
2461 struct ll_sb_info *sbi = ll_i2sbi(inode);
2462 struct ldlm_res_id res_id =
2463 { .name = { fid_seq(ll_inode2fid(inode)),
2464 fid_oid(ll_inode2fid(inode)),
2465 fid_ver(ll_inode2fid(inode)),
2467 struct ldlm_enqueue_info einfo = { LDLM_FLOCK, 0, NULL,
2468 ldlm_flock_completion_ast, NULL, file_lock };
2469 struct lustre_handle lockh = {0};
2470 ldlm_policy_data_t flock;
2475 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu file_lock=%p\n",
2476 inode->i_ino, file_lock);
2478 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_FLOCK, 1);
2480 if (file_lock->fl_flags & FL_FLOCK) {
2481 LASSERT((cmd == F_SETLKW) || (cmd == F_SETLK));
2482 /* set missing params for flock() calls */
2483 file_lock->fl_end = OFFSET_MAX;
2484 file_lock->fl_pid = current->tgid;
2486 flock.l_flock.pid = file_lock->fl_pid;
2487 flock.l_flock.start = file_lock->fl_start;
2488 flock.l_flock.end = file_lock->fl_end;
2490 switch (file_lock->fl_type) {
2492 einfo.ei_mode = LCK_PR;
2495 /* An unlock request may or may not have any relation to
2496 * existing locks so we may not be able to pass a lock handle
2497 * via a normal ldlm_lock_cancel() request. The request may even
2498 * unlock a byte range in the middle of an existing lock. In
2499 * order to process an unlock request we need all of the same
2500 * information that is given with a normal read or write record
2501 * lock request. To avoid creating another ldlm unlock (cancel)
2502 * message we'll treat a LCK_NL flock request as an unlock. */
2503 einfo.ei_mode = LCK_NL;
2506 einfo.ei_mode = LCK_PW;
2509 CERROR("unknown fcntl lock type: %d\n", file_lock->fl_type);
2524 flags = LDLM_FL_BLOCK_NOWAIT;
2530 flags = LDLM_FL_TEST_LOCK;
2531 /* Save the old mode so that if the mode in the lock changes we
2532 * can decrement the appropriate reader or writer refcount. */
2533 file_lock->fl_type = einfo.ei_mode;
2536 CERROR("unknown fcntl lock command: %d\n", cmd);
2540 CDEBUG(D_DLMTRACE, "inode=%lu, pid=%u, flags=%#x, mode=%u, "
2541 "start="LPU64", end="LPU64"\n", inode->i_ino, flock.l_flock.pid,
2542 flags, einfo.ei_mode, flock.l_flock.start, flock.l_flock.end);
2544 rc = ldlm_cli_enqueue(sbi->ll_md_exp, NULL, &einfo, &res_id,
2545 &flock, &flags, NULL, 0, NULL, &lockh, 0);
2546 if ((file_lock->fl_flags & FL_FLOCK) && (rc == 0))
2547 ll_flock_lock_file_wait(file, file_lock, (cmd == F_SETLKW));
2548 #ifdef HAVE_F_OP_FLOCK
2549 if ((file_lock->fl_flags & FL_POSIX) && (rc == 0) &&
2550 !(flags & LDLM_FL_TEST_LOCK))
2551 posix_lock_file_wait(file, file_lock);
2557 int ll_file_noflock(struct file *file, int cmd, struct file_lock *file_lock)
2564 int ll_have_md_lock(struct inode *inode, __u64 bits)
2566 struct lustre_handle lockh;
2567 ldlm_policy_data_t policy = { .l_inodebits = {bits}};
2575 fid = &ll_i2info(inode)->lli_fid;
2576 CDEBUG(D_INFO, "trying to match res "DFID"\n", PFID(fid));
2578 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING | LDLM_FL_TEST_LOCK;
2579 if (md_lock_match(ll_i2mdexp(inode), flags, fid, LDLM_IBITS, &policy,
2580 LCK_CR|LCK_CW|LCK_PR|LCK_PW, &lockh)) {
2586 ldlm_mode_t ll_take_md_lock(struct inode *inode, __u64 bits,
2587 struct lustre_handle *lockh)
2589 ldlm_policy_data_t policy = { .l_inodebits = {bits}};
2595 fid = &ll_i2info(inode)->lli_fid;
