xref: /openbmc/linux/drivers/block/drbd/drbd_worker.c (revision c4c3c32d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3    drbd_worker.c
4 
5    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6 
7    Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
8    Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
9    Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
10 
11 
12 */
13 
14 #include <linux/module.h>
15 #include <linux/drbd.h>
16 #include <linux/sched/signal.h>
17 #include <linux/wait.h>
18 #include <linux/mm.h>
19 #include <linux/memcontrol.h>
20 #include <linux/mm_inline.h>
21 #include <linux/slab.h>
22 #include <linux/random.h>
23 #include <linux/string.h>
24 #include <linux/scatterlist.h>
25 #include <linux/part_stat.h>
26 
27 #include "drbd_int.h"
28 #include "drbd_protocol.h"
29 #include "drbd_req.h"
30 
31 static int make_ov_request(struct drbd_peer_device *, int);
32 static int make_resync_request(struct drbd_peer_device *, int);
33 
34 /* endio handlers:
35  *   drbd_md_endio (defined here)
36  *   drbd_request_endio (defined here)
37  *   drbd_peer_request_endio (defined here)
38  *   drbd_bm_endio (defined in drbd_bitmap.c)
39  *
40  * For all these callbacks, note the following:
41  * The callbacks will be called in irq context by the IDE drivers,
42  * and in Softirqs/Tasklets/BH context by the SCSI drivers.
43  * Try to get the locking right :)
44  *
45  */
46 
47 /* used for synchronous meta data and bitmap IO
48  * submitted by drbd_md_sync_page_io()
49  */
50 void drbd_md_endio(struct bio *bio)
51 {
52 	struct drbd_device *device;
53 
54 	device = bio->bi_private;
55 	device->md_io.error = blk_status_to_errno(bio->bi_status);
56 
57 	/* special case: drbd_md_read() during drbd_adm_attach() */
58 	if (device->ldev)
59 		put_ldev(device);
60 	bio_put(bio);
61 
62 	/* We grabbed an extra reference in _drbd_md_sync_page_io() to be able
63 	 * to timeout on the lower level device, and eventually detach from it.
64 	 * If this io completion runs after that timeout expired, this
65 	 * drbd_md_put_buffer() may allow us to finally try and re-attach.
66 	 * During normal operation, this only puts that extra reference
67 	 * down to 1 again.
68 	 * Make sure we first drop the reference, and only then signal
69 	 * completion, or we may (in drbd_al_read_log()) cycle so fast into the
70 	 * next drbd_md_sync_page_io(), that we trigger the
71 	 * ASSERT(atomic_read(&device->md_io_in_use) == 1) there.
72 	 */
73 	drbd_md_put_buffer(device);
74 	device->md_io.done = 1;
75 	wake_up(&device->misc_wait);
76 }
77 
78 /* reads on behalf of the partner,
79  * "submitted" by the receiver
80  */
81 static void drbd_endio_read_sec_final(struct drbd_peer_request *peer_req) __releases(local)
82 {
83 	unsigned long flags = 0;
84 	struct drbd_peer_device *peer_device = peer_req->peer_device;
85 	struct drbd_device *device = peer_device->device;
86 
87 	spin_lock_irqsave(&device->resource->req_lock, flags);
88 	device->read_cnt += peer_req->i.size >> 9;
89 	list_del(&peer_req->w.list);
90 	if (list_empty(&device->read_ee))
91 		wake_up(&device->ee_wait);
92 	if (test_bit(__EE_WAS_ERROR, &peer_req->flags))
93 		__drbd_chk_io_error(device, DRBD_READ_ERROR);
94 	spin_unlock_irqrestore(&device->resource->req_lock, flags);
95 
96 	drbd_queue_work(&peer_device->connection->sender_work, &peer_req->w);
97 	put_ldev(device);
98 }
99 
100 /* writes on behalf of the partner, or resync writes,
101  * "submitted" by the receiver, final stage.  */
102 void drbd_endio_write_sec_final(struct drbd_peer_request *peer_req) __releases(local)
103 {
104 	unsigned long flags = 0;
105 	struct drbd_peer_device *peer_device = peer_req->peer_device;
106 	struct drbd_device *device = peer_device->device;
107 	struct drbd_connection *connection = peer_device->connection;
108 	struct drbd_interval i;
109 	int do_wake;
110 	u64 block_id;
111 	int do_al_complete_io;
112 
113 	/* after we moved peer_req to done_ee,
114 	 * we may no longer access it,
115 	 * it may be freed/reused already!
116 	 * (as soon as we release the req_lock) */
117 	i = peer_req->i;
118 	do_al_complete_io = peer_req->flags & EE_CALL_AL_COMPLETE_IO;
119 	block_id = peer_req->block_id;
120 	peer_req->flags &= ~EE_CALL_AL_COMPLETE_IO;
121 
122 	if (peer_req->flags & EE_WAS_ERROR) {
123 		/* In protocol != C, we usually do not send write acks.
124 		 * In case of a write error, send the neg ack anyways. */
125 		if (!__test_and_set_bit(__EE_SEND_WRITE_ACK, &peer_req->flags))
126 			inc_unacked(device);
127 		drbd_set_out_of_sync(peer_device, peer_req->i.sector, peer_req->i.size);
128 	}
129 
130 	spin_lock_irqsave(&device->resource->req_lock, flags);
131 	device->writ_cnt += peer_req->i.size >> 9;
132 	list_move_tail(&peer_req->w.list, &device->done_ee);
133 
134 	/*
135 	 * Do not remove from the write_requests tree here: we did not send the
136 	 * Ack yet and did not wake possibly waiting conflicting requests.
137 	 * Removed from the tree from "drbd_process_done_ee" within the
138 	 * appropriate dw.cb (e_end_block/e_end_resync_block) or from
139 	 * _drbd_clear_done_ee.
140 	 */
141 
142 	do_wake = list_empty(block_id == ID_SYNCER ? &device->sync_ee : &device->active_ee);
143 
144 	/* FIXME do we want to detach for failed REQ_OP_DISCARD?
145 	 * ((peer_req->flags & (EE_WAS_ERROR|EE_TRIM)) == EE_WAS_ERROR) */
146 	if (peer_req->flags & EE_WAS_ERROR)
147 		__drbd_chk_io_error(device, DRBD_WRITE_ERROR);
148 
149 	if (connection->cstate >= C_WF_REPORT_PARAMS) {
150 		kref_get(&device->kref); /* put is in drbd_send_acks_wf() */
151 		if (!queue_work(connection->ack_sender, &peer_device->send_acks_work))
152 			kref_put(&device->kref, drbd_destroy_device);
153 	}
154 	spin_unlock_irqrestore(&device->resource->req_lock, flags);
155 
156 	if (block_id == ID_SYNCER)
157 		drbd_rs_complete_io(device, i.sector);
158 
159 	if (do_wake)
160 		wake_up(&device->ee_wait);
161 
162 	if (do_al_complete_io)
163 		drbd_al_complete_io(device, &i);
164 
165 	put_ldev(device);
166 }
167 
168 /* writes on behalf of the partner, or resync writes,
169  * "submitted" by the receiver.
170  */
171 void drbd_peer_request_endio(struct bio *bio)
172 {
173 	struct drbd_peer_request *peer_req = bio->bi_private;
174 	struct drbd_device *device = peer_req->peer_device->device;
175 	bool is_write = bio_data_dir(bio) == WRITE;
176 	bool is_discard = bio_op(bio) == REQ_OP_WRITE_ZEROES ||
177 			  bio_op(bio) == REQ_OP_DISCARD;
178 
179 	if (bio->bi_status && drbd_ratelimit())
180 		drbd_warn(device, "%s: error=%d s=%llus\n",
181 				is_write ? (is_discard ? "discard" : "write")
182 					: "read", bio->bi_status,
183 				(unsigned long long)peer_req->i.sector);
184 
185 	if (bio->bi_status)
186 		set_bit(__EE_WAS_ERROR, &peer_req->flags);
187 
188 	bio_put(bio); /* no need for the bio anymore */
189 	if (atomic_dec_and_test(&peer_req->pending_bios)) {
190 		if (is_write)
191 			drbd_endio_write_sec_final(peer_req);
192 		else
193 			drbd_endio_read_sec_final(peer_req);
194 	}
195 }
196 
197 static void
198 drbd_panic_after_delayed_completion_of_aborted_request(struct drbd_device *device)
199 {
200 	panic("drbd%u %s/%u potential random memory corruption caused by delayed completion of aborted local request\n",
201 		device->minor, device->resource->name, device->vnr);
202 }
203 
204 /* read, readA or write requests on R_PRIMARY coming from drbd_make_request
205  */
206 void drbd_request_endio(struct bio *bio)
207 {
208 	unsigned long flags;
209 	struct drbd_request *req = bio->bi_private;
210 	struct drbd_device *device = req->device;
211 	struct bio_and_error m;
212 	enum drbd_req_event what;
213 
214 	/* If this request was aborted locally before,
215 	 * but now was completed "successfully",
216 	 * chances are that this caused arbitrary data corruption.
217 	 *
218 	 * "aborting" requests, or force-detaching the disk, is intended for
219 	 * completely blocked/hung local backing devices which do no longer
220 	 * complete requests at all, not even do error completions.  In this
221 	 * situation, usually a hard-reset and failover is the only way out.
222 	 *
223 	 * By "aborting", basically faking a local error-completion,
224 	 * we allow for a more graceful swichover by cleanly migrating services.
225 	 * Still the affected node has to be rebooted "soon".
226 	 *
227 	 * By completing these requests, we allow the upper layers to re-use
228 	 * the associated data pages.
229 	 *
230 	 * If later the local backing device "recovers", and now DMAs some data
231 	 * from disk into the original request pages, in the best case it will
232 	 * just put random data into unused pages; but typically it will corrupt
233 	 * meanwhile completely unrelated data, causing all sorts of damage.
234 	 *
235 	 * Which means delayed successful completion,
236 	 * especially for READ requests,
237 	 * is a reason to panic().
238 	 *
239 	 * We assume that a delayed *error* completion is OK,
240 	 * though we still will complain noisily about it.
