2 * @ubi: UBI device description object
3 * Copyright (c) International Business Machines Corp., 2006
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See
13 * the GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 * Authors: Artem Bityutskiy (Битюцкий Артём), Thomas Gleixner
23 * UBI wear-leveling sub-system.
25 * This sub-system is responsible for wear-leveling. It works in terms of
26 * physical eraseblocks and erase counters and knows nothing about logical
27 * eraseblocks, volumes, etc. From this sub-system's perspective all physical
28 * eraseblocks are of two types - used and free. Used physical eraseblocks are
29 * those that were "get" by the 'ubi_wl_get_peb()' function, and free physical
30 * eraseblocks are those that were put by the 'ubi_wl_put_peb()' function.
32 * Physical eraseblocks returned by 'ubi_wl_get_peb()' have only erase counter
33 * header. The rest of the physical eraseblock contains only %0xFF bytes.
35 * When physical eraseblocks are returned to the WL sub-system by means of the
36 * 'ubi_wl_put_peb()' function, they are scheduled for erasure. The erasure is
37 * done asynchronously in context of the per-UBI device background thread,
38 * which is also managed by the WL sub-system.
40 * The wear-leveling is ensured by means of moving the contents of used
41 * physical eraseblocks with low erase counter to free physical eraseblocks
42 * with high erase counter.
44 * If the WL sub-system fails to erase a physical eraseblock, it marks it as
47 * This sub-system is also responsible for scrubbing. If a bit-flip is detected
48 * in a physical eraseblock, it has to be moved. Technically this is the same
49 * as moving it for wear-leveling reasons.
51 * As it was said, for the UBI sub-system all physical eraseblocks are either
52 * "free" or "used". Free eraseblock are kept in the @wl->free RB-tree, while
53 * used eraseblocks are kept in @wl->used, @wl->erroneous, or @wl->scrub
54 * RB-trees, as well as (temporarily) in the @wl->pq queue.
56 * When the WL sub-system returns a physical eraseblock, the physical
57 * eraseblock is protected from being moved for some "time". For this reason,
58 * the physical eraseblock is not directly moved from the @wl->free tree to the
59 * @wl->used tree. There is a protection queue in between where this
60 * physical eraseblock is temporarily stored (@wl->pq).
62 * All this protection stuff is needed because:
63 * o we don't want to move physical eraseblocks just after we have given them
64 * to the user; instead, we first want to let users fill them up with data;
66 * o there is a chance that the user will put the physical eraseblock very
67 * soon, so it makes sense not to move it for some time, but wait.
69 * Physical eraseblocks stay protected only for limited time. But the "time" is
70 * measured in erase cycles in this case. This is implemented with help of the
71 * protection queue. Eraseblocks are put to the tail of this queue when they
72 * are returned by the 'ubi_wl_get_peb()', and eraseblocks are removed from the
73 * head of the queue on each erase operation (for any eraseblock). So the
74 * length of the queue defines how may (global) erase cycles PEBs are protected.
76 * To put it differently, each physical eraseblock has 2 main states: free and
77 * used. The former state corresponds to the @wl->free tree. The latter state
78 * is split up on several sub-states:
79 * o the WL movement is allowed (@wl->used tree);
80 * o the WL movement is disallowed (@wl->erroneous) because the PEB is
81 * erroneous - e.g., there was a read error;
82 * o the WL movement is temporarily prohibited (@wl->pq queue);
83 * o scrubbing is needed (@wl->scrub tree).
85 * Depending on the sub-state, wear-leveling entries of the used physical
86 * eraseblocks may be kept in one of those structures.
88 * Note, in this implementation, we keep a small in-RAM object for each physical
89 * eraseblock. This is surely not a scalable solution. But it appears to be good
90 * enough for moderately large flashes and it is simple. In future, one may
91 * re-work this sub-system and make it more scalable.
93 * At the moment this sub-system does not utilize the sequence number, which
94 * was introduced relatively recently. But it would be wise to do this because
95 * the sequence number of a logical eraseblock characterizes how old is it. For
96 * example, when we move a PEB with low erase counter, and we need to pick the
97 * target PEB, we pick a PEB with the highest EC if our PEB is "old" and we
98 * pick target PEB with an average EC if our PEB is not very "old". This is a
99 * room for future re-works of the WL sub-system.
