]>
Commit | Line | Data |
---|---|---|
b411b363 PR |
1 | /* |
2 | drbd_req.c | |
3 | ||
4 | This file is part of DRBD by Philipp Reisner and Lars Ellenberg. | |
5 | ||
6 | Copyright (C) 2001-2008, LINBIT Information Technologies GmbH. | |
7 | Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>. | |
8 | Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>. | |
9 | ||
10 | drbd is free software; you can redistribute it and/or modify | |
11 | it under the terms of the GNU General Public License as published by | |
12 | the Free Software Foundation; either version 2, or (at your option) | |
13 | any later version. | |
14 | ||
15 | drbd is distributed in the hope that it will be useful, | |
16 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
17 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
18 | GNU General Public License for more details. | |
19 | ||
20 | You should have received a copy of the GNU General Public License | |
21 | along with drbd; see the file COPYING. If not, write to | |
22 | the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. | |
23 | ||
24 | */ | |
25 | ||
b411b363 PR |
26 | #include <linux/module.h> |
27 | ||
28 | #include <linux/slab.h> | |
29 | #include <linux/drbd.h> | |
30 | #include "drbd_int.h" | |
b411b363 PR |
31 | #include "drbd_req.h" |
32 | ||
33 | ||
b30ab791 | 34 | static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size); |
57bcb6cf | 35 | |
b411b363 | 36 | /* Update disk stats at start of I/O request */ |
b30ab791 | 37 | static void _drbd_start_io_acct(struct drbd_device *device, struct drbd_request *req) |
b411b363 | 38 | { |
24480854 GZ |
39 | generic_start_io_acct(bio_data_dir(req->master_bio), req->i.size >> 9, |
40 | &device->vdisk->part0); | |
b411b363 PR |
41 | } |
42 | ||
43 | /* Update disk stats when completing request upwards */ | |
b30ab791 | 44 | static void _drbd_end_io_acct(struct drbd_device *device, struct drbd_request *req) |
b411b363 | 45 | { |
24480854 GZ |
46 | generic_end_io_acct(bio_data_dir(req->master_bio), |
47 | &device->vdisk->part0, req->start_jif); | |
b411b363 PR |
48 | } |
49 | ||
9104d31a | 50 | static struct drbd_request *drbd_req_new(struct drbd_device *device, struct bio *bio_src) |
9e204cdd AG |
51 | { |
52 | struct drbd_request *req; | |
53 | ||
23fe8f8b | 54 | req = mempool_alloc(drbd_request_mempool, GFP_NOIO); |
9e204cdd AG |
55 | if (!req) |
56 | return NULL; | |
23fe8f8b | 57 | memset(req, 0, sizeof(*req)); |
9e204cdd AG |
58 | |
59 | drbd_req_make_private_bio(req, bio_src); | |
9104d31a LE |
60 | req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0) |
61 | | (bio_op(bio_src) == REQ_OP_WRITE_SAME ? RQ_WSAME : 0) | |
62 | | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0); | |
63 | req->device = device; | |
64 | req->master_bio = bio_src; | |
65 | req->epoch = 0; | |
53840641 | 66 | |
9e204cdd | 67 | drbd_clear_interval(&req->i); |
4f024f37 KO |
68 | req->i.sector = bio_src->bi_iter.bi_sector; |
69 | req->i.size = bio_src->bi_iter.bi_size; | |
5e472264 | 70 | req->i.local = true; |
53840641 AG |
71 | req->i.waiting = false; |
72 | ||
9e204cdd AG |
73 | INIT_LIST_HEAD(&req->tl_requests); |
74 | INIT_LIST_HEAD(&req->w.list); | |
844a6ae7 LE |
75 | INIT_LIST_HEAD(&req->req_pending_master_completion); |
76 | INIT_LIST_HEAD(&req->req_pending_local); | |
9e204cdd | 77 | |
a0d856df | 78 | /* one reference to be put by __drbd_make_request */ |
b406777e | 79 | atomic_set(&req->completion_ref, 1); |
a0d856df | 80 | /* one kref as long as completion_ref > 0 */ |
b406777e | 81 | kref_init(&req->kref); |
9e204cdd AG |
82 | return req; |
83 | } | |
84 | ||
08d0dabf LE |
85 | static void drbd_remove_request_interval(struct rb_root *root, |
86 | struct drbd_request *req) | |
87 | { | |
88 | struct drbd_device *device = req->device; | |
89 | struct drbd_interval *i = &req->i; | |
90 | ||
91 | drbd_remove_interval(root, i); | |
92 | ||
93 | /* Wake up any processes waiting for this request to complete. */ | |
94 | if (i->waiting) | |
95 | wake_up(&device->misc_wait); | |
96 | } | |
97 | ||
9a278a79 | 98 | void drbd_req_destroy(struct kref *kref) |
b411b363 | 99 | { |
b406777e | 100 | struct drbd_request *req = container_of(kref, struct drbd_request, kref); |
84b8c06b | 101 | struct drbd_device *device = req->device; |
a0d856df LE |
102 | const unsigned s = req->rq_state; |
103 | ||
104 | if ((req->master_bio && !(s & RQ_POSTPONED)) || | |
105 | atomic_read(&req->completion_ref) || | |
106 | (s & RQ_LOCAL_PENDING) || | |
107 | ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) { | |
d0180171 | 108 | drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n", |
a0d856df LE |
109 | s, atomic_read(&req->completion_ref)); |
110 | return; | |
111 | } | |
288f422e | 112 | |
844a6ae7 LE |
113 | /* If called from mod_rq_state (expected normal case) or |
114 | * drbd_send_and_submit (the less likely normal path), this holds the | |
115 | * req_lock, and req->tl_requests will typicaly be on ->transfer_log, | |
116 | * though it may be still empty (never added to the transfer log). | |
117 | * | |
118 | * If called from do_retry(), we do NOT hold the req_lock, but we are | |
119 | * still allowed to unconditionally list_del(&req->tl_requests), | |
120 | * because it will be on a local on-stack list only. */ | |
2312f0b3 | 121 | list_del_init(&req->tl_requests); |
288f422e | 122 | |
08d0dabf LE |
123 | /* finally remove the request from the conflict detection |
124 | * respective block_id verification interval tree. */ | |
125 | if (!drbd_interval_empty(&req->i)) { | |
126 | struct rb_root *root; | |
127 | ||
128 | if (s & RQ_WRITE) | |
129 | root = &device->write_requests; | |
130 | else | |
131 | root = &device->read_requests; | |
132 | drbd_remove_request_interval(root, req); | |
133 | } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0) | |
134 | drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n", | |
135 | s, (unsigned long long)req->i.sector, req->i.size); | |
136 | ||
b411b363 PR |
137 | /* if it was a write, we may have to set the corresponding |
138 | * bit(s) out-of-sync first. If it had a local part, we need to | |
139 | * release the reference to the activity log. */ | |
b406777e | 140 | if (s & RQ_WRITE) { |
b411b363 PR |
141 | /* Set out-of-sync unless both OK flags are set |
142 | * (local only or remote failed). | |
143 | * Other places where we set out-of-sync: | |
144 | * READ with local io-error */ | |
b411b363 | 145 | |
70f17b6b LE |
146 | /* There is a special case: |
147 | * we may notice late that IO was suspended, | |
148 | * and postpone, or schedule for retry, a write, | |
149 | * before it even was submitted or sent. | |
150 | * In that case we do not want to touch the bitmap at all. | |
151 | */ | |
152 | if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) { | |
d7644018 | 153 | if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK)) |
b30ab791 | 154 | drbd_set_out_of_sync(device, req->i.sector, req->i.size); |
b411b363 | 155 | |
d7644018 | 156 | if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS)) |
b30ab791 | 157 | drbd_set_in_sync(device, req->i.sector, req->i.size); |
d7644018 | 158 | } |
b411b363 PR |
159 | |
160 | /* one might be tempted to move the drbd_al_complete_io | |
fcefa62e | 161 | * to the local io completion callback drbd_request_endio. |
b411b363 PR |
162 | * but, if this was a mirror write, we may only |
163 | * drbd_al_complete_io after this is RQ_NET_DONE, | |
164 | * otherwise the extent could be dropped from the al | |
165 | * before it has actually been written on the peer. | |
166 | * if we crash before our peer knows about the request, | |
167 | * but after the extent has been dropped from the al, | |
168 | * we would forget to resync the corresponding extent. | |
169 | */ | |
76590cd1 | 170 | if (s & RQ_IN_ACT_LOG) { |
b30ab791 AG |
171 | if (get_ldev_if_state(device, D_FAILED)) { |
172 | drbd_al_complete_io(device, &req->i); | |
173 | put_ldev(device); | |
b411b363 | 174 | } else if (__ratelimit(&drbd_ratelimit_state)) { |
d0180171 | 175 | drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), " |
181286ad LE |
176 | "but my Disk seems to have failed :(\n", |
177 | (unsigned long long) req->i.sector, req->i.size); | |
b411b363 PR |
178 | } |
179 | } | |
180 | } | |
181 | ||
9a278a79 | 182 | mempool_free(req, drbd_request_mempool); |
b411b363 PR |
183 | } |
184 | ||
bde89a9e AG |
185 | static void wake_all_senders(struct drbd_connection *connection) |
186 | { | |
187 | wake_up(&connection->sender_work.q_wait); | |
b411b363 PR |
188 | } |
189 | ||
b6dd1a89 | 190 | /* must hold resource->req_lock */ |
bde89a9e | 191 | void start_new_tl_epoch(struct drbd_connection *connection) |
b411b363 | 192 | { |
99b4d8fe | 193 | /* no point closing an epoch, if it is empty, anyways. */ |
bde89a9e | 194 | if (connection->current_tle_writes == 0) |
99b4d8fe | 195 | return; |
b411b363 | 196 | |
bde89a9e AG |
197 | connection->current_tle_writes = 0; |
198 | atomic_inc(&connection->current_tle_nr); | |
199 | wake_all_senders(connection); | |
b411b363 PR |
200 | } |
201 | ||
b30ab791 | 202 | void complete_master_bio(struct drbd_device *device, |
b411b363 PR |
203 | struct bio_and_error *m) |
204 | { | |
4246a0b6 CH |
205 | m->bio->bi_error = m->error; |
206 | bio_endio(m->bio); | |
b30ab791 | 207 | dec_ap_bio(device); |
b411b363 PR |
208 | } |
209 | ||
53840641 | 210 | |
b411b363 PR |
211 | /* Helper for __req_mod(). |
212 | * Set m->bio to the master bio, if it is fit to be completed, | |
213 | * or leave it alone (it is initialized to NULL in __req_mod), | |
214 | * if it has already been completed, or cannot be completed yet. | |