241 	 */
242 	if (unlikely(req->rq_state & RQ_LOCAL_ABORTED)) {
243 		if (drbd_ratelimit())
244 			drbd_emerg(device, "delayed completion of aborted local request; disk-timeout may be too aggressive\n");
245 
246 		if (!bio->bi_status)
247 			drbd_panic_after_delayed_completion_of_aborted_request(device);
248 	}
249 
250 	/* to avoid recursion in __req_mod */
251 	if (unlikely(bio->bi_status)) {
252 		switch (bio_op(bio)) {
253 		case REQ_OP_WRITE_ZEROES:
254 		case REQ_OP_DISCARD:
255 			if (bio->bi_status == BLK_STS_NOTSUPP)
256 				what = DISCARD_COMPLETED_NOTSUPP;
257 			else
258 				what = DISCARD_COMPLETED_WITH_ERROR;
259 			break;
260 		case REQ_OP_READ:
261 			if (bio->bi_opf & REQ_RAHEAD)
262 				what = READ_AHEAD_COMPLETED_WITH_ERROR;
263 			else
264 				what = READ_COMPLETED_WITH_ERROR;
265 			break;
266 		default:
267 			what = WRITE_COMPLETED_WITH_ERROR;
268 			break;
269 		}
270 	} else {
271 		what = COMPLETED_OK;
272 	}
273 
274 	req->private_bio = ERR_PTR(blk_status_to_errno(bio->bi_status));
275 	bio_put(bio);
276 
277 	/* not req_mod(), we need irqsave here! */
278 	spin_lock_irqsave(&device->resource->req_lock, flags);
279 	__req_mod(req, what, NULL, &m);
280 	spin_unlock_irqrestore(&device->resource->req_lock, flags);
281 	put_ldev(device);
282 
283 	if (m.bio)
284 		complete_master_bio(device, &m);
285 }
286 
287 void drbd_csum_ee(struct crypto_shash *tfm, struct drbd_peer_request *peer_req, void *digest)
288 {
289 	SHASH_DESC_ON_STACK(desc, tfm);
290 	struct page *page = peer_req->pages;
291 	struct page *tmp;
292 	unsigned len;
293 	void *src;
294 
295 	desc->tfm = tfm;
296 
297 	crypto_shash_init(desc);
298 
299 	src = kmap_atomic(page);
300 	while ((tmp = page_chain_next(page))) {
301 		/* all but the last page will be fully used */
302 		crypto_shash_update(desc, src, PAGE_SIZE);
303 		kunmap_atomic(src);
304 		page = tmp;
305 		src = kmap_atomic(page);
306 	}
307 	/* and now the last, possibly only partially used page */
308 	len = peer_req->i.size & (PAGE_SIZE - 1);
309 	crypto_shash_update(desc, src, len ?: PAGE_SIZE);
310 	kunmap_atomic(src);
311 
312 	crypto_shash_final(desc, digest);
313 	shash_desc_zero(desc);
314 }
315 
316 void drbd_csum_bio(struct crypto_shash *tfm, struct bio *bio, void *digest)
317 {
318 	SHASH_DESC_ON_STACK(desc, tfm);
319 	struct bio_vec bvec;
320 	struct bvec_iter iter;
321 
322 	desc->tfm = tfm;
323 
324 	crypto_shash_init(desc);
325 
326 	bio_for_each_segment(bvec, bio, iter) {
327 		u8 *src;
328 
329 		src = bvec_kmap_local(&bvec);
330 		crypto_shash_update(desc, src, bvec.bv_len);
331 		kunmap_local(src);
332 	}
333 	crypto_shash_final(desc, digest);
334 	shash_desc_zero(desc);
335 }
336 
337 /* MAYBE merge common code with w_e_end_ov_req */
338 static int w_e_send_csum(struct drbd_work *w, int cancel)
339 {
340 	struct drbd_peer_request *peer_req = container_of(w, struct drbd_peer_request, w);
341 	struct drbd_peer_device *peer_device = peer_req->peer_device;
342 	struct drbd_device *device = peer_device->device;
343 	int digest_size;
344 	void *digest;
345 	int err = 0;
346 
347 	if (unlikely(cancel))
348 		goto out;
349 
350 	if (unlikely((peer_req->flags & EE_WAS_ERROR) != 0))
351 		goto out;
352 
353 	digest_size = crypto_shash_digestsize(peer_device->connection->csums_tfm);
354 	digest = kmalloc(digest_size, GFP_NOIO);
355 	if (digest) {
356 		sector_t sector = peer_req->i.sector;
357 		unsigned int size = peer_req->i.size;
358 		drbd_csum_ee(peer_device->connection->csums_tfm, peer_req, digest);
359 		/* Free peer_req and pages before send.
360 		 * In case we block on congestion, we could otherwise run into
361 		 * some distributed deadlock, if the other side blocks on
362 		 * congestion as well, because our receiver blocks in
363 		 * drbd_alloc_pages due to pp_in_use > max_buffers. */
364 		drbd_free_peer_req(device, peer_req);
365 		peer_req = NULL;
366 		inc_rs_pending(peer_device);
367 		err = drbd_send_drequest_csum(peer_device, sector, size,
368 					      digest, digest_size,
369 					      P_CSUM_RS_REQUEST);
370 		kfree(digest);
371 	} else {
372 		drbd_err(device, "kmalloc() of digest failed.\n");
373 		err = -ENOMEM;
374 	}
375 
376 out:
377 	if (peer_req)
378 		drbd_free_peer_req(device, peer_req);
379 
380 	if (unlikely(err))
381 		drbd_err(device, "drbd_send_drequest(..., csum) failed\n");
382 	return err;
383 }
384 
385 #define GFP_TRY	(__GFP_HIGHMEM | __GFP_NOWARN)
386 
387 static int read_for_csum(struct drbd_peer_device *peer_device, sector_t sector, int size)
388 {
389 	struct drbd_device *device = peer_device->device;
390 	struct drbd_peer_request *peer_req;
391 
392 	if (!get_ldev(device))
393 		return -EIO;
394 
395 	/* GFP_TRY, because if there is no memory available right now, this may
396 	 * be rescheduled for later. It is "only" background resync, after all. */
397 	peer_req = drbd_alloc_peer_req(peer_device, ID_SYNCER /* unused */, sector,
398 				       size, size, GFP_TRY);
399 	if (!peer_req)
400 		goto defer;
401 
402 	peer_req->w.cb = w_e_send_csum;
403 	peer_req->opf = REQ_OP_READ;
404 	spin_lock_irq(&device->resource->req_lock);
405 	list_add_tail(&peer_req->w.list, &device->read_ee);
406 	spin_unlock_irq(&device->resource->req_lock);
407 
408 	atomic_add(size >> 9, &device->rs_sect_ev);
409 	if (drbd_submit_peer_request(peer_req) == 0)
410 		return 0;
411 
412 	/* If it failed because of ENOMEM, retry should help.  If it failed
413 	 * because bio_add_page failed (probably broken lower level driver),
414 	 * retry may or may not help.
415 	 * If it does not, you may need to force disconnect. */
416 	spin_lock_irq(&device->resource->req_lock);
417 	list_del(&peer_req->w.list);
418 	spin_unlock_irq(&device->resource->req_lock);
419 
420 	drbd_free_peer_req(device, peer_req);
421 defer:
422 	put_ldev(device);
423 	return -EAGAIN;
424 }
425 
426 int w_resync_timer(struct drbd_work *w, int cancel)
427 {
428 	struct drbd_device *device =
429 		container_of(w, struct drbd_device, resync_work);
430 
431 	switch (device->state.conn) {
432 	case C_VERIFY_S:
433 		make_ov_request(first_peer_device(device), cancel);
434 		break;
435 	case C_SYNC_TARGET:
436 		make_resync_request(first_peer_device(device), cancel);
437 		break;
438 	}
439 
440 	return 0;
441 }
442 
443 void resync_timer_fn(struct timer_list *t)
444 {
445 	struct drbd_device *device = from_timer(device, t, resync_timer);
446 
447 	drbd_queue_work_if_unqueued(
448 		&first_peer_device(device)->connection->sender_work,
449 		&device->resync_work);
450 }
451 
452 static void fifo_set(struct fifo_buffer *fb, int value)
453 {
454 	int i;
455 
456 	for (i = 0; i < fb->size; i++)
457 		fb->values[i] = value;
458 }
459 
460 static int fifo_push(struct fifo_buffer *fb, int value)
461 {
462 	int ov;
463 
464 	ov = fb->values[fb->head_index];
465 	fb->values[fb->head_index++] = value;
466 
467 	if (fb->head_index >= fb->size)
468 		fb->head_index = 0;
469 
470 	return ov;
471 }
472 
473 static void fifo_add_val(struct fifo_buffer *fb, int value)
474 {
475 	int i;
476 
477 	for (i = 0; i < fb->size; i++)
478 		fb->values[i] += value;
479 }
480 
481 struct fifo_buffer *fifo_alloc(unsigned int fifo_size)
482 {
483 	struct fifo_buffer *fb;
484 
485 	fb = kzalloc(struct_size(fb, values, fifo_size), GFP_NOIO);
486 	if (!fb)
487 		return NULL;
488 
489 	fb->head_index = 0;
490 	fb->size = fifo_size;
491 	fb->total = 0;
492 
493 	return fb;
494 }
495 
496 static int drbd_rs_controller(struct drbd_peer_device *peer_device, unsigned int sect_in)
497 {
498 	struct drbd_device *device = peer_device->device;
499 	struct disk_conf *dc;
500 	unsigned int want;     /* The number of sectors we want in-flight */
501 	int req_sect; /* Number of sectors to request in this turn */
502 	int correction; /* Number of sectors more we need in-flight */
503 	int cps; /* correction per invocation of drbd_rs_controller() */
504 	int steps; /* Number of time steps to plan ahead */
505 	int curr_corr;
506 	int max_sect;
507 	struct fifo_buffer *plan;
508 
509 	dc = rcu_dereference(device->ldev->disk_conf);
510 	plan = rcu_dereference(device->rs_plan_s);
511 
512 	steps = plan->size; /* (dc->c_plan_ahead * 10 * SLEEP_TIME) / HZ; */
513 
514 	if (device->rs_in_flight + sect_in == 0) { /* At start of resync */
515 		want = ((dc->resync_rate * 2 * SLEEP_TIME) / HZ) * steps;
516 	} else { /* normal path */
517 		want = dc->c_fill_target ? dc->c_fill_target :
518 			sect_in * dc->c_delay_target * HZ / (SLEEP_TIME * 10);
519 	}
520 
521 	correction = want - device->rs_in_flight - plan->total;
522 
523 	/* Plan ahead */
524 	cps = correction / steps;
525 	fifo_add_val(plan, cps);
526 	plan->total += cps * steps;
527 
528 	/* What we do in this step */
529 	curr_corr = fifo_push(plan, 0);
530 	plan->total -= curr_corr;
531 
532 	req_sect = sect_in + curr_corr;
533 	if (req_sect < 0)
534 		req_sect = 0;
535 
536 	max_sect = (dc->c_max_rate * 2 * SLEEP_TIME) / HZ;
537 	if (req_sect > max_sect)
538 		req_sect = max_sect;
539 
540 	/*
541 	drbd_warn(device, "si=%u if=%d wa=%u co=%d st=%d cps=%d pl=%d cc=%d rs=%d\n",
542 		 sect_in, device->rs_in_flight, want, correction,
543 		 steps, cps, device->rs_planed, curr_corr, req_sect);
544 	*/
545 
546 	return req_sect;
547 }
548 
549 static int drbd_rs_number_requests(struct drbd_peer_device *peer_device)
550 {
551 	struct drbd_device *device = peer_device->device;
552 	unsigned int sect_in;  /* Number of sectors that came in since the last turn */
553 	int number, mxb;
554 
555 	sect_in = atomic_xchg(&device->rs_sect_in, 0);
556 	device->rs_in_flight -= sect_in;
557 
558 	rcu_read_lock();
559 	mxb = drbd_get_max_buffers(device) / 2;
560 	if (rcu_dereference(device->rs_plan_s)->size) {
561 		number = drbd_rs_controller(peer_device, sect_in) >> (BM_BLOCK_SHIFT - 9);
562 		device->c_sync_rate = number * HZ * (BM_BLOCK_SIZE / 1024) / SLEEP_TIME;
563 	} else {
564 		device->c_sync_rate = rcu_dereference(device->ldev->disk_conf)->resync_rate;
565 		number = SLEEP_TIME * device->c_sync_rate  / ((BM_BLOCK_SIZE / 1024) * HZ);
566 	}
567 	rcu_read_unlock();
568 
569 	/* Don't have more than "max-buffers"/2 in-flight.
