xref: /openbmc/linux/drivers/md/raid1.c (revision 1f98a13f623e0ef666690a18c1250335fc6d7ef1)
1 /*
2  * raid1.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5  *
6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7  *
8  * RAID-1 management functions.
9  *
10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11  *
12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14  *
15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16  * bitmapped intelligence in resync:
17  *
18  *      - bitmap marked during normal i/o
19  *      - bitmap used to skip nondirty blocks during sync
20  *
21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22  * - persistent bitmap code
23  *
24  * This program is free software; you can redistribute it and/or modify
25  * it under the terms of the GNU General Public License as published by
26  * the Free Software Foundation; either version 2, or (at your option)
27  * any later version.
28  *
29  * You should have received a copy of the GNU General Public License
30  * (for example /usr/src/linux/COPYING); if not, write to the Free
31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32  */
33 
34 #include <linux/delay.h>
35 #include <linux/blkdev.h>
36 #include <linux/seq_file.h>
37 #include "md.h"
38 #include "raid1.h"
39 #include "bitmap.h"
40 
41 #define DEBUG 0
42 #if DEBUG
43 #define PRINTK(x...) printk(x)
44 #else
45 #define PRINTK(x...)
46 #endif
47 
48 /*
49  * Number of guaranteed r1bios in case of extreme VM load:
50  */
51 #define	NR_RAID1_BIOS 256
52 
53 
54 static void unplug_slaves(mddev_t *mddev);
55 
56 static void allow_barrier(conf_t *conf);
57 static void lower_barrier(conf_t *conf);
58 
59 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
60 {
61 	struct pool_info *pi = data;
62 	r1bio_t *r1_bio;
63 	int size = offsetof(r1bio_t, bios[pi->raid_disks]);
64 
65 	/* allocate a r1bio with room for raid_disks entries in the bios array */
66 	r1_bio = kzalloc(size, gfp_flags);
67 	if (!r1_bio)
68 		unplug_slaves(pi->mddev);
69 
70 	return r1_bio;
71 }
72 
73 static void r1bio_pool_free(void *r1_bio, void *data)
74 {
75 	kfree(r1_bio);
76 }
77 
78 #define RESYNC_BLOCK_SIZE (64*1024)
79 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
80 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
81 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
82 #define RESYNC_WINDOW (2048*1024)
83 
84 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
85 {
86 	struct pool_info *pi = data;
87 	struct page *page;
88 	r1bio_t *r1_bio;
89 	struct bio *bio;
90 	int i, j;
91 
92 	r1_bio = r1bio_pool_alloc(gfp_flags, pi);
93 	if (!r1_bio) {
94 		unplug_slaves(pi->mddev);
95 		return NULL;
96 	}
97 
98 	/*
99 	 * Allocate bios : 1 for reading, n-1 for writing
100 	 */
101 	for (j = pi->raid_disks ; j-- ; ) {
102 		bio = bio_alloc(gfp_flags, RESYNC_PAGES);
103 		if (!bio)
104 			goto out_free_bio;
105 		r1_bio->bios[j] = bio;
106 	}
107 	/*
108 	 * Allocate RESYNC_PAGES data pages and attach them to
109 	 * the first bio.
110 	 * If this is a user-requested check/repair, allocate
111 	 * RESYNC_PAGES for each bio.
112 	 */
113 	if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
114 		j = pi->raid_disks;
115 	else
116 		j = 1;
117 	while(j--) {
118 		bio = r1_bio->bios[j];
119 		for (i = 0; i < RESYNC_PAGES; i++) {
120 			page = alloc_page(gfp_flags);
121 			if (unlikely(!page))
122 				goto out_free_pages;
123 
124 			bio->bi_io_vec[i].bv_page = page;
125 			bio->bi_vcnt = i+1;
126 		}
127 	}
128 	/* If not user-requests, copy the page pointers to all bios */
129 	if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
130 		for (i=0; i<RESYNC_PAGES ; i++)
131 			for (j=1; j<pi->raid_disks; j++)
132 				r1_bio->bios[j]->bi_io_vec[i].bv_page =
133 					r1_bio->bios[0]->bi_io_vec[i].bv_page;
134 	}
135 
136 	r1_bio->master_bio = NULL;
137 
138 	return r1_bio;
139 
140 out_free_pages:
141 	for (j=0 ; j < pi->raid_disks; j++)
142 		for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
143 			put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
144 	j = -1;
145 out_free_bio:
146 	while ( ++j < pi->raid_disks )
147 		bio_put(r1_bio->bios[j]);
148 	r1bio_pool_free(r1_bio, data);
149 	return NULL;
150 }
151 
152 static void r1buf_pool_free(void *__r1_bio, void *data)
153 {
154 	struct pool_info *pi = data;
155 	int i,j;
156 	r1bio_t *r1bio = __r1_bio;
157 
158 	for (i = 0; i < RESYNC_PAGES; i++)
159 		for (j = pi->raid_disks; j-- ;) {
160 			if (j == 0 ||
161 			    r1bio->bios[j]->bi_io_vec[i].bv_page !=
162 			    r1bio->bios[0]->bi_io_vec[i].bv_page)
163 				safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
164 		}
165 	for (i=0 ; i < pi->raid_disks; i++)
166 		bio_put(r1bio->bios[i]);
167 
168 	r1bio_pool_free(r1bio, data);
169 }
170 
171 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
172 {
173 	int i;
174 
175 	for (i = 0; i < conf->raid_disks; i++) {
176 		struct bio **bio = r1_bio->bios + i;
177 		if (*bio && *bio != IO_BLOCKED)
178 			bio_put(*bio);
179 		*bio = NULL;
180 	}
181 }
182 
183 static void free_r1bio(r1bio_t *r1_bio)
184 {
185 	conf_t *conf = r1_bio->mddev->private;
186 
187 	/*
188 	 * Wake up any possible resync thread that waits for the device
189 	 * to go idle.
190 	 */
191 	allow_barrier(conf);
192 
193 	put_all_bios(conf, r1_bio);
194 	mempool_free(r1_bio, conf->r1bio_pool);
195 }
196 
197 static void put_buf(r1bio_t *r1_bio)
198 {
199 	conf_t *conf = r1_bio->mddev->private;
200 	int i;
201 
202 	for (i=0; i<conf->raid_disks; i++) {
203 		struct bio *bio = r1_bio->bios[i];
204 		if (bio->bi_end_io)
205 			rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
206 	}
207 
208 	mempool_free(r1_bio, conf->r1buf_pool);
209 
210 	lower_barrier(conf);
211 }
212 
213 static void reschedule_retry(r1bio_t *r1_bio)
214 {
215 	unsigned long flags;
216 	mddev_t *mddev = r1_bio->mddev;
217 	conf_t *conf = mddev->private;
218 
219 	spin_lock_irqsave(&conf->device_lock, flags);
220 	list_add(&r1_bio->retry_list, &conf->retry_list);
221 	conf->nr_queued ++;
222 	spin_unlock_irqrestore(&conf->device_lock, flags);
223 
224 	wake_up(&conf->wait_barrier);
225 	md_wakeup_thread(mddev->thread);
226 }
227 
228 /*
229  * raid_end_bio_io() is called when we have finished servicing a mirrored
230  * operation and are ready to return a success/failure code to the buffer
231  * cache layer.
232  */
233 static void raid_end_bio_io(r1bio_t *r1_bio)
234 {
235 	struct bio *bio = r1_bio->master_bio;
236 
237 	/* if nobody has done the final endio yet, do it now */
238 	if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
239 		PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
240 			(bio_data_dir(bio) == WRITE) ? "write" : "read",
241 			(unsigned long long) bio->bi_sector,
242 			(unsigned long long) bio->bi_sector +
243 				(bio->bi_size >> 9) - 1);
244 
245 		bio_endio(bio,
246 			test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
247 	}
248 	free_r1bio(r1_bio);
249 }
250 
251 /*
252  * Update disk head position estimator based on IRQ completion info.
253  */
254 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
255 {
256 	conf_t *conf = r1_bio->mddev->private;
257 
258 	conf->mirrors[disk].head_position =
259 		r1_bio->sector + (r1_bio->sectors);
260 }
261 
262 static void raid1_end_read_request(struct bio *bio, int error)
263 {
264 	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
265 	r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
266 	int mirror;
267 	conf_t *conf = r1_bio->mddev->private;
268 
269 	mirror = r1_bio->read_disk;
270 	/*
271 	 * this branch is our 'one mirror IO has finished' event handler:
272 	 */
273 	update_head_pos(mirror, r1_bio);
274 
275 	if (uptodate)
276 		set_bit(R1BIO_Uptodate, &r1_bio->state);
277 	else {
278 		/* If all other devices have failed, we want to return
279 		 * the error upwards rather than fail the last device.
280 		 * Here we redefine "uptodate" to mean "Don't want to retry"
281 		 */
282 		unsigned long flags;
283 		spin_lock_irqsave(&conf->device_lock, flags);
284 		if (r1_bio->mddev->degraded == conf->raid_disks ||
285 		    (r1_bio->mddev->degraded == conf->raid_disks-1 &&
286 		     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
287 			uptodate = 1;
288 		spin_unlock_irqrestore(&conf->device_lock, flags);
289 	}
290 
291 	if (uptodate)
292 		raid_end_bio_io(r1_bio);
293 	else {
294 		/*
295 		 * oops, read error:
296 		 */
297 		char b[BDEVNAME_SIZE];
298 		if (printk_ratelimit())
299 			printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
300 			       bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
301 		reschedule_retry(r1_bio);
302 	}
303 
304 	rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
305 }
306 
307 static void raid1_end_write_request(struct bio *bio, int error)
308 {
309 	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310 	r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
311 	int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
312 	conf_t *conf = r1_bio->mddev->private;
313 	struct bio *to_put = NULL;
314 
315 
316 	for (mirror = 0; mirror < conf->raid_disks; mirror++)
317 		if (r1_bio->bios[mirror] == bio)
318 			break;
319 
320 	if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
321 		set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
322 		set_bit(R1BIO_BarrierRetry, &r1_bio->state);
323 		r1_bio->mddev->barriers_work = 0;
324 		/* Don't rdev_dec_pending in this branch - keep it for the retry */
325 	} else {
326 		/*
327 		 * this branch is our 'one mirror IO has finished' event handler:
328 		 */
329 		r1_bio->bios[mirror] = NULL;
330 		to_put = bio;
331 		if (!uptodate) {
332 			md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
333 			/* an I/O failed, we can't clear the bitmap */
334 			set_bit(R1BIO_Degraded, &r1_bio->state);
335 		} else
336 			/*
337 			 * Set R1BIO_Uptodate in our master bio, so that
338 			 * we will return a good error code for to the higher
339 			 * levels even if IO on some other mirrored buffer fails.
