This patch contains the code changes to check the commit id in
[kvmfornfv.git] / kernel / 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 sub-system.
23  *
24  * This sub-system is responsible for wear-leveling. It works in terms of
25  * physical eraseblocks and erase counters and knows nothing about logical
26  * eraseblocks, volumes, etc. From this sub-system's perspective all physical
27  * eraseblocks are of two types - used and free. Used physical eraseblocks are
28  * those that were "get" by the 'ubi_wl_get_peb()' function, and free physical
29  * eraseblocks are 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 sub-system 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 sub-system.
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  * If the WL sub-system fails to erase a physical eraseblock, it marks it as
44  * bad.
45  *
46  * This sub-system is also responsible for scrubbing. If a bit-flip is detected
47  * in a physical eraseblock, it has to be moved. Technically this is the same
48  * as moving it for wear-leveling reasons.
49  *
50  * As it was said, for the UBI sub-system all physical eraseblocks are either
51  * "free" or "used". Free eraseblock are kept in the @wl->free RB-tree, while
52  * used eraseblocks are kept in @wl->used, @wl->erroneous, or @wl->scrub
53  * RB-trees, as well as (temporarily) in the @wl->pq queue.
54  *
55  * When the WL sub-system returns a physical eraseblock, the physical
56  * eraseblock is protected from being moved for some "time". For this reason,
57  * the physical eraseblock is not directly moved from the @wl->free tree to the
58  * @wl->used tree. There is a protection queue in between where this
59  * physical eraseblock is temporarily stored (@wl->pq).
60  *
61  * All this protection stuff is needed because:
62  *  o we don't want to move physical eraseblocks just after we have given them
63  *    to the user; instead, we first want to let users fill them up with data;
64  *
65  *  o there is a chance that the user will put the physical eraseblock very
66  *    soon, so it makes sense not to move it for some time, but wait.
67  *
68  * Physical eraseblocks stay protected only for limited time. But the "time" is
69  * measured in erase cycles in this case. This is implemented with help of the
70  * protection queue. Eraseblocks are put to the tail of this queue when they
71  * are returned by the 'ubi_wl_get_peb()', and eraseblocks are removed from the
72  * head of the queue on each erase operation (for any eraseblock). So the
73  * length of the queue defines how may (global) erase cycles PEBs are protected.
74  *
75  * To put it differently, each physical eraseblock has 2 main states: free and
76  * used. The former state corresponds to the @wl->free tree. The latter state
77  * is split up on several sub-states:
78  * o the WL movement is allowed (@wl->used tree);
79  * o the WL movement is disallowed (@wl->erroneous) because the PEB is
80  *   erroneous - e.g., there was a read error;
81  * o the WL movement is temporarily prohibited (@wl->pq queue);
82  * o scrubbing is needed (@wl->scrub tree).
83  *
84  * Depending on the sub-state, wear-leveling entries of the used physical
85  * eraseblocks may be kept in one of those structures.
86  *
87  * Note, in this implementation, we keep a small in-RAM object for each physical
88  * eraseblock. This is surely not a scalable solution. But it appears to be good
89  * enough for moderately large flashes and it is simple. In future, one may
90  * re-work this sub-system and make it more scalable.
91  *
92  * At the moment this sub-system does not utilize the sequence number, which
93  * was introduced relatively recently. But it would be wise to do this because
94  * the sequence number of a logical eraseblock characterizes how old is it. For
95  * example, when we move a PEB with low erase counter, and we need to pick the
96  * target PEB, we pick a PEB with the highest EC if our PEB is "old" and we
97  * pick target PEB with an average EC if our PEB is not very "old". This is a
98  * room for future re-works of the WL sub-system.
99  */
100
101 #include <linux/slab.h>
102 #include <linux/crc32.h>
103 #include <linux/freezer.h>
104 #include <linux/kthread.h>
105 #include "ubi.h"
106 #include "wl.h"
107
108 /* Number of physical eraseblocks reserved for wear-leveling purposes */
109 #define WL_RESERVED_PEBS 1
110
111 /*
112  * Maximum difference between two erase counters. If this threshold is
113  * exceeded, the WL sub-system starts moving data from used physical
114  * eraseblocks with low erase counter to free physical eraseblocks with high
115  * erase counter.
116  */
117 #define UBI_WL_THRESHOLD CONFIG_MTD_UBI_WL_THRESHOLD
118
119 /*
120  * When a physical eraseblock is moved, the WL sub-system has to pick the target
121  * physical eraseblock to move to. The simplest way would be just to pick the
122  * one with the highest erase counter. But in certain workloads this could lead
123  * to an unlimited wear of one or few physical eraseblock. Indeed, imagine a
124  * situation when the picked physical eraseblock is constantly erased after the
125  * data is written to it. So, we have a constant which limits the highest erase
126  * counter of the free physical eraseblock to pick. Namely, the WL sub-system
127  * does not pick eraseblocks with erase counter greater than the lowest erase
128  * counter plus %WL_FREE_MAX_DIFF.
129  */
130 #define WL_FREE_MAX_DIFF (2*UBI_WL_THRESHOLD)
131
132 /*
133  * Maximum number of consecutive background thread failures which is enough to
134  * switch to read-only mode.
135  */
136 #define WL_MAX_FAILURES 32
137
138 static int self_check_ec(struct ubi_device *ubi, int pnum, int ec);
139 static int self_check_in_wl_tree(const struct ubi_device *ubi,
140                                  struct ubi_wl_entry *e, struct rb_root *root);
141 static int self_check_in_pq(const struct ubi_device *ubi,
142                             struct ubi_wl_entry *e);
143
144 /**
145  * wl_tree_add - add a wear-leveling entry to a WL RB-tree.
146  * @e: the wear-leveling entry to add
147  * @root: the root of the tree
148  *
149  * Note, we use (erase counter, physical eraseblock number) pairs as keys in
150  * the @ubi->used and @ubi->free RB-trees.
151  */
152 static void wl_tree_add(struct ubi_wl_entry *e, struct rb_root *root)
153 {
154         struct rb_node **p, *parent = NULL;
155
156         p = &root->rb_node;
157         while (*p) {
158                 struct ubi_wl_entry *e1;
159
160                 parent = *p;
161                 e1 = rb_entry(parent, struct ubi_wl_entry, u.rb);
162
163                 if (e->ec < e1->ec)
164                         p = &(*p)->rb_left;
165                 else if (e->ec > e1->ec)
166                         p = &(*p)->rb_right;
167                 else {
168                         ubi_assert(e->pnum != e1->pnum);
169                         if (e->pnum < e1->pnum)
170                                 p = &(*p)->rb_left;
171                         else
172                                 p = &(*p)->rb_right;
173                 }
174         }
175
176         rb_link_node(&e->u.rb, parent, p);
177         rb_insert_color(&e->u.rb, root);
178 }
179
180 /**
181  * wl_tree_destroy - destroy a wear-leveling entry.
182  * @ubi: UBI device description object
183  * @e: the wear-leveling entry to add
184  *
185  * This function destroys a wear leveling entry and removes
186  * the reference from the lookup table.
187  */
188 static void wl_entry_destroy(struct ubi_device *ubi, struct ubi_wl_entry *e)
189 {
190         ubi->lookuptbl[e->pnum] = NULL;
191         kmem_cache_free(ubi_wl_entry_slab, e);
192 }
193
194 /**
195  * do_work - do one pending work.
196  * @ubi: UBI device description object
197  *
198  * This function returns zero in case of success and a negative error code in
199  * case of failure.
200  */
201 static int do_work(struct ubi_device *ubi)
202 {
203         int err;
204         struct ubi_work *wrk;
205
206         cond_resched();
207
208         /*
209          * @ubi->work_sem is used to synchronize with the workers. Workers take
210          * it in read mode, so many of them may be doing works at a time. But
211          * the queue flush code has to be sure the whole queue of works is
212          * done, and it takes the mutex in write mode.
213          */
214         down_read(&ubi->work_sem);
215         spin_lock(&ubi->wl_lock);
216         if (list_empty(&ubi->works)) {
217                 spin_unlock(&ubi->wl_lock);
218                 up_read(&ubi->work_sem);
219                 return 0;
220         }
221
222         wrk = list_entry(ubi->works.next, struct ubi_work, list);
223         list_del(&wrk->list);
224         ubi->works_count -= 1;
225         ubi_assert(ubi->works_count >= 0);
226         spin_unlock(&ubi->wl_lock);
227
228         /*
229          * Call the worker function. Do not touch the work structure
230          * after this call as it will have been freed or reused by that
231          * time by the worker function.
232          */
233         err = wrk->func(ubi, wrk, 0);
234         if (err)
235                 ubi_err(ubi, "work failed with error code %d", err);
236         up_read(&ubi->work_sem);
237
238         return err;
239 }
240
241 /**
242  * in_wl_tree - check if wear-leveling entry is present in a WL RB-tree.