570 	 * Otherwise we may cause the remote site to stall on drbd_alloc_pages(),
571 	 * potentially causing a distributed deadlock on congestion during
572 	 * online-verify or (checksum-based) resync, if max-buffers,
573 	 * socket buffer sizes and resync rate settings are mis-configured. */
574 
575 	/* note that "number" is in units of "BM_BLOCK_SIZE" (which is 4k),
576 	 * mxb (as used here, and in drbd_alloc_pages on the peer) is
577 	 * "number of pages" (typically also 4k),
578 	 * but "rs_in_flight" is in "sectors" (512 Byte). */
579 	if (mxb - device->rs_in_flight/8 < number)
580 		number = mxb - device->rs_in_flight/8;
581 
582 	return number;
583 }
584 
585 static int make_resync_request(struct drbd_peer_device *const peer_device, int cancel)
586 {
587 	struct drbd_device *const device = peer_device->device;
588 	struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
589 	unsigned long bit;
590 	sector_t sector;
591 	const sector_t capacity = get_capacity(device->vdisk);
592 	int max_bio_size;
593 	int number, rollback_i, size;
594 	int align, requeue = 0;
595 	int i = 0;
596 	int discard_granularity = 0;
597 
598 	if (unlikely(cancel))
599 		return 0;
600 
601 	if (device->rs_total == 0) {
602 		/* empty resync? */
603 		drbd_resync_finished(peer_device);
604 		return 0;
605 	}
606 
607 	if (!get_ldev(device)) {
608 		/* Since we only need to access device->rsync a
609 		   get_ldev_if_state(device,D_FAILED) would be sufficient, but
610 		   to continue resync with a broken disk makes no sense at
611 		   all */
612 		drbd_err(device, "Disk broke down during resync!\n");
613 		return 0;
614 	}
615 
616 	if (connection->agreed_features & DRBD_FF_THIN_RESYNC) {
617 		rcu_read_lock();
618 		discard_granularity = rcu_dereference(device->ldev->disk_conf)->rs_discard_granularity;
619 		rcu_read_unlock();
620 	}
621 
622 	max_bio_size = queue_max_hw_sectors(device->rq_queue) << 9;
623 	number = drbd_rs_number_requests(peer_device);
624 	if (number <= 0)
625 		goto requeue;
626 
627 	for (i = 0; i < number; i++) {
628 		/* Stop generating RS requests when half of the send buffer is filled,
629 		 * but notify TCP that we'd like to have more space. */
630 		mutex_lock(&connection->data.mutex);
631 		if (connection->data.socket) {
632 			struct sock *sk = connection->data.socket->sk;
633 			int queued = sk->sk_wmem_queued;
634 			int sndbuf = sk->sk_sndbuf;
635 			if (queued > sndbuf / 2) {
636 				requeue = 1;
637 				if (sk->sk_socket)
638 					set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
639 			}
640 		} else
641 			requeue = 1;
642 		mutex_unlock(&connection->data.mutex);
643 		if (requeue)
644 			goto requeue;
645 
646 next_sector:
647 		size = BM_BLOCK_SIZE;
648 		bit  = drbd_bm_find_next(device, device->bm_resync_fo);
649 
650 		if (bit == DRBD_END_OF_BITMAP) {
651 			device->bm_resync_fo = drbd_bm_bits(device);
652 			put_ldev(device);
653 			return 0;
654 		}
655 
656 		sector = BM_BIT_TO_SECT(bit);
657 
658 		if (drbd_try_rs_begin_io(peer_device, sector)) {
659 			device->bm_resync_fo = bit;
660 			goto requeue;
661 		}
662 		device->bm_resync_fo = bit + 1;
663 
664 		if (unlikely(drbd_bm_test_bit(device, bit) == 0)) {
665 			drbd_rs_complete_io(device, sector);
666 			goto next_sector;
667 		}
668 
669 #if DRBD_MAX_BIO_SIZE > BM_BLOCK_SIZE
670 		/* try to find some adjacent bits.
671 		 * we stop if we have already the maximum req size.
672 		 *
673 		 * Additionally always align bigger requests, in order to
674 		 * be prepared for all stripe sizes of software RAIDs.
675 		 */
676 		align = 1;
677 		rollback_i = i;
678 		while (i < number) {
679 			if (size + BM_BLOCK_SIZE > max_bio_size)
680 				break;
681 
682 			/* Be always aligned */
683 			if (sector & ((1<<(align+3))-1))
684 				break;
685 
686 			if (discard_granularity && size == discard_granularity)
687 				break;
688 
689 			/* do not cross extent boundaries */
690 			if (((bit+1) & BM_BLOCKS_PER_BM_EXT_MASK) == 0)
691 				break;
692 			/* now, is it actually dirty, after all?
693 			 * caution, drbd_bm_test_bit is tri-state for some
694 			 * obscure reason; ( b == 0 ) would get the out-of-band
695 			 * only accidentally right because of the "oddly sized"
696 			 * adjustment below */
697 			if (drbd_bm_test_bit(device, bit+1) != 1)
698 				break;
699 			bit++;
700 			size += BM_BLOCK_SIZE;
701 			if ((BM_BLOCK_SIZE << align) <= size)
702 				align++;
703 			i++;
704 		}
705 		/* if we merged some,
706 		 * reset the offset to start the next drbd_bm_find_next from */
707 		if (size > BM_BLOCK_SIZE)
708 			device->bm_resync_fo = bit + 1;
709 #endif
710 
711 		/* adjust very last sectors, in case we are oddly sized */
712 		if (sector + (size>>9) > capacity)
713 			size = (capacity-sector)<<9;
714 
715 		if (device->use_csums) {
716 			switch (read_for_csum(peer_device, sector, size)) {
717 			case -EIO: /* Disk failure */
718 				put_ldev(device);
719 				return -EIO;
720 			case -EAGAIN: /* allocation failed, or ldev busy */
721 				drbd_rs_complete_io(device, sector);
722 				device->bm_resync_fo = BM_SECT_TO_BIT(sector);
723 				i = rollback_i;
724 				goto requeue;
725 			case 0:
726 				/* everything ok */
727 				break;
728 			default:
729 				BUG();
730 			}
731 		} else {
732 			int err;
733 
734 			inc_rs_pending(peer_device);
735 			err = drbd_send_drequest(peer_device,
736 						 size == discard_granularity ? P_RS_THIN_REQ : P_RS_DATA_REQUEST,
737 						 sector, size, ID_SYNCER);
738 			if (err) {
739 				drbd_err(device, "drbd_send_drequest() failed, aborting...\n");
740 				dec_rs_pending(peer_device);
741 				put_ldev(device);
742 				return err;
743 			}
744 		}
745 	}
746 
747 	if (device->bm_resync_fo >= drbd_bm_bits(device)) {
748 		/* last syncer _request_ was sent,
749 		 * but the P_RS_DATA_REPLY not yet received.  sync will end (and
750 		 * next sync group will resume), as soon as we receive the last
751 		 * resync data block, and the last bit is cleared.
752 		 * until then resync "work" is "inactive" ...
753 		 */
754 		put_ldev(device);
755 		return 0;
756 	}
757 
758  requeue:
759 	device->rs_in_flight += (i << (BM_BLOCK_SHIFT - 9));
760 	mod_timer(&device->resync_timer, jiffies + SLEEP_TIME);
761 	put_ldev(device);
762 	return 0;
763 }
764 
765 static int make_ov_request(struct drbd_peer_device *peer_device, int cancel)
766 {
767 	struct drbd_device *device = peer_device->device;
768 	int number, i, size;
769 	sector_t sector;
770 	const sector_t capacity = get_capacity(device->vdisk);
771 	bool stop_sector_reached = false;
772 
773 	if (unlikely(cancel))
774 		return 1;
775 
776 	number = drbd_rs_number_requests(peer_device);
777 
778 	sector = device->ov_position;
779 	for (i = 0; i < number; i++) {
780 		if (sector >= capacity)
781 			return 1;
782 
783 		/* We check for "finished" only in the reply path:
784 		 * w_e_end_ov_reply().
785 		 * We need to send at least one request out. */
786 		stop_sector_reached = i > 0
787 			&& verify_can_do_stop_sector(device)
788 			&& sector >= device->ov_stop_sector;
789 		if (stop_sector_reached)
790 			break;
791 
792 		size = BM_BLOCK_SIZE;
793 
794 		if (drbd_try_rs_begin_io(peer_device, sector)) {
795 			device->ov_position = sector;
796 			goto requeue;
797 		}
798 
799 		if (sector + (size>>9) > capacity)
800 			size = (capacity-sector)<<9;
801 
802 		inc_rs_pending(peer_device);
803 		if (drbd_send_ov_request(first_peer_device(device), sector, size)) {
804 			dec_rs_pending(peer_device);
805 			return 0;
806 		}
807 		sector += BM_SECT_PER_BIT;
808 	}
809 	device->ov_position = sector;
810 
811  requeue:
812 	device->rs_in_flight += (i << (BM_BLOCK_SHIFT - 9));
813 	if (i == 0 || !stop_sector_reached)
814 		mod_timer(&device->resync_timer, jiffies + SLEEP_TIME);
815 	return 1;
816 }
817 
818 int w_ov_finished(struct drbd_work *w, int cancel)
819 {
820 	struct drbd_device_work *dw =
821 		container_of(w, struct drbd_device_work, w);
822 	struct drbd_device *device = dw->device;
823 	kfree(dw);
824 	ov_out_of_sync_print(first_peer_device(device));
825 	drbd_resync_finished(first_peer_device(device));
826 
827 	return 0;
828 }
829 
830 static int w_resync_finished(struct drbd_work *w, int cancel)
831 {
832 	struct drbd_device_work *dw =
833 		container_of(w, struct drbd_device_work, w);
834 	struct drbd_device *device = dw->device;
835 	kfree(dw);
836 
837 	drbd_resync_finished(first_peer_device(device));
838 
839 	return 0;
840 }
841 
842 static void ping_peer(struct drbd_device *device)
843 {
844 	struct drbd_connection *connection = first_peer_device(device)->connection;
845 
846 	clear_bit(GOT_PING_ACK, &connection->flags);
847 	request_ping(connection);
848 	wait_event(connection->ping_wait,
849 		   test_bit(GOT_PING_ACK, &connection->flags) || device->state.conn < C_CONNECTED);
850 }
851 
852 int drbd_resync_finished(struct drbd_peer_device *peer_device)
853 {
854 	struct drbd_device *device = peer_device->device;
855 	struct drbd_connection *connection = peer_device->connection;
856 	unsigned long db, dt, dbdt;
857 	unsigned long n_oos;
858 	union drbd_state os, ns;
859 	struct drbd_device_work *dw;
860 	char *khelper_cmd = NULL;
861 	int verify_done = 0;
862 
863 	/* Remove all elements from the resync LRU. Since future actions
864 	 * might set bits in the (main) bitmap, then the entries in the
865 	 * resync LRU would be wrong. */
866 	if (drbd_rs_del_all(device)) {
867 		/* In case this is not possible now, most probably because
868 		 * there are P_RS_DATA_REPLY Packets lingering on the worker's
869 		 * queue (or even the read operations for those packets
870 		 * is not finished by now).   Retry in 100ms. */
871 
872 		schedule_timeout_interruptible(HZ / 10);
873 		dw = kmalloc(sizeof(struct drbd_device_work), GFP_ATOMIC);
874 		if (dw) {
875 			dw->w.cb = w_resync_finished;
876 			dw->device = device;
877 			drbd_queue_work(&connection->sender_work, &dw->w);
878 			return 1;
879 		}
880 		drbd_err(device, "Warn failed to drbd_rs_del_all() and to kmalloc(dw).\n");
881 	}
882 
883 	dt = (jiffies - device->rs_start - device->rs_paused) / HZ;
884 	if (dt <= 0)
885 		dt = 1;
886 
887 	db = device->rs_total;
