1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* Handle fileserver selection and rotation.
3 *
4 * Copyright (C) 2017 Red Hat, Inc. All Rights Reserved.
5 * Written by David Howells (dhowells@redhat.com)
6 */
7
8 #include <linux/kernel.h>
9 #include <linux/slab.h>
10 #include <linux/fs.h>
11 #include <linux/sched.h>
12 #include <linux/delay.h>
13 #include <linux/sched/signal.h>
14 #include "internal.h"
15 #include "afs_fs.h"
16
17 /*
18 * Begin an operation on the fileserver.
19 *
20 * Fileserver operations are serialised on the server by vnode, so we serialise
21 * them here also using the io_lock.
22 */
afs_begin_vnode_operation(struct afs_fs_cursor * fc,struct afs_vnode * vnode,struct key * key,bool intr)23 bool afs_begin_vnode_operation(struct afs_fs_cursor *fc, struct afs_vnode *vnode,
24 struct key *key, bool intr)
25 {
26 memset(fc, 0, sizeof(*fc));
27 fc->vnode = vnode;
28 fc->key = key;
29 fc->ac.error = SHRT_MAX;
30 fc->error = -EDESTADDRREQ;
31
32 if (intr) {
33 fc->flags |= AFS_FS_CURSOR_INTR;
34 if (mutex_lock_interruptible(&vnode->io_lock) < 0) {
35 fc->error = -EINTR;
36 fc->flags |= AFS_FS_CURSOR_STOP;
37 return false;
38 }
39 } else {
40 mutex_lock(&vnode->io_lock);
41 }
42
43 if (vnode->lock_state != AFS_VNODE_LOCK_NONE)
44 fc->flags |= AFS_FS_CURSOR_CUR_ONLY;
45 return true;
46 }
47
48 /*
49 * Begin iteration through a server list, starting with the vnode's last used
50 * server if possible, or the last recorded good server if not.
51 */
afs_start_fs_iteration(struct afs_fs_cursor * fc,struct afs_vnode * vnode)52 static bool afs_start_fs_iteration(struct afs_fs_cursor *fc,
53 struct afs_vnode *vnode)
54 {
55 struct afs_cb_interest *cbi;
56 int i;
57
58 read_lock(&vnode->volume->servers_lock);
59 fc->server_list = afs_get_serverlist(vnode->volume->servers);
60 read_unlock(&vnode->volume->servers_lock);
61
62 fc->untried = (1UL << fc->server_list->nr_servers) - 1;
63 fc->index = READ_ONCE(fc->server_list->preferred);
64
65 cbi = rcu_dereference_protected(vnode->cb_interest,
66 lockdep_is_held(&vnode->io_lock));
67 if (cbi) {
68 /* See if the vnode's preferred record is still available */
69 for (i = 0; i < fc->server_list->nr_servers; i++) {
70 if (fc->server_list->servers[i].cb_interest == cbi) {
71 fc->index = i;
72 goto found_interest;
73 }
74 }
75
76 /* If we have a lock outstanding on a server that's no longer
77 * serving this vnode, then we can't switch to another server
78 * and have to return an error.
79 */
80 if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
81 fc->error = -ESTALE;
82 return false;
83 }
84
85 /* Note that the callback promise is effectively broken */
86 write_seqlock(&vnode->cb_lock);
87 ASSERTCMP(cbi, ==, rcu_access_pointer(vnode->cb_interest));
88 rcu_assign_pointer(vnode->cb_interest, NULL);
89 if (test_and_clear_bit(AFS_VNODE_CB_PROMISED, &vnode->flags))
90 vnode->cb_break++;
91 write_sequnlock(&vnode->cb_lock);
92
93 afs_put_cb_interest(afs_v2net(vnode), cbi);
94 cbi = NULL;
95 }
96
97 found_interest:
98 return true;
99 }
100
101 /*
102 * Post volume busy note.
103 */
afs_busy(struct afs_volume * volume,u32 abort_code)104 static void afs_busy(struct afs_volume *volume, u32 abort_code)
105 {
106 const char *m;
107
108 switch (abort_code) {
109 case VOFFLINE: m = "offline"; break;
110 case VRESTARTING: m = "restarting"; break;
111 case VSALVAGING: m = "being salvaged"; break;
112 default: m = "busy"; break;
113 }
114
115 pr_notice("kAFS: Volume %llu '%s' is %s\n", volume->vid, volume->name, m);
116 }
117
118 /*
119 * Sleep and retry the operation to the same fileserver.
