1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
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 <lustre_mdc.h>
46 #include <lustre_lite.h>
47 #include "llite_internal.h"
49 /* record that a write is in flight */
50 void llap_write_pending(struct inode *inode, struct ll_async_page *llap)
52 struct ll_inode_info *lli = ll_i2info(inode);
55 spin_lock(&lli->lli_lock);
56 lli->lli_flags |= LLIF_SOM_DIRTY;
57 if (llap && list_empty(&llap->llap_pending_write))
58 list_add(&llap->llap_pending_write,
59 &lli->lli_pending_write_llaps);
60 spin_unlock(&lli->lli_lock);
64 /* record that a write has completed */
65 int llap_write_complete(struct inode *inode, struct ll_async_page *llap)
67 struct ll_inode_info *lli = ll_i2info(inode);
71 spin_lock(&lli->lli_lock);
72 if (llap && !list_empty(&llap->llap_pending_write)) {
73 list_del_init(&llap->llap_pending_write);
76 spin_unlock(&lli->lli_lock);
80 /* Queue DONE_WRITING if
81 * - done writing is allowed;
82 * - inode has no no dirty pages; */
83 void ll_queue_done_writing(struct inode *inode, unsigned long flags)
85 struct ll_inode_info *lli = ll_i2info(inode);
87 spin_lock(&lli->lli_lock);
88 lli->lli_flags |= flags;
90 if ((lli->lli_flags & LLIF_DONE_WRITING) &&
91 list_empty(&lli->lli_pending_write_llaps)) {
92 struct ll_close_queue *lcq = ll_i2sbi(inode)->ll_lcq;
94 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
95 CWARN("ino %lu/%u(flags %lu) som valid it just after "
97 inode->i_ino, inode->i_generation,
99 /* DONE_WRITING is allowed and inode has no dirty page. */
100 spin_lock(&lcq->lcq_lock);
102 LASSERT(list_empty(&lli->lli_close_list));
103 CDEBUG(D_INODE, "adding inode %lu/%u to close list\n",
104 inode->i_ino, inode->i_generation);
105 list_add_tail(&lli->lli_close_list, &lcq->lcq_head);
107 /* Avoid a concurrent insertion into the close thread queue:
108 * an inode is already in the close thread, open(), write(),
109 * close() happen, epoch is closed as the inode is marked as
110 * LLIF_EPOCH_PENDING. When pages are written inode should not
111 * be inserted into the queue again, clear this flag to avoid
113 lli->lli_flags &= ~LLIF_DONE_WRITING;
115 wake_up(&lcq->lcq_waitq);
116 spin_unlock(&lcq->lcq_lock);
118 spin_unlock(&lli->lli_lock);
121 /* Close epoch and send Size-on-MDS attribute update if possible.
122 * Call this under @lli->lli_lock spinlock. */
123 void ll_epoch_close(struct inode *inode, struct md_op_data *op_data,
124 struct obd_client_handle **och, unsigned long flags)
126 struct ll_inode_info *lli = ll_i2info(inode);
129 spin_lock(&lli->lli_lock);
130 if (!(list_empty(&lli->lli_pending_write_llaps))) {
131 if (!(lli->lli_flags & LLIF_EPOCH_PENDING)) {
132 LASSERT(*och != NULL);
133 LASSERT(lli->lli_pending_och == NULL);
134 /* Inode is dirty and there is no pending write done
135 * request yet, DONE_WRITE is to be sent later. */
136 lli->lli_flags |= LLIF_EPOCH_PENDING;
137 lli->lli_pending_och = *och;
138 spin_unlock(&lli->lli_lock);
140 inode = igrab(inode);
144 if (flags & LLIF_DONE_WRITING) {
145 /* Some pages are still dirty, it is early to send
146 * DONE_WRITE. Wait untill all pages will be flushed
147 * and try DONE_WRITE again later. */
148 LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
149 lli->lli_flags |= LLIF_DONE_WRITING;
150 spin_unlock(&lli->lli_lock);
152 inode = igrab(inode);
157 CDEBUG(D_INODE, "Epoch "LPU64" closed on "DFID"\n",
158 ll_i2info(inode)->lli_ioepoch, PFID(&lli->lli_fid));
159 op_data->op_flags |= MF_EPOCH_CLOSE;
161 if (flags & LLIF_DONE_WRITING) {
162 LASSERT(lli->lli_flags & LLIF_SOM_DIRTY);
163 LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
164 *och = lli->lli_pending_och;
165 lli->lli_pending_och = NULL;
166 lli->lli_flags &= ~LLIF_EPOCH_PENDING;
168 /* Pack Size-on-MDS inode attributes only if they has changed */
169 if (!(lli->lli_flags & LLIF_SOM_DIRTY)) {
170 spin_unlock(&lli->lli_lock);
174 /* There is a pending DONE_WRITE -- close epoch with no
175 * attribute change. */
176 if (lli->lli_flags & LLIF_EPOCH_PENDING) {
177 spin_unlock(&lli->lli_lock);
182 LASSERT(list_empty(&lli->lli_pending_write_llaps));
183 lli->lli_flags &= ~LLIF_SOM_DIRTY;
184 spin_unlock(&lli->lli_lock);
185 op_data->op_flags |= MF_SOM_CHANGE;
187 /* Check if Size-on-MDS attributes are valid. */
