Loading...
1/*
2 * Copyright (c) 2002, 2007 Red Hat, Inc. All rights reserved.
3 *
4 * This software may be freely redistributed under the terms of the
5 * GNU General Public License.
6 *
7 * You should have received a copy of the GNU General Public License
8 * along with this program; if not, write to the Free Software
9 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
10 *
11 * Authors: David Woodhouse <dwmw2@infradead.org>
12 * David Howells <dhowells@redhat.com>
13 *
14 */
15
16#include <linux/kernel.h>
17#include <linux/module.h>
18#include <linux/init.h>
19#include <linux/circ_buf.h>
20#include <linux/sched.h>
21#include "internal.h"
22
23/*
24 * Handle invalidation of an mmap'd file. We invalidate all the PTEs referring
25 * to the pages in this file's pagecache, forcing the kernel to go through
26 * ->fault() or ->page_mkwrite() - at which point we can handle invalidation
27 * more fully.
28 */
29void afs_invalidate_mmap_work(struct work_struct *work)
30{
31 struct afs_vnode *vnode = container_of(work, struct afs_vnode, cb_work);
32
33 unmap_mapping_pages(vnode->netfs.inode.i_mapping, 0, 0, false);
34}
35
36static void afs_volume_init_callback(struct afs_volume *volume)
37{
38 struct afs_vnode *vnode;
39
40 down_read(&volume->open_mmaps_lock);
41
42 list_for_each_entry(vnode, &volume->open_mmaps, cb_mmap_link) {
43 if (vnode->cb_v_check != atomic_read(&volume->cb_v_break)) {
44 atomic64_set(&vnode->cb_expires_at, AFS_NO_CB_PROMISE);
45 queue_work(system_unbound_wq, &vnode->cb_work);
46 }
47 }
48
49 up_read(&volume->open_mmaps_lock);
50}
51
52/*
53 * Allow the fileserver to request callback state (re-)initialisation.
54 * Unfortunately, UUIDs are not guaranteed unique.
55 */
56void afs_init_callback_state(struct afs_server *server)
57{
58 struct afs_server_entry *se;
59
60 down_read(&server->cell->vs_lock);
61
62 list_for_each_entry(se, &server->volumes, slink) {
63 se->cb_expires_at = AFS_NO_CB_PROMISE;
64 se->volume->cb_expires_at = AFS_NO_CB_PROMISE;
65 trace_afs_cb_v_break(se->volume->vid, atomic_read(&se->volume->cb_v_break),
66 afs_cb_break_for_s_reinit);
67 if (!list_empty(&se->volume->open_mmaps))
68 afs_volume_init_callback(se->volume);
69 }
70
71 up_read(&server->cell->vs_lock);
72}
73
74/*
75 * actually break a callback
76 */
77void __afs_break_callback(struct afs_vnode *vnode, enum afs_cb_break_reason reason)
78{
79 _enter("");
80
81 clear_bit(AFS_VNODE_NEW_CONTENT, &vnode->flags);
82 if (atomic64_xchg(&vnode->cb_expires_at, AFS_NO_CB_PROMISE) != AFS_NO_CB_PROMISE) {
83 vnode->cb_break++;
84 vnode->cb_v_check = atomic_read(&vnode->volume->cb_v_break);
85 afs_clear_permits(vnode);
86
87 if (vnode->lock_state == AFS_VNODE_LOCK_WAITING_FOR_CB)
88 afs_lock_may_be_available(vnode);
89
90 if (reason != afs_cb_break_for_deleted &&
91 vnode->status.type == AFS_FTYPE_FILE &&
92 atomic_read(&vnode->cb_nr_mmap))
93 queue_work(system_unbound_wq, &vnode->cb_work);
94
95 trace_afs_cb_break(&vnode->fid, vnode->cb_break, reason, true);
96 } else {
97 trace_afs_cb_break(&vnode->fid, vnode->cb_break, reason, false);
98 }
99}
100
101void afs_break_callback(struct afs_vnode *vnode, enum afs_cb_break_reason reason)
102{
103 write_seqlock(&vnode->cb_lock);
104 __afs_break_callback(vnode, reason);
105 write_sequnlock(&vnode->cb_lock);
106}
107
108/*
109 * Look up a volume by volume ID under RCU conditions.