120 */
afs_sleep_and_retry(struct afs_fs_cursor * fc)121 static bool afs_sleep_and_retry(struct afs_fs_cursor *fc)
122 {
123 if (fc->flags & AFS_FS_CURSOR_INTR) {
124 msleep_interruptible(1000);
125 if (signal_pending(current)) {
126 fc->error = -ERESTARTSYS;
127 return false;
128 }
129 } else {
130 msleep(1000);
131 }
132
133 return true;
134 }
135
136 /*
137 * Select the fileserver to use. May be called multiple times to rotate
138 * through the fileservers.
139 */
afs_select_fileserver(struct afs_fs_cursor * fc)140 bool afs_select_fileserver(struct afs_fs_cursor *fc)
141 {
142 struct afs_addr_list *alist;
143 struct afs_server *server;
144 struct afs_vnode *vnode = fc->vnode;
145 struct afs_error e;
146 u32 rtt;
147 int error = fc->ac.error, i;
148
149 _enter("%lx[%d],%lx[%d],%d,%d",
150 fc->untried, fc->index,
151 fc->ac.tried, fc->ac.index,
152 error, fc->ac.abort_code);
153
154 if (fc->flags & AFS_FS_CURSOR_STOP) {
155 _leave(" = f [stopped]");
156 return false;
157 }
158
159 fc->nr_iterations++;
160
161 /* Evaluate the result of the previous operation, if there was one. */
162 switch (error) {
163 case SHRT_MAX:
164 goto start;
165
166 case 0:
167 default:
168 /* Success or local failure. Stop. */
169 fc->error = error;
170 fc->flags |= AFS_FS_CURSOR_STOP;
171 _leave(" = f [okay/local %d]", error);
172 return false;
173
174 case -ECONNABORTED:
175 /* The far side rejected the operation on some grounds. This
176 * might involve the server being busy or the volume having been moved.
177 */
178 switch (fc->ac.abort_code) {
179 case VNOVOL:
180 /* This fileserver doesn't know about the volume.
181 * - May indicate that the VL is wrong - retry once and compare
182 * the results.
183 * - May indicate that the fileserver couldn't attach to the vol.
184 */
185 if (fc->flags & AFS_FS_CURSOR_VNOVOL) {
186 fc->error = -EREMOTEIO;
187 goto next_server;
188 }
189
190 write_lock(&vnode->volume->servers_lock);
191 fc->server_list->vnovol_mask |= 1 << fc->index;
192 write_unlock(&vnode->volume->servers_lock);
193
194 set_bit(AFS_VOLUME_NEEDS_UPDATE, &vnode->volume->flags);
195 error = afs_check_volume_status(vnode->volume, fc->key);
196 if (error < 0)
197 goto failed_set_error;
198
199 if (test_bit(AFS_VOLUME_DELETED, &vnode->volume->flags)) {
200 fc->error = -ENOMEDIUM;
201 goto failed;
202 }
203
204 /* If the server list didn't change, then assume that
205 * it's the fileserver having trouble.
206 */
207 if (vnode->volume->servers == fc->server_list) {
208 fc->error = -EREMOTEIO;
209 goto next_server;
210 }
211
212 /* Try again */
213 fc->flags |= AFS_FS_CURSOR_VNOVOL;
214 _leave(" = t [vnovol]");
215 return true;
216
217 case VSALVAGE: /* TODO: Should this return an error or iterate? */
218 case VVOLEXISTS:
219 case VNOSERVICE:
220 case VONLINE:
221 case VDISKFULL:
222 case VOVERQUOTA:
223 fc->error = afs_abort_to_error(fc->ac.abort_code);
224 goto next_server;
225
226 case VOFFLINE:
227 if (!test_and_set_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags)) {
228 afs_busy(vnode->volume, fc->ac.abort_code);
229 clear_bit(AFS_VOLUME_BUSY, &vnode->volume->flags);
230 }
231 if (fc->flags & AFS_FS_CURSOR_NO_VSLEEP) {
232 fc->error = -EADV;
233 goto failed;
234 }
235 if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
236 fc->error = -ESTALE;
237 goto failed;
238 }
239 goto busy;
240
241 case VSALVAGING:
242 case VRESTARTING:
243 case VBUSY:
244 /* Retry after going round all the servers unless we
245 * have a file lock we need to maintain.