102 #include <linux/slab.h>
103 #include <linux/crc32.h>
104 #include <linux/freezer.h>
105 #include <linux/kthread.h>
108 /* Number of physical eraseblocks reserved for wear-leveling purposes */
109 #define WL_RESERVED_PEBS 1
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
117 #define UBI_WL_THRESHOLD CONFIG_MTD_UBI_WL_THRESHOLD
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.
130 #define WL_FREE_MAX_DIFF (2*UBI_WL_THRESHOLD)
133 * Maximum number of consecutive background thread failures which is enough to
134 * switch to read-only mode.
136 #define WL_MAX_FAILURES 32
139 * struct ubi_work - UBI work description data structure.
140 * @list: a link in the list of pending works
141 * @func: worker function
142 * @e: physical eraseblock to erase
143 * @vol_id: the volume ID on which this erasure is being performed
144 * @lnum: the logical eraseblock number
145 * @torture: if the physical eraseblock has to be tortured
147 * The @func pointer points to the worker function. If the @cancel argument is
148 * not zero, the worker has to free the resources and exit immediately. The
149 * worker has to return zero in case of success and a negative error code in
153 struct list_head list;
154 int (*func)(struct ubi_device *ubi, struct ubi_work *wrk, int cancel);
155 /* The below fields are only relevant to erasure works */
156 struct ubi_wl_entry *e;
162 static int self_check_ec(struct ubi_device *ubi, int pnum, int ec);
163 static int self_check_in_wl_tree(const struct ubi_device *ubi,
164 struct ubi_wl_entry *e, struct rb_root *root);
165 static int self_check_in_pq(const struct ubi_device *ubi,
166 struct ubi_wl_entry *e);
169 * wl_tree_add - add a wear-leveling entry to a WL RB-tree.
170 * @e: the wear-leveling entry to add
171 * @root: the root of the tree
173 * Note, we use (erase counter, physical eraseblock number) pairs as keys in
174 * the @ubi->used and @ubi->free RB-trees.
176 static void wl_tree_add(struct ubi_wl_entry *e, struct rb_root *root)
178 struct rb_node **p, *parent = NULL;
182 struct ubi_wl_entry *e1;
185 e1 = rb_entry(parent, struct ubi_wl_entry, u.rb);
189 else if (e->ec > e1->ec)
192 ubi_assert(e->pnum != e1->pnum);
193 if (e->pnum < e1->pnum)
200 rb_link_node(&e->u.rb, parent, p);
201 rb_insert_color(&e->u.rb, root);
205 * do_work - do one pending work.
206 * @ubi: UBI device description object
208 * This function returns zero in case of success and a negative error code in
211 static int do_work(struct ubi_device *ubi)
214 struct ubi_work *wrk;
219 * @ubi->work_sem is used to synchronize with the workers. Workers take
220 * it in read mode, so many of them may be doing works at a time. But
221 * the queue flush code has to be sure the whole queue of works is
222 * done, and it takes the mutex in write mode.
224 down_read(&ubi->work_sem);
225 spin_lock(&ubi->wl_lock);
226 if (list_empty(&ubi->works)) {
227 spin_unlock(&ubi->wl_lock);
228 up_read(&ubi->work_sem);
232 wrk = list_entry(ubi->works.next, struct ubi_work, list);
233 list_del(&wrk->list);
234 ubi->works_count -= 1;
235 ubi_assert(ubi->works_count >= 0);
236 spin_unlock(&ubi->wl_lock);
239 * Call the worker function. Do not touch the work structure
240 * after this call as it will have been freed or reused by that
241 * time by the worker function.
243 err = wrk->func(ubi, wrk, 0);
245 ubi_err("work failed with error code %d", err);
246 up_read(&ubi->work_sem);
252 * produce_free_peb - produce a free physical eraseblock.
253 * @ubi: UBI device description object
255 * This function tries to make a free PEB by means of synchronous execution of
256 * pending works. This may be needed if, for example the background thread is
257 * disabled. Returns zero in case of success and a negative error code in case
260 static int produce_free_peb(struct ubi_device *ubi)
264 spin_lock(&ubi->wl_lock);
265 while (!ubi->free.rb_node) {
266 spin_unlock(&ubi->wl_lock);
268 dbg_wl("do one work synchronously");
273 spin_lock(&ubi->wl_lock);
275 spin_unlock(&ubi->wl_lock);
281 * in_wl_tree - check if wear-leveling entry is present in a WL RB-tree.
