]> git.kernelconcepts.de Git - karo-tx-uboot.git/blob - drivers/mtd/ubi/wl.c
Merge branch 'fixes' into cleanups
[karo-tx-uboot.git] / drivers / mtd / ubi / wl.c
1 /*
2  * Copyright (c) International Business Machines Corp., 2006
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of the GNU General Public License as published by
6  * the Free Software Foundation; either version 2 of the License, or
7  * (at your option) any later version.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See
12  * the GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write to the Free Software
16  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
17  *
18  * Authors: Artem Bityutskiy (Битюцкий Артём), Thomas Gleixner
19  */
20
21 /*
22  * UBI wear-leveling unit.
23  *
24  * This unit is responsible for wear-leveling. It works in terms of physical
25  * eraseblocks and erase counters and knows nothing about logical eraseblocks,
26  * volumes, etc. From this unit's perspective all physical eraseblocks are of
27  * two types - used and free. Used physical eraseblocks are those that were
28  * "get" by the 'ubi_wl_get_peb()' function, and free physical eraseblocks are
29  * those that were put by the 'ubi_wl_put_peb()' function.
30  *
31  * Physical eraseblocks returned by 'ubi_wl_get_peb()' have only erase counter
32  * header. The rest of the physical eraseblock contains only 0xFF bytes.
33  *
34  * When physical eraseblocks are returned to the WL unit by means of the
35  * 'ubi_wl_put_peb()' function, they are scheduled for erasure. The erasure is
36  * done asynchronously in context of the per-UBI device background thread,
37  * which is also managed by the WL unit.
38  *
39  * The wear-leveling is ensured by means of moving the contents of used
40  * physical eraseblocks with low erase counter to free physical eraseblocks
41  * with high erase counter.
42  *
43  * The 'ubi_wl_get_peb()' function accepts data type hints which help to pick
44  * an "optimal" physical eraseblock. For example, when it is known that the
45  * physical eraseblock will be "put" soon because it contains short-term data,
46  * the WL unit may pick a free physical eraseblock with low erase counter, and
47  * so forth.
48  *
49  * If the WL unit fails to erase a physical eraseblock, it marks it as bad.
50  *
51  * This unit is also responsible for scrubbing. If a bit-flip is detected in a
52  * physical eraseblock, it has to be moved. Technically this is the same as
53  * moving it for wear-leveling reasons.
54  *
55  * As it was said, for the UBI unit all physical eraseblocks are either "free"
56  * or "used". Free eraseblock are kept in the @wl->free RB-tree, while used
57  * eraseblocks are kept in a set of different RB-trees: @wl->used,
58  * @wl->prot.pnum, @wl->prot.aec, and @wl->scrub.
59  *
60  * Note, in this implementation, we keep a small in-RAM object for each physical
61  * eraseblock. This is surely not a scalable solution. But it appears to be good
62  * enough for moderately large flashes and it is simple. In future, one may
63  * re-work this unit and make it more scalable.
64  *
65  * At the moment this unit does not utilize the sequence number, which was
66  * introduced relatively recently. But it would be wise to do this because the
67  * sequence number of a logical eraseblock characterizes how old is it. For
68  * example, when we move a PEB with low erase counter, and we need to pick the
69  * target PEB, we pick a PEB with the highest EC if our PEB is "old" and we
70  * pick target PEB with an average EC if our PEB is not very "old". This is a
71  * room for future re-works of the WL unit.
72  *
73  * FIXME: looks too complex, should be simplified (later).
74  */
75
76 #ifdef UBI_LINUX
77 #include <linux/slab.h>
78 #include <linux/crc32.h>
79 #include <linux/freezer.h>
80 #include <linux/kthread.h>
81 #endif
82
83 #include <ubi_uboot.h>
84 #include "ubi.h"
85
86 /* Number of physical eraseblocks reserved for wear-leveling purposes */
87 #define WL_RESERVED_PEBS 1
88
89 /*
90  * How many erase cycles are short term, unknown, and long term physical
91  * eraseblocks protected.
92  */
93 #define ST_PROTECTION 16
94 #define U_PROTECTION  10
95 #define LT_PROTECTION 4
96
97 /*
98  * Maximum difference between two erase counters. If this threshold is
99  * exceeded, the WL unit starts moving data from used physical eraseblocks with
100  * low erase counter to free physical eraseblocks with high erase counter.
101  */
102 #define UBI_WL_THRESHOLD CONFIG_MTD_UBI_WL_THRESHOLD
103
104 /*
105  * When a physical eraseblock is moved, the WL unit has to pick the target
106  * physical eraseblock to move to. The simplest way would be just to pick the
107  * one with the highest erase counter. But in certain workloads this could lead
108  * to an unlimited wear of one or few physical eraseblock. Indeed, imagine a
109  * situation when the picked physical eraseblock is constantly erased after the
110  * data is written to it. So, we have a constant which limits the highest erase
111  * counter of the free physical eraseblock to pick. Namely, the WL unit does
112  * not pick eraseblocks with erase counter greater then the lowest erase
113  * counter plus %WL_FREE_MAX_DIFF.
114  */
115 #define WL_FREE_MAX_DIFF (2*UBI_WL_THRESHOLD)
116
117 /*
118  * Maximum number of consecutive background thread failures which is enough to
119  * switch to read-only mode.
120  */
121 #define WL_MAX_FAILURES 32
122
123 /**
124  * struct ubi_wl_prot_entry - PEB protection entry.
125  * @rb_pnum: link in the @wl->prot.pnum RB-tree
126  * @rb_aec: link in the @wl->prot.aec RB-tree
127  * @abs_ec: the absolute erase counter value when the protection ends
128  * @e: the wear-leveling entry of the physical eraseblock under protection
129  *
130  * When the WL unit returns a physical eraseblock, the physical eraseblock is
131  * protected from being moved for some "time". For this reason, the physical
132  * eraseblock is not directly moved from the @wl->free tree to the @wl->used
133  * tree. There is one more tree in between where this physical eraseblock is
134  * temporarily stored (@wl->prot).
135  *
136  * All this protection stuff is needed because:
137  *  o we don't want to move physical eraseblocks just after we have given them
138  *    to the user; instead, we first want to let users fill them up with data;
139  *
140  *  o there is a chance that the user will put the physical eraseblock very
141  *    soon, so it makes sense not to move it for some time, but wait; this is
142  *    especially important in case of "short term" physical eraseblocks.
143  *
144  * Physical eraseblocks stay protected only for limited time. But the "time" is
145  * measured in erase cycles in this case. This is implemented with help of the
146  * absolute erase counter (@wl->abs_ec). When it reaches certain value, the
147  * physical eraseblocks are moved from the protection trees (@wl->prot.*) to
148  * the @wl->used tree.
149  *
150  * Protected physical eraseblocks are searched by physical eraseblock number
151  * (when they are put) and by the absolute erase counter (to check if it is
152  * time to move them to the @wl->used tree). So there are actually 2 RB-trees
153  * storing the protected physical eraseblocks: @wl->prot.pnum and
154  * @wl->prot.aec. They are referred to as the "protection" trees. The
155  * first one is indexed by the physical eraseblock number. The second one is
156  * indexed by the absolute erase counter. Both trees store
157  * &struct ubi_wl_prot_entry objects.
158  *
159  * Each physical eraseblock has 2 main states: free and used. The former state
160  * corresponds to the @wl->free tree. The latter state is split up on several
161  * sub-states:
162  * o the WL movement is allowed (@wl->used tree);
163  * o the WL movement is temporarily prohibited (@wl->prot.pnum and
164  * @wl->prot.aec trees);
165  * o scrubbing is needed (@wl->scrub tree).
