blob: caaca9e178bc2bdd15e960596c569d79aa86212a [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
NeilBrown16a53ec2006-06-26 00:27:38 -07005 * Copyright (C) 2002, 2003 H. Peter Anvin
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
NeilBrown16a53ec2006-06-26 00:27:38 -07007 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
Linus Torvalds1da177e2005-04-16 15:20:36 -070010 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
NeilBrownae3c20c2006-07-10 04:44:17 -070021/*
22 * BITMAP UNPLUGGING:
23 *
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
26 * explanation.
27 *
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->bm_write is the number of the last batch successfully written.
31 * conf->bm_flush is the number of the last batch that was closed to
32 * new additions.
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is bm_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
39 * batch.
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
43 * miss any bits.
44 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Linus Torvalds1da177e2005-04-16 15:20:36 -070046#include <linux/module.h>
47#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070048#include <linux/highmem.h>
49#include <linux/bitops.h>
NeilBrownf6705572006-03-27 01:18:11 -080050#include <linux/kthread.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070051#include <asm/atomic.h>
NeilBrown16a53ec2006-06-26 00:27:38 -070052#include "raid6.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070053
NeilBrown72626682005-09-09 16:23:54 -070054#include <linux/raid/bitmap.h>
Dan Williams91c00922007-01-02 13:52:30 -070055#include <linux/async_tx.h>
NeilBrown72626682005-09-09 16:23:54 -070056
Linus Torvalds1da177e2005-04-16 15:20:36 -070057/*
58 * Stripe cache
59 */
60
61#define NR_STRIPES 256
62#define STRIPE_SIZE PAGE_SIZE
63#define STRIPE_SHIFT (PAGE_SHIFT - 9)
64#define STRIPE_SECTORS (STRIPE_SIZE>>9)
65#define IO_THRESHOLD 1
NeilBrownfccddba2006-01-06 00:20:33 -080066#define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
Linus Torvalds1da177e2005-04-16 15:20:36 -070067#define HASH_MASK (NR_HASH - 1)
68
NeilBrownfccddba2006-01-06 00:20:33 -080069#define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
Linus Torvalds1da177e2005-04-16 15:20:36 -070070
71/* bio's attached to a stripe+device for I/O are linked together in bi_sector
72 * order without overlap. There may be several bio's per stripe+device, and
73 * a bio could span several devices.
74 * When walking this list for a particular stripe+device, we must never proceed
75 * beyond a bio that extends past this device, as the next bio might no longer
76 * be valid.
77 * This macro is used to determine the 'next' bio in the list, given the sector
78 * of the current stripe+device
79 */
80#define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
81/*
82 * The following can be used to debug the driver
83 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070084#define RAID5_PARANOIA 1
85#if RAID5_PARANOIA && defined(CONFIG_SMP)
86# define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87#else
88# define CHECK_DEVLOCK()
89#endif
90
Dan Williams45b42332007-07-09 11:56:43 -070091#ifdef DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#define inline
93#define __inline__
94#endif
95
NeilBrown16a53ec2006-06-26 00:27:38 -070096#if !RAID6_USE_EMPTY_ZERO_PAGE
97/* In .bss so it's zeroed */
98const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
99#endif
100
101static inline int raid6_next_disk(int disk, int raid_disks)
102{
103 disk++;
104 return (disk < raid_disks) ? disk : 0;
105}
Dan Williamsa4456852007-07-09 11:56:43 -0700106
107static void return_io(struct bio *return_bi)
108{
109 struct bio *bi = return_bi;
110 while (bi) {
Dan Williamsa4456852007-07-09 11:56:43 -0700111
112 return_bi = bi->bi_next;
113 bi->bi_next = NULL;
114 bi->bi_size = 0;
NeilBrown6712ecf2007-09-27 12:47:43 +0200115 bi->bi_end_io(bi,
Dan Williamsa4456852007-07-09 11:56:43 -0700116 test_bit(BIO_UPTODATE, &bi->bi_flags)
117 ? 0 : -EIO);
118 bi = return_bi;
119 }
120}
121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122static void print_raid5_conf (raid5_conf_t *conf);
123
Arjan van de Ven858119e2006-01-14 13:20:43 -0800124static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125{
126 if (atomic_dec_and_test(&sh->count)) {
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +0200127 BUG_ON(!list_empty(&sh->lru));
128 BUG_ON(atomic_read(&conf->active_stripes)==0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 if (test_bit(STRIPE_HANDLE, &sh->state)) {
NeilBrown7c785b72006-07-10 04:44:16 -0700130 if (test_bit(STRIPE_DELAYED, &sh->state)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 list_add_tail(&sh->lru, &conf->delayed_list);
NeilBrown7c785b72006-07-10 04:44:16 -0700132 blk_plug_device(conf->mddev->queue);
133 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
NeilBrownae3c20c2006-07-10 04:44:17 -0700134 sh->bm_seq - conf->seq_write > 0) {
NeilBrown72626682005-09-09 16:23:54 -0700135 list_add_tail(&sh->lru, &conf->bitmap_list);
NeilBrown7c785b72006-07-10 04:44:16 -0700136 blk_plug_device(conf->mddev->queue);
137 } else {
NeilBrown72626682005-09-09 16:23:54 -0700138 clear_bit(STRIPE_BIT_DELAY, &sh->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 list_add_tail(&sh->lru, &conf->handle_list);
NeilBrown72626682005-09-09 16:23:54 -0700140 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 md_wakeup_thread(conf->mddev->thread);
142 } else {
Dan Williamsd84e0f12007-01-02 13:52:30 -0700143 BUG_ON(sh->ops.pending);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
145 atomic_dec(&conf->preread_active_stripes);
146 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
147 md_wakeup_thread(conf->mddev->thread);
148 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 atomic_dec(&conf->active_stripes);
NeilBrownccfcc3c2006-03-27 01:18:09 -0800150 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
151 list_add_tail(&sh->lru, &conf->inactive_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 wake_up(&conf->wait_for_stripe);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -0800153 if (conf->retry_read_aligned)
154 md_wakeup_thread(conf->mddev->thread);
NeilBrownccfcc3c2006-03-27 01:18:09 -0800155 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 }
157 }
158}
159static void release_stripe(struct stripe_head *sh)
160{
161 raid5_conf_t *conf = sh->raid_conf;
162 unsigned long flags;
NeilBrown16a53ec2006-06-26 00:27:38 -0700163
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 spin_lock_irqsave(&conf->device_lock, flags);
165 __release_stripe(conf, sh);
166 spin_unlock_irqrestore(&conf->device_lock, flags);
167}
168
NeilBrownfccddba2006-01-06 00:20:33 -0800169static inline void remove_hash(struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170{
Dan Williams45b42332007-07-09 11:56:43 -0700171 pr_debug("remove_hash(), stripe %llu\n",
172 (unsigned long long)sh->sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173
NeilBrownfccddba2006-01-06 00:20:33 -0800174 hlist_del_init(&sh->hash);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175}
176
NeilBrown16a53ec2006-06-26 00:27:38 -0700177static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178{
NeilBrownfccddba2006-01-06 00:20:33 -0800179 struct hlist_head *hp = stripe_hash(conf, sh->sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180
Dan Williams45b42332007-07-09 11:56:43 -0700181 pr_debug("insert_hash(), stripe %llu\n",
182 (unsigned long long)sh->sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183
184 CHECK_DEVLOCK();
NeilBrownfccddba2006-01-06 00:20:33 -0800185 hlist_add_head(&sh->hash, hp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186}
187
188
189/* find an idle stripe, make sure it is unhashed, and return it. */
190static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
191{
192 struct stripe_head *sh = NULL;
193 struct list_head *first;
194
195 CHECK_DEVLOCK();
196 if (list_empty(&conf->inactive_list))
197 goto out;
198 first = conf->inactive_list.next;
199 sh = list_entry(first, struct stripe_head, lru);
200 list_del_init(first);
201 remove_hash(sh);
202 atomic_inc(&conf->active_stripes);
203out:
204 return sh;
205}
206
207static void shrink_buffers(struct stripe_head *sh, int num)
208{
209 struct page *p;
210 int i;
211
212 for (i=0; i<num ; i++) {
213 p = sh->dev[i].page;
214 if (!p)
215 continue;
216 sh->dev[i].page = NULL;
NeilBrown2d1f3b52006-01-06 00:20:31 -0800217 put_page(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218 }
219}
220
221static int grow_buffers(struct stripe_head *sh, int num)
222{
223 int i;
224
225 for (i=0; i<num; i++) {
226 struct page *page;
227
228 if (!(page = alloc_page(GFP_KERNEL))) {
229 return 1;
230 }
231 sh->dev[i].page = page;
232 }
233 return 0;
234}
235
236static void raid5_build_block (struct stripe_head *sh, int i);
237
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800238static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239{
240 raid5_conf_t *conf = sh->raid_conf;
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800241 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +0200243 BUG_ON(atomic_read(&sh->count) != 0);
244 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
Dan Williamsd84e0f12007-01-02 13:52:30 -0700245 BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
246
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 CHECK_DEVLOCK();
Dan Williams45b42332007-07-09 11:56:43 -0700248 pr_debug("init_stripe called, stripe %llu\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 (unsigned long long)sh->sector);
250
251 remove_hash(sh);
NeilBrown16a53ec2006-06-26 00:27:38 -0700252
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 sh->sector = sector;
254 sh->pd_idx = pd_idx;
255 sh->state = 0;
256
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800257 sh->disks = disks;
258
259 for (i = sh->disks; i--; ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 struct r5dev *dev = &sh->dev[i];
261
Dan Williamsd84e0f12007-01-02 13:52:30 -0700262 if (dev->toread || dev->read || dev->towrite || dev->written ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 test_bit(R5_LOCKED, &dev->flags)) {
Dan Williamsd84e0f12007-01-02 13:52:30 -0700264 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 (unsigned long long)sh->sector, i, dev->toread,
Dan Williamsd84e0f12007-01-02 13:52:30 -0700266 dev->read, dev->towrite, dev->written,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 test_bit(R5_LOCKED, &dev->flags));
268 BUG();
269 }
270 dev->flags = 0;
271 raid5_build_block(sh, i);
272 }
273 insert_hash(conf, sh);
274}
275
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800276static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277{
278 struct stripe_head *sh;
NeilBrownfccddba2006-01-06 00:20:33 -0800279 struct hlist_node *hn;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280
281 CHECK_DEVLOCK();
Dan Williams45b42332007-07-09 11:56:43 -0700282 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
NeilBrownfccddba2006-01-06 00:20:33 -0800283 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800284 if (sh->sector == sector && sh->disks == disks)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285 return sh;
Dan Williams45b42332007-07-09 11:56:43 -0700286 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 return NULL;
288}
289
290static void unplug_slaves(mddev_t *mddev);
Jens Axboe165125e2007-07-24 09:28:11 +0200291static void raid5_unplug_device(struct request_queue *q);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800293static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
294 int pd_idx, int noblock)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295{
296 struct stripe_head *sh;
297
Dan Williams45b42332007-07-09 11:56:43 -0700298 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299
300 spin_lock_irq(&conf->device_lock);
301
302 do {
NeilBrown72626682005-09-09 16:23:54 -0700303 wait_event_lock_irq(conf->wait_for_stripe,
304 conf->quiesce == 0,
305 conf->device_lock, /* nothing */);
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800306 sh = __find_stripe(conf, sector, disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 if (!sh) {
308 if (!conf->inactive_blocked)
309 sh = get_free_stripe(conf);
310 if (noblock && sh == NULL)
311 break;
312 if (!sh) {
313 conf->inactive_blocked = 1;
314 wait_event_lock_irq(conf->wait_for_stripe,
315 !list_empty(&conf->inactive_list) &&
NeilBrown50368052005-12-12 02:39:17 -0800316 (atomic_read(&conf->active_stripes)
317 < (conf->max_nr_stripes *3/4)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 || !conf->inactive_blocked),
319 conf->device_lock,
NeilBrownf4370782006-07-10 04:44:14 -0700320 raid5_unplug_device(conf->mddev->queue)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 );
322 conf->inactive_blocked = 0;
323 } else
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800324 init_stripe(sh, sector, pd_idx, disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 } else {
326 if (atomic_read(&sh->count)) {
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +0200327 BUG_ON(!list_empty(&sh->lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 } else {
329 if (!test_bit(STRIPE_HANDLE, &sh->state))
330 atomic_inc(&conf->active_stripes);
NeilBrownff4e8d92006-07-10 04:44:16 -0700331 if (list_empty(&sh->lru) &&
332 !test_bit(STRIPE_EXPANDING, &sh->state))
NeilBrown16a53ec2006-06-26 00:27:38 -0700333 BUG();
334 list_del_init(&sh->lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 }
336 }
337 } while (sh == NULL);
338
339 if (sh)
340 atomic_inc(&sh->count);
341
342 spin_unlock_irq(&conf->device_lock);
343 return sh;
344}
345
Dan Williamsd84e0f12007-01-02 13:52:30 -0700346/* test_and_ack_op() ensures that we only dequeue an operation once */
347#define test_and_ack_op(op, pend) \
348do { \
349 if (test_bit(op, &sh->ops.pending) && \
350 !test_bit(op, &sh->ops.complete)) { \
351 if (test_and_set_bit(op, &sh->ops.ack)) \
352 clear_bit(op, &pend); \
353 else \
354 ack++; \
355 } else \
356 clear_bit(op, &pend); \
357} while (0)
358
359/* find new work to run, do not resubmit work that is already
360 * in flight
361 */
362static unsigned long get_stripe_work(struct stripe_head *sh)
363{
364 unsigned long pending;
365 int ack = 0;
366
367 pending = sh->ops.pending;
368
369 test_and_ack_op(STRIPE_OP_BIOFILL, pending);
370 test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
371 test_and_ack_op(STRIPE_OP_PREXOR, pending);
372 test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
373 test_and_ack_op(STRIPE_OP_POSTXOR, pending);
374 test_and_ack_op(STRIPE_OP_CHECK, pending);
375 if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
376 ack++;
377
378 sh->ops.count -= ack;
379 BUG_ON(sh->ops.count < 0);
380
381 return pending;
382}
383
NeilBrown6712ecf2007-09-27 12:47:43 +0200384static void
385raid5_end_read_request(struct bio *bi, int error);
386static void
387raid5_end_write_request(struct bio *bi, int error);
Dan Williams91c00922007-01-02 13:52:30 -0700388
389static void ops_run_io(struct stripe_head *sh)
390{
391 raid5_conf_t *conf = sh->raid_conf;
392 int i, disks = sh->disks;
393
394 might_sleep();
395
396 for (i = disks; i--; ) {
397 int rw;
398 struct bio *bi;
399 mdk_rdev_t *rdev;
400 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
401 rw = WRITE;
402 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
403 rw = READ;
404 else
405 continue;
406
407 bi = &sh->dev[i].req;
408
409 bi->bi_rw = rw;
410 if (rw == WRITE)
411 bi->bi_end_io = raid5_end_write_request;
412 else
413 bi->bi_end_io = raid5_end_read_request;
414
415 rcu_read_lock();
416 rdev = rcu_dereference(conf->disks[i].rdev);
417 if (rdev && test_bit(Faulty, &rdev->flags))
418 rdev = NULL;
419 if (rdev)
420 atomic_inc(&rdev->nr_pending);
421 rcu_read_unlock();
422
423 if (rdev) {
424 if (test_bit(STRIPE_SYNCING, &sh->state) ||
425 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
426 test_bit(STRIPE_EXPAND_READY, &sh->state))
427 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
428
429 bi->bi_bdev = rdev->bdev;
430 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
431 __FUNCTION__, (unsigned long long)sh->sector,
432 bi->bi_rw, i);
433 atomic_inc(&sh->count);
434 bi->bi_sector = sh->sector + rdev->data_offset;
435 bi->bi_flags = 1 << BIO_UPTODATE;
436 bi->bi_vcnt = 1;
437 bi->bi_max_vecs = 1;
438 bi->bi_idx = 0;
439 bi->bi_io_vec = &sh->dev[i].vec;
440 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
441 bi->bi_io_vec[0].bv_offset = 0;
442 bi->bi_size = STRIPE_SIZE;
443 bi->bi_next = NULL;
444 if (rw == WRITE &&
445 test_bit(R5_ReWrite, &sh->dev[i].flags))
446 atomic_add(STRIPE_SECTORS,
447 &rdev->corrected_errors);
448 generic_make_request(bi);
449 } else {
450 if (rw == WRITE)
451 set_bit(STRIPE_DEGRADED, &sh->state);
452 pr_debug("skip op %ld on disc %d for sector %llu\n",
453 bi->bi_rw, i, (unsigned long long)sh->sector);
454 clear_bit(R5_LOCKED, &sh->dev[i].flags);
455 set_bit(STRIPE_HANDLE, &sh->state);
456 }
457 }
458}
459
460static struct dma_async_tx_descriptor *
461async_copy_data(int frombio, struct bio *bio, struct page *page,
462 sector_t sector, struct dma_async_tx_descriptor *tx)
463{
464 struct bio_vec *bvl;
465 struct page *bio_page;
466 int i;
467 int page_offset;
468
469 if (bio->bi_sector >= sector)
470 page_offset = (signed)(bio->bi_sector - sector) * 512;
471 else
472 page_offset = (signed)(sector - bio->bi_sector) * -512;
473 bio_for_each_segment(bvl, bio, i) {
474 int len = bio_iovec_idx(bio, i)->bv_len;
475 int clen;
476 int b_offset = 0;
477
478 if (page_offset < 0) {
479 b_offset = -page_offset;
480 page_offset += b_offset;
481 len -= b_offset;
482 }
483
484 if (len > 0 && page_offset + len > STRIPE_SIZE)
485 clen = STRIPE_SIZE - page_offset;
486 else
487 clen = len;
488
489 if (clen > 0) {
490 b_offset += bio_iovec_idx(bio, i)->bv_offset;
491 bio_page = bio_iovec_idx(bio, i)->bv_page;
492 if (frombio)
493 tx = async_memcpy(page, bio_page, page_offset,
494 b_offset, clen,
Dan Williamseb0645a2007-07-20 00:31:46 -0700495 ASYNC_TX_DEP_ACK,
Dan Williams91c00922007-01-02 13:52:30 -0700496 tx, NULL, NULL);
497 else
498 tx = async_memcpy(bio_page, page, b_offset,
499 page_offset, clen,
Dan Williamseb0645a2007-07-20 00:31:46 -0700500 ASYNC_TX_DEP_ACK,
Dan Williams91c00922007-01-02 13:52:30 -0700501 tx, NULL, NULL);
502 }
503 if (clen < len) /* hit end of page */
504 break;
505 page_offset += len;
506 }
507
508 return tx;
509}
510
511static void ops_complete_biofill(void *stripe_head_ref)
512{
513 struct stripe_head *sh = stripe_head_ref;
514 struct bio *return_bi = NULL;
515 raid5_conf_t *conf = sh->raid_conf;
Dan Williamse4d84902007-09-24 10:06:13 -0700516 int i;
Dan Williams91c00922007-01-02 13:52:30 -0700517
518 pr_debug("%s: stripe %llu\n", __FUNCTION__,
519 (unsigned long long)sh->sector);
520
521 /* clear completed biofills */
522 for (i = sh->disks; i--; ) {
523 struct r5dev *dev = &sh->dev[i];
Dan Williams91c00922007-01-02 13:52:30 -0700524
525 /* acknowledge completion of a biofill operation */
Dan Williamse4d84902007-09-24 10:06:13 -0700526 /* and check if we need to reply to a read request,
527 * new R5_Wantfill requests are held off until
528 * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
529 */
530 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
Dan Williams91c00922007-01-02 13:52:30 -0700531 struct bio *rbi, *rbi2;
Dan Williams91c00922007-01-02 13:52:30 -0700532
533 /* The access to dev->read is outside of the
534 * spin_lock_irq(&conf->device_lock), but is protected
535 * by the STRIPE_OP_BIOFILL pending bit
536 */
537 BUG_ON(!dev->read);
538 rbi = dev->read;
539 dev->read = NULL;
540 while (rbi && rbi->bi_sector <
541 dev->sector + STRIPE_SECTORS) {
542 rbi2 = r5_next_bio(rbi, dev->sector);
543 spin_lock_irq(&conf->device_lock);
544 if (--rbi->bi_phys_segments == 0) {
545 rbi->bi_next = return_bi;
546 return_bi = rbi;
547 }
548 spin_unlock_irq(&conf->device_lock);
549 rbi = rbi2;
550 }
551 }
552 }
553 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
554 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
555
556 return_io(return_bi);
557
Dan Williamse4d84902007-09-24 10:06:13 -0700558 set_bit(STRIPE_HANDLE, &sh->state);
Dan Williams91c00922007-01-02 13:52:30 -0700559 release_stripe(sh);
560}
561
562static void ops_run_biofill(struct stripe_head *sh)
563{
564 struct dma_async_tx_descriptor *tx = NULL;
565 raid5_conf_t *conf = sh->raid_conf;
566 int i;
567
568 pr_debug("%s: stripe %llu\n", __FUNCTION__,
569 (unsigned long long)sh->sector);
570
571 for (i = sh->disks; i--; ) {
572 struct r5dev *dev = &sh->dev[i];
573 if (test_bit(R5_Wantfill, &dev->flags)) {
574 struct bio *rbi;
575 spin_lock_irq(&conf->device_lock);
576 dev->read = rbi = dev->toread;
577 dev->toread = NULL;
578 spin_unlock_irq(&conf->device_lock);
579 while (rbi && rbi->bi_sector <
580 dev->sector + STRIPE_SECTORS) {
581 tx = async_copy_data(0, rbi, dev->page,
582 dev->sector, tx);
583 rbi = r5_next_bio(rbi, dev->sector);
584 }
585 }
586 }
587
588 atomic_inc(&sh->count);
589 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
590 ops_complete_biofill, sh);
591}
592
593static void ops_complete_compute5(void *stripe_head_ref)
594{
595 struct stripe_head *sh = stripe_head_ref;
596 int target = sh->ops.target;
597 struct r5dev *tgt = &sh->dev[target];
598
599 pr_debug("%s: stripe %llu\n", __FUNCTION__,
600 (unsigned long long)sh->sector);
601
602 set_bit(R5_UPTODATE, &tgt->flags);
603 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
604 clear_bit(R5_Wantcompute, &tgt->flags);
605 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
606 set_bit(STRIPE_HANDLE, &sh->state);
607 release_stripe(sh);
608}
609
610static struct dma_async_tx_descriptor *
611ops_run_compute5(struct stripe_head *sh, unsigned long pending)
612{
613 /* kernel stack size limits the total number of disks */
614 int disks = sh->disks;
615 struct page *xor_srcs[disks];
616 int target = sh->ops.target;
617 struct r5dev *tgt = &sh->dev[target];
618 struct page *xor_dest = tgt->page;
619 int count = 0;
620 struct dma_async_tx_descriptor *tx;
621 int i;
622
623 pr_debug("%s: stripe %llu block: %d\n",
624 __FUNCTION__, (unsigned long long)sh->sector, target);
625 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
626
627 for (i = disks; i--; )
628 if (i != target)
629 xor_srcs[count++] = sh->dev[i].page;
630
631 atomic_inc(&sh->count);
632
633 if (unlikely(count == 1))
634 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
635 0, NULL, ops_complete_compute5, sh);
636 else
637 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
638 ASYNC_TX_XOR_ZERO_DST, NULL,
639 ops_complete_compute5, sh);
640
641 /* ack now if postxor is not set to be run */
642 if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
643 async_tx_ack(tx);
644
645 return tx;
646}
647
648static void ops_complete_prexor(void *stripe_head_ref)
649{
650 struct stripe_head *sh = stripe_head_ref;
651
652 pr_debug("%s: stripe %llu\n", __FUNCTION__,
653 (unsigned long long)sh->sector);
654
655 set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
656}
657
658static struct dma_async_tx_descriptor *
659ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
660{
661 /* kernel stack size limits the total number of disks */
662 int disks = sh->disks;
663 struct page *xor_srcs[disks];
664 int count = 0, pd_idx = sh->pd_idx, i;
665
666 /* existing parity data subtracted */
667 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
668
669 pr_debug("%s: stripe %llu\n", __FUNCTION__,
670 (unsigned long long)sh->sector);
671
672 for (i = disks; i--; ) {
673 struct r5dev *dev = &sh->dev[i];
674 /* Only process blocks that are known to be uptodate */
675 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
676 xor_srcs[count++] = dev->page;
677 }
678
679 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
680 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
681 ops_complete_prexor, sh);
682
683 return tx;
684}
685
686static struct dma_async_tx_descriptor *
687ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
688{
689 int disks = sh->disks;
690 int pd_idx = sh->pd_idx, i;
691
692 /* check if prexor is active which means only process blocks
693 * that are part of a read-modify-write (Wantprexor)
694 */
695 int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
696
697 pr_debug("%s: stripe %llu\n", __FUNCTION__,
698 (unsigned long long)sh->sector);
699
700 for (i = disks; i--; ) {
701 struct r5dev *dev = &sh->dev[i];
702 struct bio *chosen;
703 int towrite;
704
705 towrite = 0;
706 if (prexor) { /* rmw */
707 if (dev->towrite &&
708 test_bit(R5_Wantprexor, &dev->flags))
709 towrite = 1;
710 } else { /* rcw */
711 if (i != pd_idx && dev->towrite &&
712 test_bit(R5_LOCKED, &dev->flags))
713 towrite = 1;
714 }
715
716 if (towrite) {
717 struct bio *wbi;
718
719 spin_lock(&sh->lock);
720 chosen = dev->towrite;
721 dev->towrite = NULL;
722 BUG_ON(dev->written);
723 wbi = dev->written = chosen;
724 spin_unlock(&sh->lock);
725
726 while (wbi && wbi->bi_sector <
727 dev->sector + STRIPE_SECTORS) {
728 tx = async_copy_data(1, wbi, dev->page,
729 dev->sector, tx);
730 wbi = r5_next_bio(wbi, dev->sector);
731 }
732 }
733 }
734
735 return tx;
736}
737
738static void ops_complete_postxor(void *stripe_head_ref)
739{
740 struct stripe_head *sh = stripe_head_ref;
741
742 pr_debug("%s: stripe %llu\n", __FUNCTION__,
743 (unsigned long long)sh->sector);
744
745 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
746 set_bit(STRIPE_HANDLE, &sh->state);
747 release_stripe(sh);
748}
749
750static void ops_complete_write(void *stripe_head_ref)
751{
752 struct stripe_head *sh = stripe_head_ref;
753 int disks = sh->disks, i, pd_idx = sh->pd_idx;
754
755 pr_debug("%s: stripe %llu\n", __FUNCTION__,
756 (unsigned long long)sh->sector);
757
758 for (i = disks; i--; ) {
759 struct r5dev *dev = &sh->dev[i];
760 if (dev->written || i == pd_idx)
761 set_bit(R5_UPTODATE, &dev->flags);
762 }
763
764 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
765 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
766
767 set_bit(STRIPE_HANDLE, &sh->state);
768 release_stripe(sh);
769}
770
771static void
772ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
773{
774 /* kernel stack size limits the total number of disks */
775 int disks = sh->disks;
776 struct page *xor_srcs[disks];
777
778 int count = 0, pd_idx = sh->pd_idx, i;
779 struct page *xor_dest;
780 int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
781 unsigned long flags;
782 dma_async_tx_callback callback;
783
784 pr_debug("%s: stripe %llu\n", __FUNCTION__,
785 (unsigned long long)sh->sector);
786
787 /* check if prexor is active which means only process blocks
788 * that are part of a read-modify-write (written)
789 */
790 if (prexor) {
791 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
792 for (i = disks; i--; ) {
793 struct r5dev *dev = &sh->dev[i];
794 if (dev->written)
795 xor_srcs[count++] = dev->page;
796 }
797 } else {
798 xor_dest = sh->dev[pd_idx].page;
799 for (i = disks; i--; ) {
800 struct r5dev *dev = &sh->dev[i];
801 if (i != pd_idx)
802 xor_srcs[count++] = dev->page;
803 }
804 }
805
806 /* check whether this postxor is part of a write */
807 callback = test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending) ?