110 */
111static struct afs_volume *afs_lookup_volume_rcu(struct afs_cell *cell,
112 afs_volid_t vid)
113{
114 struct afs_volume *volume = NULL;
115 struct rb_node *p;
116 int seq = 1;
117
118 for (;;) {
119 /* Unfortunately, rbtree walking doesn't give reliable results
120 * under just the RCU read lock, so we have to check for
121 * changes.
122 */
123 seq++; /* 2 on the 1st/lockless path, otherwise odd */
124 read_seqbegin_or_lock(&cell->volume_lock, &seq);
125
126 p = rcu_dereference_raw(cell->volumes.rb_node);
127 while (p) {
128 volume = rb_entry(p, struct afs_volume, cell_node);
129
130 if (volume->vid < vid)
131 p = rcu_dereference_raw(p->rb_left);
132 else if (volume->vid > vid)
133 p = rcu_dereference_raw(p->rb_right);
134 else
135 break;
136 volume = NULL;
137 }
138
139 if (volume && afs_try_get_volume(volume, afs_volume_trace_get_callback))
140 break;
141 if (!need_seqretry(&cell->volume_lock, seq))
142 break;
143 seq |= 1; /* Want a lock next time */
144 }
145
146 done_seqretry(&cell->volume_lock, seq);
147 return volume;
148}
149
150/*
151 * Allow the fileserver to break callbacks at the volume-level. This is
152 * typically done when, for example, a R/W volume is snapshotted to a R/O
153 * volume (the only way to change an R/O volume). It may also, however, happen
154 * when a volserver takes control of a volume (offlining it, moving it, etc.).
155 *
156 * Every file in that volume will need to be reevaluated.
157 */
158static void afs_break_volume_callback(struct afs_server *server,
159 struct afs_volume *volume)
160 __releases(RCU)
161{
162 struct afs_server_list *slist = rcu_dereference(volume->servers);
163 unsigned int i, cb_v_break;
164
165 write_lock(&volume->cb_v_break_lock);
166
167 for (i = 0; i < slist->nr_servers; i++)
168 if (slist->servers[i].server == server)
169 slist->servers[i].cb_expires_at = AFS_NO_CB_PROMISE;
170 volume->cb_expires_at = AFS_NO_CB_PROMISE;
171
172 cb_v_break = atomic_inc_return_release(&volume->cb_v_break);
173 trace_afs_cb_v_break(volume->vid, cb_v_break, afs_cb_break_for_volume_callback);
174
175 write_unlock(&volume->cb_v_break_lock);
176 rcu_read_unlock();
177
178 if (!list_empty(&volume->open_mmaps))
179 afs_volume_init_callback(volume);
180}
181
182/*
183 * allow the fileserver to explicitly break one callback
184 * - happens when
185 * - the backing file is changed
186 * - a lock is released
187 */
188static void afs_break_one_callback(struct afs_server *server,
189 struct afs_volume *volume,
190 struct afs_fid *fid)
191{
192 struct super_block *sb;
193 struct afs_vnode *vnode;
194 struct inode *inode;
195
196 /* See if we can find a matching inode - even an I_NEW inode needs to
197 * be marked as it can have its callback broken before we finish
198 * setting up the local inode.
199 */
200 sb = rcu_dereference(volume->sb);
201 if (!sb)
202 return;
203
204 inode = find_inode_rcu(sb, fid->vnode, afs_ilookup5_test_by_fid, fid);
205 if (inode) {
206 vnode = AFS_FS_I(inode);
207 afs_break_callback(vnode, afs_cb_break_for_callback);
208 } else {
209 trace_afs_cb_miss(fid, afs_cb_break_for_callback);
210 }
211}
212
213static void afs_break_some_callbacks(struct afs_server *server,
214 struct afs_callback_break *cbb,
215 size_t *_count)
216{
217 struct afs_callback_break *residue = cbb;
218 struct afs_volume *volume;
219 afs_volid_t vid = cbb->fid.vid;
220 size_t i;
221
222 rcu_read_lock();
223 volume = afs_lookup_volume_rcu(server->cell, vid);
224 if (cbb->fid.vnode == 0 && cbb->fid.unique == 0) {
225 afs_break_volume_callback(server, volume);
226 *_count -= 1;
227 if (*_count)
228 memmove(cbb, cbb + 1, sizeof(*cbb) * *_count);
229 } else {
230 /* TODO: Find all matching volumes if we couldn't match the server and
231 * break them anyway.