166  *
167  * Depending on the sub-state, wear-leveling entries of the used physical
168  * eraseblocks may be kept in one of those trees.
169  */
170 struct ubi_wl_prot_entry {
171         struct rb_node rb_pnum;
172         struct rb_node rb_aec;
173         unsigned long long abs_ec;
174         struct ubi_wl_entry *e;
175 };
176
177 /**
178  * struct ubi_work - UBI work description data structure.
179  * @list: a link in the list of pending works
180  * @func: worker function
181  * @priv: private data of the worker function
182  *
183  * @e: physical eraseblock to erase
184  * @torture: if the physical eraseblock has to be tortured
185  *
186  * The @func pointer points to the worker function. If the @cancel argument is
187  * not zero, the worker has to free the resources and exit immediately. The
188  * worker has to return zero in case of success and a negative error code in
189  * case of failure.
190  */
191 struct ubi_work {
192         struct list_head list;
193         int (*func)(struct ubi_device *ubi, struct ubi_work *wrk, int cancel);
194         /* The below fields are only relevant to erasure works */
195         struct ubi_wl_entry *e;
196         int torture;
197 };
198
199 #ifdef CONFIG_MTD_UBI_DEBUG_PARANOID
200 static int paranoid_check_ec(struct ubi_device *ubi, int pnum, int ec);
201 static int paranoid_check_in_wl_tree(struct ubi_wl_entry *e,
202                                      struct rb_root *root);
203 #else
204 #define paranoid_check_ec(ubi, pnum, ec) 0
205 #define paranoid_check_in_wl_tree(e, root)
206 #endif
207
208 /**
209  * wl_tree_add - add a wear-leveling entry to a WL RB-tree.
210  * @e: the wear-leveling entry to add
211  * @root: the root of the tree
212  *
213  * Note, we use (erase counter, physical eraseblock number) pairs as keys in
214  * the @ubi->used and @ubi->free RB-trees.
215  */
216 static void wl_tree_add(struct ubi_wl_entry *e, struct rb_root *root)
217 {
218         struct rb_node **p, *parent = NULL;
219
220         p = &root->rb_node;
221         while (*p) {
222                 struct ubi_wl_entry *e1;
223
224                 parent = *p;
225                 e1 = rb_entry(parent, struct ubi_wl_entry, rb);
226
227                 if (e->ec < e1->ec)
228                         p = &(*p)->rb_left;
229                 else if (e->ec > e1->ec)
230                         p = &(*p)->rb_right;
231                 else {
232                         ubi_assert(e->pnum != e1->pnum);
233                         if (e->pnum < e1->pnum)
234                                 p = &(*p)->rb_left;
235                         else
236                                 p = &(*p)->rb_right;
237                 }
238         }
239
240         rb_link_node(&e->rb, parent, p);
241         rb_insert_color(&e->rb, root);
242 }
243
244 /**
245  * do_work - do one pending work.
246  * @ubi: UBI device description object
247  *
248  * This function returns zero in case of success and a negative error code in
249  * case of failure.
250  */
251 static int do_work(struct ubi_device *ubi)
252 {
253         int err;
254         struct ubi_work *wrk;
255
256         cond_resched();
257
258         /*
259          * @ubi->work_sem is used to synchronize with the workers. Workers take
260          * it in read mode, so many of them may be doing works at a time. But
261          * the queue flush code has to be sure the whole queue of works is
262          * done, and it takes the mutex in write mode.
263          */
264         down_read(&ubi->work_sem);
265         spin_lock(&ubi->wl_lock);
266         if (list_empty(&ubi->works)) {
267                 spin_unlock(&ubi->wl_lock);
268                 up_read(&ubi->work_sem);
269                 return 0;
270         }
271
272         wrk = list_entry(ubi->works.next, struct ubi_work, list);
273         list_del(&wrk->list);
274         ubi->works_count -= 1;
275         ubi_assert(ubi->works_count >= 0);
276         spin_unlock(&ubi->wl_lock);
277
278         /*
279          * Call the worker function. Do not touch the work structure
280          * after this call as it will have been freed or reused by that
281          * time by the worker function.
282          */
283         err = wrk->func(ubi, wrk, 0);
284         if (err)
285                 ubi_err("work failed with error code %d", err);
286         up_read(&ubi->work_sem);
287
288         return err;
289 }
290
291 /**
292  * produce_free_peb - produce a free physical eraseblock.
293  * @ubi: UBI device description object
294  *
295  * This function tries to make a free PEB by means of synchronous execution of
296  * pending works. This may be needed if, for example the background thread is
297  * disabled. Returns zero in case of success and a negative error code in case
298  * of failure.
299  */
300 static int produce_free_peb(struct ubi_device *ubi)
301 {
302         int err;
303
304         spin_lock(&ubi->wl_lock);
305         while (!ubi->free.rb_node) {
306                 spin_unlock(&ubi->wl_lock);
307
308                 dbg_wl("do one work synchronously");
309                 err = do_work(ubi);
310                 if (err)
311                         return err;
312
313                 spin_lock(&ubi->wl_lock);
314         }
315         spin_unlock(&ubi->wl_lock);
316
317         return 0;
318 }
319
320 /**
321  * in_wl_tree - check if wear-leveling entry is present in a WL RB-tree.
322  * @e: the wear-leveling entry to check
323  * @root: the root of the tree
324  *
325  * This function returns non-zero if @e is in the @root RB-tree and zero if it
326  * is not.
327  */
328 static int in_wl_tree(struct ubi_wl_entry *e, struct rb_root *root)
329 {
330         struct rb_node *p;
331
332         p = root->rb_node;
333         while (p) {
334                 struct ubi_wl_entry *e1;
335
336                 e1 = rb_entry(p, struct ubi_wl_entry, rb);
337
338                 if (e->pnum == e1->pnum) {
339                         ubi_assert(e == e1);
340                         return 1;
341                 }
342
343                 if (e->ec < e1->ec)
344                         p = p->rb_left;
345                 else if (e->ec > e1->ec)
346                         p = p->rb_right;
347                 else {
348                         ubi_assert(e->pnum != e1->pnum);
349                         if (e->pnum < e1->pnum)
350                                 p = p->rb_left;
351                         else
352                                 p = p->rb_right;
353                 }
354         }
355
356         return 0;
357 }
358
359 /**
360  * prot_tree_add - add physical eraseblock to protection trees.
361  * @ubi: UBI device description object
362  * @e: the physical eraseblock to add
363  * @pe: protection entry object to use
364  * @abs_ec: absolute erase counter value when this physical eraseblock has
365  * to be removed from the protection trees.
366  *
367  * @wl->lock has to be locked.
368  */
369 static void prot_tree_add(struct ubi_device *ubi, struct ubi_wl_entry *e,
370                           struct ubi_wl_prot_entry *pe, int abs_ec)
371 {
372         struct rb_node **p, *parent = NULL;
373         struct ubi_wl_prot_entry *pe1;
374
375         pe->e = e;
376         pe->abs_ec = ubi->abs_ec + abs_ec;
377
378         p = &ubi->prot.pnum.rb_node;
379         while (*p) {
380                 parent = *p;
381                 pe1 = rb_entry(parent, struct ubi_wl_prot_entry, rb_pnum);
382
383                 if (e->pnum < pe1->e->pnum)
384                         p = &(*p)->rb_left;
385                 else
386                         p = &(*p)->rb_right;
387         }
388         rb_link_node(&pe->rb_pnum, parent, p);
389         rb_insert_color(&pe->rb_pnum, &ubi->prot.pnum);
390
391         p = &ubi->prot.aec.rb_node;
392         parent = NULL;
393         while (*p) {
394                 parent = *p;
395                 pe1 = rb_entry(parent, struct ubi_wl_prot_entry, rb_aec);
396
397                 if (pe->abs_ec < pe1->abs_ec)
398                         p = &(*p)->rb_left;
399                 else
400                         p = &(*p)->rb_right;
401         }
402         rb_link_node(&pe->rb_aec, parent, p);
403         rb_insert_color(&pe->rb_aec, &ubi->prot.aec);
404 }
405
406 /**
407  * find_wl_entry - find wear-leveling entry closest to certain erase counter.