808 ops_complete_write : ops_complete_postxor;
809
810 /* 1/ if we prexor'd then the dest is reused as a source
811 * 2/ if we did not prexor then we are redoing the parity
812 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
813 * for the synchronous xor case
814 */
815 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
816 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
817
818 atomic_inc(&sh->count);
819
820 if (unlikely(count == 1)) {
821 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
822 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
823 flags, tx, callback, sh);
824 } else
825 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
826 flags, tx, callback, sh);
827}
828
829static void ops_complete_check(void *stripe_head_ref)
830{
831 struct stripe_head *sh = stripe_head_ref;
832 int pd_idx = sh->pd_idx;
833
834 pr_debug("%s: stripe %llu\n", __FUNCTION__,
835 (unsigned long long)sh->sector);
836
837 if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
838 sh->ops.zero_sum_result == 0)
839 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
840
841 set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
842 set_bit(STRIPE_HANDLE, &sh->state);
843 release_stripe(sh);
844}
845
846static void ops_run_check(struct stripe_head *sh)
847{
848 /* kernel stack size limits the total number of disks */
849 int disks = sh->disks;
850 struct page *xor_srcs[disks];
851 struct dma_async_tx_descriptor *tx;
852
853 int count = 0, pd_idx = sh->pd_idx, i;
854 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
855
856 pr_debug("%s: stripe %llu\n", __FUNCTION__,
857 (unsigned long long)sh->sector);
858
859 for (i = disks; i--; ) {
860 struct r5dev *dev = &sh->dev[i];
861 if (i != pd_idx)
862 xor_srcs[count++] = dev->page;
863 }
864
865 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
866 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
867
868 if (tx)
869 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
870 else
871 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
872
873 atomic_inc(&sh->count);
874 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
875 ops_complete_check, sh);
876}
877
878static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
879{
880 int overlap_clear = 0, i, disks = sh->disks;
881 struct dma_async_tx_descriptor *tx = NULL;
882
883 if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
884 ops_run_biofill(sh);
885 overlap_clear++;
886 }
887
888 if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
889 tx = ops_run_compute5(sh, pending);
890
891 if (test_bit(STRIPE_OP_PREXOR, &pending))
892 tx = ops_run_prexor(sh, tx);
893
894 if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
895 tx = ops_run_biodrain(sh, tx);
896 overlap_clear++;
897 }
898
899 if (test_bit(STRIPE_OP_POSTXOR, &pending))
900 ops_run_postxor(sh, tx);
901
902 if (test_bit(STRIPE_OP_CHECK, &pending))
903 ops_run_check(sh);
904
905 if (test_bit(STRIPE_OP_IO, &pending))
906 ops_run_io(sh);
907
908 if (overlap_clear)
909 for (i = disks; i--; ) {
910 struct r5dev *dev = &sh->dev[i];
911 if (test_and_clear_bit(R5_Overlap, &dev->flags))
912 wake_up(&sh->raid_conf->wait_for_overlap);
913 }
914}
915
NeilBrown3f294f42005-11-08 21:39:25 -0800916static int grow_one_stripe(raid5_conf_t *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917{
918 struct stripe_head *sh;
NeilBrown3f294f42005-11-08 21:39:25 -0800919 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
920 if (!sh)
921 return 0;
922 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
923 sh->raid_conf = conf;
924 spin_lock_init(&sh->lock);
925
926 if (grow_buffers(sh, conf->raid_disks)) {
927 shrink_buffers(sh, conf->raid_disks);
928 kmem_cache_free(conf->slab_cache, sh);
929 return 0;
930 }
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800931 sh->disks = conf->raid_disks;
NeilBrown3f294f42005-11-08 21:39:25 -0800932 /* we just created an active stripe so... */
933 atomic_set(&sh->count, 1);
934 atomic_inc(&conf->active_stripes);
935 INIT_LIST_HEAD(&sh->lru);
936 release_stripe(sh);
937 return 1;
938}
939
940static int grow_stripes(raid5_conf_t *conf, int num)
941{
Christoph Lametere18b8902006-12-06 20:33:20 -0800942 struct kmem_cache *sc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 int devs = conf->raid_disks;
944
NeilBrown42b9beb2007-05-09 02:35:37 -0700945 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
946 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
NeilBrownad01c9e2006-03-27 01:18:07 -0800947 conf->active_name = 0;
948 sc = kmem_cache_create(conf->cache_name[conf->active_name],
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
Paul Mundt20c2df82007-07-20 10:11:58 +0900950 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 if (!sc)
952 return 1;
953 conf->slab_cache = sc;
NeilBrownad01c9e2006-03-27 01:18:07 -0800954 conf->pool_size = devs;
NeilBrown16a53ec2006-06-26 00:27:38 -0700955 while (num--)
NeilBrown3f294f42005-11-08 21:39:25 -0800956 if (!grow_one_stripe(conf))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 return 0;
959}
NeilBrown29269552006-03-27 01:18:10 -0800960
961#ifdef CONFIG_MD_RAID5_RESHAPE
NeilBrownad01c9e2006-03-27 01:18:07 -0800962static int resize_stripes(raid5_conf_t *conf, int newsize)
963{
964 /* Make all the stripes able to hold 'newsize' devices.
965 * New slots in each stripe get 'page' set to a new page.
966 *
967 * This happens in stages:
968 * 1/ create a new kmem_cache and allocate the required number of
969 * stripe_heads.
970 * 2/ gather all the old stripe_heads and tranfer the pages across
971 * to the new stripe_heads. This will have the side effect of
972 * freezing the array as once all stripe_heads have been collected,
973 * no IO will be possible. Old stripe heads are freed once their
974 * pages have been transferred over, and the old kmem_cache is
975 * freed when all stripes are done.
976 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
977 * we simple return a failre status - no need to clean anything up.
978 * 4/ allocate new pages for the new slots in the new stripe_heads.
979 * If this fails, we don't bother trying the shrink the
980 * stripe_heads down again, we just leave them as they are.
981 * As each stripe_head is processed the new one is released into
982 * active service.
983 *
984 * Once step2 is started, we cannot afford to wait for a write,
985 * so we use GFP_NOIO allocations.
986 */
987 struct stripe_head *osh, *nsh;
988 LIST_HEAD(newstripes);
989 struct disk_info *ndisks;
990 int err = 0;
Christoph Lametere18b8902006-12-06 20:33:20 -0800991 struct kmem_cache *sc;
NeilBrownad01c9e2006-03-27 01:18:07 -0800992 int i;
993
994 if (newsize <= conf->pool_size)
995 return 0; /* never bother to shrink */
996
NeilBrown2a2275d2007-01-26 00:57:11 -0800997 md_allow_write(conf->mddev);
998
NeilBrownad01c9e2006-03-27 01:18:07 -0800999 /* Step 1 */
1000 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1001 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
Paul Mundt20c2df82007-07-20 10:11:58 +09001002 0, 0, NULL);
NeilBrownad01c9e2006-03-27 01:18:07 -08001003 if (!sc)
1004 return -ENOMEM;
1005
1006 for (i = conf->max_nr_stripes; i; i--) {
1007 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1008 if (!nsh)
1009 break;
1010
1011 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1012
1013 nsh->raid_conf = conf;
1014 spin_lock_init(&nsh->lock);
1015
1016 list_add(&nsh->lru, &newstripes);
1017 }
1018 if (i) {
1019 /* didn't get enough, give up */
1020 while (!list_empty(&newstripes)) {
1021 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1022 list_del(&nsh->lru);
1023 kmem_cache_free(sc, nsh);
1024 }
1025 kmem_cache_destroy(sc);
1026 return -ENOMEM;
1027 }
1028 /* Step 2 - Must use GFP_NOIO now.
1029 * OK, we have enough stripes, start collecting inactive
1030 * stripes and copying them over
1031 */
1032 list_for_each_entry(nsh, &newstripes, lru) {
1033 spin_lock_irq(&conf->device_lock);
1034 wait_event_lock_irq(conf->wait_for_stripe,
1035 !list_empty(&conf->inactive_list),
1036 conf->device_lock,
NeilBrownb3b46be2006-03-27 01:18:16 -08001037 unplug_slaves(conf->mddev)
NeilBrownad01c9e2006-03-27 01:18:07 -08001038 );
1039 osh = get_free_stripe(conf);
1040 spin_unlock_irq(&conf->device_lock);
1041 atomic_set(&nsh->count, 1);
1042 for(i=0; i<conf->pool_size; i++)
1043 nsh->dev[i].page = osh->dev[i].page;
1044 for( ; i<newsize; i++)
1045 nsh->dev[i].page = NULL;
1046 kmem_cache_free(conf->slab_cache, osh);
1047 }
1048 kmem_cache_destroy(conf->slab_cache);
1049
1050 /* Step 3.
1051 * At this point, we are holding all the stripes so the array
1052 * is completely stalled, so now is a good time to resize
1053 * conf->disks.
1054 */
1055 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1056 if (ndisks) {
1057 for (i=0; i<conf->raid_disks; i++)
1058 ndisks[i] = conf->disks[i];
1059 kfree(conf->disks);
1060 conf->disks = ndisks;
1061 } else
1062 err = -ENOMEM;
1063
1064 /* Step 4, return new stripes to service */
1065 while(!list_empty(&newstripes)) {
1066 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1067 list_del_init(&nsh->lru);
1068 for (i=conf->raid_disks; i < newsize; i++)
1069 if (nsh->dev[i].page == NULL) {
1070 struct page *p = alloc_page(GFP_NOIO);
1071 nsh->dev[i].page = p;
1072 if (!p)
1073 err = -ENOMEM;
1074 }
1075 release_stripe(nsh);
1076 }
1077 /* critical section pass, GFP_NOIO no longer needed */
1078
1079 conf->slab_cache = sc;
1080 conf->active_name = 1-conf->active_name;
1081 conf->pool_size = newsize;
1082 return err;
1083}
NeilBrown29269552006-03-27 01:18:10 -08001084#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085
NeilBrown3f294f42005-11-08 21:39:25 -08001086static int drop_one_stripe(raid5_conf_t *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087{
1088 struct stripe_head *sh;
1089
NeilBrown3f294f42005-11-08 21:39:25 -08001090 spin_lock_irq(&conf->device_lock);
1091 sh = get_free_stripe(conf);
1092 spin_unlock_irq(&conf->device_lock);
1093 if (!sh)
1094 return 0;
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +02001095 BUG_ON(atomic_read(&sh->count));
NeilBrownad01c9e2006-03-27 01:18:07 -08001096 shrink_buffers(sh, conf->pool_size);
NeilBrown3f294f42005-11-08 21:39:25 -08001097 kmem_cache_free(conf->slab_cache, sh);
1098 atomic_dec(&conf->active_stripes);
1099 return 1;
1100}
1101
1102static void shrink_stripes(raid5_conf_t *conf)
1103{
1104 while (drop_one_stripe(conf))
1105 ;
1106
NeilBrown29fc7e32006-02-03 03:03:41 -08001107 if (conf->slab_cache)
1108 kmem_cache_destroy(conf->slab_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 conf->slab_cache = NULL;
1110}
1111
NeilBrown6712ecf2007-09-27 12:47:43 +02001112static void raid5_end_read_request(struct bio * bi, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113{
1114 struct stripe_head *sh = bi->bi_private;
1115 raid5_conf_t *conf = sh->raid_conf;
NeilBrown7ecaa1e2006-03-27 01:18:08 -08001116 int disks = sh->disks, i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
NeilBrownd6950432006-07-10 04:44:20 -07001118 char b[BDEVNAME_SIZE];
1119 mdk_rdev_t *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121
1122 for (i=0 ; i<disks; i++)
1123 if (bi == &sh->dev[i].req)
1124 break;
1125
Dan Williams45b42332007-07-09 11:56:43 -07001126 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1127 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 uptodate);
1129 if (i == disks) {
1130 BUG();
NeilBrown6712ecf2007-09-27 12:47:43 +02001131 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 }
1133
1134 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 set_bit(R5_UPTODATE, &sh->dev[i].flags);
NeilBrown4e5314b2005-11-08 21:39:22 -08001136 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
NeilBrownd6950432006-07-10 04:44:20 -07001137 rdev = conf->disks[i].rdev;
1138 printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1139 mdname(conf->mddev), STRIPE_SECTORS,
1140 (unsigned long long)sh->sector + rdev->data_offset,
1141 bdevname(rdev->bdev, b));
NeilBrown4e5314b2005-11-08 21:39:22 -08001142 clear_bit(R5_ReadError, &sh->dev[i].flags);
1143 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1144 }
NeilBrownba22dcb2005-11-08 21:39:31 -08001145 if (atomic_read(&conf->disks[i].rdev->read_errors))
1146 atomic_set(&conf->disks[i].rdev->read_errors, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 } else {
NeilBrownd6950432006-07-10 04:44:20 -07001148 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
NeilBrownba22dcb2005-11-08 21:39:31 -08001149 int retry = 0;
NeilBrownd6950432006-07-10 04:44:20 -07001150 rdev = conf->disks[i].rdev;
1151
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
NeilBrownd6950432006-07-10 04:44:20 -07001153 atomic_inc(&rdev->read_errors);
NeilBrownba22dcb2005-11-08 21:39:31 -08001154 if (conf->mddev->degraded)
NeilBrownd6950432006-07-10 04:44:20 -07001155 printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1156 mdname(conf->mddev),
1157 (unsigned long long)sh->sector + rdev->data_offset,
1158 bdn);
NeilBrownba22dcb2005-11-08 21:39:31 -08001159 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
NeilBrown4e5314b2005-11-08 21:39:22 -08001160 /* Oh, no!!! */
NeilBrownd6950432006-07-10 04:44:20 -07001161 printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1162 mdname(conf->mddev),
1163 (unsigned long long)sh->sector + rdev->data_offset,
1164 bdn);
1165 else if (atomic_read(&rdev->read_errors)
NeilBrownba22dcb2005-11-08 21:39:31 -08001166 > conf->max_nr_stripes)
NeilBrown14f8d262006-01-06 00:20:14 -08001167 printk(KERN_WARNING
NeilBrownd6950432006-07-10 04:44:20 -07001168 "raid5:%s: Too many read errors, failing device %s.\n",
1169 mdname(conf->mddev), bdn);
NeilBrownba22dcb2005-11-08 21:39:31 -08001170 else
1171 retry = 1;
1172 if (retry)
1173 set_bit(R5_ReadError, &sh->dev[i].flags);
1174 else {
NeilBrown4e5314b2005-11-08 21:39:22 -08001175 clear_bit(R5_ReadError, &sh->dev[i].flags);
1176 clear_bit(R5_ReWrite, &sh->dev[i].flags);
NeilBrownd6950432006-07-10 04:44:20 -07001177 md_error(conf->mddev, rdev);
NeilBrownba22dcb2005-11-08 21:39:31 -08001178 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 }
1180 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1182 set_bit(STRIPE_HANDLE, &sh->state);
1183 release_stripe(sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184}
1185
NeilBrown6712ecf2007-09-27 12:47:43 +02001186static void raid5_end_write_request (struct bio *bi, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187{
1188 struct stripe_head *sh = bi->bi_private;
1189 raid5_conf_t *conf = sh->raid_conf;
NeilBrown7ecaa1e2006-03-27 01:18:08 -08001190 int disks = sh->disks, i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1192
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 for (i=0 ; i<disks; i++)
1194 if (bi == &sh->dev[i].req)
1195 break;
1196
Dan Williams45b42332007-07-09 11:56:43 -07001197 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1199 uptodate);
1200 if (i == disks) {
1201 BUG();
NeilBrown6712ecf2007-09-27 12:47:43 +02001202 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 }
1204
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 if (!uptodate)
1206 md_error(conf->mddev, conf->disks[i].rdev);
1207
1208 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1209
1210 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1211 set_bit(STRIPE_HANDLE, &sh->state);
NeilBrownc04be0a2006-10-03 01:15:53 -07001212 release_stripe(sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213}
1214
1215
1216static sector_t compute_blocknr(struct stripe_head *sh, int i);
1217
1218static void raid5_build_block (struct stripe_head *sh, int i)
1219{
1220 struct r5dev *dev = &sh->dev[i];
1221
1222 bio_init(&dev->req);
1223 dev->req.bi_io_vec = &dev->vec;
1224 dev->req.bi_vcnt++;
1225 dev->req.bi_max_vecs++;
1226 dev->vec.bv_page = dev->page;
1227 dev->vec.bv_len = STRIPE_SIZE;
1228 dev->vec.bv_offset = 0;
1229
1230 dev->req.bi_sector = sh->sector;
1231 dev->req.bi_private = sh;
1232
1233 dev->flags = 0;
NeilBrown16a53ec2006-06-26 00:27:38 -07001234 dev->sector = compute_blocknr(sh, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235}
1236
1237static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1238{
1239 char b[BDEVNAME_SIZE];
1240 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
Dan Williams45b42332007-07-09 11:56:43 -07001241 pr_debug("raid5: error called\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242
NeilBrownb2d444d2005-11-08 21:39:31 -08001243 if (!test_bit(Faulty, &rdev->flags)) {
NeilBrown850b2b42006-10-03 01:15:46 -07001244 set_bit(MD_CHANGE_DEVS, &mddev->flags);
NeilBrownc04be0a2006-10-03 01:15:53 -07001245 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1246 unsigned long flags;
1247 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 mddev->degraded++;
NeilBrownc04be0a2006-10-03 01:15:53 -07001249 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 /*
1251 * if recovery was running, make sure it aborts.
1252 */
1253 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1254 }
NeilBrownb2d444d2005-11-08 21:39:31 -08001255 set_bit(Faulty, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 printk (KERN_ALERT
1257 "raid5: Disk failure on %s, disabling device."
1258 " Operation continuing on %d devices\n",
NeilBrown02c2de82006-10-03 01:15:47 -07001259 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 }
NeilBrown16a53ec2006-06-26 00:27:38 -07001261}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262
1263/*
1264 * Input: a 'big' sector number,
1265 * Output: index of the data and parity disk, and the sector # in them.
1266 */
1267static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1268 unsigned int data_disks, unsigned int * dd_idx,
1269 unsigned int * pd_idx, raid5_conf_t *conf)
1270{
1271 long stripe;
1272 unsigned long chunk_number;
1273 unsigned int chunk_offset;
1274 sector_t new_sector;
1275 int sectors_per_chunk = conf->chunk_size >> 9;
1276
1277 /* First compute the information on this sector */
1278
1279 /*
1280 * Compute the chunk number and the sector offset inside the chunk
1281 */
1282 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1283 chunk_number = r_sector;
1284 BUG_ON(r_sector != chunk_number);
1285
1286 /*
1287 * Compute the stripe number
1288 */
1289 stripe = chunk_number / data_disks;
1290
1291 /*
1292 * Compute the data disk and parity disk indexes inside the stripe
1293 */
1294 *dd_idx = chunk_number % data_disks;
1295
1296 /*
1297 * Select the parity disk based on the user selected algorithm.
