1 /*
2 * GPL HEADER START
3 *
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
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.
9 *
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).
15 *
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
19 *
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
22 * have any questions.
23 *
24 * GPL HEADER END
25 */
26 /*
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
29 *
30 * Copyright (c) 2011, 2012, Intel Corporation.
31 */
32 /*
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
35 *
36 * lustre/llite/llite_close.c
37 *
38 * Lustre Lite routines to issue a secondary close after writeback
39 */
40
41 #include <linux/module.h>
42
43 #define DEBUG_SUBSYSTEM S_LLITE
44
45 #include "../include/lustre_lite.h"
46 #include "llite_internal.h"
47
48 /** records that a write is in flight */
vvp_write_pending(struct ccc_object * club,struct ccc_page * page)49 void vvp_write_pending(struct ccc_object *club, struct ccc_page *page)
50 {
51 struct ll_inode_info *lli = ll_i2info(club->cob_inode);
52
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);
59 }
60
61 /** records that a write has completed */
vvp_write_complete(struct ccc_object * club,struct ccc_page * page)62 void vvp_write_complete(struct ccc_object *club, struct ccc_page *page)
63 {
64 struct ll_inode_info *lli = ll_i2info(club->cob_inode);
65 int rc = 0;
66
67 spin_lock(&lli->lli_lock);
68 if (page != NULL && !list_empty(&page->cpg_pending_linkage)) {
69 list_del_init(&page->cpg_pending_linkage);
70 rc = 1;
71 }
72 spin_unlock(&lli->lli_lock);
73 if (rc)
74 ll_queue_done_writing(club->cob_inode, 0);
75 }
76
77 /** Queues DONE_WRITING if
78 * - done writing is allowed;
79 * - inode has no no dirty pages; */
ll_queue_done_writing(struct inode * inode,unsigned long flags)80 void ll_queue_done_writing(struct inode *inode, unsigned long flags)
81 {
82 struct ll_inode_info *lli = ll_i2info(inode);
83 struct ccc_object *club = cl2ccc(ll_i2info(inode)->lli_clob);
84
85 spin_lock(&lli->lli_lock);
86 lli->lli_flags |= flags;
87
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;
91
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,
95 lli->lli_flags);
96 /* DONE_WRITING is allowed and inode has no dirty page. */
97 spin_lock(&lcq->lcq_lock);
98
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);
103
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
109 * it. */
110 lli->lli_flags &= ~LLIF_DONE_WRITING;
111
112 wake_up(&lcq->lcq_waitq);
113 spin_unlock(&lcq->lcq_lock);
114 }
115 spin_unlock(&lli->lli_lock);
116 }
117
118 /** Pack SOM attributes info @opdata for CLOSE, DONE_WRITING rpc. */
ll_done_writing_attr(struct inode * inode,struct md_op_data * op_data)119 void ll_done_writing_attr(struct inode *inode, struct md_op_data *op_data)
120 {
121 struct ll_inode_info *lli = ll_i2info(inode);
122
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,
128 lli->lli_flags);
129
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;
134 }
135 }
136
137 /** Closes ioepoch and packs Size-on-MDS attribute if needed into @op_data. */
ll_ioepoch_close(struct inode * inode,struct md_op_data * op_data,struct obd_client_handle ** och,unsigned long flags)138 void ll_ioepoch_close(struct inode *inode, struct md_op_data *op_data,
139 struct obd_client_handle **och, unsigned long flags)
140 {
141 struct ll_inode_info *lli = ll_i2info(inode);
142 struct ccc_object *club = cl2ccc(ll_i2info(inode)->lli_clob);
143
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);
154
155 inode = igrab(inode);
156 LASSERT(inode);
157 goto out;
158 }
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);
166
167 inode = igrab(inode);
168 LASSERT(inode);
169 goto out;
170 }
171 }
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;
175
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;
182 } else {
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);
186 goto out;
187 }
188
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);
193 goto out;
194 }
195 }
196
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);
201
202 out:
203 return;
204 }
205
206 /**
207 * Cliens updates SOM attributes on MDS (including llog cookies):
208 * obd_getattr with no lock and md_setattr.
209 */
ll_som_update(struct inode * inode,struct md_op_data * op_data)210 int ll_som_update(struct inode *inode, struct md_op_data *op_data)
211 {
212 struct ll_inode_info *lli = ll_i2info(inode);
213 struct ptlrpc_request *request = NULL;
214 __u32 old_flags;
215 struct obdo *oa;
216 int rc;
217
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,
222 lli->lli_flags);
223
224 oa = kmem_cache_alloc(obdo_cachep, GFP_NOFS | __GFP_ZERO);
225 if (!oa) {
226 CERROR("can't allocate memory for Size-on-MDS update.\n");
227 return -ENOMEM;
228 }
229
230 old_flags = op_data->op_flags;
231 op_data->op_flags = MF_SOM_CHANGE;
232
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);
237 if (rc) {
238 oa->o_valid = 0;
239 if (rc != -ENOENT)
240 CERROR("inode_getattr failed (%d): unable to send a Size-on-MDS attribute update for inode %lu/%u\n",
241 rc, inode->i_ino,
242 inode->i_generation);
243 } else {
244 CDEBUG(D_INODE, "Size-on-MDS update on "DFID"\n",
245 PFID(&lli->lli_fid));
246 }
247 /* Install attributes into op_data. */
248 md_from_obdo(op_data, oa, oa->o_valid);
249 }
250
251 rc = md_setattr(ll_i2sbi(inode)->ll_md_exp, op_data,
252 NULL, 0, NULL, 0, &request, NULL);
253 ptlrpc_req_finished(request);
254
255 kmem_cache_free(obdo_cachep, oa);
256 return rc;
257 }
258
259 /**
260 * Closes the ioepoch and packs all the attributes into @op_data for
261 * DONE_WRITING rpc.
