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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/llite/llite_close.c
38 * Lustre Lite routines to issue a secondary close after writeback
41 #include <linux/module.h>
43 #define DEBUG_SUBSYSTEM S_LLITE
45 #include "../include/lustre_lite.h"
46 #include "llite_internal.h"
48 /** records that a write is in flight */
49 void vvp_write_pending(struct ccc_object *club, struct ccc_page *page)
51 struct ll_inode_info *lli = ll_i2info(club->cob_inode);
53 spin_lock(&lli->lli_lock);
54 lli->lli_flags |= LLIF_SOM_DIRTY;
55 if (page != NULL && list_empty(&page->cpg_pending_linkage))
56 list_add(&page->cpg_pending_linkage,
57 &club->cob_pending_list);
58 spin_unlock(&lli->lli_lock);
61 /** records that a write has completed */
62 void vvp_write_complete(struct ccc_object *club, struct ccc_page *page)
64 struct ll_inode_info *lli = ll_i2info(club->cob_inode);
67 spin_lock(&lli->lli_lock);
68 if (page != NULL && !list_empty(&page->cpg_pending_linkage)) {
69 list_del_init(&page->cpg_pending_linkage);
72 spin_unlock(&lli->lli_lock);
74 ll_queue_done_writing(club->cob_inode, 0);
77 /** Queues DONE_WRITING if
78 * - done writing is allowed;
79 * - inode has no no dirty pages; */
80 void ll_queue_done_writing(struct inode *inode, unsigned long flags)
82 struct ll_inode_info *lli = ll_i2info(inode);
83 struct ccc_object *club = cl2ccc(ll_i2info(inode)->lli_clob);
85 spin_lock(&lli->lli_lock);
86 lli->lli_flags |= flags;
88 if ((lli->lli_flags & LLIF_DONE_WRITING) &&
89 list_empty(&club->cob_pending_list)) {
90 struct ll_close_queue *lcq = ll_i2sbi(inode)->ll_lcq;
92 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
93 CWARN("ino %lu/%u(flags %u) som valid it just after recovery\n",
94 inode->i_ino, inode->i_generation,
96 /* DONE_WRITING is allowed and inode has no dirty page. */
97 spin_lock(&lcq->lcq_lock);
99 LASSERT(list_empty(&lli->lli_close_list));
100 CDEBUG(D_INODE, "adding inode %lu/%u to close list\n",
101 inode->i_ino, inode->i_generation);
102 list_add_tail(&lli->lli_close_list, &lcq->lcq_head);
104 /* Avoid a concurrent insertion into the close thread queue:
105 * an inode is already in the close thread, open(), write(),
106 * close() happen, epoch is closed as the inode is marked as
107 * LLIF_EPOCH_PENDING. When pages are written inode should not
108 * be inserted into the queue again, clear this flag to avoid
110 lli->lli_flags &= ~LLIF_DONE_WRITING;
112 wake_up(&lcq->lcq_waitq);
113 spin_unlock(&lcq->lcq_lock);
115 spin_unlock(&lli->lli_lock);
118 /** Pack SOM attributes info @opdata for CLOSE, DONE_WRITING rpc. */
119 void ll_done_writing_attr(struct inode *inode, struct md_op_data *op_data)
121 struct ll_inode_info *lli = ll_i2info(inode);
123 op_data->op_flags |= MF_SOM_CHANGE;
124 /* Check if Size-on-MDS attributes are valid. */
125 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
126 CERROR("ino %lu/%u(flags %u) som valid it just after recovery\n",
127 inode->i_ino, inode->i_generation,
130 if (!cl_local_size(inode)) {
131 /* Send Size-on-MDS Attributes if valid. */
132 op_data->op_attr.ia_valid |= ATTR_MTIME_SET | ATTR_CTIME_SET |
133 ATTR_ATIME_SET | ATTR_SIZE | ATTR_BLOCKS;
137 /** Closes ioepoch and packs Size-on-MDS attribute if needed into @op_data. */
138 void ll_ioepoch_close(struct inode *inode, struct md_op_data *op_data,
139 struct obd_client_handle **och, unsigned long flags)
141 struct ll_inode_info *lli = ll_i2info(inode);
142 struct ccc_object *club = cl2ccc(ll_i2info(inode)->lli_clob);
144 spin_lock(&lli->lli_lock);
145 if (!(list_empty(&club->cob_pending_list))) {
146 if (!(lli->lli_flags & LLIF_EPOCH_PENDING)) {
147 LASSERT(*och != NULL);