1298 */
NeilBrown16a53ec2006-06-26 00:27:38 -07001299 switch(conf->level) {
1300 case 4:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 *pd_idx = data_disks;
NeilBrown16a53ec2006-06-26 00:27:38 -07001302 break;
1303 case 5:
1304 switch (conf->algorithm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 case ALGORITHM_LEFT_ASYMMETRIC:
1306 *pd_idx = data_disks - stripe % raid_disks;
1307 if (*dd_idx >= *pd_idx)
1308 (*dd_idx)++;
1309 break;
1310 case ALGORITHM_RIGHT_ASYMMETRIC:
1311 *pd_idx = stripe % raid_disks;
1312 if (*dd_idx >= *pd_idx)
1313 (*dd_idx)++;
1314 break;
1315 case ALGORITHM_LEFT_SYMMETRIC:
1316 *pd_idx = data_disks - stripe % raid_disks;
1317 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1318 break;
1319 case ALGORITHM_RIGHT_SYMMETRIC:
1320 *pd_idx = stripe % raid_disks;
1321 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1322 break;
1323 default:
NeilBrown14f8d262006-01-06 00:20:14 -08001324 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 conf->algorithm);
NeilBrown16a53ec2006-06-26 00:27:38 -07001326 }
1327 break;
1328 case 6:
1329
1330 /**** FIX THIS ****/
1331 switch (conf->algorithm) {
1332 case ALGORITHM_LEFT_ASYMMETRIC:
1333 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1334 if (*pd_idx == raid_disks-1)
1335 (*dd_idx)++; /* Q D D D P */
1336 else if (*dd_idx >= *pd_idx)
1337 (*dd_idx) += 2; /* D D P Q D */
1338 break;
1339 case ALGORITHM_RIGHT_ASYMMETRIC:
1340 *pd_idx = stripe % raid_disks;
1341 if (*pd_idx == raid_disks-1)
1342 (*dd_idx)++; /* Q D D D P */
1343 else if (*dd_idx >= *pd_idx)
1344 (*dd_idx) += 2; /* D D P Q D */
1345 break;
1346 case ALGORITHM_LEFT_SYMMETRIC:
1347 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1348 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1349 break;
1350 case ALGORITHM_RIGHT_SYMMETRIC:
1351 *pd_idx = stripe % raid_disks;
1352 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1353 break;
1354 default:
1355 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1356 conf->algorithm);
1357 }
1358 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 }
1360
1361 /*
1362 * Finally, compute the new sector number
1363 */
1364 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1365 return new_sector;
1366}
1367
1368
1369static sector_t compute_blocknr(struct stripe_head *sh, int i)
1370{
1371 raid5_conf_t *conf = sh->raid_conf;
NeilBrownb875e532006-12-10 02:20:49 -08001372 int raid_disks = sh->disks;
1373 int data_disks = raid_disks - conf->max_degraded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 sector_t new_sector = sh->sector, check;
1375 int sectors_per_chunk = conf->chunk_size >> 9;
1376 sector_t stripe;
1377 int chunk_offset;
1378 int chunk_number, dummy1, dummy2, dd_idx = i;
1379 sector_t r_sector;
1380
NeilBrown16a53ec2006-06-26 00:27:38 -07001381
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1383 stripe = new_sector;
1384 BUG_ON(new_sector != stripe);
1385
NeilBrown16a53ec2006-06-26 00:27:38 -07001386 if (i == sh->pd_idx)
1387 return 0;
1388 switch(conf->level) {
1389 case 4: break;
1390 case 5:
1391 switch (conf->algorithm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 case ALGORITHM_LEFT_ASYMMETRIC:
1393 case ALGORITHM_RIGHT_ASYMMETRIC:
1394 if (i > sh->pd_idx)
1395 i--;
1396 break;
1397 case ALGORITHM_LEFT_SYMMETRIC:
1398 case ALGORITHM_RIGHT_SYMMETRIC:
1399 if (i < sh->pd_idx)
1400 i += raid_disks;
1401 i -= (sh->pd_idx + 1);
1402 break;
1403 default:
NeilBrown14f8d262006-01-06 00:20:14 -08001404 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
NeilBrown16a53ec2006-06-26 00:27:38 -07001405 conf->algorithm);
1406 }
1407 break;
1408 case 6:
NeilBrown16a53ec2006-06-26 00:27:38 -07001409 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1410 return 0; /* It is the Q disk */
1411 switch (conf->algorithm) {
1412 case ALGORITHM_LEFT_ASYMMETRIC:
1413 case ALGORITHM_RIGHT_ASYMMETRIC:
1414 if (sh->pd_idx == raid_disks-1)
1415 i--; /* Q D D D P */
1416 else if (i > sh->pd_idx)
1417 i -= 2; /* D D P Q D */
1418 break;
1419 case ALGORITHM_LEFT_SYMMETRIC:
1420 case ALGORITHM_RIGHT_SYMMETRIC:
1421 if (sh->pd_idx == raid_disks-1)
1422 i--; /* Q D D D P */
1423 else {
1424 /* D D P Q D */
1425 if (i < sh->pd_idx)
1426 i += raid_disks;
1427 i -= (sh->pd_idx + 2);
1428 }
1429 break;
1430 default:
1431 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 conf->algorithm);
NeilBrown16a53ec2006-06-26 00:27:38 -07001433 }
1434 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435 }
1436
1437 chunk_number = stripe * data_disks + i;
1438 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1439
1440 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1441 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
NeilBrown14f8d262006-01-06 00:20:14 -08001442 printk(KERN_ERR "compute_blocknr: map not correct\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443 return 0;
1444 }
1445 return r_sector;
1446}
1447
1448
1449
1450/*
NeilBrown16a53ec2006-06-26 00:27:38 -07001451 * Copy data between a page in the stripe cache, and one or more bion
1452 * The page could align with the middle of the bio, or there could be
1453 * several bion, each with several bio_vecs, which cover part of the page
1454 * Multiple bion are linked together on bi_next. There may be extras
1455 * at the end of this list. We ignore them.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 */
1457static void copy_data(int frombio, struct bio *bio,
1458 struct page *page,
1459 sector_t sector)
1460{
1461 char *pa = page_address(page);
1462 struct bio_vec *bvl;
1463 int i;
1464 int page_offset;
1465
1466 if (bio->bi_sector >= sector)
1467 page_offset = (signed)(bio->bi_sector - sector) * 512;
1468 else
1469 page_offset = (signed)(sector - bio->bi_sector) * -512;
1470 bio_for_each_segment(bvl, bio, i) {
1471 int len = bio_iovec_idx(bio,i)->bv_len;
1472 int clen;
1473 int b_offset = 0;
1474
1475 if (page_offset < 0) {
1476 b_offset = -page_offset;
1477 page_offset += b_offset;
1478 len -= b_offset;
1479 }
1480
1481 if (len > 0 && page_offset + len > STRIPE_SIZE)
1482 clen = STRIPE_SIZE - page_offset;
1483 else clen = len;
NeilBrown16a53ec2006-06-26 00:27:38 -07001484
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 if (clen > 0) {
1486 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1487 if (frombio)
1488 memcpy(pa+page_offset, ba+b_offset, clen);
1489 else
1490 memcpy(ba+b_offset, pa+page_offset, clen);
1491 __bio_kunmap_atomic(ba, KM_USER0);
1492 }
1493 if (clen < len) /* hit end of page */
1494 break;
1495 page_offset += len;
1496 }
1497}
1498
Dan Williams9bc89cd2007-01-02 11:10:44 -07001499#define check_xor() do { \
1500 if (count == MAX_XOR_BLOCKS) { \
1501 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1502 count = 0; \
1503 } \
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504 } while(0)
1505
NeilBrown16a53ec2006-06-26 00:27:38 -07001506static void compute_parity6(struct stripe_head *sh, int method)
1507{
1508 raid6_conf_t *conf = sh->raid_conf;
NeilBrownf4168852007-02-28 20:11:53 -08001509 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
NeilBrown16a53ec2006-06-26 00:27:38 -07001510 struct bio *chosen;
1511 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1512 void *ptrs[disks];
1513
1514 qd_idx = raid6_next_disk(pd_idx, disks);
1515 d0_idx = raid6_next_disk(qd_idx, disks);
1516
Dan Williams45b42332007-07-09 11:56:43 -07001517 pr_debug("compute_parity, stripe %llu, method %d\n",
NeilBrown16a53ec2006-06-26 00:27:38 -07001518 (unsigned long long)sh->sector, method);
1519
1520 switch(method) {
1521 case READ_MODIFY_WRITE:
1522 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1523 case RECONSTRUCT_WRITE:
1524 for (i= disks; i-- ;)
1525 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1526 chosen = sh->dev[i].towrite;
1527 sh->dev[i].towrite = NULL;
1528
1529 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1530 wake_up(&conf->wait_for_overlap);
1531
Eric Sesterhenn52e5f9d2006-10-03 23:33:23 +02001532 BUG_ON(sh->dev[i].written);
NeilBrown16a53ec2006-06-26 00:27:38 -07001533 sh->dev[i].written = chosen;
1534 }
1535 break;
1536 case CHECK_PARITY:
1537 BUG(); /* Not implemented yet */
1538 }
1539
1540 for (i = disks; i--;)
1541 if (sh->dev[i].written) {
1542 sector_t sector = sh->dev[i].sector;
1543 struct bio *wbi = sh->dev[i].written;
1544 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1545 copy_data(1, wbi, sh->dev[i].page, sector);
1546 wbi = r5_next_bio(wbi, sector);
1547 }
1548
1549 set_bit(R5_LOCKED, &sh->dev[i].flags);
1550 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1551 }
1552
1553// switch(method) {
1554// case RECONSTRUCT_WRITE:
1555// case CHECK_PARITY:
1556// case UPDATE_PARITY:
1557 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1558 /* FIX: Is this ordering of drives even remotely optimal? */
1559 count = 0;
1560 i = d0_idx;
1561 do {
1562 ptrs[count++] = page_address(sh->dev[i].page);
1563 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1564 printk("block %d/%d not uptodate on parity calc\n", i,count);
1565 i = raid6_next_disk(i, disks);
1566 } while ( i != d0_idx );
1567// break;
1568// }
1569
1570 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1571
1572 switch(method) {
1573 case RECONSTRUCT_WRITE:
1574 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1575 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1576 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1577 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1578 break;
1579 case UPDATE_PARITY:
1580 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1581 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1582 break;
1583 }
1584}
1585
1586
1587/* Compute one missing block */
1588static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1589{
NeilBrownf4168852007-02-28 20:11:53 -08001590 int i, count, disks = sh->disks;
Dan Williams9bc89cd2007-01-02 11:10:44 -07001591 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
NeilBrown16a53ec2006-06-26 00:27:38 -07001592 int pd_idx = sh->pd_idx;
1593 int qd_idx = raid6_next_disk(pd_idx, disks);
1594
Dan Williams45b42332007-07-09 11:56:43 -07001595 pr_debug("compute_block_1, stripe %llu, idx %d\n",
NeilBrown16a53ec2006-06-26 00:27:38 -07001596 (unsigned long long)sh->sector, dd_idx);
1597
1598 if ( dd_idx == qd_idx ) {
1599 /* We're actually computing the Q drive */
1600 compute_parity6(sh, UPDATE_PARITY);
1601 } else {
Dan Williams9bc89cd2007-01-02 11:10:44 -07001602 dest = page_address(sh->dev[dd_idx].page);
1603 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1604 count = 0;
NeilBrown16a53ec2006-06-26 00:27:38 -07001605 for (i = disks ; i--; ) {
1606 if (i == dd_idx || i == qd_idx)
1607 continue;
1608 p = page_address(sh->dev[i].page);
1609 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1610 ptr[count++] = p;
1611 else
1612 printk("compute_block() %d, stripe %llu, %d"
1613 " not present\n", dd_idx,
1614 (unsigned long long)sh->sector, i);
1615
1616 check_xor();
1617 }
Dan Williams9bc89cd2007-01-02 11:10:44 -07001618 if (count)
1619 xor_blocks(count, STRIPE_SIZE, dest, ptr);
NeilBrown16a53ec2006-06-26 00:27:38 -07001620 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1621 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1622 }
1623}
1624
1625/* Compute two missing blocks */
1626static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1627{
NeilBrownf4168852007-02-28 20:11:53 -08001628 int i, count, disks = sh->disks;
NeilBrown16a53ec2006-06-26 00:27:38 -07001629 int pd_idx = sh->pd_idx;
1630 int qd_idx = raid6_next_disk(pd_idx, disks);
1631 int d0_idx = raid6_next_disk(qd_idx, disks);
1632 int faila, failb;
1633
1634 /* faila and failb are disk numbers relative to d0_idx */
1635 /* pd_idx become disks-2 and qd_idx become disks-1 */
1636 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1637 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1638
1639 BUG_ON(faila == failb);
1640 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1641
Dan Williams45b42332007-07-09 11:56:43 -07001642 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
NeilBrown16a53ec2006-06-26 00:27:38 -07001643 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1644
1645 if ( failb == disks-1 ) {
1646 /* Q disk is one of the missing disks */
1647 if ( faila == disks-2 ) {
1648 /* Missing P+Q, just recompute */
1649 compute_parity6(sh, UPDATE_PARITY);
1650 return;
1651 } else {
1652 /* We're missing D+Q; recompute D from P */
1653 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1654 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1655 return;
1656 }
1657 }
1658
1659 /* We're missing D+P or D+D; build pointer table */
1660 {
1661 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1662 void *ptrs[disks];
1663
1664 count = 0;
1665 i = d0_idx;
1666 do {
1667 ptrs[count++] = page_address(sh->dev[i].page);
1668 i = raid6_next_disk(i, disks);
1669 if (i != dd_idx1 && i != dd_idx2 &&
1670 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1671 printk("compute_2 with missing block %d/%d\n", count, i);
1672 } while ( i != d0_idx );
1673
1674 if ( failb == disks-2 ) {
1675 /* We're missing D+P. */
1676 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1677 } else {
1678 /* We're missing D+D. */
1679 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1680 }
1681
1682 /* Both the above update both missing blocks */
1683 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1684 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1685 }
1686}
1687
Dan Williamse33129d2007-01-02 13:52:30 -07001688static int
1689handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1690{
1691 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1692 int locked = 0;
NeilBrown16a53ec2006-06-26 00:27:38 -07001693
Dan Williamse33129d2007-01-02 13:52:30 -07001694 if (rcw) {
1695 /* if we are not expanding this is a proper write request, and
1696 * there will be bios with new data to be drained into the
1697 * stripe cache
1698 */
1699 if (!expand) {
1700 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1701 sh->ops.count++;
1702 }
1703
1704 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1705 sh->ops.count++;
1706
1707 for (i = disks; i--; ) {
1708 struct r5dev *dev = &sh->dev[i];
1709
1710 if (dev->towrite) {
1711 set_bit(R5_LOCKED, &dev->flags);
1712 if (!expand)
1713 clear_bit(R5_UPTODATE, &dev->flags);
1714 locked++;
1715 }
1716 }
1717 } else {
1718 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1719 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1720
1721 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1722 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1723 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1724
1725 sh->ops.count += 3;
1726
1727 for (i = disks; i--; ) {
1728 struct r5dev *dev = &sh->dev[i];
1729 if (i == pd_idx)
1730 continue;
1731
1732 /* For a read-modify write there may be blocks that are
1733 * locked for reading while others are ready to be
1734 * written so we distinguish these blocks by the
1735 * R5_Wantprexor bit
1736 */
1737 if (dev->towrite &&
1738 (test_bit(R5_UPTODATE, &dev->flags) ||
1739 test_bit(R5_Wantcompute, &dev->flags))) {
1740 set_bit(R5_Wantprexor, &dev->flags);
1741 set_bit(R5_LOCKED, &dev->flags);
1742 clear_bit(R5_UPTODATE, &dev->flags);
1743 locked++;
1744 }
1745 }
1746 }
1747
1748 /* keep the parity disk locked while asynchronous operations
1749 * are in flight
1750 */
1751 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1752 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1753 locked++;
1754
1755 pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1756 __FUNCTION__, (unsigned long long)sh->sector,
1757 locked, sh->ops.pending);
1758
1759 return locked;
1760}
NeilBrown16a53ec2006-06-26 00:27:38 -07001761
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762/*
1763 * Each stripe/dev can have one or more bion attached.
NeilBrown16a53ec2006-06-26 00:27:38 -07001764 * toread/towrite point to the first in a chain.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 * The bi_next chain must be in order.
1766 */
1767static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1768{
1769 struct bio **bip;
1770 raid5_conf_t *conf = sh->raid_conf;
NeilBrown72626682005-09-09 16:23:54 -07001771 int firstwrite=0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772
Dan Williams45b42332007-07-09 11:56:43 -07001773 pr_debug("adding bh b#%llu to stripe s#%llu\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 (unsigned long long)bi->bi_sector,
1775 (unsigned long long)sh->sector);
1776
1777
1778 spin_lock(&sh->lock);
1779 spin_lock_irq(&conf->device_lock);
NeilBrown72626682005-09-09 16:23:54 -07001780 if (forwrite) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 bip = &sh->dev[dd_idx].towrite;
NeilBrown72626682005-09-09 16:23:54 -07001782 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1783 firstwrite = 1;
1784 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 bip = &sh->dev[dd_idx].toread;
1786 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1787 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1788 goto overlap;
1789 bip = & (*bip)->bi_next;
1790 }
1791 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1792 goto overlap;
1793
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +02001794 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 if (*bip)
1796 bi->bi_next = *bip;
1797 *bip = bi;
1798 bi->bi_phys_segments ++;
1799 spin_unlock_irq(&conf->device_lock);
1800 spin_unlock(&sh->lock);
1801
Dan Williams45b42332007-07-09 11:56:43 -07001802 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 (unsigned long long)bi->bi_sector,
1804 (unsigned long long)sh->sector, dd_idx);
1805
NeilBrown72626682005-09-09 16:23:54 -07001806 if (conf->mddev->bitmap && firstwrite) {
NeilBrown72626682005-09-09 16:23:54 -07001807 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1808 STRIPE_SECTORS, 0);
NeilBrownae3c20c2006-07-10 04:44:17 -07001809 sh->bm_seq = conf->seq_flush+1;
NeilBrown72626682005-09-09 16:23:54 -07001810 set_bit(STRIPE_BIT_DELAY, &sh->state);
1811 }
1812
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 if (forwrite) {
1814 /* check if page is covered */
1815 sector_t sector = sh->dev[dd_idx].sector;
1816 for (bi=sh->dev[dd_idx].towrite;
1817 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1818 bi && bi->bi_sector <= sector;
1819 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1820 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1821 sector = bi->bi_sector + (bi->bi_size>>9);
1822 }
1823 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1824 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1825 }
1826 return 1;
1827
1828 overlap:
1829 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1830 spin_unlock_irq(&conf->device_lock);
1831 spin_unlock(&sh->lock);
1832 return 0;
1833}
1834
NeilBrown29269552006-03-27 01:18:10 -08001835static void end_reshape(raid5_conf_t *conf);
1836
NeilBrown16a53ec2006-06-26 00:27:38 -07001837static int page_is_zero(struct page *p)
1838{
1839 char *a = page_address(p);
1840 return ((*(u32*)a) == 0 &&
1841 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1842}
1843
NeilBrownccfcc3c2006-03-27 01:18:09 -08001844static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1845{
1846 int sectors_per_chunk = conf->chunk_size >> 9;
NeilBrownccfcc3c2006-03-27 01:18:09 -08001847 int pd_idx, dd_idx;
Coywolf Qi Hunt2d2063c2006-10-03 01:15:50 -07001848 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1849
NeilBrownb875e532006-12-10 02:20:49 -08001850 raid5_compute_sector(stripe * (disks - conf->max_degraded)
1851 *sectors_per_chunk + chunk_offset,
1852 disks, disks - conf->max_degraded,
1853 &dd_idx, &pd_idx, conf);
NeilBrownccfcc3c2006-03-27 01:18:09 -08001854 return pd_idx;
1855}
1856
Dan Williamsa4456852007-07-09 11:56:43 -07001857static void
1858handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1859 struct stripe_head_state *s, int disks,
1860 struct bio **return_bi)
1861{
1862 int i;
1863 for (i = disks; i--; ) {
1864 struct bio *bi;
1865 int bitmap_end = 0;
1866
1867 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1868 mdk_rdev_t *rdev;
1869 rcu_read_lock();
1870 rdev = rcu_dereference(conf->disks[i].rdev);
1871 if (rdev && test_bit(In_sync, &rdev->flags))
1872 /* multiple read failures in one stripe */
1873 md_error(conf->mddev, rdev);
1874 rcu_read_unlock();
1875 }
1876 spin_lock_irq(&conf->device_lock);
1877 /* fail all writes first */
1878 bi = sh->dev[i].towrite;
1879 sh->dev[i].towrite = NULL;
1880 if (bi) {
1881 s->to_write--;
1882 bitmap_end = 1;
1883 }
1884
1885 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1886 wake_up(&conf->wait_for_overlap);
1887
1888 while (bi && bi->bi_sector <
1889 sh->dev[i].sector + STRIPE_SECTORS) {
1890 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1891 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1892 if (--bi->bi_phys_segments == 0) {
1893 md_write_end(conf->mddev);
1894 bi->bi_next = *return_bi;
1895 *return_bi = bi;
1896 }
1897 bi = nextbi;
1898 }
1899 /* and fail all 'written' */
1900 bi = sh->dev[i].written;
1901 sh->dev[i].written = NULL;
1902 if (bi) bitmap_end = 1;
1903 while (bi && bi->bi_sector <
1904 sh->dev[i].sector + STRIPE_SECTORS) {
1905 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1906 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1907 if (--bi->bi_phys_segments == 0) {
1908 md_write_end(conf->mddev);
1909 bi->bi_next = *return_bi;
1910 *return_bi = bi;
1911 }
1912 bi = bi2;
1913 }
1914
Dan Williamsb5e98d62007-01-02 13:52:31 -07001915 /* fail any reads if this device is non-operational and
1916 * the data has not reached the cache yet.
1917 */
1918 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1919 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1920 test_bit(R5_ReadError, &sh->dev[i].flags))) {
Dan Williamsa4456852007-07-09 11:56:43 -07001921 bi = sh->dev[i].toread;
1922 sh->dev[i].toread = NULL;
1923 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1924 wake_up(&conf->wait_for_overlap);
1925 if (bi) s->to_read--;
1926 while (bi && bi->bi_sector <
1927 sh->dev[i].sector + STRIPE_SECTORS) {
1928 struct bio *nextbi =
1929 r5_next_bio(bi, sh->dev[i].sector);
1930 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1931 if (--bi->bi_phys_segments == 0) {
1932 bi->bi_next = *return_bi;
1933 *return_bi = bi;
1934 }
1935 bi = nextbi;
1936 }
1937 }
1938 spin_unlock_irq(&conf->device_lock);
1939 if (bitmap_end)
1940 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1941 STRIPE_SECTORS, 0, 0);
1942 }
1943
1944}
1945
Dan Williamsf38e1212007-01-02 13:52:30 -07001946/* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1947 * to process
1948 */
1949static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1950 struct stripe_head_state *s, int disk_idx, int disks)
1951{
1952 struct r5dev *dev = &sh->dev[disk_idx];
1953 struct r5dev *failed_dev = &sh->dev[s->failed_num];
1954
1955 /* don't schedule compute operations or reads on the parity block while
1956 * a check is in flight
1957 */
1958 if ((disk_idx == sh->pd_idx) &&
1959 test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1960 return ~0;
1961
1962 /* is the data in this block needed, and can we get it? */
1963 if (!test_bit(R5_LOCKED, &dev->flags) &&
1964 !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1965 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1966 s->syncing || s->expanding || (s->failed &&
1967 (failed_dev->toread || (failed_dev->towrite &&
1968 !test_bit(R5_OVERWRITE, &failed_dev->flags)
1969 ))))) {
1970 /* 1/ We would like to get this block, possibly by computing it,
1971 * but we might not be able to.
1972 *
1973 * 2/ Since parity check operations potentially make the parity
1974 * block !uptodate it will need to be refreshed before any
1975 * compute operations on data disks are scheduled.
1976 *
1977 * 3/ We hold off parity block re-reads until check operations
1978 * have quiesced.
1979 */
1980 if ((s->uptodate == disks - 1) &&
1981 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1982 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1983 set_bit(R5_Wantcompute, &dev->flags);
1984 sh->ops.target = disk_idx;
1985 s->req_compute = 1;
1986 sh->ops.count++;
1987 /* Careful: from this point on 'uptodate' is in the eye
1988 * of raid5_run_ops which services 'compute' operations
1989 * before writes. R5_Wantcompute flags a block that will
1990 * be R5_UPTODATE by the time it is needed for a
1991 * subsequent operation.