262 */
ll_prepare_done_writing(struct inode * inode,struct md_op_data * op_data,struct obd_client_handle ** och)263 static void ll_prepare_done_writing(struct inode *inode,
264 struct md_op_data *op_data,
265 struct obd_client_handle **och)
266 {
267 ll_ioepoch_close(inode, op_data, och, LLIF_DONE_WRITING);
268 /* If there is no @och, we do not do D_W yet. */
269 if (*och == NULL)
270 return;
271
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);
275 }
276
277 /** Send a DONE_WRITING rpc. */
ll_done_writing(struct inode * inode)278 static void ll_done_writing(struct inode *inode)
279 {
280 struct obd_client_handle *och = NULL;
281 struct md_op_data *op_data;
282 int rc;
283
284 LASSERT(exp_connect_som(ll_i2mdexp(inode)));
285
286 op_data = kzalloc(sizeof(*op_data), GFP_NOFS);
287 if (!op_data)
288 return;
289
290 ll_prepare_done_writing(inode, op_data, &och);
291 /* If there is no @och, we do not do D_W yet. */
292 if (och == NULL)
293 goto out;
294
295 rc = md_done_writing(ll_i2sbi(inode)->ll_md_exp, op_data, NULL);
296 if (rc == -EAGAIN)
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);
300 else if (rc)
301 CERROR("inode %lu mdc done_writing failed: rc = %d\n",
302 inode->i_ino, rc);
303 out:
304 ll_finish_md_op_data(op_data);
305 if (och) {
306 md_clear_open_replay_data(ll_i2sbi(inode)->ll_md_exp, och);
307 kfree(och);
308 }
309 }
310
ll_close_next_lli(struct ll_close_queue * lcq)311 static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
312 {
313 struct ll_inode_info *lli = NULL;
314
315 spin_lock(&lcq->lcq_lock);
316
317 if (!list_empty(&lcq->lcq_head)) {
318 lli = list_entry(lcq->lcq_head.next, struct ll_inode_info,
319 lli_close_list);
320 list_del_init(&lli->lli_close_list);
321 } else if (atomic_read(&lcq->lcq_stop))
322 lli = ERR_PTR(-EALREADY);
323
324 spin_unlock(&lcq->lcq_lock);
325 return lli;
326 }
327
ll_close_thread(void * arg)328 static int ll_close_thread(void *arg)
329 {
330 struct ll_close_queue *lcq = arg;
331
332 complete(&lcq->lcq_comp);
333
334 while (1) {
335 struct l_wait_info lwi = { 0 };
336 struct ll_inode_info *lli;
337 struct inode *inode;
338
339 l_wait_event_exclusive(lcq->lcq_waitq,
340 (lli = ll_close_next_lli(lcq)) != NULL,
341 &lwi);
342 if (IS_ERR(lli))
343 break;
344
345 inode = ll_info2i(lli);
346 CDEBUG(D_INFO, "done_writing for inode %lu/%u\n",
347 inode->i_ino, inode->i_generation);
348 ll_done_writing(inode);
349 iput(inode);
350 }
351
352 CDEBUG(D_INFO, "ll_close exiting\n");
353 complete(&lcq->lcq_comp);
354 return 0;
355 }
356
ll_close_thread_start(struct ll_close_queue ** lcq_ret)357 int ll_close_thread_start(struct ll_close_queue **lcq_ret)
358 {
359 struct ll_close_queue *lcq;
360 struct task_struct *task;
361
362 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CLOSE_THREAD))
363 return -EINTR;
364
365 lcq = kzalloc(sizeof(*lcq), GFP_NOFS);
366 if (!lcq)
367 return -ENOMEM;
368
369 spin_lock_init(&lcq->lcq_lock);
370 INIT_LIST_HEAD(&lcq->lcq_head);
371 init_waitqueue_head(&lcq->lcq_waitq);
372 init_completion(&lcq->lcq_comp);
373
374 task = kthread_run(ll_close_thread, lcq, "ll_close");
375 if (IS_ERR(task)) {
376 kfree(lcq);
377 return PTR_ERR(task);
378 }
379
380 wait_for_completion(&lcq->lcq_comp);
381 *lcq_ret = lcq;
382 return 0;
383 }
384
ll_close_thread_shutdown(struct ll_close_queue * lcq)385 void ll_close_thread_shutdown(struct ll_close_queue *lcq)
386 {
387 init_completion(&lcq->lcq_comp);
388 atomic_inc(&lcq->lcq_stop);
389 wake_up(&lcq->lcq_waitq);
390 wait_for_completion(&lcq->lcq_comp);
391 kfree(lcq);
392 }
393