215 | * If m->bio is set, the error status to be returned is placed in m->error. | |
216 | */ | |
6870ca6d | 217 | static |
a0d856df | 218 | void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m) |
b411b363 | 219 | { |
a0d856df | 220 | const unsigned s = req->rq_state; |
84b8c06b | 221 | struct drbd_device *device = req->device; |
a0d856df LE |
222 | int rw; |
223 | int error, ok; | |
b411b363 | 224 | |
b411b363 PR |
225 | /* we must not complete the master bio, while it is |
226 | * still being processed by _drbd_send_zc_bio (drbd_send_dblock) | |
227 | * not yet acknowledged by the peer | |
228 | * not yet completed by the local io subsystem | |
229 | * these flags may get cleared in any order by | |
230 | * the worker, | |
231 | * the receiver, | |
232 | * the bio_endio completion callbacks. | |
233 | */ | |
a0d856df LE |
234 | if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) || |
235 | (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) || | |
236 | (s & RQ_COMPLETION_SUSP)) { | |
d0180171 | 237 | drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s); |
b411b363 | 238 | return; |
a0d856df LE |
239 | } |
240 | ||
241 | if (!req->master_bio) { | |
d0180171 | 242 | drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n"); |
b411b363 | 243 | return; |
a0d856df | 244 | } |
b411b363 | 245 | |
a0d856df | 246 | rw = bio_rw(req->master_bio); |
b411b363 | 247 | |
a0d856df LE |
248 | /* |
249 | * figure out whether to report success or failure. | |
250 | * | |
251 | * report success when at least one of the operations succeeded. | |
252 | * or, to put the other way, | |
253 | * only report failure, when both operations failed. | |
254 | * | |
255 | * what to do about the failures is handled elsewhere. | |
256 | * what we need to do here is just: complete the master_bio. | |
257 | * | |
258 | * local completion error, if any, has been stored as ERR_PTR | |
259 | * in private_bio within drbd_request_endio. | |
260 | */ | |
261 | ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK); | |
262 | error = PTR_ERR(req->private_bio); | |
b411b363 | 263 | |
a0d856df LE |
264 | /* Before we can signal completion to the upper layers, |
265 | * we may need to close the current transfer log epoch. | |
266 | * We are within the request lock, so we can simply compare | |
267 | * the request epoch number with the current transfer log | |
268 | * epoch number. If they match, increase the current_tle_nr, | |
269 | * and reset the transfer log epoch write_cnt. | |
270 | */ | |
271 | if (rw == WRITE && | |
a6b32bc3 AG |
272 | req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr)) |
273 | start_new_tl_epoch(first_peer_device(device)->connection); | |
b411b363 | 274 | |
a0d856df | 275 | /* Update disk stats */ |
b30ab791 | 276 | _drbd_end_io_acct(device, req); |
b411b363 | 277 | |
a0d856df LE |
278 | /* If READ failed, |
279 | * have it be pushed back to the retry work queue, | |
280 | * so it will re-enter __drbd_make_request(), | |
281 | * and be re-assigned to a suitable local or remote path, | |
282 | * or failed if we do not have access to good data anymore. | |
283 | * | |
284 | * Unless it was failed early by __drbd_make_request(), | |
285 | * because no path was available, in which case | |
286 | * it was not even added to the transfer_log. | |
287 | * | |
288 | * READA may fail, and will not be retried. | |
289 | * | |
290 | * WRITE should have used all available paths already. | |
291 | */ | |
292 | if (!ok && rw == READ && !list_empty(&req->tl_requests)) | |
293 | req->rq_state |= RQ_POSTPONED; | |
b411b363 | 294 | |
a0d856df | 295 | if (!(req->rq_state & RQ_POSTPONED)) { |
b411b363 PR |
296 | m->error = ok ? 0 : (error ?: -EIO); |
297 | m->bio = req->master_bio; | |
298 | req->master_bio = NULL; | |
08d0dabf LE |
299 | /* We leave it in the tree, to be able to verify later |
300 | * write-acks in protocol != C during resync. | |
301 | * But we mark it as "complete", so it won't be counted as | |
302 | * conflict in a multi-primary setup. */ | |
303 | req->i.completed = true; | |
b411b363 | 304 | } |
08d0dabf LE |
305 | |
306 | if (req->i.waiting) | |
307 | wake_up(&device->misc_wait); | |
844a6ae7 LE |
308 | |
309 | /* Either we are about to complete to upper layers, | |
310 | * or we will restart this request. | |
311 | * In either case, the request object will be destroyed soon, | |
312 | * so better remove it from all lists. */ | |
313 | list_del_init(&req->req_pending_master_completion); | |
b411b363 | 314 | } |
b411b363 | 315 | |
844a6ae7 | 316 | /* still holds resource->req_lock */ |
a0d856df | 317 | static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put) |
cfa03415 | 318 | { |
84b8c06b | 319 | struct drbd_device *device = req->device; |
0b0ba1ef | 320 | D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED)); |
a0d856df LE |
321 | |
322 | if (!atomic_sub_and_test(put, &req->completion_ref)) | |
323 | return 0; | |
2b4dd36f | 324 | |
a0d856df | 325 | drbd_req_complete(req, m); |
9a278a79 LE |
326 | |
327 | if (req->rq_state & RQ_POSTPONED) { | |
328 | /* don't destroy the req object just yet, | |
329 | * but queue it for retry */ | |
330 | drbd_restart_request(req); | |
331 | return 0; | |
b411b363 | 332 | } |
9a278a79 | 333 | |
a0d856df | 334 | return 1; |
b411b363 PR |
335 | } |
336 | ||
7753a4c1 LE |
337 | static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req) |
338 | { | |
339 | struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; | |
340 | if (!connection) | |
341 | return; | |
342 | if (connection->req_next == NULL) | |
343 | connection->req_next = req; | |
344 | } | |
345 | ||
346 | static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req) | |
347 | { | |
348 | struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; | |
349 | if (!connection) | |
350 | return; | |
351 | if (connection->req_next != req) | |
352 | return; | |
353 | list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) { | |
354 | const unsigned s = req->rq_state; | |
355 | if (s & RQ_NET_QUEUED) | |
356 | break; | |
357 | } | |
358 | if (&req->tl_requests == &connection->transfer_log) | |
359 | req = NULL; | |
360 | connection->req_next = req; | |
361 | } | |
362 | ||
363 | static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req) | |
364 | { | |
365 | struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; | |
366 | if (!connection) | |
367 | return; | |
368 | if (connection->req_ack_pending == NULL) | |
369 | connection->req_ack_pending = req; | |
370 | } | |
371 | ||
372 | static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req) | |
373 | { | |
374 | struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; | |
375 | if (!connection) | |
376 | return; | |
377 | if (connection->req_ack_pending != req) | |
378 | return; | |
379 | list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) { | |
380 | const unsigned s = req->rq_state; | |
381 | if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING)) | |
382 | break; | |
383 | } | |
384 | if (&req->tl_requests == &connection->transfer_log) | |
385 | req = NULL; | |
386 | connection->req_ack_pending = req; | |
387 | } | |
388 | ||
389 | static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req) | |
390 | { | |
391 | struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; | |
392 | if (!connection) | |
393 | return; | |
394 | if (connection->req_not_net_done == NULL) | |
395 | connection->req_not_net_done = req; | |
396 | } | |
397 | ||
398 | static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req) | |
399 | { | |
400 | struct drbd_connection *connection = peer_device ? peer_device->connection : NULL; | |
401 | if (!connection) | |
402 | return; | |
403 | if (connection->req_not_net_done != req) | |
404 | return; | |
405 | list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) { | |
406 | const unsigned s = req->rq_state; | |
407 | if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE)) | |
408 | break; | |
409 | } | |
410 | if (&req->tl_requests == &connection->transfer_log) | |
411 | req = NULL; | |
412 | connection->req_not_net_done = req; | |
413 | } | |
414 | ||
a0d856df LE |
415 | /* I'd like this to be the only place that manipulates |
416 | * req->completion_ref and req->kref. */ | |
417 | static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m, | |
418 | int clear, int set) | |
cfa03415 | 419 | { |
84b8c06b | 420 | struct drbd_device *device = req->device; |
7753a4c1 | 421 | struct drbd_peer_device *peer_device = first_peer_device(device); |
a0d856df LE |
422 | unsigned s = req->rq_state; |
423 | int c_put = 0; | |
424 | int k_put = 0; | |
cfa03415 | 425 | |
b30ab791 | 426 | if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP)) |
5af2e8ce | 427 | set |= RQ_COMPLETION_SUSP; |
cfa03415 | 428 | |
a0d856df | 429 | /* apply */ |
b411b363 | 430 | |
a0d856df LE |
431 | req->rq_state &= ~clear; |
432 | req->rq_state |= set; | |
b411b363 | 433 | |
a0d856df LE |
434 | /* no change? */ |
435 | if (req->rq_state == s) | |
436 | return; | |
b411b363 | 437 | |
a0d856df LE |
438 | /* intent: get references */ |
439 | ||
440 | if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING)) | |
441 | atomic_inc(&req->completion_ref); | |
442 | ||
443 | if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) { | |
b30ab791 | 444 | inc_ap_pending(device); |
a0d856df | 445 | atomic_inc(&req->completion_ref); |
b411b363 PR |
446 | } |
447 | ||
e5f891b2 | 448 | if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) { |
a0d856df | 449 | atomic_inc(&req->completion_ref); |
7753a4c1 | 450 | set_if_null_req_next(peer_device, req); |
e5f891b2 | 451 | } |
a0d856df LE |
452 | |
453 | if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK)) | |