1992 */
1993 s->uptodate++;
1994 return 0; /* uptodate + compute == disks */
1995 } else if ((s->uptodate < disks - 1) &&
1996 test_bit(R5_Insync, &dev->flags)) {
1997 /* Note: we hold off compute operations while checks are
1998 * in flight, but we still prefer 'compute' over 'read'
1999 * hence we only read if (uptodate < * disks-1)
2000 */
2001 set_bit(R5_LOCKED, &dev->flags);
2002 set_bit(R5_Wantread, &dev->flags);
2003 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2004 sh->ops.count++;
2005 s->locked++;
2006 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2007 s->syncing);
2008 }
2009 }
2010
2011 return ~0;
2012}
2013
Dan Williamsa4456852007-07-09 11:56:43 -07002014static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2015 struct stripe_head_state *s, int disks)
2016{
2017 int i;
Dan Williamsf38e1212007-01-02 13:52:30 -07002018
2019 /* Clear completed compute operations. Parity recovery
2020 * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2021 * later on in this routine
2022 */
2023 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2024 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2025 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2026 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2027 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2028 }
2029
2030 /* look for blocks to read/compute, skip this if a compute
2031 * is already in flight, or if the stripe contents are in the
2032 * midst of changing due to a write
2033 */
2034 if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2035 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2036 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2037 for (i = disks; i--; )
2038 if (__handle_issuing_new_read_requests5(
2039 sh, s, i, disks) == 0)
2040 break;
Dan Williamsa4456852007-07-09 11:56:43 -07002041 }
2042 set_bit(STRIPE_HANDLE, &sh->state);
2043}
2044
2045static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2046 struct stripe_head_state *s, struct r6_state *r6s,
2047 int disks)
2048{
2049 int i;
2050 for (i = disks; i--; ) {
2051 struct r5dev *dev = &sh->dev[i];
2052 if (!test_bit(R5_LOCKED, &dev->flags) &&
2053 !test_bit(R5_UPTODATE, &dev->flags) &&
2054 (dev->toread || (dev->towrite &&
2055 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2056 s->syncing || s->expanding ||
2057 (s->failed >= 1 &&
2058 (sh->dev[r6s->failed_num[0]].toread ||
2059 s->to_write)) ||
2060 (s->failed >= 2 &&
2061 (sh->dev[r6s->failed_num[1]].toread ||
2062 s->to_write)))) {
2063 /* we would like to get this block, possibly
2064 * by computing it, but we might not be able to
2065 */
2066 if (s->uptodate == disks-1) {
Dan Williams45b42332007-07-09 11:56:43 -07002067 pr_debug("Computing stripe %llu block %d\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002068 (unsigned long long)sh->sector, i);
2069 compute_block_1(sh, i, 0);
2070 s->uptodate++;
2071 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2072 /* Computing 2-failure is *very* expensive; only
2073 * do it if failed >= 2
2074 */
2075 int other;
2076 for (other = disks; other--; ) {
2077 if (other == i)
2078 continue;
2079 if (!test_bit(R5_UPTODATE,
2080 &sh->dev[other].flags))
2081 break;
2082 }
2083 BUG_ON(other < 0);
Dan Williams45b42332007-07-09 11:56:43 -07002084 pr_debug("Computing stripe %llu blocks %d,%d\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002085 (unsigned long long)sh->sector,
2086 i, other);
2087 compute_block_2(sh, i, other);
2088 s->uptodate += 2;
2089 } else if (test_bit(R5_Insync, &dev->flags)) {
2090 set_bit(R5_LOCKED, &dev->flags);
2091 set_bit(R5_Wantread, &dev->flags);
2092 s->locked++;
Dan Williams45b42332007-07-09 11:56:43 -07002093 pr_debug("Reading block %d (sync=%d)\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002094 i, s->syncing);
2095 }
2096 }
2097 }
2098 set_bit(STRIPE_HANDLE, &sh->state);
2099}
2100
2101
2102/* handle_completed_write_requests
2103 * any written block on an uptodate or failed drive can be returned.
2104 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2105 * never LOCKED, so we don't need to test 'failed' directly.
2106 */
2107static void handle_completed_write_requests(raid5_conf_t *conf,
2108 struct stripe_head *sh, int disks, struct bio **return_bi)
2109{
2110 int i;
2111 struct r5dev *dev;
2112
2113 for (i = disks; i--; )
2114 if (sh->dev[i].written) {
2115 dev = &sh->dev[i];
2116 if (!test_bit(R5_LOCKED, &dev->flags) &&
2117 test_bit(R5_UPTODATE, &dev->flags)) {
2118 /* We can return any write requests */
2119 struct bio *wbi, *wbi2;
2120 int bitmap_end = 0;
Dan Williams45b42332007-07-09 11:56:43 -07002121 pr_debug("Return write for disc %d\n", i);
Dan Williamsa4456852007-07-09 11:56:43 -07002122 spin_lock_irq(&conf->device_lock);
2123 wbi = dev->written;
2124 dev->written = NULL;
2125 while (wbi && wbi->bi_sector <
2126 dev->sector + STRIPE_SECTORS) {
2127 wbi2 = r5_next_bio(wbi, dev->sector);
2128 if (--wbi->bi_phys_segments == 0) {
2129 md_write_end(conf->mddev);
2130 wbi->bi_next = *return_bi;
2131 *return_bi = wbi;
2132 }
2133 wbi = wbi2;
2134 }
2135 if (dev->towrite == NULL)
2136 bitmap_end = 1;
2137 spin_unlock_irq(&conf->device_lock);
2138 if (bitmap_end)
2139 bitmap_endwrite(conf->mddev->bitmap,
2140 sh->sector,
2141 STRIPE_SECTORS,
2142 !test_bit(STRIPE_DEGRADED, &sh->state),
2143 0);
2144 }
2145 }
2146}
2147
2148static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2149 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2150{
2151 int rmw = 0, rcw = 0, i;
2152 for (i = disks; i--; ) {
2153 /* would I have to read this buffer for read_modify_write */
2154 struct r5dev *dev = &sh->dev[i];
2155 if ((dev->towrite || i == sh->pd_idx) &&
2156 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07002157 !(test_bit(R5_UPTODATE, &dev->flags) ||
2158 test_bit(R5_Wantcompute, &dev->flags))) {
Dan Williamsa4456852007-07-09 11:56:43 -07002159 if (test_bit(R5_Insync, &dev->flags))
2160 rmw++;
2161 else
2162 rmw += 2*disks; /* cannot read it */
2163 }
2164 /* Would I have to read this buffer for reconstruct_write */
2165 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2166 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07002167 !(test_bit(R5_UPTODATE, &dev->flags) ||
2168 test_bit(R5_Wantcompute, &dev->flags))) {
2169 if (test_bit(R5_Insync, &dev->flags)) rcw++;
Dan Williamsa4456852007-07-09 11:56:43 -07002170 else
2171 rcw += 2*disks;
2172 }
2173 }
Dan Williams45b42332007-07-09 11:56:43 -07002174 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002175 (unsigned long long)sh->sector, rmw, rcw);
2176 set_bit(STRIPE_HANDLE, &sh->state);
2177 if (rmw < rcw && rmw > 0)
2178 /* prefer read-modify-write, but need to get some data */
2179 for (i = disks; i--; ) {
2180 struct r5dev *dev = &sh->dev[i];
2181 if ((dev->towrite || i == sh->pd_idx) &&
2182 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07002183 !(test_bit(R5_UPTODATE, &dev->flags) ||
2184 test_bit(R5_Wantcompute, &dev->flags)) &&
Dan Williamsa4456852007-07-09 11:56:43 -07002185 test_bit(R5_Insync, &dev->flags)) {
2186 if (
2187 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
Dan Williams45b42332007-07-09 11:56:43 -07002188 pr_debug("Read_old block "
Dan Williamsa4456852007-07-09 11:56:43 -07002189 "%d for r-m-w\n", i);
2190 set_bit(R5_LOCKED, &dev->flags);
2191 set_bit(R5_Wantread, &dev->flags);
Dan Williams830ea012007-01-02 13:52:31 -07002192 if (!test_and_set_bit(
2193 STRIPE_OP_IO, &sh->ops.pending))
2194 sh->ops.count++;
Dan Williamsa4456852007-07-09 11:56:43 -07002195 s->locked++;
2196 } else {
2197 set_bit(STRIPE_DELAYED, &sh->state);
2198 set_bit(STRIPE_HANDLE, &sh->state);
2199 }
2200 }
2201 }
2202 if (rcw <= rmw && rcw > 0)
2203 /* want reconstruct write, but need to get some data */
2204 for (i = disks; i--; ) {
2205 struct r5dev *dev = &sh->dev[i];
2206 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2207 i != sh->pd_idx &&
2208 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07002209 !(test_bit(R5_UPTODATE, &dev->flags) ||
2210 test_bit(R5_Wantcompute, &dev->flags)) &&
Dan Williamsa4456852007-07-09 11:56:43 -07002211 test_bit(R5_Insync, &dev->flags)) {
2212 if (
2213 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
Dan Williams45b42332007-07-09 11:56:43 -07002214 pr_debug("Read_old block "
Dan Williamsa4456852007-07-09 11:56:43 -07002215 "%d for Reconstruct\n", i);
2216 set_bit(R5_LOCKED, &dev->flags);
2217 set_bit(R5_Wantread, &dev->flags);
Dan Williams830ea012007-01-02 13:52:31 -07002218 if (!test_and_set_bit(
2219 STRIPE_OP_IO, &sh->ops.pending))
2220 sh->ops.count++;
Dan Williamsa4456852007-07-09 11:56:43 -07002221 s->locked++;
2222 } else {
2223 set_bit(STRIPE_DELAYED, &sh->state);
2224 set_bit(STRIPE_HANDLE, &sh->state);
2225 }
2226 }
2227 }
2228 /* now if nothing is locked, and if we have enough data,
2229 * we can start a write request
2230 */
Dan Williamsf38e1212007-01-02 13:52:30 -07002231 /* since handle_stripe can be called at any time we need to handle the
2232 * case where a compute block operation has been submitted and then a
2233 * subsequent call wants to start a write request. raid5_run_ops only
2234 * handles the case where compute block and postxor are requested
2235 * simultaneously. If this is not the case then new writes need to be
2236 * held off until the compute completes.
2237 */
2238 if ((s->req_compute ||
2239 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2240 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2241 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
Dan Williamse33129d2007-01-02 13:52:30 -07002242 s->locked += handle_write_operations5(sh, rcw == 0, 0);
Dan Williamsa4456852007-07-09 11:56:43 -07002243}
2244
2245static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2246 struct stripe_head *sh, struct stripe_head_state *s,
2247 struct r6_state *r6s, int disks)
2248{
2249 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2250 int qd_idx = r6s->qd_idx;
2251 for (i = disks; i--; ) {
2252 struct r5dev *dev = &sh->dev[i];
2253 /* Would I have to read this buffer for reconstruct_write */
2254 if (!test_bit(R5_OVERWRITE, &dev->flags)
2255 && i != pd_idx && i != qd_idx
2256 && (!test_bit(R5_LOCKED, &dev->flags)
2257 ) &&
2258 !test_bit(R5_UPTODATE, &dev->flags)) {
2259 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2260 else {
Dan Williams45b42332007-07-09 11:56:43 -07002261 pr_debug("raid6: must_compute: "
Dan Williamsa4456852007-07-09 11:56:43 -07002262 "disk %d flags=%#lx\n", i, dev->flags);
2263 must_compute++;
2264 }
2265 }
2266 }
Dan Williams45b42332007-07-09 11:56:43 -07002267 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002268 (unsigned long long)sh->sector, rcw, must_compute);
2269 set_bit(STRIPE_HANDLE, &sh->state);
2270
2271 if (rcw > 0)
2272 /* want reconstruct write, but need to get some data */
2273 for (i = disks; i--; ) {
2274 struct r5dev *dev = &sh->dev[i];
2275 if (!test_bit(R5_OVERWRITE, &dev->flags)
2276 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2277 && !test_bit(R5_LOCKED, &dev->flags) &&
2278 !test_bit(R5_UPTODATE, &dev->flags) &&
2279 test_bit(R5_Insync, &dev->flags)) {
2280 if (
2281 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
Dan Williams45b42332007-07-09 11:56:43 -07002282 pr_debug("Read_old stripe %llu "
Dan Williamsa4456852007-07-09 11:56:43 -07002283 "block %d for Reconstruct\n",
2284 (unsigned long long)sh->sector, i);
2285 set_bit(R5_LOCKED, &dev->flags);
2286 set_bit(R5_Wantread, &dev->flags);
2287 s->locked++;
2288 } else {
Dan Williams45b42332007-07-09 11:56:43 -07002289 pr_debug("Request delayed stripe %llu "
Dan Williamsa4456852007-07-09 11:56:43 -07002290 "block %d for Reconstruct\n",
2291 (unsigned long long)sh->sector, i);
2292 set_bit(STRIPE_DELAYED, &sh->state);
2293 set_bit(STRIPE_HANDLE, &sh->state);
2294 }
2295 }
2296 }
2297 /* now if nothing is locked, and if we have enough data, we can start a
2298 * write request
2299 */
2300 if (s->locked == 0 && rcw == 0 &&
2301 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2302 if (must_compute > 0) {
2303 /* We have failed blocks and need to compute them */
2304 switch (s->failed) {
2305 case 0:
2306 BUG();
2307 case 1:
2308 compute_block_1(sh, r6s->failed_num[0], 0);
2309 break;
2310 case 2:
2311 compute_block_2(sh, r6s->failed_num[0],
2312 r6s->failed_num[1]);
2313 break;
2314 default: /* This request should have been failed? */
2315 BUG();
2316 }
2317 }
2318
Dan Williams45b42332007-07-09 11:56:43 -07002319 pr_debug("Computing parity for stripe %llu\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002320 (unsigned long long)sh->sector);
2321 compute_parity6(sh, RECONSTRUCT_WRITE);
2322 /* now every locked buffer is ready to be written */
2323 for (i = disks; i--; )
2324 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
Dan Williams45b42332007-07-09 11:56:43 -07002325 pr_debug("Writing stripe %llu block %d\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002326 (unsigned long long)sh->sector, i);
2327 s->locked++;
2328 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2329 }
2330 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2331 set_bit(STRIPE_INSYNC, &sh->state);
2332
2333 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2334 atomic_dec(&conf->preread_active_stripes);
2335 if (atomic_read(&conf->preread_active_stripes) <
2336 IO_THRESHOLD)
2337 md_wakeup_thread(conf->mddev->thread);
2338 }
2339 }
2340}
2341
2342static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2343 struct stripe_head_state *s, int disks)
2344{
2345 set_bit(STRIPE_HANDLE, &sh->state);
Dan Williamse89f8962007-01-02 13:52:31 -07002346 /* Take one of the following actions:
2347 * 1/ start a check parity operation if (uptodate == disks)
2348 * 2/ finish a check parity operation and act on the result
2349 * 3/ skip to the writeback section if we previously
2350 * initiated a recovery operation
2351 */
2352 if (s->failed == 0 &&
2353 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2354 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2355 BUG_ON(s->uptodate != disks);
2356 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2357 sh->ops.count++;
2358 s->uptodate--;
2359 } else if (
2360 test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2361 clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2362 clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2363
2364 if (sh->ops.zero_sum_result == 0)
2365 /* parity is correct (on disc,
2366 * not in buffer any more)
2367 */
Dan Williamsa4456852007-07-09 11:56:43 -07002368 set_bit(STRIPE_INSYNC, &sh->state);
2369 else {
Dan Williamse89f8962007-01-02 13:52:31 -07002370 conf->mddev->resync_mismatches +=
2371 STRIPE_SECTORS;
2372 if (test_bit(
2373 MD_RECOVERY_CHECK, &conf->mddev->recovery))
2374 /* don't try to repair!! */
2375 set_bit(STRIPE_INSYNC, &sh->state);
2376 else {
2377 set_bit(STRIPE_OP_COMPUTE_BLK,
2378 &sh->ops.pending);
2379 set_bit(STRIPE_OP_MOD_REPAIR_PD,
2380 &sh->ops.pending);
2381 set_bit(R5_Wantcompute,
2382 &sh->dev[sh->pd_idx].flags);
2383 sh->ops.target = sh->pd_idx;
2384 sh->ops.count++;
2385 s->uptodate++;
2386 }
Dan Williamsa4456852007-07-09 11:56:43 -07002387 }
2388 }
2389 }
Dan Williamse89f8962007-01-02 13:52:31 -07002390
2391 /* check if we can clear a parity disk reconstruct */
2392 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2393 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2394
2395 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2396 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2397 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2398 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2399 }
2400
2401 /* Wait for check parity and compute block operations to complete
2402 * before write-back
2403 */
2404 if (!test_bit(STRIPE_INSYNC, &sh->state) &&
2405 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2406 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002407 struct r5dev *dev;
2408 /* either failed parity check, or recovery is happening */
2409 if (s->failed == 0)
2410 s->failed_num = sh->pd_idx;
2411 dev = &sh->dev[s->failed_num];
2412 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2413 BUG_ON(s->uptodate != disks);
2414
2415 set_bit(R5_LOCKED, &dev->flags);
2416 set_bit(R5_Wantwrite, &dev->flags);
Dan Williams830ea012007-01-02 13:52:31 -07002417 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2418 sh->ops.count++;
2419
Dan Williamsa4456852007-07-09 11:56:43 -07002420 clear_bit(STRIPE_DEGRADED, &sh->state);
2421 s->locked++;
2422 set_bit(STRIPE_INSYNC, &sh->state);
2423 }
2424}
2425
2426
2427static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2428 struct stripe_head_state *s,
2429 struct r6_state *r6s, struct page *tmp_page,
2430 int disks)
2431{
2432 int update_p = 0, update_q = 0;
2433 struct r5dev *dev;
2434 int pd_idx = sh->pd_idx;
2435 int qd_idx = r6s->qd_idx;
2436
2437 set_bit(STRIPE_HANDLE, &sh->state);
2438
2439 BUG_ON(s->failed > 2);
2440 BUG_ON(s->uptodate < disks);
2441 /* Want to check and possibly repair P and Q.
2442 * However there could be one 'failed' device, in which
2443 * case we can only check one of them, possibly using the
2444 * other to generate missing data
2445 */
2446
2447 /* If !tmp_page, we cannot do the calculations,
2448 * but as we have set STRIPE_HANDLE, we will soon be called
2449 * by stripe_handle with a tmp_page - just wait until then.
2450 */
2451 if (tmp_page) {
2452 if (s->failed == r6s->q_failed) {
2453 /* The only possible failed device holds 'Q', so it
2454 * makes sense to check P (If anything else were failed,
2455 * we would have used P to recreate it).
2456 */
2457 compute_block_1(sh, pd_idx, 1);
2458 if (!page_is_zero(sh->dev[pd_idx].page)) {
2459 compute_block_1(sh, pd_idx, 0);
2460 update_p = 1;
2461 }
2462 }
2463 if (!r6s->q_failed && s->failed < 2) {
2464 /* q is not failed, and we didn't use it to generate
2465 * anything, so it makes sense to check it
2466 */
2467 memcpy(page_address(tmp_page),
2468 page_address(sh->dev[qd_idx].page),
2469 STRIPE_SIZE);
2470 compute_parity6(sh, UPDATE_PARITY);
2471 if (memcmp(page_address(tmp_page),
2472 page_address(sh->dev[qd_idx].page),
2473 STRIPE_SIZE) != 0) {
2474 clear_bit(STRIPE_INSYNC, &sh->state);
2475 update_q = 1;
2476 }
2477 }
2478 if (update_p || update_q) {
2479 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2480 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2481 /* don't try to repair!! */
2482 update_p = update_q = 0;
2483 }
2484
2485 /* now write out any block on a failed drive,
2486 * or P or Q if they need it
2487 */
2488
2489 if (s->failed == 2) {
2490 dev = &sh->dev[r6s->failed_num[1]];
2491 s->locked++;
2492 set_bit(R5_LOCKED, &dev->flags);
2493 set_bit(R5_Wantwrite, &dev->flags);
2494 }
2495 if (s->failed >= 1) {
2496 dev = &sh->dev[r6s->failed_num[0]];
2497 s->locked++;
2498 set_bit(R5_LOCKED, &dev->flags);
2499 set_bit(R5_Wantwrite, &dev->flags);
2500 }
2501
2502 if (update_p) {
2503 dev = &sh->dev[pd_idx];
2504 s->locked++;
2505 set_bit(R5_LOCKED, &dev->flags);
2506 set_bit(R5_Wantwrite, &dev->flags);
2507 }
2508 if (update_q) {
2509 dev = &sh->dev[qd_idx];
2510 s->locked++;
2511 set_bit(R5_LOCKED, &dev->flags);
2512 set_bit(R5_Wantwrite, &dev->flags);
2513 }
2514 clear_bit(STRIPE_DEGRADED, &sh->state);
2515
2516 set_bit(STRIPE_INSYNC, &sh->state);
2517 }
2518}
2519
2520static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2521 struct r6_state *r6s)
2522{
2523 int i;
2524
2525 /* We have read all the blocks in this stripe and now we need to
2526 * copy some of them into a target stripe for expand.
2527 */
Dan Williamsf0a50d32007-01-02 13:52:31 -07002528 struct dma_async_tx_descriptor *tx = NULL;
Dan Williamsa4456852007-07-09 11:56:43 -07002529 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2530 for (i = 0; i < sh->disks; i++)
NeilBrowna2e08552007-09-11 15:23:36 -07002531 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002532 int dd_idx, pd_idx, j;
2533 struct stripe_head *sh2;
2534
2535 sector_t bn = compute_blocknr(sh, i);
2536 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2537 conf->raid_disks -
2538 conf->max_degraded, &dd_idx,
2539 &pd_idx, conf);
2540 sh2 = get_active_stripe(conf, s, conf->raid_disks,
2541 pd_idx, 1);
2542 if (sh2 == NULL)
2543 /* so far only the early blocks of this stripe
2544 * have been requested. When later blocks
2545 * get requested, we will try again
2546 */
2547 continue;
2548 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2549 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2550 /* must have already done this block */
2551 release_stripe(sh2);
2552 continue;
2553 }
Dan Williamsf0a50d32007-01-02 13:52:31 -07002554
2555 /* place all the copies on one channel */
2556 tx = async_memcpy(sh2->dev[dd_idx].page,
2557 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2558 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2559
Dan Williamsa4456852007-07-09 11:56:43 -07002560 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2561 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2562 for (j = 0; j < conf->raid_disks; j++)
2563 if (j != sh2->pd_idx &&
NeilBrowna2e08552007-09-11 15:23:36 -07002564 (!r6s || j != raid6_next_disk(sh2->pd_idx,
2565 sh2->disks)) &&
Dan Williamsa4456852007-07-09 11:56:43 -07002566 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2567 break;
2568 if (j == conf->raid_disks) {
2569 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2570 set_bit(STRIPE_HANDLE, &sh2->state);
2571 }
2572 release_stripe(sh2);
Dan Williamsf0a50d32007-01-02 13:52:31 -07002573
Dan Williamsa4456852007-07-09 11:56:43 -07002574 }
NeilBrowna2e08552007-09-11 15:23:36 -07002575 /* done submitting copies, wait for them to complete */
2576 if (tx) {
2577 async_tx_ack(tx);
2578 dma_wait_for_async_tx(tx);
2579 }
Dan Williamsa4456852007-07-09 11:56:43 -07002580}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581
2582/*
2583 * handle_stripe - do things to a stripe.
2584 *
2585 * We lock the stripe and then examine the state of various bits
2586 * to see what needs to be done.
2587 * Possible results:
2588 * return some read request which now have data
2589 * return some write requests which are safely on disc
2590 * schedule a read on some buffers
2591 * schedule a write of some buffers
2592 * return confirmation of parity correctness
2593 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 * buffers are taken off read_list or write_list, and bh_cache buffers
2595 * get BH_Lock set before the stripe lock is released.
