1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Lustre Lite routines to issue a secondary close after writeback
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #include <linux/module.h>
26 #define DEBUG_SUBSYSTEM S_LLITE
28 #include <lustre_mdc.h>
29 #include <lustre_lite.h>
30 #include "llite_internal.h"
32 /* record that a write is in flight */
33 void llap_write_pending(struct inode *inode, struct ll_async_page *llap)
35 struct ll_inode_info *lli = ll_i2info(inode);
38 spin_lock(&lli->lli_lock);
39 lli->lli_flags |= LLIF_SOM_DIRTY;
40 if (llap && list_empty(&llap->llap_pending_write))
41 list_add(&llap->llap_pending_write,
42 &lli->lli_pending_write_llaps);
43 spin_unlock(&lli->lli_lock);
47 /* record that a write has completed */
48 int llap_write_complete(struct inode *inode, struct ll_async_page *llap)
50 struct ll_inode_info *lli = ll_i2info(inode);
54 spin_lock(&lli->lli_lock);
55 if (llap && !list_empty(&llap->llap_pending_write)) {
56 list_del_init(&llap->llap_pending_write);
59 spin_unlock(&lli->lli_lock);
63 /* Queue DONE_WRITING if
64 * - done writing is allowed;
65 * - inode has no no dirty pages; */
66 void ll_queue_done_writing(struct inode *inode, unsigned long flags)
68 struct ll_inode_info *lli = ll_i2info(inode);
70 spin_lock(&lli->lli_lock);
71 lli->lli_flags |= flags;
73 if ((lli->lli_flags & LLIF_DONE_WRITING) &&
74 list_empty(&lli->lli_pending_write_llaps)) {
75 struct ll_close_queue *lcq = ll_i2sbi(inode)->ll_lcq;
77 /* DONE_WRITING is allowed and inode has no dirty page. */
78 spin_lock(&lcq->lcq_lock);
81 * XXX: Seems sometimes it is possible to try to add inode more
82 * than once to close thread queue. Not sure if that is correct
83 * from caller POV, but it looks to me logically to check this
84 * here and just do nothing if inode is already on the queue.
85 * Hope Vitaly will correct me if I'm wrong. --umka
87 if (list_empty(&lli->lli_close_list)) {
88 CDEBUG(D_INODE, "adding inode %lu/%u to close list\n",
89 inode->i_ino, inode->i_generation);
90 list_add_tail(&lli->lli_close_list, &lcq->lcq_head);
92 CWARN("Inode %lu/%u is already queued for done writing!\n",
93 inode->i_ino, inode->i_generation);
95 wake_up(&lcq->lcq_waitq);
96 spin_unlock(&lcq->lcq_lock);
98 spin_unlock(&lli->lli_lock);
101 /* Close epoch and send Size-on-MDS attribute update if possible.
102 * Call this under @lli->lli_lock spinlock. */
103 void ll_epoch_close(struct inode *inode, struct md_op_data *op_data,
104 struct obd_client_handle **och, unsigned long flags)
106 struct ll_inode_info *lli = ll_i2info(inode);
109 spin_lock(&lli->lli_lock);
110 if (!(list_empty(&lli->lli_pending_write_llaps)) &&
111 !(lli->lli_flags & LLIF_EPOCH_PENDING)) {
112 LASSERT(*och != NULL);
113 LASSERT(lli->lli_pending_och == NULL);
114 /* Inode is dirty and there is no pending write done request
115 * yet, DONE_WRITE is to be sent later. */
116 lli->lli_flags |= LLIF_EPOCH_PENDING;
117 lli->lli_pending_och = *och;
118 spin_unlock(&lli->lli_lock);
120 inode = igrab(inode);
125 CDEBUG(D_INODE, "Epoch "LPU64" closed on "DFID"\n",
126 op_data->ioepoch, PFID(&lli->lli_fid));
127 op_data->flags |= MF_EPOCH_CLOSE;
129 if (flags & LLIF_DONE_WRITING) {
130 LASSERT(lli->lli_flags & LLIF_SOM_DIRTY);
131 *och = lli->lli_pending_och;
132 lli->lli_pending_och = NULL;
133 lli->lli_flags &= ~(LLIF_DONE_WRITING | LLIF_EPOCH_PENDING |
136 /* Pack Size-on-MDS inode attributes only if they has changed */
137 if (!(lli->lli_flags & LLIF_SOM_DIRTY)) {
138 spin_unlock(&lli->lli_lock);
142 /* There is already 1 pending DONE_WRITE, do not create another
143 * one -- close epoch with no attribute change. */