454 | kref_get(&req->kref); /* wait for the DONE */ | |
455 | ||
e5f891b2 | 456 | if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) { |
668700b4 | 457 | /* potentially already completed in the ack_receiver thread */ |
7753a4c1 | 458 | if (!(s & RQ_NET_DONE)) { |
e5f891b2 | 459 | atomic_add(req->i.size >> 9, &device->ap_in_flight); |
7753a4c1 LE |
460 | set_if_null_req_not_net_done(peer_device, req); |
461 | } | |
f85d9f2d | 462 | if (req->rq_state & RQ_NET_PENDING) |
7753a4c1 | 463 | set_if_null_req_ack_pending(peer_device, req); |
e5f891b2 | 464 | } |
a0d856df | 465 | |
5af2e8ce PR |
466 | if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP)) |
467 | atomic_inc(&req->completion_ref); | |
468 | ||
a0d856df LE |
469 | /* progress: put references */ |
470 | ||
471 | if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP)) | |
472 | ++c_put; | |
473 | ||
474 | if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) { | |
0b0ba1ef | 475 | D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING); |
a0d856df LE |
476 | /* local completion may still come in later, |
477 | * we need to keep the req object around. */ | |
478 | kref_get(&req->kref); | |
479 | ++c_put; | |
b411b363 | 480 | } |
b411b363 | 481 | |
a0d856df LE |
482 | if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) { |
483 | if (req->rq_state & RQ_LOCAL_ABORTED) | |
484 | ++k_put; | |
485 | else | |
486 | ++c_put; | |
844a6ae7 | 487 | list_del_init(&req->req_pending_local); |
a0d856df | 488 | } |
b411b363 | 489 | |
a0d856df | 490 | if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) { |
b30ab791 | 491 | dec_ap_pending(device); |
a0d856df | 492 | ++c_put; |
e5f891b2 | 493 | req->acked_jif = jiffies; |
7753a4c1 | 494 | advance_conn_req_ack_pending(peer_device, req); |
a0d856df LE |
495 | } |
496 | ||
7753a4c1 | 497 | if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) { |
a0d856df | 498 | ++c_put; |
7753a4c1 LE |
499 | advance_conn_req_next(peer_device, req); |
500 | } | |
a0d856df | 501 | |
e5f891b2 LE |
502 | if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) { |
503 | if (s & RQ_NET_SENT) | |
b30ab791 | 504 | atomic_sub(req->i.size >> 9, &device->ap_in_flight); |
e5f891b2 LE |
505 | if (s & RQ_EXP_BARR_ACK) |
506 | ++k_put; | |
507 | req->net_done_jif = jiffies; | |
7753a4c1 LE |
508 | |
509 | /* in ahead/behind mode, or just in case, | |
510 | * before we finally destroy this request, | |
511 | * the caching pointers must not reference it anymore */ | |
512 | advance_conn_req_next(peer_device, req); | |
513 | advance_conn_req_ack_pending(peer_device, req); | |
514 | advance_conn_req_not_net_done(peer_device, req); | |
a0d856df LE |
515 | } |
516 | ||
517 | /* potentially complete and destroy */ | |
518 | ||
519 | if (k_put || c_put) { | |
520 | /* Completion does it's own kref_put. If we are going to | |
521 | * kref_sub below, we need req to be still around then. */ | |
522 | int at_least = k_put + !!c_put; | |
523 | int refcount = atomic_read(&req->kref.refcount); | |
524 | if (refcount < at_least) | |
d0180171 | 525 | drbd_err(device, |
a0d856df LE |
526 | "mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n", |
527 | s, req->rq_state, refcount, at_least); | |
528 | } | |
529 | ||
530 | /* If we made progress, retry conflicting peer requests, if any. */ | |
531 | if (req->i.waiting) | |
b30ab791 | 532 | wake_up(&device->misc_wait); |
a0d856df LE |
533 | |
534 | if (c_put) | |
535 | k_put += drbd_req_put_completion_ref(req, m, c_put); | |
536 | if (k_put) | |
537 | kref_sub(&req->kref, k_put, drbd_req_destroy); | |
b411b363 PR |
538 | } |
539 | ||
b30ab791 | 540 | static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req) |
ccae7868 LE |
541 | { |
542 | char b[BDEVNAME_SIZE]; | |
543 | ||
42839f65 | 544 | if (!__ratelimit(&drbd_ratelimit_state)) |
ccae7868 LE |
545 | return; |
546 | ||
d0180171 | 547 | drbd_warn(device, "local %s IO error sector %llu+%u on %s\n", |
ccae7868 | 548 | (req->rq_state & RQ_WRITE) ? "WRITE" : "READ", |
42839f65 LE |
549 | (unsigned long long)req->i.sector, |
550 | req->i.size >> 9, | |
b30ab791 | 551 | bdevname(device->ldev->backing_bdev, b)); |
ccae7868 LE |
552 | } |
553 | ||
e5f891b2 LE |
554 | /* Helper for HANDED_OVER_TO_NETWORK. |
555 | * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)? | |
556 | * Is it also still "PENDING"? | |
557 | * --> If so, clear PENDING and set NET_OK below. | |
558 | * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster | |
559 | * (and we must not set RQ_NET_OK) */ | |
560 | static inline bool is_pending_write_protocol_A(struct drbd_request *req) | |
561 | { | |
562 | return (req->rq_state & | |
563 | (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK)) | |
564 | == (RQ_WRITE|RQ_NET_PENDING); | |
565 | } | |
566 | ||
b411b363 PR |
567 | /* obviously this could be coded as many single functions |
568 | * instead of one huge switch, | |
569 | * or by putting the code directly in the respective locations | |
570 | * (as it has been before). | |
571 | * | |
572 | * but having it this way | |
573 | * enforces that it is all in this one place, where it is easier to audit, | |
574 | * it makes it obvious that whatever "event" "happens" to a request should | |
575 | * happen "atomically" within the req_lock, | |
576 | * and it enforces that we have to think in a very structured manner | |
577 | * about the "events" that may happen to a request during its life time ... | |
578 | */ | |
2a80699f | 579 | int __req_mod(struct drbd_request *req, enum drbd_req_event what, |
b411b363 PR |
580 | struct bio_and_error *m) |
581 | { | |
44a4d551 LE |
582 | struct drbd_device *const device = req->device; |
583 | struct drbd_peer_device *const peer_device = first_peer_device(device); | |
584 | struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL; | |
44ed167d | 585 | struct net_conf *nc; |
303d1448 | 586 | int p, rv = 0; |
7be8da07 AG |
587 | |
588 | if (m) | |
589 | m->bio = NULL; | |
b411b363 | 590 | |
b411b363 PR |
591 | switch (what) { |
592 | default: | |
d0180171 | 593 | drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__); |
b411b363 PR |
594 | break; |
595 | ||
596 | /* does not happen... | |
597 | * initialization done in drbd_req_new | |
8554df1c | 598 | case CREATED: |
b411b363 PR |
599 | break; |
600 | */ | |
601 | ||
8554df1c | 602 | case TO_BE_SENT: /* via network */ |
7be8da07 | 603 | /* reached via __drbd_make_request |
b411b363 | 604 | * and from w_read_retry_remote */ |
0b0ba1ef | 605 | D_ASSERT(device, !(req->rq_state & RQ_NET_MASK)); |
44ed167d | 606 | rcu_read_lock(); |
44a4d551 | 607 | nc = rcu_dereference(connection->net_conf); |
44ed167d PR |
608 | p = nc->wire_protocol; |
609 | rcu_read_unlock(); | |
303d1448 PR |
610 | req->rq_state |= |
611 | p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK : | |
612 | p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0; | |
a0d856df | 613 | mod_rq_state(req, m, 0, RQ_NET_PENDING); |
b411b363 PR |
614 | break; |
615 | ||
8554df1c | 616 | case TO_BE_SUBMITTED: /* locally */ |
7be8da07 | 617 | /* reached via __drbd_make_request */ |
0b0ba1ef | 618 | D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK)); |
a0d856df | 619 | mod_rq_state(req, m, 0, RQ_LOCAL_PENDING); |
b411b363 PR |
620 | break; |
621 | ||
8554df1c | 622 | case COMPLETED_OK: |
2b4dd36f | 623 | if (req->rq_state & RQ_WRITE) |
b30ab791 | 624 | device->writ_cnt += req->i.size >> 9; |
b411b363 | 625 | else |
b30ab791 | 626 | device->read_cnt += req->i.size >> 9; |
b411b363 | 627 | |
a0d856df LE |
628 | mod_rq_state(req, m, RQ_LOCAL_PENDING, |
629 | RQ_LOCAL_COMPLETED|RQ_LOCAL_OK); | |
b411b363 PR |
630 | break; |
631 | ||
cdfda633 | 632 | case ABORT_DISK_IO: |
a0d856df | 633 | mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED); |
2b4dd36f PR |
634 | break; |
635 | ||
edc9f5eb | 636 | case WRITE_COMPLETED_WITH_ERROR: |
b30ab791 AG |
637 | drbd_report_io_error(device, req); |
638 | __drbd_chk_io_error(device, DRBD_WRITE_ERROR); | |
edc9f5eb | 639 | mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); |
b411b363 PR |
640 | break; |
641 | ||
8554df1c | 642 | case READ_COMPLETED_WITH_ERROR: |
b30ab791 AG |
643 | drbd_set_out_of_sync(device, req->i.sector, req->i.size); |
644 | drbd_report_io_error(device, req); | |
645 | __drbd_chk_io_error(device, DRBD_READ_ERROR); | |
a0d856df LE |
646 | /* fall through. */ |
647 | case READ_AHEAD_COMPLETED_WITH_ERROR: | |
648 | /* it is legal to fail READA, no __drbd_chk_io_error in that case. */ | |
649 | mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); | |
2f632aeb LE |
650 | break; |
651 | ||
652 | case DISCARD_COMPLETED_NOTSUPP: | |
653 | case DISCARD_COMPLETED_WITH_ERROR: | |
654 | /* I'd rather not detach from local disk just because it | |
655 | * failed a REQ_DISCARD. */ | |
656 | mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); | |
4439c400 | 657 | break; |
b411b363 | 658 | |
8554df1c | 659 | case QUEUE_FOR_NET_READ: |
b411b363 PR |
660 | /* READ or READA, and |
661 | * no local disk, | |
662 | * or target area marked as invalid, | |
663 | * or just got an io-error. */ | |
7be8da07 | 664 | /* from __drbd_make_request |
b411b363 PR |
665 | * or from bio_endio during read io-error recovery */ |
666 | ||
6870ca6d LE |
667 | /* So we can verify the handle in the answer packet. |
668 | * Corresponding drbd_remove_request_interval is in | |
a0d856df | 669 | * drbd_req_complete() */ |
0b0ba1ef | 670 | D_ASSERT(device, drbd_interval_empty(&req->i)); |
b30ab791 | 671 | drbd_insert_interval(&device->read_requests, &req->i); |
b411b363 | 672 | |
b30ab791 | 673 | set_bit(UNPLUG_REMOTE, &device->flags); |
b411b363 | 674 | |
0b0ba1ef AG |
