4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2012, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/llite/llite_close.c
34 * Lustre Lite routines to issue a secondary close after writeback
37 #include <linux/module.h>
39 #define DEBUG_SUBSYSTEM S_LLITE
41 #include "../include/lustre_lite.h"
42 #include "llite_internal.h"
44 /** records that a write is in flight */
45 void vvp_write_pending(struct vvp_object *club, struct vvp_page *page)
47 struct ll_inode_info *lli = ll_i2info(club->vob_inode);
49 spin_lock(&lli->lli_lock);
50 lli->lli_flags |= LLIF_SOM_DIRTY;
51 if (page && list_empty(&page->vpg_pending_linkage))
52 list_add(&page->vpg_pending_linkage, &club->vob_pending_list);
53 spin_unlock(&lli->lli_lock);
56 /** records that a write has completed */
57 void vvp_write_complete(struct vvp_object *club, struct vvp_page *page)
59 struct ll_inode_info *lli = ll_i2info(club->vob_inode);
62 spin_lock(&lli->lli_lock);
63 if (page && !list_empty(&page->vpg_pending_linkage)) {
64 list_del_init(&page->vpg_pending_linkage);
67 spin_unlock(&lli->lli_lock);
69 ll_queue_done_writing(club->vob_inode, 0);
72 /** Queues DONE_WRITING if
73 * - done writing is allowed;
74 * - inode has no no dirty pages;
76 void ll_queue_done_writing(struct inode *inode, unsigned long flags)
78 struct ll_inode_info *lli = ll_i2info(inode);
79 struct vvp_object *club = cl2vvp(ll_i2info(inode)->lli_clob);
81 spin_lock(&lli->lli_lock);
82 lli->lli_flags |= flags;
84 if ((lli->lli_flags & LLIF_DONE_WRITING) &&
85 list_empty(&club->vob_pending_list)) {
86 struct ll_close_queue *lcq = ll_i2sbi(inode)->ll_lcq;
88 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
89 CWARN("%s: file "DFID"(flags %u) Size-on-MDS valid, done writing allowed and no diry pages\n",
90 ll_get_fsname(inode->i_sb, NULL, 0),
91 PFID(ll_inode2fid(inode)), lli->lli_flags);
92 /* DONE_WRITING is allowed and inode has no dirty page. */
93 spin_lock(&lcq->lcq_lock);
95 LASSERT(list_empty(&lli->lli_close_list));
96 CDEBUG(D_INODE, "adding inode "DFID" to close list\n",
97 PFID(ll_inode2fid(inode)));
98 list_add_tail(&lli->lli_close_list, &lcq->lcq_head);
100 /* Avoid a concurrent insertion into the close thread queue:
101 * an inode is already in the close thread, open(), write(),
102 * close() happen, epoch is closed as the inode is marked as
103 * LLIF_EPOCH_PENDING. When pages are written inode should not
104 * be inserted into the queue again, clear this flag to avoid
107 lli->lli_flags &= ~LLIF_DONE_WRITING;
109 wake_up(&lcq->lcq_waitq);
110 spin_unlock(&lcq->lcq_lock);
112 spin_unlock(&lli->lli_lock);
115 /** Pack SOM attributes info @opdata for CLOSE, DONE_WRITING rpc. */
116 void ll_done_writing_attr(struct inode *inode, struct md_op_data *op_data)
118 struct ll_inode_info *lli = ll_i2info(inode);
120 op_data->op_flags |= MF_SOM_CHANGE;
121 /* Check if Size-on-MDS attributes are valid. */
122 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
123 CERROR("%s: inode "DFID"(flags %u) MDS holds lock on Size-on-MDS attributes\n",
124 ll_get_fsname(inode->i_sb, NULL, 0),
125 PFID(ll_inode2fid(inode)), lli->lli_flags);
127 if (!cl_local_size(inode)) {
128 /* Send Size-on-MDS Attributes if valid. */
129 op_data->op_attr.ia_valid |= ATTR_MTIME_SET | ATTR_CTIME_SET |
130 ATTR_ATIME_SET | ATTR_SIZE | ATTR_BLOCKS;
134 /** Closes ioepoch and packs Size-on-MDS attribute if needed into @op_data. */
135 void ll_ioepoch_close(struct inode *inode, struct md_op_data *op_data,
136 struct obd_client_handle **och, unsigned long flags)
138 struct ll_inode_info *lli = ll_i2info(inode);
139 struct vvp_object *club = cl2vvp(ll_i2info(inode)->lli_clob);
141 spin_lock(&lli->lli_lock);
142 if (!(list_empty(&club->vob_pending_list))) {
143 if (!(lli->lli_flags & LLIF_EPOCH_PENDING)) {
145 LASSERT(!lli->lli_pending_och);
146 /* Inode is dirty and there is no pending write done
147 * request yet, DONE_WRITE is to be sent later.