340 			 *
341 			 * The 'master' represents the composite IO operation to
342 			 * user-side. So if something waits for IO, then it will
343 			 * wait for the 'master' bio.
344 			 */
345 			set_bit(R1BIO_Uptodate, &r1_bio->state);
346 
347 		update_head_pos(mirror, r1_bio);
348 
349 		if (behind) {
350 			if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
351 				atomic_dec(&r1_bio->behind_remaining);
352 
353 			/* In behind mode, we ACK the master bio once the I/O has safely
354 			 * reached all non-writemostly disks. Setting the Returned bit
355 			 * ensures that this gets done only once -- we don't ever want to
356 			 * return -EIO here, instead we'll wait */
357 
358 			if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
359 			    test_bit(R1BIO_Uptodate, &r1_bio->state)) {
360 				/* Maybe we can return now */
361 				if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
362 					struct bio *mbio = r1_bio->master_bio;
363 					PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
364 					       (unsigned long long) mbio->bi_sector,
365 					       (unsigned long long) mbio->bi_sector +
366 					       (mbio->bi_size >> 9) - 1);
367 					bio_endio(mbio, 0);
368 				}
369 			}
370 		}
371 		rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
372 	}
373 	/*
374 	 *
375 	 * Let's see if all mirrored write operations have finished
376 	 * already.
377 	 */
378 	if (atomic_dec_and_test(&r1_bio->remaining)) {
379 		if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
380 			reschedule_retry(r1_bio);
381 		else {
382 			/* it really is the end of this request */
383 			if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
384 				/* free extra copy of the data pages */
385 				int i = bio->bi_vcnt;
386 				while (i--)
387 					safe_put_page(bio->bi_io_vec[i].bv_page);
388 			}
389 			/* clear the bitmap if all writes complete successfully */
390 			bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
391 					r1_bio->sectors,
392 					!test_bit(R1BIO_Degraded, &r1_bio->state),
393 					behind);
394 			md_write_end(r1_bio->mddev);
395 			raid_end_bio_io(r1_bio);
396 		}
397 	}
398 
399 	if (to_put)
400 		bio_put(to_put);
401 }
402 
403 
404 /*
405  * This routine returns the disk from which the requested read should
406  * be done. There is a per-array 'next expected sequential IO' sector
407  * number - if this matches on the next IO then we use the last disk.
408  * There is also a per-disk 'last know head position' sector that is
409  * maintained from IRQ contexts, both the normal and the resync IO
410  * completion handlers update this position correctly. If there is no
411  * perfect sequential match then we pick the disk whose head is closest.
412  *
413  * If there are 2 mirrors in the same 2 devices, performance degrades
414  * because position is mirror, not device based.
415  *
416  * The rdev for the device selected will have nr_pending incremented.
417  */
418 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
419 {
420 	const unsigned long this_sector = r1_bio->sector;
421 	int new_disk = conf->last_used, disk = new_disk;
422 	int wonly_disk = -1;
423 	const int sectors = r1_bio->sectors;
424 	sector_t new_distance, current_distance;
425 	mdk_rdev_t *rdev;
426 
427 	rcu_read_lock();
428 	/*
429 	 * Check if we can balance. We can balance on the whole
430 	 * device if no resync is going on, or below the resync window.
431 	 * We take the first readable disk when above the resync window.
432 	 */
433  retry:
434 	if (conf->mddev->recovery_cp < MaxSector &&
435 	    (this_sector + sectors >= conf->next_resync)) {
436 		/* Choose the first operation device, for consistancy */
437 		new_disk = 0;
438 
439 		for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
440 		     r1_bio->bios[new_disk] == IO_BLOCKED ||
441 		     !rdev || !test_bit(In_sync, &rdev->flags)
442 			     || test_bit(WriteMostly, &rdev->flags);
443 		     rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
444 
445 			if (rdev && test_bit(In_sync, &rdev->flags) &&
446 				r1_bio->bios[new_disk] != IO_BLOCKED)
447 				wonly_disk = new_disk;
448 
449 			if (new_disk == conf->raid_disks - 1) {
450 				new_disk = wonly_disk;
451 				break;
452 			}
453 		}
454 		goto rb_out;
455 	}
456 
457 
458 	/* make sure the disk is operational */
459 	for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
460 	     r1_bio->bios[new_disk] == IO_BLOCKED ||
461 	     !rdev || !test_bit(In_sync, &rdev->flags) ||
462 		     test_bit(WriteMostly, &rdev->flags);
463 	     rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
464 
465 		if (rdev && test_bit(In_sync, &rdev->flags) &&
466 		    r1_bio->bios[new_disk] != IO_BLOCKED)
467 			wonly_disk = new_disk;
468 
469 		if (new_disk <= 0)
470 			new_disk = conf->raid_disks;
471 		new_disk--;
472 		if (new_disk == disk) {
473 			new_disk = wonly_disk;
474 			break;
475 		}
476 	}
477 
478 	if (new_disk < 0)
479 		goto rb_out;
480 
481 	disk = new_disk;
482 	/* now disk == new_disk == starting point for search */
483 
484 	/*
485 	 * Don't change to another disk for sequential reads:
486 	 */
487 	if (conf->next_seq_sect == this_sector)
488 		goto rb_out;
489 	if (this_sector == conf->mirrors[new_disk].head_position)
490 		goto rb_out;
491 
492 	current_distance = abs(this_sector - conf->mirrors[disk].head_position);
493 
494 	/* Find the disk whose head is closest */
495 
496 	do {
497 		if (disk <= 0)
498 			disk = conf->raid_disks;
499 		disk--;
500 
501 		rdev = rcu_dereference(conf->mirrors[disk].rdev);
502 
503 		if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
504 		    !test_bit(In_sync, &rdev->flags) ||
505 		    test_bit(WriteMostly, &rdev->flags))
506 			continue;
507 
508 		if (!atomic_read(&rdev->nr_pending)) {
509 			new_disk = disk;
510 			break;
511 		}
512 		new_distance = abs(this_sector - conf->mirrors[disk].head_position);
513 		if (new_distance < current_distance) {
514 			current_distance = new_distance;
515 			new_disk = disk;
516 		}
517 	} while (disk != conf->last_used);
518 
519  rb_out:
520 
521 
522 	if (new_disk >= 0) {
523 		rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
524 		if (!rdev)
525 			goto retry;
526 		atomic_inc(&rdev->nr_pending);
527 		if (!test_bit(In_sync, &rdev->flags)) {
528 			/* cannot risk returning a device that failed
529 			 * before we inc'ed nr_pending
530 			 */
531 			rdev_dec_pending(rdev, conf->mddev);
532 			goto retry;
533 		}
534 		conf->next_seq_sect = this_sector + sectors;
535 		conf->last_used = new_disk;
536 	}
537 	rcu_read_unlock();
538 
539 	return new_disk;
540 }
541 
542 static void unplug_slaves(mddev_t *mddev)
543 {
544 	conf_t *conf = mddev->private;
545 	int i;
546 
547 	rcu_read_lock();
548 	for (i=0; i<mddev->raid_disks; i++) {
549 		mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
550 		if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
551 			struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
552 
553 			atomic_inc(&rdev->nr_pending);
554 			rcu_read_unlock();
555 
556 			blk_unplug(r_queue);
557 
558 			rdev_dec_pending(rdev, mddev);
559 			rcu_read_lock();
560 		}
561 	}
562 	rcu_read_unlock();
563 }
564 
565 static void raid1_unplug(struct request_queue *q)
566 {
567 	mddev_t *mddev = q->queuedata;
568 
569 	unplug_slaves(mddev);
570 	md_wakeup_thread(mddev->thread);
571 }
572 
573 static int raid1_congested(void *data, int bits)
574 {
575 	mddev_t *mddev = data;
576 	conf_t *conf = mddev->private;
577 	int i, ret = 0;
578 
579 	rcu_read_lock();
580 	for (i = 0; i < mddev->raid_disks; i++) {
581 		mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
582 		if (rdev && !test_bit(Faulty, &rdev->flags)) {
583 			struct request_queue *q = bdev_get_queue(rdev->bdev);
584 
585 			/* Note the '|| 1' - when read_balance prefers
586 			 * non-congested targets, it can be removed
587 			 */
588 			if ((bits & (1<<BDI_async_congested)) || 1)
589 				ret |= bdi_congested(&q->backing_dev_info, bits);
590 			else
591 				ret &= bdi_congested(&q->backing_dev_info, bits);
592 		}
593 	}
594 	rcu_read_unlock();
595 	return ret;
596 }
597 
598 
599 static int flush_pending_writes(conf_t *conf)
600 {
601 	/* Any writes that have been queued but are awaiting
602 	 * bitmap updates get flushed here.
603 	 * We return 1 if any requests were actually submitted.
604 	 */
605 	int rv = 0;
606 
607 	spin_lock_irq(&conf->device_lock);
608 
609 	if (conf->pending_bio_list.head) {
610 		struct bio *bio;
611 		bio = bio_list_get(&conf->pending_bio_list);
612 		blk_remove_plug(conf->mddev->queue);
613 		spin_unlock_irq(&conf->device_lock);
614 		/* flush any pending bitmap writes to
615 		 * disk before proceeding w/ I/O */
616 		bitmap_unplug(conf->mddev->bitmap);
617 
618 		while (bio) { /* submit pending writes */
619 			struct bio *next = bio->bi_next;
620 			bio->bi_next = NULL;
621 			generic_make_request(bio);
622 			bio = next;
623 		}
624 		rv = 1;
625 	} else
626 		spin_unlock_irq(&conf->device_lock);
627 	return rv;
628 }
629 
630 /* Barriers....