2596 CDEBUG(D_INFO, "trying to match res "DFID"\n", PFID(fid));
2598 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
2599 rc = md_lock_match(ll_i2mdexp(inode), flags, fid, LDLM_IBITS, &policy,
2600 LCK_CR|LCK_CW|LCK_PR|LCK_PW, lockh);
2604 static int ll_inode_revalidate_fini(struct inode *inode, int rc) {
2605 if (rc == -ENOENT) { /* Already unlinked. Just update nlink
2606 * and return success */
2608 /* This path cannot be hit for regular files unless in
2609 * case of obscure races, so no need to to validate
2611 if (!S_ISREG(inode->i_mode) &&
2612 !S_ISDIR(inode->i_mode))
2617 CERROR("failure %d inode %lu\n", rc, inode->i_ino);
2625 int ll_inode_revalidate_it(struct dentry *dentry, struct lookup_intent *it)
2627 struct inode *inode = dentry->d_inode;
2628 struct ptlrpc_request *req = NULL;
2629 struct ll_sb_info *sbi;
2630 struct obd_export *exp;
2635 CERROR("REPORT THIS LINE TO PETER\n");
2638 sbi = ll_i2sbi(inode);
2640 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),name=%s\n",
2641 inode->i_ino, inode->i_generation, inode, dentry->d_name.name);
2643 exp = ll_i2mdexp(inode);
2645 if (exp->exp_connect_flags & OBD_CONNECT_ATTRFID) {
2646 struct lookup_intent oit = { .it_op = IT_GETATTR };
2647 struct md_op_data *op_data;
2649 /* Call getattr by fid, so do not provide name at all. */
2650 op_data = ll_prep_md_op_data(NULL, dentry->d_parent->d_inode,
2651 dentry->d_inode, NULL, 0, 0,
2652 LUSTRE_OPC_ANY, NULL);
2653 if (IS_ERR(op_data))
2654 RETURN(PTR_ERR(op_data));
2656 oit.it_flags |= O_CHECK_STALE;
2657 rc = md_intent_lock(exp, op_data, NULL, 0,
2658 /* we are not interested in name
2661 ll_md_blocking_ast, 0);
2662 ll_finish_md_op_data(op_data);
2663 oit.it_flags &= ~O_CHECK_STALE;
2665 rc = ll_inode_revalidate_fini(inode, rc);
2669 rc = ll_revalidate_it_finish(req, &oit, dentry);
2671 ll_intent_release(&oit);
2675 /* Unlinked? Unhash dentry, so it is not picked up later by
2676 do_lookup() -> ll_revalidate_it(). We cannot use d_drop
2677 here to preserve get_cwd functionality on 2.6.
2679 if (!dentry->d_inode->i_nlink) {
2680 spin_lock(&dcache_lock);
2681 ll_drop_dentry(dentry);
2682 spin_unlock(&dcache_lock);
2685 ll_lookup_finish_locks(&oit, dentry);
2686 } else if (!ll_have_md_lock(dentry->d_inode, MDS_INODELOCK_UPDATE |
2687 MDS_INODELOCK_LOOKUP)) {
2688 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
2689 obd_valid valid = OBD_MD_FLGETATTR;
2690 struct obd_capa *oc;
2693 if (S_ISREG(inode->i_mode)) {
2694 rc = ll_get_max_mdsize(sbi, &ealen);
2697 valid |= OBD_MD_FLEASIZE | OBD_MD_FLMODEASIZE;
2699 /* Once OBD_CONNECT_ATTRFID is not supported, we can't find one
2700 * capa for this inode. Because we only keep capas of dirs
2702 oc = ll_mdscapa_get(inode);
2703 rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode), oc, valid,
2707 rc = ll_inode_revalidate_fini(inode, rc);
2711 rc = ll_prep_inode(&inode, req, NULL);
2716 /* if object not yet allocated, don't validate size */
2717 if (ll_i2info(inode)->lli_smd == NULL)
2720 /* ll_glimpse_size will prefer locally cached writes if they extend
2722 rc = ll_glimpse_size(inode, 0);
2725 ptlrpc_req_finished(req);
2729 int ll_getattr_it(struct vfsmount *mnt, struct dentry *de,
2730 struct lookup_intent *it, struct kstat *stat)