675 | D_ASSERT(device, req->rq_state & RQ_NET_PENDING); |
676 | D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0); | |
a0d856df | 677 | mod_rq_state(req, m, 0, RQ_NET_QUEUED); |
4439c400 | 678 | req->w.cb = w_send_read_req; |
44a4d551 | 679 | drbd_queue_work(&connection->sender_work, |
84b8c06b | 680 | &req->w); |
b411b363 PR |
681 | break; |
682 | ||
8554df1c | 683 | case QUEUE_FOR_NET_WRITE: |
b411b363 | 684 | /* assert something? */ |
7be8da07 | 685 | /* from __drbd_make_request only */ |
b411b363 | 686 | |
6870ca6d | 687 | /* Corresponding drbd_remove_request_interval is in |
a0d856df | 688 | * drbd_req_complete() */ |
0b0ba1ef | 689 | D_ASSERT(device, drbd_interval_empty(&req->i)); |
b30ab791 | 690 | drbd_insert_interval(&device->write_requests, &req->i); |
b411b363 PR |
691 | |
692 | /* NOTE | |
693 | * In case the req ended up on the transfer log before being | |
694 | * queued on the worker, it could lead to this request being | |
695 | * missed during cleanup after connection loss. | |
696 | * So we have to do both operations here, | |
697 | * within the same lock that protects the transfer log. | |
698 | * | |
699 | * _req_add_to_epoch(req); this has to be after the | |
700 | * _maybe_start_new_epoch(req); which happened in | |
7be8da07 | 701 | * __drbd_make_request, because we now may set the bit |
b411b363 PR |
702 | * again ourselves to close the current epoch. |
703 | * | |
704 | * Add req to the (now) current epoch (barrier). */ | |
705 | ||
83c38830 LE |
706 | /* otherwise we may lose an unplug, which may cause some remote |
707 | * io-scheduler timeout to expire, increasing maximum latency, | |
708 | * hurting performance. */ | |
b30ab791 | 709 | set_bit(UNPLUG_REMOTE, &device->flags); |
b411b363 PR |
710 | |
711 | /* queue work item to send data */ | |
0b0ba1ef | 712 | D_ASSERT(device, req->rq_state & RQ_NET_PENDING); |
a0d856df | 713 | mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK); |
b411b363 | 714 | req->w.cb = w_send_dblock; |
44a4d551 | 715 | drbd_queue_work(&connection->sender_work, |
84b8c06b | 716 | &req->w); |
b411b363 PR |
717 | |
718 | /* close the epoch, in case it outgrew the limit */ | |
44ed167d | 719 | rcu_read_lock(); |
44a4d551 | 720 | nc = rcu_dereference(connection->net_conf); |
44ed167d PR |
721 | p = nc->max_epoch_size; |
722 | rcu_read_unlock(); | |
44a4d551 LE |
723 | if (connection->current_tle_writes >= p) |
724 | start_new_tl_epoch(connection); | |
b411b363 PR |
725 | |
726 | break; | |
727 | ||
8554df1c | 728 | case QUEUE_FOR_SEND_OOS: |
a0d856df | 729 | mod_rq_state(req, m, 0, RQ_NET_QUEUED); |
8f7bed77 | 730 | req->w.cb = w_send_out_of_sync; |
44a4d551 | 731 | drbd_queue_work(&connection->sender_work, |
84b8c06b | 732 | &req->w); |
73a01a18 PR |
733 | break; |
734 | ||
ea9d6729 | 735 | case READ_RETRY_REMOTE_CANCELED: |
8554df1c | 736 | case SEND_CANCELED: |
8554df1c | 737 | case SEND_FAILED: |
b411b363 PR |
738 | /* real cleanup will be done from tl_clear. just update flags |
739 | * so it is no longer marked as on the worker queue */ | |
a0d856df | 740 | mod_rq_state(req, m, RQ_NET_QUEUED, 0); |
b411b363 PR |
741 | break; |
742 | ||
8554df1c | 743 | case HANDED_OVER_TO_NETWORK: |
b411b363 | 744 | /* assert something? */ |
e5f891b2 | 745 | if (is_pending_write_protocol_A(req)) |
b411b363 PR |
746 | /* this is what is dangerous about protocol A: |
747 | * pretend it was successfully written on the peer. */ | |
e5f891b2 LE |
748 | mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING, |
749 | RQ_NET_SENT|RQ_NET_OK); | |
750 | else | |
751 | mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT); | |
752 | /* It is still not yet RQ_NET_DONE until the | |
753 | * corresponding epoch barrier got acked as well, | |
754 | * so we know what to dirty on connection loss. */ | |
6d49e101 LE |
755 | break; |
756 | ||
27a434fe | 757 | case OOS_HANDED_TO_NETWORK: |
6d49e101 LE |
758 | /* Was not set PENDING, no longer QUEUED, so is now DONE |
759 | * as far as this connection is concerned. */ | |
a0d856df | 760 | mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE); |
b411b363 PR |
761 | break; |
762 | ||
8554df1c | 763 | case CONNECTION_LOST_WHILE_PENDING: |
b411b363 | 764 | /* transfer log cleanup after connection loss */ |
a0d856df LE |
765 | mod_rq_state(req, m, |
766 | RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP, | |
767 | RQ_NET_DONE); | |
b411b363 PR |
768 | break; |
769 | ||
d4dabbe2 LE |
770 | case CONFLICT_RESOLVED: |
771 | /* for superseded conflicting writes of multiple primaries, | |
b411b363 | 772 | * there is no need to keep anything in the tl, potential |
934722a2 LE |
773 | * node crashes are covered by the activity log. |
774 | * | |
775 | * If this request had been marked as RQ_POSTPONED before, | |
d4dabbe2 | 776 | * it will actually not be completed, but "restarted", |
934722a2 | 777 | * resubmitted from the retry worker context. */ |
0b0ba1ef AG |
778 | D_ASSERT(device, req->rq_state & RQ_NET_PENDING); |
779 | D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK); | |
934722a2 LE |
780 | mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK); |
781 | break; | |
782 | ||
0afd569a | 783 | case WRITE_ACKED_BY_PEER_AND_SIS: |
934722a2 | 784 | req->rq_state |= RQ_NET_SIS; |
8554df1c | 785 | case WRITE_ACKED_BY_PEER: |
08d0dabf LE |
786 | /* Normal operation protocol C: successfully written on peer. |
787 | * During resync, even in protocol != C, | |
788 | * we requested an explicit write ack anyways. | |
789 | * Which means we cannot even assert anything here. | |
d64957c9 | 790 | * Nothing more to do here. |
b411b363 | 791 | * We want to keep the tl in place for all protocols, to cater |
d64957c9 | 792 | * for volatile write-back caches on lower level devices. */ |
303d1448 | 793 | goto ack_common; |
8554df1c | 794 | case RECV_ACKED_BY_PEER: |
0b0ba1ef | 795 | D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK); |
b411b363 | 796 | /* protocol B; pretends to be successfully written on peer. |
8554df1c | 797 | * see also notes above in HANDED_OVER_TO_NETWORK about |
b411b363 | 798 | * protocol != C */ |
303d1448 | 799 | ack_common: |
a0d856df | 800 | mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK); |
b411b363 PR |
801 | break; |
802 | ||
7be8da07 | 803 | case POSTPONE_WRITE: |
0b0ba1ef | 804 | D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK); |
303d1448 | 805 | /* If this node has already detected the write conflict, the |
7be8da07 AG |
806 | * worker will be waiting on misc_wait. Wake it up once this |
807 | * request has completed locally. | |
808 | */ | |
0b0ba1ef | 809 | D_ASSERT(device, req->rq_state & RQ_NET_PENDING); |
7be8da07 | 810 | req->rq_state |= RQ_POSTPONED; |
a0d856df | 811 | if (req->i.waiting) |
b30ab791 | 812 | wake_up(&device->misc_wait); |
a0d856df LE |
813 | /* Do not clear RQ_NET_PENDING. This request will make further |
814 | * progress via restart_conflicting_writes() or | |
815 | * fail_postponed_requests(). Hopefully. */ | |
7be8da07 | 816 | break; |
b411b363 | 817 | |
8554df1c | 818 | case NEG_ACKED: |
46e21bba | 819 | mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0); |
b411b363 PR |
820 | break; |
821 | ||
8554df1c | 822 | case FAIL_FROZEN_DISK_IO: |
265be2d0 PR |
823 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) |
824 | break; | |
a0d856df | 825 | mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); |
265be2d0 PR |
826 | break; |
827 | ||
8554df1c | 828 | case RESTART_FROZEN_DISK_IO: |
265be2d0 PR |
829 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) |
830 | break; | |
831 | ||
a0d856df LE |
832 | mod_rq_state(req, m, |
833 | RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED, | |
834 | RQ_LOCAL_PENDING); | |
265be2d0 PR |
835 | |
836 | rv = MR_READ; | |
837 | if (bio_data_dir(req->master_bio) == WRITE) | |
838 | rv = MR_WRITE; | |
839 | ||
b30ab791 | 840 | get_ldev(device); /* always succeeds in this call path */ |
265be2d0 | 841 | req->w.cb = w_restart_disk_io; |
44a4d551 | 842 | drbd_queue_work(&connection->sender_work, |
84b8c06b | 843 | &req->w); |
265be2d0 PR |
844 | break; |
845 | ||
8554df1c | 846 | case RESEND: |
509fc019 PR |
847 | /* Simply complete (local only) READs. */ |
848 | if (!(req->rq_state & RQ_WRITE) && !req->w.cb) { | |
8a0bab2a | 849 | mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); |
509fc019 PR |
850 | break; |
851 | } | |
852 | ||
11b58e73 | 853 | /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK |
a0d856df LE |
854 | before the connection loss (B&C only); only P_BARRIER_ACK |
855 | (or the local completion?) was missing when we suspended. | |
6870ca6d LE |
856 | Throwing them out of the TL here by pretending we got a BARRIER_ACK. |
857 | During connection handshake, we ensure that the peer was not rebooted. */ | |
11b58e73 | 858 | if (!(req->rq_state & RQ_NET_OK)) { |
84b8c06b | 859 | /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync? |
a0d856df LE |
860 | * in that case we must not set RQ_NET_PENDING. */ |
861 | ||
862 | mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING); | |
11b58e73 | 863 | if (req->w.cb) { |
44a4d551 LE |
864 | /* w.cb expected to be w_send_dblock, or w_send_read_req */ |
865 | drbd_queue_work(&connection->sender_work, | |
84b8c06b | 866 | &req->w); |
11b58e73 | 867 | rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ; |
a0d856df | 868 | } /* else: FIXME can this happen? */ |
11b58e73 PR |
869 | break; |
870 | } | |
8554df1c | 871 | /* else, fall through to BARRIER_ACKED */ |
11b58e73 | 872 | |
8554df1c | 873 | case BARRIER_ACKED: |
a0d856df | 874 | /* barrier ack for READ requests does not make sense */ |
288f422e PR |
875 | if (!(req->rq_state & RQ_WRITE)) |
876 | break; | |
877 | ||