2596 *
2597 */
Dan Williamsa4456852007-07-09 11:56:43 -07002598
NeilBrown16a53ec2006-06-26 00:27:38 -07002599static void handle_stripe5(struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600{
2601 raid5_conf_t *conf = sh->raid_conf;
Dan Williamsa4456852007-07-09 11:56:43 -07002602 int disks = sh->disks, i;
2603 struct bio *return_bi = NULL;
2604 struct stripe_head_state s;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 struct r5dev *dev;
Dan Williamsd84e0f12007-01-02 13:52:30 -07002606 unsigned long pending = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607
Dan Williamsa4456852007-07-09 11:56:43 -07002608 memset(&s, 0, sizeof(s));
Dan Williamsd84e0f12007-01-02 13:52:30 -07002609 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2610 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2611 atomic_read(&sh->count), sh->pd_idx,
2612 sh->ops.pending, sh->ops.ack, sh->ops.complete);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613
2614 spin_lock(&sh->lock);
2615 clear_bit(STRIPE_HANDLE, &sh->state);
2616 clear_bit(STRIPE_DELAYED, &sh->state);
2617
Dan Williamsa4456852007-07-09 11:56:43 -07002618 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2619 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2620 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 /* Now to look around and see what can be done */
2622
NeilBrown9910f162006-01-06 00:20:24 -08002623 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624 for (i=disks; i--; ) {
2625 mdk_rdev_t *rdev;
Dan Williamsa4456852007-07-09 11:56:43 -07002626 struct r5dev *dev = &sh->dev[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 clear_bit(R5_Insync, &dev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628
Dan Williamsb5e98d62007-01-02 13:52:31 -07002629 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2630 "written %p\n", i, dev->flags, dev->toread, dev->read,
2631 dev->towrite, dev->written);
2632
2633 /* maybe we can request a biofill operation
2634 *
2635 * new wantfill requests are only permitted while
2636 * STRIPE_OP_BIOFILL is clear
2637 */
2638 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2639 !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2640 set_bit(R5_Wantfill, &dev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641
2642 /* now count some things */
Dan Williamsa4456852007-07-09 11:56:43 -07002643 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2644 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
Dan Williamsf38e1212007-01-02 13:52:30 -07002645 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646
Dan Williamsb5e98d62007-01-02 13:52:31 -07002647 if (test_bit(R5_Wantfill, &dev->flags))
2648 s.to_fill++;
2649 else if (dev->toread)
Dan Williamsa4456852007-07-09 11:56:43 -07002650 s.to_read++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 if (dev->towrite) {
Dan Williamsa4456852007-07-09 11:56:43 -07002652 s.to_write++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 if (!test_bit(R5_OVERWRITE, &dev->flags))
Dan Williamsa4456852007-07-09 11:56:43 -07002654 s.non_overwrite++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 }
Dan Williamsa4456852007-07-09 11:56:43 -07002656 if (dev->written)
2657 s.written++;
NeilBrown9910f162006-01-06 00:20:24 -08002658 rdev = rcu_dereference(conf->disks[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -08002659 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
NeilBrown14f8d262006-01-06 00:20:14 -08002660 /* The ReadError flag will just be confusing now */
NeilBrown4e5314b2005-11-08 21:39:22 -08002661 clear_bit(R5_ReadError, &dev->flags);
2662 clear_bit(R5_ReWrite, &dev->flags);
2663 }
NeilBrownb2d444d2005-11-08 21:39:31 -08002664 if (!rdev || !test_bit(In_sync, &rdev->flags)
NeilBrown4e5314b2005-11-08 21:39:22 -08002665 || test_bit(R5_ReadError, &dev->flags)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002666 s.failed++;
2667 s.failed_num = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668 } else
2669 set_bit(R5_Insync, &dev->flags);
2670 }
NeilBrown9910f162006-01-06 00:20:24 -08002671 rcu_read_unlock();
Dan Williamsb5e98d62007-01-02 13:52:31 -07002672
2673 if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2674 sh->ops.count++;
2675
Dan Williams45b42332007-07-09 11:56:43 -07002676 pr_debug("locked=%d uptodate=%d to_read=%d"
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 " to_write=%d failed=%d failed_num=%d\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002678 s.locked, s.uptodate, s.to_read, s.to_write,
2679 s.failed, s.failed_num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 /* check if the array has lost two devices and, if so, some requests might
2681 * need to be failed
2682 */
Dan Williamsa4456852007-07-09 11:56:43 -07002683 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2684 handle_requests_to_failed_array(conf, sh, &s, disks,
2685 &return_bi);
2686 if (s.failed > 1 && s.syncing) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2688 clear_bit(STRIPE_SYNCING, &sh->state);
Dan Williamsa4456852007-07-09 11:56:43 -07002689 s.syncing = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 }
2691
2692 /* might be able to return some write requests if the parity block
2693 * is safe, or on a failed drive
2694 */
2695 dev = &sh->dev[sh->pd_idx];
Dan Williamsa4456852007-07-09 11:56:43 -07002696 if ( s.written &&
2697 ((test_bit(R5_Insync, &dev->flags) &&
2698 !test_bit(R5_LOCKED, &dev->flags) &&
2699 test_bit(R5_UPTODATE, &dev->flags)) ||
2700 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2701 handle_completed_write_requests(conf, sh, disks, &return_bi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702
2703 /* Now we might consider reading some blocks, either to check/generate
2704 * parity, or to satisfy requests
2705 * or to load a block that is being partially written.
2706 */
Dan Williamsa4456852007-07-09 11:56:43 -07002707 if (s.to_read || s.non_overwrite ||
Dan Williamsf38e1212007-01-02 13:52:30 -07002708 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2709 test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
Dan Williamsa4456852007-07-09 11:56:43 -07002710 handle_issuing_new_read_requests5(sh, &s, disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711
Dan Williamse33129d2007-01-02 13:52:30 -07002712 /* Now we check to see if any write operations have recently
2713 * completed
2714 */
2715
2716 /* leave prexor set until postxor is done, allows us to distinguish
2717 * a rmw from a rcw during biodrain
2718 */
2719 if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2720 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2721
2722 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2723 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2724 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2725
2726 for (i = disks; i--; )
2727 clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2728 }
2729
2730 /* if only POSTXOR is set then this is an 'expand' postxor */
2731 if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2732 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2733
2734 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2735 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2736 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2737
2738 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2739 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2740 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2741
2742 /* All the 'written' buffers and the parity block are ready to
2743 * be written back to disk
2744 */
2745 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2746 for (i = disks; i--; ) {
2747 dev = &sh->dev[i];
2748 if (test_bit(R5_LOCKED, &dev->flags) &&
2749 (i == sh->pd_idx || dev->written)) {
2750 pr_debug("Writing block %d\n", i);
2751 set_bit(R5_Wantwrite, &dev->flags);
2752 if (!test_and_set_bit(
2753 STRIPE_OP_IO, &sh->ops.pending))
2754 sh->ops.count++;
2755 if (!test_bit(R5_Insync, &dev->flags) ||
2756 (i == sh->pd_idx && s.failed == 0))
2757 set_bit(STRIPE_INSYNC, &sh->state);
2758 }
2759 }
2760 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2761 atomic_dec(&conf->preread_active_stripes);
2762 if (atomic_read(&conf->preread_active_stripes) <
2763 IO_THRESHOLD)
2764 md_wakeup_thread(conf->mddev->thread);
2765 }
2766 }
2767
2768 /* Now to consider new write requests and what else, if anything
2769 * should be read. We do not handle new writes when:
2770 * 1/ A 'write' operation (copy+xor) is already in flight.
2771 * 2/ A 'check' operation is in flight, as it may clobber the parity
2772 * block.
2773 */
2774 if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2775 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
Dan Williamsa4456852007-07-09 11:56:43 -07002776 handle_issuing_new_write_requests5(conf, sh, &s, disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777
2778 /* maybe we need to check and possibly fix the parity for this stripe
Dan Williamse89f8962007-01-02 13:52:31 -07002779 * Any reads will already have been scheduled, so we just see if enough
2780 * data is available. The parity check is held off while parity
2781 * dependent operations are in flight.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 */
Dan Williamse89f8962007-01-02 13:52:31 -07002783 if ((s.syncing && s.locked == 0 &&
2784 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2785 !test_bit(STRIPE_INSYNC, &sh->state)) ||
2786 test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2787 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
Dan Williamsa4456852007-07-09 11:56:43 -07002788 handle_parity_checks5(conf, sh, &s, disks);
Dan Williamse89f8962007-01-02 13:52:31 -07002789
Dan Williamsa4456852007-07-09 11:56:43 -07002790 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2792 clear_bit(STRIPE_SYNCING, &sh->state);
2793 }
NeilBrown4e5314b2005-11-08 21:39:22 -08002794
2795 /* If the failed drive is just a ReadError, then we might need to progress
2796 * the repair/check process
2797 */
Dan Williamsa4456852007-07-09 11:56:43 -07002798 if (s.failed == 1 && !conf->mddev->ro &&
2799 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2800 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2801 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
NeilBrown4e5314b2005-11-08 21:39:22 -08002802 ) {
Dan Williamsa4456852007-07-09 11:56:43 -07002803 dev = &sh->dev[s.failed_num];
NeilBrown4e5314b2005-11-08 21:39:22 -08002804 if (!test_bit(R5_ReWrite, &dev->flags)) {
2805 set_bit(R5_Wantwrite, &dev->flags);
Dan Williams830ea012007-01-02 13:52:31 -07002806 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2807 sh->ops.count++;
NeilBrown4e5314b2005-11-08 21:39:22 -08002808 set_bit(R5_ReWrite, &dev->flags);
2809 set_bit(R5_LOCKED, &dev->flags);
Dan Williamsa4456852007-07-09 11:56:43 -07002810 s.locked++;
NeilBrown4e5314b2005-11-08 21:39:22 -08002811 } else {
2812 /* let's read it back */
2813 set_bit(R5_Wantread, &dev->flags);
Dan Williams830ea012007-01-02 13:52:31 -07002814 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2815 sh->ops.count++;
NeilBrown4e5314b2005-11-08 21:39:22 -08002816 set_bit(R5_LOCKED, &dev->flags);
Dan Williamsa4456852007-07-09 11:56:43 -07002817 s.locked++;
NeilBrown4e5314b2005-11-08 21:39:22 -08002818 }
2819 }
2820
Dan Williamsf0a50d32007-01-02 13:52:31 -07002821 /* Finish postxor operations initiated by the expansion
2822 * process
2823 */
2824 if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2825 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2826
2827 clear_bit(STRIPE_EXPANDING, &sh->state);
2828
2829 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2830 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2831 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2832
2833 for (i = conf->raid_disks; i--; ) {
2834 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2835 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2836 sh->ops.count++;
2837 }
2838 }
2839
2840 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2841 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
NeilBrownccfcc3c2006-03-27 01:18:09 -08002842 /* Need to write out all blocks after computing parity */
2843 sh->disks = conf->raid_disks;
Dan Williamsf0a50d32007-01-02 13:52:31 -07002844 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2845 conf->raid_disks);
NeilBrowna2e08552007-09-11 15:23:36 -07002846 s.locked += handle_write_operations5(sh, 1, 1);
Dan Williamsf0a50d32007-01-02 13:52:31 -07002847 } else if (s.expanded &&
2848 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
NeilBrownccfcc3c2006-03-27 01:18:09 -08002849 clear_bit(STRIPE_EXPAND_READY, &sh->state);
NeilBrownf6705572006-03-27 01:18:11 -08002850 atomic_dec(&conf->reshape_stripes);
NeilBrownccfcc3c2006-03-27 01:18:09 -08002851 wake_up(&conf->wait_for_overlap);
2852 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2853 }
2854
Dan Williamsa4456852007-07-09 11:56:43 -07002855 if (s.expanding && s.locked == 0)
2856 handle_stripe_expansion(conf, sh, NULL);
NeilBrownccfcc3c2006-03-27 01:18:09 -08002857
Dan Williamsd84e0f12007-01-02 13:52:30 -07002858 if (sh->ops.count)
2859 pending = get_stripe_work(sh);
2860
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 spin_unlock(&sh->lock);
2862
Dan Williamsd84e0f12007-01-02 13:52:30 -07002863 if (pending)
2864 raid5_run_ops(sh, pending);
2865
Dan Williamsa4456852007-07-09 11:56:43 -07002866 return_io(return_bi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868}
2869
NeilBrown16a53ec2006-06-26 00:27:38 -07002870static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2871{
2872 raid6_conf_t *conf = sh->raid_conf;
NeilBrownf4168852007-02-28 20:11:53 -08002873 int disks = sh->disks;
Dan Williamsa4456852007-07-09 11:56:43 -07002874 struct bio *return_bi = NULL;
2875 int i, pd_idx = sh->pd_idx;
2876 struct stripe_head_state s;
2877 struct r6_state r6s;
NeilBrown16a53ec2006-06-26 00:27:38 -07002878 struct r5dev *dev, *pdev, *qdev;
NeilBrown16a53ec2006-06-26 00:27:38 -07002879
Dan Williamsa4456852007-07-09 11:56:43 -07002880 r6s.qd_idx = raid6_next_disk(pd_idx, disks);
Dan Williams45b42332007-07-09 11:56:43 -07002881 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
Dan Williamsa4456852007-07-09 11:56:43 -07002882 "pd_idx=%d, qd_idx=%d\n",
2883 (unsigned long long)sh->sector, sh->state,
2884 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2885 memset(&s, 0, sizeof(s));
NeilBrown16a53ec2006-06-26 00:27:38 -07002886
2887 spin_lock(&sh->lock);
2888 clear_bit(STRIPE_HANDLE, &sh->state);
2889 clear_bit(STRIPE_DELAYED, &sh->state);
2890
Dan Williamsa4456852007-07-09 11:56:43 -07002891 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2892 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2893 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
NeilBrown16a53ec2006-06-26 00:27:38 -07002894 /* Now to look around and see what can be done */
2895
2896 rcu_read_lock();
2897 for (i=disks; i--; ) {
2898 mdk_rdev_t *rdev;
2899 dev = &sh->dev[i];
2900 clear_bit(R5_Insync, &dev->flags);
2901
Dan Williams45b42332007-07-09 11:56:43 -07002902 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
NeilBrown16a53ec2006-06-26 00:27:38 -07002903 i, dev->flags, dev->toread, dev->towrite, dev->written);
2904 /* maybe we can reply to a read */
2905 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2906 struct bio *rbi, *rbi2;
Dan Williams45b42332007-07-09 11:56:43 -07002907 pr_debug("Return read for disc %d\n", i);
NeilBrown16a53ec2006-06-26 00:27:38 -07002908 spin_lock_irq(&conf->device_lock);
2909 rbi = dev->toread;
2910 dev->toread = NULL;
2911 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2912 wake_up(&conf->wait_for_overlap);
2913 spin_unlock_irq(&conf->device_lock);
2914 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2915 copy_data(0, rbi, dev->page, dev->sector);
2916 rbi2 = r5_next_bio(rbi, dev->sector);
2917 spin_lock_irq(&conf->device_lock);
2918 if (--rbi->bi_phys_segments == 0) {
2919 rbi->bi_next = return_bi;
2920 return_bi = rbi;
2921 }
2922 spin_unlock_irq(&conf->device_lock);
2923 rbi = rbi2;
2924 }
2925 }
2926
2927 /* now count some things */
Dan Williamsa4456852007-07-09 11:56:43 -07002928 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2929 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
NeilBrown16a53ec2006-06-26 00:27:38 -07002930
2931
Dan Williamsa4456852007-07-09 11:56:43 -07002932 if (dev->toread)
2933 s.to_read++;
NeilBrown16a53ec2006-06-26 00:27:38 -07002934 if (dev->towrite) {
Dan Williamsa4456852007-07-09 11:56:43 -07002935 s.to_write++;
NeilBrown16a53ec2006-06-26 00:27:38 -07002936 if (!test_bit(R5_OVERWRITE, &dev->flags))
Dan Williamsa4456852007-07-09 11:56:43 -07002937 s.non_overwrite++;
NeilBrown16a53ec2006-06-26 00:27:38 -07002938 }
Dan Williamsa4456852007-07-09 11:56:43 -07002939 if (dev->written)
2940 s.written++;
NeilBrown16a53ec2006-06-26 00:27:38 -07002941 rdev = rcu_dereference(conf->disks[i].rdev);
2942 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2943 /* The ReadError flag will just be confusing now */
2944 clear_bit(R5_ReadError, &dev->flags);
2945 clear_bit(R5_ReWrite, &dev->flags);
2946 }
2947 if (!rdev || !test_bit(In_sync, &rdev->flags)
2948 || test_bit(R5_ReadError, &dev->flags)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002949 if (s.failed < 2)
2950 r6s.failed_num[s.failed] = i;
2951 s.failed++;
NeilBrown16a53ec2006-06-26 00:27:38 -07002952 } else
2953 set_bit(R5_Insync, &dev->flags);
2954 }
2955 rcu_read_unlock();
Dan Williams45b42332007-07-09 11:56:43 -07002956 pr_debug("locked=%d uptodate=%d to_read=%d"
NeilBrown16a53ec2006-06-26 00:27:38 -07002957 " to_write=%d failed=%d failed_num=%d,%d\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002958 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2959 r6s.failed_num[0], r6s.failed_num[1]);
2960 /* check if the array has lost >2 devices and, if so, some requests
2961 * might need to be failed
NeilBrown16a53ec2006-06-26 00:27:38 -07002962 */
Dan Williamsa4456852007-07-09 11:56:43 -07002963 if (s.failed > 2 && s.to_read+s.to_write+s.written)
2964 handle_requests_to_failed_array(conf, sh, &s, disks,
2965 &return_bi);
2966 if (s.failed > 2 && s.syncing) {
NeilBrown16a53ec2006-06-26 00:27:38 -07002967 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2968 clear_bit(STRIPE_SYNCING, &sh->state);
Dan Williamsa4456852007-07-09 11:56:43 -07002969 s.syncing = 0;
NeilBrown16a53ec2006-06-26 00:27:38 -07002970 }
2971
2972 /*
2973 * might be able to return some write requests if the parity blocks
2974 * are safe, or on a failed drive
2975 */
2976 pdev = &sh->dev[pd_idx];
Dan Williamsa4456852007-07-09 11:56:43 -07002977 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2978 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2979 qdev = &sh->dev[r6s.qd_idx];
2980 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2981 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
NeilBrown16a53ec2006-06-26 00:27:38 -07002982
Dan Williamsa4456852007-07-09 11:56:43 -07002983 if ( s.written &&
2984 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
NeilBrown16a53ec2006-06-26 00:27:38 -07002985 && !test_bit(R5_LOCKED, &pdev->flags)
Dan Williamsa4456852007-07-09 11:56:43 -07002986 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
2987 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
NeilBrown16a53ec2006-06-26 00:27:38 -07002988 && !test_bit(R5_LOCKED, &qdev->flags)
Dan Williamsa4456852007-07-09 11:56:43 -07002989 && test_bit(R5_UPTODATE, &qdev->flags)))))
2990 handle_completed_write_requests(conf, sh, disks, &return_bi);
NeilBrown16a53ec2006-06-26 00:27:38 -07002991
2992 /* Now we might consider reading some blocks, either to check/generate
2993 * parity, or to satisfy requests
2994 * or to load a block that is being partially written.
2995 */
Dan Williamsa4456852007-07-09 11:56:43 -07002996 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
2997 (s.syncing && (s.uptodate < disks)) || s.expanding)
2998 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
NeilBrown16a53ec2006-06-26 00:27:38 -07002999
3000 /* now to consider writing and what else, if anything should be read */
Dan Williamsa4456852007-07-09 11:56:43 -07003001 if (s.to_write)
3002 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
NeilBrown16a53ec2006-06-26 00:27:38 -07003003
3004 /* maybe we need to check and possibly fix the parity for this stripe
Dan Williamsa4456852007-07-09 11:56:43 -07003005 * Any reads will already have been scheduled, so we just see if enough
3006 * data is available
NeilBrown16a53ec2006-06-26 00:27:38 -07003007 */
Dan Williamsa4456852007-07-09 11:56:43 -07003008 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3009 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
NeilBrown16a53ec2006-06-26 00:27:38 -07003010
Dan Williamsa4456852007-07-09 11:56:43 -07003011 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
NeilBrown16a53ec2006-06-26 00:27:38 -07003012 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3013 clear_bit(STRIPE_SYNCING, &sh->state);
3014 }
3015
3016 /* If the failed drives are just a ReadError, then we might need
3017 * to progress the repair/check process
3018 */
Dan Williamsa4456852007-07-09 11:56:43 -07003019 if (s.failed <= 2 && !conf->mddev->ro)
3020 for (i = 0; i < s.failed; i++) {
3021 dev = &sh->dev[r6s.failed_num[i]];
NeilBrown16a53ec2006-06-26 00:27:38 -07003022 if (test_bit(R5_ReadError, &dev->flags)
3023 && !test_bit(R5_LOCKED, &dev->flags)
3024 && test_bit(R5_UPTODATE, &dev->flags)
3025 ) {
3026 if (!test_bit(R5_ReWrite, &dev->flags)) {
3027 set_bit(R5_Wantwrite, &dev->flags);
3028 set_bit(R5_ReWrite, &dev->flags);
3029 set_bit(R5_LOCKED, &dev->flags);
3030 } else {
3031 /* let's read it back */
3032 set_bit(R5_Wantread, &dev->flags);
3033 set_bit(R5_LOCKED, &dev->flags);
3034 }
3035 }
3036 }
NeilBrownf4168852007-02-28 20:11:53 -08003037
Dan Williamsa4456852007-07-09 11:56:43 -07003038 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
NeilBrownf4168852007-02-28 20:11:53 -08003039 /* Need to write out all blocks after computing P&Q */
3040 sh->disks = conf->raid_disks;
3041 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3042 conf->raid_disks);
3043 compute_parity6(sh, RECONSTRUCT_WRITE);
3044 for (i = conf->raid_disks ; i-- ; ) {
3045 set_bit(R5_LOCKED, &sh->dev[i].flags);
Dan Williamsa4456852007-07-09 11:56:43 -07003046 s.locked++;
NeilBrownf4168852007-02-28 20:11:53 -08003047 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3048 }
3049 clear_bit(STRIPE_EXPANDING, &sh->state);
Dan Williamsa4456852007-07-09 11:56:43 -07003050 } else if (s.expanded) {
NeilBrownf4168852007-02-28 20:11:53 -08003051 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3052 atomic_dec(&conf->reshape_stripes);
3053 wake_up(&conf->wait_for_overlap);
3054 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3055 }
3056
Dan Williamsa4456852007-07-09 11:56:43 -07003057 if (s.expanding && s.locked == 0)
3058 handle_stripe_expansion(conf, sh, &r6s);
NeilBrownf4168852007-02-28 20:11:53 -08003059
NeilBrown16a53ec2006-06-26 00:27:38 -07003060 spin_unlock(&sh->lock);
3061
Dan Williamsa4456852007-07-09 11:56:43 -07003062 return_io(return_bi);
NeilBrown16a53ec2006-06-26 00:27:38 -07003063
NeilBrown16a53ec2006-06-26 00:27:38 -07003064 for (i=disks; i-- ;) {
3065 int rw;
3066 struct bio *bi;
3067 mdk_rdev_t *rdev;
3068 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
NeilBrown802ba062006-12-13 00:34:13 -08003069 rw = WRITE;
NeilBrown16a53ec2006-06-26 00:27:38 -07003070 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
NeilBrown802ba062006-12-13 00:34:13 -08003071 rw = READ;
NeilBrown16a53ec2006-06-26 00:27:38 -07003072 else
3073 continue;
3074
3075 bi = &sh->dev[i].req;
3076
3077 bi->bi_rw = rw;
NeilBrown802ba062006-12-13 00:34:13 -08003078 if (rw == WRITE)
NeilBrown16a53ec2006-06-26 00:27:38 -07003079 bi->bi_end_io = raid5_end_write_request;
3080 else
3081 bi->bi_end_io = raid5_end_read_request;
3082
3083 rcu_read_lock();
3084 rdev = rcu_dereference(conf->disks[i].rdev);
3085 if (rdev && test_bit(Faulty, &rdev->flags))
3086 rdev = NULL;
3087 if (rdev)
3088 atomic_inc(&rdev->nr_pending);
3089 rcu_read_unlock();
3090
3091 if (rdev) {
Dan Williamsa4456852007-07-09 11:56:43 -07003092 if (s.syncing || s.expanding || s.expanded)
NeilBrown16a53ec2006-06-26 00:27:38 -07003093 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3094
3095 bi->bi_bdev = rdev->bdev;
Dan Williams45b42332007-07-09 11:56:43 -07003096 pr_debug("for %llu schedule op %ld on disc %d\n",
NeilBrown16a53ec2006-06-26 00:27:38 -07003097 (unsigned long long)sh->sector, bi->bi_rw, i);
3098 atomic_inc(&sh->count);
3099 bi->bi_sector = sh->sector + rdev->data_offset;
3100 bi->bi_flags = 1 << BIO_UPTODATE;
3101 bi->bi_vcnt = 1;
3102 bi->bi_max_vecs = 1;
3103 bi->bi_idx = 0;
3104 bi->bi_io_vec = &sh->dev[i].vec;
3105 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3106 bi->bi_io_vec[0].bv_offset = 0;
3107 bi->bi_size = STRIPE_SIZE;
3108 bi->bi_next = NULL;
3109 if (rw == WRITE &&
3110 test_bit(R5_ReWrite, &sh->dev[i].flags))
3111 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3112 generic_make_request(bi);
3113 } else {
NeilBrown802ba062006-12-13 00:34:13 -08003114 if (rw == WRITE)
NeilBrown16a53ec2006-06-26 00:27:38 -07003115 set_bit(STRIPE_DEGRADED, &sh->state);
Dan Williams45b42332007-07-09 11:56:43 -07003116 pr_debug("skip op %ld on disc %d for sector %llu\n",
NeilBrown16a53ec2006-06-26 00:27:38 -07003117 bi->bi_rw, i, (unsigned long long)sh->sector);
3118 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3119 set_bit(STRIPE_HANDLE, &sh->state);
3120 }
3121 }
3122}
3123
3124static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3125{
3126 if (sh->raid_conf->level == 6)
3127 handle_stripe6(sh, tmp_page);
3128 else
3129 handle_stripe5(sh);
3130}
3131
3132
3133
Arjan van de Ven858119e2006-01-14 13:20:43 -08003134static void raid5_activate_delayed(raid5_conf_t *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003135{
3136 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3137 while (!list_empty(&conf->delayed_list)) {
3138 struct list_head *l = conf->delayed_list.next;
3139 struct stripe_head *sh;
3140 sh = list_entry(l, struct stripe_head, lru);
3141 list_del_init(l);
3142 clear_bit(STRIPE_DELAYED, &sh->state);
3143 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3144 atomic_inc(&conf->preread_active_stripes);
3145 list_add_tail(&sh->lru, &conf->handle_list);
3146 }
3147 }
3148}
3149
Arjan van de Ven858119e2006-01-14 13:20:43 -08003150static void activate_bit_delay(raid5_conf_t *conf)
NeilBrown72626682005-09-09 16:23:54 -07003151{
3152 /* device_lock is held */
3153 struct list_head head;
3154 list_add(&head, &conf->bitmap_list);
3155 list_del_init(&conf->bitmap_list);
3156 while (!list_empty(&head)) {
3157 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3158 list_del_init(&sh->lru);
3159 atomic_inc(&sh->count);
3160 __release_stripe(conf, sh);
3161 }
3162}
3163
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164static void unplug_slaves(mddev_t *mddev)
3165{
3166 raid5_conf_t *conf = mddev_to_conf(mddev);
3167 int i;
3168
3169 rcu_read_lock();
3170 for (i=0; i<mddev->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -08003171 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -08003172 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
Jens Axboe165125e2007-07-24 09:28:11 +02003173 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174
3175 atomic_inc(&rdev->nr_pending);
3176 rcu_read_unlock();
3177
3178 if (r_queue->unplug_fn)
3179 r_queue->unplug_fn(r_queue);
3180
3181 rdev_dec_pending(rdev, mddev);
3182 rcu_read_lock();
3183 }
3184 }
3185 rcu_read_unlock();
3186}
3187
Jens Axboe165125e2007-07-24 09:28:11 +02003188static void raid5_unplug_device(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189{
3190 mddev_t *mddev = q->queuedata;
3191 raid5_conf_t *conf = mddev_to_conf(mddev);
3192 unsigned long flags;
3193
3194 spin_lock_irqsave(&conf->device_lock, flags);
3195
NeilBrown72626682005-09-09 16:23:54 -07003196 if (blk_remove_plug(q)) {
3197 conf->seq_flush++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198 raid5_activate_delayed(conf);
NeilBrown72626682005-09-09 16:23:54 -07003199 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200 md_wakeup_thread(mddev->thread);
3201
3202 spin_unlock_irqrestore(&conf->device_lock, flags);
3203
3204 unplug_slaves(mddev);
3205}
3206
Jens Axboe165125e2007-07-24 09:28:11 +02003207static int raid5_issue_flush(struct request_queue *q, struct gendisk *disk,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003208 sector_t *error_sector)
3209{
3210 mddev_t *mddev = q->queuedata;
3211 raid5_conf_t *conf = mddev_to_conf(mddev);
3212 int i, ret = 0;
3213
3214 rcu_read_lock();
3215 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -08003216 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -08003217 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 struct block_device *bdev = rdev->bdev;
Jens Axboe165125e2007-07-24 09:28:11 +02003219 struct request_queue *r_queue = bdev_get_queue(bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220
3221 if (!r_queue->issue_flush_fn)
3222 ret = -EOPNOTSUPP;
3223 else {
3224 atomic_inc(&rdev->nr_pending);
3225 rcu_read_unlock();
3226 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
3227 error_sector);
3228 rdev_dec_pending(rdev, mddev);
3229 rcu_read_lock();
3230 }
3231 }
3232 }
3233 rcu_read_unlock();
3234 return ret;
3235}
3236
NeilBrownf022b2f2006-10-03 01:15:56 -07003237static int raid5_congested(void *data, int bits)
3238{
3239 mddev_t *mddev = data;
3240 raid5_conf_t *conf = mddev_to_conf(mddev);
3241
3242 /* No difference between reads and writes. Just check
3243 * how busy the stripe_cache is
3244 */
3245 if (conf->inactive_blocked)
3246 return 1;
3247 if (conf->quiesce)
3248 return 1;
3249 if (list_empty_careful(&conf->inactive_list))
3250 return 1;
3251
3252 return 0;
3253}
3254
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08003255/* We want read requests to align with chunks where possible,
3256 * but write requests don't need to.