144 if (lli->lli_flags & LLIF_EPOCH_PENDING) {
145 spin_unlock(&lli->lli_lock);
150 spin_unlock(&lli->lli_lock);
151 op_data->flags |= MF_SOM_CHANGE;
153 /* Check if Size-on-MDS attributes are valid. */
154 LASSERT(!(lli->lli_flags & LLIF_MDS_SIZE_LOCK));
155 if (!ll_local_size(inode)) {
156 /* Send Size-on-MDS Attributes if valid. */
157 op_data->attr.ia_valid |= ATTR_MTIME_SET | ATTR_CTIME_SET |
158 ATTR_SIZE | ATTR_BLOCKS;
165 int ll_sizeonmds_update(struct inode *inode, struct lustre_handle *fh)
167 struct ll_inode_info *lli = ll_i2info(inode);
168 struct md_op_data *op_data;
173 LASSERT(!(lli->lli_flags & LLIF_MDS_SIZE_LOCK));
176 OBD_ALLOC_PTR(op_data);
177 if (!oa || !op_data) {
178 CERROR("can't allocate memory for Size-on-MDS update.\n");
181 rc = ll_inode_getattr(inode, oa);
183 CERROR("inode_getattr failed (%d): unable to send a "
184 "Size-on-MDS attribute update for inode %lu/%u\n",
185 rc, inode->i_ino, inode->i_generation);
188 CDEBUG(D_INODE, "Size-on-MDS update on "DFID"\n", PFID(&lli->lli_fid));
190 md_from_obdo(op_data, oa, oa->o_valid);
191 memcpy(&op_data->handle, fh, sizeof(*fh));
193 op_data->ioepoch = lli->lli_ioepoch;
194 op_data->flags |= MF_SOM_CHANGE;
196 rc = ll_md_setattr(inode, op_data);
202 ll_finish_md_op_data(op_data);
206 /* Send a DONE_WRITING rpc, pack Size-on-MDS attributes into it, if possible */
207 static void ll_done_writing(struct inode *inode)
209 struct obd_client_handle *och = NULL;
210 struct md_op_data *op_data;
214 OBD_ALLOC_PTR(op_data);
215 if (op_data == NULL) {
216 CERROR("can't allocate op_data\n");
221 ll_epoch_close(inode, op_data, &och, LLIF_DONE_WRITING);
222 ll_pack_inode2opdata(inode, op_data, &och->och_fh);
224 rc = md_done_writing(ll_i2sbi(inode)->ll_md_exp, op_data, och);
225 ll_finish_md_op_data(op_data);
227 /* MDS has instructed us to obtain Size-on-MDS attribute from
228 * OSTs and send setattr to back to MDS. */
229 rc = ll_sizeonmds_update(inode, &och->och_fh);
231 CERROR("inode %lu mdc done_writing failed: rc = %d\n",
238 static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
240 struct ll_inode_info *lli = NULL;
242 spin_lock(&lcq->lcq_lock);
244 if (!list_empty(&lcq->lcq_head)) {
245 lli = list_entry(lcq->lcq_head.next, struct ll_inode_info,
247 list_del_init(&lli->lli_close_list);
248 } else if (atomic_read(&lcq->lcq_stop))
249 lli = ERR_PTR(-EALREADY);
251 spin_unlock(&lcq->lcq_lock);
255 static int ll_close_thread(void *arg)
257 struct ll_close_queue *lcq = arg;
261 char name[CFS_CURPROC_COMM_MAX];
262 snprintf(name, sizeof(name) - 1, "ll_close");
266 complete(&lcq->lcq_comp);
269 struct l_wait_info lwi = { 0 };
270 struct ll_inode_info *lli;
273 l_wait_event_exclusive(lcq->lcq_waitq,
274 (lli = ll_close_next_lli(lcq)) != NULL,
279 inode = ll_info2i(lli);
280 CDEBUG(D_INFO, "done_writting for inode %lu/%u\n",
281 inode->i_ino, inode->i_generation);
282 ll_done_writing(inode);
286 CDEBUG(D_INFO, "ll_close exiting\n");
287 complete(&lcq->lcq_comp);
291 int ll_close_thread_start(struct ll_close_queue **lcq_ret)
293 struct ll_close_queue *lcq;
296 OBD_ALLOC(lcq, sizeof(*lcq));
300 spin_lock_init(&lcq->lcq_lock);
301 INIT_LIST_HEAD(&lcq->lcq_head);
302 init_waitqueue_head(&lcq->lcq_waitq);
303 init_completion(&lcq->lcq_comp);
305 pid = kernel_thread(ll_close_thread, lcq, 0);
307 OBD_FREE(lcq, sizeof(*lcq));
311 wait_for_completion(&lcq->lcq_comp);
316 void ll_close_thread_shutdown(struct ll_close_queue *lcq)
318 init_completion(&lcq->lcq_comp);
319 atomic_inc(&lcq->lcq_stop);
320 wake_up(&lcq->lcq_waitq);
321 wait_for_completion(&lcq->lcq_comp);
322 OBD_FREE(lcq, sizeof(*lcq));