246 */
247 if (fc->flags & AFS_FS_CURSOR_NO_VSLEEP) {
248 fc->error = -EBUSY;
249 goto failed;
250 }
251 if (!test_and_set_bit(AFS_VOLUME_BUSY, &vnode->volume->flags)) {
252 afs_busy(vnode->volume, fc->ac.abort_code);
253 clear_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags);
254 }
255 busy:
256 if (fc->flags & AFS_FS_CURSOR_CUR_ONLY) {
257 if (!afs_sleep_and_retry(fc))
258 goto failed;
259
260 /* Retry with same server & address */
261 _leave(" = t [vbusy]");
262 return true;
263 }
264
265 fc->flags |= AFS_FS_CURSOR_VBUSY;
266 goto next_server;
267
268 case VMOVED:
269 /* The volume migrated to another server. We consider
270 * consider all locks and callbacks broken and request
271 * an update from the VLDB.
272 *
273 * We also limit the number of VMOVED hops we will
274 * honour, just in case someone sets up a loop.
275 */
276 if (fc->flags & AFS_FS_CURSOR_VMOVED) {
277 fc->error = -EREMOTEIO;
278 goto failed;
279 }
280 fc->flags |= AFS_FS_CURSOR_VMOVED;
281
282 set_bit(AFS_VOLUME_WAIT, &vnode->volume->flags);
283 set_bit(AFS_VOLUME_NEEDS_UPDATE, &vnode->volume->flags);
284 error = afs_check_volume_status(vnode->volume, fc->key);
285 if (error < 0)
286 goto failed_set_error;
287
288 /* If the server list didn't change, then the VLDB is
289 * out of sync with the fileservers. This is hopefully
290 * a temporary condition, however, so we don't want to
291 * permanently block access to the file.
292 *
293 * TODO: Try other fileservers if we can.
294 *
295 * TODO: Retry a few times with sleeps.
296 */
297 if (vnode->volume->servers == fc->server_list) {
298 fc->error = -ENOMEDIUM;
299 goto failed;
300 }
301
302 goto restart_from_beginning;
303
304 default:
305 clear_bit(AFS_VOLUME_OFFLINE, &vnode->volume->flags);
306 clear_bit(AFS_VOLUME_BUSY, &vnode->volume->flags);
307 fc->error = afs_abort_to_error(fc->ac.abort_code);
308 goto failed;
309 }
310
311 case -ETIMEDOUT:
312 case -ETIME:
313 if (fc->error != -EDESTADDRREQ)
314 goto iterate_address;
315 /* Fall through */
316 case -ERFKILL:
317 case -EADDRNOTAVAIL:
318 case -ENETUNREACH:
319 case -EHOSTUNREACH:
320 case -EHOSTDOWN:
321 case -ECONNREFUSED:
322 _debug("no conn");
323 fc->error = error;
324 goto iterate_address;
325
326 case -ECONNRESET:
327 _debug("call reset");
328 fc->error = error;
329 goto failed;
330 }
331
332 restart_from_beginning:
333 _debug("restart");
334 afs_end_cursor(&fc->ac);
335 afs_put_cb_interest(afs_v2net(vnode), fc->cbi);
336 fc->cbi = NULL;
337 afs_put_serverlist(afs_v2net(vnode), fc->server_list);
338 fc->server_list = NULL;
339 start:
340 _debug("start");
341 /* See if we need to do an update of the volume record. Note that the
342 * volume may have moved or even have been deleted.
343 */
344 error = afs_check_volume_status(vnode->volume, fc->key);
345 if (error < 0)
346 goto failed_set_error;
347
348 if (!afs_start_fs_iteration(fc, vnode))
349 goto failed;
350
351 _debug("__ VOL %llx __", vnode->volume->vid);
352 error = afs_probe_fileservers(afs_v2net(vnode), fc->key, fc->server_list);
353 if (error < 0)
354 goto failed_set_error;
355
356 pick_server:
357 _debug("pick [%lx]", fc->untried);
358
359 error = afs_wait_for_fs_probes(fc->server_list, fc->untried);
360 if (error < 0)
361 goto failed_set_error;
362
363 /* Pick the untried server with the lowest RTT. If we have outstanding
364 * callbacks, we stick with the server we're already using if we can.
365 */
366 if (fc->cbi) {
367 _debug("cbi %u", fc->index);
368 if (test_bit(fc->index, &fc->untried))
369 goto selected_server;
370 afs_put_cb_interest(afs_v2net(vnode), fc->cbi);
371 fc->cbi = NULL;
372 _debug("nocbi");
373 }
374
375 fc->index = -1;
376 rtt = U32_MAX;
377 for (i = 0; i < fc->server_list->nr_servers; i++) {
378 struct afs_server *s = fc->server_list->servers[i].server;
379
380 if (!test_bit(i, &fc->untried) || !s->probe.responded)
381 continue;
382 if (s->probe.rtt < rtt) {
383 fc->index = i;
384 rtt = s->probe.rtt;
385 }
386 }
387
388 if (fc->index == -1)
389 goto no_more_servers;
390
391 selected_server:
392 _debug("use %d", fc->index);
393 __clear_bit(fc->index, &fc->untried);
394
395 /* We're starting on a different fileserver from the list. We need to
396 * check it, create a callback intercept, find its address list and
397 * probe its capabilities before we use it.