631  * Sometimes we need to suspend IO while we do something else,
632  * either some resync/recovery, or reconfigure the array.
633  * To do this we raise a 'barrier'.
634  * The 'barrier' is a counter that can be raised multiple times
635  * to count how many activities are happening which preclude
636  * normal IO.
637  * We can only raise the barrier if there is no pending IO.
638  * i.e. if nr_pending == 0.
639  * We choose only to raise the barrier if no-one is waiting for the
640  * barrier to go down.  This means that as soon as an IO request
641  * is ready, no other operations which require a barrier will start
642  * until the IO request has had a chance.
643  *
644  * So: regular IO calls 'wait_barrier'.  When that returns there
645  *    is no backgroup IO happening,  It must arrange to call
646  *    allow_barrier when it has finished its IO.
647  * backgroup IO calls must call raise_barrier.  Once that returns
648  *    there is no normal IO happeing.  It must arrange to call
649  *    lower_barrier when the particular background IO completes.
650  */
651 #define RESYNC_DEPTH 32
652 
653 static void raise_barrier(conf_t *conf)
654 {
655 	spin_lock_irq(&conf->resync_lock);
656 
657 	/* Wait until no block IO is waiting */
658 	wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
659 			    conf->resync_lock,
660 			    raid1_unplug(conf->mddev->queue));
661 
662 	/* block any new IO from starting */
663 	conf->barrier++;
664 
665 	/* No wait for all pending IO to complete */
666 	wait_event_lock_irq(conf->wait_barrier,
667 			    !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
668 			    conf->resync_lock,
669 			    raid1_unplug(conf->mddev->queue));
670 
671 	spin_unlock_irq(&conf->resync_lock);
672 }
673 
674 static void lower_barrier(conf_t *conf)
675 {
676 	unsigned long flags;
677 	spin_lock_irqsave(&conf->resync_lock, flags);
678 	conf->barrier--;
679 	spin_unlock_irqrestore(&conf->resync_lock, flags);
680 	wake_up(&conf->wait_barrier);
681 }
682 
683 static void wait_barrier(conf_t *conf)
684 {
685 	spin_lock_irq(&conf->resync_lock);
686 	if (conf->barrier) {
687 		conf->nr_waiting++;
688 		wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
689 				    conf->resync_lock,
690 				    raid1_unplug(conf->mddev->queue));
691 		conf->nr_waiting--;
692 	}
693 	conf->nr_pending++;
694 	spin_unlock_irq(&conf->resync_lock);
695 }
696 
697 static void allow_barrier(conf_t *conf)
698 {
699 	unsigned long flags;
700 	spin_lock_irqsave(&conf->resync_lock, flags);
701 	conf->nr_pending--;
702 	spin_unlock_irqrestore(&conf->resync_lock, flags);
703 	wake_up(&conf->wait_barrier);
704 }
705 
706 static void freeze_array(conf_t *conf)
707 {
708 	/* stop syncio and normal IO and wait for everything to
709 	 * go quite.
710 	 * We increment barrier and nr_waiting, and then
711 	 * wait until nr_pending match nr_queued+1
712 	 * This is called in the context of one normal IO request
713 	 * that has failed. Thus any sync request that might be pending
714 	 * will be blocked by nr_pending, and we need to wait for
715 	 * pending IO requests to complete or be queued for re-try.
716 	 * Thus the number queued (nr_queued) plus this request (1)
717 	 * must match the number of pending IOs (nr_pending) before
718 	 * we continue.
719 	 */
720 	spin_lock_irq(&conf->resync_lock);
721 	conf->barrier++;
722 	conf->nr_waiting++;
723 	wait_event_lock_irq(conf->wait_barrier,
724 			    conf->nr_pending == conf->nr_queued+1,
725 			    conf->resync_lock,
726 			    ({ flush_pending_writes(conf);
727 			       raid1_unplug(conf->mddev->queue); }));
728 	spin_unlock_irq(&conf->resync_lock);
729 }
730 static void unfreeze_array(conf_t *conf)
731 {
732 	/* reverse the effect of the freeze */
733 	spin_lock_irq(&conf->resync_lock);
734 	conf->barrier--;
735 	conf->nr_waiting--;
736 	wake_up(&conf->wait_barrier);
737 	spin_unlock_irq(&conf->resync_lock);
738 }
739 
740 
741 /* duplicate the data pages for behind I/O */
742 static struct page **alloc_behind_pages(struct bio *bio)
743 {
744 	int i;
745 	struct bio_vec *bvec;
746 	struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
747 					GFP_NOIO);
748 	if (unlikely(!pages))
749 		goto do_sync_io;
750 
751 	bio_for_each_segment(bvec, bio, i) {
752 		pages[i] = alloc_page(GFP_NOIO);
753 		if (unlikely(!pages[i]))
754 			goto do_sync_io;
755 		memcpy(kmap(pages[i]) + bvec->bv_offset,
756 			kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
757 		kunmap(pages[i]);
758 		kunmap(bvec->bv_page);
759 	}
760 
761 	return pages;
762 
763 do_sync_io:
764 	if (pages)
765 		for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
766 			put_page(pages[i]);
767 	kfree(pages);
768 	PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
769 	return NULL;
770 }
771 
772 static int make_request(struct request_queue *q, struct bio * bio)
773 {
774 	mddev_t *mddev = q->queuedata;
775 	conf_t *conf = mddev->private;
776 	mirror_info_t *mirror;
777 	r1bio_t *r1_bio;
778 	struct bio *read_bio;
779 	int i, targets = 0, disks;
780 	struct bitmap *bitmap;
781 	unsigned long flags;
782 	struct bio_list bl;
783 	struct page **behind_pages = NULL;
784 	const int rw = bio_data_dir(bio);
785 	const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
786 	int cpu;
787 	bool do_barriers;
788 	mdk_rdev_t *blocked_rdev;
789 
790 	/*
791 	 * Register the new request and wait if the reconstruction
792 	 * thread has put up a bar for new requests.
793 	 * Continue immediately if no resync is active currently.
794 	 * We test barriers_work *after* md_write_start as md_write_start
795 	 * may cause the first superblock write, and that will check out
796 	 * if barriers work.
797 	 */
798 
799 	md_write_start(mddev, bio); /* wait on superblock update early */
800 
801 	if (unlikely(!mddev->barriers_work &&
802 		     bio_rw_flagged(bio, BIO_RW_BARRIER))) {
803 		if (rw == WRITE)
804 			md_write_end(mddev);
805 		bio_endio(bio, -EOPNOTSUPP);
806 		return 0;
807 	}
808 
809 	wait_barrier(conf);
810 
811 	bitmap = mddev->bitmap;
812 
813 	cpu = part_stat_lock();
814 	part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
815 	part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
816 		      bio_sectors(bio));
817 	part_stat_unlock();
818 
819 	/*
820 	 * make_request() can abort the operation when READA is being
821 	 * used and no empty request is available.