888 	/* adjust for verify start and stop sectors, respective reached position */
889 	if (device->state.conn == C_VERIFY_S || device->state.conn == C_VERIFY_T)
890 		db -= device->ov_left;
891 
892 	dbdt = Bit2KB(db/dt);
893 	device->rs_paused /= HZ;
894 
895 	if (!get_ldev(device))
896 		goto out;
897 
898 	ping_peer(device);
899 
900 	spin_lock_irq(&device->resource->req_lock);
901 	os = drbd_read_state(device);
902 
903 	verify_done = (os.conn == C_VERIFY_S || os.conn == C_VERIFY_T);
904 
905 	/* This protects us against multiple calls (that can happen in the presence
906 	   of application IO), and against connectivity loss just before we arrive here. */
907 	if (os.conn <= C_CONNECTED)
908 		goto out_unlock;
909 
910 	ns = os;
911 	ns.conn = C_CONNECTED;
912 
913 	drbd_info(device, "%s done (total %lu sec; paused %lu sec; %lu K/sec)\n",
914 	     verify_done ? "Online verify" : "Resync",
915 	     dt + device->rs_paused, device->rs_paused, dbdt);
916 
917 	n_oos = drbd_bm_total_weight(device);
918 
919 	if (os.conn == C_VERIFY_S || os.conn == C_VERIFY_T) {
920 		if (n_oos) {
921 			drbd_alert(device, "Online verify found %lu %dk block out of sync!\n",
922 			      n_oos, Bit2KB(1));
923 			khelper_cmd = "out-of-sync";
924 		}
925 	} else {
926 		D_ASSERT(device, (n_oos - device->rs_failed) == 0);
927 
928 		if (os.conn == C_SYNC_TARGET || os.conn == C_PAUSED_SYNC_T)
929 			khelper_cmd = "after-resync-target";
930 
931 		if (device->use_csums && device->rs_total) {
932 			const unsigned long s = device->rs_same_csum;
933 			const unsigned long t = device->rs_total;
934 			const int ratio =
935 				(t == 0)     ? 0 :
936 			(t < 100000) ? ((s*100)/t) : (s/(t/100));
937 			drbd_info(device, "%u %% had equal checksums, eliminated: %luK; "
938 			     "transferred %luK total %luK\n",
939 			     ratio,
940 			     Bit2KB(device->rs_same_csum),
941 			     Bit2KB(device->rs_total - device->rs_same_csum),
942 			     Bit2KB(device->rs_total));
943 		}
944 	}
945 
946 	if (device->rs_failed) {
947 		drbd_info(device, "            %lu failed blocks\n", device->rs_failed);
948 
949 		if (os.conn == C_SYNC_TARGET || os.conn == C_PAUSED_SYNC_T) {
950 			ns.disk = D_INCONSISTENT;
951 			ns.pdsk = D_UP_TO_DATE;
952 		} else {
953 			ns.disk = D_UP_TO_DATE;
954 			ns.pdsk = D_INCONSISTENT;
955 		}
956 	} else {
957 		ns.disk = D_UP_TO_DATE;
958 		ns.pdsk = D_UP_TO_DATE;
959 
960 		if (os.conn == C_SYNC_TARGET || os.conn == C_PAUSED_SYNC_T) {
961 			if (device->p_uuid) {
962 				int i;
963 				for (i = UI_BITMAP ; i <= UI_HISTORY_END ; i++)
964 					_drbd_uuid_set(device, i, device->p_uuid[i]);
965 				drbd_uuid_set(device, UI_BITMAP, device->ldev->md.uuid[UI_CURRENT]);
966 				_drbd_uuid_set(device, UI_CURRENT, device->p_uuid[UI_CURRENT]);
967 			} else {
968 				drbd_err(device, "device->p_uuid is NULL! BUG\n");
969 			}
970 		}
971 
972 		if (!(os.conn == C_VERIFY_S || os.conn == C_VERIFY_T)) {
973 			/* for verify runs, we don't update uuids here,
974 			 * so there would be nothing to report. */
975 			drbd_uuid_set_bm(device, 0UL);
976 			drbd_print_uuids(device, "updated UUIDs");
977 			if (device->p_uuid) {
978 				/* Now the two UUID sets are equal, update what we
979 				 * know of the peer. */
980 				int i;
981 				for (i = UI_CURRENT ; i <= UI_HISTORY_END ; i++)
982 					device->p_uuid[i] = device->ldev->md.uuid[i];
983 			}
984 		}
985 	}
986 
987 	_drbd_set_state(device, ns, CS_VERBOSE, NULL);
988 out_unlock:
989 	spin_unlock_irq(&device->resource->req_lock);
990 
991 	/* If we have been sync source, and have an effective fencing-policy,
992 	 * once *all* volumes are back in sync, call "unfence". */
993 	if (os.conn == C_SYNC_SOURCE) {
994 		enum drbd_disk_state disk_state = D_MASK;
995 		enum drbd_disk_state pdsk_state = D_MASK;
996 		enum drbd_fencing_p fp = FP_DONT_CARE;
997 
998 		rcu_read_lock();
999 		fp = rcu_dereference(device->ldev->disk_conf)->fencing;
1000 		if (fp != FP_DONT_CARE) {
1001 			struct drbd_peer_device *peer_device;
1002 			int vnr;
1003 			idr_for_each_entry(&connection->peer_devices, peer_device, vnr) {
1004 				struct drbd_device *device = peer_device->device;
1005 				disk_state = min_t(enum drbd_disk_state, disk_state, device->state.disk);
1006 				pdsk_state = min_t(enum drbd_disk_state, pdsk_state, device->state.pdsk);
1007 			}
1008 		}
1009 		rcu_read_unlock();
1010 		if (disk_state == D_UP_TO_DATE && pdsk_state == D_UP_TO_DATE)
1011 			conn_khelper(connection, "unfence-peer");
1012 	}
1013 
1014 	put_ldev(device);
1015 out:
1016 	device->rs_total  = 0;
1017 	device->rs_failed = 0;
1018 	device->rs_paused = 0;
1019 
1020 	/* reset start sector, if we reached end of device */
1021 	if (verify_done && device->ov_left == 0)
1022 		device->ov_start_sector = 0;
1023 
1024 	drbd_md_sync(device);
1025 
1026 	if (khelper_cmd)
1027 		drbd_khelper(device, khelper_cmd);
1028 
1029 	return 1;
1030 }
1031 
1032 /* helper */
1033 static void move_to_net_ee_or_free(struct drbd_device *device, struct drbd_peer_request *peer_req)
1034 {
1035 	if (drbd_peer_req_has_active_page(peer_req)) {
1036 		/* This might happen if sendpage() has not finished */
1037 		int i = PFN_UP(peer_req->i.size);
1038 		atomic_add(i, &device->pp_in_use_by_net);
1039 		atomic_sub(i, &device->pp_in_use);
1040 		spin_lock_irq(&device->resource->req_lock);
1041 		list_add_tail(&peer_req->w.list, &device->net_ee);
1042 		spin_unlock_irq(&device->resource->req_lock);
1043 		wake_up(&drbd_pp_wait);
1044 	} else
1045 		drbd_free_peer_req(device, peer_req);
1046 }
1047 
1048 /**
1049  * w_e_end_data_req() - Worker callback, to send a P_DATA_REPLY packet in response to a P_DATA_REQUEST
1050  * @w:		work object.
1051  * @cancel:	The connection will be closed anyways
1052  */
1053 int w_e_end_data_req(struct drbd_work *w, int cancel)
1054 {
1055 	struct drbd_peer_request *peer_req = container_of(w, struct drbd_peer_request, w);
1056 	struct drbd_peer_device *peer_device = peer_req->peer_device;
1057 	struct drbd_device *device = peer_device->device;
1058 	int err;
1059 
1060 	if (unlikely(cancel)) {
1061 		drbd_free_peer_req(device, peer_req);
1062 		dec_unacked(device);
1063 		return 0;
1064 	}
1065 
1066 	if (likely((peer_req->flags & EE_WAS_ERROR) == 0)) {
1067 		err = drbd_send_block(peer_device, P_DATA_REPLY, peer_req);
1068 	} else {
1069 		if (drbd_ratelimit())
1070 			drbd_err(device, "Sending NegDReply. sector=%llus.\n",
1071 			    (unsigned long long)peer_req->i.sector);
1072 
1073 		err = drbd_send_ack(peer_device, P_NEG_DREPLY, peer_req);
1074 	}
1075 
1076 	dec_unacked(device);
1077 
1078 	move_to_net_ee_or_free(device, peer_req);
1079 
1080 	if (unlikely(err))
1081 		drbd_err(device, "drbd_send_block() failed\n");
1082 	return err;
1083 }
1084 
1085 static bool all_zero(struct drbd_peer_request *peer_req)
1086 {
1087 	struct page *page = peer_req->pages;
1088 	unsigned int len = peer_req->i.size;
1089 
1090 	page_chain_for_each(page) {
1091 		unsigned int l = min_t(unsigned int, len, PAGE_SIZE);
1092 		unsigned int i, words = l / sizeof(long);
1093 		unsigned long *d;
1094 
1095 		d = kmap_atomic(page);
1096 		for (i = 0; i < words; i++) {
1097 			if (d[i]) {
1098 				kunmap_atomic(d);
1099 				return false;
1100 			}
1101 		}
1102 		kunmap_atomic(d);
1103 		len -= l;
1104 	}
1105 
1106 	return true;
1107 }
1108 
1109 /**
1110  * w_e_end_rsdata_req() - Worker callback to send a P_RS_DATA_REPLY packet in response to a P_RS_DATA_REQUEST
1111  * @w:		work object.
1112  * @cancel:	The connection will be closed anyways
1113  */
1114 int w_e_end_rsdata_req(struct drbd_work *w, int cancel)
1115 {
1116 	struct drbd_peer_request *peer_req = container_of(w, struct drbd_peer_request, w);
1117 	struct drbd_peer_device *peer_device = peer_req->peer_device;
1118 	struct drbd_device *device = peer_device->device;
1119 	int err;
1120 
1121 	if (unlikely(cancel)) {
1122 		drbd_free_peer_req(device, peer_req);
1123 		dec_unacked(device);
1124 		return 0;
1125 	}
1126 
1127 	if (get_ldev_if_state(device, D_FAILED)) {
1128 		drbd_rs_complete_io(device, peer_req->i.sector);
1129 		put_ldev(device);
1130 	}
1131 
1132 	if (device->state.conn == C_AHEAD) {
1133 		err = drbd_send_ack(peer_device, P_RS_CANCEL, peer_req);
1134 	} else if (likely((peer_req->flags & EE_WAS_ERROR) == 0)) {
1135 		if (likely(device->state.pdsk >= D_INCONSISTENT)) {
1136 			inc_rs_pending(peer_device);
1137 			if (peer_req->flags & EE_RS_THIN_REQ && all_zero(peer_req))
1138 				err = drbd_send_rs_deallocated(peer_device, peer_req);
1139 			else
1140 				err = drbd_send_block(peer_device, P_RS_DATA_REPLY, peer_req);
1141 		} else {
1142 			if (drbd_ratelimit())
1143 				drbd_err(device, "Not sending RSDataReply, "
1144 				    "partner DISKLESS!\n");
1145 			err = 0;
1146 		}
1147 	} else {
1148 		if (drbd_ratelimit())
1149 			drbd_err(device, "Sending NegRSDReply. sector %llus.\n",
1150 			    (unsigned long long)peer_req->i.sector);
1151 
1152 		err = drbd_send_ack(peer_device, P_NEG_RS_DREPLY, peer_req);
1153 
1154 		/* update resync data with failure */
1155 		drbd_rs_failed_io(peer_device, peer_req->i.sector, peer_req->i.size);
1156 	}
1157 
1158 	dec_unacked(device);
1159 
1160 	move_to_net_ee_or_free(device, peer_req);
1161 
1162 	if (unlikely(err))
1163 		drbd_err(device, "drbd_send_block() failed\n");
1164 	return err;
1165 }
1166 
1167 int w_e_end_csum_rs_req(struct drbd_work *w, int cancel)
1168 {
1169 	struct drbd_peer_request *peer_req = container_of(w, struct drbd_peer_request, w);
1170 	struct drbd_peer_device *peer_device = peer_req->peer_device;
1171 	struct drbd_device *device = peer_device->device;
1172 	struct digest_info *di;
1173 	int digest_size;
1174 	void *digest = NULL;
1175 	int err, eq = 0;
1176 
1177 	if (unlikely(cancel)) {
1178 		drbd_free_peer_req(device, peer_req);
1179 		dec_unacked(device);
1180 		return 0;
1181 	}
1182 
1183 	if (get_ldev(device)) {
1184 		drbd_rs_complete_io(device, peer_req->i.sector);
1185 		put_ldev(device);
1186 	}
1187 
1188 	di = peer_req->digest;
1189 
1190 	if (likely((peer_req->flags & EE_WAS_ERROR) == 0)) {
1191 		/* quick hack to try to avoid a race against reconfiguration.