149 lli->lli_flags |= LLIF_EPOCH_PENDING;
150 lli->lli_pending_och = *och;
151 spin_unlock(&lli->lli_lock);
153 inode = igrab(inode);
157 if (flags & LLIF_DONE_WRITING) {
158 /* Some pages are still dirty, it is early to send
159 * DONE_WRITE. Wait until all pages will be flushed
160 * and try DONE_WRITE again later.
162 LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
163 lli->lli_flags |= LLIF_DONE_WRITING;
164 spin_unlock(&lli->lli_lock);
166 inode = igrab(inode);
171 CDEBUG(D_INODE, "Epoch %llu closed on "DFID"\n",
172 ll_i2info(inode)->lli_ioepoch, PFID(&lli->lli_fid));
173 op_data->op_flags |= MF_EPOCH_CLOSE;
175 if (flags & LLIF_DONE_WRITING) {
176 LASSERT(lli->lli_flags & LLIF_SOM_DIRTY);
177 LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
178 *och = lli->lli_pending_och;
179 lli->lli_pending_och = NULL;
180 lli->lli_flags &= ~LLIF_EPOCH_PENDING;
182 /* Pack Size-on-MDS inode attributes only if they has changed */
183 if (!(lli->lli_flags & LLIF_SOM_DIRTY)) {
184 spin_unlock(&lli->lli_lock);
188 /* There is a pending DONE_WRITE -- close epoch with no
191 if (lli->lli_flags & LLIF_EPOCH_PENDING) {
192 spin_unlock(&lli->lli_lock);
197 LASSERT(list_empty(&club->vob_pending_list));
198 lli->lli_flags &= ~LLIF_SOM_DIRTY;
199 spin_unlock(&lli->lli_lock);
200 ll_done_writing_attr(inode, op_data);
207 * Cliens updates SOM attributes on MDS (including llog cookies):
208 * obd_getattr with no lock and md_setattr.
210 int ll_som_update(struct inode *inode, struct md_op_data *op_data)
212 struct ll_inode_info *lli = ll_i2info(inode);
213 struct ptlrpc_request *request = NULL;
219 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
220 CERROR("%s: inode "DFID"(flags %u) MDS holds lock on Size-on-MDS attributes\n",
221 ll_get_fsname(inode->i_sb, NULL, 0),
222 PFID(ll_inode2fid(inode)), lli->lli_flags);
224 oa = kmem_cache_zalloc(obdo_cachep, GFP_NOFS);
226 CERROR("can't allocate memory for Size-on-MDS update.\n");
230 old_flags = op_data->op_flags;
231 op_data->op_flags = MF_SOM_CHANGE;
233 /* If inode is already in another epoch, skip getattr from OSTs. */
234 if (lli->lli_ioepoch == op_data->op_ioepoch) {
235 rc = ll_inode_getattr(inode, oa, op_data->op_ioepoch,
236 old_flags & MF_GETATTR_LOCK);
240 CERROR("%s: inode_getattr failed - unable to send a Size-on-MDS attribute update for inode "DFID": rc = %d\n",
241 ll_get_fsname(inode->i_sb, NULL, 0),
242 PFID(ll_inode2fid(inode)), rc);
244 CDEBUG(D_INODE, "Size-on-MDS update on "DFID"\n",
245 PFID(&lli->lli_fid));
247 /* Install attributes into op_data. */
248 md_from_obdo(op_data, oa, oa->o_valid);
251 rc = md_setattr(ll_i2sbi(inode)->ll_md_exp, op_data,
252 NULL, 0, NULL, 0, &request, NULL);
253 ptlrpc_req_finished(request);
255 kmem_cache_free(obdo_cachep, oa);
260 * Closes the ioepoch and packs all the attributes into @op_data for