243  * @e: the wear-leveling entry to check
244  * @root: the root of the tree
245  *
246  * This function returns non-zero if @e is in the @root RB-tree and zero if it
247  * is not.
248  */
249 static int in_wl_tree(struct ubi_wl_entry *e, struct rb_root *root)
250 {
251         struct rb_node *p;
252
253         p = root->rb_node;
254         while (p) {
255                 struct ubi_wl_entry *e1;
256
257                 e1 = rb_entry(p, struct ubi_wl_entry, u.rb);
258
259                 if (e->pnum == e1->pnum) {
260                         ubi_assert(e == e1);
261                         return 1;
262                 }
263
264                 if (e->ec < e1->ec)
265                         p = p->rb_left;
266                 else if (e->ec > e1->ec)
267                         p = p->rb_right;
268                 else {
269                         ubi_assert(e->pnum != e1->pnum);
270                         if (e->pnum < e1->pnum)
271                                 p = p->rb_left;
272                         else
273                                 p = p->rb_right;
274                 }
275         }
276
277         return 0;
278 }
279
280 /**
281  * prot_queue_add - add physical eraseblock to the protection queue.
282  * @ubi: UBI device description object
283  * @e: the physical eraseblock to add
284  *
285  * This function adds @e to the tail of the protection queue @ubi->pq, where
286  * @e will stay for %UBI_PROT_QUEUE_LEN erase operations and will be
287  * temporarily protected from the wear-leveling worker. Note, @wl->lock has to
288  * be locked.
289  */
290 static void prot_queue_add(struct ubi_device *ubi, struct ubi_wl_entry *e)
291 {
292         int pq_tail = ubi->pq_head - 1;
293
294         if (pq_tail < 0)
295                 pq_tail = UBI_PROT_QUEUE_LEN - 1;
296         ubi_assert(pq_tail >= 0 && pq_tail < UBI_PROT_QUEUE_LEN);
297         list_add_tail(&e->u.list, &ubi->pq[pq_tail]);
298         dbg_wl("added PEB %d EC %d to the protection queue", e->pnum, e->ec);
299 }
300
301 /**
302  * find_wl_entry - find wear-leveling entry closest to certain erase counter.
303  * @ubi: UBI device description object
304  * @root: the RB-tree where to look for
305  * @diff: maximum possible difference from the smallest erase counter
306  *
307  * This function looks for a wear leveling entry with erase counter closest to
308  * min + @diff, where min is the smallest erase counter.
309  */
310 static struct ubi_wl_entry *find_wl_entry(struct ubi_device *ubi,
311                                           struct rb_root *root, int diff)
312 {
313         struct rb_node *p;
314         struct ubi_wl_entry *e, *prev_e = NULL;
315         int max;
316
317         e = rb_entry(rb_first(root), struct ubi_wl_entry, u.rb);
318         max = e->ec + diff;
319
320         p = root->rb_node;
321         while (p) {
322                 struct ubi_wl_entry *e1;
323
324                 e1 = rb_entry(p, struct ubi_wl_entry, u.rb);
325                 if (e1->ec >= max)
326                         p = p->rb_left;
327                 else {
328                         p = p->rb_right;
329                         prev_e = e;
330                         e = e1;
331                 }
332         }
333
334         /* If no fastmap has been written and this WL entry can be used
335          * as anchor PEB, hold it back and return the second best WL entry
336          * such that fastmap can use the anchor PEB later. */
337         if (prev_e && !ubi->fm_disabled &&
338             !ubi->fm && e->pnum < UBI_FM_MAX_START)
339                 return prev_e;
340
341         return e;
342 }
343
344 /**
345  * find_mean_wl_entry - find wear-leveling entry with medium erase counter.
346  * @ubi: UBI device description object
347  * @root: the RB-tree where to look for
348  *
349  * This function looks for a wear leveling entry with medium erase counter,
350  * but not greater or equivalent than the lowest erase counter plus
351  * %WL_FREE_MAX_DIFF/2.
352  */
353 static struct ubi_wl_entry *find_mean_wl_entry(struct ubi_device *ubi,
354                                                struct rb_root *root)
355 {
356         struct ubi_wl_entry *e, *first, *last;
357
358         first = rb_entry(rb_first(root), struct ubi_wl_entry, u.rb);
359         last = rb_entry(rb_last(root), struct ubi_wl_entry, u.rb);
360
361         if (last->ec - first->ec < WL_FREE_MAX_DIFF) {
362                 e = rb_entry(root->rb_node, struct ubi_wl_entry, u.rb);
363
364                 /* If no fastmap has been written and this WL entry can be used
365                  * as anchor PEB, hold it back and return the second best
366                  * WL entry such that fastmap can use the anchor PEB later. */
367                 e = may_reserve_for_fm(ubi, e, root);
368         } else
369                 e = find_wl_entry(ubi, root, WL_FREE_MAX_DIFF/2);
370
371         return e;
372 }
373
374 /**
375  * wl_get_wle - get a mean wl entry to be used by ubi_wl_get_peb() or
376  * refill_wl_user_pool().
377  * @ubi: UBI device description object
378  *
379  * This function returns a a wear leveling entry in case of success and
380  * NULL in case of failure.
381  */
382 static struct ubi_wl_entry *wl_get_wle(struct ubi_device *ubi)
383 {
384         struct ubi_wl_entry *e;
385
386         e = find_mean_wl_entry(ubi, &ubi->free);
387         if (!e) {
388                 ubi_err(ubi, "no free eraseblocks");
389                 return NULL;
390         }
391
392         self_check_in_wl_tree(ubi, e, &ubi->free);
393
394         /*
395          * Move the physical eraseblock to the protection queue where it will
396          * be protected from being moved for some time.
397          */
398         rb_erase(&e->u.rb, &ubi->free);
399         ubi->free_count--;
400         dbg_wl("PEB %d EC %d", e->pnum, e->ec);
401
402         return e;
403 }
404
405 /**
406  * prot_queue_del - remove a physical eraseblock from the protection queue.
407  * @ubi: UBI device description object
408  * @pnum: the physical eraseblock to remove
409  *
410  * This function deletes PEB @pnum from the protection queue and returns zero
411  * in case of success and %-ENODEV if the PEB was not found.
412  */
413 static int prot_queue_del(struct ubi_device *ubi, int pnum)
414 {
415         struct ubi_wl_entry *e;
416
417         e = ubi->lookuptbl[pnum];
418         if (!e)
419                 return -ENODEV;
420
421         if (self_check_in_pq(ubi, e))
422                 return -ENODEV;
423
424         list_del(&e->u.list);
425         dbg_wl("deleted PEB %d from the protection queue", e->pnum);
426         return 0;
427 }
428
429 /**
430  * sync_erase - synchronously erase a physical eraseblock.
431  * @ubi: UBI device description object
432  * @e: the the physical eraseblock to erase
433  * @torture: if the physical eraseblock has to be tortured
434  *
435  * This function returns zero in case of success and a negative error code in
436  * case of failure.
437  */
438 static int sync_erase(struct ubi_device *ubi, struct ubi_wl_entry *e,
439                       int torture)
440 {
441         int err;
442         struct ubi_ec_hdr *ec_hdr;
443         unsigned long long ec = e->ec;
444
445         dbg_wl("erase PEB %d, old EC %llu", e->pnum, ec);
446
447         err = self_check_ec(ubi, e->pnum, e->ec);
448         if (err)
449                 return -EINVAL;
450
451         ec_hdr = kzalloc(ubi->ec_hdr_alsize, GFP_NOFS);
452         if (!ec_hdr)
453                 return -ENOMEM;
454
455         err = ubi_io_sync_erase(ubi, e->pnum, torture);
456         if (err < 0)
457                 goto out_free;
458
459         ec += err;
460         if (ec > UBI_MAX_ERASECOUNTER) {
461                 /*
462                  * Erase counter overflow. Upgrade UBI and use 64-bit
463                  * erase counters internally.
464                  */
465                 ubi_err(ubi, "erase counter overflow at PEB %d, EC %llu",
466                         e->pnum, ec);
467                 err = -EINVAL;
468                 goto out_free;
469         }
470
471         dbg_wl("erased PEB %d, new EC %llu", e->pnum, ec);
472
473         ec_hdr->ec = cpu_to_be64(ec);
474
475         err = ubi_io_write_ec_hdr(ubi, e->pnum, ec_hdr);
476         if (err)
477                 goto out_free;
478
479         e->ec = ec;
480         spin_lock(&ubi->wl_lock);
481         if (e->ec > ubi->max_ec)
482                 ubi->max_ec = e->ec;
483         spin_unlock(&ubi->wl_lock);
484
485 out_free:
486         kfree(ec_hdr);
487         return err;
488 }
489
490 /**
491  * serve_prot_queue - check if it is time to stop protecting PEBs.