3257 */
Jens Axboe165125e2007-07-24 09:28:11 +02003258static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08003259{
3260 mddev_t *mddev = q->queuedata;
3261 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3262 int max;
3263 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3264 unsigned int bio_sectors = bio->bi_size >> 9;
3265
NeilBrown802ba062006-12-13 00:34:13 -08003266 if (bio_data_dir(bio) == WRITE)
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08003267 return biovec->bv_len; /* always allow writes to be mergeable */
3268
3269 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3270 if (max < 0) max = 0;
3271 if (max <= biovec->bv_len && bio_sectors == 0)
3272 return biovec->bv_len;
3273 else
3274 return max;
3275}
3276
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003277
3278static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3279{
3280 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3281 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3282 unsigned int bio_sectors = bio->bi_size >> 9;
3283
3284 return chunk_sectors >=
3285 ((sector & (chunk_sectors - 1)) + bio_sectors);
3286}
3287
3288/*
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003289 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3290 * later sampled by raid5d.
3291 */
3292static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3293{
3294 unsigned long flags;
3295
3296 spin_lock_irqsave(&conf->device_lock, flags);
3297
3298 bi->bi_next = conf->retry_read_aligned_list;
3299 conf->retry_read_aligned_list = bi;
3300
3301 spin_unlock_irqrestore(&conf->device_lock, flags);
3302 md_wakeup_thread(conf->mddev->thread);
3303}
3304
3305
3306static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3307{
3308 struct bio *bi;
3309
3310 bi = conf->retry_read_aligned;
3311 if (bi) {
3312 conf->retry_read_aligned = NULL;
3313 return bi;
3314 }
3315 bi = conf->retry_read_aligned_list;
3316 if(bi) {
Neil Brown387bb172007-02-08 14:20:29 -08003317 conf->retry_read_aligned_list = bi->bi_next;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003318 bi->bi_next = NULL;
3319 bi->bi_phys_segments = 1; /* biased count of active stripes */
3320 bi->bi_hw_segments = 0; /* count of processed stripes */
3321 }
3322
3323 return bi;
3324}
3325
3326
3327/*
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003328 * The "raid5_align_endio" should check if the read succeeded and if it
3329 * did, call bio_endio on the original bio (having bio_put the new bio
3330 * first).
3331 * If the read failed..
3332 */
NeilBrown6712ecf2007-09-27 12:47:43 +02003333static void raid5_align_endio(struct bio *bi, int error)
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003334{
3335 struct bio* raid_bi = bi->bi_private;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003336 mddev_t *mddev;
3337 raid5_conf_t *conf;
3338 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3339 mdk_rdev_t *rdev;
3340
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003341 bio_put(bi);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003342
3343 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3344 conf = mddev_to_conf(mddev);
3345 rdev = (void*)raid_bi->bi_next;
3346 raid_bi->bi_next = NULL;
3347
3348 rdev_dec_pending(rdev, conf->mddev);
3349
3350 if (!error && uptodate) {
NeilBrown6712ecf2007-09-27 12:47:43 +02003351 bio_endio(raid_bi, 0);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003352 if (atomic_dec_and_test(&conf->active_aligned_reads))
3353 wake_up(&conf->wait_for_stripe);
NeilBrown6712ecf2007-09-27 12:47:43 +02003354 return;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003355 }
3356
3357
Dan Williams45b42332007-07-09 11:56:43 -07003358 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003359
3360 add_bio_to_retry(raid_bi, conf);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003361}
3362
Neil Brown387bb172007-02-08 14:20:29 -08003363static int bio_fits_rdev(struct bio *bi)
3364{
Jens Axboe165125e2007-07-24 09:28:11 +02003365 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
Neil Brown387bb172007-02-08 14:20:29 -08003366
3367 if ((bi->bi_size>>9) > q->max_sectors)
3368 return 0;
3369 blk_recount_segments(q, bi);
3370 if (bi->bi_phys_segments > q->max_phys_segments ||
3371 bi->bi_hw_segments > q->max_hw_segments)
3372 return 0;
3373
3374 if (q->merge_bvec_fn)
3375 /* it's too hard to apply the merge_bvec_fn at this stage,
3376 * just just give up
3377 */
3378 return 0;
3379
3380 return 1;
3381}
3382
3383
Jens Axboe165125e2007-07-24 09:28:11 +02003384static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003385{
3386 mddev_t *mddev = q->queuedata;
3387 raid5_conf_t *conf = mddev_to_conf(mddev);
3388 const unsigned int raid_disks = conf->raid_disks;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003389 const unsigned int data_disks = raid_disks - conf->max_degraded;
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003390 unsigned int dd_idx, pd_idx;
3391 struct bio* align_bi;
3392 mdk_rdev_t *rdev;
3393
3394 if (!in_chunk_boundary(mddev, raid_bio)) {
Dan Williams45b42332007-07-09 11:56:43 -07003395 pr_debug("chunk_aligned_read : non aligned\n");
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003396 return 0;
3397 }
3398 /*
3399 * use bio_clone to make a copy of the bio
3400 */
3401 align_bi = bio_clone(raid_bio, GFP_NOIO);
3402 if (!align_bi)
3403 return 0;
3404 /*
3405 * set bi_end_io to a new function, and set bi_private to the
3406 * original bio.
3407 */
3408 align_bi->bi_end_io = raid5_align_endio;
3409 align_bi->bi_private = raid_bio;
3410 /*
3411 * compute position
3412 */
3413 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
3414 raid_disks,
3415 data_disks,
3416 &dd_idx,
3417 &pd_idx,
3418 conf);
3419
3420 rcu_read_lock();
3421 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3422 if (rdev && test_bit(In_sync, &rdev->flags)) {
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003423 atomic_inc(&rdev->nr_pending);
3424 rcu_read_unlock();
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003425 raid_bio->bi_next = (void*)rdev;
3426 align_bi->bi_bdev = rdev->bdev;
3427 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3428 align_bi->bi_sector += rdev->data_offset;
3429
Neil Brown387bb172007-02-08 14:20:29 -08003430 if (!bio_fits_rdev(align_bi)) {
3431 /* too big in some way */
3432 bio_put(align_bi);
3433 rdev_dec_pending(rdev, mddev);
3434 return 0;
3435 }
3436
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003437 spin_lock_irq(&conf->device_lock);
3438 wait_event_lock_irq(conf->wait_for_stripe,
3439 conf->quiesce == 0,
3440 conf->device_lock, /* nothing */);
3441 atomic_inc(&conf->active_aligned_reads);
3442 spin_unlock_irq(&conf->device_lock);
3443
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003444 generic_make_request(align_bi);
3445 return 1;
3446 } else {
3447 rcu_read_unlock();
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003448 bio_put(align_bi);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003449 return 0;
3450 }
3451}
3452
3453
Jens Axboe165125e2007-07-24 09:28:11 +02003454static int make_request(struct request_queue *q, struct bio * bi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455{
3456 mddev_t *mddev = q->queuedata;
3457 raid5_conf_t *conf = mddev_to_conf(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458 unsigned int dd_idx, pd_idx;
3459 sector_t new_sector;
3460 sector_t logical_sector, last_sector;
3461 struct stripe_head *sh;
Jens Axboea3623572005-11-01 09:26:16 +01003462 const int rw = bio_data_dir(bi);
NeilBrownf6344752006-03-27 01:18:17 -08003463 int remaining;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464
NeilBrowne5dcdd82005-09-09 16:23:41 -07003465 if (unlikely(bio_barrier(bi))) {
NeilBrown6712ecf2007-09-27 12:47:43 +02003466 bio_endio(bi, -EOPNOTSUPP);
NeilBrowne5dcdd82005-09-09 16:23:41 -07003467 return 0;
3468 }
3469
NeilBrown3d310eb2005-06-21 17:17:26 -07003470 md_write_start(mddev, bi);
NeilBrown06d91a52005-06-21 17:17:12 -07003471
Jens Axboea3623572005-11-01 09:26:16 +01003472 disk_stat_inc(mddev->gendisk, ios[rw]);
3473 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474
NeilBrown802ba062006-12-13 00:34:13 -08003475 if (rw == READ &&
Raz Ben-Jehuda(caro)52488612006-12-10 02:20:48 -08003476 mddev->reshape_position == MaxSector &&
3477 chunk_aligned_read(q,bi))
3478 return 0;
3479
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3481 last_sector = bi->bi_sector + (bi->bi_size>>9);
3482 bi->bi_next = NULL;
3483 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
NeilBrown06d91a52005-06-21 17:17:12 -07003484
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3486 DEFINE_WAIT(w);
NeilBrown16a53ec2006-06-26 00:27:38 -07003487 int disks, data_disks;
NeilBrownb578d552006-03-27 01:18:12 -08003488
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003489 retry:
NeilBrownb578d552006-03-27 01:18:12 -08003490 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003491 if (likely(conf->expand_progress == MaxSector))
3492 disks = conf->raid_disks;
3493 else {
NeilBrowndf8e7f72006-03-27 01:18:15 -08003494 /* spinlock is needed as expand_progress may be
3495 * 64bit on a 32bit platform, and so it might be
3496 * possible to see a half-updated value
3497 * Ofcourse expand_progress could change after
3498 * the lock is dropped, so once we get a reference
3499 * to the stripe that we think it is, we will have
3500 * to check again.
3501 */
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003502 spin_lock_irq(&conf->device_lock);
3503 disks = conf->raid_disks;
3504 if (logical_sector >= conf->expand_progress)
3505 disks = conf->previous_raid_disks;
NeilBrownb578d552006-03-27 01:18:12 -08003506 else {
3507 if (logical_sector >= conf->expand_lo) {
3508 spin_unlock_irq(&conf->device_lock);
3509 schedule();
3510 goto retry;
3511 }
3512 }
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003513 spin_unlock_irq(&conf->device_lock);
3514 }
NeilBrown16a53ec2006-06-26 00:27:38 -07003515 data_disks = disks - conf->max_degraded;
3516
3517 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003518 &dd_idx, &pd_idx, conf);
Dan Williams45b42332007-07-09 11:56:43 -07003519 pr_debug("raid5: make_request, sector %llu logical %llu\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 (unsigned long long)new_sector,
3521 (unsigned long long)logical_sector);
3522
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003523 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524 if (sh) {
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003525 if (unlikely(conf->expand_progress != MaxSector)) {
3526 /* expansion might have moved on while waiting for a
NeilBrowndf8e7f72006-03-27 01:18:15 -08003527 * stripe, so we must do the range check again.
3528 * Expansion could still move past after this
3529 * test, but as we are holding a reference to
3530 * 'sh', we know that if that happens,
3531 * STRIPE_EXPANDING will get set and the expansion
3532 * won't proceed until we finish with the stripe.
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003533 */
3534 int must_retry = 0;
3535 spin_lock_irq(&conf->device_lock);
3536 if (logical_sector < conf->expand_progress &&
3537 disks == conf->previous_raid_disks)
3538 /* mismatch, need to try again */
3539 must_retry = 1;
3540 spin_unlock_irq(&conf->device_lock);
3541 if (must_retry) {
3542 release_stripe(sh);
3543 goto retry;
3544 }
3545 }
NeilBrowne464eaf2006-03-27 01:18:14 -08003546 /* FIXME what if we get a false positive because these
3547 * are being updated.
3548 */
3549 if (logical_sector >= mddev->suspend_lo &&
3550 logical_sector < mddev->suspend_hi) {
3551 release_stripe(sh);
3552 schedule();
3553 goto retry;
3554 }
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003555
3556 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3557 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3558 /* Stripe is busy expanding or
3559 * add failed due to overlap. Flush everything
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 * and wait a while
3561 */
3562 raid5_unplug_device(mddev->queue);
3563 release_stripe(sh);
3564 schedule();
3565 goto retry;
3566 }
3567 finish_wait(&conf->wait_for_overlap, &w);
NeilBrown16a53ec2006-06-26 00:27:38 -07003568 handle_stripe(sh, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569 release_stripe(sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 } else {
3571 /* cannot get stripe for read-ahead, just give-up */
3572 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3573 finish_wait(&conf->wait_for_overlap, &w);
3574 break;
3575 }
3576
3577 }
3578 spin_lock_irq(&conf->device_lock);
NeilBrownf6344752006-03-27 01:18:17 -08003579 remaining = --bi->bi_phys_segments;
3580 spin_unlock_irq(&conf->device_lock);
3581 if (remaining == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582
NeilBrown16a53ec2006-06-26 00:27:38 -07003583 if ( rw == WRITE )
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584 md_write_end(mddev);
NeilBrown6712ecf2007-09-27 12:47:43 +02003585
3586 bi->bi_end_io(bi,
NeilBrownc2b00852006-12-10 02:20:51 -08003587 test_bit(BIO_UPTODATE, &bi->bi_flags)
3588 ? 0 : -EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590 return 0;
3591}
3592
NeilBrown52c03292006-06-26 00:27:43 -07003593static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594{
NeilBrown52c03292006-06-26 00:27:43 -07003595 /* reshaping is quite different to recovery/resync so it is
3596 * handled quite separately ... here.
3597 *
3598 * On each call to sync_request, we gather one chunk worth of
3599 * destination stripes and flag them as expanding.
3600 * Then we find all the source stripes and request reads.
3601 * As the reads complete, handle_stripe will copy the data
3602 * into the destination stripe and release that stripe.
3603 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3605 struct stripe_head *sh;
NeilBrownccfcc3c2006-03-27 01:18:09 -08003606 int pd_idx;
3607 sector_t first_sector, last_sector;
NeilBrownf4168852007-02-28 20:11:53 -08003608 int raid_disks = conf->previous_raid_disks;
3609 int data_disks = raid_disks - conf->max_degraded;
3610 int new_data_disks = conf->raid_disks - conf->max_degraded;
NeilBrown52c03292006-06-26 00:27:43 -07003611 int i;
3612 int dd_idx;
3613 sector_t writepos, safepos, gap;
3614
3615 if (sector_nr == 0 &&
3616 conf->expand_progress != 0) {
3617 /* restarting in the middle, skip the initial sectors */
3618 sector_nr = conf->expand_progress;
NeilBrownf4168852007-02-28 20:11:53 -08003619 sector_div(sector_nr, new_data_disks);
NeilBrown52c03292006-06-26 00:27:43 -07003620 *skipped = 1;
3621 return sector_nr;
3622 }
3623
3624 /* we update the metadata when there is more than 3Meg
3625 * in the block range (that is rather arbitrary, should
3626 * probably be time based) or when the data about to be
3627 * copied would over-write the source of the data at
3628 * the front of the range.
3629 * i.e. one new_stripe forward from expand_progress new_maps
3630 * to after where expand_lo old_maps to
3631 */
3632 writepos = conf->expand_progress +
NeilBrownf4168852007-02-28 20:11:53 -08003633 conf->chunk_size/512*(new_data_disks);
3634 sector_div(writepos, new_data_disks);
NeilBrown52c03292006-06-26 00:27:43 -07003635 safepos = conf->expand_lo;
NeilBrownf4168852007-02-28 20:11:53 -08003636 sector_div(safepos, data_disks);
NeilBrown52c03292006-06-26 00:27:43 -07003637 gap = conf->expand_progress - conf->expand_lo;
3638
3639 if (writepos >= safepos ||
NeilBrownf4168852007-02-28 20:11:53 -08003640 gap > (new_data_disks)*3000*2 /*3Meg*/) {
NeilBrown52c03292006-06-26 00:27:43 -07003641 /* Cannot proceed until we've updated the superblock... */
3642 wait_event(conf->wait_for_overlap,
3643 atomic_read(&conf->reshape_stripes)==0);
3644 mddev->reshape_position = conf->expand_progress;
NeilBrown850b2b42006-10-03 01:15:46 -07003645 set_bit(MD_CHANGE_DEVS, &mddev->flags);
NeilBrown52c03292006-06-26 00:27:43 -07003646 md_wakeup_thread(mddev->thread);
NeilBrown850b2b42006-10-03 01:15:46 -07003647 wait_event(mddev->sb_wait, mddev->flags == 0 ||
NeilBrown52c03292006-06-26 00:27:43 -07003648 kthread_should_stop());
3649 spin_lock_irq(&conf->device_lock);
3650 conf->expand_lo = mddev->reshape_position;
3651 spin_unlock_irq(&conf->device_lock);
3652 wake_up(&conf->wait_for_overlap);
3653 }
3654
3655 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3656 int j;
3657 int skipped = 0;
3658 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3659 sh = get_active_stripe(conf, sector_nr+i,
3660 conf->raid_disks, pd_idx, 0);
3661 set_bit(STRIPE_EXPANDING, &sh->state);
3662 atomic_inc(&conf->reshape_stripes);
3663 /* If any of this stripe is beyond the end of the old
3664 * array, then we need to zero those blocks
3665 */
3666 for (j=sh->disks; j--;) {
3667 sector_t s;
3668 if (j == sh->pd_idx)
3669 continue;
NeilBrownf4168852007-02-28 20:11:53 -08003670 if (conf->level == 6 &&
3671 j == raid6_next_disk(sh->pd_idx, sh->disks))
3672 continue;
NeilBrown52c03292006-06-26 00:27:43 -07003673 s = compute_blocknr(sh, j);
3674 if (s < (mddev->array_size<<1)) {
3675 skipped = 1;
3676 continue;
3677 }
3678 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3679 set_bit(R5_Expanded, &sh->dev[j].flags);
3680 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3681 }
3682 if (!skipped) {
3683 set_bit(STRIPE_EXPAND_READY, &sh->state);
3684 set_bit(STRIPE_HANDLE, &sh->state);
3685 }
3686 release_stripe(sh);
3687 }
3688 spin_lock_irq(&conf->device_lock);
NeilBrown6d3baf22007-03-05 00:30:44 -08003689 conf->expand_progress = (sector_nr + i) * new_data_disks;
NeilBrown52c03292006-06-26 00:27:43 -07003690 spin_unlock_irq(&conf->device_lock);
3691 /* Ok, those stripe are ready. We can start scheduling
3692 * reads on the source stripes.
3693 * The source stripes are determined by mapping the first and last
3694 * block on the destination stripes.