282 * @e: the wear-leveling entry to check
283 * @root: the root of the tree
285 * This function returns non-zero if @e is in the @root RB-tree and zero if it
288 static int in_wl_tree(struct ubi_wl_entry *e, struct rb_root *root)
294 struct ubi_wl_entry *e1;
296 e1 = rb_entry(p, struct ubi_wl_entry, u.rb);
298 if (e->pnum == e1->pnum) {
305 else if (e->ec > e1->ec)
308 ubi_assert(e->pnum != e1->pnum);
309 if (e->pnum < e1->pnum)
320 * prot_queue_add - add physical eraseblock to the protection queue.
321 * @ubi: UBI device description object
322 * @e: the physical eraseblock to add
324 * This function adds @e to the tail of the protection queue @ubi->pq, where
325 * @e will stay for %UBI_PROT_QUEUE_LEN erase operations and will be
326 * temporarily protected from the wear-leveling worker. Note, @wl->lock has to
329 static void prot_queue_add(struct ubi_device *ubi, struct ubi_wl_entry *e)
331 int pq_tail = ubi->pq_head - 1;
334 pq_tail = UBI_PROT_QUEUE_LEN - 1;
335 ubi_assert(pq_tail >= 0 && pq_tail < UBI_PROT_QUEUE_LEN);
336 list_add_tail(&e->u.list, &ubi->pq[pq_tail]);
337 dbg_wl("added PEB %d EC %d to the protection queue", e->pnum, e->ec);
341 * find_wl_entry - find wear-leveling entry closest to certain erase counter.
342 * @root: the RB-tree where to look for
343 * @diff: maximum possible difference from the smallest erase counter
345 * This function looks for a wear leveling entry with erase counter closest to
346 * min + @diff, where min is the smallest erase counter.
348 static struct ubi_wl_entry *find_wl_entry(struct rb_root *root, int diff)
351 struct ubi_wl_entry *e;
354 e = rb_entry(rb_first(root), struct ubi_wl_entry, u.rb);
359 struct ubi_wl_entry *e1;
361 e1 = rb_entry(p, struct ubi_wl_entry, u.rb);
374 * ubi_wl_get_peb - get a physical eraseblock.
375 * @ubi: UBI device description object
377 * This function returns a physical eraseblock in case of success and a
378 * negative error code in case of failure. Might sleep.
380 int ubi_wl_get_peb(struct ubi_device *ubi)
383 struct ubi_wl_entry *e, *first, *last;
386 spin_lock(&ubi->wl_lock);
387 if (!ubi->free.rb_node) {
388 if (ubi->works_count == 0) {
389 ubi_assert(list_empty(&ubi->works));
390 ubi_err("no free eraseblocks");
391 spin_unlock(&ubi->wl_lock);
394 spin_unlock(&ubi->wl_lock);
396 err = produce_free_peb(ubi);
402 first = rb_entry(rb_first(&ubi->free), struct ubi_wl_entry, u.rb);
403 last = rb_entry(rb_last(&ubi->free), struct ubi_wl_entry, u.rb);
405 if (last->ec - first->ec < WL_FREE_MAX_DIFF)
406 e = rb_entry(ubi->free.rb_node, struct ubi_wl_entry, u.rb);
408 e = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF/2);
410 self_check_in_wl_tree(ubi, e, &ubi->free);
413 * Move the physical eraseblock to the protection queue where it will
414 * be protected from being moved for some time.
416 rb_erase(&e->u.rb, &ubi->free);
417 dbg_wl("PEB %d EC %d", e->pnum, e->ec);
418 prot_queue_add(ubi, e);
419 spin_unlock(&ubi->wl_lock);
421 err = ubi_self_check_all_ff(ubi, e->pnum, ubi->vid_hdr_aloffset,
422 ubi->peb_size - ubi->vid_hdr_aloffset);
424 ubi_err("new PEB %d does not contain all 0xFF bytes", e->pnum);
432 * prot_queue_del - remove a physical eraseblock from the protection queue.
433 * @ubi: UBI device description object
434 * @pnum: the physical eraseblock to remove
436 * This function deletes PEB @pnum from the protection queue and returns zero
437 * in case of success and %-ENODEV if the PEB was not found.