408  * @root: the RB-tree where to look for
409  * @max: highest possible erase counter
410  *
411  * This function looks for a wear leveling entry with erase counter closest to
412  * @max and less then @max.
413  */
414 static struct ubi_wl_entry *find_wl_entry(struct rb_root *root, int max)
415 {
416         struct rb_node *p;
417         struct ubi_wl_entry *e;
418
419         e = rb_entry(rb_first(root), struct ubi_wl_entry, rb);
420         max += e->ec;
421
422         p = root->rb_node;
423         while (p) {
424                 struct ubi_wl_entry *e1;
425
426                 e1 = rb_entry(p, struct ubi_wl_entry, rb);
427                 if (e1->ec >= max)
428                         p = p->rb_left;
429                 else {
430                         p = p->rb_right;
431                         e = e1;
432                 }
433         }
434
435         return e;
436 }
437
438 /**
439  * ubi_wl_get_peb - get a physical eraseblock.
440  * @ubi: UBI device description object
441  * @dtype: type of data which will be stored in this physical eraseblock
442  *
443  * This function returns a physical eraseblock in case of success and a
444  * negative error code in case of failure. Might sleep.
445  */
446 int ubi_wl_get_peb(struct ubi_device *ubi, int dtype)
447 {
448         int err, protect, medium_ec;
449         struct ubi_wl_entry *e, *first, *last;
450         struct ubi_wl_prot_entry *pe;
451
452         ubi_assert(dtype == UBI_LONGTERM || dtype == UBI_SHORTTERM ||
453                    dtype == UBI_UNKNOWN);
454
455         pe = kmalloc(sizeof(struct ubi_wl_prot_entry), GFP_NOFS);
456         if (!pe)
457                 return -ENOMEM;
458
459 retry:
460         spin_lock(&ubi->wl_lock);
461         if (!ubi->free.rb_node) {
462                 if (ubi->works_count == 0) {
463                         ubi_assert(list_empty(&ubi->works));
464                         ubi_err("no free eraseblocks");
465                         spin_unlock(&ubi->wl_lock);
466                         kfree(pe);
467                         return -ENOSPC;
468                 }
469                 spin_unlock(&ubi->wl_lock);
470
471                 err = produce_free_peb(ubi);
472                 if (err < 0) {
473                         kfree(pe);
474                         return err;
475                 }
476                 goto retry;
477         }
478
479         switch (dtype) {
480                 case UBI_LONGTERM:
481                         /*
482                          * For long term data we pick a physical eraseblock
483                          * with high erase counter. But the highest erase
484                          * counter we can pick is bounded by the the lowest
485                          * erase counter plus %WL_FREE_MAX_DIFF.
486                          */
487                         e = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
488                         protect = LT_PROTECTION;
489                         break;
490                 case UBI_UNKNOWN:
491                         /*
492                          * For unknown data we pick a physical eraseblock with
493                          * medium erase counter. But we by no means can pick a
494                          * physical eraseblock with erase counter greater or
495                          * equivalent than the lowest erase counter plus
496                          * %WL_FREE_MAX_DIFF.
497                          */
498                         first = rb_entry(rb_first(&ubi->free),
499                                          struct ubi_wl_entry, rb);
500                         last = rb_entry(rb_last(&ubi->free),
501                                         struct ubi_wl_entry, rb);
502
503                         if (last->ec - first->ec < WL_FREE_MAX_DIFF)
504                                 e = rb_entry(ubi->free.rb_node,
505                                                 struct ubi_wl_entry, rb);
506                         else {
507                                 medium_ec = (first->ec + WL_FREE_MAX_DIFF)/2;
508                                 e = find_wl_entry(&ubi->free, medium_ec);
509                         }
510                         protect = U_PROTECTION;
511                         break;
512                 case UBI_SHORTTERM:
513                         /*
514                          * For short term data we pick a physical eraseblock
515                          * with the lowest erase counter as we expect it will
516                          * be erased soon.
517                          */
518                         e = rb_entry(rb_first(&ubi->free),
519                                      struct ubi_wl_entry, rb);
520                         protect = ST_PROTECTION;
521                         break;
522                 default:
523                         protect = 0;
524                         e = NULL;
525                         BUG();
526         }
527
528         /*
529          * Move the physical eraseblock to the protection trees where it will
530          * be protected from being moved for some time.
531          */
532         paranoid_check_in_wl_tree(e, &ubi->free);
533         rb_erase(&e->rb, &ubi->free);
534         prot_tree_add(ubi, e, pe, protect);
535
536         dbg_wl("PEB %d EC %d, protection %d", e->pnum, e->ec, protect);
537         spin_unlock(&ubi->wl_lock);
538
539         return e->pnum;
540 }
541
542 /**
543  * prot_tree_del - remove a physical eraseblock from the protection trees
544  * @ubi: UBI device description object
545  * @pnum: the physical eraseblock to remove
546  *
547  * This function returns PEB @pnum from the protection trees and returns zero
548  * in case of success and %-ENODEV if the PEB was not found in the protection
549  * trees.
550  */
551 static int prot_tree_del(struct ubi_device *ubi, int pnum)
552 {
553         struct rb_node *p;
554         struct ubi_wl_prot_entry *pe = NULL;
555
556         p = ubi->prot.pnum.rb_node;
557         while (p) {
558
559                 pe = rb_entry(p, struct ubi_wl_prot_entry, rb_pnum);
560
561                 if (pnum == pe->e->pnum)
562                         goto found;
563
564                 if (pnum < pe->e->pnum)
565                         p = p->rb_left;
566                 else
567                         p = p->rb_right;
568         }
569
570         return -ENODEV;
571
572 found:
573         ubi_assert(pe->e->pnum == pnum);
574         rb_erase(&pe->rb_aec, &ubi->prot.aec);
575         rb_erase(&pe->rb_pnum, &ubi->prot.pnum);
576         kfree(pe);
577         return 0;
578 }
579
580 /**
581  * sync_erase - synchronously erase a physical eraseblock.
582  * @ubi: UBI device description object
583  * @e: the the physical eraseblock to erase
584  * @torture: if the physical eraseblock has to be tortured
585  *
586  * This function returns zero in case of success and a negative error code in
587  * case of failure.
588  */
589 static int sync_erase(struct ubi_device *ubi, struct ubi_wl_entry *e, int torture)
590 {
591         int err;
592         struct ubi_ec_hdr *ec_hdr;
593         unsigned long long ec = e->ec;
594
595         dbg_wl("erase PEB %d, old EC %llu", e->pnum, ec);
596
597         err = paranoid_check_ec(ubi, e->pnum, e->ec);
598         if (err > 0)
599                 return -EINVAL;
600
601         ec_hdr = kzalloc(ubi->ec_hdr_alsize, GFP_NOFS);
602         if (!ec_hdr)
603                 return -ENOMEM;
604
605         err = ubi_io_sync_erase(ubi, e->pnum, torture);
606         if (err < 0)
607                 goto out_free;
608
609         ec += err;
610         if (ec > UBI_MAX_ERASECOUNTER) {
611                 /*
612                  * Erase counter overflow. Upgrade UBI and use 64-bit
613                  * erase counters internally.