232 */
233
234 for (i = *_count; i > 0; cbb++, i--) {
235 if (cbb->fid.vid == vid) {
236 _debug("- Fid { vl=%08llx n=%llu u=%u }",
237 cbb->fid.vid,
238 cbb->fid.vnode,
239 cbb->fid.unique);
240 --*_count;
241 if (volume)
242 afs_break_one_callback(server, volume, &cbb->fid);
243 } else {
244 *residue++ = *cbb;
245 }
246 }
247 rcu_read_unlock();
248 }
249
250 afs_put_volume(volume, afs_volume_trace_put_callback);
251}
252
253/*
254 * allow the fileserver to break callback promises
255 */
256void afs_break_callbacks(struct afs_server *server, size_t count,
257 struct afs_callback_break *callbacks)
258{
259 _enter("%p,%zu,", server, count);
260
261 ASSERT(server != NULL);
262
263 while (count > 0)
264 afs_break_some_callbacks(server, callbacks, &count);
265}
1/*
2 * Copyright (c) 2002, 2007 Red Hat, Inc. All rights reserved.
3 *
4 * This software may be freely redistributed under the terms of the
5 * GNU General Public License.
6 *
7 * You should have received a copy of the GNU General Public License
8 * along with this program; if not, write to the Free Software
9 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
10 *
11 * Authors: David Woodhouse <dwmw2@infradead.org>
12 * David Howells <dhowells@redhat.com>
13 *
14 */
15
16#include <linux/kernel.h>
17#include <linux/module.h>
18#include <linux/init.h>
19#include <linux/circ_buf.h>
20#include <linux/sched.h>
21#include "internal.h"
22
23#if 0
24unsigned afs_vnode_update_timeout = 10;
25#endif /* 0 */
26
27#define afs_breakring_space(server) \
28 CIRC_SPACE((server)->cb_break_head, (server)->cb_break_tail, \
29 ARRAY_SIZE((server)->cb_break))
30
31//static void afs_callback_updater(struct work_struct *);
32
33static struct workqueue_struct *afs_callback_update_worker;
34
35/*
36 * allow the fileserver to request callback state (re-)initialisation
37 */
38void afs_init_callback_state(struct afs_server *server)
39{
40 struct afs_vnode *vnode;
41
42 _enter("{%p}", server);
43
44 spin_lock(&server->cb_lock);
45
46 /* kill all the promises on record from this server */
47 while (!RB_EMPTY_ROOT(&server->cb_promises)) {
48 vnode = rb_entry(server->cb_promises.rb_node,
49 struct afs_vnode, cb_promise);
50 _debug("UNPROMISE { vid=%x:%u uq=%u}",
51 vnode->fid.vid, vnode->fid.vnode, vnode->fid.unique);
52 rb_erase(&vnode->cb_promise, &server->cb_promises);
53 vnode->cb_promised = false;
54 }
55
56 spin_unlock(&server->cb_lock);
57 _leave("");
58}
59
60/*
61 * handle the data invalidation side of a callback being broken
62 */
63void afs_broken_callback_work(struct work_struct *work)
64{
65 struct afs_vnode *vnode =
66 container_of(work, struct afs_vnode, cb_broken_work);
67
68 _enter("");
69
70 if (test_bit(AFS_VNODE_DELETED, &vnode->flags))
71 return;
72
73 /* we're only interested in dealing with a broken callback on *this*
74 * vnode and only if no-one else has dealt with it yet */
75 if (!mutex_trylock(&vnode->validate_lock))
76 return; /* someone else is dealing with it */
77
78 if (test_bit(AFS_VNODE_CB_BROKEN, &vnode->flags)) {
79 if (S_ISDIR(vnode->vfs_inode.i_mode))
80 afs_clear_permits(vnode);
81
82 if (afs_vnode_fetch_status(vnode, NULL, NULL) < 0)
83 goto out;
84
85 if (test_bit(AFS_VNODE_DELETED, &vnode->flags))
86 goto out;
87
88 /* if the vnode's data version number changed then its contents
89 * are different */