439 static int prot_queue_del(struct ubi_device *ubi, int pnum)
441 struct ubi_wl_entry *e;
443 e = ubi->lookuptbl[pnum];
447 if (self_check_in_pq(ubi, e))
450 list_del(&e->u.list);
451 dbg_wl("deleted PEB %d from the protection queue", e->pnum);
456 * sync_erase - synchronously erase a physical eraseblock.
457 * @ubi: UBI device description object
458 * @e: the the physical eraseblock to erase
459 * @torture: if the physical eraseblock has to be tortured
461 * This function returns zero in case of success and a negative error code in
464 static int sync_erase(struct ubi_device *ubi, struct ubi_wl_entry *e,
468 struct ubi_ec_hdr *ec_hdr;
469 unsigned long long ec = e->ec;
471 dbg_wl("erase PEB %d, old EC %llu", e->pnum, ec);
473 err = self_check_ec(ubi, e->pnum, e->ec);
477 ec_hdr = kzalloc(ubi->ec_hdr_alsize, GFP_NOFS);
481 err = ubi_io_sync_erase(ubi, e->pnum, torture);
486 if (ec > UBI_MAX_ERASECOUNTER) {
488 * Erase counter overflow. Upgrade UBI and use 64-bit
489 * erase counters internally.
491 ubi_err("erase counter overflow at PEB %d, EC %llu",
497 dbg_wl("erased PEB %d, new EC %llu", e->pnum, ec);
499 ec_hdr->ec = cpu_to_be64(ec);
501 err = ubi_io_write_ec_hdr(ubi, e->pnum, ec_hdr);
506 spin_lock(&ubi->wl_lock);
507 if (e->ec > ubi->max_ec)
509 spin_unlock(&ubi->wl_lock);
517 * serve_prot_queue - check if it is time to stop protecting PEBs.
518 * @ubi: UBI device description object
520 * This function is called after each erase operation and removes PEBs from the
521 * tail of the protection queue. These PEBs have been protected for long enough
522 * and should be moved to the used tree.
524 static void serve_prot_queue(struct ubi_device *ubi)
526 struct ubi_wl_entry *e, *tmp;
530 * There may be several protected physical eraseblock to remove,
535 spin_lock(&ubi->wl_lock);
536 list_for_each_entry_safe(e, tmp, &ubi->pq[ubi->pq_head], u.list) {
537 dbg_wl("PEB %d EC %d protection over, move to used tree",
540 list_del(&e->u.list);
541 wl_tree_add(e, &ubi->used);
544 * Let's be nice and avoid holding the spinlock for
547 spin_unlock(&ubi->wl_lock);
554 if (ubi->pq_head == UBI_PROT_QUEUE_LEN)
556 ubi_assert(ubi->pq_head >= 0 && ubi->pq_head < UBI_PROT_QUEUE_LEN);
557 spin_unlock(&ubi->wl_lock);
561 * schedule_ubi_work - schedule a work.
562 * @ubi: UBI device description object
563 * @wrk: the work to schedule
565 * This function adds a work defined by @wrk to the tail of the pending works
568 static void schedule_ubi_work(struct ubi_device *ubi, struct ubi_work *wrk)
570 spin_lock(&ubi->wl_lock);
571 list_add_tail(&wrk->list, &ubi->works);
572 ubi_assert(ubi->works_count >= 0);
573 ubi->works_count += 1;
574 if (ubi->thread_enabled && !ubi_dbg_is_bgt_disabled(ubi))
575 wake_up_process(ubi->bgt_thread);
576 spin_unlock(&ubi->wl_lock);
579 static int erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk,
583 * schedule_erase - schedule an erase work.
584 * @ubi: UBI device description object
585 * @e: the WL entry of the physical eraseblock to erase
586 * @vol_id: the volume ID that last used this PEB
587 * @lnum: the last used logical eraseblock number for the PEB
588 * @torture: if the physical eraseblock has to be tortured
590 * This function returns zero in case of success and a %-ENOMEM in case of
593 static int schedule_erase(struct ubi_device *ubi, struct ubi_wl_entry *e,
594 int vol_id, int lnum, int torture)
596 struct ubi_work *wl_wrk;
598 dbg_wl("schedule erasure of PEB %d, EC %d, torture %d",
599 e->pnum, e->ec, torture);
601 wl_wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
605 wl_wrk->func = &erase_worker;
607 wl_wrk->vol_id = vol_id;
609 wl_wrk->torture = torture;
611 schedule_ubi_work(ubi, wl_wrk);
616 * wear_leveling_worker - wear-leveling worker function.