b411b363 | 878 | if (req->rq_state & RQ_NET_PENDING) { |
a209b4ae | 879 | /* barrier came in before all requests were acked. |
b411b363 PR |
880 | * this is bad, because if the connection is lost now, |
881 | * we won't be able to clean them up... */ | |
d0180171 | 882 | drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n"); |
b411b363 | 883 | } |
a0d856df LE |
884 | /* Allowed to complete requests, even while suspended. |
885 | * As this is called for all requests within a matching epoch, | |
886 | * we need to filter, and only set RQ_NET_DONE for those that | |
887 | * have actually been on the wire. */ | |
888 | mod_rq_state(req, m, RQ_COMPLETION_SUSP, | |
889 | (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0); | |
b411b363 PR |
890 | break; |
891 | ||
8554df1c | 892 | case DATA_RECEIVED: |
0b0ba1ef | 893 | D_ASSERT(device, req->rq_state & RQ_NET_PENDING); |
a0d856df | 894 | mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE); |
b411b363 | 895 | break; |
7074e4a7 LE |
896 | |
897 | case QUEUE_AS_DRBD_BARRIER: | |
44a4d551 | 898 | start_new_tl_epoch(connection); |
7074e4a7 LE |
899 | mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE); |
900 | break; | |
b411b363 | 901 | }; |
2a80699f PR |
902 | |
903 | return rv; | |
b411b363 PR |
904 | } |
905 | ||
906 | /* we may do a local read if: | |
907 | * - we are consistent (of course), | |
908 | * - or we are generally inconsistent, | |
909 | * BUT we are still/already IN SYNC for this area. | |
910 | * since size may be bigger than BM_BLOCK_SIZE, | |
911 | * we may need to check several bits. | |
912 | */ | |
b30ab791 | 913 | static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size) |
b411b363 PR |
914 | { |
915 | unsigned long sbnr, ebnr; | |
916 | sector_t esector, nr_sectors; | |
917 | ||
b30ab791 | 918 | if (device->state.disk == D_UP_TO_DATE) |
0da34df0 | 919 | return true; |
b30ab791 | 920 | if (device->state.disk != D_INCONSISTENT) |
0da34df0 | 921 | return false; |
b411b363 | 922 | esector = sector + (size >> 9) - 1; |
b30ab791 | 923 | nr_sectors = drbd_get_capacity(device->this_bdev); |
0b0ba1ef AG |
924 | D_ASSERT(device, sector < nr_sectors); |
925 | D_ASSERT(device, esector < nr_sectors); | |
b411b363 PR |
926 | |
927 | sbnr = BM_SECT_TO_BIT(sector); | |
928 | ebnr = BM_SECT_TO_BIT(esector); | |
929 | ||
b30ab791 | 930 | return drbd_bm_count_bits(device, sbnr, ebnr) == 0; |
b411b363 PR |
931 | } |
932 | ||
b30ab791 | 933 | static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector, |
5da9c836 | 934 | enum drbd_read_balancing rbm) |
380207d0 | 935 | { |
380207d0 | 936 | struct backing_dev_info *bdi; |
d60de03a | 937 | int stripe_shift; |
380207d0 | 938 | |
380207d0 PR |
939 | switch (rbm) { |
940 | case RB_CONGESTED_REMOTE: | |
b30ab791 | 941 | bdi = &device->ldev->backing_bdev->bd_disk->queue->backing_dev_info; |
380207d0 PR |
942 | return bdi_read_congested(bdi); |
943 | case RB_LEAST_PENDING: | |
b30ab791 AG |
944 | return atomic_read(&device->local_cnt) > |
945 | atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt); | |
d60de03a PR |
946 | case RB_32K_STRIPING: /* stripe_shift = 15 */ |
947 | case RB_64K_STRIPING: | |
948 | case RB_128K_STRIPING: | |
949 | case RB_256K_STRIPING: | |
950 | case RB_512K_STRIPING: | |
951 | case RB_1M_STRIPING: /* stripe_shift = 20 */ | |
952 | stripe_shift = (rbm - RB_32K_STRIPING + 15); | |
953 | return (sector >> (stripe_shift - 9)) & 1; | |
380207d0 | 954 | case RB_ROUND_ROBIN: |
b30ab791 | 955 | return test_and_change_bit(READ_BALANCE_RR, &device->flags); |
380207d0 PR |
956 | case RB_PREFER_REMOTE: |
957 | return true; | |
958 | case RB_PREFER_LOCAL: | |
959 | default: | |
960 | return false; | |
961 | } | |
962 | } | |
963 | ||
6024fece AG |
964 | /* |
965 | * complete_conflicting_writes - wait for any conflicting write requests | |
966 | * | |
967 | * The write_requests tree contains all active write requests which we | |
968 | * currently know about. Wait for any requests to complete which conflict with | |
969 | * the new one. | |
648e46b5 LE |
970 | * |
971 | * Only way out: remove the conflicting intervals from the tree. | |
6024fece | 972 | */ |
648e46b5 | 973 | static void complete_conflicting_writes(struct drbd_request *req) |
6024fece | 974 | { |
648e46b5 | 975 | DEFINE_WAIT(wait); |
84b8c06b | 976 | struct drbd_device *device = req->device; |
648e46b5 LE |
977 | struct drbd_interval *i; |
978 | sector_t sector = req->i.sector; | |
979 | int size = req->i.size; | |
980 | ||
648e46b5 | 981 | for (;;) { |
1b228c98 LE |
982 | drbd_for_each_overlap(i, &device->write_requests, sector, size) { |
983 | /* Ignore, if already completed to upper layers. */ | |
984 | if (i->completed) | |
985 | continue; | |
986 | /* Handle the first found overlap. After the schedule | |
987 | * we have to restart the tree walk. */ | |
648e46b5 | 988 | break; |
1b228c98 LE |
989 | } |
990 | if (!i) /* if any */ | |
991 | break; | |
992 | ||
648e46b5 | 993 | /* Indicate to wake up device->misc_wait on progress. */ |
1b228c98 | 994 | prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE); |
648e46b5 | 995 | i->waiting = true; |
0500813f | 996 | spin_unlock_irq(&device->resource->req_lock); |
648e46b5 | 997 | schedule(); |
0500813f | 998 | spin_lock_irq(&device->resource->req_lock); |
6024fece | 999 | } |
b30ab791 | 1000 | finish_wait(&device->misc_wait, &wait); |
b411b363 PR |
1001 | } |
1002 | ||
7e5fec31 | 1003 | /* called within req_lock */ |
b30ab791 | 1004 | static void maybe_pull_ahead(struct drbd_device *device) |
0d5934e3 | 1005 | { |
a6b32bc3 | 1006 | struct drbd_connection *connection = first_peer_device(device)->connection; |
5da9c836 LE |
1007 | struct net_conf *nc; |
1008 | bool congested = false; | |
1009 | enum drbd_on_congestion on_congestion; | |
1010 | ||
607f25e5 | 1011 | rcu_read_lock(); |
bde89a9e | 1012 | nc = rcu_dereference(connection->net_conf); |
5da9c836 | 1013 | on_congestion = nc ? nc->on_congestion : OC_BLOCK; |
607f25e5 | 1014 | rcu_read_unlock(); |
5da9c836 | 1015 | if (on_congestion == OC_BLOCK || |
bde89a9e | 1016 | connection->agreed_pro_version < 96) |
3b9ef85e | 1017 | return; |
0d5934e3 | 1018 | |
0c066bc3 LE |
1019 | if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD) |
1020 | return; /* nothing to do ... */ | |
1021 | ||
0d5934e3 LE |
1022 | /* If I don't even have good local storage, we can not reasonably try |
1023 | * to pull ahead of the peer. We also need the local reference to make | |
b30ab791 | 1024 | * sure device->act_log is there. |
0d5934e3 | 1025 | */ |
b30ab791 | 1026 | if (!get_ldev_if_state(device, D_UP_TO_DATE)) |
0d5934e3 LE |
1027 | return; |
1028 | ||
5da9c836 | 1029 | if (nc->cong_fill && |
b30ab791 | 1030 | atomic_read(&device->ap_in_flight) >= nc->cong_fill) { |
d0180171 | 1031 | drbd_info(device, "Congestion-fill threshold reached\n"); |
5da9c836 | 1032 | congested = true; |
0d5934e3 LE |
1033 | } |
1034 | ||
b30ab791 | 1035 | if (device->act_log->used >= nc->cong_extents) { |
d0180171 | 1036 | drbd_info(device, "Congestion-extents threshold reached\n"); |
5da9c836 | 1037 | congested = true; |
0d5934e3 LE |
1038 | } |
1039 | ||
1040 | if (congested) { | |
99b4d8fe | 1041 | /* start a new epoch for non-mirrored writes */ |
a6b32bc3 | 1042 | start_new_tl_epoch(first_peer_device(device)->connection); |
0d5934e3 | 1043 | |
5da9c836 | 1044 | if (on_congestion == OC_PULL_AHEAD) |
b30ab791 | 1045 | _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL); |
5da9c836 | 1046 | else /*nc->on_congestion == OC_DISCONNECT */ |
b30ab791 | 1047 | _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL); |
0d5934e3 | 1048 | } |
b30ab791 | 1049 | put_ldev(device); |
0d5934e3 LE |
1050 | } |
1051 | ||
5da9c836 LE |
1052 | /* If this returns false, and req->private_bio is still set, |
1053 | * this should be submitted locally. | |
1054 | * | |
1055 | * If it returns false, but req->private_bio is not set, | |
1056 | * we do not have access to good data :( | |
1057 | * | |
1058 | * Otherwise, this destroys req->private_bio, if any, | |
1059 | * and returns true. | |
1060 | */ | |
1061 | static bool do_remote_read(struct drbd_request *req) | |
1062 | { | |
84b8c06b | 1063 | struct drbd_device *device = req->device; |
5da9c836 LE |
1064 | enum drbd_read_balancing rbm; |
1065 | ||
1066 | if (req->private_bio) { | |
b30ab791 | 1067 | if (!drbd_may_do_local_read(device, |
5da9c836 LE |
1068 | req->i.sector, req->i.size)) { |
1069 | bio_put(req->private_bio); | |
1070 | req->private_bio = NULL; | |
b30ab791 | 1071 | put_ldev(device); |
5da9c836 LE |
1072 | } |
1073 | } | |
1074 | ||
b30ab791 | 1075 | if (device->state.pdsk != D_UP_TO_DATE) |
5da9c836 LE |
1076 | return false; |
1077 | ||
a0d856df LE |
1078 | if (req->private_bio == NULL) |
1079 | return true; | |
1080 | ||
5da9c836 LE |
1081 | /* TODO: improve read balancing decisions, take into account drbd |
1082 | * protocol, pending requests etc. */ | |
1083 | ||
1084 | rcu_read_lock(); | |
b30ab791 | 1085 | rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing; |
5da9c836 LE |
1086 | rcu_read_unlock(); |
1087 | ||
1088 | if (rbm == RB_PREFER_LOCAL && req->private_bio) | |
1089 | return false; /* submit locally */ | |
1090 | ||
b30ab791 | 1091 | if (remote_due_to_read_balancing(device, req->i.sector, rbm)) { |
5da9c836 LE |
1092 | if (req->private_bio) { |
1093 | bio_put(req->private_bio); | |
1094 | req->private_bio = NULL; | |
b30ab791 | 1095 | put_ldev(device); |
5da9c836 LE |
1096 | } |
1097 | return true; | |
1098 | } | |
1099 | ||
1100 | return false; | |
1101 | } | |
1102 | ||
2e9ffde6 AG |
1103 | bool drbd_should_do_remote(union drbd_dev_state s) |
1104 | { | |
1105 | return s.pdsk == D_UP_TO_DATE || | |