90 if (test_and_clear_bit(AFS_VNODE_ZAP_DATA, &vnode->flags))
91 afs_zap_data(vnode);
92 }
93
94out:
95 mutex_unlock(&vnode->validate_lock);
96
97 /* avoid the potential race whereby the mutex_trylock() in this
98 * function happens again between the clear_bit() and the
99 * mutex_unlock() */
100 if (test_bit(AFS_VNODE_CB_BROKEN, &vnode->flags)) {
101 _debug("requeue");
102 queue_work(afs_callback_update_worker, &vnode->cb_broken_work);
103 }
104 _leave("");
105}
106
107/*
108 * actually break a callback
109 */
110static void afs_break_callback(struct afs_server *server,
111 struct afs_vnode *vnode)
112{
113 _enter("");
114
115 set_bit(AFS_VNODE_CB_BROKEN, &vnode->flags);
116
117 if (vnode->cb_promised) {
118 spin_lock(&vnode->lock);
119
120 _debug("break callback");
121
122 spin_lock(&server->cb_lock);
123 if (vnode->cb_promised) {
124 rb_erase(&vnode->cb_promise, &server->cb_promises);
125 vnode->cb_promised = false;
126 }
127 spin_unlock(&server->cb_lock);
128
129 queue_work(afs_callback_update_worker, &vnode->cb_broken_work);
130 if (list_empty(&vnode->granted_locks) &&
131 !list_empty(&vnode->pending_locks))
132 afs_lock_may_be_available(vnode);
133 spin_unlock(&vnode->lock);
134 }
135}
136
137/*
138 * allow the fileserver to explicitly break one callback
139 * - happens when
140 * - the backing file is changed
141 * - a lock is released
142 */
143static void afs_break_one_callback(struct afs_server *server,
144 struct afs_fid *fid)
145{
146 struct afs_vnode *vnode;
147 struct rb_node *p;
148
149 _debug("find");
150 spin_lock(&server->fs_lock);
151 p = server->fs_vnodes.rb_node;
152 while (p) {
153 vnode = rb_entry(p, struct afs_vnode, server_rb);
154 if (fid->vid < vnode->fid.vid)
155 p = p->rb_left;
156 else if (fid->vid > vnode->fid.vid)
157 p = p->rb_right;
158 else if (fid->vnode < vnode->fid.vnode)
159 p = p->rb_left;
160 else if (fid->vnode > vnode->fid.vnode)
161 p = p->rb_right;
162 else if (fid->unique < vnode->fid.unique)
163 p = p->rb_left;
164 else if (fid->unique > vnode->fid.unique)
165 p = p->rb_right;
166 else
167 goto found;
168 }
169
170 /* not found so we just ignore it (it may have moved to another
171 * server) */
172not_available:
173 _debug("not avail");
174 spin_unlock(&server->fs_lock);
175 _leave("");
176 return;
177
178found:
179 _debug("found");
180 ASSERTCMP(server, ==, vnode->server);
181
182 if (!igrab(AFS_VNODE_TO_I(vnode)))
183 goto not_available;
184 spin_unlock(&server->fs_lock);
185
186 afs_break_callback(server, vnode);
187 iput(&vnode->vfs_inode);
188 _leave("");
189}
190
191/*
192 * allow the fileserver to break callback promises
193 */
194void afs_break_callbacks(struct afs_server *server, size_t count,
195 struct afs_callback callbacks[])
196{
197 _enter("%p,%zu,", server, count);
198
199 ASSERT(server != NULL);
200 ASSERTCMP(count, <=, AFSCBMAX);
201
202 for (; count > 0; callbacks++, count--) {
203 _debug("- Fid { vl=%08x n=%u u=%u } CB { v=%u x=%u t=%u }",
204 callbacks->fid.vid,
205 callbacks->fid.vnode,
206 callbacks->fid.unique,
207 callbacks->version,
208 callbacks->expiry,
209 callbacks->type
210 );
211 afs_break_one_callback(server, &callbacks->fid);
212 }
213
214 _leave("");
215 return;
216}
217
218/*
219 * record the callback for breaking
220 * - the caller must hold server->cb_lock
221 */
222static void afs_do_give_up_callback(struct afs_server *server,