822 	 *
823 	 */
824 	r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
825 
826 	r1_bio->master_bio = bio;
827 	r1_bio->sectors = bio->bi_size >> 9;
828 	r1_bio->state = 0;
829 	r1_bio->mddev = mddev;
830 	r1_bio->sector = bio->bi_sector;
831 
832 	if (rw == READ) {
833 		/*
834 		 * read balancing logic:
835 		 */
836 		int rdisk = read_balance(conf, r1_bio);
837 
838 		if (rdisk < 0) {
839 			/* couldn't find anywhere to read from */
840 			raid_end_bio_io(r1_bio);
841 			return 0;
842 		}
843 		mirror = conf->mirrors + rdisk;
844 
845 		r1_bio->read_disk = rdisk;
846 
847 		read_bio = bio_clone(bio, GFP_NOIO);
848 
849 		r1_bio->bios[rdisk] = read_bio;
850 
851 		read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
852 		read_bio->bi_bdev = mirror->rdev->bdev;
853 		read_bio->bi_end_io = raid1_end_read_request;
854 		read_bio->bi_rw = READ | do_sync;
855 		read_bio->bi_private = r1_bio;
856 
857 		generic_make_request(read_bio);
858 		return 0;
859 	}
860 
861 	/*
862 	 * WRITE:
863 	 */
864 	/* first select target devices under spinlock and
865 	 * inc refcount on their rdev.  Record them by setting
866 	 * bios[x] to bio
867 	 */
868 	disks = conf->raid_disks;
869 #if 0
870 	{ static int first=1;
871 	if (first) printk("First Write sector %llu disks %d\n",
872 			  (unsigned long long)r1_bio->sector, disks);
873 	first = 0;
874 	}
875 #endif
876  retry_write:
877 	blocked_rdev = NULL;
878 	rcu_read_lock();
879 	for (i = 0;  i < disks; i++) {
880 		mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
881 		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
882 			atomic_inc(&rdev->nr_pending);
883 			blocked_rdev = rdev;
884 			break;
885 		}
886 		if (rdev && !test_bit(Faulty, &rdev->flags)) {
887 			atomic_inc(&rdev->nr_pending);
888 			if (test_bit(Faulty, &rdev->flags)) {
889 				rdev_dec_pending(rdev, mddev);
890 				r1_bio->bios[i] = NULL;
891 			} else
892 				r1_bio->bios[i] = bio;
893 			targets++;
894 		} else
895 			r1_bio->bios[i] = NULL;
896 	}
897 	rcu_read_unlock();
898 
899 	if (unlikely(blocked_rdev)) {
900 		/* Wait for this device to become unblocked */
901 		int j;
902 
903 		for (j = 0; j < i; j++)
904 			if (r1_bio->bios[j])
905 				rdev_dec_pending(conf->mirrors[j].rdev, mddev);
906 
907 		allow_barrier(conf);
908 		md_wait_for_blocked_rdev(blocked_rdev, mddev);
909 		wait_barrier(conf);
910 		goto retry_write;
911 	}
912 
913 	BUG_ON(targets == 0); /* we never fail the last device */
914 
915 	if (targets < conf->raid_disks) {
916 		/* array is degraded, we will not clear the bitmap
917 		 * on I/O completion (see raid1_end_write_request) */
918 		set_bit(R1BIO_Degraded, &r1_bio->state);
919 	}
920 
921 	/* do behind I/O ? */
922 	if (bitmap &&
923 	    atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
924 	    (behind_pages = alloc_behind_pages(bio)) != NULL)
925 		set_bit(R1BIO_BehindIO, &r1_bio->state);
926 
927 	atomic_set(&r1_bio->remaining, 0);
928 	atomic_set(&r1_bio->behind_remaining, 0);
929 
930 	do_barriers = bio_rw_flagged(bio, BIO_RW_BARRIER);
931 	if (do_barriers)
932 		set_bit(R1BIO_Barrier, &r1_bio->state);
933 
934 	bio_list_init(&bl);
935 	for (i = 0; i < disks; i++) {
936 		struct bio *mbio;
937 		if (!r1_bio->bios[i])
938 			continue;
939 
940 		mbio = bio_clone(bio, GFP_NOIO);
941 		r1_bio->bios[i] = mbio;
942 
943 		mbio->bi_sector	= r1_bio->sector + conf->mirrors[i].rdev->data_offset;
944 		mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
945 		mbio->bi_end_io	= raid1_end_write_request;
946 		mbio->bi_rw = WRITE | do_barriers | do_sync;
947 		mbio->bi_private = r1_bio;
948 
949 		if (behind_pages) {
950 			struct bio_vec *bvec;
951 			int j;
952 
953 			/* Yes, I really want the '__' version so that
954 			 * we clear any unused pointer in the io_vec, rather
955 			 * than leave them unchanged.  This is important
956 			 * because when we come to free the pages, we won't
957 			 * know the originial bi_idx, so we just free
958 			 * them all
959 			 */
960 			__bio_for_each_segment(bvec, mbio, j, 0)
961 				bvec->bv_page = behind_pages[j];
962 			if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
963 				atomic_inc(&r1_bio->behind_remaining);
964 		}
965 
966 		atomic_inc(&r1_bio->remaining);
967 
968 		bio_list_add(&bl, mbio);
969 	}
970 	kfree(behind_pages); /* the behind pages are attached to the bios now */
971 
972 	bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
973 				test_bit(R1BIO_BehindIO, &r1_bio->state));
974 	spin_lock_irqsave(&conf->device_lock, flags);
975 	bio_list_merge(&conf->pending_bio_list, &bl);
976 	bio_list_init(&bl);
977 
978 	blk_plug_device(mddev->queue);
979 	spin_unlock_irqrestore(&conf->device_lock, flags);
980 
981 	/* In case raid1d snuck into freeze_array */
982 	wake_up(&conf->wait_barrier);
983 
984 	if (do_sync)
985 		md_wakeup_thread(mddev->thread);
986 #if 0
987 	while ((bio = bio_list_pop(&bl)) != NULL)
988 		generic_make_request(bio);
989 #endif
990 
991 	return 0;
992 }
993 
994 static void status(struct seq_file *seq, mddev_t *mddev)
995 {
996 	conf_t *conf = mddev->private;
997 	int i;
998 
999 	seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1000 		   conf->raid_disks - mddev->degraded);
1001 	rcu_read_lock();
1002 	for (i = 0; i < conf->raid_disks; i++) {
1003 		mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1004 		seq_printf(seq, "%s",
1005 			   rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1006 	}
1007 	rcu_read_unlock();
1008 	seq_printf(seq, "]");
1009 }
1010 
1011 
1012 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1013 {
1014 	char b[BDEVNAME_SIZE];
1015 	conf_t *conf = mddev->private;
1016 
1017 	/*
1018 	 * If it is not operational, then we have already marked it as dead
1019 	 * else if it is the last working disks, ignore the error, let the
1020 	 * next level up know.
1021 	 * else mark the drive as failed
1022 	 */
1023 	if (test_bit(In_sync, &rdev->flags)
1024 	    && (conf->raid_disks - mddev->degraded) == 1) {
1025 		/*
1026 		 * Don't fail the drive, act as though we were just a
1027 		 * normal single drive.
1028 		 * However don't try a recovery from this drive as
1029 		 * it is very likely to fail.
1030 		 */
1031 		mddev->recovery_disabled = 1;
1032 		return;
1033 	}
1034 	if (test_and_clear_bit(In_sync, &rdev->flags)) {
1035 		unsigned long flags;
1036 		spin_lock_irqsave(&conf->device_lock, flags);
1037 		mddev->degraded++;
1038 		set_bit(Faulty, &rdev->flags);
1039 		spin_unlock_irqrestore(&conf->device_lock, flags);
1040 		/*
1041 		 * if recovery is running, make sure it aborts.
1042 		 */
1043 		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1044 	} else
1045 		set_bit(Faulty, &rdev->flags);
1046 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
1047 	printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1048 		"raid1: Operation continuing on %d devices.\n",
1049 		bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1050 }
1051 
1052 static void print_conf(conf_t *conf)
1053 {
1054 	int i;
1055 
1056 	printk("RAID1 conf printout:\n");
1057 	if (!conf) {
1058 		printk("(!conf)\n");
1059 		return;
1060 	}
1061 	printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1062 		conf->raid_disks);
1063 
1064 	rcu_read_lock();
1065 	for (i = 0; i < conf->raid_disks; i++) {
1066 		char b[BDEVNAME_SIZE];
1067 		mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1068 		if (rdev)
1069 			printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1070 			       i, !test_bit(In_sync, &rdev->flags),
1071 			       !test_bit(Faulty, &rdev->flags),
1072 			       bdevname(rdev->bdev,b));
1073 	}
1074 	rcu_read_unlock();
1075 }
1076 
1077 static void close_sync(conf_t *conf)
1078 {
1079 	wait_barrier(conf);
1080 	allow_barrier(conf);
1081 
1082 	mempool_destroy(conf->r1buf_pool);
1083 	conf->r1buf_pool = NULL;
1084 }
1085 
1086 static int raid1_spare_active(mddev_t *mddev)
1087 {
1088 	int i;
1089 	conf_t *conf = mddev->private;
1090 
1091 	/*
1092 	 * Find all failed disks within the RAID1 configuration
1093 	 * and mark them readable.
1094 	 * Called under mddev lock, so rcu protection not needed.
1095 	 */
1096 	for (i = 0; i < conf->raid_disks; i++) {
1097 		mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1098 		if (rdev
1099 		    && !test_bit(Faulty, &rdev->flags)
1100 		    && !test_and_set_bit(In_sync, &rdev->flags)) {
1101 			unsigned long flags;
1102 			spin_lock_irqsave(&conf->device_lock, flags);
1103 			mddev->degraded--;
1104 			spin_unlock_irqrestore(&conf->device_lock, flags);
1105 		}
1106 	}
1107 
1108 	print_conf(conf);
1109 	return 0;
1110 }
1111 
1112 
1113 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1114 {
1115 	conf_t *conf = mddev->private;
1116 	int err = -EEXIST;
1117 	int mirror = 0;
1118 	mirror_info_t *p;
1119 	int first = 0;
1120 	int last = mddev->raid_disks - 1;
1121 
1122 	if (rdev->raid_disk >= 0)
1123 		first = last = rdev->raid_disk;
1124 
1125 	for (mirror = first; mirror <= last; mirror++)
1126 		if ( !(p=conf->mirrors+mirror)->rdev) {
1127 
1128 			disk_stack_limits(mddev->gendisk, rdev->bdev,
1129 					  rdev->data_offset << 9);
1130 			/* as we don't honour merge_bvec_fn, we must never risk
1131 			 * violating it, so limit ->max_sector to one PAGE, as
1132 			 * a one page request is never in violation.
1133 			 */
1134 			if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1135 			    queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
1136 				blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1137 
1138 			p->head_position = 0;
1139 			rdev->raid_disk = mirror;
1140 			err = 0;
1141 			/* As all devices are equivalent, we don't need a full recovery
1142 			 * if this was recently any drive of the array
1143 			 */
1144 			if (rdev->saved_raid_disk < 0)
1145 				conf->fullsync = 1;
1146 			rcu_assign_pointer(p->rdev, rdev);
1147 			break;
1148 		}
1149 	md_integrity_add_rdev(rdev, mddev);
1150 	print_conf(conf);
1151 	return err;
1152 }
1153 
1154 static int raid1_remove_disk(mddev_t *mddev, int number)
1155 {
1156 	conf_t *conf = mddev->private;
1157 	int err = 0;
1158 	mdk_rdev_t *rdev;
1159 	mirror_info_t *p = conf->mirrors+ number;
1160 
1161 	print_conf(conf);
1162 	rdev = p->rdev;
1163 	if (rdev) {
1164 		if (test_bit(In_sync, &rdev->flags) ||
1165 		    atomic_read(&rdev->nr_pending)) {
1166 			err = -EBUSY;
1167 			goto abort;
1168 		}
1169 		/* Only remove non-faulty devices is recovery
1170 		 * is not possible.
1171 		 */
1172 		if (!test_bit(Faulty, &rdev->flags) &&
1173 		    mddev->degraded < conf->raid_disks) {
1174 			err = -EBUSY;
1175 			goto abort;
1176 		}
1177 		p->rdev = NULL;
1178 		synchronize_rcu();
1179 		if (atomic_read(&rdev->nr_pending)) {
1180 			/* lost the race, try later */
1181 			err = -EBUSY;
1182 			p->rdev = rdev;
1183 			goto abort;
1184 		}
1185 		md_integrity_register(mddev);
1186 	}
1187 abort:
1188 
1189 	print_conf(conf);
1190 	return err;
1191 }
1192 
1193 
1194 static void end_sync_read(struct bio *bio, int error)
1195 {
1196 	r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1197 	int i;
1198 
1199 	for (i=r1_bio->mddev->raid_disks; i--; )
1200 		if (r1_bio->bios[i] == bio)
1201 			break;
1202 	BUG_ON(i < 0);
1203 	update_head_pos(i, r1_bio);
1204 	/*
1205 	 * we have read a block, now it needs to be re-written,
1206 	 * or re-read if the read failed.