492  * @ubi: UBI device description object
493  *
494  * This function is called after each erase operation and removes PEBs from the
495  * tail of the protection queue. These PEBs have been protected for long enough
496  * and should be moved to the used tree.
497  */
498 static void serve_prot_queue(struct ubi_device *ubi)
499 {
500         struct ubi_wl_entry *e, *tmp;
501         int count;
502
503         /*
504          * There may be several protected physical eraseblock to remove,
505          * process them all.
506          */
507 repeat:
508         count = 0;
509         spin_lock(&ubi->wl_lock);
510         list_for_each_entry_safe(e, tmp, &ubi->pq[ubi->pq_head], u.list) {
511                 dbg_wl("PEB %d EC %d protection over, move to used tree",
512                         e->pnum, e->ec);
513
514                 list_del(&e->u.list);
515                 wl_tree_add(e, &ubi->used);
516                 if (count++ > 32) {
517                         /*
518                          * Let's be nice and avoid holding the spinlock for
519                          * too long.
520                          */
521                         spin_unlock(&ubi->wl_lock);
522                         cond_resched();
523                         goto repeat;
524                 }
525         }
526
527         ubi->pq_head += 1;
528         if (ubi->pq_head == UBI_PROT_QUEUE_LEN)
529                 ubi->pq_head = 0;
530         ubi_assert(ubi->pq_head >= 0 && ubi->pq_head < UBI_PROT_QUEUE_LEN);
531         spin_unlock(&ubi->wl_lock);
532 }
533
534 /**
535  * __schedule_ubi_work - schedule a work.
536  * @ubi: UBI device description object
537  * @wrk: the work to schedule
538  *
539  * This function adds a work defined by @wrk to the tail of the pending works
540  * list. Can only be used if ubi->work_sem is already held in read mode!
541  */
542 static void __schedule_ubi_work(struct ubi_device *ubi, struct ubi_work *wrk)
543 {
544         spin_lock(&ubi->wl_lock);
545         list_add_tail(&wrk->list, &ubi->works);
546         ubi_assert(ubi->works_count >= 0);
547         ubi->works_count += 1;
548         if (ubi->thread_enabled && !ubi_dbg_is_bgt_disabled(ubi))
549                 wake_up_process(ubi->bgt_thread);
550         spin_unlock(&ubi->wl_lock);
551 }
552
553 /**
554  * schedule_ubi_work - schedule a work.
555  * @ubi: UBI device description object
556  * @wrk: the work to schedule
557  *
558  * This function adds a work defined by @wrk to the tail of the pending works
559  * list.
560  */
561 static void schedule_ubi_work(struct ubi_device *ubi, struct ubi_work *wrk)
562 {
563         down_read(&ubi->work_sem);
564         __schedule_ubi_work(ubi, wrk);
565         up_read(&ubi->work_sem);
566 }
567
568 static int erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk,
569                         int shutdown);
570
571 /**
572  * schedule_erase - schedule an erase work.
573  * @ubi: UBI device description object
574  * @e: the WL entry of the physical eraseblock to erase
575  * @vol_id: the volume ID that last used this PEB
576  * @lnum: the last used logical eraseblock number for the PEB
577  * @torture: if the physical eraseblock has to be tortured
578  *
579  * This function returns zero in case of success and a %-ENOMEM in case of
580  * failure.
581  */
582 static int schedule_erase(struct ubi_device *ubi, struct ubi_wl_entry *e,
583                           int vol_id, int lnum, int torture)
584 {
585         struct ubi_work *wl_wrk;
586
587         ubi_assert(e);
588
589         dbg_wl("schedule erasure of PEB %d, EC %d, torture %d",
590                e->pnum, e->ec, torture);
591
592         wl_wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
593         if (!wl_wrk)
594                 return -ENOMEM;
595
596         wl_wrk->func = &erase_worker;
597         wl_wrk->e = e;
598         wl_wrk->vol_id = vol_id;
599         wl_wrk->lnum = lnum;
600         wl_wrk->torture = torture;
601
602         schedule_ubi_work(ubi, wl_wrk);
603         return 0;
604 }
605
606 static int __erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk);
607 /**
608  * do_sync_erase - run the erase worker synchronously.
609  * @ubi: UBI device description object
610  * @e: the WL entry of the physical eraseblock to erase
611  * @vol_id: the volume ID that last used this PEB
612  * @lnum: the last used logical eraseblock number for the PEB
613  * @torture: if the physical eraseblock has to be tortured
614  *
615  */
616 static int do_sync_erase(struct ubi_device *ubi, struct ubi_wl_entry *e,
617                          int vol_id, int lnum, int torture)
618 {
619         struct ubi_work wl_wrk;
620
621         dbg_wl("sync erase of PEB %i", e->pnum);
622
623         wl_wrk.e = e;
624         wl_wrk.vol_id = vol_id;
625         wl_wrk.lnum = lnum;
626         wl_wrk.torture = torture;
627
628         return __erase_worker(ubi, &wl_wrk);
629 }
630
631 /**
632  * wear_leveling_worker - wear-leveling worker function.
633  * @ubi: UBI device description object
634  * @wrk: the work object
635  * @shutdown: non-zero if the worker has to free memory and exit
636  * because the WL-subsystem is shutting down
637  *
638  * This function copies a more worn out physical eraseblock to a less worn out
639  * one. Returns zero in case of success and a negative error code in case of
640  * failure.
641  */
642 static int wear_leveling_worker(struct ubi_device *ubi, struct ubi_work *wrk,
643                                 int shutdown)
644 {
645         int err, scrubbing = 0, torture = 0, protect = 0, erroneous = 0;
646         int erase = 0, keep = 0, vol_id = -1, lnum = -1;
647 #ifdef CONFIG_MTD_UBI_FASTMAP
648         int anchor = wrk->anchor;
649 #endif
650         struct ubi_wl_entry *e1, *e2;
651         struct ubi_vid_hdr *vid_hdr;
652
653         kfree(wrk);
654         if (shutdown)
655                 return 0;
656
657         vid_hdr = ubi_zalloc_vid_hdr(ubi, GFP_NOFS);
658         if (!vid_hdr)
659                 return -ENOMEM;
660
661         mutex_lock(&ubi->move_mutex);
662         spin_lock(&ubi->wl_lock);
663         ubi_assert(!ubi->move_from && !ubi->move_to);
664         ubi_assert(!ubi->move_to_put);
665
666         if (!ubi->free.rb_node ||
667             (!ubi->used.rb_node && !ubi->scrub.rb_node)) {
668                 /*
669                  * No free physical eraseblocks? Well, they must be waiting in
670                  * the queue to be erased. Cancel movement - it will be
671                  * triggered again when a free physical eraseblock appears.
672                  *
673                  * No used physical eraseblocks? They must be temporarily
674                  * protected from being moved. They will be moved to the
675                  * @ubi->used tree later and the wear-leveling will be
676                  * triggered again.
677                  */
678                 dbg_wl("cancel WL, a list is empty: free %d, used %d",
679                        !ubi->free.rb_node, !ubi->used.rb_node);
680                 goto out_cancel;
681         }
682
683 #ifdef CONFIG_MTD_UBI_FASTMAP
684         /* Check whether we need to produce an anchor PEB */
685         if (!anchor)
686                 anchor = !anchor_pebs_avalible(&ubi->free);
687
688         if (anchor) {
689                 e1 = find_anchor_wl_entry(&ubi->used);
690                 if (!e1)
691                         goto out_cancel;
692                 e2 = get_peb_for_wl(ubi);
693                 if (!e2)
694                         goto out_cancel;
695
696                 self_check_in_wl_tree(ubi, e1, &ubi->used);
697                 rb_erase(&e1->u.rb, &ubi->used);
698                 dbg_wl("anchor-move PEB %d to PEB %d", e1->pnum, e2->pnum);
699         } else if (!ubi->scrub.rb_node) {
700 #else
701         if (!ubi->scrub.rb_node) {
702 #endif
703                 /*
704                  * Now pick the least worn-out used physical eraseblock and a
705                  * highly worn-out free physical eraseblock. If the erase
706                  * counters differ much enough, start wear-leveling.