223 struct afs_vnode *vnode)
224{
225 struct afs_callback *cb;
226
227 _enter("%p,%p", server, vnode);
228
229 cb = &server->cb_break[server->cb_break_head];
230 cb->fid = vnode->fid;
231 cb->version = vnode->cb_version;
232 cb->expiry = vnode->cb_expiry;
233 cb->type = vnode->cb_type;
234 smp_wmb();
235 server->cb_break_head =
236 (server->cb_break_head + 1) &
237 (ARRAY_SIZE(server->cb_break) - 1);
238
239 /* defer the breaking of callbacks to try and collect as many as
240 * possible to ship in one operation */
241 switch (atomic_inc_return(&server->cb_break_n)) {
242 case 1 ... AFSCBMAX - 1:
243 queue_delayed_work(afs_callback_update_worker,
244 &server->cb_break_work, HZ * 2);
245 break;
246 case AFSCBMAX:
247 afs_flush_callback_breaks(server);
248 break;
249 default:
250 break;
251 }
252
253 ASSERT(server->cb_promises.rb_node != NULL);
254 rb_erase(&vnode->cb_promise, &server->cb_promises);
255 vnode->cb_promised = false;
256 _leave("");
257}
258
259/*
260 * discard the callback on a deleted item
261 */
262void afs_discard_callback_on_delete(struct afs_vnode *vnode)
263{
264 struct afs_server *server = vnode->server;
265
266 _enter("%d", vnode->cb_promised);
267
268 if (!vnode->cb_promised) {
269 _leave(" [not promised]");
270 return;
271 }
272
273 ASSERT(server != NULL);
274
275 spin_lock(&server->cb_lock);
276 if (vnode->cb_promised) {
277 ASSERT(server->cb_promises.rb_node != NULL);
278 rb_erase(&vnode->cb_promise, &server->cb_promises);
279 vnode->cb_promised = false;
280 }
281 spin_unlock(&server->cb_lock);
282 _leave("");
283}
284
285/*
286 * give up the callback registered for a vnode on the file server when the
287 * inode is being cleared
288 */
289void afs_give_up_callback(struct afs_vnode *vnode)
290{
291 struct afs_server *server = vnode->server;
292
293 DECLARE_WAITQUEUE(myself, current);
294
295 _enter("%d", vnode->cb_promised);
296
297 _debug("GIVE UP INODE %p", &vnode->vfs_inode);
298
299 if (!vnode->cb_promised) {
300 _leave(" [not promised]");
301 return;
302 }
303
304 ASSERT(server != NULL);
305
306 spin_lock(&server->cb_lock);
307 if (vnode->cb_promised && afs_breakring_space(server) == 0) {
308 add_wait_queue(&server->cb_break_waitq, &myself);
309 for (;;) {
310 set_current_state(TASK_UNINTERRUPTIBLE);
311 if (!vnode->cb_promised ||
312 afs_breakring_space(server) != 0)
313 break;
314 spin_unlock(&server->cb_lock);
315 schedule();
316 spin_lock(&server->cb_lock);
317 }
318 remove_wait_queue(&server->cb_break_waitq, &myself);
319 __set_current_state(TASK_RUNNING);
320 }
321
322 /* of course, it's always possible for the server to break this vnode's
323 * callback first... */
324 if (vnode->cb_promised)
325 afs_do_give_up_callback(server, vnode);
326
327 spin_unlock(&server->cb_lock);
328 _leave("");
329}
330
331/*
332 * dispatch a deferred give up callbacks operation
333 */
334void afs_dispatch_give_up_callbacks(struct work_struct *work)
335{
336 struct afs_server *server =
337 container_of(work, struct afs_server, cb_break_work.work);
338
339 _enter("");
340
341 /* tell the fileserver to discard the callback promises it has
342 * - in the event of ENOMEM or some other error, we just forget that we
343 * had callbacks entirely, and the server will call us later to break
344 * them
345 */
346 afs_fs_give_up_callbacks(server, &afs_async_call);
347}
348
349/*