1207 	 * We don't do much here, just schedule handling by raid1d
1208 	 */
1209 	if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1210 		set_bit(R1BIO_Uptodate, &r1_bio->state);
1211 
1212 	if (atomic_dec_and_test(&r1_bio->remaining))
1213 		reschedule_retry(r1_bio);
1214 }
1215 
1216 static void end_sync_write(struct bio *bio, int error)
1217 {
1218 	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1219 	r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1220 	mddev_t *mddev = r1_bio->mddev;
1221 	conf_t *conf = mddev->private;
1222 	int i;
1223 	int mirror=0;
1224 
1225 	for (i = 0; i < conf->raid_disks; i++)
1226 		if (r1_bio->bios[i] == bio) {
1227 			mirror = i;
1228 			break;
1229 		}
1230 	if (!uptodate) {
1231 		int sync_blocks = 0;
1232 		sector_t s = r1_bio->sector;
1233 		long sectors_to_go = r1_bio->sectors;
1234 		/* make sure these bits doesn't get cleared. */
1235 		do {
1236 			bitmap_end_sync(mddev->bitmap, s,
1237 					&sync_blocks, 1);
1238 			s += sync_blocks;
1239 			sectors_to_go -= sync_blocks;
1240 		} while (sectors_to_go > 0);
1241 		md_error(mddev, conf->mirrors[mirror].rdev);
1242 	}
1243 
1244 	update_head_pos(mirror, r1_bio);
1245 
1246 	if (atomic_dec_and_test(&r1_bio->remaining)) {
1247 		sector_t s = r1_bio->sectors;
1248 		put_buf(r1_bio);
1249 		md_done_sync(mddev, s, uptodate);
1250 	}
1251 }
1252 
1253 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1254 {
1255 	conf_t *conf = mddev->private;
1256 	int i;
1257 	int disks = conf->raid_disks;
1258 	struct bio *bio, *wbio;
1259 
1260 	bio = r1_bio->bios[r1_bio->read_disk];
1261 
1262 
1263 	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1264 		/* We have read all readable devices.  If we haven't
1265 		 * got the block, then there is no hope left.
1266 		 * If we have, then we want to do a comparison
1267 		 * and skip the write if everything is the same.
1268 		 * If any blocks failed to read, then we need to
1269 		 * attempt an over-write
1270 		 */
1271 		int primary;
1272 		if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1273 			for (i=0; i<mddev->raid_disks; i++)
1274 				if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1275 					md_error(mddev, conf->mirrors[i].rdev);
1276 
1277 			md_done_sync(mddev, r1_bio->sectors, 1);
1278 			put_buf(r1_bio);
1279 			return;
1280 		}
1281 		for (primary=0; primary<mddev->raid_disks; primary++)
1282 			if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1283 			    test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1284 				r1_bio->bios[primary]->bi_end_io = NULL;
1285 				rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1286 				break;
1287 			}
1288 		r1_bio->read_disk = primary;
1289 		for (i=0; i<mddev->raid_disks; i++)
1290 			if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1291 				int j;
1292 				int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1293 				struct bio *pbio = r1_bio->bios[primary];
1294 				struct bio *sbio = r1_bio->bios[i];
1295 
1296 				if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1297 					for (j = vcnt; j-- ; ) {
1298 						struct page *p, *s;
1299 						p = pbio->bi_io_vec[j].bv_page;
1300 						s = sbio->bi_io_vec[j].bv_page;
1301 						if (memcmp(page_address(p),
1302 							   page_address(s),
1303 							   PAGE_SIZE))
1304 							break;
1305 					}
1306 				} else
1307 					j = 0;
1308 				if (j >= 0)
1309 					mddev->resync_mismatches += r1_bio->sectors;
1310 				if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1311 					      && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1312 					sbio->bi_end_io = NULL;
1313 					rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1314 				} else {
1315 					/* fixup the bio for reuse */
1316 					int size;
1317 					sbio->bi_vcnt = vcnt;
1318 					sbio->bi_size = r1_bio->sectors << 9;
1319 					sbio->bi_idx = 0;
1320 					sbio->bi_phys_segments = 0;
1321 					sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1322 					sbio->bi_flags |= 1 << BIO_UPTODATE;
1323 					sbio->bi_next = NULL;
1324 					sbio->bi_sector = r1_bio->sector +
1325 						conf->mirrors[i].rdev->data_offset;
1326 					sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1327 					size = sbio->bi_size;
1328 					for (j = 0; j < vcnt ; j++) {
1329 						struct bio_vec *bi;
1330 						bi = &sbio->bi_io_vec[j];
1331 						bi->bv_offset = 0;
1332 						if (size > PAGE_SIZE)
1333 							bi->bv_len = PAGE_SIZE;
1334 						else
1335 							bi->bv_len = size;
1336 						size -= PAGE_SIZE;
1337 						memcpy(page_address(bi->bv_page),
1338 						       page_address(pbio->bi_io_vec[j].bv_page),
1339 						       PAGE_SIZE);
1340 					}
1341 
1342 				}
1343 			}
1344 	}
1345 	if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1346 		/* ouch - failed to read all of that.
1347 		 * Try some synchronous reads of other devices to get
1348 		 * good data, much like with normal read errors.  Only
1349 		 * read into the pages we already have so we don't
1350 		 * need to re-issue the read request.
1351 		 * We don't need to freeze the array, because being in an
1352 		 * active sync request, there is no normal IO, and
1353 		 * no overlapping syncs.
1354 		 */
1355 		sector_t sect = r1_bio->sector;
1356 		int sectors = r1_bio->sectors;
1357 		int idx = 0;
1358 
1359 		while(sectors) {
1360 			int s = sectors;
1361 			int d = r1_bio->read_disk;
1362 			int success = 0;
1363 			mdk_rdev_t *rdev;
1364 
1365 			if (s > (PAGE_SIZE>>9))
1366 				s = PAGE_SIZE >> 9;
1367 			do {
1368 				if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1369 					/* No rcu protection needed here devices
1370 					 * can only be removed when no resync is
1371 					 * active, and resync is currently active
1372 					 */
1373 					rdev = conf->mirrors[d].rdev;
1374 					if (sync_page_io(rdev->bdev,
1375 							 sect + rdev->data_offset,
1376 							 s<<9,
1377 							 bio->bi_io_vec[idx].bv_page,
1378 							 READ)) {
1379 						success = 1;
1380 						break;
1381 					}
1382 				}
1383 				d++;
1384 				if (d == conf->raid_disks)
1385 					d = 0;
1386 			} while (!success && d != r1_bio->read_disk);
1387 
1388 			if (success) {
1389 				int start = d;
1390 				/* write it back and re-read */
1391 				set_bit(R1BIO_Uptodate, &r1_bio->state);
1392 				while (d != r1_bio->read_disk) {
1393 					if (d == 0)
1394 						d = conf->raid_disks;
1395 					d--;
1396 					if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1397 						continue;
1398 					rdev = conf->mirrors[d].rdev;
1399 					atomic_add(s, &rdev->corrected_errors);
1400 					if (sync_page_io(rdev->bdev,
1401 							 sect + rdev->data_offset,
1402 							 s<<9,
1403 							 bio->bi_io_vec[idx].bv_page,
1404 							 WRITE) == 0)
1405 						md_error(mddev, rdev);
1406 				}
1407 				d = start;
1408 				while (d != r1_bio->read_disk) {
1409 					if (d == 0)
1410 						d = conf->raid_disks;
1411 					d--;
1412 					if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1413 						continue;
1414 					rdev = conf->mirrors[d].rdev;
1415 					if (sync_page_io(rdev->bdev,
1416 							 sect + rdev->data_offset,
1417 							 s<<9,
1418 							 bio->bi_io_vec[idx].bv_page,
1419 							 READ) == 0)
1420 						md_error(mddev, rdev);
1421 				}
1422 			} else {
1423 				char b[BDEVNAME_SIZE];
1424 				/* Cannot read from anywhere, array is toast */
1425 				md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1426 				printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1427 				       " for block %llu\n",
1428 				       bdevname(bio->bi_bdev,b),
1429 				       (unsigned long long)r1_bio->sector);
1430 				md_done_sync(mddev, r1_bio->sectors, 0);
1431 				put_buf(r1_bio);
1432 				return;
1433 			}
1434 			sectors -= s;
1435 			sect += s;
1436 			idx ++;
1437 		}
1438 	}
1439 
1440 	/*
1441 	 * schedule writes
1442 	 */
1443 	atomic_set(&r1_bio->remaining, 1);
1444 	for (i = 0; i < disks ; i++) {
1445 		wbio = r1_bio->bios[i];
1446 		if (wbio->bi_end_io == NULL ||
1447 		    (wbio->bi_end_io == end_sync_read &&
1448 		     (i == r1_bio->read_disk ||
1449 		      !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1450 			continue;
1451 
1452 		wbio->bi_rw = WRITE;
1453 		wbio->bi_end_io = end_sync_write;
1454 		atomic_inc(&r1_bio->remaining);
1455 		md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1456 
1457 		generic_make_request(wbio);
1458 	}
1459 
1460 	if (atomic_dec_and_test(&r1_bio->remaining)) {
1461 		/* if we're here, all write(s) have completed, so clean up */
1462 		md_done_sync(mddev, r1_bio->sectors, 1);
1463 		put_buf(r1_bio);
1464 	}
1465 }
1466 
1467 /*
1468  * This is a kernel thread which:
1469  *
1470  *	1.	Retries failed read operations on working mirrors.
1471  *	2.	Updates the raid superblock when problems encounter.
1472  *	3.	Performs writes following reads for array syncronising.