1192 		 * a real fix would be much more involved,
1193 		 * introducing more locking mechanisms */
1194 		if (peer_device->connection->csums_tfm) {
1195 			digest_size = crypto_shash_digestsize(peer_device->connection->csums_tfm);
1196 			D_ASSERT(device, digest_size == di->digest_size);
1197 			digest = kmalloc(digest_size, GFP_NOIO);
1198 		}
1199 		if (digest) {
1200 			drbd_csum_ee(peer_device->connection->csums_tfm, peer_req, digest);
1201 			eq = !memcmp(digest, di->digest, digest_size);
1202 			kfree(digest);
1203 		}
1204 
1205 		if (eq) {
1206 			drbd_set_in_sync(peer_device, peer_req->i.sector, peer_req->i.size);
1207 			/* rs_same_csums unit is BM_BLOCK_SIZE */
1208 			device->rs_same_csum += peer_req->i.size >> BM_BLOCK_SHIFT;
1209 			err = drbd_send_ack(peer_device, P_RS_IS_IN_SYNC, peer_req);
1210 		} else {
1211 			inc_rs_pending(peer_device);
1212 			peer_req->block_id = ID_SYNCER; /* By setting block_id, digest pointer becomes invalid! */
1213 			peer_req->flags &= ~EE_HAS_DIGEST; /* This peer request no longer has a digest pointer */
1214 			kfree(di);
1215 			err = drbd_send_block(peer_device, P_RS_DATA_REPLY, peer_req);
1216 		}
1217 	} else {
1218 		err = drbd_send_ack(peer_device, P_NEG_RS_DREPLY, peer_req);
1219 		if (drbd_ratelimit())
1220 			drbd_err(device, "Sending NegDReply. I guess it gets messy.\n");
1221 	}
1222 
1223 	dec_unacked(device);
1224 	move_to_net_ee_or_free(device, peer_req);
1225 
1226 	if (unlikely(err))
1227 		drbd_err(device, "drbd_send_block/ack() failed\n");
1228 	return err;
1229 }
1230 
1231 int w_e_end_ov_req(struct drbd_work *w, int cancel)
1232 {
1233 	struct drbd_peer_request *peer_req = container_of(w, struct drbd_peer_request, w);
1234 	struct drbd_peer_device *peer_device = peer_req->peer_device;
1235 	struct drbd_device *device = peer_device->device;
1236 	sector_t sector = peer_req->i.sector;
1237 	unsigned int size = peer_req->i.size;
1238 	int digest_size;
1239 	void *digest;
1240 	int err = 0;
1241 
1242 	if (unlikely(cancel))
1243 		goto out;
1244 
1245 	digest_size = crypto_shash_digestsize(peer_device->connection->verify_tfm);
1246 	digest = kmalloc(digest_size, GFP_NOIO);
1247 	if (!digest) {
1248 		err = 1;	/* terminate the connection in case the allocation failed */
1249 		goto out;
1250 	}
1251 
1252 	if (likely(!(peer_req->flags & EE_WAS_ERROR)))
1253 		drbd_csum_ee(peer_device->connection->verify_tfm, peer_req, digest);
1254 	else
1255 		memset(digest, 0, digest_size);
1256 
1257 	/* Free e and pages before send.
1258 	 * In case we block on congestion, we could otherwise run into
1259 	 * some distributed deadlock, if the other side blocks on
1260 	 * congestion as well, because our receiver blocks in
1261 	 * drbd_alloc_pages due to pp_in_use > max_buffers. */
1262 	drbd_free_peer_req(device, peer_req);
1263 	peer_req = NULL;
1264 	inc_rs_pending(peer_device);
1265 	err = drbd_send_drequest_csum(peer_device, sector, size, digest, digest_size, P_OV_REPLY);
1266 	if (err)
1267 		dec_rs_pending(peer_device);
1268 	kfree(digest);
1269 
1270 out:
1271 	if (peer_req)
1272 		drbd_free_peer_req(device, peer_req);
1273 	dec_unacked(device);
1274 	return err;
1275 }
1276 
1277 void drbd_ov_out_of_sync_found(struct drbd_peer_device *peer_device, sector_t sector, int size)
1278 {
1279 	struct drbd_device *device = peer_device->device;
1280 	if (device->ov_last_oos_start + device->ov_last_oos_size == sector) {
1281 		device->ov_last_oos_size += size>>9;
1282 	} else {
1283 		device->ov_last_oos_start = sector;
1284 		device->ov_last_oos_size = size>>9;
1285 	}
1286 	drbd_set_out_of_sync(peer_device, sector, size);
1287 }
1288 
1289 int w_e_end_ov_reply(struct drbd_work *w, int cancel)
1290 {
1291 	struct drbd_peer_request *peer_req = container_of(w, struct drbd_peer_request, w);
1292 	struct drbd_peer_device *peer_device = peer_req->peer_device;
1293 	struct drbd_device *device = peer_device->device;
1294 	struct digest_info *di;
1295 	void *digest;
1296 	sector_t sector = peer_req->i.sector;
1297 	unsigned int size = peer_req->i.size;
1298 	int digest_size;
1299 	int err, eq = 0;
1300 	bool stop_sector_reached = false;
1301 
1302 	if (unlikely(cancel)) {
1303 		drbd_free_peer_req(device, peer_req);
1304 		dec_unacked(device);
1305 		return 0;
1306 	}
1307 
1308 	/* after "cancel", because after drbd_disconnect/drbd_rs_cancel_all
1309 	 * the resync lru has been cleaned up already */
1310 	if (get_ldev(device)) {
1311 		drbd_rs_complete_io(device, peer_req->i.sector);
1312 		put_ldev(device);
1313 	}
1314 
1315 	di = peer_req->digest;
1316 
1317 	if (likely((peer_req->flags & EE_WAS_ERROR) == 0)) {
1318 		digest_size = crypto_shash_digestsize(peer_device->connection->verify_tfm);
1319 		digest = kmalloc(digest_size, GFP_NOIO);
1320 		if (digest) {
1321 			drbd_csum_ee(peer_device->connection->verify_tfm, peer_req, digest);
1322 
1323 			D_ASSERT(device, digest_size == di->digest_size);
1324 			eq = !memcmp(digest, di->digest, digest_size);
1325 			kfree(digest);
1326 		}
1327 	}
1328 
1329 	/* Free peer_req and pages before send.
1330 	 * In case we block on congestion, we could otherwise run into
1331 	 * some distributed deadlock, if the other side blocks on
1332 	 * congestion as well, because our receiver blocks in
1333 	 * drbd_alloc_pages due to pp_in_use > max_buffers. */
1334 	drbd_free_peer_req(device, peer_req);
1335 	if (!eq)
1336 		drbd_ov_out_of_sync_found(peer_device, sector, size);
1337 	else
1338 		ov_out_of_sync_print(peer_device);
1339 
1340 	err = drbd_send_ack_ex(peer_device, P_OV_RESULT, sector, size,
1341 			       eq ? ID_IN_SYNC : ID_OUT_OF_SYNC);
1342 
1343 	dec_unacked(device);
1344 
1345 	--device->ov_left;
1346 
1347 	/* let's advance progress step marks only for every other megabyte */
1348 	if ((device->ov_left & 0x200) == 0x200)
1349 		drbd_advance_rs_marks(peer_device, device->ov_left);
1350 
1351 	stop_sector_reached = verify_can_do_stop_sector(device) &&
1352 		(sector + (size>>9)) >= device->ov_stop_sector;
1353 
1354 	if (device->ov_left == 0 || stop_sector_reached) {
1355 		ov_out_of_sync_print(peer_device);
1356 		drbd_resync_finished(peer_device);
1357 	}
1358 
1359 	return err;
1360 }
1361 
1362 /* FIXME
1363  * We need to track the number of pending barrier acks,
1364  * and to be able to wait for them.
1365  * See also comment in drbd_adm_attach before drbd_suspend_io.
1366  */
1367 static int drbd_send_barrier(struct drbd_connection *connection)
1368 {
1369 	struct p_barrier *p;
1370 	struct drbd_socket *sock;
1371 
1372 	sock = &connection->data;
1373 	p = conn_prepare_command(connection, sock);
1374 	if (!p)
1375 		return -EIO;
1376 	p->barrier = connection->send.current_epoch_nr;
1377 	p->pad = 0;
1378 	connection->send.current_epoch_writes = 0;
1379 	connection->send.last_sent_barrier_jif = jiffies;
1380 
1381 	return conn_send_command(connection, sock, P_BARRIER, sizeof(*p), NULL, 0);
1382 }
1383 
1384 static int pd_send_unplug_remote(struct drbd_peer_device *pd)
1385 {
1386 	struct drbd_socket *sock = &pd->connection->data;
1387 	if (!drbd_prepare_command(pd, sock))
1388 		return -EIO;
1389 	return drbd_send_command(pd, sock, P_UNPLUG_REMOTE, 0, NULL, 0);
1390 }
1391 
1392 int w_send_write_hint(struct drbd_work *w, int cancel)
1393 {
1394 	struct drbd_device *device =
1395 		container_of(w, struct drbd_device, unplug_work);
1396 
1397 	if (cancel)
1398 		return 0;
1399 	return pd_send_unplug_remote(first_peer_device(device));
1400 }
1401 
1402 static void re_init_if_first_write(struct drbd_connection *connection, unsigned int epoch)
1403 {
1404 	if (!connection->send.seen_any_write_yet) {
1405 		connection->send.seen_any_write_yet = true;
1406 		connection->send.current_epoch_nr = epoch;
1407 		connection->send.current_epoch_writes = 0;
1408 		connection->send.last_sent_barrier_jif = jiffies;
1409 	}
1410 }
1411 
1412 static void maybe_send_barrier(struct drbd_connection *connection, unsigned int epoch)
1413 {
1414 	/* re-init if first write on this connection */
1415 	if (!connection->send.seen_any_write_yet)
1416 		return;
1417 	if (connection->send.current_epoch_nr != epoch) {
1418 		if (connection->send.current_epoch_writes)
1419 			drbd_send_barrier(connection);
1420 		connection->send.current_epoch_nr = epoch;
1421 	}
1422 }
1423 
1424 int w_send_out_of_sync(struct drbd_work *w, int cancel)
1425 {
1426 	struct drbd_request *req = container_of(w, struct drbd_request, w);
1427 	struct drbd_device *device = req->device;
1428 	struct drbd_peer_device *const peer_device = first_peer_device(device);
1429 	struct drbd_connection *const connection = peer_device->connection;
1430 	int err;
1431 
1432 	if (unlikely(cancel)) {
1433 		req_mod(req, SEND_CANCELED, peer_device);
1434 		return 0;
1435 	}
1436 	req->pre_send_jif = jiffies;
1437 
1438 	/* this time, no connection->send.current_epoch_writes++;
1439 	 * If it was sent, it was the closing barrier for the last
1440 	 * replicated epoch, before we went into AHEAD mode.