614                  */
615                 ubi_err("erase counter overflow at PEB %d, EC %llu",
616                         e->pnum, ec);
617                 err = -EINVAL;
618                 goto out_free;
619         }
620
621         dbg_wl("erased PEB %d, new EC %llu", e->pnum, ec);
622
623         ec_hdr->ec = cpu_to_be64(ec);
624
625         err = ubi_io_write_ec_hdr(ubi, e->pnum, ec_hdr);
626         if (err)
627                 goto out_free;
628
629         e->ec = ec;
630         spin_lock(&ubi->wl_lock);
631         if (e->ec > ubi->max_ec)
632                 ubi->max_ec = e->ec;
633         spin_unlock(&ubi->wl_lock);
634
635 out_free:
636         kfree(ec_hdr);
637         return err;
638 }
639
640 /**
641  * check_protection_over - check if it is time to stop protecting some
642  * physical eraseblocks.
643  * @ubi: UBI device description object
644  *
645  * This function is called after each erase operation, when the absolute erase
646  * counter is incremented, to check if some physical eraseblock  have not to be
647  * protected any longer. These physical eraseblocks are moved from the
648  * protection trees to the used tree.
649  */
650 static void check_protection_over(struct ubi_device *ubi)
651 {
652         struct ubi_wl_prot_entry *pe;
653
654         /*
655          * There may be several protected physical eraseblock to remove,
656          * process them all.
657          */
658         while (1) {
659                 spin_lock(&ubi->wl_lock);
660                 if (!ubi->prot.aec.rb_node) {
661                         spin_unlock(&ubi->wl_lock);
662                         break;
663                 }
664
665                 pe = rb_entry(rb_first(&ubi->prot.aec),
666                               struct ubi_wl_prot_entry, rb_aec);
667
668                 if (pe->abs_ec > ubi->abs_ec) {
669                         spin_unlock(&ubi->wl_lock);
670                         break;
671                 }
672
673                 dbg_wl("PEB %d protection over, abs_ec %llu, PEB abs_ec %llu",
674                        pe->e->pnum, ubi->abs_ec, pe->abs_ec);
675                 rb_erase(&pe->rb_aec, &ubi->prot.aec);
676                 rb_erase(&pe->rb_pnum, &ubi->prot.pnum);
677                 wl_tree_add(pe->e, &ubi->used);
678                 spin_unlock(&ubi->wl_lock);
679
680                 kfree(pe);
681                 cond_resched();
682         }
683 }
684
685 /**
686  * schedule_ubi_work - schedule a work.
687  * @ubi: UBI device description object
688  * @wrk: the work to schedule
689  *
690  * This function enqueues a work defined by @wrk to the tail of the pending
691  * works list.
692  */
693 static void schedule_ubi_work(struct ubi_device *ubi, struct ubi_work *wrk)
694 {
695         spin_lock(&ubi->wl_lock);
696         list_add_tail(&wrk->list, &ubi->works);
697         ubi_assert(ubi->works_count >= 0);
698         ubi->works_count += 1;
699         if (ubi->thread_enabled)
700                 wake_up_process(ubi->bgt_thread);
701         spin_unlock(&ubi->wl_lock);
702 }
703
704 static int erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk,
705                         int cancel);
706
707 /**
708  * schedule_erase - schedule an erase work.
709  * @ubi: UBI device description object
710  * @e: the WL entry of the physical eraseblock to erase
711  * @torture: if the physical eraseblock has to be tortured
712  *
713  * This function returns zero in case of success and a %-ENOMEM in case of
714  * failure.
715  */
716 static int schedule_erase(struct ubi_device *ubi, struct ubi_wl_entry *e,
717                           int torture)
718 {
719         struct ubi_work *wl_wrk;
720
721         dbg_wl("schedule erasure of PEB %d, EC %d, torture %d",
722                e->pnum, e->ec, torture);
723
724         wl_wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
725         if (!wl_wrk)
726                 return -ENOMEM;
727
728         wl_wrk->func = &erase_worker;
729         wl_wrk->e = e;
730         wl_wrk->torture = torture;
731
732         schedule_ubi_work(ubi, wl_wrk);
733         return 0;
734 }
735
736 /**
737  * wear_leveling_worker - wear-leveling worker function.
738  * @ubi: UBI device description object
739  * @wrk: the work object
740  * @cancel: non-zero if the worker has to free memory and exit
741  *
742  * This function copies a more worn out physical eraseblock to a less worn out
743  * one. Returns zero in case of success and a negative error code in case of
744  * failure.
745  */
746 static int wear_leveling_worker(struct ubi_device *ubi, struct ubi_work *wrk,
747                                 int cancel)
748 {
749         int err, put = 0, scrubbing = 0, protect = 0;
750         struct ubi_wl_prot_entry *uninitialized_var(pe);
751         struct ubi_wl_entry *e1, *e2;
752         struct ubi_vid_hdr *vid_hdr;
753
754         kfree(wrk);
755
756         if (cancel)
757                 return 0;
758
759         vid_hdr = ubi_zalloc_vid_hdr(ubi, GFP_NOFS);
760         if (!vid_hdr)
761                 return -ENOMEM;
762
763         mutex_lock(&ubi->move_mutex);
764         spin_lock(&ubi->wl_lock);
765         ubi_assert(!ubi->move_from && !ubi->move_to);
766         ubi_assert(!ubi->move_to_put);
767
768         if (!ubi->free.rb_node ||
769             (!ubi->used.rb_node && !ubi->scrub.rb_node)) {
770                 /*
771                  * No free physical eraseblocks? Well, they must be waiting in
772                  * the queue to be erased. Cancel movement - it will be
773                  * triggered again when a free physical eraseblock appears.
774                  *
775                  * No used physical eraseblocks? They must be temporarily
776                  * protected from being moved. They will be moved to the
777                  * @ubi->used tree later and the wear-leveling will be
778                  * triggered again.
779                  */
780                 dbg_wl("cancel WL, a list is empty: free %d, used %d",
781                        !ubi->free.rb_node, !ubi->used.rb_node);
782                 goto out_cancel;
783         }
784
785         if (!ubi->scrub.rb_node) {
786                 /*
787                  * Now pick the least worn-out used physical eraseblock and a
788                  * highly worn-out free physical eraseblock. If the erase
789                  * counters differ much enough, start wear-leveling.
790                  */
791                 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, rb);
792                 e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
793
794                 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD)) {
795                         dbg_wl("no WL needed: min used EC %d, max free EC %d",
796                                e1->ec, e2->ec);
797                         goto out_cancel;
798                 }
799                 paranoid_check_in_wl_tree(e1, &ubi->used);
800                 rb_erase(&e1->rb, &ubi->used);
801                 dbg_wl("move PEB %d EC %d to PEB %d EC %d",
802                        e1->pnum, e1->ec, e2->pnum, e2->ec);
803         } else {
804                 /* Perform scrubbing */
805                 scrubbing = 1;
806                 e1 = rb_entry(rb_first(&ubi->scrub), struct ubi_wl_entry, rb);
807                 e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
808                 paranoid_check_in_wl_tree(e1, &ubi->scrub);
809                 rb_erase(&e1->rb, &ubi->scrub);
810                 dbg_wl("scrub PEB %d to PEB %d", e1->pnum, e2->pnum);
811         }
812
813         paranoid_check_in_wl_tree(e2, &ubi->free);
814         rb_erase(&e2->rb, &ubi->free);
815         ubi->move_from = e1;
816         ubi->move_to = e2;
817         spin_unlock(&ubi->wl_lock);
818
819         /*
820          * Now we are going to copy physical eraseblock @e1->pnum to @e2->pnum.
821          * We so far do not know which logical eraseblock our physical
822          * eraseblock (@e1) belongs to. We have to read the volume identifier
823          * header first.