1106 | (s.pdsk >= D_INCONSISTENT && | |
1107 | s.conn >= C_WF_BITMAP_T && | |
1108 | s.conn < C_AHEAD); | |
1109 | /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T. | |
1110 | That is equivalent since before 96 IO was frozen in the C_WF_BITMAP* | |
1111 | states. */ | |
1112 | } | |
1113 | ||
1114 | static bool drbd_should_send_out_of_sync(union drbd_dev_state s) | |
1115 | { | |
1116 | return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S; | |
1117 | /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary | |
1118 | since we enter state C_AHEAD only if proto >= 96 */ | |
1119 | } | |
1120 | ||
5da9c836 LE |
1121 | /* returns number of connections (== 1, for drbd 8.4) |
1122 | * expected to actually write this data, | |
1123 | * which does NOT include those that we are L_AHEAD for. */ | |
1124 | static int drbd_process_write_request(struct drbd_request *req) | |
1125 | { | |
84b8c06b | 1126 | struct drbd_device *device = req->device; |
5da9c836 LE |
1127 | int remote, send_oos; |
1128 | ||
b30ab791 AG |
1129 | remote = drbd_should_do_remote(device->state); |
1130 | send_oos = drbd_should_send_out_of_sync(device->state); | |
5da9c836 | 1131 | |
519b6d3e LE |
1132 | /* Need to replicate writes. Unless it is an empty flush, |
1133 | * which is better mapped to a DRBD P_BARRIER packet, | |
1134 | * also for drbd wire protocol compatibility reasons. | |
1135 | * If this was a flush, just start a new epoch. | |
1136 | * Unless the current epoch was empty anyways, or we are not currently | |
1137 | * replicating, in which case there is no point. */ | |
1138 | if (unlikely(req->i.size == 0)) { | |
1139 | /* The only size==0 bios we expect are empty flushes. */ | |
28a8f0d3 | 1140 | D_ASSERT(device, req->master_bio->bi_rw & REQ_PREFLUSH); |
99b4d8fe | 1141 | if (remote) |
7074e4a7 LE |
1142 | _req_mod(req, QUEUE_AS_DRBD_BARRIER); |
1143 | return remote; | |
519b6d3e LE |
1144 | } |
1145 | ||
5da9c836 LE |
1146 | if (!remote && !send_oos) |
1147 | return 0; | |
1148 | ||
0b0ba1ef | 1149 | D_ASSERT(device, !(remote && send_oos)); |
5da9c836 LE |
1150 | |
1151 | if (remote) { | |
1152 | _req_mod(req, TO_BE_SENT); | |
1153 | _req_mod(req, QUEUE_FOR_NET_WRITE); | |
b30ab791 | 1154 | } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size)) |
5da9c836 LE |
1155 | _req_mod(req, QUEUE_FOR_SEND_OOS); |
1156 | ||
1157 | return remote; | |
1158 | } | |
1159 | ||
7435e901 LE |
1160 | static void drbd_process_discard_req(struct drbd_request *req) |
1161 | { | |
1162 | int err = drbd_issue_discard_or_zero_out(req->device, | |
1163 | req->i.sector, req->i.size >> 9, true); | |
1164 | ||
1165 | if (err) | |
1166 | req->private_bio->bi_error = -EIO; | |
1167 | bio_endio(req->private_bio); | |
1168 | } | |
1169 | ||
5da9c836 LE |
1170 | static void |
1171 | drbd_submit_req_private_bio(struct drbd_request *req) | |
1172 | { | |
84b8c06b | 1173 | struct drbd_device *device = req->device; |
5da9c836 LE |
1174 | struct bio *bio = req->private_bio; |
1175 | const int rw = bio_rw(bio); | |
1176 | ||
b30ab791 | 1177 | bio->bi_bdev = device->ldev->backing_bdev; |
5da9c836 LE |
1178 | |
1179 | /* State may have changed since we grabbed our reference on the | |
1180 | * ->ldev member. Double check, and short-circuit to endio. | |
1181 | * In case the last activity log transaction failed to get on | |
1182 | * stable storage, and this is a WRITE, we may not even submit | |
1183 | * this bio. */ | |
b30ab791 AG |
1184 | if (get_ldev(device)) { |
1185 | if (drbd_insert_fault(device, | |
5da9c836 LE |
1186 | rw == WRITE ? DRBD_FAULT_DT_WR |
1187 | : rw == READ ? DRBD_FAULT_DT_RD | |
1188 | : DRBD_FAULT_DT_RA)) | |
4246a0b6 | 1189 | bio_io_error(bio); |
7435e901 LE |
1190 | else if (bio_op(bio) == REQ_OP_DISCARD) |
1191 | drbd_process_discard_req(req); | |
5da9c836 LE |
1192 | else |
1193 | generic_make_request(bio); | |
b30ab791 | 1194 | put_ldev(device); |
5da9c836 | 1195 | } else |
4246a0b6 | 1196 | bio_io_error(bio); |
5da9c836 LE |
1197 | } |
1198 | ||
b30ab791 | 1199 | static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req) |
779b3fe4 | 1200 | { |
844a6ae7 | 1201 | spin_lock_irq(&device->resource->req_lock); |
b30ab791 | 1202 | list_add_tail(&req->tl_requests, &device->submit.writes); |
844a6ae7 LE |
1203 | list_add_tail(&req->req_pending_master_completion, |
1204 | &device->pending_master_completion[1 /* WRITE */]); | |
1205 | spin_unlock_irq(&device->resource->req_lock); | |
b30ab791 | 1206 | queue_work(device->submit.wq, &device->submit.worker); |
f5b90b6b LE |
1207 | /* do_submit() may sleep internally on al_wait, too */ |
1208 | wake_up(&device->al_wait); | |
779b3fe4 LE |
1209 | } |
1210 | ||
6d9febe2 LE |
1211 | /* returns the new drbd_request pointer, if the caller is expected to |
1212 | * drbd_send_and_submit() it (to save latency), or NULL if we queued the | |
1213 | * request on the submitter thread. | |
1214 | * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request. | |
1215 | */ | |
01cd2636 | 1216 | static struct drbd_request * |
e5f891b2 | 1217 | drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_jif) |
b411b363 | 1218 | { |
6d9febe2 | 1219 | const int rw = bio_data_dir(bio); |
b411b363 | 1220 | struct drbd_request *req; |
b411b363 PR |
1221 | |
1222 | /* allocate outside of all locks; */ | |
b30ab791 | 1223 | req = drbd_req_new(device, bio); |
b411b363 | 1224 | if (!req) { |
b30ab791 | 1225 | dec_ap_bio(device); |
b411b363 PR |
1226 | /* only pass the error to the upper layers. |
1227 | * if user cannot handle io errors, that's not our business. */ | |
d0180171 | 1228 | drbd_err(device, "could not kmalloc() req\n"); |
4246a0b6 CH |
1229 | bio->bi_error = -ENOMEM; |
1230 | bio_endio(bio); | |
6d9febe2 | 1231 | return ERR_PTR(-ENOMEM); |
b411b363 | 1232 | } |
e5f891b2 | 1233 | req->start_jif = start_jif; |
b411b363 | 1234 | |
b30ab791 | 1235 | if (!get_ldev(device)) { |
5da9c836 | 1236 | bio_put(req->private_bio); |
b411b363 PR |
1237 | req->private_bio = NULL; |
1238 | } | |
b411b363 | 1239 | |
7e8c288f | 1240 | /* Update disk stats */ |
b30ab791 | 1241 | _drbd_start_io_acct(device, req); |
7e8c288f | 1242 | |
7435e901 LE |
1243 | /* process discards always from our submitter thread */ |
1244 | if (bio_op(bio) & REQ_OP_DISCARD) | |
1245 | goto queue_for_submitter_thread; | |
1246 | ||
519b6d3e | 1247 | if (rw == WRITE && req->private_bio && req->i.size |
b30ab791 | 1248 | && !test_bit(AL_SUSPENDED, &device->flags)) { |
7435e901 LE |
1249 | if (!drbd_al_begin_io_fastpath(device, &req->i)) |
1250 | goto queue_for_submitter_thread; | |
0778286a | 1251 | req->rq_state |= RQ_IN_ACT_LOG; |
e5f891b2 | 1252 | req->in_actlog_jif = jiffies; |
0778286a | 1253 | } |
6d9febe2 | 1254 | return req; |
7435e901 LE |
1255 | |
1256 | queue_for_submitter_thread: | |
1257 | atomic_inc(&device->ap_actlog_cnt); | |
1258 | drbd_queue_write(device, req); | |
1259 | return NULL; | |
6d9febe2 LE |
1260 | } |
1261 | ||
0ead5cca LE |
1262 | /* Require at least one path to current data. |
1263 | * We don't want to allow writes on C_STANDALONE D_INCONSISTENT: | |
1264 | * We would not allow to read what was written, | |
1265 | * we would not have bumped the data generation uuids, | |
1266 | * we would cause data divergence for all the wrong reasons. | |
1267 | * | |
1268 | * If we don't see at least one D_UP_TO_DATE, we will fail this request, | |
1269 | * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO, | |
1270 | * and queues for retry later. | |
1271 | */ | |
1272 | static bool may_do_writes(struct drbd_device *device) | |
1273 | { | |
1274 | const union drbd_dev_state s = device->state; | |
1275 | return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE; | |
1276 | } | |
1277 | ||
b30ab791 | 1278 | static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req) |
6d9febe2 | 1279 | { |
35b5ed5b | 1280 | struct drbd_resource *resource = device->resource; |
6d9febe2 LE |
1281 | const int rw = bio_rw(req->master_bio); |
1282 | struct bio_and_error m = { NULL, }; | |
1283 | bool no_remote = false; | |
35b5ed5b | 1284 | bool submit_private_bio = false; |
6d9febe2 | 1285 | |
35b5ed5b | 1286 | spin_lock_irq(&resource->req_lock); |
6024fece | 1287 | if (rw == WRITE) { |
648e46b5 LE |
1288 | /* This may temporarily give up the req_lock, |
1289 | * but will re-aquire it before it returns here. | |
1290 | * Needs to be before the check on drbd_suspended() */ | |
1291 | complete_conflicting_writes(req); | |
607f25e5 LE |
1292 | /* no more giving up req_lock from now on! */ |
1293 | ||
1294 | /* check for congestion, and potentially stop sending | |
1295 | * full data updates, but start sending "dirty bits" only. */ | |
b30ab791 | 1296 | maybe_pull_ahead(device); |
b411b363 PR |
1297 | } |
1298 | ||
9a25a04c | 1299 | |
b30ab791 | 1300 | if (drbd_suspended(device)) { |
5da9c836 LE |
1301 | /* push back and retry: */ |
1302 | req->rq_state |= RQ_POSTPONED; | |
1303 | if (req->private_bio) { | |
1304 | bio_put(req->private_bio); | |
1305 | req->private_bio = NULL; | |
b30ab791 | 1306 | put_ldev(device); |
b411b363 | 1307 | } |
5da9c836 | 1308 | goto out; |
b411b363 PR |
1309 | } |
1310 | ||
5da9c836 LE |
1311 | /* We fail READ/READA early, if we can not serve it. |
1312 | * We must do this before req is registered on any lists. | |
a0d856df | 1313 | * Otherwise, drbd_req_complete() will queue failed READ for retry. */ |
5da9c836 LE |
1314 | if (rw != WRITE) { |
1315 | if (!do_remote_read(req) && !req->private_bio) | |
1316 | goto nodata; | |
b411b363 PR |
1317 | } |
1318 | ||
b6dd1a89 | 1319 | /* which transfer log epoch does this belong to? */ |
a6b32bc3 | 1320 | req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr); |