2732 struct inode *inode = de->d_inode;
2735 res = ll_inode_revalidate_it(de, it);
2736 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_GETATTR, 1);
2741 stat->dev = inode->i_sb->s_dev;
2742 stat->ino = inode->i_ino;
2743 stat->mode = inode->i_mode;
2744 stat->nlink = inode->i_nlink;
2745 stat->uid = inode->i_uid;
2746 stat->gid = inode->i_gid;
2747 stat->rdev = kdev_t_to_nr(inode->i_rdev);
2748 stat->atime = inode->i_atime;
2749 stat->mtime = inode->i_mtime;
2750 stat->ctime = inode->i_ctime;
2751 #ifdef HAVE_INODE_BLKSIZE
2752 stat->blksize = inode->i_blksize;
2754 stat->blksize = 1 << inode->i_blkbits;
2757 ll_inode_size_lock(inode, 0);
2758 stat->size = i_size_read(inode);
2759 stat->blocks = inode->i_blocks;
2760 ll_inode_size_unlock(inode, 0);
2764 int ll_getattr(struct vfsmount *mnt, struct dentry *de, struct kstat *stat)
2766 struct lookup_intent it = { .it_op = IT_GETATTR };
2768 return ll_getattr_it(mnt, de, &it, stat);
2772 int lustre_check_acl(struct inode *inode, int mask)
2774 #ifdef CONFIG_FS_POSIX_ACL
2775 struct ll_inode_info *lli = ll_i2info(inode);
2776 struct posix_acl *acl;
2780 spin_lock(&lli->lli_lock);
2781 acl = posix_acl_dup(lli->lli_posix_acl);
2782 spin_unlock(&lli->lli_lock);
2787 rc = posix_acl_permission(inode, acl, mask);
2788 posix_acl_release(acl);
2796 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,10))
2797 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
2799 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), mask %o\n",
2800 inode->i_ino, inode->i_generation, inode, mask);
2801 if (ll_i2sbi(inode)->ll_flags & LL_SBI_RMT_CLIENT)
2802 return lustre_check_remote_perm(inode, mask);
2804 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_INODE_PERM, 1);
2805 return generic_permission(inode, mask, lustre_check_acl);
2808 int ll_inode_permission(struct inode *inode, int mask, struct nameidata *nd)
2810 int mode = inode->i_mode;
2813 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), mask %o\n",
2814 inode->i_ino, inode->i_generation, inode, mask);
2816 if (ll_i2sbi(inode)->ll_flags & LL_SBI_RMT_CLIENT)
2817 return lustre_check_remote_perm(inode, mask);
2819 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_INODE_PERM, 1);
2821 if ((mask & MAY_WRITE) && IS_RDONLY(inode) &&
2822 (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
2824 if ((mask & MAY_WRITE) && IS_IMMUTABLE(inode))
2826 if (current->fsuid == inode->i_uid) {
2829 if (((mode >> 3) & mask & S_IRWXO) != mask)
2831 rc = lustre_check_acl(inode, mask);
2835 goto check_capabilities;
2839 if (in_group_p(inode->i_gid))
2842 if ((mode & mask & S_IRWXO) == mask)
2846 if (!(mask & MAY_EXEC) ||
2847 (inode->i_mode & S_IXUGO) || S_ISDIR(inode->i_mode))
2848 if (capable(CAP_DAC_OVERRIDE))
2851 if (capable(CAP_DAC_READ_SEARCH) && ((mask == MAY_READ) ||
2852 (S_ISDIR(inode->i_mode) && !(mask & MAY_WRITE))))
2859 /* -o localflock - only provides locally consistent flock locks */
2860 struct file_operations ll_file_operations = {
2861 .read = ll_file_read,
2862 .write = ll_file_write,
2863 .ioctl = ll_file_ioctl,
2864 .open = ll_file_open,
2865 .release = ll_file_release,
2866 .mmap = ll_file_mmap,
2867 .llseek = ll_file_seek,
2868 .sendfile = ll_file_sendfile,