824          *
825          * Note, we are protected from this PEB being unmapped and erased. The
826          * 'ubi_wl_put_peb()' would wait for moving to be finished if the PEB
827          * which is being moved was unmapped.
828          */
829
830         err = ubi_io_read_vid_hdr(ubi, e1->pnum, vid_hdr, 0);
831         if (err && err != UBI_IO_BITFLIPS) {
832                 if (err == UBI_IO_PEB_FREE) {
833                         /*
834                          * We are trying to move PEB without a VID header. UBI
835                          * always write VID headers shortly after the PEB was
836                          * given, so we have a situation when it did not have
837                          * chance to write it down because it was preempted.
838                          * Just re-schedule the work, so that next time it will
839                          * likely have the VID header in place.
840                          */
841                         dbg_wl("PEB %d has no VID header", e1->pnum);
842                         goto out_not_moved;
843                 }
844
845                 ubi_err("error %d while reading VID header from PEB %d",
846                         err, e1->pnum);
847                 if (err > 0)
848                         err = -EIO;
849                 goto out_error;
850         }
851
852         err = ubi_eba_copy_leb(ubi, e1->pnum, e2->pnum, vid_hdr);
853         if (err) {
854
855                 if (err < 0)
856                         goto out_error;
857                 if (err == 1)
858                         goto out_not_moved;
859
860                 /*
861                  * For some reason the LEB was not moved - it might be because
862                  * the volume is being deleted. We should prevent this PEB from
863                  * being selected for wear-levelling movement for some "time",
864                  * so put it to the protection tree.
865                  */
866
867                 dbg_wl("cancelled moving PEB %d", e1->pnum);
868                 pe = kmalloc(sizeof(struct ubi_wl_prot_entry), GFP_NOFS);
869                 if (!pe) {
870                         err = -ENOMEM;
871                         goto out_error;
872                 }
873
874                 protect = 1;
875         }
876
877         ubi_free_vid_hdr(ubi, vid_hdr);
878         spin_lock(&ubi->wl_lock);
879         if (protect)
880                 prot_tree_add(ubi, e1, pe, protect);
881         if (!ubi->move_to_put)
882                 wl_tree_add(e2, &ubi->used);
883         else
884                 put = 1;
885         ubi->move_from = ubi->move_to = NULL;
886         ubi->move_to_put = ubi->wl_scheduled = 0;
887         spin_unlock(&ubi->wl_lock);
888
889         if (put) {
890                 /*
891                  * Well, the target PEB was put meanwhile, schedule it for
892                  * erasure.
893                  */
894                 dbg_wl("PEB %d was put meanwhile, erase", e2->pnum);
895                 err = schedule_erase(ubi, e2, 0);
896                 if (err)
897                         goto out_error;
898         }
899
900         if (!protect) {
901                 err = schedule_erase(ubi, e1, 0);
902                 if (err)
903                         goto out_error;
904         }
905
906
907         dbg_wl("done");
908         mutex_unlock(&ubi->move_mutex);
909         return 0;
910
911         /*
912          * For some reasons the LEB was not moved, might be an error, might be
913          * something else. @e1 was not changed, so return it back. @e2 might
914          * be changed, schedule it for erasure.
915          */
916 out_not_moved:
917         ubi_free_vid_hdr(ubi, vid_hdr);
918         spin_lock(&ubi->wl_lock);
919         if (scrubbing)
920                 wl_tree_add(e1, &ubi->scrub);
921         else
922                 wl_tree_add(e1, &ubi->used);
923         ubi->move_from = ubi->move_to = NULL;
924         ubi->move_to_put = ubi->wl_scheduled = 0;
925         spin_unlock(&ubi->wl_lock);
926
927         err = schedule_erase(ubi, e2, 0);
928         if (err)
929                 goto out_error;
930
931         mutex_unlock(&ubi->move_mutex);
932         return 0;
933
934 out_error:
935         ubi_err("error %d while moving PEB %d to PEB %d",
936                 err, e1->pnum, e2->pnum);
937
938         ubi_free_vid_hdr(ubi, vid_hdr);
939         spin_lock(&ubi->wl_lock);
940         ubi->move_from = ubi->move_to = NULL;
941         ubi->move_to_put = ubi->wl_scheduled = 0;
942         spin_unlock(&ubi->wl_lock);
943
944         kmem_cache_free(ubi_wl_entry_slab, e1);
945         kmem_cache_free(ubi_wl_entry_slab, e2);
946         ubi_ro_mode(ubi);
947
948         mutex_unlock(&ubi->move_mutex);
949         return err;
950
951 out_cancel:
952         ubi->wl_scheduled = 0;
953         spin_unlock(&ubi->wl_lock);
954         mutex_unlock(&ubi->move_mutex);
955         ubi_free_vid_hdr(ubi, vid_hdr);
956         return 0;
957 }
958
959 /**
960  * ensure_wear_leveling - schedule wear-leveling if it is needed.
961  * @ubi: UBI device description object
962  *
963  * This function checks if it is time to start wear-leveling and schedules it
964  * if yes. This function returns zero in case of success and a negative error
965  * code in case of failure.
966  */
967 static int ensure_wear_leveling(struct ubi_device *ubi)
968 {
969         int err = 0;
970         struct ubi_wl_entry *e1;
971         struct ubi_wl_entry *e2;
972         struct ubi_work *wrk;
973
974         spin_lock(&ubi->wl_lock);
975         if (ubi->wl_scheduled)
976                 /* Wear-leveling is already in the work queue */
977                 goto out_unlock;
978
979         /*
980          * If the ubi->scrub tree is not empty, scrubbing is needed, and the
981          * the WL worker has to be scheduled anyway.
982          */
983         if (!ubi->scrub.rb_node) {
984                 if (!ubi->used.rb_node || !ubi->free.rb_node)
985                         /* No physical eraseblocks - no deal */
986                         goto out_unlock;
987
988                 /*
989                  * We schedule wear-leveling only if the difference between the
990                  * lowest erase counter of used physical eraseblocks and a high
991                  * erase counter of free physical eraseblocks is greater then
992                  * %UBI_WL_THRESHOLD.
993                  */
994                 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, rb);
995                 e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
996
997                 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD))
998                         goto out_unlock;
999                 dbg_wl("schedule wear-leveling");
1000         } else
1001                 dbg_wl("schedule scrubbing");
1002
1003         ubi->wl_scheduled = 1;
1004         spin_unlock(&ubi->wl_lock);
1005
1006         wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
1007         if (!wrk) {
1008                 err = -ENOMEM;
1009                 goto out_cancel;
1010         }
1011
1012         wrk->func = &wear_leveling_worker;
1013         schedule_ubi_work(ubi, wrk);
1014         return err;
1015
1016 out_cancel:
1017         spin_lock(&ubi->wl_lock);
1018         ubi->wl_scheduled = 0;
1019 out_unlock:
1020         spin_unlock(&ubi->wl_lock);
1021         return err;
1022 }
1023
1024 /**
1025  * erase_worker - physical eraseblock erase worker function.
1026  * @ubi: UBI device description object
1027  * @wl_wrk: the work object
1028  * @cancel: non-zero if the worker has to free memory and exit
1029  *
1030  * This function erases a physical eraseblock and perform torture testing if
1031  * needed. It also takes care about marking the physical eraseblock bad if
1032  * needed. Returns zero in case of success and a negative error code in case of
1033  * failure.