707                  */
708                 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, u.rb);
709                 e2 = get_peb_for_wl(ubi);
710                 if (!e2)
711                         goto out_cancel;
712
713                 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD)) {
714                         dbg_wl("no WL needed: min used EC %d, max free EC %d",
715                                e1->ec, e2->ec);
716
717                         /* Give the unused PEB back */
718                         wl_tree_add(e2, &ubi->free);
719                         ubi->free_count++;
720                         goto out_cancel;
721                 }
722                 self_check_in_wl_tree(ubi, e1, &ubi->used);
723                 rb_erase(&e1->u.rb, &ubi->used);
724                 dbg_wl("move PEB %d EC %d to PEB %d EC %d",
725                        e1->pnum, e1->ec, e2->pnum, e2->ec);
726         } else {
727                 /* Perform scrubbing */
728                 scrubbing = 1;
729                 e1 = rb_entry(rb_first(&ubi->scrub), struct ubi_wl_entry, u.rb);
730                 e2 = get_peb_for_wl(ubi);
731                 if (!e2)
732                         goto out_cancel;
733
734                 self_check_in_wl_tree(ubi, e1, &ubi->scrub);
735                 rb_erase(&e1->u.rb, &ubi->scrub);
736                 dbg_wl("scrub PEB %d to PEB %d", e1->pnum, e2->pnum);
737         }
738
739         ubi->move_from = e1;
740         ubi->move_to = e2;
741         spin_unlock(&ubi->wl_lock);
742
743         /*
744          * Now we are going to copy physical eraseblock @e1->pnum to @e2->pnum.
745          * We so far do not know which logical eraseblock our physical
746          * eraseblock (@e1) belongs to. We have to read the volume identifier
747          * header first.
748          *
749          * Note, we are protected from this PEB being unmapped and erased. The
750          * 'ubi_wl_put_peb()' would wait for moving to be finished if the PEB
751          * which is being moved was unmapped.
752          */
753
754         err = ubi_io_read_vid_hdr(ubi, e1->pnum, vid_hdr, 0);
755         if (err && err != UBI_IO_BITFLIPS) {
756                 if (err == UBI_IO_FF) {
757                         /*
758                          * We are trying to move PEB without a VID header. UBI
759                          * always write VID headers shortly after the PEB was
760                          * given, so we have a situation when it has not yet
761                          * had a chance to write it, because it was preempted.
762                          * So add this PEB to the protection queue so far,
763                          * because presumably more data will be written there
764                          * (including the missing VID header), and then we'll
765                          * move it.
766                          */
767                         dbg_wl("PEB %d has no VID header", e1->pnum);
768                         protect = 1;
769                         goto out_not_moved;
770                 } else if (err == UBI_IO_FF_BITFLIPS) {
771                         /*
772                          * The same situation as %UBI_IO_FF, but bit-flips were
773                          * detected. It is better to schedule this PEB for
774                          * scrubbing.
775                          */
776                         dbg_wl("PEB %d has no VID header but has bit-flips",
777                                e1->pnum);
778                         scrubbing = 1;
779                         goto out_not_moved;
780                 } else if (ubi->fast_attach && err == UBI_IO_BAD_HDR_EBADMSG) {
781                         /*
782                          * While a full scan would detect interrupted erasures
783                          * at attach time we can face them here when attached from
784                          * Fastmap.
785                          */
786                         dbg_wl("PEB %d has ECC errors, maybe from an interrupted erasure",
787                                e1->pnum);
788                         erase = 1;
789                         goto out_not_moved;
790                 }
791
792                 ubi_err(ubi, "error %d while reading VID header from PEB %d",
793                         err, e1->pnum);
794                 goto out_error;
795         }
796
797         vol_id = be32_to_cpu(vid_hdr->vol_id);
798         lnum = be32_to_cpu(vid_hdr->lnum);
799
800         err = ubi_eba_copy_leb(ubi, e1->pnum, e2->pnum, vid_hdr);
801         if (err) {
802                 if (err == MOVE_CANCEL_RACE) {
803                         /*
804                          * The LEB has not been moved because the volume is
805                          * being deleted or the PEB has been put meanwhile. We
806                          * should prevent this PEB from being selected for
807                          * wear-leveling movement again, so put it to the
808                          * protection queue.
809                          */
810                         protect = 1;
811                         goto out_not_moved;
812                 }
813                 if (err == MOVE_RETRY) {
814                         scrubbing = 1;
815                         goto out_not_moved;
816                 }
817                 if (err == MOVE_TARGET_BITFLIPS || err == MOVE_TARGET_WR_ERR ||
818                     err == MOVE_TARGET_RD_ERR) {
819                         /*
820                          * Target PEB had bit-flips or write error - torture it.
821                          */
822                         torture = 1;
823                         keep = 1;
824                         goto out_not_moved;
825                 }
826
827                 if (err == MOVE_SOURCE_RD_ERR) {
828                         /*
829                          * An error happened while reading the source PEB. Do
830                          * not switch to R/O mode in this case, and give the
831                          * upper layers a possibility to recover from this,
832                          * e.g. by unmapping corresponding LEB. Instead, just
833                          * put this PEB to the @ubi->erroneous list to prevent
834                          * UBI from trying to move it over and over again.
835                          */
836                         if (ubi->erroneous_peb_count > ubi->max_erroneous) {
837                                 ubi_err(ubi, "too many erroneous eraseblocks (%d)",
838                                         ubi->erroneous_peb_count);
839                                 goto out_error;
840                         }
841                         erroneous = 1;
842                         goto out_not_moved;
843                 }
844
845                 if (err < 0)
846                         goto out_error;
847
848                 ubi_assert(0);
849         }
850
851         /* The PEB has been successfully moved */
852         if (scrubbing)
853                 ubi_msg(ubi, "scrubbed PEB %d (LEB %d:%d), data moved to PEB %d",
854                         e1->pnum, vol_id, lnum, e2->pnum);
855         ubi_free_vid_hdr(ubi, vid_hdr);
856
857         spin_lock(&ubi->wl_lock);
858         if (!ubi->move_to_put) {
859                 wl_tree_add(e2, &ubi->used);
860                 e2 = NULL;
861         }
862         ubi->move_from = ubi->move_to = NULL;
863         ubi->move_to_put = ubi->wl_scheduled = 0;
864         spin_unlock(&ubi->wl_lock);
865
866         err = do_sync_erase(ubi, e1, vol_id, lnum, 0);
867         if (err) {
868                 if (e2)
869                         wl_entry_destroy(ubi, e2);
870                 goto out_ro;
871         }
872
873         if (e2) {
874                 /*
875                  * Well, the target PEB was put meanwhile, schedule it for
876                  * erasure.
877                  */
878                 dbg_wl("PEB %d (LEB %d:%d) was put meanwhile, erase",
879                        e2->pnum, vol_id, lnum);
880                 err = do_sync_erase(ubi, e2, vol_id, lnum, 0);
881                 if (err)
882                         goto out_ro;
883         }
884
885         dbg_wl("done");
886         mutex_unlock(&ubi->move_mutex);
887         return 0;
888
889         /*
890          * For some reasons the LEB was not moved, might be an error, might be
891          * something else. @e1 was not changed, so return it back. @e2 might
892          * have been changed, schedule it for erasure.
893          */
894 out_not_moved:
895         if (vol_id != -1)
896                 dbg_wl("cancel moving PEB %d (LEB %d:%d) to PEB %d (%d)",
897                        e1->pnum, vol_id, lnum, e2->pnum, err);
898         else
899                 dbg_wl("cancel moving PEB %d to PEB %d (%d)",
900                        e1->pnum, e2->pnum, err);
901         spin_lock(&ubi->wl_lock);
902         if (protect)
903                 prot_queue_add(ubi, e1);
904         else if (erroneous) {
905                 wl_tree_add(e1, &ubi->erroneous);
906                 ubi->erroneous_peb_count += 1;
907         } else if (scrubbing)
908                 wl_tree_add(e1, &ubi->scrub);
909         else if (keep)
910                 wl_tree_add(e1, &ubi->used);
911         ubi_assert(!ubi->move_to_put);
912         ubi->move_from = ubi->move_to = NULL;
913         ubi->wl_scheduled = 0;
914         spin_unlock(&ubi->wl_lock);
915
916         ubi_free_vid_hdr(ubi, vid_hdr);
917         err = do_sync_erase(ubi, e2, vol_id, lnum, torture);
918         if (err)
919                 goto out_ro;
920
921         if (erase) {
922                 err = do_sync_erase(ubi, e1, vol_id, lnum, 1);
923                 if (err)
924                         goto out_ro;
925         }
926
927         mutex_unlock(&ubi->move_mutex);
928         return 0;
929
930 out_error:
931         if (vol_id != -1)
932                 ubi_err(ubi, "error %d while moving PEB %d to PEB %d",
933                         err, e1->pnum, e2->pnum);
934         else
935                 ubi_err(ubi, "error %d while moving PEB %d (LEB %d:%d) to PEB %d",
936                         err, e1->pnum, vol_id, lnum, e2->pnum);
937         spin_lock(&ubi->wl_lock);
938         ubi->move_from = ubi->move_to = NULL;
939         ubi->move_to_put = ubi->wl_scheduled = 0;
940         spin_unlock(&ubi->wl_lock);
941
942         ubi_free_vid_hdr(ubi, vid_hdr);
943         wl_entry_destroy(ubi, e1);
944         wl_entry_destroy(ubi, e2);
945
946 out_ro:
947         ubi_ro_mode(ubi);
948         mutex_unlock(&ubi->move_mutex);
949         ubi_assert(err != 0);
950         return err < 0 ? err : -EIO;
951
952 out_cancel:
953         ubi->wl_scheduled = 0;
954         spin_unlock(&ubi->wl_lock);
955         mutex_unlock(&ubi->move_mutex);
956         ubi_free_vid_hdr(ubi, vid_hdr);
957         return 0;
958 }
959
960 /**
961  * ensure_wear_leveling - schedule wear-leveling if it is needed.