617 * @ubi: UBI device description object
618 * @wrk: the work object
619 * @cancel: non-zero if the worker has to free memory and exit
621 * This function copies a more worn out physical eraseblock to a less worn out
622 * one. Returns zero in case of success and a negative error code in case of
625 static int wear_leveling_worker(struct ubi_device *ubi, struct ubi_work *wrk,
628 int err, scrubbing = 0, torture = 0, protect = 0, erroneous = 0;
629 int vol_id = -1, uninitialized_var(lnum);
630 struct ubi_wl_entry *e1, *e2;
631 struct ubi_vid_hdr *vid_hdr;
637 vid_hdr = ubi_zalloc_vid_hdr(ubi, GFP_NOFS);
641 mutex_lock(&ubi->move_mutex);
642 spin_lock(&ubi->wl_lock);
643 ubi_assert(!ubi->move_from && !ubi->move_to);
644 ubi_assert(!ubi->move_to_put);
646 if (!ubi->free.rb_node ||
647 (!ubi->used.rb_node && !ubi->scrub.rb_node)) {
649 * No free physical eraseblocks? Well, they must be waiting in
650 * the queue to be erased. Cancel movement - it will be
651 * triggered again when a free physical eraseblock appears.
653 * No used physical eraseblocks? They must be temporarily
654 * protected from being moved. They will be moved to the
655 * @ubi->used tree later and the wear-leveling will be
658 dbg_wl("cancel WL, a list is empty: free %d, used %d",
659 !ubi->free.rb_node, !ubi->used.rb_node);
663 if (!ubi->scrub.rb_node) {
665 * Now pick the least worn-out used physical eraseblock and a
666 * highly worn-out free physical eraseblock. If the erase
667 * counters differ much enough, start wear-leveling.
669 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, u.rb);
670 e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
672 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD)) {
673 dbg_wl("no WL needed: min used EC %d, max free EC %d",
677 self_check_in_wl_tree(ubi, e1, &ubi->used);
678 rb_erase(&e1->u.rb, &ubi->used);
679 dbg_wl("move PEB %d EC %d to PEB %d EC %d",
680 e1->pnum, e1->ec, e2->pnum, e2->ec);
682 /* Perform scrubbing */
684 e1 = rb_entry(rb_first(&ubi->scrub), struct ubi_wl_entry, u.rb);
685 e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
686 self_check_in_wl_tree(ubi, e1, &ubi->scrub);
687 rb_erase(&e1->u.rb, &ubi->scrub);
688 dbg_wl("scrub PEB %d to PEB %d", e1->pnum, e2->pnum);
691 self_check_in_wl_tree(ubi, e2, &ubi->free);
692 rb_erase(&e2->u.rb, &ubi->free);
695 spin_unlock(&ubi->wl_lock);
698 * Now we are going to copy physical eraseblock @e1->pnum to @e2->pnum.
699 * We so far do not know which logical eraseblock our physical
700 * eraseblock (@e1) belongs to. We have to read the volume identifier
703 * Note, we are protected from this PEB being unmapped and erased. The
704 * 'ubi_wl_put_peb()' would wait for moving to be finished if the PEB
705 * which is being moved was unmapped.
708 err = ubi_io_read_vid_hdr(ubi, e1->pnum, vid_hdr, 0);
709 if (err && err != UBI_IO_BITFLIPS) {
710 if (err == UBI_IO_FF) {
712 * We are trying to move PEB without a VID header. UBI
713 * always write VID headers shortly after the PEB was
714 * given, so we have a situation when it has not yet
715 * had a chance to write it, because it was preempted.