288f422e | 1321 | |
227f052f LE |
1322 | /* no point in adding empty flushes to the transfer log, |
1323 | * they are mapped to drbd barriers already. */ | |
99b4d8fe LE |
1324 | if (likely(req->i.size!=0)) { |
1325 | if (rw == WRITE) | |
a6b32bc3 | 1326 | first_peer_device(device)->connection->current_tle_writes++; |
288f422e | 1327 | |
a6b32bc3 | 1328 | list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log); |
b411b363 | 1329 | } |
67531718 | 1330 | |
5da9c836 | 1331 | if (rw == WRITE) { |
0ead5cca LE |
1332 | if (req->private_bio && !may_do_writes(device)) { |
1333 | bio_put(req->private_bio); | |
1334 | req->private_bio = NULL; | |
1335 | put_ldev(device); | |
1336 | goto nodata; | |
1337 | } | |
5da9c836 LE |
1338 | if (!drbd_process_write_request(req)) |
1339 | no_remote = true; | |
1340 | } else { | |
1341 | /* We either have a private_bio, or we can read from remote. | |
1342 | * Otherwise we had done the goto nodata above. */ | |
1343 | if (req->private_bio == NULL) { | |
1344 | _req_mod(req, TO_BE_SENT); | |
1345 | _req_mod(req, QUEUE_FOR_NET_READ); | |
6719fb03 | 1346 | } else |
5da9c836 | 1347 | no_remote = true; |
b411b363 PR |
1348 | } |
1349 | ||
844a6ae7 LE |
1350 | /* If it took the fast path in drbd_request_prepare, add it here. |
1351 | * The slow path has added it already. */ | |
1352 | if (list_empty(&req->req_pending_master_completion)) | |
1353 | list_add_tail(&req->req_pending_master_completion, | |
1354 | &device->pending_master_completion[rw == WRITE]); | |
5da9c836 LE |
1355 | if (req->private_bio) { |
1356 | /* needs to be marked within the same spinlock */ | |
05cbbb39 | 1357 | req->pre_submit_jif = jiffies; |
844a6ae7 LE |
1358 | list_add_tail(&req->req_pending_local, |
1359 | &device->pending_completion[rw == WRITE]); | |
5da9c836 LE |
1360 | _req_mod(req, TO_BE_SUBMITTED); |
1361 | /* but we need to give up the spinlock to submit */ | |
35b5ed5b | 1362 | submit_private_bio = true; |
5da9c836 LE |
1363 | } else if (no_remote) { |
1364 | nodata: | |
1365 | if (__ratelimit(&drbd_ratelimit_state)) | |
d0180171 | 1366 | drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n", |
42839f65 | 1367 | (unsigned long long)req->i.sector, req->i.size >> 9); |
5da9c836 | 1368 | /* A write may have been queued for send_oos, however. |
a0d856df | 1369 | * So we can not simply free it, we must go through drbd_req_put_completion_ref() */ |
b411b363 | 1370 | } |
b411b363 | 1371 | |
5da9c836 | 1372 | out: |
a0d856df LE |
1373 | if (drbd_req_put_completion_ref(req, &m, 1)) |
1374 | kref_put(&req->kref, drbd_req_destroy); | |
35b5ed5b LE |
1375 | spin_unlock_irq(&resource->req_lock); |
1376 | ||
1377 | /* Even though above is a kref_put(), this is safe. | |
1378 | * As long as we still need to submit our private bio, | |
1379 | * we hold a completion ref, and the request cannot disappear. | |
1380 | * If however this request did not even have a private bio to submit | |
1381 | * (e.g. remote read), req may already be invalid now. | |
1382 | * That's why we cannot check on req->private_bio. */ | |
1383 | if (submit_private_bio) | |
1384 | drbd_submit_req_private_bio(req); | |
5da9c836 | 1385 | if (m.bio) |
b30ab791 | 1386 | complete_master_bio(device, &m); |
6d9febe2 LE |
1387 | } |
1388 | ||
e5f891b2 | 1389 | void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_jif) |
6d9febe2 | 1390 | { |
e5f891b2 | 1391 | struct drbd_request *req = drbd_request_prepare(device, bio, start_jif); |
6d9febe2 LE |
1392 | if (IS_ERR_OR_NULL(req)) |
1393 | return; | |
b30ab791 | 1394 | drbd_send_and_submit(device, req); |
b411b363 PR |
1395 | } |
1396 | ||
b30ab791 | 1397 | static void submit_fast_path(struct drbd_device *device, struct list_head *incoming) |
113fef9e | 1398 | { |
08a1ddab LE |
1399 | struct drbd_request *req, *tmp; |
1400 | list_for_each_entry_safe(req, tmp, incoming, tl_requests) { | |
1401 | const int rw = bio_data_dir(req->master_bio); | |
113fef9e | 1402 | |
08a1ddab LE |
1403 | if (rw == WRITE /* rw != WRITE should not even end up here! */ |
1404 | && req->private_bio && req->i.size | |
b30ab791 AG |
1405 | && !test_bit(AL_SUSPENDED, &device->flags)) { |
1406 | if (!drbd_al_begin_io_fastpath(device, &req->i)) | |
08a1ddab LE |
1407 | continue; |
1408 | ||
1409 | req->rq_state |= RQ_IN_ACT_LOG; | |
e5f891b2 | 1410 | req->in_actlog_jif = jiffies; |
ad3fee79 | 1411 | atomic_dec(&device->ap_actlog_cnt); |
08a1ddab LE |
1412 | } |
1413 | ||
1414 | list_del_init(&req->tl_requests); | |
b30ab791 | 1415 | drbd_send_and_submit(device, req); |
113fef9e | 1416 | } |
113fef9e LE |
1417 | } |
1418 | ||
b30ab791 | 1419 | static bool prepare_al_transaction_nonblock(struct drbd_device *device, |
08a1ddab | 1420 | struct list_head *incoming, |
f5b90b6b LE |
1421 | struct list_head *pending, |
1422 | struct list_head *later) | |
08a1ddab LE |
1423 | { |
1424 | struct drbd_request *req, *tmp; | |
1425 | int wake = 0; | |
1426 | int err; | |
1427 | ||
b30ab791 | 1428 | spin_lock_irq(&device->al_lock); |
08a1ddab | 1429 | list_for_each_entry_safe(req, tmp, incoming, tl_requests) { |
b30ab791 | 1430 | err = drbd_al_begin_io_nonblock(device, &req->i); |
f5b90b6b LE |
1431 | if (err == -ENOBUFS) |
1432 | break; | |
08a1ddab LE |
1433 | if (err == -EBUSY) |
1434 | wake = 1; | |
1435 | if (err) | |
f5b90b6b LE |
1436 | list_move_tail(&req->tl_requests, later); |
1437 | else | |
1438 | list_move_tail(&req->tl_requests, pending); | |
08a1ddab | 1439 | } |
b30ab791 | 1440 | spin_unlock_irq(&device->al_lock); |
08a1ddab | 1441 | if (wake) |
b30ab791 | 1442 | wake_up(&device->al_wait); |
08a1ddab LE |
1443 | return !list_empty(pending); |
1444 | } | |
113fef9e | 1445 | |
f5b90b6b LE |
1446 | void send_and_submit_pending(struct drbd_device *device, struct list_head *pending) |
1447 | { | |
1448 | struct drbd_request *req, *tmp; | |
1449 | ||
1450 | list_for_each_entry_safe(req, tmp, pending, tl_requests) { | |
1451 | req->rq_state |= RQ_IN_ACT_LOG; | |
1452 | req->in_actlog_jif = jiffies; | |
1453 | atomic_dec(&device->ap_actlog_cnt); | |
1454 | list_del_init(&req->tl_requests); | |
1455 | drbd_send_and_submit(device, req); | |
1456 | } | |
1457 | } | |
1458 | ||
113fef9e LE |
1459 | void do_submit(struct work_struct *ws) |
1460 | { | |
b30ab791 | 1461 | struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker); |
f5b90b6b LE |
1462 | LIST_HEAD(incoming); /* from drbd_make_request() */ |
1463 | LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */ | |
1464 | LIST_HEAD(busy); /* blocked by resync requests */ | |
1465 | ||
1466 | /* grab new incoming requests */ | |
1467 | spin_lock_irq(&device->resource->req_lock); | |
1468 | list_splice_tail_init(&device->submit.writes, &incoming); | |
1469 | spin_unlock_irq(&device->resource->req_lock); | |
113fef9e | 1470 | |
08a1ddab | 1471 | for (;;) { |
f5b90b6b | 1472 | DEFINE_WAIT(wait); |
113fef9e | 1473 | |
f5b90b6b LE |
1474 | /* move used-to-be-busy back to front of incoming */ |
1475 | list_splice_init(&busy, &incoming); | |
b30ab791 | 1476 | submit_fast_path(device, &incoming); |
08a1ddab LE |
1477 | if (list_empty(&incoming)) |
1478 | break; | |
1479 | ||
45ad07b3 | 1480 | for (;;) { |
f5b90b6b LE |
1481 | prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE); |
1482 | ||
1483 | list_splice_init(&busy, &incoming); | |
1484 | prepare_al_transaction_nonblock(device, &incoming, &pending, &busy); | |
1485 | if (!list_empty(&pending)) | |
1486 | break; | |
1487 | ||
1488 | schedule(); | |
1489 | ||
1490 | /* If all currently "hot" activity log extents are kept busy by | |
1491 | * incoming requests, we still must not totally starve new | |
1492 | * requests to "cold" extents. | |
1493 | * Something left on &incoming means there had not been | |
1494 | * enough update slots available, and the activity log | |
1495 | * has been marked as "starving". | |
1496 | * | |
1497 | * Try again now, without looking for new requests, | |
1498 | * effectively blocking all new requests until we made | |
1499 | * at least _some_ progress with what we currently have. | |
1500 | */ | |
1501 | if (!list_empty(&incoming)) | |
1502 | continue; | |
1503 | ||
1504 | /* Nothing moved to pending, but nothing left | |
1505 | * on incoming: all moved to busy! | |
1506 | * Grab new and iterate. */ | |
1507 | spin_lock_irq(&device->resource->req_lock); | |
1508 | list_splice_tail_init(&device->submit.writes, &incoming); | |
1509 | spin_unlock_irq(&device->resource->req_lock); | |
1510 | } | |
1511 | finish_wait(&device->al_wait, &wait); | |
1512 | ||
1513 | /* If the transaction was full, before all incoming requests | |
1514 | * had been processed, skip ahead to commit, and iterate | |
1515 | * without splicing in more incoming requests from upper layers. | |
1516 | * | |
1517 | * Else, if all incoming have been processed, | |
1518 | * they have become either "pending" (to be submitted after | |
1519 | * next transaction commit) or "busy" (blocked by resync). | |
1520 | * | |
1521 | * Maybe more was queued, while we prepared the transaction? | |
1522 | * Try to stuff those into this transaction as well. | |
1523 | * Be strictly non-blocking here, | |
1524 | * we already have something to commit. | |
1525 | * | |
1526 | * Commit if we don't make any more progres. | |
1527 | */ | |
1528 | ||
1529 | while (list_empty(&incoming)) { | |
45ad07b3 LE |
1530 | LIST_HEAD(more_pending); |
1531 | LIST_HEAD(more_incoming); | |
1532 | bool made_progress; | |
1533 | ||
1534 | /* It is ok to look outside the lock, | |
1535 | * it's only an optimization anyways */ | |
b30ab791 | 1536 | if (list_empty(&device->submit.writes)) |
45ad07b3 LE |
1537 | break; |
1538 | ||
844a6ae7 | 1539 | spin_lock_irq(&device->resource->req_lock); |