1473  */
1474 
1475 static void fix_read_error(conf_t *conf, int read_disk,
1476 			   sector_t sect, int sectors)
1477 {
1478 	mddev_t *mddev = conf->mddev;
1479 	while(sectors) {
1480 		int s = sectors;
1481 		int d = read_disk;
1482 		int success = 0;
1483 		int start;
1484 		mdk_rdev_t *rdev;
1485 
1486 		if (s > (PAGE_SIZE>>9))
1487 			s = PAGE_SIZE >> 9;
1488 
1489 		do {
1490 			/* Note: no rcu protection needed here
1491 			 * as this is synchronous in the raid1d thread
1492 			 * which is the thread that might remove
1493 			 * a device.  If raid1d ever becomes multi-threaded....
1494 			 */
1495 			rdev = conf->mirrors[d].rdev;
1496 			if (rdev &&
1497 			    test_bit(In_sync, &rdev->flags) &&
1498 			    sync_page_io(rdev->bdev,
1499 					 sect + rdev->data_offset,
1500 					 s<<9,
1501 					 conf->tmppage, READ))
1502 				success = 1;
1503 			else {
1504 				d++;
1505 				if (d == conf->raid_disks)
1506 					d = 0;
1507 			}
1508 		} while (!success && d != read_disk);
1509 
1510 		if (!success) {
1511 			/* Cannot read from anywhere -- bye bye array */
1512 			md_error(mddev, conf->mirrors[read_disk].rdev);
1513 			break;
1514 		}
1515 		/* write it back and re-read */
1516 		start = d;
1517 		while (d != read_disk) {
1518 			if (d==0)
1519 				d = conf->raid_disks;
1520 			d--;
1521 			rdev = conf->mirrors[d].rdev;
1522 			if (rdev &&
1523 			    test_bit(In_sync, &rdev->flags)) {
1524 				if (sync_page_io(rdev->bdev,
1525 						 sect + rdev->data_offset,
1526 						 s<<9, conf->tmppage, WRITE)
1527 				    == 0)
1528 					/* Well, this device is dead */
1529 					md_error(mddev, rdev);
1530 			}
1531 		}
1532 		d = start;
1533 		while (d != read_disk) {
1534 			char b[BDEVNAME_SIZE];
1535 			if (d==0)
1536 				d = conf->raid_disks;
1537 			d--;
1538 			rdev = conf->mirrors[d].rdev;
1539 			if (rdev &&
1540 			    test_bit(In_sync, &rdev->flags)) {
1541 				if (sync_page_io(rdev->bdev,
1542 						 sect + rdev->data_offset,
1543 						 s<<9, conf->tmppage, READ)
1544 				    == 0)
1545 					/* Well, this device is dead */
1546 					md_error(mddev, rdev);
1547 				else {
1548 					atomic_add(s, &rdev->corrected_errors);
1549 					printk(KERN_INFO
1550 					       "raid1:%s: read error corrected "
1551 					       "(%d sectors at %llu on %s)\n",
1552 					       mdname(mddev), s,
1553 					       (unsigned long long)(sect +
1554 					           rdev->data_offset),
1555 					       bdevname(rdev->bdev, b));
1556 				}
1557 			}
1558 		}
1559 		sectors -= s;
1560 		sect += s;
1561 	}
1562 }
1563 
1564 static void raid1d(mddev_t *mddev)
1565 {
1566 	r1bio_t *r1_bio;
1567 	struct bio *bio;
1568 	unsigned long flags;
1569 	conf_t *conf = mddev->private;
1570 	struct list_head *head = &conf->retry_list;
1571 	int unplug=0;
1572 	mdk_rdev_t *rdev;
1573 
1574 	md_check_recovery(mddev);
1575 
1576 	for (;;) {
1577 		char b[BDEVNAME_SIZE];
1578 
1579 		unplug += flush_pending_writes(conf);
1580 
1581 		spin_lock_irqsave(&conf->device_lock, flags);
1582 		if (list_empty(head)) {
1583 			spin_unlock_irqrestore(&conf->device_lock, flags);
1584 			break;
1585 		}
1586 		r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1587 		list_del(head->prev);
1588 		conf->nr_queued--;
1589 		spin_unlock_irqrestore(&conf->device_lock, flags);
1590 
1591 		mddev = r1_bio->mddev;
1592 		conf = mddev->private;
1593 		if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1594 			sync_request_write(mddev, r1_bio);
1595 			unplug = 1;
1596 		} else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1597 			/* some requests in the r1bio were BIO_RW_BARRIER
1598 			 * requests which failed with -EOPNOTSUPP.  Hohumm..
1599 			 * Better resubmit without the barrier.
1600 			 * We know which devices to resubmit for, because
1601 			 * all others have had their bios[] entry cleared.
1602 			 * We already have a nr_pending reference on these rdevs.
1603 			 */
1604 			int i;
1605 			const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1606 			clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1607 			clear_bit(R1BIO_Barrier, &r1_bio->state);
1608 			for (i=0; i < conf->raid_disks; i++)
1609 				if (r1_bio->bios[i])
1610 					atomic_inc(&r1_bio->remaining);
1611 			for (i=0; i < conf->raid_disks; i++)
1612 				if (r1_bio->bios[i]) {
1613 					struct bio_vec *bvec;
1614 					int j;
1615 
1616 					bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1617 					/* copy pages from the failed bio, as
1618 					 * this might be a write-behind device */
1619 					__bio_for_each_segment(bvec, bio, j, 0)
1620 						bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1621 					bio_put(r1_bio->bios[i]);
1622 					bio->bi_sector = r1_bio->sector +
1623 						conf->mirrors[i].rdev->data_offset;
1624 					bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1625 					bio->bi_end_io = raid1_end_write_request;
1626 					bio->bi_rw = WRITE | do_sync;
1627 					bio->bi_private = r1_bio;
1628 					r1_bio->bios[i] = bio;
1629 					generic_make_request(bio);
1630 				}
1631 		} else {
1632 			int disk;
1633 
1634 			/* we got a read error. Maybe the drive is bad.  Maybe just
1635 			 * the block and we can fix it.
1636 			 * We freeze all other IO, and try reading the block from
1637 			 * other devices.  When we find one, we re-write
1638 			 * and check it that fixes the read error.
1639 			 * This is all done synchronously while the array is
1640 			 * frozen
1641 			 */
1642 			if (mddev->ro == 0) {
1643 				freeze_array(conf);
1644 				fix_read_error(conf, r1_bio->read_disk,
1645 					       r1_bio->sector,
1646 					       r1_bio->sectors);
1647 				unfreeze_array(conf);
1648 			}
1649 
1650 			bio = r1_bio->bios[r1_bio->read_disk];
1651 			if ((disk=read_balance(conf, r1_bio)) == -1 ||
1652 			    disk == r1_bio->read_disk) {
1653 				printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1654 				       " read error for block %llu\n",
1655 				       bdevname(bio->bi_bdev,b),
1656 				       (unsigned long long)r1_bio->sector);
1657 				raid_end_bio_io(r1_bio);
1658 			} else {
1659 				const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1660 				r1_bio->bios[r1_bio->read_disk] =
1661 					mddev->ro ? IO_BLOCKED : NULL;
1662 				r1_bio->read_disk = disk;
1663 				bio_put(bio);
1664 				bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1665 				r1_bio->bios[r1_bio->read_disk] = bio;
1666 				rdev = conf->mirrors[disk].rdev;
1667 				if (printk_ratelimit())
1668 					printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1669 					       " another mirror\n",
1670 					       bdevname(rdev->bdev,b),
1671 					       (unsigned long long)r1_bio->sector);
1672 				bio->bi_sector = r1_bio->sector + rdev->data_offset;
1673 				bio->bi_bdev = rdev->bdev;
1674 				bio->bi_end_io = raid1_end_read_request;
1675 				bio->bi_rw = READ | do_sync;
1676 				bio->bi_private = r1_bio;
1677 				unplug = 1;
1678 				generic_make_request(bio);
1679 			}
1680 		}
1681 	}
1682 	if (unplug)
1683 		unplug_slaves(mddev);
1684 }
1685 
1686 
1687 static int init_resync(conf_t *conf)
1688 {
1689 	int buffs;
1690 
1691 	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1692 	BUG_ON(conf->r1buf_pool);
1693 	conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1694 					  conf->poolinfo);
1695 	if (!conf->r1buf_pool)
1696 		return -ENOMEM;
1697 	conf->next_resync = 0;
1698 	return 0;
1699 }
1700 
1701 /*
1702  * perform a "sync" on one "block"
1703  *
1704  * We need to make sure that no normal I/O request - particularly write
1705  * requests - conflict with active sync requests.
1706  *
1707  * This is achieved by tracking pending requests and a 'barrier' concept
1708  * that can be installed to exclude normal IO requests.
1709  */
1710 
1711 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1712 {
1713 	conf_t *conf = mddev->private;
1714 	r1bio_t *r1_bio;
1715 	struct bio *bio;
1716 	sector_t max_sector, nr_sectors;
1717 	int disk = -1;
1718 	int i;
1719 	int wonly = -1;
1720 	int write_targets = 0, read_targets = 0;
1721 	int sync_blocks;
1722 	int still_degraded = 0;
1723 
1724 	if (!conf->r1buf_pool)
1725 	{
1726 /*
1727 		printk("sync start - bitmap %p\n", mddev->bitmap);
1728 */
1729 		if (init_resync(conf))
1730 			return 0;
1731 	}
1732 
1733 	max_sector = mddev->dev_sectors;
1734 	if (sector_nr >= max_sector) {
1735 		/* If we aborted, we need to abort the
1736 		 * sync on the 'current' bitmap chunk (there will
1737 		 * only be one in raid1 resync.
1738 		 * We can find the current addess in mddev->curr_resync
1739 		 */
1740 		if (mddev->curr_resync < max_sector) /* aborted */
1741 			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1742 						&sync_blocks, 1);
1743 		else /* completed sync */
1744 			conf->fullsync = 0;
1745 
1746 		bitmap_close_sync(mddev->bitmap);
1747 		close_sync(conf);
1748 		return 0;
1749 	}
1750 
1751 	if (mddev->bitmap == NULL &&
1752 	    mddev->recovery_cp == MaxSector &&
1753 	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1754 	    conf->fullsync == 0) {
1755 		*skipped = 1;
1756 		return max_sector - sector_nr;
1757 	}
1758 	/* before building a request, check if we can skip these blocks..