1441 	 * No more barriers will be sent, until we leave AHEAD mode again. */
1442 	maybe_send_barrier(connection, req->epoch);
1443 
1444 	err = drbd_send_out_of_sync(peer_device, req);
1445 	req_mod(req, OOS_HANDED_TO_NETWORK, peer_device);
1446 
1447 	return err;
1448 }
1449 
1450 /**
1451  * w_send_dblock() - Worker callback to send a P_DATA packet in order to mirror a write request
1452  * @w:		work object.
1453  * @cancel:	The connection will be closed anyways
1454  */
1455 int w_send_dblock(struct drbd_work *w, int cancel)
1456 {
1457 	struct drbd_request *req = container_of(w, struct drbd_request, w);
1458 	struct drbd_device *device = req->device;
1459 	struct drbd_peer_device *const peer_device = first_peer_device(device);
1460 	struct drbd_connection *connection = peer_device->connection;
1461 	bool do_send_unplug = req->rq_state & RQ_UNPLUG;
1462 	int err;
1463 
1464 	if (unlikely(cancel)) {
1465 		req_mod(req, SEND_CANCELED, peer_device);
1466 		return 0;
1467 	}
1468 	req->pre_send_jif = jiffies;
1469 
1470 	re_init_if_first_write(connection, req->epoch);
1471 	maybe_send_barrier(connection, req->epoch);
1472 	connection->send.current_epoch_writes++;
1473 
1474 	err = drbd_send_dblock(peer_device, req);
1475 	req_mod(req, err ? SEND_FAILED : HANDED_OVER_TO_NETWORK, peer_device);
1476 
1477 	if (do_send_unplug && !err)
1478 		pd_send_unplug_remote(peer_device);
1479 
1480 	return err;
1481 }
1482 
1483 /**
1484  * w_send_read_req() - Worker callback to send a read request (P_DATA_REQUEST) packet
1485  * @w:		work object.
1486  * @cancel:	The connection will be closed anyways
1487  */
1488 int w_send_read_req(struct drbd_work *w, int cancel)
1489 {
1490 	struct drbd_request *req = container_of(w, struct drbd_request, w);
1491 	struct drbd_device *device = req->device;
1492 	struct drbd_peer_device *const peer_device = first_peer_device(device);
1493 	struct drbd_connection *connection = peer_device->connection;
1494 	bool do_send_unplug = req->rq_state & RQ_UNPLUG;
1495 	int err;
1496 
1497 	if (unlikely(cancel)) {
1498 		req_mod(req, SEND_CANCELED, peer_device);
1499 		return 0;
1500 	}
1501 	req->pre_send_jif = jiffies;
1502 
1503 	/* Even read requests may close a write epoch,
1504 	 * if there was any yet. */
1505 	maybe_send_barrier(connection, req->epoch);
1506 
1507 	err = drbd_send_drequest(peer_device, P_DATA_REQUEST, req->i.sector, req->i.size,
1508 				 (unsigned long)req);
1509 
1510 	req_mod(req, err ? SEND_FAILED : HANDED_OVER_TO_NETWORK, peer_device);
1511 
1512 	if (do_send_unplug && !err)
1513 		pd_send_unplug_remote(peer_device);
1514 
1515 	return err;
1516 }
1517 
1518 int w_restart_disk_io(struct drbd_work *w, int cancel)
1519 {
1520 	struct drbd_request *req = container_of(w, struct drbd_request, w);
1521 	struct drbd_device *device = req->device;
1522 
1523 	if (bio_data_dir(req->master_bio) == WRITE && req->rq_state & RQ_IN_ACT_LOG)
1524 		drbd_al_begin_io(device, &req->i);
1525 
1526 	req->private_bio = bio_alloc_clone(device->ldev->backing_bdev,
1527 					   req->master_bio, GFP_NOIO,
1528 					  &drbd_io_bio_set);
1529 	req->private_bio->bi_private = req;
1530 	req->private_bio->bi_end_io = drbd_request_endio;
1531 	submit_bio_noacct(req->private_bio);
1532 
1533 	return 0;
1534 }
1535 
1536 static int _drbd_may_sync_now(struct drbd_device *device)
1537 {
1538 	struct drbd_device *odev = device;
1539 	int resync_after;
1540 
1541 	while (1) {
1542 		if (!odev->ldev || odev->state.disk == D_DISKLESS)
1543 			return 1;
1544 		rcu_read_lock();
1545 		resync_after = rcu_dereference(odev->ldev->disk_conf)->resync_after;
1546 		rcu_read_unlock();
1547 		if (resync_after == -1)
1548 			return 1;
1549 		odev = minor_to_device(resync_after);
1550 		if (!odev)
1551 			return 1;
1552 		if ((odev->state.conn >= C_SYNC_SOURCE &&
1553 		     odev->state.conn <= C_PAUSED_SYNC_T) ||
1554 		    odev->state.aftr_isp || odev->state.peer_isp ||
1555 		    odev->state.user_isp)
1556 			return 0;
1557 	}
1558 }
1559 
1560 /**
1561  * drbd_pause_after() - Pause resync on all devices that may not resync now
1562  * @device:	DRBD device.
1563  *
1564  * Called from process context only (admin command and after_state_ch).
1565  */
1566 static bool drbd_pause_after(struct drbd_device *device)
1567 {
1568 	bool changed = false;
1569 	struct drbd_device *odev;
1570 	int i;
1571 
1572 	rcu_read_lock();
1573 	idr_for_each_entry(&drbd_devices, odev, i) {
1574 		if (odev->state.conn == C_STANDALONE && odev->state.disk == D_DISKLESS)
1575 			continue;
1576 		if (!_drbd_may_sync_now(odev) &&
1577 		    _drbd_set_state(_NS(odev, aftr_isp, 1),
1578 				    CS_HARD, NULL) != SS_NOTHING_TO_DO)
1579 			changed = true;
1580 	}
1581 	rcu_read_unlock();
1582 
1583 	return changed;
1584 }
1585 
1586 /**
1587  * drbd_resume_next() - Resume resync on all devices that may resync now
1588  * @device:	DRBD device.
1589  *
1590  * Called from process context only (admin command and worker).
1591  */
1592 static bool drbd_resume_next(struct drbd_device *device)
1593 {
1594 	bool changed = false;
1595 	struct drbd_device *odev;
1596 	int i;
1597 
1598 	rcu_read_lock();
1599 	idr_for_each_entry(&drbd_devices, odev, i) {
1600 		if (odev->state.conn == C_STANDALONE && odev->state.disk == D_DISKLESS)
1601 			continue;
1602 		if (odev->state.aftr_isp) {
1603 			if (_drbd_may_sync_now(odev) &&
1604 			    _drbd_set_state(_NS(odev, aftr_isp, 0),
1605 					    CS_HARD, NULL) != SS_NOTHING_TO_DO)
1606 				changed = true;
1607 		}
1608 	}
1609 	rcu_read_unlock();
1610 	return changed;
1611 }
1612 
1613 void resume_next_sg(struct drbd_device *device)
1614 {
1615 	lock_all_resources();
1616 	drbd_resume_next(device);
1617 	unlock_all_resources();
1618 }
1619 
1620 void suspend_other_sg(struct drbd_device *device)
1621 {
1622 	lock_all_resources();
1623 	drbd_pause_after(device);
1624 	unlock_all_resources();
1625 }
1626 
1627 /* caller must lock_all_resources() */
1628 enum drbd_ret_code drbd_resync_after_valid(struct drbd_device *device, int o_minor)
1629 {
1630 	struct drbd_device *odev;
1631 	int resync_after;
1632 
1633 	if (o_minor == -1)
1634 		return NO_ERROR;
1635 	if (o_minor < -1 || o_minor > MINORMASK)
1636 		return ERR_RESYNC_AFTER;
1637 
1638 	/* check for loops */
1639 	odev = minor_to_device(o_minor);
1640 	while (1) {
1641 		if (odev == device)
1642 			return ERR_RESYNC_AFTER_CYCLE;
1643 
1644 		/* You are free to depend on diskless, non-existing,
1645 		 * or not yet/no longer existing minors.
1646 		 * We only reject dependency loops.
1647 		 * We cannot follow the dependency chain beyond a detached or
1648 		 * missing minor.
1649 		 */
1650 		if (!odev || !odev->ldev || odev->state.disk == D_DISKLESS)
1651 			return NO_ERROR;
1652 
1653 		rcu_read_lock();
1654 		resync_after = rcu_dereference(odev->ldev->disk_conf)->resync_after;
1655 		rcu_read_unlock();
1656 		/* dependency chain ends here, no cycles. */
1657 		if (resync_after == -1)
1658 			return NO_ERROR;
1659 
1660 		/* follow the dependency chain */
1661 		odev = minor_to_device(resync_after);
1662 	}
1663 }
1664 
1665 /* caller must lock_all_resources() */
1666 void drbd_resync_after_changed(struct drbd_device *device)
1667 {
1668 	int changed;
1669 
1670 	do {
1671 		changed  = drbd_pause_after(device);
1672 		changed |= drbd_resume_next(device);
1673 	} while (changed);
1674 }
1675 
1676 void drbd_rs_controller_reset(struct drbd_peer_device *peer_device)
1677 {
1678 	struct drbd_device *device = peer_device->device;
1679 	struct gendisk *disk = device->ldev->backing_bdev->bd_disk;
1680 	struct fifo_buffer *plan;
1681 
1682 	atomic_set(&device->rs_sect_in, 0);
1683 	atomic_set(&device->rs_sect_ev, 0);
1684 	device->rs_in_flight = 0;
1685 	device->rs_last_events =
1686 		(int)part_stat_read_accum(disk->part0, sectors);
1687 
1688 	/* Updating the RCU protected object in place is necessary since
1689 	   this function gets called from atomic context.