716 * So add this PEB to the protection queue so far,
717 * because presumably more data will be written there
718 * (including the missing VID header), and then we'll
721 dbg_wl("PEB %d has no VID header", e1->pnum);
724 } else if (err == UBI_IO_FF_BITFLIPS) {
726 * The same situation as %UBI_IO_FF, but bit-flips were
727 * detected. It is better to schedule this PEB for
730 dbg_wl("PEB %d has no VID header but has bit-flips",
736 ubi_err("error %d while reading VID header from PEB %d",
741 vol_id = be32_to_cpu(vid_hdr->vol_id);
742 lnum = be32_to_cpu(vid_hdr->lnum);
744 err = ubi_eba_copy_leb(ubi, e1->pnum, e2->pnum, vid_hdr);
746 if (err == MOVE_CANCEL_RACE) {
748 * The LEB has not been moved because the volume is
749 * being deleted or the PEB has been put meanwhile. We
750 * should prevent this PEB from being selected for
751 * wear-leveling movement again, so put it to the
757 if (err == MOVE_RETRY) {
761 if (err == MOVE_TARGET_BITFLIPS || err == MOVE_TARGET_WR_ERR ||
762 err == MOVE_TARGET_RD_ERR) {
764 * Target PEB had bit-flips or write error - torture it.
770 if (err == MOVE_SOURCE_RD_ERR) {
772 * An error happened while reading the source PEB. Do
773 * not switch to R/O mode in this case, and give the
774 * upper layers a possibility to recover from this,
775 * e.g. by unmapping corresponding LEB. Instead, just
776 * put this PEB to the @ubi->erroneous list to prevent
777 * UBI from trying to move it over and over again.
779 if (ubi->erroneous_peb_count > ubi->max_erroneous) {
780 ubi_err("too many erroneous eraseblocks (%d)",
781 ubi->erroneous_peb_count);
794 /* The PEB has been successfully moved */
796 ubi_msg("scrubbed PEB %d (LEB %d:%d), data moved to PEB %d",
797 e1->pnum, vol_id, lnum, e2->pnum);
798 ubi_free_vid_hdr(ubi, vid_hdr);
800 spin_lock(&ubi->wl_lock);
801 if (!ubi->move_to_put) {
802 wl_tree_add(e2, &ubi->used);
805 ubi->move_from = ubi->move_to = NULL;
806 ubi->move_to_put = ubi->wl_scheduled = 0;
807 spin_unlock(&ubi->wl_lock);
809 err = schedule_erase(ubi, e1, vol_id, lnum, 0);
811 kmem_cache_free(ubi_wl_entry_slab, e1);
813 kmem_cache_free(ubi_wl_entry_slab, e2);
819 * Well, the target PEB was put meanwhile, schedule it for
822 dbg_wl("PEB %d (LEB %d:%d) was put meanwhile, erase",
823 e2->pnum, vol_id, lnum);
824 err = schedule_erase(ubi, e2, vol_id, lnum, 0);
826 kmem_cache_free(ubi_wl_entry_slab, e2);
832 mutex_unlock(&ubi->move_mutex);
836 * For some reasons the LEB was not moved, might be an error, might be
837 * something else. @e1 was not changed, so return it back. @e2 might
838 * have been changed, schedule it for erasure.
842 dbg_wl("cancel moving PEB %d (LEB %d:%d) to PEB %d (%d)",
843 e1->pnum, vol_id, lnum, e2->pnum, err);
845 dbg_wl("cancel moving PEB %d to PEB %d (%d)",
846 e1->pnum, e2->pnum, err);
847 spin_lock(&ubi->wl_lock);
849 prot_queue_add(ubi, e1);
850 else if (erroneous) {
851 wl_tree_add(e1, &ubi->erroneous);
852 ubi->erroneous_peb_count += 1;
853 } else if (scrubbing)
854 wl_tree_add(e1, &ubi->scrub);
856 wl_tree_add(e1, &ubi->used);
857 ubi_assert(!ubi->move_to_put);
858 ubi->move_from = ubi->move_to = NULL;
859 ubi->wl_scheduled = 0;
860 spin_unlock(&ubi->wl_lock);
862 ubi_free_vid_hdr(ubi, vid_hdr);
863 err = schedule_erase(ubi, e2, vol_id, lnum, torture);
865 kmem_cache_free(ubi_wl_entry_slab, e2);
868 mutex_unlock(&ubi->move_mutex);
873 ubi_err("error %d while moving PEB %d to PEB %d",
874 err, e1->pnum, e2->pnum);
876 ubi_err("error %d while moving PEB %d (LEB %d:%d) to PEB %d",
877 err, e1->pnum, vol_id, lnum, e2->pnum);
878 spin_lock(&ubi->wl_lock);
879 ubi->move_from = ubi->move_to = NULL;
880 ubi->move_to_put = ubi->wl_scheduled = 0;
881 spin_unlock(&ubi->wl_lock);
883 ubi_free_vid_hdr(ubi, vid_hdr);
884 kmem_cache_free(ubi_wl_entry_slab, e1);
885 kmem_cache_free(ubi_wl_entry_slab, e2);
889 mutex_unlock(&ubi->move_mutex);
890 ubi_assert(err != 0);
891 return err < 0 ? err : -EIO;
894 ubi->wl_scheduled = 0;
895 spin_unlock(&ubi->wl_lock);
896 mutex_unlock(&ubi->move_mutex);
897 ubi_free_vid_hdr(ubi, vid_hdr);
902 * ensure_wear_leveling - schedule wear-leveling if it is needed.