962  * @ubi: UBI device description object
963  * @nested: set to non-zero if this function is called from UBI worker
964  *
965  * This function checks if it is time to start wear-leveling and schedules it
966  * if yes. This function returns zero in case of success and a negative error
967  * code in case of failure.
968  */
969 static int ensure_wear_leveling(struct ubi_device *ubi, int nested)
970 {
971         int err = 0;
972         struct ubi_wl_entry *e1;
973         struct ubi_wl_entry *e2;
974         struct ubi_work *wrk;
975
976         spin_lock(&ubi->wl_lock);
977         if (ubi->wl_scheduled)
978                 /* Wear-leveling is already in the work queue */
979                 goto out_unlock;
980
981         /*
982          * If the ubi->scrub tree is not empty, scrubbing is needed, and the
983          * the WL worker has to be scheduled anyway.
984          */
985         if (!ubi->scrub.rb_node) {
986                 if (!ubi->used.rb_node || !ubi->free.rb_node)
987                         /* No physical eraseblocks - no deal */
988                         goto out_unlock;
989
990                 /*
991                  * We schedule wear-leveling only if the difference between the
992                  * lowest erase counter of used physical eraseblocks and a high
993                  * erase counter of free physical eraseblocks is greater than
994                  * %UBI_WL_THRESHOLD.
995                  */
996                 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, u.rb);
997                 e2 = find_wl_entry(ubi, &ubi->free, WL_FREE_MAX_DIFF);
998
999                 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD))
1000                         goto out_unlock;
1001                 dbg_wl("schedule wear-leveling");
1002         } else
1003                 dbg_wl("schedule scrubbing");
1004
1005         ubi->wl_scheduled = 1;
1006         spin_unlock(&ubi->wl_lock);
1007
1008         wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
1009         if (!wrk) {
1010                 err = -ENOMEM;
1011                 goto out_cancel;
1012         }
1013
1014         wrk->anchor = 0;
1015         wrk->func = &wear_leveling_worker;
1016         if (nested)
1017                 __schedule_ubi_work(ubi, wrk);
1018         else
1019                 schedule_ubi_work(ubi, wrk);
1020         return err;
1021
1022 out_cancel:
1023         spin_lock(&ubi->wl_lock);
1024         ubi->wl_scheduled = 0;
1025 out_unlock:
1026         spin_unlock(&ubi->wl_lock);
1027         return err;
1028 }
1029
1030 /**
1031  * __erase_worker - physical eraseblock erase worker function.
1032  * @ubi: UBI device description object
1033  * @wl_wrk: the work object
1034  * @shutdown: non-zero if the worker has to free memory and exit
1035  * because the WL sub-system is shutting down
1036  *
1037  * This function erases a physical eraseblock and perform torture testing if
1038  * needed. It also takes care about marking the physical eraseblock bad if
1039  * needed. Returns zero in case of success and a negative error code in case of
1040  * failure.
1041  */
1042 static int __erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk)
1043 {
1044         struct ubi_wl_entry *e = wl_wrk->e;
1045         int pnum = e->pnum;
1046         int vol_id = wl_wrk->vol_id;
1047         int lnum = wl_wrk->lnum;
1048         int err, available_consumed = 0;
1049
1050         dbg_wl("erase PEB %d EC %d LEB %d:%d",
1051                pnum, e->ec, wl_wrk->vol_id, wl_wrk->lnum);
1052
1053         err = sync_erase(ubi, e, wl_wrk->torture);
1054         if (!err) {
1055                 spin_lock(&ubi->wl_lock);
1056                 wl_tree_add(e, &ubi->free);
1057                 ubi->free_count++;
1058                 spin_unlock(&ubi->wl_lock);
1059
1060                 /*
1061                  * One more erase operation has happened, take care about
1062                  * protected physical eraseblocks.
1063                  */
1064                 serve_prot_queue(ubi);
1065
1066                 /* And take care about wear-leveling */
1067                 err = ensure_wear_leveling(ubi, 1);
1068                 return err;
1069         }
1070
1071         ubi_err(ubi, "failed to erase PEB %d, error %d", pnum, err);
1072
1073         if (err == -EINTR || err == -ENOMEM || err == -EAGAIN ||
1074             err == -EBUSY) {
1075                 int err1;
1076
1077                 /* Re-schedule the LEB for erasure */
1078                 err1 = schedule_erase(ubi, e, vol_id, lnum, 0);
1079                 if (err1) {
1080                         wl_entry_destroy(ubi, e);
1081                         err = err1;
1082                         goto out_ro;
1083                 }
1084                 return err;
1085         }
1086
1087         wl_entry_destroy(ubi, e);
1088         if (err != -EIO)
1089                 /*
1090                  * If this is not %-EIO, we have no idea what to do. Scheduling
1091                  * this physical eraseblock for erasure again would cause
1092                  * errors again and again. Well, lets switch to R/O mode.
1093                  */
1094                 goto out_ro;
1095
1096         /* It is %-EIO, the PEB went bad */
1097
1098         if (!ubi->bad_allowed) {
1099                 ubi_err(ubi, "bad physical eraseblock %d detected", pnum);
1100                 goto out_ro;
1101         }
1102
1103         spin_lock(&ubi->volumes_lock);
1104         if (ubi->beb_rsvd_pebs == 0) {
1105                 if (ubi->avail_pebs == 0) {
1106                         spin_unlock(&ubi->volumes_lock);
1107                         ubi_err(ubi, "no reserved/available physical eraseblocks");
1108                         goto out_ro;
1109                 }
1110                 ubi->avail_pebs -= 1;
1111                 available_consumed = 1;
1112         }
1113         spin_unlock(&ubi->volumes_lock);
1114
1115         ubi_msg(ubi, "mark PEB %d as bad", pnum);
1116         err = ubi_io_mark_bad(ubi, pnum);
1117         if (err)
1118                 goto out_ro;
1119
1120         spin_lock(&ubi->volumes_lock);
1121         if (ubi->beb_rsvd_pebs > 0) {
1122                 if (available_consumed) {
1123                         /*
1124                          * The amount of reserved PEBs increased since we last
1125                          * checked.
1126                          */
1127                         ubi->avail_pebs += 1;
1128                         available_consumed = 0;
1129                 }
1130                 ubi->beb_rsvd_pebs -= 1;
1131         }
1132         ubi->bad_peb_count += 1;
1133         ubi->good_peb_count -= 1;
1134         ubi_calculate_reserved(ubi);
1135         if (available_consumed)
1136                 ubi_warn(ubi, "no PEBs in the reserved pool, used an available PEB");
1137         else if (ubi->beb_rsvd_pebs)
1138                 ubi_msg(ubi, "%d PEBs left in the reserve",
1139                         ubi->beb_rsvd_pebs);
1140         else
1141                 ubi_warn(ubi, "last PEB from the reserve was used");
1142         spin_unlock(&ubi->volumes_lock);
1143
1144         return err;
1145
1146 out_ro:
1147         if (available_consumed) {
1148                 spin_lock(&ubi->volumes_lock);
1149                 ubi->avail_pebs += 1;
1150                 spin_unlock(&ubi->volumes_lock);
1151         }
1152         ubi_ro_mode(ubi);
1153         return err;
1154 }
1155
1156 static int erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk,
1157                           int shutdown)
1158 {
1159         int ret;
1160
1161         if (shutdown) {
1162                 struct ubi_wl_entry *e = wl_wrk->e;
1163
1164                 dbg_wl("cancel erasure of PEB %d EC %d", e->pnum, e->ec);
1165                 kfree(wl_wrk);
1166                 wl_entry_destroy(ubi, e);
1167                 return 0;
1168         }
1169
1170         ret = __erase_worker(ubi, wl_wrk);
1171         kfree(wl_wrk);
1172         return ret;
1173 }
1174
1175 /**
1176  * ubi_wl_put_peb - return a PEB to the wear-leveling sub-system.