398 */
399 ASSERTCMP(fc->ac.alist, ==, NULL);
400 server = fc->server_list->servers[fc->index].server;
401
402 if (!afs_check_server_record(fc, server))
403 goto failed;
404
405 _debug("USING SERVER: %pU", &server->uuid);
406
407 /* Make sure we've got a callback interest record for this server. We
408 * have to link it in before we send the request as we can be sent a
409 * break request before we've finished decoding the reply and
410 * installing the vnode.
411 */
412 error = afs_register_server_cb_interest(vnode, fc->server_list,
413 fc->index);
414 if (error < 0)
415 goto failed_set_error;
416
417 fc->cbi = afs_get_cb_interest(
418 rcu_dereference_protected(vnode->cb_interest,
419 lockdep_is_held(&vnode->io_lock)));
420
421 read_lock(&server->fs_lock);
422 alist = rcu_dereference_protected(server->addresses,
423 lockdep_is_held(&server->fs_lock));
424 afs_get_addrlist(alist);
425 read_unlock(&server->fs_lock);
426
427 memset(&fc->ac, 0, sizeof(fc->ac));
428
429 if (!fc->ac.alist)
430 fc->ac.alist = alist;
431 else
432 afs_put_addrlist(alist);
433
434 fc->ac.index = -1;
435
436 iterate_address:
437 ASSERT(fc->ac.alist);
438 /* Iterate over the current server's address list to try and find an
439 * address on which it will respond to us.
440 */
441 if (!afs_iterate_addresses(&fc->ac))
442 goto next_server;
443
444 _debug("address [%u] %u/%u", fc->index, fc->ac.index, fc->ac.alist->nr_addrs);
445
446 _leave(" = t");
447 return true;
448
449 next_server:
450 _debug("next");
451 afs_end_cursor(&fc->ac);
452 goto pick_server;
453
454 no_more_servers:
455 /* That's all the servers poked to no good effect. Try again if some
456 * of them were busy.
457 */
458 if (fc->flags & AFS_FS_CURSOR_VBUSY)
459 goto restart_from_beginning;
460
461 e.error = -EDESTADDRREQ;
462 e.responded = false;
463 for (i = 0; i < fc->server_list->nr_servers; i++) {
464 struct afs_server *s = fc->server_list->servers[i].server;
465
466 afs_prioritise_error(&e, READ_ONCE(s->probe.error),
467 s->probe.abort_code);
468 }
469
470 error = e.error;
471
472 failed_set_error:
473 fc->error = error;
474 failed:
475 fc->flags |= AFS_FS_CURSOR_STOP;
476 afs_end_cursor(&fc->ac);
477 _leave(" = f [failed %d]", fc->error);
478 return false;
479 }
480
481 /*
482 * Select the same fileserver we used for a vnode before and only that
483 * fileserver. We use this when we have a lock on that file, which is backed
484 * only by the fileserver we obtained it from.
485 */
afs_select_current_fileserver(struct afs_fs_cursor * fc)486 bool afs_select_current_fileserver(struct afs_fs_cursor *fc)
487 {
488 struct afs_vnode *vnode = fc->vnode;
489 struct afs_cb_interest *cbi;
490 struct afs_addr_list *alist;
491 int error = fc->ac.error;
492
493 _enter("");
494
495 cbi = rcu_dereference_protected(vnode->cb_interest,
496 lockdep_is_held(&vnode->io_lock));
497
498 switch (error) {
499 case SHRT_MAX:
500 if (!cbi) {
501 fc->error = -ESTALE;
502 fc->flags |= AFS_FS_CURSOR_STOP;
503 return false;
504 }
505
506 fc->cbi = afs_get_cb_interest(cbi);
507
508 read_lock(&cbi->server->fs_lock);
509 alist = rcu_dereference_protected(cbi->server->addresses,
510 lockdep_is_held(&cbi->server->fs_lock));
511 afs_get_addrlist(alist);
512 read_unlock(&cbi->server->fs_lock);
513 if (!alist) {
514 fc->error = -ESTALE;
515 fc->flags |= AFS_FS_CURSOR_STOP;
516 return false;
517 }
518
519 memset(&fc->ac, 0, sizeof(fc->ac));
520 fc->ac.alist = alist;
521 fc->ac.index = -1;
522 goto iterate_address;
523
524 case 0:
525 default:
526 /* Success or local failure. Stop. */
527 fc->error = error;
528 fc->flags |= AFS_FS_CURSOR_STOP;
529 _leave(" = f [okay/local %d]", error);
530 return false;
531
532 case -ECONNABORTED:
533 fc->error = afs_abort_to_error(fc->ac.abort_code);
534 fc->flags |= AFS_FS_CURSOR_STOP;
535 _leave(" = f [abort]");
536 return false;
537
538 case -ERFKILL:
539 case -EADDRNOTAVAIL:
540 case -ENETUNREACH:
541 case -EHOSTUNREACH:
542 case -EHOSTDOWN:
543 case -ECONNREFUSED:
544 case -ETIMEDOUT:
545 case -ETIME:
546 _debug("no conn");
547 fc->error = error;
548 goto iterate_address;
549 }
550
551 iterate_address:
552 /* Iterate over the current server's address list to try and find an
553 * address on which it will respond to us.