3695 */
NeilBrown52c03292006-06-26 00:27:43 -07003696 first_sector =
NeilBrownf4168852007-02-28 20:11:53 -08003697 raid5_compute_sector(sector_nr*(new_data_disks),
NeilBrown52c03292006-06-26 00:27:43 -07003698 raid_disks, data_disks,
3699 &dd_idx, &pd_idx, conf);
3700 last_sector =
3701 raid5_compute_sector((sector_nr+conf->chunk_size/512)
NeilBrownf4168852007-02-28 20:11:53 -08003702 *(new_data_disks) -1,
NeilBrown52c03292006-06-26 00:27:43 -07003703 raid_disks, data_disks,
3704 &dd_idx, &pd_idx, conf);
3705 if (last_sector >= (mddev->size<<1))
3706 last_sector = (mddev->size<<1)-1;
3707 while (first_sector <= last_sector) {
NeilBrownf4168852007-02-28 20:11:53 -08003708 pd_idx = stripe_to_pdidx(first_sector, conf,
3709 conf->previous_raid_disks);
NeilBrown52c03292006-06-26 00:27:43 -07003710 sh = get_active_stripe(conf, first_sector,
3711 conf->previous_raid_disks, pd_idx, 0);
3712 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3713 set_bit(STRIPE_HANDLE, &sh->state);
3714 release_stripe(sh);
3715 first_sector += STRIPE_SECTORS;
3716 }
3717 return conf->chunk_size>>9;
3718}
3719
3720/* FIXME go_faster isn't used */
3721static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3722{
3723 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3724 struct stripe_head *sh;
3725 int pd_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726 int raid_disks = conf->raid_disks;
NeilBrown72626682005-09-09 16:23:54 -07003727 sector_t max_sector = mddev->size << 1;
3728 int sync_blocks;
NeilBrown16a53ec2006-06-26 00:27:38 -07003729 int still_degraded = 0;
3730 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731
NeilBrown72626682005-09-09 16:23:54 -07003732 if (sector_nr >= max_sector) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003733 /* just being told to finish up .. nothing much to do */
3734 unplug_slaves(mddev);
NeilBrown29269552006-03-27 01:18:10 -08003735 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3736 end_reshape(conf);
3737 return 0;
3738 }
NeilBrown72626682005-09-09 16:23:54 -07003739
3740 if (mddev->curr_resync < max_sector) /* aborted */
3741 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3742 &sync_blocks, 1);
NeilBrown16a53ec2006-06-26 00:27:38 -07003743 else /* completed sync */
NeilBrown72626682005-09-09 16:23:54 -07003744 conf->fullsync = 0;
3745 bitmap_close_sync(mddev->bitmap);
3746
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747 return 0;
3748 }
NeilBrownccfcc3c2006-03-27 01:18:09 -08003749
NeilBrown52c03292006-06-26 00:27:43 -07003750 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3751 return reshape_request(mddev, sector_nr, skipped);
NeilBrownf6705572006-03-27 01:18:11 -08003752
NeilBrown16a53ec2006-06-26 00:27:38 -07003753 /* if there is too many failed drives and we are trying
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754 * to resync, then assert that we are finished, because there is
3755 * nothing we can do.
3756 */
NeilBrown3285edf2006-06-26 00:27:55 -07003757 if (mddev->degraded >= conf->max_degraded &&
NeilBrown16a53ec2006-06-26 00:27:38 -07003758 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
NeilBrown57afd892005-06-21 17:17:13 -07003759 sector_t rv = (mddev->size << 1) - sector_nr;
3760 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761 return rv;
3762 }
NeilBrown72626682005-09-09 16:23:54 -07003763 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
NeilBrown3855ad92005-11-08 21:39:38 -08003764 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
NeilBrown72626682005-09-09 16:23:54 -07003765 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3766 /* we can skip this block, and probably more */
3767 sync_blocks /= STRIPE_SECTORS;
3768 *skipped = 1;
3769 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3770 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771
NeilBrownccfcc3c2006-03-27 01:18:09 -08003772 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003773 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003774 if (sh == NULL) {
NeilBrown7ecaa1e2006-03-27 01:18:08 -08003775 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776 /* make sure we don't swamp the stripe cache if someone else
NeilBrown16a53ec2006-06-26 00:27:38 -07003777 * is trying to get access
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778 */
Nishanth Aravamudan66c006a2005-11-07 01:01:17 -08003779 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003780 }
NeilBrown16a53ec2006-06-26 00:27:38 -07003781 /* Need to check if array will still be degraded after recovery/resync
3782 * We don't need to check the 'failed' flag as when that gets set,
3783 * recovery aborts.
3784 */
3785 for (i=0; i<mddev->raid_disks; i++)
3786 if (conf->disks[i].rdev == NULL)
3787 still_degraded = 1;
3788
3789 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3790
3791 spin_lock(&sh->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792 set_bit(STRIPE_SYNCING, &sh->state);
3793 clear_bit(STRIPE_INSYNC, &sh->state);
3794 spin_unlock(&sh->lock);
3795
NeilBrown16a53ec2006-06-26 00:27:38 -07003796 handle_stripe(sh, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797 release_stripe(sh);
3798
3799 return STRIPE_SECTORS;
3800}
3801
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003802static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3803{
3804 /* We may not be able to submit a whole bio at once as there
3805 * may not be enough stripe_heads available.
3806 * We cannot pre-allocate enough stripe_heads as we may need
3807 * more than exist in the cache (if we allow ever large chunks).
3808 * So we do one stripe head at a time and record in
3809 * ->bi_hw_segments how many have been done.
3810 *
3811 * We *know* that this entire raid_bio is in one chunk, so
3812 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3813 */
3814 struct stripe_head *sh;
3815 int dd_idx, pd_idx;
3816 sector_t sector, logical_sector, last_sector;
3817 int scnt = 0;
3818 int remaining;
3819 int handled = 0;
3820
3821 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3822 sector = raid5_compute_sector( logical_sector,
3823 conf->raid_disks,
3824 conf->raid_disks - conf->max_degraded,
3825 &dd_idx,
3826 &pd_idx,
3827 conf);
3828 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3829
3830 for (; logical_sector < last_sector;
Neil Brown387bb172007-02-08 14:20:29 -08003831 logical_sector += STRIPE_SECTORS,
3832 sector += STRIPE_SECTORS,
3833 scnt++) {
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003834
3835 if (scnt < raid_bio->bi_hw_segments)
3836 /* already done this stripe */
3837 continue;
3838
3839 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3840
3841 if (!sh) {
3842 /* failed to get a stripe - must wait */
3843 raid_bio->bi_hw_segments = scnt;
3844 conf->retry_read_aligned = raid_bio;
3845 return handled;
3846 }
3847
3848 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
Neil Brown387bb172007-02-08 14:20:29 -08003849 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3850 release_stripe(sh);
3851 raid_bio->bi_hw_segments = scnt;
3852 conf->retry_read_aligned = raid_bio;
3853 return handled;
3854 }
3855
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003856 handle_stripe(sh, NULL);
3857 release_stripe(sh);
3858 handled++;
3859 }
3860 spin_lock_irq(&conf->device_lock);
3861 remaining = --raid_bio->bi_phys_segments;
3862 spin_unlock_irq(&conf->device_lock);
3863 if (remaining == 0) {
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003864
NeilBrown6712ecf2007-09-27 12:47:43 +02003865 raid_bio->bi_end_io(raid_bio,
NeilBrownc2b00852006-12-10 02:20:51 -08003866 test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3867 ? 0 : -EIO);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003868 }
3869 if (atomic_dec_and_test(&conf->active_aligned_reads))
3870 wake_up(&conf->wait_for_stripe);
3871 return handled;
3872}
3873
3874
3875
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876/*
3877 * This is our raid5 kernel thread.
3878 *
3879 * We scan the hash table for stripes which can be handled now.
3880 * During the scan, completed stripes are saved for us by the interrupt
3881 * handler, so that they will not have to wait for our next wakeup.
3882 */
3883static void raid5d (mddev_t *mddev)
3884{
3885 struct stripe_head *sh;
3886 raid5_conf_t *conf = mddev_to_conf(mddev);
3887 int handled;
3888
Dan Williams45b42332007-07-09 11:56:43 -07003889 pr_debug("+++ raid5d active\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890
3891 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892
3893 handled = 0;
3894 spin_lock_irq(&conf->device_lock);
3895 while (1) {
3896 struct list_head *first;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003897 struct bio *bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898
NeilBrownae3c20c2006-07-10 04:44:17 -07003899 if (conf->seq_flush != conf->seq_write) {
NeilBrown72626682005-09-09 16:23:54 -07003900 int seq = conf->seq_flush;
NeilBrown700e4322005-11-28 13:44:10 -08003901 spin_unlock_irq(&conf->device_lock);
NeilBrown72626682005-09-09 16:23:54 -07003902 bitmap_unplug(mddev->bitmap);
NeilBrown700e4322005-11-28 13:44:10 -08003903 spin_lock_irq(&conf->device_lock);
NeilBrown72626682005-09-09 16:23:54 -07003904 conf->seq_write = seq;
3905 activate_bit_delay(conf);
3906 }
3907
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908 if (list_empty(&conf->handle_list) &&
3909 atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
3910 !blk_queue_plugged(mddev->queue) &&
3911 !list_empty(&conf->delayed_list))
3912 raid5_activate_delayed(conf);
3913
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003914 while ((bio = remove_bio_from_retry(conf))) {
3915 int ok;
3916 spin_unlock_irq(&conf->device_lock);
3917 ok = retry_aligned_read(conf, bio);
3918 spin_lock_irq(&conf->device_lock);
3919 if (!ok)
3920 break;
3921 handled++;
3922 }
3923
Dan Williamsd84e0f12007-01-02 13:52:30 -07003924 if (list_empty(&conf->handle_list)) {
3925 async_tx_issue_pending_all();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926 break;
Dan Williamsd84e0f12007-01-02 13:52:30 -07003927 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928
3929 first = conf->handle_list.next;
3930 sh = list_entry(first, struct stripe_head, lru);
3931
3932 list_del_init(first);
3933 atomic_inc(&sh->count);
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +02003934 BUG_ON(atomic_read(&sh->count)!= 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003935 spin_unlock_irq(&conf->device_lock);
3936
3937 handled++;
NeilBrown16a53ec2006-06-26 00:27:38 -07003938 handle_stripe(sh, conf->spare_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939 release_stripe(sh);
3940
3941 spin_lock_irq(&conf->device_lock);
3942 }
Dan Williams45b42332007-07-09 11:56:43 -07003943 pr_debug("%d stripes handled\n", handled);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944
3945 spin_unlock_irq(&conf->device_lock);
3946
3947 unplug_slaves(mddev);
3948
Dan Williams45b42332007-07-09 11:56:43 -07003949 pr_debug("--- raid5d inactive\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950}
3951
NeilBrown3f294f42005-11-08 21:39:25 -08003952static ssize_t
NeilBrown007583c2005-11-08 21:39:30 -08003953raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
NeilBrown3f294f42005-11-08 21:39:25 -08003954{
NeilBrown007583c2005-11-08 21:39:30 -08003955 raid5_conf_t *conf = mddev_to_conf(mddev);
NeilBrown96de1e62005-11-08 21:39:39 -08003956 if (conf)
3957 return sprintf(page, "%d\n", conf->max_nr_stripes);
3958 else
3959 return 0;
NeilBrown3f294f42005-11-08 21:39:25 -08003960}
3961
3962static ssize_t
NeilBrown007583c2005-11-08 21:39:30 -08003963raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
NeilBrown3f294f42005-11-08 21:39:25 -08003964{
NeilBrown007583c2005-11-08 21:39:30 -08003965 raid5_conf_t *conf = mddev_to_conf(mddev);
NeilBrown3f294f42005-11-08 21:39:25 -08003966 char *end;
3967 int new;
3968 if (len >= PAGE_SIZE)
3969 return -EINVAL;
NeilBrown96de1e62005-11-08 21:39:39 -08003970 if (!conf)
3971 return -ENODEV;
NeilBrown3f294f42005-11-08 21:39:25 -08003972
3973 new = simple_strtoul(page, &end, 10);
3974 if (!*page || (*end && *end != '\n') )
3975 return -EINVAL;
3976 if (new <= 16 || new > 32768)
3977 return -EINVAL;
3978 while (new < conf->max_nr_stripes) {
3979 if (drop_one_stripe(conf))
3980 conf->max_nr_stripes--;
3981 else
3982 break;
3983 }
NeilBrown2a2275d2007-01-26 00:57:11 -08003984 md_allow_write(mddev);
NeilBrown3f294f42005-11-08 21:39:25 -08003985 while (new > conf->max_nr_stripes) {
3986 if (grow_one_stripe(conf))
3987 conf->max_nr_stripes++;
3988 else break;
3989 }
3990 return len;
3991}
NeilBrown007583c2005-11-08 21:39:30 -08003992
NeilBrown96de1e62005-11-08 21:39:39 -08003993static struct md_sysfs_entry
3994raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3995 raid5_show_stripe_cache_size,
3996 raid5_store_stripe_cache_size);
NeilBrown3f294f42005-11-08 21:39:25 -08003997
3998static ssize_t
NeilBrown96de1e62005-11-08 21:39:39 -08003999stripe_cache_active_show(mddev_t *mddev, char *page)
NeilBrown3f294f42005-11-08 21:39:25 -08004000{
NeilBrown007583c2005-11-08 21:39:30 -08004001 raid5_conf_t *conf = mddev_to_conf(mddev);
NeilBrown96de1e62005-11-08 21:39:39 -08004002 if (conf)
4003 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4004 else
4005 return 0;
NeilBrown3f294f42005-11-08 21:39:25 -08004006}
4007
NeilBrown96de1e62005-11-08 21:39:39 -08004008static struct md_sysfs_entry
4009raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
NeilBrown3f294f42005-11-08 21:39:25 -08004010
NeilBrown007583c2005-11-08 21:39:30 -08004011static struct attribute *raid5_attrs[] = {
NeilBrown3f294f42005-11-08 21:39:25 -08004012 &raid5_stripecache_size.attr,
4013 &raid5_stripecache_active.attr,
4014 NULL,
4015};
NeilBrown007583c2005-11-08 21:39:30 -08004016static struct attribute_group raid5_attrs_group = {
4017 .name = NULL,
4018 .attrs = raid5_attrs,
NeilBrown3f294f42005-11-08 21:39:25 -08004019};
4020
NeilBrown72626682005-09-09 16:23:54 -07004021static int run(mddev_t *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022{
4023 raid5_conf_t *conf;
4024 int raid_disk, memory;
4025 mdk_rdev_t *rdev;
4026 struct disk_info *disk;
4027 struct list_head *tmp;
NeilBrown02c2de82006-10-03 01:15:47 -07004028 int working_disks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029
NeilBrown16a53ec2006-06-26 00:27:38 -07004030 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4031 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
NeilBrown14f8d262006-01-06 00:20:14 -08004032 mdname(mddev), mddev->level);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033 return -EIO;
4034 }
4035
NeilBrownf6705572006-03-27 01:18:11 -08004036 if (mddev->reshape_position != MaxSector) {
4037 /* Check that we can continue the reshape.
4038 * Currently only disks can change, it must
4039 * increase, and we must be past the point where
4040 * a stripe over-writes itself
4041 */
4042 sector_t here_new, here_old;
4043 int old_disks;
NeilBrownf4168852007-02-28 20:11:53 -08004044 int max_degraded = (mddev->level == 5 ? 1 : 2);
NeilBrownf6705572006-03-27 01:18:11 -08004045
4046 if (mddev->new_level != mddev->level ||
4047 mddev->new_layout != mddev->layout ||
4048 mddev->new_chunk != mddev->chunk_size) {
NeilBrownf4168852007-02-28 20:11:53 -08004049 printk(KERN_ERR "raid5: %s: unsupported reshape "
4050 "required - aborting.\n",
NeilBrownf6705572006-03-27 01:18:11 -08004051 mdname(mddev));
4052 return -EINVAL;
4053 }
4054 if (mddev->delta_disks <= 0) {
NeilBrownf4168852007-02-28 20:11:53 -08004055 printk(KERN_ERR "raid5: %s: unsupported reshape "
4056 "(reduce disks) required - aborting.\n",
NeilBrownf6705572006-03-27 01:18:11 -08004057 mdname(mddev));
4058 return -EINVAL;
4059 }
4060 old_disks = mddev->raid_disks - mddev->delta_disks;
4061 /* reshape_position must be on a new-stripe boundary, and one
NeilBrownf4168852007-02-28 20:11:53 -08004062 * further up in new geometry must map after here in old
4063 * geometry.
NeilBrownf6705572006-03-27 01:18:11 -08004064 */
4065 here_new = mddev->reshape_position;
NeilBrownf4168852007-02-28 20:11:53 -08004066 if (sector_div(here_new, (mddev->chunk_size>>9)*
4067 (mddev->raid_disks - max_degraded))) {
4068 printk(KERN_ERR "raid5: reshape_position not "
4069 "on a stripe boundary\n");
NeilBrownf6705572006-03-27 01:18:11 -08004070 return -EINVAL;
4071 }
4072 /* here_new is the stripe we will write to */
4073 here_old = mddev->reshape_position;
NeilBrownf4168852007-02-28 20:11:53 -08004074 sector_div(here_old, (mddev->chunk_size>>9)*
4075 (old_disks-max_degraded));
4076 /* here_old is the first stripe that we might need to read
4077 * from */
NeilBrownf6705572006-03-27 01:18:11 -08004078 if (here_new >= here_old) {
4079 /* Reading from the same stripe as writing to - bad */
NeilBrownf4168852007-02-28 20:11:53 -08004080 printk(KERN_ERR "raid5: reshape_position too early for "
4081 "auto-recovery - aborting.\n");
NeilBrownf6705572006-03-27 01:18:11 -08004082 return -EINVAL;
4083 }
4084 printk(KERN_INFO "raid5: reshape will continue\n");
4085 /* OK, we should be able to continue; */
4086 }
4087
4088
NeilBrownb55e6bf2006-03-27 01:18:06 -08004089 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090 if ((conf = mddev->private) == NULL)
4091 goto abort;
NeilBrownf6705572006-03-27 01:18:11 -08004092 if (mddev->reshape_position == MaxSector) {
4093 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4094 } else {
4095 conf->raid_disks = mddev->raid_disks;
4096 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4097 }
4098
4099 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
NeilBrownb55e6bf2006-03-27 01:18:06 -08004100 GFP_KERNEL);
4101 if (!conf->disks)
4102 goto abort;
NeilBrown9ffae0c2006-01-06 00:20:32 -08004103
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104 conf->mddev = mddev;
4105
NeilBrownfccddba2006-01-06 00:20:33 -08004106 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 goto abort;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004108
NeilBrown16a53ec2006-06-26 00:27:38 -07004109 if (mddev->level == 6) {
4110 conf->spare_page = alloc_page(GFP_KERNEL);
4111 if (!conf->spare_page)
4112 goto abort;
4113 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114 spin_lock_init(&conf->device_lock);
4115 init_waitqueue_head(&conf->wait_for_stripe);
4116 init_waitqueue_head(&conf->wait_for_overlap);
4117 INIT_LIST_HEAD(&conf->handle_list);
4118 INIT_LIST_HEAD(&conf->delayed_list);
NeilBrown72626682005-09-09 16:23:54 -07004119 INIT_LIST_HEAD(&conf->bitmap_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120 INIT_LIST_HEAD(&conf->inactive_list);
4121 atomic_set(&conf->active_stripes, 0);
4122 atomic_set(&conf->preread_active_stripes, 0);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004123 atomic_set(&conf->active_aligned_reads, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124
Dan Williams45b42332007-07-09 11:56:43 -07004125 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004126
4127 ITERATE_RDEV(mddev,rdev,tmp) {
4128 raid_disk = rdev->raid_disk;
NeilBrownf6705572006-03-27 01:18:11 -08004129 if (raid_disk >= conf->raid_disks
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130 || raid_disk < 0)
4131 continue;
4132 disk = conf->disks + raid_disk;
4133
4134 disk->rdev = rdev;
4135
NeilBrownb2d444d2005-11-08 21:39:31 -08004136 if (test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004137 char b[BDEVNAME_SIZE];
4138 printk(KERN_INFO "raid5: device %s operational as raid"
4139 " disk %d\n", bdevname(rdev->bdev,b),
4140 raid_disk);
NeilBrown02c2de82006-10-03 01:15:47 -07004141 working_disks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142 }
4143 }
4144
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145 /*
NeilBrown16a53ec2006-06-26 00:27:38 -07004146 * 0 for a fully functional array, 1 or 2 for a degraded array.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147 */
NeilBrown02c2de82006-10-03 01:15:47 -07004148 mddev->degraded = conf->raid_disks - working_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004149 conf->mddev = mddev;
4150 conf->chunk_size = mddev->chunk_size;
4151 conf->level = mddev->level;
NeilBrown16a53ec2006-06-26 00:27:38 -07004152 if (conf->level == 6)
4153 conf->max_degraded = 2;
4154 else
4155 conf->max_degraded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156 conf->algorithm = mddev->layout;
4157 conf->max_nr_stripes = NR_STRIPES;
NeilBrownf6705572006-03-27 01:18:11 -08004158 conf->expand_progress = mddev->reshape_position;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159
4160 /* device size must be a multiple of chunk size */
4161 mddev->size &= ~(mddev->chunk_size/1024 -1);
NeilBrownb1581562005-07-31 22:34:50 -07004162 mddev->resync_max_sectors = mddev->size << 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163
NeilBrown16a53ec2006-06-26 00:27:38 -07004164 if (conf->level == 6 && conf->raid_disks < 4) {
4165 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4166 mdname(mddev), conf->raid_disks);
4167 goto abort;
4168 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169 if (!conf->chunk_size || conf->chunk_size % 4) {
4170 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4171 conf->chunk_size, mdname(mddev));
4172 goto abort;
4173 }
4174 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4175 printk(KERN_ERR
4176 "raid5: unsupported parity algorithm %d for %s\n",
4177 conf->algorithm, mdname(mddev));
4178 goto abort;
4179 }
NeilBrown16a53ec2006-06-26 00:27:38 -07004180 if (mddev->degraded > conf->max_degraded) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181 printk(KERN_ERR "raid5: not enough operational devices for %s"
4182 " (%d/%d failed)\n",
NeilBrown02c2de82006-10-03 01:15:47 -07004183 mdname(mddev), mddev->degraded, conf->raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004184 goto abort;
4185 }
4186
NeilBrown16a53ec2006-06-26 00:27:38 -07004187 if (mddev->degraded > 0 &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004188 mddev->recovery_cp != MaxSector) {
NeilBrown6ff8d8ec2006-01-06 00:20:15 -08004189 if (mddev->ok_start_degraded)
4190 printk(KERN_WARNING
4191 "raid5: starting dirty degraded array: %s"
4192 "- data corruption possible.\n",
4193 mdname(mddev));
4194 else {
4195 printk(KERN_ERR
4196 "raid5: cannot start dirty degraded array for %s\n",
4197 mdname(mddev));
4198 goto abort;
4199 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200 }
4201
4202 {
4203 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4204 if (!mddev->thread) {
4205 printk(KERN_ERR
4206 "raid5: couldn't allocate thread for %s\n",
4207 mdname(mddev));
4208 goto abort;
4209 }
4210 }
NeilBrown50368052005-12-12 02:39:17 -08004211 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4213 if (grow_stripes(conf, conf->max_nr_stripes)) {
4214 printk(KERN_ERR
4215 "raid5: couldn't allocate %dkB for buffers\n", memory);
4216 shrink_stripes(conf);
4217 md_unregister_thread(mddev->thread);
4218 goto abort;
4219 } else
4220 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4221 memory, mdname(mddev));
4222
4223 if (mddev->degraded == 0)
4224 printk("raid5: raid level %d set %s active with %d out of %d"
4225 " devices, algorithm %d\n", conf->level, mdname(mddev),
4226 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4227 conf->algorithm);