188 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
189 CWARN("ino %lu/%u(flags %lu) som valid it just after "
191 inode->i_ino, inode->i_generation, lli->lli_flags);
193 if (!ll_local_size(inode)) {
194 /* Send Size-on-MDS Attributes if valid. Atime is sent along
195 * with all the attributes. */
196 op_data->op_attr.ia_valid |= ATTR_MTIME_SET | ATTR_CTIME_SET |
197 ATTR_ATIME_SET | ATTR_SIZE | ATTR_BLOCKS;
204 int ll_sizeonmds_update(struct inode *inode, struct md_open_data *mod,
205 struct lustre_handle *fh, __u64 ioepoch)
207 struct ll_inode_info *lli = ll_i2info(inode);
208 struct md_op_data *op_data;
213 /* LASSERT(!(lli->lli_flags & LLIF_MDS_SIZE_LOCK)); */
214 /* After recovery that can be valid. */
215 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
216 CWARN("ino %lu/%u(flags %lu) som valid it just after "
217 "recovery\n", inode->i_ino, inode->i_generation,
221 OBD_ALLOC_PTR(op_data);
222 if (!oa || !op_data) {
223 CERROR("can't allocate memory for Size-on-MDS update.\n");
226 rc = ll_inode_getattr(inode, oa);
229 CDEBUG(D_INODE, "objid "LPX64" is already destroyed\n",
230 lli->lli_smd->lsm_object_id);
232 CERROR("inode_getattr failed (%d): unable to send a "
233 "Size-on-MDS attribute update for inode %lu/%u\n",
234 rc, inode->i_ino, inode->i_generation);
237 CDEBUG(D_INODE, "Size-on-MDS update on "DFID"\n", PFID(&lli->lli_fid));
239 md_from_obdo(op_data, oa, oa->o_valid);
240 memcpy(&op_data->op_handle, fh, sizeof(*fh));
242 op_data->op_ioepoch = ioepoch;
243 op_data->op_flags |= MF_SOM_CHANGE;
245 rc = ll_md_setattr(inode, op_data, &mod);
251 ll_finish_md_op_data(op_data);
255 /* Send a DONE_WRITING rpc, pack Size-on-MDS attributes into it, if possible */
256 static void ll_done_writing(struct inode *inode)
258 struct obd_client_handle *och = NULL;
259 struct md_op_data *op_data;
263 LASSERT(ll_i2mdexp(inode)->exp_connect_flags & OBD_CONNECT_SOM);
265 OBD_ALLOC_PTR(op_data);
266 if (op_data == NULL) {
267 CERROR("can't allocate op_data\n");
272 ll_epoch_close(inode, op_data, &och, LLIF_DONE_WRITING);
273 /* If there is no @och, we do not do D_W yet. */
277 ll_pack_inode2opdata(inode, op_data, &och->och_fh);
279 rc = md_done_writing(ll_i2sbi(inode)->ll_md_exp, op_data, och->och_mod);
281 /* MDS has instructed us to obtain Size-on-MDS attribute from
282 * OSTs and send setattr to back to MDS. */
283 rc = ll_sizeonmds_update(inode, och->och_mod,
284 &och->och_fh, op_data->op_ioepoch);
286 CERROR("inode %lu mdc done_writing failed: rc = %d\n",
290 ll_finish_md_op_data(op_data);
292 md_clear_open_replay_data(ll_i2sbi(inode)->ll_md_exp, och);
298 static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
300 struct ll_inode_info *lli = NULL;
302 spin_lock(&lcq->lcq_lock);
304 if (!list_empty(&lcq->lcq_head)) {
305 lli = list_entry(lcq->lcq_head.next, struct ll_inode_info,
307 list_del_init(&lli->lli_close_list);
308 } else if (atomic_read(&lcq->lcq_stop))
309 lli = ERR_PTR(-EALREADY);
311 spin_unlock(&lcq->lcq_lock);
315 static int ll_close_thread(void *arg)
317 struct ll_close_queue *lcq = arg;
321 char name[CFS_CURPROC_COMM_MAX];
322 snprintf(name, sizeof(name) - 1, "ll_close");
326 complete(&lcq->lcq_comp);
329 struct l_wait_info lwi = { 0 };
330 struct ll_inode_info *lli;
333 l_wait_event_exclusive(lcq->lcq_waitq,
334 (lli = ll_close_next_lli(lcq)) != NULL,
339 inode = ll_info2i(lli);
340 CDEBUG(D_INFO, "done_writting for inode %lu/%u\n",
341 inode->i_ino, inode->i_generation);
342 ll_done_writing(inode);
346 CDEBUG(D_INFO, "ll_close exiting\n");
347 complete(&lcq->lcq_comp);
351 int ll_close_thread_start(struct ll_close_queue **lcq_ret)
353 struct ll_close_queue *lcq;
356 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CLOSE_THREAD))
359 OBD_ALLOC(lcq, sizeof(*lcq));
363 spin_lock_init(&lcq->lcq_lock);
364 INIT_LIST_HEAD(&lcq->lcq_head);
365 init_waitqueue_head(&lcq->lcq_waitq);
366 init_completion(&lcq->lcq_comp);
368 pid = kernel_thread(ll_close_thread, lcq, 0);
370 OBD_FREE(lcq, sizeof(*lcq));
374 wait_for_completion(&lcq->lcq_comp);
379 void ll_close_thread_shutdown(struct ll_close_queue *lcq)
381 init_completion(&lcq->lcq_comp);
382 atomic_inc(&lcq->lcq_stop);
383 wake_up(&lcq->lcq_waitq);
384 wait_for_completion(&lcq->lcq_comp);
385 OBD_FREE(lcq, sizeof(*lcq));