b30ab791 | 1540 | list_splice_tail_init(&device->submit.writes, &more_incoming); |
844a6ae7 | 1541 | spin_unlock_irq(&device->resource->req_lock); |
45ad07b3 LE |
1542 | |
1543 | if (list_empty(&more_incoming)) | |
1544 | break; | |
1545 | ||
f5b90b6b | 1546 | made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy); |
45ad07b3 LE |
1547 | |
1548 | list_splice_tail_init(&more_pending, &pending); | |
1549 | list_splice_tail_init(&more_incoming, &incoming); | |
45ad07b3 LE |
1550 | if (!made_progress) |
1551 | break; | |
1552 | } | |
08a1ddab | 1553 | |
f5b90b6b LE |
1554 | drbd_al_begin_io_commit(device); |
1555 | send_and_submit_pending(device, &pending); | |
113fef9e LE |
1556 | } |
1557 | } | |
1558 | ||
dece1635 | 1559 | blk_qc_t drbd_make_request(struct request_queue *q, struct bio *bio) |
b411b363 | 1560 | { |
b30ab791 | 1561 | struct drbd_device *device = (struct drbd_device *) q->queuedata; |
e5f891b2 | 1562 | unsigned long start_jif; |
b411b363 | 1563 | |
54efd50b KO |
1564 | blk_queue_split(q, &bio, q->bio_split); |
1565 | ||
e5f891b2 | 1566 | start_jif = jiffies; |
aeda1cd6 | 1567 | |
b411b363 PR |
1568 | /* |
1569 | * what we "blindly" assume: | |
1570 | */ | |
0b0ba1ef | 1571 | D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512)); |
b411b363 | 1572 | |
b30ab791 | 1573 | inc_ap_bio(device); |
e5f891b2 | 1574 | __drbd_make_request(device, bio, start_jif); |
dece1635 | 1575 | return BLK_QC_T_NONE; |
b411b363 PR |
1576 | } |
1577 | ||
84d34f2f LE |
1578 | static bool net_timeout_reached(struct drbd_request *net_req, |
1579 | struct drbd_connection *connection, | |
1580 | unsigned long now, unsigned long ent, | |
1581 | unsigned int ko_count, unsigned int timeout) | |
1582 | { | |
1583 | struct drbd_device *device = net_req->device; | |
1584 | ||
1585 | if (!time_after(now, net_req->pre_send_jif + ent)) | |
1586 | return false; | |
1587 | ||
1588 | if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) | |
1589 | return false; | |
1590 | ||
1591 | if (net_req->rq_state & RQ_NET_PENDING) { | |
1592 | drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n", | |
1593 | jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout); | |
1594 | return true; | |
1595 | } | |
1596 | ||
1597 | /* We received an ACK already (or are using protocol A), | |
1598 | * but are waiting for the epoch closing barrier ack. | |
1599 | * Check if we sent the barrier already. We should not blame the peer | |
1600 | * for being unresponsive, if we did not even ask it yet. */ | |
1601 | if (net_req->epoch == connection->send.current_epoch_nr) { | |
1602 | drbd_warn(device, | |
1603 | "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n", | |
1604 | jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout); | |
1605 | return false; | |
1606 | } | |
1607 | ||
1608 | /* Worst case: we may have been blocked for whatever reason, then | |
1609 | * suddenly are able to send a lot of requests (and epoch separating | |
1610 | * barriers) in quick succession. | |
1611 | * The timestamp of the net_req may be much too old and not correspond | |
1612 | * to the sending time of the relevant unack'ed barrier packet, so | |
1613 | * would trigger a spurious timeout. The latest barrier packet may | |
1614 | * have a too recent timestamp to trigger the timeout, potentially miss | |
1615 | * a timeout. Right now we don't have a place to conveniently store | |
1616 | * these timestamps. | |
1617 | * But in this particular situation, the application requests are still | |
1618 | * completed to upper layers, DRBD should still "feel" responsive. | |
1619 | * No need yet to kill this connection, it may still recover. | |
1620 | * If not, eventually we will have queued enough into the network for | |
1621 | * us to block. From that point of view, the timestamp of the last sent | |
1622 | * barrier packet is relevant enough. | |
1623 | */ | |
1624 | if (time_after(now, connection->send.last_sent_barrier_jif + ent)) { | |
1625 | drbd_warn(device, "Remote failed to answer a P_BARRIER (sent at %lu jif; now=%lu jif) within %ums > ko-count (%u) * timeout (%u * 0.1s)\n", | |
1626 | connection->send.last_sent_barrier_jif, now, | |
1627 | jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout); | |
1628 | return true; | |
1629 | } | |
1630 | return false; | |
1631 | } | |
1632 | ||
1633 | /* A request is considered timed out, if | |
1634 | * - we have some effective timeout from the configuration, | |
1635 | * with some state restrictions applied, | |
1636 | * - the oldest request is waiting for a response from the network | |
1637 | * resp. the local disk, | |
1638 | * - the oldest request is in fact older than the effective timeout, | |
1639 | * - the connection was established (resp. disk was attached) | |
1640 | * for longer than the timeout already. | |
1641 | * Note that for 32bit jiffies and very stable connections/disks, | |
1642 | * we may have a wrap around, which is catched by | |
1643 | * !time_in_range(now, last_..._jif, last_..._jif + timeout). | |
1644 | * | |
1645 | * Side effect: once per 32bit wrap-around interval, which means every | |
1646 | * ~198 days with 250 HZ, we have a window where the timeout would need | |
1647 | * to expire twice (worst case) to become effective. Good enough. | |
1648 | */ | |
1649 | ||
7fde2be9 PR |
1650 | void request_timer_fn(unsigned long data) |
1651 | { | |
b30ab791 | 1652 | struct drbd_device *device = (struct drbd_device *) data; |
a6b32bc3 | 1653 | struct drbd_connection *connection = first_peer_device(device)->connection; |
7753a4c1 | 1654 | struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */ |
44ed167d | 1655 | struct net_conf *nc; |
7753a4c1 | 1656 | unsigned long oldest_submit_jif; |
dfa8bedb | 1657 | unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */ |
ba280c09 | 1658 | unsigned long now; |
84d34f2f | 1659 | unsigned int ko_count = 0, timeout = 0; |
7fde2be9 | 1660 | |
44ed167d | 1661 | rcu_read_lock(); |
bde89a9e | 1662 | nc = rcu_dereference(connection->net_conf); |
84d34f2f LE |
1663 | if (nc && device->state.conn >= C_WF_REPORT_PARAMS) { |
1664 | ko_count = nc->ko_count; | |
1665 | timeout = nc->timeout; | |
1666 | } | |
cdfda633 | 1667 | |
b30ab791 AG |
1668 | if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */ |
1669 | dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10; | |
1670 | put_ldev(device); | |
dfa8bedb | 1671 | } |
44ed167d | 1672 | rcu_read_unlock(); |
7fde2be9 | 1673 | |
84d34f2f LE |
1674 | |
1675 | ent = timeout * HZ/10 * ko_count; | |
dfa8bedb PR |
1676 | et = min_not_zero(dt, ent); |
1677 | ||
ba280c09 | 1678 | if (!et) |
7fde2be9 PR |
1679 | return; /* Recurring timer stopped */ |
1680 | ||
ba280c09 | 1681 | now = jiffies; |
7753a4c1 | 1682 | nt = now + et; |
ba280c09 | 1683 | |
0500813f | 1684 | spin_lock_irq(&device->resource->req_lock); |
7753a4c1 LE |
1685 | req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local); |
1686 | req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local); | |
84d34f2f | 1687 | |
7753a4c1 | 1688 | /* maybe the oldest request waiting for the peer is in fact still |
84d34f2f LE |
1689 | * blocking in tcp sendmsg. That's ok, though, that's handled via the |
1690 | * socket send timeout, requesting a ping, and bumping ko-count in | |
1691 | * we_should_drop_the_connection(). | |
1692 | */ | |
1693 | ||
1694 | /* check the oldest request we did successfully sent, | |
1695 | * but which is still waiting for an ACK. */ | |
1696 | req_peer = connection->req_ack_pending; | |
1697 | ||
1698 | /* if we don't have such request (e.g. protocoll A) | |
1699 | * check the oldest requests which is still waiting on its epoch | |
1700 | * closing barrier ack. */ | |
1701 | if (!req_peer) | |
1702 | req_peer = connection->req_not_net_done; | |
7753a4c1 LE |
1703 | |
1704 | /* evaluate the oldest peer request only in one timer! */ | |
1705 | if (req_peer && req_peer->device != device) | |
1706 | req_peer = NULL; | |
1707 | ||
1708 | /* do we have something to evaluate? */ | |
1709 | if (req_peer == NULL && req_write == NULL && req_read == NULL) | |
1710 | goto out; | |
1711 | ||
1712 | oldest_submit_jif = | |
1713 | (req_write && req_read) | |
1714 | ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif) | |
1715 | ? req_write->pre_submit_jif : req_read->pre_submit_jif ) | |
1716 | : req_write ? req_write->pre_submit_jif | |
1717 | : req_read ? req_read->pre_submit_jif : now; | |
7fde2be9 | 1718 | |
84d34f2f | 1719 | if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout)) |
9581f97a | 1720 | _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD); |
84d34f2f | 1721 | |
7753a4c1 LE |
1722 | if (dt && oldest_submit_jif != now && |
1723 | time_after(now, oldest_submit_jif + dt) && | |
b30ab791 | 1724 | !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) { |
d0180171 | 1725 | drbd_warn(device, "Local backing device failed to meet the disk-timeout\n"); |
b30ab791 | 1726 | __drbd_chk_io_error(device, DRBD_FORCE_DETACH); |
dfa8bedb | 1727 | } |
08535466 LE |
1728 | |
1729 | /* Reschedule timer for the nearest not already expired timeout. | |
1730 | * Fallback to now + min(effective network timeout, disk timeout). */ | |
7753a4c1 LE |
1731 | ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent)) |
1732 | ? req_peer->pre_send_jif + ent : now + et; | |
1733 | dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt)) | |
1734 | ? oldest_submit_jif + dt : now + et; | |
08535466 | 1735 | nt = time_before(ent, dt) ? ent : dt; |
7753a4c1 | 1736 | out: |
8d4ba3f0 | 1737 | spin_unlock_irq(&device->resource->req_lock); |
b30ab791 | 1738 | mod_timer(&device->request_timer, nt); |
7fde2be9 | 1739 | } |