2872 struct file_operations ll_file_operations_flock = {
2873 .read = ll_file_read,
2874 .write = ll_file_write,
2875 .ioctl = ll_file_ioctl,
2876 .open = ll_file_open,
2877 .release = ll_file_release,
2878 .mmap = ll_file_mmap,
2879 .llseek = ll_file_seek,
2880 .sendfile = ll_file_sendfile,
2882 #ifdef HAVE_F_OP_FLOCK
2883 .flock = ll_file_flock,
2885 .lock = ll_file_flock
2888 /* These are for -o noflock - to return ENOSYS on flock calls */
2889 struct file_operations ll_file_operations_noflock = {
2890 .read = ll_file_read,
2891 .write = ll_file_write,
2892 .ioctl = ll_file_ioctl,
2893 .open = ll_file_open,
2894 .release = ll_file_release,
2895 .mmap = ll_file_mmap,
2896 .llseek = ll_file_seek,
2897 .sendfile = ll_file_sendfile,
2899 #ifdef HAVE_F_OP_FLOCK
2900 .flock = ll_file_noflock,
2902 .lock = ll_file_noflock
2905 struct inode_operations ll_file_inode_operations = {
2906 #ifdef HAVE_VFS_INTENT_PATCHES
2907 .setattr_raw = ll_setattr_raw,
2909 .setattr = ll_setattr,
2910 .truncate = ll_truncate,
2911 .getattr = ll_getattr,
2912 .permission = ll_inode_permission,
2913 .setxattr = ll_setxattr,
2914 .getxattr = ll_getxattr,
2915 .listxattr = ll_listxattr,
2916 .removexattr = ll_removexattr,
2919 /* dynamic ioctl number support routins */
2920 static struct llioc_ctl_data {
2921 struct rw_semaphore ioc_sem;
2922 struct list_head ioc_head;
2924 __RWSEM_INITIALIZER(llioc.ioc_sem),
2925 CFS_LIST_HEAD_INIT(llioc.ioc_head)
2930 struct list_head iocd_list;
2931 unsigned int iocd_size;
2932 llioc_callback_t iocd_cb;
2933 unsigned int iocd_count;
2934 unsigned int iocd_cmd[0];
2937 void *ll_iocontrol_register(llioc_callback_t cb, int count, unsigned int *cmd)
2940 struct llioc_data *in_data = NULL;
2943 if (cb == NULL || cmd == NULL ||
2944 count > LLIOC_MAX_CMD || count < 0)
2947 size = sizeof(*in_data) + count * sizeof(unsigned int);
2948 OBD_ALLOC(in_data, size);
2949 if (in_data == NULL)
2952 memset(in_data, 0, sizeof(*in_data));
2953 in_data->iocd_size = size;
2954 in_data->iocd_cb = cb;
2955 in_data->iocd_count = count;
2956 memcpy(in_data->iocd_cmd, cmd, sizeof(unsigned int) * count);
2958 down_write(&llioc.ioc_sem);
2959 list_add_tail(&in_data->iocd_list, &llioc.ioc_head);
2960 up_write(&llioc.ioc_sem);
2965 void ll_iocontrol_unregister(void *magic)
2967 struct llioc_data *tmp;
2972 down_write(&llioc.ioc_sem);
2973 list_for_each_entry(tmp, &llioc.ioc_head, iocd_list) {
2975 unsigned int size = tmp->iocd_size;
2977 list_del(&tmp->iocd_list);
2978 up_write(&llioc.ioc_sem);
2980 OBD_FREE(tmp, size);
2984 up_write(&llioc.ioc_sem);
2986 CWARN("didn't find iocontrol register block with magic: %p\n", magic);
2989 EXPORT_SYMBOL(ll_iocontrol_register);
2990 EXPORT_SYMBOL(ll_iocontrol_unregister);
2992 enum llioc_iter ll_iocontrol_call(struct inode *inode, struct file *file,
2993 unsigned int cmd, unsigned long arg, int *rcp)
2995 enum llioc_iter ret = LLIOC_CONT;
2996 struct llioc_data *data;
2997 int rc = -EINVAL, i;
2999 down_read(&llioc.ioc_sem);
3000 list_for_each_entry(data, &llioc.ioc_head, iocd_list) {
3001 for (i = 0; i < data->iocd_count; i++) {
3002 if (cmd != data->iocd_cmd[i])
3005 ret = data->iocd_cb(inode, file, cmd, arg, data, &rc);
3009 if (ret == LLIOC_STOP)
3012 up_read(&llioc.ioc_sem);