1034  */
1035 static int erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk,
1036                         int cancel)
1037 {
1038         struct ubi_wl_entry *e = wl_wrk->e;
1039         int pnum = e->pnum, err, need;
1040
1041         if (cancel) {
1042                 dbg_wl("cancel erasure of PEB %d EC %d", pnum, e->ec);
1043                 kfree(wl_wrk);
1044                 kmem_cache_free(ubi_wl_entry_slab, e);
1045                 return 0;
1046         }
1047
1048         dbg_wl("erase PEB %d EC %d", pnum, e->ec);
1049
1050         err = sync_erase(ubi, e, wl_wrk->torture);
1051         if (!err) {
1052                 /* Fine, we've erased it successfully */
1053                 kfree(wl_wrk);
1054
1055                 spin_lock(&ubi->wl_lock);
1056                 ubi->abs_ec += 1;
1057                 wl_tree_add(e, &ubi->free);
1058                 spin_unlock(&ubi->wl_lock);
1059
1060                 /*
1061                  * One more erase operation has happened, take care about protected
1062                  * physical eraseblocks.
1063                  */
1064                 check_protection_over(ubi);
1065
1066                 /* And take care about wear-leveling */
1067                 err = ensure_wear_leveling(ubi);
1068                 return err;
1069         }
1070
1071         ubi_err("failed to erase PEB %d, error %d", pnum, err);
1072         kfree(wl_wrk);
1073         kmem_cache_free(ubi_wl_entry_slab, e);
1074
1075         if (err == -EINTR || err == -ENOMEM || err == -EAGAIN ||
1076             err == -EBUSY) {
1077                 int err1;
1078
1079                 /* Re-schedule the LEB for erasure */
1080                 err1 = schedule_erase(ubi, e, 0);
1081                 if (err1) {
1082                         err = err1;
1083                         goto out_ro;
1084                 }
1085                 return err;
1086         } else if (err != -EIO) {
1087                 /*
1088                  * If this is not %-EIO, we have no idea what to do. Scheduling
1089                  * this physical eraseblock for erasure again would cause
1090                  * errors again and again. Well, lets switch to RO mode.
1091                  */
1092                 goto out_ro;
1093         }
1094
1095         /* It is %-EIO, the PEB went bad */
1096
1097         if (!ubi->bad_allowed) {
1098                 ubi_err("bad physical eraseblock %d detected", pnum);
1099                 goto out_ro;
1100         }
1101
1102         spin_lock(&ubi->volumes_lock);
1103         need = ubi->beb_rsvd_level - ubi->beb_rsvd_pebs + 1;
1104         if (need > 0) {
1105                 need = ubi->avail_pebs >= need ? need : ubi->avail_pebs;
1106                 ubi->avail_pebs -= need;
1107                 ubi->rsvd_pebs += need;
1108                 ubi->beb_rsvd_pebs += need;
1109                 if (need > 0)
1110                         ubi_msg("reserve more %d PEBs", need);
1111         }
1112
1113         if (ubi->beb_rsvd_pebs == 0) {
1114                 spin_unlock(&ubi->volumes_lock);
1115                 ubi_err("no reserved physical eraseblocks");
1116                 goto out_ro;
1117         }
1118
1119         spin_unlock(&ubi->volumes_lock);
1120         ubi_msg("mark PEB %d as bad", pnum);
1121
1122         err = ubi_io_mark_bad(ubi, pnum);
1123         if (err)
1124                 goto out_ro;
1125
1126         spin_lock(&ubi->volumes_lock);
1127         ubi->beb_rsvd_pebs -= 1;
1128         ubi->bad_peb_count += 1;
1129         ubi->good_peb_count -= 1;
1130         ubi_calculate_reserved(ubi);
1131         if (ubi->beb_rsvd_pebs == 0)
1132                 ubi_warn("last PEB from the reserved pool was used");
1133         spin_unlock(&ubi->volumes_lock);
1134
1135         return err;
1136
1137 out_ro:
1138         ubi_ro_mode(ubi);
1139         return err;
1140 }
1141
1142 /**
1143  * ubi_wl_put_peb - return a physical eraseblock to the wear-leveling unit.
1144  * @ubi: UBI device description object
1145  * @pnum: physical eraseblock to return
1146  * @torture: if this physical eraseblock has to be tortured
1147  *
1148  * This function is called to return physical eraseblock @pnum to the pool of
1149  * free physical eraseblocks. The @torture flag has to be set if an I/O error
1150  * occurred to this @pnum and it has to be tested. This function returns zero
1151  * in case of success, and a negative error code in case of failure.
1152  */
1153 int ubi_wl_put_peb(struct ubi_device *ubi, int pnum, int torture)
1154 {
1155         int err;
1156         struct ubi_wl_entry *e;
1157
1158         dbg_wl("PEB %d", pnum);
1159         ubi_assert(pnum >= 0);
1160         ubi_assert(pnum < ubi->peb_count);
1161
1162 retry:
1163         spin_lock(&ubi->wl_lock);
1164         e = ubi->lookuptbl[pnum];
1165         if (e == ubi->move_from) {
1166                 /*
1167                  * User is putting the physical eraseblock which was selected to
1168                  * be moved. It will be scheduled for erasure in the
1169                  * wear-leveling worker.
1170                  */
1171                 dbg_wl("PEB %d is being moved, wait", pnum);
1172                 spin_unlock(&ubi->wl_lock);
1173
1174                 /* Wait for the WL worker by taking the @ubi->move_mutex */
1175                 mutex_lock(&ubi->move_mutex);
1176                 mutex_unlock(&ubi->move_mutex);
1177                 goto retry;
1178         } else if (e == ubi->move_to) {
1179                 /*
1180                  * User is putting the physical eraseblock which was selected
1181                  * as the target the data is moved to. It may happen if the EBA
1182                  * unit already re-mapped the LEB in 'ubi_eba_copy_leb()' but
1183                  * the WL unit has not put the PEB to the "used" tree yet, but
1184                  * it is about to do this. So we just set a flag which will
1185                  * tell the WL worker that the PEB is not needed anymore and
1186                  * should be scheduled for erasure.
1187                  */
1188                 dbg_wl("PEB %d is the target of data moving", pnum);
1189                 ubi_assert(!ubi->move_to_put);
1190                 ubi->move_to_put = 1;
1191                 spin_unlock(&ubi->wl_lock);
1192                 return 0;
1193         } else {
1194                 if (in_wl_tree(e, &ubi->used)) {
1195                         paranoid_check_in_wl_tree(e, &ubi->used);
1196                         rb_erase(&e->rb, &ubi->used);
1197                 } else if (in_wl_tree(e, &ubi->scrub)) {
1198                         paranoid_check_in_wl_tree(e, &ubi->scrub);
1199                         rb_erase(&e->rb, &ubi->scrub);
1200                 } else {
1201                         err = prot_tree_del(ubi, e->pnum);
1202                         if (err) {
1203                                 ubi_err("PEB %d not found", pnum);
1204                                 ubi_ro_mode(ubi);
1205                                 spin_unlock(&ubi->wl_lock);
1206                                 return err;
1207                         }
1208                 }
1209         }
1210         spin_unlock(&ubi->wl_lock);
1211
1212         err = schedule_erase(ubi, e, torture);
1213         if (err) {
1214                 spin_lock(&ubi->wl_lock);
1215                 wl_tree_add(e, &ubi->used);
1216                 spin_unlock(&ubi->wl_lock);
1217         }
1218
1219         return err;
1220 }
1221
1222 /**
1223  * ubi_wl_scrub_peb - schedule a physical eraseblock for scrubbing.
1224  * @ubi: UBI device description object
1225  * @pnum: the physical eraseblock to schedule
1226  *
1227  * If a bit-flip in a physical eraseblock is detected, this physical eraseblock
1228  * needs scrubbing. This function schedules a physical eraseblock for
1229  * scrubbing which is done in background. This function returns zero in case of
1230  * success and a negative error code in case of failure.