148 LASSERT(lli->lli_pending_och == NULL);
149 /* Inode is dirty and there is no pending write done
150 * request yet, DONE_WRITE is to be sent later. */
151 lli->lli_flags |= LLIF_EPOCH_PENDING;
152 lli->lli_pending_och = *och;
153 spin_unlock(&lli->lli_lock);
155 inode = igrab(inode);
159 if (flags & LLIF_DONE_WRITING) {
160 /* Some pages are still dirty, it is early to send
161 * DONE_WRITE. Wait until all pages will be flushed
162 * and try DONE_WRITE again later. */
163 LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
164 lli->lli_flags |= LLIF_DONE_WRITING;
165 spin_unlock(&lli->lli_lock);
167 inode = igrab(inode);
172 CDEBUG(D_INODE, "Epoch %llu closed on "DFID"\n",
173 ll_i2info(inode)->lli_ioepoch, PFID(&lli->lli_fid));
174 op_data->op_flags |= MF_EPOCH_CLOSE;
176 if (flags & LLIF_DONE_WRITING) {
177 LASSERT(lli->lli_flags & LLIF_SOM_DIRTY);
178 LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
179 *och = lli->lli_pending_och;
180 lli->lli_pending_och = NULL;
181 lli->lli_flags &= ~LLIF_EPOCH_PENDING;
183 /* Pack Size-on-MDS inode attributes only if they has changed */
184 if (!(lli->lli_flags & LLIF_SOM_DIRTY)) {
185 spin_unlock(&lli->lli_lock);
189 /* There is a pending DONE_WRITE -- close epoch with no
190 * attribute change. */
191 if (lli->lli_flags & LLIF_EPOCH_PENDING) {
192 spin_unlock(&lli->lli_lock);
197 LASSERT(list_empty(&club->cob_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;
218 LASSERT(op_data != NULL);
219 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
220 CERROR("ino %lu/%u(flags %u) som valid it just after recovery\n",
221 inode->i_ino, inode->i_generation,
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("inode_getattr failed (%d): unable to send a Size-on-MDS attribute update for inode %lu/%u\n",
242 inode->i_generation);
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);
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. */
299 rc = ll_som_update(inode, op_data);
301 CERROR("inode %lu mdc done_writing failed: rc = %d\n",
305 ll_finish_md_op_data(op_data);
307 md_clear_open_replay_data(ll_i2sbi(inode)->ll_md_exp, och);
312 static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
314 struct ll_inode_info *lli = NULL;
316 spin_lock(&lcq->lcq_lock);
318 if (!list_empty(&lcq->lcq_head)) {
319 lli = list_entry(lcq->lcq_head.next, struct ll_inode_info,
321 list_del_init(&lli->lli_close_list);
322 } else if (atomic_read(&lcq->lcq_stop))
323 lli = ERR_PTR(-EALREADY);
325 spin_unlock(&lcq->lcq_lock);
329 static int ll_close_thread(void *arg)
331 struct ll_close_queue *lcq = arg;
333 complete(&lcq->lcq_comp);
336 struct l_wait_info lwi = { 0 };
337 struct ll_inode_info *lli;
340 l_wait_event_exclusive(lcq->lcq_waitq,
341 (lli = ll_close_next_lli(lcq)) != NULL,
346 inode = ll_info2i(lli);
347 CDEBUG(D_INFO, "done_writing for inode %lu/%u\n",
348 inode->i_ino, inode->i_generation);
349 ll_done_writing(inode);
353 CDEBUG(D_INFO, "ll_close exiting\n");
354 complete(&lcq->lcq_comp);
358 int ll_close_thread_start(struct ll_close_queue **lcq_ret)
360 struct ll_close_queue *lcq;
361 struct task_struct *task;
363 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CLOSE_THREAD))
366 lcq = kzalloc(sizeof(*lcq), GFP_NOFS);
370 spin_lock_init(&lcq->lcq_lock);
371 INIT_LIST_HEAD(&lcq->lcq_head);
372 init_waitqueue_head(&lcq->lcq_waitq);
373 init_completion(&lcq->lcq_comp);
375 task = kthread_run(ll_close_thread, lcq, "ll_close");
378 return PTR_ERR(task);
381 wait_for_completion(&lcq->lcq_comp);
386 void ll_close_thread_shutdown(struct ll_close_queue *lcq)
388 init_completion(&lcq->lcq_comp);
389 atomic_inc(&lcq->lcq_stop);
390 wake_up(&lcq->lcq_waitq);
391 wait_for_completion(&lcq->lcq_comp);