903 * @ubi: UBI device description object
905 * This function checks if it is time to start wear-leveling and schedules it
906 * if yes. This function returns zero in case of success and a negative error
907 * code in case of failure.
909 static int ensure_wear_leveling(struct ubi_device *ubi)
912 struct ubi_wl_entry *e1;
913 struct ubi_wl_entry *e2;
914 struct ubi_work *wrk;
916 spin_lock(&ubi->wl_lock);
917 if (ubi->wl_scheduled)
918 /* Wear-leveling is already in the work queue */
922 * If the ubi->scrub tree is not empty, scrubbing is needed, and the
923 * the WL worker has to be scheduled anyway.
925 if (!ubi->scrub.rb_node) {
926 if (!ubi->used.rb_node || !ubi->free.rb_node)
927 /* No physical eraseblocks - no deal */
931 * We schedule wear-leveling only if the difference between the
932 * lowest erase counter of used physical eraseblocks and a high
933 * erase counter of free physical eraseblocks is greater than
936 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, u.rb);
937 e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
939 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD))
941 dbg_wl("schedule wear-leveling");
943 dbg_wl("schedule scrubbing");
945 ubi->wl_scheduled = 1;
946 spin_unlock(&ubi->wl_lock);
948 wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
954 wrk->func = &wear_leveling_worker;
955 schedule_ubi_work(ubi, wrk);
959 spin_lock(&ubi->wl_lock);
960 ubi->wl_scheduled = 0;
962 spin_unlock(&ubi->wl_lock);
967 * erase_worker - physical eraseblock erase worker function.
968 * @ubi: UBI device description object
969 * @wl_wrk: the work object
970 * @cancel: non-zero if the worker has to free memory and exit
972 * This function erases a physical eraseblock and perform torture testing if
973 * needed. It also takes care about marking the physical eraseblock bad if
974 * needed. Returns zero in case of success and a negative error code in case of
977 static int erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk,
980 struct ubi_wl_entry *e = wl_wrk->e;
982 int vol_id = wl_wrk->vol_id;
983 int lnum = wl_wrk->lnum;
984 int err, available_consumed = 0;
987 dbg_wl("cancel erasure of PEB %d EC %d", pnum, e->ec);
989 kmem_cache_free(ubi_wl_entry_slab, e);
993 dbg_wl("erase PEB %d EC %d LEB %d:%d",
994 pnum, e->ec, wl_wrk->vol_id, wl_wrk->lnum);
996 err = sync_erase(ubi, e, wl_wrk->torture);
998 /* Fine, we've erased it successfully */
1001 spin_lock(&ubi->wl_lock);
1002 wl_tree_add(e, &ubi->free);
1003 spin_unlock(&ubi->wl_lock);
1006 * One more erase operation has happened, take care about
1007 * protected physical eraseblocks.
1009 serve_prot_queue(ubi);
1011 /* And take care about wear-leveling */
1012 err = ensure_wear_leveling(ubi);
1016 ubi_err("failed to erase PEB %d, error %d", pnum, err);
1019 if (err == -EINTR || err == -ENOMEM || err == -EAGAIN ||
1023 /* Re-schedule the LEB for erasure */
1024 err1 = schedule_erase(ubi, e, vol_id, lnum, 0);
1032 kmem_cache_free(ubi_wl_entry_slab, e);
1035 * If this is not %-EIO, we have no idea what to do. Scheduling
1036 * this physical eraseblock for erasure again would cause
1037 * errors again and again. Well, lets switch to R/O mode.