1231  */
1232 int ubi_wl_scrub_peb(struct ubi_device *ubi, int pnum)
1233 {
1234         struct ubi_wl_entry *e;
1235
1236         ubi_msg("schedule PEB %d for scrubbing", pnum);
1237
1238 retry:
1239         spin_lock(&ubi->wl_lock);
1240         e = ubi->lookuptbl[pnum];
1241         if (e == ubi->move_from || in_wl_tree(e, &ubi->scrub)) {
1242                 spin_unlock(&ubi->wl_lock);
1243                 return 0;
1244         }
1245
1246         if (e == ubi->move_to) {
1247                 /*
1248                  * This physical eraseblock was used to move data to. The data
1249                  * was moved but the PEB was not yet inserted to the proper
1250                  * tree. We should just wait a little and let the WL worker
1251                  * proceed.
1252                  */
1253                 spin_unlock(&ubi->wl_lock);
1254                 dbg_wl("the PEB %d is not in proper tree, retry", pnum);
1255                 yield();
1256                 goto retry;
1257         }
1258
1259         if (in_wl_tree(e, &ubi->used)) {
1260                 paranoid_check_in_wl_tree(e, &ubi->used);
1261                 rb_erase(&e->rb, &ubi->used);
1262         } else {
1263                 int err;
1264
1265                 err = prot_tree_del(ubi, e->pnum);
1266                 if (err) {
1267                         ubi_err("PEB %d not found", pnum);
1268                         ubi_ro_mode(ubi);
1269                         spin_unlock(&ubi->wl_lock);
1270                         return err;
1271                 }
1272         }
1273
1274         wl_tree_add(e, &ubi->scrub);
1275         spin_unlock(&ubi->wl_lock);
1276
1277         /*
1278          * Technically scrubbing is the same as wear-leveling, so it is done
1279          * by the WL worker.
1280          */
1281         return ensure_wear_leveling(ubi);
1282 }
1283
1284 /**
1285  * ubi_wl_flush - flush all pending works.
1286  * @ubi: UBI device description object
1287  *
1288  * This function returns zero in case of success and a negative error code in
1289  * case of failure.
1290  */
1291 int ubi_wl_flush(struct ubi_device *ubi)
1292 {
1293         int err;
1294
1295         /*
1296          * Erase while the pending works queue is not empty, but not more then
1297          * the number of currently pending works.
1298          */
1299         dbg_wl("flush (%d pending works)", ubi->works_count);
1300         while (ubi->works_count) {
1301                 err = do_work(ubi);
1302                 if (err)
1303                         return err;
1304         }
1305
1306         /*
1307          * Make sure all the works which have been done in parallel are
1308          * finished.
1309          */
1310         down_write(&ubi->work_sem);
1311         up_write(&ubi->work_sem);
1312
1313         /*
1314          * And in case last was the WL worker and it cancelled the LEB
1315          * movement, flush again.
1316          */
1317         while (ubi->works_count) {
1318                 dbg_wl("flush more (%d pending works)", ubi->works_count);
1319                 err = do_work(ubi);
1320                 if (err)
1321                         return err;
1322         }
1323
1324         return 0;
1325 }
1326
1327 /**
1328  * tree_destroy - destroy an RB-tree.
1329  * @root: the root of the tree to destroy
1330  */
1331 static void tree_destroy(struct rb_root *root)
1332 {
1333         struct rb_node *rb;
1334         struct ubi_wl_entry *e;
1335
1336         rb = root->rb_node;
1337         while (rb) {
1338                 if (rb->rb_left)
1339                         rb = rb->rb_left;
1340                 else if (rb->rb_right)
1341                         rb = rb->rb_right;
1342                 else {
1343                         e = rb_entry(rb, struct ubi_wl_entry, rb);
1344
1345                         rb = rb_parent(rb);
1346                         if (rb) {
1347                                 if (rb->rb_left == &e->rb)
1348                                         rb->rb_left = NULL;
1349                                 else
1350                                         rb->rb_right = NULL;
1351                         }
1352
1353                         kmem_cache_free(ubi_wl_entry_slab, e);
1354                 }
1355         }
1356 }
1357
1358 /**
1359  * ubi_thread - UBI background thread.
1360  * @u: the UBI device description object pointer
1361  */
1362 int ubi_thread(void *u)
1363 {
1364         int failures = 0;
1365         struct ubi_device *ubi = u;
1366
1367         ubi_msg("background thread \"%s\" started, PID %d",
1368                 ubi->bgt_name, task_pid_nr(current));
1369
1370         set_freezable();
1371         for (;;) {
1372                 int err;
1373
1374                 if (kthread_should_stop())
1375                         break;
1376
1377                 if (try_to_freeze())
1378                         continue;
1379
1380                 spin_lock(&ubi->wl_lock);
1381                 if (list_empty(&ubi->works) || ubi->ro_mode ||
1382                                !ubi->thread_enabled) {
1383                         set_current_state(TASK_INTERRUPTIBLE);
1384                         spin_unlock(&ubi->wl_lock);
1385                         schedule();
1386                         continue;
1387                 }
1388                 spin_unlock(&ubi->wl_lock);
1389
1390                 err = do_work(ubi);
1391                 if (err) {
1392                         ubi_err("%s: work failed with error code %d",
1393                                 ubi->bgt_name, err);
1394                         if (failures++ > WL_MAX_FAILURES) {
1395                                 /*
1396                                  * Too many failures, disable the thread and
1397                                  * switch to read-only mode.
1398                                  */
1399                                 ubi_msg("%s: %d consecutive failures",
1400                                         ubi->bgt_name, WL_MAX_FAILURES);
1401                                 ubi_ro_mode(ubi);
1402                                 break;
1403                         }
1404                 } else
1405                         failures = 0;
1406
1407                 cond_resched();
1408         }
1409
1410         dbg_wl("background thread \"%s\" is killed", ubi->bgt_name);
1411         return 0;
1412 }
1413
1414 /**
1415  * cancel_pending - cancel all pending works.
1416  * @ubi: UBI device description object
1417  */
1418 static void cancel_pending(struct ubi_device *ubi)
1419 {
1420         while (!list_empty(&ubi->works)) {
1421                 struct ubi_work *wrk;
1422
1423                 wrk = list_entry(ubi->works.next, struct ubi_work, list);
1424                 list_del(&wrk->list);
1425                 wrk->func(ubi, wrk, 1);
1426                 ubi->works_count -= 1;
1427                 ubi_assert(ubi->works_count >= 0);
1428         }
1429 }
1430
1431 /**
1432  * ubi_wl_init_scan - initialize the wear-leveling unit using scanning
1433  * information.
1434  * @ubi: UBI device description object
1435  * @si: scanning information
1436  *
1437  * This function returns zero in case of success, and a negative error code in
1438  * case of failure.