1690 	   It is valid since all other updates also lead to an completely
1691 	   empty fifo */
1692 	rcu_read_lock();
1693 	plan = rcu_dereference(device->rs_plan_s);
1694 	plan->total = 0;
1695 	fifo_set(plan, 0);
1696 	rcu_read_unlock();
1697 }
1698 
1699 void start_resync_timer_fn(struct timer_list *t)
1700 {
1701 	struct drbd_device *device = from_timer(device, t, start_resync_timer);
1702 	drbd_device_post_work(device, RS_START);
1703 }
1704 
1705 static void do_start_resync(struct drbd_device *device)
1706 {
1707 	if (atomic_read(&device->unacked_cnt) || atomic_read(&device->rs_pending_cnt)) {
1708 		drbd_warn(device, "postponing start_resync ...\n");
1709 		device->start_resync_timer.expires = jiffies + HZ/10;
1710 		add_timer(&device->start_resync_timer);
1711 		return;
1712 	}
1713 
1714 	drbd_start_resync(device, C_SYNC_SOURCE);
1715 	clear_bit(AHEAD_TO_SYNC_SOURCE, &device->flags);
1716 }
1717 
1718 static bool use_checksum_based_resync(struct drbd_connection *connection, struct drbd_device *device)
1719 {
1720 	bool csums_after_crash_only;
1721 	rcu_read_lock();
1722 	csums_after_crash_only = rcu_dereference(connection->net_conf)->csums_after_crash_only;
1723 	rcu_read_unlock();
1724 	return connection->agreed_pro_version >= 89 &&		/* supported? */
1725 		connection->csums_tfm &&			/* configured? */
1726 		(csums_after_crash_only == false		/* use for each resync? */
1727 		 || test_bit(CRASHED_PRIMARY, &device->flags));	/* or only after Primary crash? */
1728 }
1729 
1730 /**
1731  * drbd_start_resync() - Start the resync process
1732  * @device:	DRBD device.
1733  * @side:	Either C_SYNC_SOURCE or C_SYNC_TARGET
1734  *
1735  * This function might bring you directly into one of the
1736  * C_PAUSED_SYNC_* states.
1737  */
1738 void drbd_start_resync(struct drbd_device *device, enum drbd_conns side)
1739 {
1740 	struct drbd_peer_device *peer_device = first_peer_device(device);
1741 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
1742 	union drbd_state ns;
1743 	int r;
1744 
1745 	if (device->state.conn >= C_SYNC_SOURCE && device->state.conn < C_AHEAD) {
1746 		drbd_err(device, "Resync already running!\n");
1747 		return;
1748 	}
1749 
1750 	if (!connection) {
1751 		drbd_err(device, "No connection to peer, aborting!\n");
1752 		return;
1753 	}
1754 
1755 	if (!test_bit(B_RS_H_DONE, &device->flags)) {
1756 		if (side == C_SYNC_TARGET) {
1757 			/* Since application IO was locked out during C_WF_BITMAP_T and
1758 			   C_WF_SYNC_UUID we are still unmodified. Before going to C_SYNC_TARGET
1759 			   we check that we might make the data inconsistent. */
1760 			r = drbd_khelper(device, "before-resync-target");
1761 			r = (r >> 8) & 0xff;
1762 			if (r > 0) {
1763 				drbd_info(device, "before-resync-target handler returned %d, "
1764 					 "dropping connection.\n", r);
1765 				conn_request_state(connection, NS(conn, C_DISCONNECTING), CS_HARD);
1766 				return;
1767 			}
1768 		} else /* C_SYNC_SOURCE */ {
1769 			r = drbd_khelper(device, "before-resync-source");
1770 			r = (r >> 8) & 0xff;
1771 			if (r > 0) {
1772 				if (r == 3) {
1773 					drbd_info(device, "before-resync-source handler returned %d, "
1774 						 "ignoring. Old userland tools?", r);
1775 				} else {
1776 					drbd_info(device, "before-resync-source handler returned %d, "
1777 						 "dropping connection.\n", r);
1778 					conn_request_state(connection,
1779 							   NS(conn, C_DISCONNECTING), CS_HARD);
1780 					return;
1781 				}
1782 			}
1783 		}
1784 	}
1785 
1786 	if (current == connection->worker.task) {
1787 		/* The worker should not sleep waiting for state_mutex,
1788 		   that can take long */
1789 		if (!mutex_trylock(device->state_mutex)) {
1790 			set_bit(B_RS_H_DONE, &device->flags);
1791 			device->start_resync_timer.expires = jiffies + HZ/5;
1792 			add_timer(&device->start_resync_timer);
1793 			return;
1794 		}
1795 	} else {
1796 		mutex_lock(device->state_mutex);
1797 	}
1798 
1799 	lock_all_resources();
1800 	clear_bit(B_RS_H_DONE, &device->flags);
1801 	/* Did some connection breakage or IO error race with us? */
1802 	if (device->state.conn < C_CONNECTED
1803 	|| !get_ldev_if_state(device, D_NEGOTIATING)) {
1804 		unlock_all_resources();
1805 		goto out;
1806 	}
1807 
1808 	ns = drbd_read_state(device);
1809 
1810 	ns.aftr_isp = !_drbd_may_sync_now(device);
1811 
1812 	ns.conn = side;
1813 
1814 	if (side == C_SYNC_TARGET)
1815 		ns.disk = D_INCONSISTENT;
1816 	else /* side == C_SYNC_SOURCE */
1817 		ns.pdsk = D_INCONSISTENT;
1818 
1819 	r = _drbd_set_state(device, ns, CS_VERBOSE, NULL);
1820 	ns = drbd_read_state(device);
1821 
1822 	if (ns.conn < C_CONNECTED)
1823 		r = SS_UNKNOWN_ERROR;
1824 
1825 	if (r == SS_SUCCESS) {
1826 		unsigned long tw = drbd_bm_total_weight(device);
1827 		unsigned long now = jiffies;
1828 		int i;
1829 
1830 		device->rs_failed    = 0;
1831 		device->rs_paused    = 0;
1832 		device->rs_same_csum = 0;
1833 		device->rs_last_sect_ev = 0;
1834 		device->rs_total     = tw;
1835 		device->rs_start     = now;
1836 		for (i = 0; i < DRBD_SYNC_MARKS; i++) {
1837 			device->rs_mark_left[i] = tw;
1838 			device->rs_mark_time[i] = now;
1839 		}
1840 		drbd_pause_after(device);
1841 		/* Forget potentially stale cached per resync extent bit-counts.
1842 		 * Open coded drbd_rs_cancel_all(device), we already have IRQs
1843 		 * disabled, and know the disk state is ok. */
1844 		spin_lock(&device->al_lock);
1845 		lc_reset(device->resync);
1846 		device->resync_locked = 0;
1847 		device->resync_wenr = LC_FREE;
1848 		spin_unlock(&device->al_lock);
1849 	}
1850 	unlock_all_resources();
1851 
1852 	if (r == SS_SUCCESS) {
1853 		wake_up(&device->al_wait); /* for lc_reset() above */
1854 		/* reset rs_last_bcast when a resync or verify is started,
1855 		 * to deal with potential jiffies wrap. */
1856 		device->rs_last_bcast = jiffies - HZ;
1857 
1858 		drbd_info(device, "Began resync as %s (will sync %lu KB [%lu bits set]).\n",
1859 		     drbd_conn_str(ns.conn),
1860 		     (unsigned long) device->rs_total << (BM_BLOCK_SHIFT-10),
1861 		     (unsigned long) device->rs_total);
1862 		if (side == C_SYNC_TARGET) {
1863 			device->bm_resync_fo = 0;
1864 			device->use_csums = use_checksum_based_resync(connection, device);
1865 		} else {
1866 			device->use_csums = false;
1867 		}
1868 
1869 		/* Since protocol 96, we must serialize drbd_gen_and_send_sync_uuid
1870 		 * with w_send_oos, or the sync target will get confused as to
1871 		 * how much bits to resync.  We cannot do that always, because for an
1872 		 * empty resync and protocol < 95, we need to do it here, as we call
1873 		 * drbd_resync_finished from here in that case.
1874 		 * We drbd_gen_and_send_sync_uuid here for protocol < 96,
1875 		 * and from after_state_ch otherwise. */
1876 		if (side == C_SYNC_SOURCE && connection->agreed_pro_version < 96)
1877 			drbd_gen_and_send_sync_uuid(peer_device);
1878 
1879 		if (connection->agreed_pro_version < 95 && device->rs_total == 0) {
1880 			/* This still has a race (about when exactly the peers
1881 			 * detect connection loss) that can lead to a full sync
1882 			 * on next handshake. In 8.3.9 we fixed this with explicit
1883 			 * resync-finished notifications, but the fix
1884 			 * introduces a protocol change.  Sleeping for some
1885 			 * time longer than the ping interval + timeout on the
1886 			 * SyncSource, to give the SyncTarget the chance to
1887 			 * detect connection loss, then waiting for a ping
1888 			 * response (implicit in drbd_resync_finished) reduces
1889 			 * the race considerably, but does not solve it. */
1890 			if (side == C_SYNC_SOURCE) {
1891 				struct net_conf *nc;
1892 				int timeo;
1893 
1894 				rcu_read_lock();
1895 				nc = rcu_dereference(connection->net_conf);
1896 				timeo = nc->ping_int * HZ + nc->ping_timeo * HZ / 9;
1897 				rcu_read_unlock();
1898 				schedule_timeout_interruptible(timeo);
1899 			}
1900 			drbd_resync_finished(peer_device);
1901 		}
1902 
1903 		drbd_rs_controller_reset(peer_device);
1904 		/* ns.conn may already be != device->state.conn,
1905 		 * we may have been paused in between, or become paused until
1906 		 * the timer triggers.
1907 		 * No matter, that is handled in resync_timer_fn() */
1908 		if (ns.conn == C_SYNC_TARGET)
1909 			mod_timer(&device->resync_timer, jiffies);
1910 
1911 		drbd_md_sync(device);
1912 	}
1913 	put_ldev(device);
1914 out:
1915 	mutex_unlock(device->state_mutex);
1916 }
1917 
1918 static void update_on_disk_bitmap(struct drbd_peer_device *peer_device, bool resync_done)
1919 {
1920 	struct drbd_device *device = peer_device->device;
1921 	struct sib_info sib = { .sib_reason = SIB_SYNC_PROGRESS, };
1922 	device->rs_last_bcast = jiffies;
1923 
1924 	if (!get_ldev(device))
1925 		return;
1926 
1927 	drbd_bm_write_lazy(device, 0);
1928 	if (resync_done && is_sync_state(device->state.conn))
1929 		drbd_resync_finished(peer_device);
1930 
1931 	drbd_bcast_event(device, &sib);
1932 	/* update timestamp, in case it took a while to write out stuff */
1933 	device->rs_last_bcast = jiffies;
1934 	put_ldev(device);
1935 }
1936 
1937 static void drbd_ldev_destroy(struct drbd_device *device)
1938 {
1939 	lc_destroy(device->resync);
1940 	device->resync = NULL;
1941 	lc_destroy(device->act_log);
1942 	device->act_log = NULL;
1943 
1944 	__acquire(local);
1945 	drbd_backing_dev_free(device, device->ldev);
1946 	device->ldev = NULL;
1947 	__release(local);
1948 
1949 	clear_bit(GOING_DISKLESS, &device->flags);
1950 	wake_up(&device->misc_wait);
1951 }
1952 
1953 static void go_diskless(struct drbd_device *device)
1954 {
1955 	struct drbd_peer_device *peer_device = first_peer_device(device);
1956 	D_ASSERT(device, device->state.disk == D_FAILED);
1957 	/* we cannot assert local_cnt == 0 here, as get_ldev_if_state will
1958 	 * inc/dec it frequently. Once we are D_DISKLESS, no one will touch
1959 	 * the protected members anymore, though, so once put_ldev reaches zero
1960 	 * again, it will be safe to free them. */
1961 
1962 	/* Try to write changed bitmap pages, read errors may have just
1963 	 * set some bits outside the area covered by the activity log.
1964 	 *
1965 	 * If we have an IO error during the bitmap writeout,
1966 	 * we will want a full sync next time, just in case.
1967 	 * (Do we want a specific meta data flag for this?)