350 * flush the outstanding callback breaks on a server
351 */
352void afs_flush_callback_breaks(struct afs_server *server)
353{
354 cancel_delayed_work(&server->cb_break_work);
355 queue_delayed_work(afs_callback_update_worker,
356 &server->cb_break_work, 0);
357}
358
359#if 0
360/*
361 * update a bunch of callbacks
362 */
363static void afs_callback_updater(struct work_struct *work)
364{
365 struct afs_server *server;
366 struct afs_vnode *vnode, *xvnode;
367 time_t now;
368 long timeout;
369 int ret;
370
371 server = container_of(work, struct afs_server, updater);
372
373 _enter("");
374
375 now = get_seconds();
376
377 /* find the first vnode to update */
378 spin_lock(&server->cb_lock);
379 for (;;) {
380 if (RB_EMPTY_ROOT(&server->cb_promises)) {
381 spin_unlock(&server->cb_lock);
382 _leave(" [nothing]");
383 return;
384 }
385
386 vnode = rb_entry(rb_first(&server->cb_promises),
387 struct afs_vnode, cb_promise);
388 if (atomic_read(&vnode->usage) > 0)
389 break;
390 rb_erase(&vnode->cb_promise, &server->cb_promises);
391 vnode->cb_promised = false;
392 }
393
394 timeout = vnode->update_at - now;
395 if (timeout > 0) {
396 queue_delayed_work(afs_vnode_update_worker,
397 &afs_vnode_update, timeout * HZ);
398 spin_unlock(&server->cb_lock);
399 _leave(" [nothing]");
400 return;
401 }
402
403 list_del_init(&vnode->update);
404 atomic_inc(&vnode->usage);
405 spin_unlock(&server->cb_lock);
406
407 /* we can now perform the update */
408 _debug("update %s", vnode->vldb.name);
409 vnode->state = AFS_VL_UPDATING;
410 vnode->upd_rej_cnt = 0;
411 vnode->upd_busy_cnt = 0;
412
413 ret = afs_vnode_update_record(vl, &vldb);
414 switch (ret) {
415 case 0:
416 afs_vnode_apply_update(vl, &vldb);
417 vnode->state = AFS_VL_UPDATING;
418 break;
419 case -ENOMEDIUM:
420 vnode->state = AFS_VL_VOLUME_DELETED;
421 break;
422 default:
423 vnode->state = AFS_VL_UNCERTAIN;
424 break;
425 }
426
427 /* and then reschedule */
428 _debug("reschedule");
429 vnode->update_at = get_seconds() + afs_vnode_update_timeout;
430
431 spin_lock(&server->cb_lock);
432
433 if (!list_empty(&server->cb_promises)) {
434 /* next update in 10 minutes, but wait at least 1 second more
435 * than the newest record already queued so that we don't spam
436 * the VL server suddenly with lots of requests
437 */
438 xvnode = list_entry(server->cb_promises.prev,
439 struct afs_vnode, update);
440 if (vnode->update_at <= xvnode->update_at)
441 vnode->update_at = xvnode->update_at + 1;
442 xvnode = list_entry(server->cb_promises.next,
443 struct afs_vnode, update);
444 timeout = xvnode->update_at - now;
445 if (timeout < 0)
446 timeout = 0;
447 } else {
448 timeout = afs_vnode_update_timeout;
449 }
450
451 list_add_tail(&vnode->update, &server->cb_promises);
452
453 _debug("timeout %ld", timeout);
454 queue_delayed_work(afs_vnode_update_worker,
455 &afs_vnode_update, timeout * HZ);
456 spin_unlock(&server->cb_lock);
457 afs_put_vnode(vl);
458}
459#endif
460
461/*
462 * initialise the callback update process
463 */
464int __init afs_callback_update_init(void)
465{
466 afs_callback_update_worker =
467 create_singlethread_workqueue("kafs_callbackd");
468 return afs_callback_update_worker ? 0 : -ENOMEM;
469}
470
471/*
472 * shut down the callback update process
473 */
474void afs_callback_update_kill(void)
475{
476 destroy_workqueue(afs_callback_update_worker);
477}