1439  */
1440 int ubi_wl_init_scan(struct ubi_device *ubi, struct ubi_scan_info *si)
1441 {
1442         int err;
1443         struct rb_node *rb1, *rb2;
1444         struct ubi_scan_volume *sv;
1445         struct ubi_scan_leb *seb, *tmp;
1446         struct ubi_wl_entry *e;
1447
1448
1449         ubi->used = ubi->free = ubi->scrub = RB_ROOT;
1450         ubi->prot.pnum = ubi->prot.aec = RB_ROOT;
1451         spin_lock_init(&ubi->wl_lock);
1452         mutex_init(&ubi->move_mutex);
1453         init_rwsem(&ubi->work_sem);
1454         ubi->max_ec = si->max_ec;
1455         INIT_LIST_HEAD(&ubi->works);
1456
1457         sprintf(ubi->bgt_name, UBI_BGT_NAME_PATTERN, ubi->ubi_num);
1458
1459         err = -ENOMEM;
1460         ubi->lookuptbl = kzalloc(ubi->peb_count * sizeof(void *), GFP_KERNEL);
1461         if (!ubi->lookuptbl)
1462                 return err;
1463
1464         list_for_each_entry_safe(seb, tmp, &si->erase, u.list) {
1465                 cond_resched();
1466
1467                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1468                 if (!e)
1469                         goto out_free;
1470
1471                 e->pnum = seb->pnum;
1472                 e->ec = seb->ec;
1473                 ubi->lookuptbl[e->pnum] = e;
1474                 if (schedule_erase(ubi, e, 0)) {
1475                         kmem_cache_free(ubi_wl_entry_slab, e);
1476                         goto out_free;
1477                 }
1478         }
1479
1480         list_for_each_entry(seb, &si->free, u.list) {
1481                 cond_resched();
1482
1483                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1484                 if (!e)
1485                         goto out_free;
1486
1487                 e->pnum = seb->pnum;
1488                 e->ec = seb->ec;
1489                 ubi_assert(e->ec >= 0);
1490                 wl_tree_add(e, &ubi->free);
1491                 ubi->lookuptbl[e->pnum] = e;
1492         }
1493
1494         list_for_each_entry(seb, &si->corr, u.list) {
1495                 cond_resched();
1496
1497                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1498                 if (!e)
1499                         goto out_free;
1500
1501                 e->pnum = seb->pnum;
1502                 e->ec = seb->ec;
1503                 ubi->lookuptbl[e->pnum] = e;
1504                 if (schedule_erase(ubi, e, 0)) {
1505                         kmem_cache_free(ubi_wl_entry_slab, e);
1506                         goto out_free;
1507                 }
1508         }
1509
1510         ubi_rb_for_each_entry(rb1, sv, &si->volumes, rb) {
1511                 ubi_rb_for_each_entry(rb2, seb, &sv->root, u.rb) {
1512                         cond_resched();
1513
1514                         e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1515                         if (!e)
1516                                 goto out_free;
1517
1518                         e->pnum = seb->pnum;
1519                         e->ec = seb->ec;
1520                         ubi->lookuptbl[e->pnum] = e;
1521                         if (!seb->scrub) {
1522                                 dbg_wl("add PEB %d EC %d to the used tree",
1523                                        e->pnum, e->ec);
1524                                 wl_tree_add(e, &ubi->used);
1525                         } else {
1526                                 dbg_wl("add PEB %d EC %d to the scrub tree",
1527                                        e->pnum, e->ec);
1528                                 wl_tree_add(e, &ubi->scrub);
1529                         }
1530                 }
1531         }
1532
1533         if (ubi->avail_pebs < WL_RESERVED_PEBS) {
1534                 ubi_err("no enough physical eraseblocks (%d, need %d)",
1535                         ubi->avail_pebs, WL_RESERVED_PEBS);
1536                 goto out_free;
1537         }
1538         ubi->avail_pebs -= WL_RESERVED_PEBS;
1539         ubi->rsvd_pebs += WL_RESERVED_PEBS;
1540
1541         /* Schedule wear-leveling if needed */
1542         err = ensure_wear_leveling(ubi);
1543         if (err)
1544                 goto out_free;
1545
1546         return 0;
1547
1548 out_free:
1549         cancel_pending(ubi);
1550         tree_destroy(&ubi->used);
1551         tree_destroy(&ubi->free);
1552         tree_destroy(&ubi->scrub);
1553         kfree(ubi->lookuptbl);
1554         return err;
1555 }
1556
1557 /**
1558  * protection_trees_destroy - destroy the protection RB-trees.
1559  * @ubi: UBI device description object
1560  */
1561 static void protection_trees_destroy(struct ubi_device *ubi)
1562 {
1563         struct rb_node *rb;
1564         struct ubi_wl_prot_entry *pe;
1565
1566         rb = ubi->prot.aec.rb_node;
1567         while (rb) {
1568                 if (rb->rb_left)
1569                         rb = rb->rb_left;
1570                 else if (rb->rb_right)
1571                         rb = rb->rb_right;
1572                 else {
1573                         pe = rb_entry(rb, struct ubi_wl_prot_entry, rb_aec);
1574
1575                         rb = rb_parent(rb);
1576                         if (rb) {
1577                                 if (rb->rb_left == &pe->rb_aec)
1578                                         rb->rb_left = NULL;
1579                                 else
1580                                         rb->rb_right = NULL;
1581                         }
1582
1583                         kmem_cache_free(ubi_wl_entry_slab, pe->e);
1584                         kfree(pe);
1585                 }
1586         }
1587 }
1588
1589 /**
1590  * ubi_wl_close - close the wear-leveling unit.
1591  * @ubi: UBI device description object
1592  */
1593 void ubi_wl_close(struct ubi_device *ubi)
1594 {
1595         dbg_wl("close the UBI wear-leveling unit");
1596
1597         cancel_pending(ubi);
1598         protection_trees_destroy(ubi);
1599         tree_destroy(&ubi->used);
1600         tree_destroy(&ubi->free);
1601         tree_destroy(&ubi->scrub);
1602         kfree(ubi->lookuptbl);
1603 }
1604
1605 #ifdef CONFIG_MTD_UBI_DEBUG_PARANOID
1606
1607 /**
1608  * paranoid_check_ec - make sure that the erase counter of a physical eraseblock
1609  * is correct.
1610  * @ubi: UBI device description object
1611  * @pnum: the physical eraseblock number to check
1612  * @ec: the erase counter to check
1613  *
1614  * This function returns zero if the erase counter of physical eraseblock @pnum
1615  * is equivalent to @ec, %1 if not, and a negative error code if an error
1616  * occurred.
1617  */
1618 static int paranoid_check_ec(struct ubi_device *ubi, int pnum, int ec)
1619 {
1620         int err;
1621         long long read_ec;
1622         struct ubi_ec_hdr *ec_hdr;
1623
1624         ec_hdr = kzalloc(ubi->ec_hdr_alsize, GFP_NOFS);
1625         if (!ec_hdr)
1626                 return -ENOMEM;
1627
1628         err = ubi_io_read_ec_hdr(ubi, pnum, ec_hdr, 0);
1629         if (err && err != UBI_IO_BITFLIPS) {
1630                 /* The header does not have to exist */
1631                 err = 0;
1632                 goto out_free;
1633         }
1634
1635         read_ec = be64_to_cpu(ec_hdr->ec);
1636         if (ec != read_ec) {
1637                 ubi_err("paranoid check failed for PEB %d", pnum);
1638                 ubi_err("read EC is %lld, should be %d", read_ec, ec);
1639                 ubi_dbg_dump_stack();
1640                 err = 1;
1641         } else
1642                 err = 0;
1643
1644 out_free:
1645         kfree(ec_hdr);
1646         return err;
1647 }
1648
1649 /**
1650  * paranoid_check_in_wl_tree - make sure that a wear-leveling entry is present
1651  * in a WL RB-tree.
1652  * @e: the wear-leveling entry to check
1653  * @root: the root of the tree
1654  *
1655  * This function returns zero if @e is in the @root RB-tree and %1 if it
1656  * is not.
1657  */
1658 static int paranoid_check_in_wl_tree(struct ubi_wl_entry *e,
1659                                      struct rb_root *root)
1660 {
1661         if (in_wl_tree(e, root))
1662                 return 0;
1663
1664         ubi_err("paranoid check failed for PEB %d, EC %d, RB-tree %p ",
1665                 e->pnum, e->ec, root);
1666         ubi_dbg_dump_stack();
1667         return 1;
1668 }
1669
1670 #endif /* CONFIG_MTD_UBI_DEBUG_PARANOID */