1968 	 *
1969 	 * If that does not make it to stable storage either,
1970 	 * we cannot do anything about that anymore.
1971 	 *
1972 	 * We still need to check if both bitmap and ldev are present, we may
1973 	 * end up here after a failed attach, before ldev was even assigned.
1974 	 */
1975 	if (device->bitmap && device->ldev) {
1976 		/* An interrupted resync or similar is allowed to recounts bits
1977 		 * while we detach.
1978 		 * Any modifications would not be expected anymore, though.
1979 		 */
1980 		if (drbd_bitmap_io_from_worker(device, drbd_bm_write,
1981 					"detach", BM_LOCKED_TEST_ALLOWED, peer_device)) {
1982 			if (test_bit(WAS_READ_ERROR, &device->flags)) {
1983 				drbd_md_set_flag(device, MDF_FULL_SYNC);
1984 				drbd_md_sync(device);
1985 			}
1986 		}
1987 	}
1988 
1989 	drbd_force_state(device, NS(disk, D_DISKLESS));
1990 }
1991 
1992 static int do_md_sync(struct drbd_device *device)
1993 {
1994 	drbd_warn(device, "md_sync_timer expired! Worker calls drbd_md_sync().\n");
1995 	drbd_md_sync(device);
1996 	return 0;
1997 }
1998 
1999 /* only called from drbd_worker thread, no locking */
2000 void __update_timing_details(
2001 		struct drbd_thread_timing_details *tdp,
2002 		unsigned int *cb_nr,
2003 		void *cb,
2004 		const char *fn, const unsigned int line)
2005 {
2006 	unsigned int i = *cb_nr % DRBD_THREAD_DETAILS_HIST;
2007 	struct drbd_thread_timing_details *td = tdp + i;
2008 
2009 	td->start_jif = jiffies;
2010 	td->cb_addr = cb;
2011 	td->caller_fn = fn;
2012 	td->line = line;
2013 	td->cb_nr = *cb_nr;
2014 
2015 	i = (i+1) % DRBD_THREAD_DETAILS_HIST;
2016 	td = tdp + i;
2017 	memset(td, 0, sizeof(*td));
2018 
2019 	++(*cb_nr);
2020 }
2021 
2022 static void do_device_work(struct drbd_device *device, const unsigned long todo)
2023 {
2024 	if (test_bit(MD_SYNC, &todo))
2025 		do_md_sync(device);
2026 	if (test_bit(RS_DONE, &todo) ||
2027 	    test_bit(RS_PROGRESS, &todo))
2028 		update_on_disk_bitmap(first_peer_device(device), test_bit(RS_DONE, &todo));
2029 	if (test_bit(GO_DISKLESS, &todo))
2030 		go_diskless(device);
2031 	if (test_bit(DESTROY_DISK, &todo))
2032 		drbd_ldev_destroy(device);
2033 	if (test_bit(RS_START, &todo))
2034 		do_start_resync(device);
2035 }
2036 
2037 #define DRBD_DEVICE_WORK_MASK	\
2038 	((1UL << GO_DISKLESS)	\
2039 	|(1UL << DESTROY_DISK)	\
2040 	|(1UL << MD_SYNC)	\
2041 	|(1UL << RS_START)	\
2042 	|(1UL << RS_PROGRESS)	\
2043 	|(1UL << RS_DONE)	\
2044 	)
2045 
2046 static unsigned long get_work_bits(unsigned long *flags)
2047 {
2048 	unsigned long old, new;
2049 	do {
2050 		old = *flags;
2051 		new = old & ~DRBD_DEVICE_WORK_MASK;
2052 	} while (cmpxchg(flags, old, new) != old);
2053 	return old & DRBD_DEVICE_WORK_MASK;
2054 }
2055 
2056 static void do_unqueued_work(struct drbd_connection *connection)
2057 {
2058 	struct drbd_peer_device *peer_device;
2059 	int vnr;
2060 
2061 	rcu_read_lock();
2062 	idr_for_each_entry(&connection->peer_devices, peer_device, vnr) {
2063 		struct drbd_device *device = peer_device->device;
2064 		unsigned long todo = get_work_bits(&device->flags);
2065 		if (!todo)
2066 			continue;
2067 
2068 		kref_get(&device->kref);
2069 		rcu_read_unlock();
2070 		do_device_work(device, todo);
2071 		kref_put(&device->kref, drbd_destroy_device);
2072 		rcu_read_lock();
2073 	}
2074 	rcu_read_unlock();
2075 }
2076 
2077 static bool dequeue_work_batch(struct drbd_work_queue *queue, struct list_head *work_list)
2078 {
2079 	spin_lock_irq(&queue->q_lock);
2080 	list_splice_tail_init(&queue->q, work_list);
2081 	spin_unlock_irq(&queue->q_lock);
2082 	return !list_empty(work_list);
2083 }
2084 
2085 static void wait_for_work(struct drbd_connection *connection, struct list_head *work_list)
2086 {
2087 	DEFINE_WAIT(wait);
2088 	struct net_conf *nc;
2089 	int uncork, cork;
2090 
2091 	dequeue_work_batch(&connection->sender_work, work_list);
2092 	if (!list_empty(work_list))
2093 		return;
2094 
2095 	/* Still nothing to do?
2096 	 * Maybe we still need to close the current epoch,
2097 	 * even if no new requests are queued yet.
2098 	 *
2099 	 * Also, poke TCP, just in case.
2100 	 * Then wait for new work (or signal). */
2101 	rcu_read_lock();
2102 	nc = rcu_dereference(connection->net_conf);
2103 	uncork = nc ? nc->tcp_cork : 0;
2104 	rcu_read_unlock();
2105 	if (uncork) {
2106 		mutex_lock(&connection->data.mutex);
2107 		if (connection->data.socket)
2108 			tcp_sock_set_cork(connection->data.socket->sk, false);
2109 		mutex_unlock(&connection->data.mutex);
2110 	}
2111 
2112 	for (;;) {
2113 		int send_barrier;
2114 		prepare_to_wait(&connection->sender_work.q_wait, &wait, TASK_INTERRUPTIBLE);
2115 		spin_lock_irq(&connection->resource->req_lock);
2116 		spin_lock(&connection->sender_work.q_lock);	/* FIXME get rid of this one? */
2117 		if (!list_empty(&connection->sender_work.q))
2118 			list_splice_tail_init(&connection->sender_work.q, work_list);
2119 		spin_unlock(&connection->sender_work.q_lock);	/* FIXME get rid of this one? */
2120 		if (!list_empty(work_list) || signal_pending(current)) {
2121 			spin_unlock_irq(&connection->resource->req_lock);
2122 			break;
2123 		}
2124 
2125 		/* We found nothing new to do, no to-be-communicated request,
2126 		 * no other work item.  We may still need to close the last
2127 		 * epoch.  Next incoming request epoch will be connection ->
2128 		 * current transfer log epoch number.  If that is different
2129 		 * from the epoch of the last request we communicated, it is
2130 		 * safe to send the epoch separating barrier now.
2131 		 */
2132 		send_barrier =
2133 			atomic_read(&connection->current_tle_nr) !=
2134 			connection->send.current_epoch_nr;
2135 		spin_unlock_irq(&connection->resource->req_lock);
2136 
2137 		if (send_barrier)
2138 			maybe_send_barrier(connection,
2139 					connection->send.current_epoch_nr + 1);
2140 
2141 		if (test_bit(DEVICE_WORK_PENDING, &connection->flags))
2142 			break;
2143 
2144 		/* drbd_send() may have called flush_signals() */
2145 		if (get_t_state(&connection->worker) != RUNNING)
2146 			break;
2147 
2148 		schedule();
2149 		/* may be woken up for other things but new work, too,
2150 		 * e.g. if the current epoch got closed.
2151 		 * In which case we send the barrier above. */
2152 	}
2153 	finish_wait(&connection->sender_work.q_wait, &wait);
2154 
2155 	/* someone may have changed the config while we have been waiting above. */
2156 	rcu_read_lock();
2157 	nc = rcu_dereference(connection->net_conf);
2158 	cork = nc ? nc->tcp_cork : 0;
2159 	rcu_read_unlock();
2160 	mutex_lock(&connection->data.mutex);
2161 	if (connection->data.socket) {
2162 		if (cork)
2163 			tcp_sock_set_cork(connection->data.socket->sk, true);
2164 		else if (!uncork)
2165 			tcp_sock_set_cork(connection->data.socket->sk, false);
2166 	}
2167 	mutex_unlock(&connection->data.mutex);
2168 }
2169 
2170 int drbd_worker(struct drbd_thread *thi)
2171 {
2172 	struct drbd_connection *connection = thi->connection;
2173 	struct drbd_work *w = NULL;
2174 	struct drbd_peer_device *peer_device;
2175 	LIST_HEAD(work_list);
2176 	int vnr;
2177 
2178 	while (get_t_state(thi) == RUNNING) {
2179 		drbd_thread_current_set_cpu(thi);
2180 
2181 		if (list_empty(&work_list)) {
2182 			update_worker_timing_details(connection, wait_for_work);
2183 			wait_for_work(connection, &work_list);
2184 		}
2185 
2186 		if (test_and_clear_bit(DEVICE_WORK_PENDING, &connection->flags)) {
2187 			update_worker_timing_details(connection, do_unqueued_work);
2188 			do_unqueued_work(connection);
2189 		}
2190 
2191 		if (signal_pending(current)) {
2192 			flush_signals(current);
2193 			if (get_t_state(thi) == RUNNING) {
2194 				drbd_warn(connection, "Worker got an unexpected signal\n");
2195 				continue;
2196 			}
2197 			break;
2198 		}
2199 
2200 		if (get_t_state(thi) != RUNNING)
2201 			break;
2202 
2203 		if (!list_empty(&work_list)) {
2204 			w = list_first_entry(&work_list, struct drbd_work, list);
2205 			list_del_init(&w->list);
2206 			update_worker_timing_details(connection, w->cb);
2207 			if (w->cb(w, connection->cstate < C_WF_REPORT_PARAMS) == 0)
2208 				continue;
2209 			if (connection->cstate >= C_WF_REPORT_PARAMS)
2210 				conn_request_state(connection, NS(conn, C_NETWORK_FAILURE), CS_HARD);
2211 		}
2212 	}
2213 
2214 	do {
2215 		if (test_and_clear_bit(DEVICE_WORK_PENDING, &connection->flags)) {
2216 			update_worker_timing_details(connection, do_unqueued_work);
2217 			do_unqueued_work(connection);
2218 		}
2219 		if (!list_empty(&work_list)) {
2220 			w = list_first_entry(&work_list, struct drbd_work, list);
2221 			list_del_init(&w->list);
2222 			update_worker_timing_details(connection, w->cb);
2223 			w->cb(w, 1);
2224 		} else
2225 			dequeue_work_batch(&connection->sender_work, &work_list);
2226 	} while (!list_empty(&work_list) || test_bit(DEVICE_WORK_PENDING, &connection->flags));
2227 
2228 	rcu_read_lock();
2229 	idr_for_each_entry(&connection->peer_devices, peer_device, vnr) {
2230 		struct drbd_device *device = peer_device->device;
2231 		D_ASSERT(device, device->state.disk == D_DISKLESS && device->state.conn == C_STANDALONE);
2232 		kref_get(&device->kref);
2233 		rcu_read_unlock();
2234 		drbd_device_cleanup(device);
2235 		kref_put(&device->kref, drbd_destroy_device);
2236 		rcu_read_lock();
2237 	}
2238 	rcu_read_unlock();
2239 
2240 	return 0;
2241 }
2242