1759 	 * This call the bitmap_start_sync doesn't actually record anything
1760 	 */
1761 	if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1762 	    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1763 		/* We can skip this block, and probably several more */
1764 		*skipped = 1;
1765 		return sync_blocks;
1766 	}
1767 	/*
1768 	 * If there is non-resync activity waiting for a turn,
1769 	 * and resync is going fast enough,
1770 	 * then let it though before starting on this new sync request.
1771 	 */
1772 	if (!go_faster && conf->nr_waiting)
1773 		msleep_interruptible(1000);
1774 
1775 	bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1776 	raise_barrier(conf);
1777 
1778 	conf->next_resync = sector_nr;
1779 
1780 	r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1781 	rcu_read_lock();
1782 	/*
1783 	 * If we get a correctably read error during resync or recovery,
1784 	 * we might want to read from a different device.  So we
1785 	 * flag all drives that could conceivably be read from for READ,
1786 	 * and any others (which will be non-In_sync devices) for WRITE.
1787 	 * If a read fails, we try reading from something else for which READ
1788 	 * is OK.
1789 	 */
1790 
1791 	r1_bio->mddev = mddev;
1792 	r1_bio->sector = sector_nr;
1793 	r1_bio->state = 0;
1794 	set_bit(R1BIO_IsSync, &r1_bio->state);
1795 
1796 	for (i=0; i < conf->raid_disks; i++) {
1797 		mdk_rdev_t *rdev;
1798 		bio = r1_bio->bios[i];
1799 
1800 		/* take from bio_init */
1801 		bio->bi_next = NULL;
1802 		bio->bi_flags |= 1 << BIO_UPTODATE;
1803 		bio->bi_rw = READ;
1804 		bio->bi_vcnt = 0;
1805 		bio->bi_idx = 0;
1806 		bio->bi_phys_segments = 0;
1807 		bio->bi_size = 0;
1808 		bio->bi_end_io = NULL;
1809 		bio->bi_private = NULL;
1810 
1811 		rdev = rcu_dereference(conf->mirrors[i].rdev);
1812 		if (rdev == NULL ||
1813 			   test_bit(Faulty, &rdev->flags)) {
1814 			still_degraded = 1;
1815 			continue;
1816 		} else if (!test_bit(In_sync, &rdev->flags)) {
1817 			bio->bi_rw = WRITE;
1818 			bio->bi_end_io = end_sync_write;
1819 			write_targets ++;
1820 		} else {
1821 			/* may need to read from here */
1822 			bio->bi_rw = READ;
1823 			bio->bi_end_io = end_sync_read;
1824 			if (test_bit(WriteMostly, &rdev->flags)) {
1825 				if (wonly < 0)
1826 					wonly = i;
1827 			} else {
1828 				if (disk < 0)
1829 					disk = i;
1830 			}
1831 			read_targets++;
1832 		}
1833 		atomic_inc(&rdev->nr_pending);
1834 		bio->bi_sector = sector_nr + rdev->data_offset;
1835 		bio->bi_bdev = rdev->bdev;
1836 		bio->bi_private = r1_bio;
1837 	}
1838 	rcu_read_unlock();
1839 	if (disk < 0)
1840 		disk = wonly;
1841 	r1_bio->read_disk = disk;
1842 
1843 	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1844 		/* extra read targets are also write targets */
1845 		write_targets += read_targets-1;
1846 
1847 	if (write_targets == 0 || read_targets == 0) {
1848 		/* There is nowhere to write, so all non-sync
1849 		 * drives must be failed - so we are finished
1850 		 */
1851 		sector_t rv = max_sector - sector_nr;
1852 		*skipped = 1;
1853 		put_buf(r1_bio);
1854 		return rv;
1855 	}
1856 
1857 	if (max_sector > mddev->resync_max)
1858 		max_sector = mddev->resync_max; /* Don't do IO beyond here */
1859 	nr_sectors = 0;
1860 	sync_blocks = 0;
1861 	do {
1862 		struct page *page;
1863 		int len = PAGE_SIZE;
1864 		if (sector_nr + (len>>9) > max_sector)
1865 			len = (max_sector - sector_nr) << 9;
1866 		if (len == 0)
1867 			break;
1868 		if (sync_blocks == 0) {
1869 			if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1870 					       &sync_blocks, still_degraded) &&
1871 			    !conf->fullsync &&
1872 			    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1873 				break;
1874 			BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1875 			if (len > (sync_blocks<<9))
1876 				len = sync_blocks<<9;
1877 		}
1878 
1879 		for (i=0 ; i < conf->raid_disks; i++) {
1880 			bio = r1_bio->bios[i];
1881 			if (bio->bi_end_io) {
1882 				page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1883 				if (bio_add_page(bio, page, len, 0) == 0) {
1884 					/* stop here */
1885 					bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1886 					while (i > 0) {
1887 						i--;
1888 						bio = r1_bio->bios[i];
1889 						if (bio->bi_end_io==NULL)
1890 							continue;
1891 						/* remove last page from this bio */
1892 						bio->bi_vcnt--;
1893 						bio->bi_size -= len;
1894 						bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1895 					}
1896 					goto bio_full;
1897 				}
1898 			}
1899 		}
1900 		nr_sectors += len>>9;
1901 		sector_nr += len>>9;
1902 		sync_blocks -= (len>>9);
1903 	} while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1904  bio_full:
1905 	r1_bio->sectors = nr_sectors;
1906 
1907 	/* For a user-requested sync, we read all readable devices and do a
1908 	 * compare
1909 	 */
1910 	if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1911 		atomic_set(&r1_bio->remaining, read_targets);
1912 		for (i=0; i<conf->raid_disks; i++) {
1913 			bio = r1_bio->bios[i];
1914 			if (bio->bi_end_io == end_sync_read) {
1915 				md_sync_acct(bio->bi_bdev, nr_sectors);
1916 				generic_make_request(bio);
1917 			}
1918 		}
1919 	} else {
1920 		atomic_set(&r1_bio->remaining, 1);
1921 		bio = r1_bio->bios[r1_bio->read_disk];
1922 		md_sync_acct(bio->bi_bdev, nr_sectors);
1923 		generic_make_request(bio);
1924 
1925 	}
1926 	return nr_sectors;
1927 }
1928 
1929 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1930 {
1931 	if (sectors)
1932 		return sectors;
1933 
1934 	return mddev->dev_sectors;
1935 }
1936 
1937 static int run(mddev_t *mddev)
1938 {
1939 	conf_t *conf;
1940 	int i, j, disk_idx;
1941 	mirror_info_t *disk;
1942 	mdk_rdev_t *rdev;
1943 
1944 	if (mddev->level != 1) {
1945 		printk("raid1: %s: raid level not set to mirroring (%d)\n",
1946 		       mdname(mddev), mddev->level);
1947 		goto out;
1948 	}
1949 	if (mddev->reshape_position != MaxSector) {
1950 		printk("raid1: %s: reshape_position set but not supported\n",
1951 		       mdname(mddev));
1952 		goto out;
1953 	}
1954 	/*
1955 	 * copy the already verified devices into our private RAID1
1956 	 * bookkeeping area. [whatever we allocate in run(),
1957 	 * should be freed in stop()]
1958 	 */
1959 	conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1960 	mddev->private = conf;
1961 	if (!conf)
1962 		goto out_no_mem;
1963 
1964 	conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1965 				 GFP_KERNEL);
1966 	if (!conf->mirrors)
1967 		goto out_no_mem;
1968 
1969 	conf->tmppage = alloc_page(GFP_KERNEL);
1970 	if (!conf->tmppage)
1971 		goto out_no_mem;
1972 
1973 	conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1974 	if (!conf->poolinfo)
1975 		goto out_no_mem;
1976 	conf->poolinfo->mddev = mddev;
1977 	conf->poolinfo->raid_disks = mddev->raid_disks;
1978 	conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1979 					  r1bio_pool_free,
1980 					  conf->poolinfo);
1981 	if (!conf->r1bio_pool)
1982 		goto out_no_mem;
1983 
1984 	spin_lock_init(&conf->device_lock);
1985 	mddev->queue->queue_lock = &conf->device_lock;
1986 
1987 	list_for_each_entry(rdev, &mddev->disks, same_set) {
1988 		disk_idx = rdev->raid_disk;
1989 		if (disk_idx >= mddev->raid_disks
1990 		    || disk_idx < 0)
1991 			continue;
1992 		disk = conf->mirrors + disk_idx;
1993 
1994 		disk->rdev = rdev;
1995 		disk_stack_limits(mddev->gendisk, rdev->bdev,
1996 				  rdev->data_offset << 9);
1997 		/* as we don't honour merge_bvec_fn, we must never risk
1998 		 * violating it, so limit ->max_sector to one PAGE, as
1999 		 * a one page request is never in violation.
2000 		 */
2001 		if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2002 		    queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
2003 			blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
2004 
2005 		disk->head_position = 0;
2006 	}
2007 	conf->raid_disks = mddev->raid_disks;
2008 	conf->mddev = mddev;
2009 	INIT_LIST_HEAD(&conf->retry_list);
2010 
2011 	spin_lock_init(&conf->resync_lock);
2012 	init_waitqueue_head(&conf->wait_barrier);
2013 
2014 	bio_list_init(&conf->pending_bio_list);
2015 	bio_list_init(&conf->flushing_bio_list);
2016 
2017 
2018 	mddev->degraded = 0;
2019 	for (i = 0; i < conf->raid_disks; i++) {
2020 
2021 		disk = conf->mirrors + i;
2022 
2023 		if (!disk->rdev ||
2024 		    !test_bit(In_sync, &disk->rdev->flags)) {
2025 			disk->head_position = 0;
2026 			mddev->degraded++;
2027 			if (disk->rdev)
2028 				conf->fullsync = 1;
2029 		}
2030 	}
2031 	if (mddev->degraded == conf->raid_disks) {
2032 		printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2033 			mdname(mddev));
2034 		goto out_free_conf;
2035 	}
2036 	if (conf->raid_disks - mddev->degraded == 1)
2037 		mddev->recovery_cp = MaxSector;
2038 
2039 	/*
2040 	 * find the first working one and use it as a starting point
2041 	 * to read balancing.