263 static void ll_prepare_done_writing(struct inode *inode,
264 struct md_op_data *op_data,
265 struct obd_client_handle **och)
267 ll_ioepoch_close(inode, op_data, och, LLIF_DONE_WRITING);
268 /* If there is no @och, we do not do D_W yet. */
272 ll_pack_inode2opdata(inode, op_data, &(*och)->och_fh);
273 ll_prep_md_op_data(op_data, inode, NULL, NULL,
274 0, 0, LUSTRE_OPC_ANY, NULL);
277 /** Send a DONE_WRITING rpc. */
278 static void ll_done_writing(struct inode *inode)
280 struct obd_client_handle *och = NULL;
281 struct md_op_data *op_data;
284 LASSERT(exp_connect_som(ll_i2mdexp(inode)));
286 op_data = kzalloc(sizeof(*op_data), GFP_NOFS);
290 ll_prepare_done_writing(inode, op_data, &och);
291 /* If there is no @och, we do not do D_W yet. */
295 rc = md_done_writing(ll_i2sbi(inode)->ll_md_exp, op_data, NULL);
297 /* MDS has instructed us to obtain Size-on-MDS attribute from
298 * OSTs and send setattr to back to MDS.
300 rc = ll_som_update(inode, op_data);
302 CERROR("%s: inode "DFID" mdc done_writing failed: rc = %d\n",
303 ll_get_fsname(inode->i_sb, NULL, 0),
304 PFID(ll_inode2fid(inode)), rc);
307 ll_finish_md_op_data(op_data);
309 md_clear_open_replay_data(ll_i2sbi(inode)->ll_md_exp, och);
314 static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
316 struct ll_inode_info *lli = NULL;
318 spin_lock(&lcq->lcq_lock);
320 if (!list_empty(&lcq->lcq_head)) {
321 lli = list_entry(lcq->lcq_head.next, struct ll_inode_info,
323 list_del_init(&lli->lli_close_list);
324 } else if (atomic_read(&lcq->lcq_stop)) {
325 lli = ERR_PTR(-EALREADY);
328 spin_unlock(&lcq->lcq_lock);
332 static int ll_close_thread(void *arg)
334 struct ll_close_queue *lcq = arg;
336 complete(&lcq->lcq_comp);
339 struct l_wait_info lwi = { 0 };
340 struct ll_inode_info *lli;
343 l_wait_event_exclusive(lcq->lcq_waitq,
344 (lli = ll_close_next_lli(lcq)) != NULL,
349 inode = ll_info2i(lli);
350 CDEBUG(D_INFO, "done_writing for inode "DFID"\n",
351 PFID(ll_inode2fid(inode)));
352 ll_done_writing(inode);
356 CDEBUG(D_INFO, "ll_close exiting\n");
357 complete(&lcq->lcq_comp);
361 int ll_close_thread_start(struct ll_close_queue **lcq_ret)
363 struct ll_close_queue *lcq;
364 struct task_struct *task;
366 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CLOSE_THREAD))
369 lcq = kzalloc(sizeof(*lcq), GFP_NOFS);
373 spin_lock_init(&lcq->lcq_lock);
374 INIT_LIST_HEAD(&lcq->lcq_head);
375 init_waitqueue_head(&lcq->lcq_waitq);
376 init_completion(&lcq->lcq_comp);
378 task = kthread_run(ll_close_thread, lcq, "ll_close");
381 return PTR_ERR(task);
384 wait_for_completion(&lcq->lcq_comp);
389 void ll_close_thread_shutdown(struct ll_close_queue *lcq)
391 init_completion(&lcq->lcq_comp);
392 atomic_inc(&lcq->lcq_stop);
393 wake_up(&lcq->lcq_waitq);
394 wait_for_completion(&lcq->lcq_comp);