1177  * @ubi: UBI device description object
1178  * @vol_id: the volume ID that last used this PEB
1179  * @lnum: the last used logical eraseblock number for the PEB
1180  * @pnum: physical eraseblock to return
1181  * @torture: if this physical eraseblock has to be tortured
1182  *
1183  * This function is called to return physical eraseblock @pnum to the pool of
1184  * free physical eraseblocks. The @torture flag has to be set if an I/O error
1185  * occurred to this @pnum and it has to be tested. This function returns zero
1186  * in case of success, and a negative error code in case of failure.
1187  */
1188 int ubi_wl_put_peb(struct ubi_device *ubi, int vol_id, int lnum,
1189                    int pnum, int torture)
1190 {
1191         int err;
1192         struct ubi_wl_entry *e;
1193
1194         dbg_wl("PEB %d", pnum);
1195         ubi_assert(pnum >= 0);
1196         ubi_assert(pnum < ubi->peb_count);
1197
1198         down_read(&ubi->fm_protect);
1199
1200 retry:
1201         spin_lock(&ubi->wl_lock);
1202         e = ubi->lookuptbl[pnum];
1203         if (e == ubi->move_from) {
1204                 /*
1205                  * User is putting the physical eraseblock which was selected to
1206                  * be moved. It will be scheduled for erasure in the
1207                  * wear-leveling worker.
1208                  */
1209                 dbg_wl("PEB %d is being moved, wait", pnum);
1210                 spin_unlock(&ubi->wl_lock);
1211
1212                 /* Wait for the WL worker by taking the @ubi->move_mutex */
1213                 mutex_lock(&ubi->move_mutex);
1214                 mutex_unlock(&ubi->move_mutex);
1215                 goto retry;
1216         } else if (e == ubi->move_to) {
1217                 /*
1218                  * User is putting the physical eraseblock which was selected
1219                  * as the target the data is moved to. It may happen if the EBA
1220                  * sub-system already re-mapped the LEB in 'ubi_eba_copy_leb()'
1221                  * but the WL sub-system has not put the PEB to the "used" tree
1222                  * yet, but it is about to do this. So we just set a flag which
1223                  * will tell the WL worker that the PEB is not needed anymore
1224                  * and should be scheduled for erasure.
1225                  */
1226                 dbg_wl("PEB %d is the target of data moving", pnum);
1227                 ubi_assert(!ubi->move_to_put);
1228                 ubi->move_to_put = 1;
1229                 spin_unlock(&ubi->wl_lock);
1230                 up_read(&ubi->fm_protect);
1231                 return 0;
1232         } else {
1233                 if (in_wl_tree(e, &ubi->used)) {
1234                         self_check_in_wl_tree(ubi, e, &ubi->used);
1235                         rb_erase(&e->u.rb, &ubi->used);
1236                 } else if (in_wl_tree(e, &ubi->scrub)) {
1237                         self_check_in_wl_tree(ubi, e, &ubi->scrub);
1238                         rb_erase(&e->u.rb, &ubi->scrub);
1239                 } else if (in_wl_tree(e, &ubi->erroneous)) {
1240                         self_check_in_wl_tree(ubi, e, &ubi->erroneous);
1241                         rb_erase(&e->u.rb, &ubi->erroneous);
1242                         ubi->erroneous_peb_count -= 1;
1243                         ubi_assert(ubi->erroneous_peb_count >= 0);
1244                         /* Erroneous PEBs should be tortured */
1245                         torture = 1;
1246                 } else {
1247                         err = prot_queue_del(ubi, e->pnum);
1248                         if (err) {
1249                                 ubi_err(ubi, "PEB %d not found", pnum);
1250                                 ubi_ro_mode(ubi);
1251                                 spin_unlock(&ubi->wl_lock);
1252                                 up_read(&ubi->fm_protect);
1253                                 return err;
1254                         }
1255                 }
1256         }
1257         spin_unlock(&ubi->wl_lock);
1258
1259         err = schedule_erase(ubi, e, vol_id, lnum, torture);
1260         if (err) {
1261                 spin_lock(&ubi->wl_lock);
1262                 wl_tree_add(e, &ubi->used);
1263                 spin_unlock(&ubi->wl_lock);
1264         }
1265
1266         up_read(&ubi->fm_protect);
1267         return err;
1268 }
1269
1270 /**
1271  * ubi_wl_scrub_peb - schedule a physical eraseblock for scrubbing.
1272  * @ubi: UBI device description object
1273  * @pnum: the physical eraseblock to schedule
1274  *
1275  * If a bit-flip in a physical eraseblock is detected, this physical eraseblock
1276  * needs scrubbing. This function schedules a physical eraseblock for
1277  * scrubbing which is done in background. This function returns zero in case of
1278  * success and a negative error code in case of failure.
1279  */
1280 int ubi_wl_scrub_peb(struct ubi_device *ubi, int pnum)
1281 {
1282         struct ubi_wl_entry *e;
1283
1284         ubi_msg(ubi, "schedule PEB %d for scrubbing", pnum);
1285
1286 retry:
1287         spin_lock(&ubi->wl_lock);
1288         e = ubi->lookuptbl[pnum];
1289         if (e == ubi->move_from || in_wl_tree(e, &ubi->scrub) ||
1290                                    in_wl_tree(e, &ubi->erroneous)) {
1291                 spin_unlock(&ubi->wl_lock);
1292                 return 0;
1293         }
1294
1295         if (e == ubi->move_to) {
1296                 /*
1297                  * This physical eraseblock was used to move data to. The data
1298                  * was moved but the PEB was not yet inserted to the proper
1299                  * tree. We should just wait a little and let the WL worker
1300                  * proceed.
1301                  */
1302                 spin_unlock(&ubi->wl_lock);
1303                 dbg_wl("the PEB %d is not in proper tree, retry", pnum);
1304                 yield();
1305                 goto retry;
1306         }
1307
1308         if (in_wl_tree(e, &ubi->used)) {
1309                 self_check_in_wl_tree(ubi, e, &ubi->used);
1310                 rb_erase(&e->u.rb, &ubi->used);
1311         } else {
1312                 int err;
1313
1314                 err = prot_queue_del(ubi, e->pnum);
1315                 if (err) {
1316                         ubi_err(ubi, "PEB %d not found", pnum);
1317                         ubi_ro_mode(ubi);
1318                         spin_unlock(&ubi->wl_lock);
1319                         return err;
1320                 }
1321         }
1322
1323         wl_tree_add(e, &ubi->scrub);
1324         spin_unlock(&ubi->wl_lock);
1325
1326         /*
1327          * Technically scrubbing is the same as wear-leveling, so it is done
1328          * by the WL worker.
1329          */
1330         return ensure_wear_leveling(ubi, 0);
1331 }
1332
1333 /**
1334  * ubi_wl_flush - flush all pending works.
1335  * @ubi: UBI device description object
1336  * @vol_id: the volume id to flush for
1337  * @lnum: the logical eraseblock number to flush for
1338  *
1339  * This function executes all pending works for a particular volume id /
1340  * logical eraseblock number pair. If either value is set to %UBI_ALL, then it
1341  * acts as a wildcard for all of the corresponding volume numbers or logical
1342  * eraseblock numbers. It returns zero in case of success and a negative error
1343  * code in case of failure.
1344  */
1345 int ubi_wl_flush(struct ubi_device *ubi, int vol_id, int lnum)
1346 {
1347         int err = 0;
1348         int found = 1;
1349
1350         /*
1351          * Erase while the pending works queue is not empty, but not more than
1352          * the number of currently pending works.
1353          */
1354         dbg_wl("flush pending work for LEB %d:%d (%d pending works)",
1355                vol_id, lnum, ubi->works_count);
1356
1357         while (found) {
1358                 struct ubi_work *wrk, *tmp;
1359                 found = 0;
1360
1361                 down_read(&ubi->work_sem);
1362                 spin_lock(&ubi->wl_lock);
1363                 list_for_each_entry_safe(wrk, tmp, &ubi->works, list) {
1364                         if ((vol_id == UBI_ALL || wrk->vol_id == vol_id) &&
1365                             (lnum == UBI_ALL || wrk->lnum == lnum)) {
1366                                 list_del(&wrk->list);
1367                                 ubi->works_count -= 1;
1368                                 ubi_assert(ubi->works_count >= 0);
1369                                 spin_unlock(&ubi->wl_lock);
1370
1371                                 err = wrk->func(ubi, wrk, 0);
1372                                 if (err) {
1373                                         up_read(&ubi->work_sem);
1374                                         return err;
1375                                 }
1376
1377                                 spin_lock(&ubi->wl_lock);
1378                                 found = 1;
1379                                 break;
1380                         }
1381                 }
1382                 spin_unlock(&ubi->wl_lock);
1383                 up_read(&ubi->work_sem);
1384         }
1385
1386         /*
1387          * Make sure all the works which have been done in parallel are
1388          * finished.