2042 	 */
2043 	for (j = 0; j < conf->raid_disks &&
2044 		     (!conf->mirrors[j].rdev ||
2045 		      !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
2046 		/* nothing */;
2047 	conf->last_used = j;
2048 
2049 
2050 	mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
2051 	if (!mddev->thread) {
2052 		printk(KERN_ERR
2053 		       "raid1: couldn't allocate thread for %s\n",
2054 		       mdname(mddev));
2055 		goto out_free_conf;
2056 	}
2057 
2058 	if (mddev->recovery_cp != MaxSector)
2059 		printk(KERN_NOTICE "raid1: %s is not clean"
2060 		       " -- starting background reconstruction\n",
2061 		       mdname(mddev));
2062 	printk(KERN_INFO
2063 		"raid1: raid set %s active with %d out of %d mirrors\n",
2064 		mdname(mddev), mddev->raid_disks - mddev->degraded,
2065 		mddev->raid_disks);
2066 	/*
2067 	 * Ok, everything is just fine now
2068 	 */
2069 	md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2070 
2071 	mddev->queue->unplug_fn = raid1_unplug;
2072 	mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2073 	mddev->queue->backing_dev_info.congested_data = mddev;
2074 	md_integrity_register(mddev);
2075 	return 0;
2076 
2077 out_no_mem:
2078 	printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2079 	       mdname(mddev));
2080 
2081 out_free_conf:
2082 	if (conf) {
2083 		if (conf->r1bio_pool)
2084 			mempool_destroy(conf->r1bio_pool);
2085 		kfree(conf->mirrors);
2086 		safe_put_page(conf->tmppage);
2087 		kfree(conf->poolinfo);
2088 		kfree(conf);
2089 		mddev->private = NULL;
2090 	}
2091 out:
2092 	return -EIO;
2093 }
2094 
2095 static int stop(mddev_t *mddev)
2096 {
2097 	conf_t *conf = mddev->private;
2098 	struct bitmap *bitmap = mddev->bitmap;
2099 	int behind_wait = 0;
2100 
2101 	/* wait for behind writes to complete */
2102 	while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2103 		behind_wait++;
2104 		printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2105 		set_current_state(TASK_UNINTERRUPTIBLE);
2106 		schedule_timeout(HZ); /* wait a second */
2107 		/* need to kick something here to make sure I/O goes? */
2108 	}
2109 
2110 	raise_barrier(conf);
2111 	lower_barrier(conf);
2112 
2113 	md_unregister_thread(mddev->thread);
2114 	mddev->thread = NULL;
2115 	blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2116 	if (conf->r1bio_pool)
2117 		mempool_destroy(conf->r1bio_pool);
2118 	kfree(conf->mirrors);
2119 	kfree(conf->poolinfo);
2120 	kfree(conf);
2121 	mddev->private = NULL;
2122 	return 0;
2123 }
2124 
2125 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2126 {
2127 	/* no resync is happening, and there is enough space
2128 	 * on all devices, so we can resize.
2129 	 * We need to make sure resync covers any new space.
2130 	 * If the array is shrinking we should possibly wait until
2131 	 * any io in the removed space completes, but it hardly seems
2132 	 * worth it.
2133 	 */
2134 	md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2135 	if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2136 		return -EINVAL;
2137 	set_capacity(mddev->gendisk, mddev->array_sectors);
2138 	mddev->changed = 1;
2139 	revalidate_disk(mddev->gendisk);
2140 	if (sectors > mddev->dev_sectors &&
2141 	    mddev->recovery_cp == MaxSector) {
2142 		mddev->recovery_cp = mddev->dev_sectors;
2143 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2144 	}
2145 	mddev->dev_sectors = sectors;
2146 	mddev->resync_max_sectors = sectors;
2147 	return 0;
2148 }
2149 
2150 static int raid1_reshape(mddev_t *mddev)
2151 {
2152 	/* We need to:
2153 	 * 1/ resize the r1bio_pool
2154 	 * 2/ resize conf->mirrors
2155 	 *
2156 	 * We allocate a new r1bio_pool if we can.
2157 	 * Then raise a device barrier and wait until all IO stops.
2158 	 * Then resize conf->mirrors and swap in the new r1bio pool.
2159 	 *
2160 	 * At the same time, we "pack" the devices so that all the missing
2161 	 * devices have the higher raid_disk numbers.
2162 	 */
2163 	mempool_t *newpool, *oldpool;
2164 	struct pool_info *newpoolinfo;
2165 	mirror_info_t *newmirrors;
2166 	conf_t *conf = mddev->private;
2167 	int cnt, raid_disks;
2168 	unsigned long flags;
2169 	int d, d2, err;
2170 
2171 	/* Cannot change chunk_size, layout, or level */
2172 	if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2173 	    mddev->layout != mddev->new_layout ||
2174 	    mddev->level != mddev->new_level) {
2175 		mddev->new_chunk_sectors = mddev->chunk_sectors;
2176 		mddev->new_layout = mddev->layout;
2177 		mddev->new_level = mddev->level;
2178 		return -EINVAL;
2179 	}
2180 
2181 	err = md_allow_write(mddev);
2182 	if (err)
2183 		return err;
2184 
2185 	raid_disks = mddev->raid_disks + mddev->delta_disks;
2186 
2187 	if (raid_disks < conf->raid_disks) {
2188 		cnt=0;
2189 		for (d= 0; d < conf->raid_disks; d++)
2190 			if (conf->mirrors[d].rdev)
2191 				cnt++;
2192 		if (cnt > raid_disks)
2193 			return -EBUSY;
2194 	}
2195 
2196 	newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2197 	if (!newpoolinfo)
2198 		return -ENOMEM;
2199 	newpoolinfo->mddev = mddev;
2200 	newpoolinfo->raid_disks = raid_disks;
2201 
2202 	newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2203 				 r1bio_pool_free, newpoolinfo);
2204 	if (!newpool) {
2205 		kfree(newpoolinfo);
2206 		return -ENOMEM;
2207 	}
2208 	newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2209 	if (!newmirrors) {
2210 		kfree(newpoolinfo);
2211 		mempool_destroy(newpool);
2212 		return -ENOMEM;
2213 	}
2214 
2215 	raise_barrier(conf);
2216 
2217 	/* ok, everything is stopped */
2218 	oldpool = conf->r1bio_pool;
2219 	conf->r1bio_pool = newpool;
2220 
2221 	for (d = d2 = 0; d < conf->raid_disks; d++) {
2222 		mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2223 		if (rdev && rdev->raid_disk != d2) {
2224 			char nm[20];
2225 			sprintf(nm, "rd%d", rdev->raid_disk);
2226 			sysfs_remove_link(&mddev->kobj, nm);
2227 			rdev->raid_disk = d2;
2228 			sprintf(nm, "rd%d", rdev->raid_disk);
2229 			sysfs_remove_link(&mddev->kobj, nm);
2230 			if (sysfs_create_link(&mddev->kobj,
2231 					      &rdev->kobj, nm))
2232 				printk(KERN_WARNING
2233 				       "md/raid1: cannot register "
2234 				       "%s for %s\n",
2235 				       nm, mdname(mddev));
2236 		}
2237 		if (rdev)
2238 			newmirrors[d2++].rdev = rdev;
2239 	}
2240 	kfree(conf->mirrors);
2241 	conf->mirrors = newmirrors;
2242 	kfree(conf->poolinfo);
2243 	conf->poolinfo = newpoolinfo;
2244 
2245 	spin_lock_irqsave(&conf->device_lock, flags);
2246 	mddev->degraded += (raid_disks - conf->raid_disks);
2247 	spin_unlock_irqrestore(&conf->device_lock, flags);
2248 	conf->raid_disks = mddev->raid_disks = raid_disks;
2249 	mddev->delta_disks = 0;
2250 
2251 	conf->last_used = 0; /* just make sure it is in-range */
2252 	lower_barrier(conf);
2253 
2254 	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2255 	md_wakeup_thread(mddev->thread);
2256 
2257 	mempool_destroy(oldpool);
2258 	return 0;
2259 }
2260 
2261 static void raid1_quiesce(mddev_t *mddev, int state)
2262 {
2263 	conf_t *conf = mddev->private;
2264 
2265 	switch(state) {
2266 	case 1:
2267 		raise_barrier(conf);
2268 		break;
2269 	case 0:
2270 		lower_barrier(conf);
2271 		break;
2272 	}
2273 }
2274 
2275 
2276 static struct mdk_personality raid1_personality =
2277 {
2278 	.name		= "raid1",
2279 	.level		= 1,
2280 	.owner		= THIS_MODULE,
2281 	.make_request	= make_request,
2282 	.run		= run,
2283 	.stop		= stop,
2284 	.status		= status,
2285 	.error_handler	= error,
2286 	.hot_add_disk	= raid1_add_disk,
2287 	.hot_remove_disk= raid1_remove_disk,
2288 	.spare_active	= raid1_spare_active,
2289 	.sync_request	= sync_request,
2290 	.resize		= raid1_resize,
2291 	.size		= raid1_size,
2292 	.check_reshape	= raid1_reshape,
2293 	.quiesce	= raid1_quiesce,
2294 };
2295 
2296 static int __init raid_init(void)
2297 {
2298 	return register_md_personality(&raid1_personality);
2299 }
2300 
2301 static void raid_exit(void)
2302 {
2303 	unregister_md_personality(&raid1_personality);
2304 }
2305 
2306 module_init(raid_init);
2307 module_exit(raid_exit);
2308 MODULE_LICENSE("GPL");
2309 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2310 MODULE_ALIAS("md-raid1");
2311 MODULE_ALIAS("md-level-1");
2312