4228 else
4229 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4230 " out of %d devices, algorithm %d\n", conf->level,
4231 mdname(mddev), mddev->raid_disks - mddev->degraded,
4232 mddev->raid_disks, conf->algorithm);
4233
4234 print_raid5_conf(conf);
4235
NeilBrownf6705572006-03-27 01:18:11 -08004236 if (conf->expand_progress != MaxSector) {
4237 printk("...ok start reshape thread\n");
NeilBrownb578d552006-03-27 01:18:12 -08004238 conf->expand_lo = conf->expand_progress;
NeilBrownf6705572006-03-27 01:18:11 -08004239 atomic_set(&conf->reshape_stripes, 0);
4240 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4241 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4242 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4243 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4244 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4245 "%s_reshape");
NeilBrownf6705572006-03-27 01:18:11 -08004246 }
4247
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248 /* read-ahead size must cover two whole stripes, which is
NeilBrown16a53ec2006-06-26 00:27:38 -07004249 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250 */
4251 {
NeilBrown16a53ec2006-06-26 00:27:38 -07004252 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4253 int stripe = data_disks *
NeilBrown8932c2e2006-06-26 00:27:36 -07004254 (mddev->chunk_size / PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004255 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4256 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4257 }
4258
4259 /* Ok, everything is just fine now */
NeilBrown5e55e2f2007-03-26 21:32:14 -08004260 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4261 printk(KERN_WARNING
4262 "raid5: failed to create sysfs attributes for %s\n",
4263 mdname(mddev));
NeilBrown7a5febe2005-05-16 21:53:16 -07004264
4265 mddev->queue->unplug_fn = raid5_unplug_device;
4266 mddev->queue->issue_flush_fn = raid5_issue_flush;
NeilBrownf022b2f2006-10-03 01:15:56 -07004267 mddev->queue->backing_dev_info.congested_data = mddev;
NeilBrown041ae522007-03-26 21:32:14 -08004268 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
NeilBrownf022b2f2006-10-03 01:15:56 -07004269
NeilBrown16a53ec2006-06-26 00:27:38 -07004270 mddev->array_size = mddev->size * (conf->previous_raid_disks -
4271 conf->max_degraded);
NeilBrown7a5febe2005-05-16 21:53:16 -07004272
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08004273 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4274
Linus Torvalds1da177e2005-04-16 15:20:36 -07004275 return 0;
4276abort:
4277 if (conf) {
4278 print_raid5_conf(conf);
NeilBrown16a53ec2006-06-26 00:27:38 -07004279 safe_put_page(conf->spare_page);
NeilBrownb55e6bf2006-03-27 01:18:06 -08004280 kfree(conf->disks);
NeilBrownfccddba2006-01-06 00:20:33 -08004281 kfree(conf->stripe_hashtbl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004282 kfree(conf);
4283 }
4284 mddev->private = NULL;
4285 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4286 return -EIO;
4287}
4288
4289
4290
NeilBrown3f294f42005-11-08 21:39:25 -08004291static int stop(mddev_t *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004292{
4293 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4294
4295 md_unregister_thread(mddev->thread);
4296 mddev->thread = NULL;
4297 shrink_stripes(conf);
NeilBrownfccddba2006-01-06 00:20:33 -08004298 kfree(conf->stripe_hashtbl);
NeilBrown041ae522007-03-26 21:32:14 -08004299 mddev->queue->backing_dev_info.congested_fn = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004300 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
NeilBrown007583c2005-11-08 21:39:30 -08004301 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
NeilBrownb55e6bf2006-03-27 01:18:06 -08004302 kfree(conf->disks);
NeilBrown96de1e62005-11-08 21:39:39 -08004303 kfree(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004304 mddev->private = NULL;
4305 return 0;
4306}
4307
Dan Williams45b42332007-07-09 11:56:43 -07004308#ifdef DEBUG
NeilBrown16a53ec2006-06-26 00:27:38 -07004309static void print_sh (struct seq_file *seq, struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004310{
4311 int i;
4312
NeilBrown16a53ec2006-06-26 00:27:38 -07004313 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4314 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4315 seq_printf(seq, "sh %llu, count %d.\n",
4316 (unsigned long long)sh->sector, atomic_read(&sh->count));
4317 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004318 for (i = 0; i < sh->disks; i++) {
NeilBrown16a53ec2006-06-26 00:27:38 -07004319 seq_printf(seq, "(cache%d: %p %ld) ",
4320 i, sh->dev[i].page, sh->dev[i].flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004321 }
NeilBrown16a53ec2006-06-26 00:27:38 -07004322 seq_printf(seq, "\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004323}
4324
NeilBrown16a53ec2006-06-26 00:27:38 -07004325static void printall (struct seq_file *seq, raid5_conf_t *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004326{
4327 struct stripe_head *sh;
NeilBrownfccddba2006-01-06 00:20:33 -08004328 struct hlist_node *hn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004329 int i;
4330
4331 spin_lock_irq(&conf->device_lock);
4332 for (i = 0; i < NR_HASH; i++) {
NeilBrownfccddba2006-01-06 00:20:33 -08004333 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004334 if (sh->raid_conf != conf)
4335 continue;
NeilBrown16a53ec2006-06-26 00:27:38 -07004336 print_sh(seq, sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004337 }
4338 }
4339 spin_unlock_irq(&conf->device_lock);
4340}
4341#endif
4342
4343static void status (struct seq_file *seq, mddev_t *mddev)
4344{
4345 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4346 int i;
4347
4348 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
NeilBrown02c2de82006-10-03 01:15:47 -07004349 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004350 for (i = 0; i < conf->raid_disks; i++)
4351 seq_printf (seq, "%s",
4352 conf->disks[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08004353 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004354 seq_printf (seq, "]");
Dan Williams45b42332007-07-09 11:56:43 -07004355#ifdef DEBUG
NeilBrown16a53ec2006-06-26 00:27:38 -07004356 seq_printf (seq, "\n");
4357 printall(seq, conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004358#endif
4359}
4360
4361static void print_raid5_conf (raid5_conf_t *conf)
4362{
4363 int i;
4364 struct disk_info *tmp;
4365
4366 printk("RAID5 conf printout:\n");
4367 if (!conf) {
4368 printk("(conf==NULL)\n");
4369 return;
4370 }
NeilBrown02c2de82006-10-03 01:15:47 -07004371 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4372 conf->raid_disks - conf->mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004373
4374 for (i = 0; i < conf->raid_disks; i++) {
4375 char b[BDEVNAME_SIZE];
4376 tmp = conf->disks + i;
4377 if (tmp->rdev)
4378 printk(" disk %d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -08004379 i, !test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -07004380 bdevname(tmp->rdev->bdev,b));
4381 }
4382}
4383
4384static int raid5_spare_active(mddev_t *mddev)
4385{
4386 int i;
4387 raid5_conf_t *conf = mddev->private;
4388 struct disk_info *tmp;
4389
4390 for (i = 0; i < conf->raid_disks; i++) {
4391 tmp = conf->disks + i;
4392 if (tmp->rdev
NeilBrownb2d444d2005-11-08 21:39:31 -08004393 && !test_bit(Faulty, &tmp->rdev->flags)
NeilBrownc04be0a2006-10-03 01:15:53 -07004394 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4395 unsigned long flags;
4396 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004397 mddev->degraded--;
NeilBrownc04be0a2006-10-03 01:15:53 -07004398 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004399 }
4400 }
4401 print_raid5_conf(conf);
4402 return 0;
4403}
4404
4405static int raid5_remove_disk(mddev_t *mddev, int number)
4406{
4407 raid5_conf_t *conf = mddev->private;
4408 int err = 0;
4409 mdk_rdev_t *rdev;
4410 struct disk_info *p = conf->disks + number;
4411
4412 print_raid5_conf(conf);
4413 rdev = p->rdev;
4414 if (rdev) {
NeilBrownb2d444d2005-11-08 21:39:31 -08004415 if (test_bit(In_sync, &rdev->flags) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004416 atomic_read(&rdev->nr_pending)) {
4417 err = -EBUSY;
4418 goto abort;
4419 }
4420 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07004421 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004422 if (atomic_read(&rdev->nr_pending)) {
4423 /* lost the race, try later */
4424 err = -EBUSY;
4425 p->rdev = rdev;
4426 }
4427 }
4428abort:
4429
4430 print_raid5_conf(conf);
4431 return err;
4432}
4433
4434static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4435{
4436 raid5_conf_t *conf = mddev->private;
4437 int found = 0;
4438 int disk;
4439 struct disk_info *p;
4440
NeilBrown16a53ec2006-06-26 00:27:38 -07004441 if (mddev->degraded > conf->max_degraded)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004442 /* no point adding a device */
4443 return 0;
4444
4445 /*
NeilBrown16a53ec2006-06-26 00:27:38 -07004446 * find the disk ... but prefer rdev->saved_raid_disk
4447 * if possible.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004448 */
NeilBrown16a53ec2006-06-26 00:27:38 -07004449 if (rdev->saved_raid_disk >= 0 &&
4450 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4451 disk = rdev->saved_raid_disk;
4452 else
4453 disk = 0;
4454 for ( ; disk < conf->raid_disks; disk++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004455 if ((p=conf->disks + disk)->rdev == NULL) {
NeilBrownb2d444d2005-11-08 21:39:31 -08004456 clear_bit(In_sync, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004457 rdev->raid_disk = disk;
4458 found = 1;
NeilBrown72626682005-09-09 16:23:54 -07004459 if (rdev->saved_raid_disk != disk)
4460 conf->fullsync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -08004461 rcu_assign_pointer(p->rdev, rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004462 break;
4463 }
4464 print_raid5_conf(conf);
4465 return found;
4466}
4467
4468static int raid5_resize(mddev_t *mddev, sector_t sectors)
4469{
4470 /* no resync is happening, and there is enough space
4471 * on all devices, so we can resize.
4472 * We need to make sure resync covers any new space.
4473 * If the array is shrinking we should possibly wait until
4474 * any io in the removed space completes, but it hardly seems
4475 * worth it.
4476 */
NeilBrown16a53ec2006-06-26 00:27:38 -07004477 raid5_conf_t *conf = mddev_to_conf(mddev);
4478
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
NeilBrown16a53ec2006-06-26 00:27:38 -07004480 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481 set_capacity(mddev->gendisk, mddev->array_size << 1);
Linus Torvalds44ce62942007-05-09 18:51:36 -07004482 mddev->changed = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004483 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
4484 mddev->recovery_cp = mddev->size << 1;
4485 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4486 }
4487 mddev->size = sectors /2;
NeilBrown4b5c7ae2005-07-27 11:43:28 -07004488 mddev->resync_max_sectors = sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004489 return 0;
4490}
4491
NeilBrown29269552006-03-27 01:18:10 -08004492#ifdef CONFIG_MD_RAID5_RESHAPE
NeilBrown63c70c42006-03-27 01:18:13 -08004493static int raid5_check_reshape(mddev_t *mddev)
NeilBrown29269552006-03-27 01:18:10 -08004494{
4495 raid5_conf_t *conf = mddev_to_conf(mddev);
4496 int err;
NeilBrown29269552006-03-27 01:18:10 -08004497
NeilBrown63c70c42006-03-27 01:18:13 -08004498 if (mddev->delta_disks < 0 ||
4499 mddev->new_level != mddev->level)
4500 return -EINVAL; /* Cannot shrink array or change level yet */
4501 if (mddev->delta_disks == 0)
NeilBrown29269552006-03-27 01:18:10 -08004502 return 0; /* nothing to do */
4503
4504 /* Can only proceed if there are plenty of stripe_heads.
4505 * We need a minimum of one full stripe,, and for sensible progress
4506 * it is best to have about 4 times that.
4507 * If we require 4 times, then the default 256 4K stripe_heads will
4508 * allow for chunk sizes up to 256K, which is probably OK.
4509 * If the chunk size is greater, user-space should request more
4510 * stripe_heads first.
4511 */
NeilBrown63c70c42006-03-27 01:18:13 -08004512 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4513 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
NeilBrown29269552006-03-27 01:18:10 -08004514 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4515 (mddev->chunk_size / STRIPE_SIZE)*4);
4516 return -ENOSPC;
4517 }
4518
NeilBrown63c70c42006-03-27 01:18:13 -08004519 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4520 if (err)
4521 return err;
4522
NeilBrownb4c4c7b2007-02-28 20:11:48 -08004523 if (mddev->degraded > conf->max_degraded)
4524 return -EINVAL;
NeilBrown63c70c42006-03-27 01:18:13 -08004525 /* looks like we might be able to manage this */
4526 return 0;
4527}
4528
4529static int raid5_start_reshape(mddev_t *mddev)
4530{
4531 raid5_conf_t *conf = mddev_to_conf(mddev);
4532 mdk_rdev_t *rdev;
4533 struct list_head *rtmp;
4534 int spares = 0;
4535 int added_devices = 0;
NeilBrownc04be0a2006-10-03 01:15:53 -07004536 unsigned long flags;
NeilBrown63c70c42006-03-27 01:18:13 -08004537
NeilBrownf4168852007-02-28 20:11:53 -08004538 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
NeilBrown63c70c42006-03-27 01:18:13 -08004539 return -EBUSY;
4540
NeilBrown29269552006-03-27 01:18:10 -08004541 ITERATE_RDEV(mddev, rdev, rtmp)
4542 if (rdev->raid_disk < 0 &&
4543 !test_bit(Faulty, &rdev->flags))
4544 spares++;
NeilBrown63c70c42006-03-27 01:18:13 -08004545
NeilBrownf4168852007-02-28 20:11:53 -08004546 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
NeilBrown29269552006-03-27 01:18:10 -08004547 /* Not enough devices even to make a degraded array
4548 * of that size
4549 */
4550 return -EINVAL;
4551
NeilBrownf6705572006-03-27 01:18:11 -08004552 atomic_set(&conf->reshape_stripes, 0);
NeilBrown29269552006-03-27 01:18:10 -08004553 spin_lock_irq(&conf->device_lock);
4554 conf->previous_raid_disks = conf->raid_disks;
NeilBrown63c70c42006-03-27 01:18:13 -08004555 conf->raid_disks += mddev->delta_disks;
NeilBrown29269552006-03-27 01:18:10 -08004556 conf->expand_progress = 0;
NeilBrownb578d552006-03-27 01:18:12 -08004557 conf->expand_lo = 0;
NeilBrown29269552006-03-27 01:18:10 -08004558 spin_unlock_irq(&conf->device_lock);
4559
4560 /* Add some new drives, as many as will fit.
4561 * We know there are enough to make the newly sized array work.
4562 */
4563 ITERATE_RDEV(mddev, rdev, rtmp)
4564 if (rdev->raid_disk < 0 &&
4565 !test_bit(Faulty, &rdev->flags)) {
4566 if (raid5_add_disk(mddev, rdev)) {
4567 char nm[20];
4568 set_bit(In_sync, &rdev->flags);
NeilBrown29269552006-03-27 01:18:10 -08004569 added_devices++;
NeilBrown5fd6c1d2006-06-26 00:27:40 -07004570 rdev->recovery_offset = 0;
NeilBrown29269552006-03-27 01:18:10 -08004571 sprintf(nm, "rd%d", rdev->raid_disk);
NeilBrown5e55e2f2007-03-26 21:32:14 -08004572 if (sysfs_create_link(&mddev->kobj,
4573 &rdev->kobj, nm))
4574 printk(KERN_WARNING
4575 "raid5: failed to create "
4576 " link %s for %s\n",
4577 nm, mdname(mddev));
NeilBrown29269552006-03-27 01:18:10 -08004578 } else
4579 break;
4580 }
4581
NeilBrownc04be0a2006-10-03 01:15:53 -07004582 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrown63c70c42006-03-27 01:18:13 -08004583 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
NeilBrownc04be0a2006-10-03 01:15:53 -07004584 spin_unlock_irqrestore(&conf->device_lock, flags);
NeilBrown63c70c42006-03-27 01:18:13 -08004585 mddev->raid_disks = conf->raid_disks;
NeilBrownf6705572006-03-27 01:18:11 -08004586 mddev->reshape_position = 0;
NeilBrown850b2b42006-10-03 01:15:46 -07004587 set_bit(MD_CHANGE_DEVS, &mddev->flags);
NeilBrownf6705572006-03-27 01:18:11 -08004588
NeilBrown29269552006-03-27 01:18:10 -08004589 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4590 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4591 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4592 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4593 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4594 "%s_reshape");
4595 if (!mddev->sync_thread) {
4596 mddev->recovery = 0;
4597 spin_lock_irq(&conf->device_lock);
4598 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4599 conf->expand_progress = MaxSector;
4600 spin_unlock_irq(&conf->device_lock);
4601 return -EAGAIN;
4602 }
4603 md_wakeup_thread(mddev->sync_thread);
4604 md_new_event(mddev);
4605 return 0;
4606}
4607#endif
4608
4609static void end_reshape(raid5_conf_t *conf)
4610{
4611 struct block_device *bdev;
4612
NeilBrownf6705572006-03-27 01:18:11 -08004613 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
NeilBrownf4168852007-02-28 20:11:53 -08004614 conf->mddev->array_size = conf->mddev->size *
4615 (conf->raid_disks - conf->max_degraded);
NeilBrownf6705572006-03-27 01:18:11 -08004616 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
Linus Torvalds44ce62942007-05-09 18:51:36 -07004617 conf->mddev->changed = 1;
NeilBrown29269552006-03-27 01:18:10 -08004618
NeilBrownf6705572006-03-27 01:18:11 -08004619 bdev = bdget_disk(conf->mddev->gendisk, 0);
4620 if (bdev) {
4621 mutex_lock(&bdev->bd_inode->i_mutex);
NeilBrown0692c6b2006-11-08 17:44:48 -08004622 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
NeilBrownf6705572006-03-27 01:18:11 -08004623 mutex_unlock(&bdev->bd_inode->i_mutex);
4624 bdput(bdev);
4625 }
4626 spin_lock_irq(&conf->device_lock);
4627 conf->expand_progress = MaxSector;
4628 spin_unlock_irq(&conf->device_lock);
4629 conf->mddev->reshape_position = MaxSector;
NeilBrown16a53ec2006-06-26 00:27:38 -07004630
4631 /* read-ahead size must cover two whole stripes, which is
4632 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4633 */
4634 {
4635 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4636 int stripe = data_disks *
4637 (conf->mddev->chunk_size / PAGE_SIZE);
4638 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4639 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4640 }
NeilBrown29269552006-03-27 01:18:10 -08004641 }
NeilBrown29269552006-03-27 01:18:10 -08004642}
4643
NeilBrown72626682005-09-09 16:23:54 -07004644static void raid5_quiesce(mddev_t *mddev, int state)
4645{
4646 raid5_conf_t *conf = mddev_to_conf(mddev);
4647
4648 switch(state) {
NeilBrowne464eaf2006-03-27 01:18:14 -08004649 case 2: /* resume for a suspend */
4650 wake_up(&conf->wait_for_overlap);
4651 break;
4652
NeilBrown72626682005-09-09 16:23:54 -07004653 case 1: /* stop all writes */
4654 spin_lock_irq(&conf->device_lock);
4655 conf->quiesce = 1;
4656 wait_event_lock_irq(conf->wait_for_stripe,
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004657 atomic_read(&conf->active_stripes) == 0 &&
4658 atomic_read(&conf->active_aligned_reads) == 0,
NeilBrown72626682005-09-09 16:23:54 -07004659 conf->device_lock, /* nothing */);
4660 spin_unlock_irq(&conf->device_lock);
4661 break;
4662
4663 case 0: /* re-enable writes */
4664 spin_lock_irq(&conf->device_lock);
4665 conf->quiesce = 0;
4666 wake_up(&conf->wait_for_stripe);
NeilBrowne464eaf2006-03-27 01:18:14 -08004667 wake_up(&conf->wait_for_overlap);
NeilBrown72626682005-09-09 16:23:54 -07004668 spin_unlock_irq(&conf->device_lock);
4669 break;
4670 }
NeilBrown72626682005-09-09 16:23:54 -07004671}
NeilBrownb15c2e52006-01-06 00:20:16 -08004672
NeilBrown16a53ec2006-06-26 00:27:38 -07004673static struct mdk_personality raid6_personality =
4674{
4675 .name = "raid6",
4676 .level = 6,
4677 .owner = THIS_MODULE,
4678 .make_request = make_request,
4679 .run = run,
4680 .stop = stop,
4681 .status = status,
4682 .error_handler = error,
4683 .hot_add_disk = raid5_add_disk,
4684 .hot_remove_disk= raid5_remove_disk,
4685 .spare_active = raid5_spare_active,
4686 .sync_request = sync_request,
4687 .resize = raid5_resize,
NeilBrownf4168852007-02-28 20:11:53 -08004688#ifdef CONFIG_MD_RAID5_RESHAPE
4689 .check_reshape = raid5_check_reshape,
4690 .start_reshape = raid5_start_reshape,
4691#endif
NeilBrown16a53ec2006-06-26 00:27:38 -07004692 .quiesce = raid5_quiesce,
4693};
NeilBrown2604b702006-01-06 00:20:36 -08004694static struct mdk_personality raid5_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07004695{
4696 .name = "raid5",
NeilBrown2604b702006-01-06 00:20:36 -08004697 .level = 5,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004698 .owner = THIS_MODULE,
4699 .make_request = make_request,
4700 .run = run,
4701 .stop = stop,
4702 .status = status,
4703 .error_handler = error,
4704 .hot_add_disk = raid5_add_disk,
4705 .hot_remove_disk= raid5_remove_disk,
4706 .spare_active = raid5_spare_active,
4707 .sync_request = sync_request,
4708 .resize = raid5_resize,
NeilBrown29269552006-03-27 01:18:10 -08004709#ifdef CONFIG_MD_RAID5_RESHAPE
NeilBrown63c70c42006-03-27 01:18:13 -08004710 .check_reshape = raid5_check_reshape,
4711 .start_reshape = raid5_start_reshape,
NeilBrown29269552006-03-27 01:18:10 -08004712#endif
NeilBrown72626682005-09-09 16:23:54 -07004713 .quiesce = raid5_quiesce,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004714};
4715
NeilBrown2604b702006-01-06 00:20:36 -08004716static struct mdk_personality raid4_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07004717{
NeilBrown2604b702006-01-06 00:20:36 -08004718 .name = "raid4",
4719 .level = 4,
4720 .owner = THIS_MODULE,
4721 .make_request = make_request,
4722 .run = run,
4723 .stop = stop,
4724 .status = status,
4725 .error_handler = error,
4726 .hot_add_disk = raid5_add_disk,
4727 .hot_remove_disk= raid5_remove_disk,
4728 .spare_active = raid5_spare_active,
4729 .sync_request = sync_request,
4730 .resize = raid5_resize,
NeilBrown3d378902007-03-26 21:32:13 -08004731#ifdef CONFIG_MD_RAID5_RESHAPE
4732 .check_reshape = raid5_check_reshape,
4733 .start_reshape = raid5_start_reshape,
4734#endif
NeilBrown2604b702006-01-06 00:20:36 -08004735 .quiesce = raid5_quiesce,
4736};
4737
4738static int __init raid5_init(void)
4739{
NeilBrown16a53ec2006-06-26 00:27:38 -07004740 int e;
4741
4742 e = raid6_select_algo();
4743 if ( e )
4744 return e;
4745 register_md_personality(&raid6_personality);
NeilBrown2604b702006-01-06 00:20:36 -08004746 register_md_personality(&raid5_personality);
4747 register_md_personality(&raid4_personality);
4748 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004749}
4750
NeilBrown2604b702006-01-06 00:20:36 -08004751static void raid5_exit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004752{
NeilBrown16a53ec2006-06-26 00:27:38 -07004753 unregister_md_personality(&raid6_personality);
NeilBrown2604b702006-01-06 00:20:36 -08004754 unregister_md_personality(&raid5_personality);
4755 unregister_md_personality(&raid4_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004756}
4757
4758module_init(raid5_init);
4759module_exit(raid5_exit);
4760MODULE_LICENSE("GPL");
4761MODULE_ALIAS("md-personality-4"); /* RAID5 */
NeilBrownd9d166c2006-01-06 00:20:51 -08004762MODULE_ALIAS("md-raid5");
4763MODULE_ALIAS("md-raid4");
NeilBrown2604b702006-01-06 00:20:36 -08004764MODULE_ALIAS("md-level-5");
4765MODULE_ALIAS("md-level-4");
NeilBrown16a53ec2006-06-26 00:27:38 -07004766MODULE_ALIAS("md-personality-8"); /* RAID6 */
4767MODULE_ALIAS("md-raid6");
4768MODULE_ALIAS("md-level-6");
4769
4770/* This used to be two separate modules, they were: */
4771MODULE_ALIAS("raid5");
4772MODULE_ALIAS("raid6");