554 */
555 if (afs_iterate_addresses(&fc->ac)) {
556 _leave(" = t");
557 return true;
558 }
559
560 afs_end_cursor(&fc->ac);
561 return false;
562 }
563
564 /*
565 * Dump cursor state in the case of the error being EDESTADDRREQ.
566 */
afs_dump_edestaddrreq(const struct afs_fs_cursor * fc)567 static void afs_dump_edestaddrreq(const struct afs_fs_cursor *fc)
568 {
569 static int count;
570 int i;
571
572 if (!IS_ENABLED(CONFIG_AFS_DEBUG_CURSOR) || count > 3)
573 return;
574 count++;
575
576 rcu_read_lock();
577
578 pr_notice("EDESTADDR occurred\n");
579 pr_notice("FC: cbb=%x cbb2=%x fl=%hx err=%hd\n",
580 fc->cb_break, fc->cb_break_2, fc->flags, fc->error);
581 pr_notice("FC: ut=%lx ix=%d ni=%u\n",
582 fc->untried, fc->index, fc->nr_iterations);
583
584 if (fc->server_list) {
585 const struct afs_server_list *sl = fc->server_list;
586 pr_notice("FC: SL nr=%u pr=%u vnov=%hx\n",
587 sl->nr_servers, sl->preferred, sl->vnovol_mask);
588 for (i = 0; i < sl->nr_servers; i++) {
589 const struct afs_server *s = sl->servers[i].server;
590 pr_notice("FC: server fl=%lx av=%u %pU\n",
591 s->flags, s->addr_version, &s->uuid);
592 if (s->addresses) {
593 const struct afs_addr_list *a =
594 rcu_dereference(s->addresses);
595 pr_notice("FC: - av=%u nr=%u/%u/%u pr=%u\n",
596 a->version,
597 a->nr_ipv4, a->nr_addrs, a->max_addrs,
598 a->preferred);
599 pr_notice("FC: - pr=%lx R=%lx F=%lx\n",
600 a->probed, a->responded, a->failed);
601 if (a == fc->ac.alist)
602 pr_notice("FC: - current\n");
603 }
604 }
605 }
606
607 pr_notice("AC: t=%lx ax=%u ac=%d er=%d r=%u ni=%u\n",
608 fc->ac.tried, fc->ac.index, fc->ac.abort_code, fc->ac.error,
609 fc->ac.responded, fc->ac.nr_iterations);
610 rcu_read_unlock();
611 }
612
613 /*
614 * Tidy up a filesystem cursor and unlock the vnode.
615 */
afs_end_vnode_operation(struct afs_fs_cursor * fc)616 int afs_end_vnode_operation(struct afs_fs_cursor *fc)
617 {
618 struct afs_net *net = afs_v2net(fc->vnode);
619
620 if (fc->error == -EDESTADDRREQ ||
621 fc->error == -EADDRNOTAVAIL ||
622 fc->error == -ENETUNREACH ||
623 fc->error == -EHOSTUNREACH)
624 afs_dump_edestaddrreq(fc);
625
626 mutex_unlock(&fc->vnode->io_lock);
627
628 afs_end_cursor(&fc->ac);
629 afs_put_cb_interest(net, fc->cbi);
630 afs_put_serverlist(net, fc->server_list);
631
632 if (fc->error == -ECONNABORTED)
633 fc->error = afs_abort_to_error(fc->ac.abort_code);
634
635 return fc->error;
636 }
637