1389          */
1390         down_write(&ubi->work_sem);
1391         up_write(&ubi->work_sem);
1392
1393         return err;
1394 }
1395
1396 /**
1397  * tree_destroy - destroy an RB-tree.
1398  * @ubi: UBI device description object
1399  * @root: the root of the tree to destroy
1400  */
1401 static void tree_destroy(struct ubi_device *ubi, struct rb_root *root)
1402 {
1403         struct rb_node *rb;
1404         struct ubi_wl_entry *e;
1405
1406         rb = root->rb_node;
1407         while (rb) {
1408                 if (rb->rb_left)
1409                         rb = rb->rb_left;
1410                 else if (rb->rb_right)
1411                         rb = rb->rb_right;
1412                 else {
1413                         e = rb_entry(rb, struct ubi_wl_entry, u.rb);
1414
1415                         rb = rb_parent(rb);
1416                         if (rb) {
1417                                 if (rb->rb_left == &e->u.rb)
1418                                         rb->rb_left = NULL;
1419                                 else
1420                                         rb->rb_right = NULL;
1421                         }
1422
1423                         wl_entry_destroy(ubi, e);
1424                 }
1425         }
1426 }
1427
1428 /**
1429  * ubi_thread - UBI background thread.
1430  * @u: the UBI device description object pointer
1431  */
1432 int ubi_thread(void *u)
1433 {
1434         int failures = 0;
1435         struct ubi_device *ubi = u;
1436
1437         ubi_msg(ubi, "background thread \"%s\" started, PID %d",
1438                 ubi->bgt_name, task_pid_nr(current));
1439
1440         set_freezable();
1441         for (;;) {
1442                 int err;
1443
1444                 if (kthread_should_stop())
1445                         break;
1446
1447                 if (try_to_freeze())
1448                         continue;
1449
1450                 spin_lock(&ubi->wl_lock);
1451                 if (list_empty(&ubi->works) || ubi->ro_mode ||
1452                     !ubi->thread_enabled || ubi_dbg_is_bgt_disabled(ubi)) {
1453                         set_current_state(TASK_INTERRUPTIBLE);
1454                         spin_unlock(&ubi->wl_lock);
1455                         schedule();
1456                         continue;
1457                 }
1458                 spin_unlock(&ubi->wl_lock);
1459
1460                 err = do_work(ubi);
1461                 if (err) {
1462                         ubi_err(ubi, "%s: work failed with error code %d",
1463                                 ubi->bgt_name, err);
1464                         if (failures++ > WL_MAX_FAILURES) {
1465                                 /*
1466                                  * Too many failures, disable the thread and
1467                                  * switch to read-only mode.
1468                                  */
1469                                 ubi_msg(ubi, "%s: %d consecutive failures",
1470                                         ubi->bgt_name, WL_MAX_FAILURES);
1471                                 ubi_ro_mode(ubi);
1472                                 ubi->thread_enabled = 0;
1473                                 continue;
1474                         }
1475                 } else
1476                         failures = 0;
1477
1478                 cond_resched();
1479         }
1480
1481         dbg_wl("background thread \"%s\" is killed", ubi->bgt_name);
1482         return 0;
1483 }
1484
1485 /**
1486  * shutdown_work - shutdown all pending works.
1487  * @ubi: UBI device description object
1488  */
1489 static void shutdown_work(struct ubi_device *ubi)
1490 {
1491 #ifdef CONFIG_MTD_UBI_FASTMAP
1492         flush_work(&ubi->fm_work);
1493 #endif
1494         while (!list_empty(&ubi->works)) {
1495                 struct ubi_work *wrk;
1496
1497                 wrk = list_entry(ubi->works.next, struct ubi_work, list);
1498                 list_del(&wrk->list);
1499                 wrk->func(ubi, wrk, 1);
1500                 ubi->works_count -= 1;
1501                 ubi_assert(ubi->works_count >= 0);
1502         }
1503 }
1504
1505 /**
1506  * ubi_wl_init - initialize the WL sub-system using attaching information.
1507  * @ubi: UBI device description object
1508  * @ai: attaching information
1509  *
1510  * This function returns zero in case of success, and a negative error code in
1511  * case of failure.
1512  */
1513 int ubi_wl_init(struct ubi_device *ubi, struct ubi_attach_info *ai)
1514 {
1515         int err, i, reserved_pebs, found_pebs = 0;
1516         struct rb_node *rb1, *rb2;
1517         struct ubi_ainf_volume *av;
1518         struct ubi_ainf_peb *aeb, *tmp;
1519         struct ubi_wl_entry *e;
1520
1521         ubi->used = ubi->erroneous = ubi->free = ubi->scrub = RB_ROOT;
1522         spin_lock_init(&ubi->wl_lock);
1523         mutex_init(&ubi->move_mutex);
1524         init_rwsem(&ubi->work_sem);
1525         ubi->max_ec = ai->max_ec;
1526         INIT_LIST_HEAD(&ubi->works);
1527
1528         sprintf(ubi->bgt_name, UBI_BGT_NAME_PATTERN, ubi->ubi_num);
1529
1530         err = -ENOMEM;
1531         ubi->lookuptbl = kzalloc(ubi->peb_count * sizeof(void *), GFP_KERNEL);
1532         if (!ubi->lookuptbl)
1533                 return err;
1534
1535         for (i = 0; i < UBI_PROT_QUEUE_LEN; i++)
1536                 INIT_LIST_HEAD(&ubi->pq[i]);
1537         ubi->pq_head = 0;
1538
1539         list_for_each_entry_safe(aeb, tmp, &ai->erase, u.list) {
1540                 cond_resched();
1541
1542                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1543                 if (!e)
1544                         goto out_free;
1545
1546                 e->pnum = aeb->pnum;
1547                 e->ec = aeb->ec;
1548                 ubi->lookuptbl[e->pnum] = e;
1549                 if (schedule_erase(ubi, e, aeb->vol_id, aeb->lnum, 0)) {
1550                         wl_entry_destroy(ubi, e);
1551                         goto out_free;
1552                 }
1553
1554                 found_pebs++;
1555         }
1556
1557         ubi->free_count = 0;
1558         list_for_each_entry(aeb, &ai->free, u.list) {
1559                 cond_resched();
1560
1561                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1562                 if (!e)
1563                         goto out_free;
1564
1565                 e->pnum = aeb->pnum;
1566                 e->ec = aeb->ec;
1567                 ubi_assert(e->ec >= 0);
1568
1569                 wl_tree_add(e, &ubi->free);
1570                 ubi->free_count++;
1571
1572                 ubi->lookuptbl[e->pnum] = e;
1573
1574                 found_pebs++;
1575         }
1576
1577         ubi_rb_for_each_entry(rb1, av, &ai->volumes, rb) {
1578                 ubi_rb_for_each_entry(rb2, aeb, &av->root, u.rb) {
1579                         cond_resched();
1580
1581                         e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1582                         if (!e)
1583                                 goto out_free;
1584
1585                         e->pnum = aeb->pnum;
1586                         e->ec = aeb->ec;
1587                         ubi->lookuptbl[e->pnum] = e;
1588
1589                         if (!aeb->scrub) {
1590                                 dbg_wl("add PEB %d EC %d to the used tree",
1591                                        e->pnum, e->ec);
1592                                 wl_tree_add(e, &ubi->used);
1593                         } else {
1594                                 dbg_wl("add PEB %d EC %d to the scrub tree",
1595                                        e->pnum, e->ec);
1596                                 wl_tree_add(e, &ubi->scrub);
1597                         }
1598
1599                         found_pebs++;
1600                 }
1601         }
1602
1603         dbg_wl("found %i PEBs", found_pebs);
1604
1605         if (ubi->fm) {
1606                 ubi_assert(ubi->good_peb_count ==
1607                            found_pebs + ubi->fm->used_blocks);
1608
1609                 for (i = 0; i < ubi->fm->used_blocks; i++) {
1610                         e = ubi->fm->e[i];
1611                         ubi->lookuptbl[e->pnum] = e;
1612                 }
1613         }
1614         else
1615                 ubi_assert(ubi->good_peb_count == found_pebs);
1616
1617         reserved_pebs = WL_RESERVED_PEBS;
1618         ubi_fastmap_init(ubi, &reserved_pebs);
1619
1620         if (ubi->avail_pebs < reserved_pebs) {
1621                 ubi_err(ubi, "no enough physical eraseblocks (%d, need %d)",
1622                         ubi->avail_pebs, reserved_pebs);
1623                 if (ubi->corr_peb_count)
1624                         ubi_err(ubi, "%d PEBs are corrupted and not used",
1625                                 ubi->corr_peb_count);
1626                 err = -ENOSPC;
1627                 goto out_free;
1628         }
1629         ubi->avail_pebs -= reserved_pebs;
1630         ubi->rsvd_pebs += reserved_pebs;
1631
1632         /* Schedule wear-leveling if needed */
1633         err = ensure_wear_leveling(ubi, 0);
1634         if (err)
1635                 goto out_free;
1636
1637         return 0;
1638
1639 out_free:
1640         shutdown_work(ubi);
1641         tree_destroy(ubi, &ubi->used);
1642         tree_destroy(ubi, &ubi->free);
1643         tree_destroy(ubi, &ubi->scrub);
1644         kfree(ubi->lookuptbl);
1645         return err;
1646 }
1647
1648 /**
1649  * protection_queue_destroy - destroy the protection queue.
1650  * @ubi: UBI device description object
1651  */
1652 static void protection_queue_destroy(struct ubi_device *ubi)
1653 {
1654         int i;
1655         struct ubi_wl_entry *e, *tmp;
1656
1657         for (i = 0; i < UBI_PROT_QUEUE_LEN; ++i) {
1658                 list_for_each_entry_safe(e, tmp, &ubi->pq[i], u.list) {
1659                         list_del(&e->u.list);
1660                         wl_entry_destroy(ubi, e);
1661                 }
1662         }
1663 }
1664
1665 /**
1666  * ubi_wl_close - close the wear-leveling sub-system.
1667  * @ubi: UBI device description object
1668  */
1669 void ubi_wl_close(struct ubi_device *ubi)
1670 {
1671         dbg_wl("close the WL sub-system");
1672         ubi_fastmap_close(ubi);
1673         shutdown_work(ubi);
1674         protection_queue_destroy(ubi);
1675         tree_destroy(ubi, &ubi->used);
1676         tree_destroy(ubi, &ubi->erroneous);
1677         tree_destroy(ubi, &ubi->free);
1678         tree_destroy(ubi, &ubi->scrub);
1679         kfree(ubi->lookuptbl);
1680 }
1681
1682 /**
1683  * self_check_ec - make sure that the erase counter of a PEB is correct.
1684  * @ubi: UBI device description object
1685  * @pnum: the physical eraseblock number to check
1686  * @ec: the erase counter to check
1687  *
1688  * This function returns zero if the erase counter of physical eraseblock @pnum
1689  * is equivalent to @ec, and a negative error code if not or if an error
1690  * occurred.
1691  */
1692 static int self_check_ec(struct ubi_device *ubi, int pnum, int ec)
1693 {
1694         int err;
1695         long long read_ec;
1696         struct ubi_ec_hdr *ec_hdr;
1697
1698         if (!ubi_dbg_chk_gen(ubi))
1699                 return 0;
1700
1701         ec_hdr = kzalloc(ubi->ec_hdr_alsize, GFP_NOFS);
1702         if (!ec_hdr)
1703                 return -ENOMEM;
1704
1705         err = ubi_io_read_ec_hdr(ubi, pnum, ec_hdr, 0);
1706         if (err && err != UBI_IO_BITFLIPS) {
1707                 /* The header does not have to exist */
1708                 err = 0;
1709                 goto out_free;
1710         }
1711
1712         read_ec = be64_to_cpu(ec_hdr->ec);
1713         if (ec != read_ec && read_ec - ec > 1) {
1714                 ubi_err(ubi, "self-check failed for PEB %d", pnum);
1715                 ubi_err(ubi, "read EC is %lld, should be %d", read_ec, ec);
1716                 dump_stack();
1717                 err = 1;
1718         } else
1719                 err = 0;
1720
1721 out_free:
1722         kfree(ec_hdr);
1723         return err;
1724 }
1725
1726 /**
1727  * self_check_in_wl_tree - check that wear-leveling entry is in WL RB-tree.
1728  * @ubi: UBI device description object
1729  * @e: the wear-leveling entry to check
1730  * @root: the root of the tree
1731  *
1732  * This function returns zero if @e is in the @root RB-tree and %-EINVAL if it
1733  * is not.
1734  */
1735 static int self_check_in_wl_tree(const struct ubi_device *ubi,
1736                                  struct ubi_wl_entry *e, struct rb_root *root)
1737 {
1738         if (!ubi_dbg_chk_gen(ubi))
1739                 return 0;
1740
1741         if (in_wl_tree(e, root))
1742                 return 0;
1743
1744         ubi_err(ubi, "self-check failed for PEB %d, EC %d, RB-tree %p ",
1745                 e->pnum, e->ec, root);
1746         dump_stack();
1747         return -EINVAL;
1748 }
1749
1750 /**
1751  * self_check_in_pq - check if wear-leveling entry is in the protection
1752  *                        queue.
1753  * @ubi: UBI device description object
1754  * @e: the wear-leveling entry to check
1755  *
1756  * This function returns zero if @e is in @ubi->pq and %-EINVAL if it is not.
1757  */
1758 static int self_check_in_pq(const struct ubi_device *ubi,
1759                             struct ubi_wl_entry *e)
1760 {
1761         struct ubi_wl_entry *p;
1762         int i;
1763
1764         if (!ubi_dbg_chk_gen(ubi))
1765                 return 0;
1766
1767         for (i = 0; i < UBI_PROT_QUEUE_LEN; ++i)
1768                 list_for_each_entry(p, &ubi->pq[i], u.list)
1769                         if (p == e)
1770                                 return 0;
1771
1772         ubi_err(ubi, "self-check failed for PEB %d, EC %d, Protect queue",
1773                 e->pnum, e->ec);
1774         dump_stack();
1775         return -EINVAL;
1776 }
1777 #ifndef CONFIG_MTD_UBI_FASTMAP
1778 static struct ubi_wl_entry *get_peb_for_wl(struct ubi_device *ubi)
1779 {
1780         struct ubi_wl_entry *e;
1781
1782         e = find_wl_entry(ubi, &ubi->free, WL_FREE_MAX_DIFF);
1783         self_check_in_wl_tree(ubi, e, &ubi->free);
1784         ubi->free_count--;
1785         ubi_assert(ubi->free_count >= 0);
1786         rb_erase(&e->u.rb, &ubi->free);
1787
1788         return e;
1789 }
1790
1791 /**
1792  * produce_free_peb - produce a free physical eraseblock.
1793  * @ubi: UBI device description object
1794  *
1795  * This function tries to make a free PEB by means of synchronous execution of
1796  * pending works. This may be needed if, for example the background thread is
1797  * disabled. Returns zero in case of success and a negative error code in case
1798  * of failure.
1799  */
1800 static int produce_free_peb(struct ubi_device *ubi)
1801 {
1802         int err;
1803
1804         while (!ubi->free.rb_node && ubi->works_count) {
1805                 spin_unlock(&ubi->wl_lock);
1806
1807                 dbg_wl("do one work synchronously");
1808                 err = do_work(ubi);
1809
1810                 spin_lock(&ubi->wl_lock);
1811                 if (err)
1812                         return err;
1813         }
1814
1815         return 0;
1816 }
1817
1818 /**
1819  * ubi_wl_get_peb - get a physical eraseblock.
1820  * @ubi: UBI device description object
1821  *
1822  * This function returns a physical eraseblock in case of success and a
1823  * negative error code in case of failure.
1824  * Returns with ubi->fm_eba_sem held in read mode!
1825  */
1826 int ubi_wl_get_peb(struct ubi_device *ubi)
1827 {
1828         int err;
1829         struct ubi_wl_entry *e;
1830
1831 retry:
1832         down_read(&ubi->fm_eba_sem);
1833         spin_lock(&ubi->wl_lock);
1834         if (!ubi->free.rb_node) {
1835                 if (ubi->works_count == 0) {
1836                         ubi_err(ubi, "no free eraseblocks");
1837                         ubi_assert(list_empty(&ubi->works));
1838                         spin_unlock(&ubi->wl_lock);
1839                         return -ENOSPC;
1840                 }
1841
1842                 err = produce_free_peb(ubi);
1843                 if (err < 0) {
1844                         spin_unlock(&ubi->wl_lock);
1845                         return err;
1846                 }
1847                 spin_unlock(&ubi->wl_lock);
1848                 up_read(&ubi->fm_eba_sem);
1849                 goto retry;
1850
1851         }
1852         e = wl_get_wle(ubi);
1853         prot_queue_add(ubi, e);
1854         spin_unlock(&ubi->wl_lock);
1855
1856         err = ubi_self_check_all_ff(ubi, e->pnum, ubi->vid_hdr_aloffset,
1857                                     ubi->peb_size - ubi->vid_hdr_aloffset);
1858         if (err) {
1859                 ubi_err(ubi, "new PEB %d does not contain all 0xFF bytes", e->pnum);
1860                 return err;
1861         }
1862
1863         return e->pnum;
1864 }
1865 #else
1866 #include "fastmap-wl.c"
1867 #endif