1041 /* It is %-EIO, the PEB went bad */
1043 if (!ubi->bad_allowed) {
1044 ubi_err("bad physical eraseblock %d detected", pnum);
1048 spin_lock(&ubi->volumes_lock);
1049 if (ubi->beb_rsvd_pebs == 0) {
1050 if (ubi->avail_pebs == 0) {
1051 spin_unlock(&ubi->volumes_lock);
1052 ubi_err("no reserved/available physical eraseblocks");
1055 ubi->avail_pebs -= 1;
1056 available_consumed = 1;
1058 spin_unlock(&ubi->volumes_lock);
1060 ubi_msg("mark PEB %d as bad", pnum);
1061 err = ubi_io_mark_bad(ubi, pnum);
1065 spin_lock(&ubi->volumes_lock);
1066 if (ubi->beb_rsvd_pebs > 0) {
1067 if (available_consumed) {
1069 * The amount of reserved PEBs increased since we last
1072 ubi->avail_pebs += 1;
1073 available_consumed = 0;
1075 ubi->beb_rsvd_pebs -= 1;
1077 ubi->bad_peb_count += 1;
1078 ubi->good_peb_count -= 1;
1079 ubi_calculate_reserved(ubi);
1080 if (available_consumed)
1081 ubi_warn("no PEBs in the reserved pool, used an available PEB");
1082 else if (ubi->beb_rsvd_pebs)
1083 ubi_msg("%d PEBs left in the reserve", ubi->beb_rsvd_pebs);
1085 ubi_warn("last PEB from the reserve was used");
1086 spin_unlock(&ubi->volumes_lock);
1091 if (available_consumed) {
1092 spin_lock(&ubi->volumes_lock);
1093 ubi->avail_pebs += 1;
1094 spin_unlock(&ubi->volumes_lock);
1101 * ubi_wl_put_peb - return a PEB to the wear-leveling sub-system.
1102 * @ubi: UBI device description object
1103 * @vol_id: the volume ID that last used this PEB
1104 * @lnum: the last used logical eraseblock number for the PEB
1105 * @pnum: physical eraseblock to return
1106 * @torture: if this physical eraseblock has to be tortured
1108 * This function is called to return physical eraseblock @pnum to the pool of
1109 * free physical eraseblocks. The @torture flag has to be set if an I/O error
1110 * occurred to this @pnum and it has to be tested. This function returns zero
1111 * in case of success, and a negative error code in case of failure.
1113 int ubi_wl_put_peb(struct ubi_device *ubi, int vol_id, int lnum,
1114 int pnum, int torture)
1117 struct ubi_wl_entry *e;
1119 dbg_wl("PEB %d", pnum);
1120 ubi_assert(pnum >= 0);
1121 ubi_assert(pnum < ubi->peb_count);
1124 spin_lock(&ubi->wl_lock);
1125 e = ubi->lookuptbl[pnum];
1126 if (e == ubi->move_from) {
1128 * User is putting the physical eraseblock which was selected to
1129 * be moved. It will be scheduled for erasure in the
1130 * wear-leveling worker.
1132 dbg_wl("PEB %d is being moved, wait", pnum);
1133 spin_unlock(&ubi->wl_lock);
1135 /* Wait for the WL worker by taking the @ubi->move_mutex */
1136 mutex_lock(&ubi->move_mutex);
1137 mutex_unlock(&ubi->move_mutex);
1139 } else if (e == ubi->move_to) {
1141 * User is putting the physical eraseblock which was selected
1142 * as the target the data is moved to. It may happen if the EBA
1143 * sub-system already re-mapped the LEB in 'ubi_eba_copy_leb()'
1144 * but the WL sub-system has not put the PEB to the "used" tree
1145 * yet, but it is about to do this. So we just set a flag which
1146 * will tell the WL worker that the PEB is not needed anymore
1147 * and should be scheduled for erasure.
1149 dbg_wl("PEB %d is the target of data moving", pnum);
1150 ubi_assert(!ubi->move_to_put);
1151 ubi->move_to_put = 1;
1152 spin_unlock(&ubi->wl_lock);
1155 if (in_wl_tree(e, &ubi->used)) {
1156 self_check_in_wl_tree(ubi, e, &ubi->used);
1157 rb_erase(&e->u.rb, &ubi->used);
1158 } else if (in_wl_tree(e, &ubi->scrub)) {
1159 self_check_in_wl_tree(ubi, e, &ubi->scrub);
1160 rb_erase(&e->u.rb, &ubi->scrub);
1161 } else if (in_wl_tree(e, &ubi->erroneous)) {
1162 self_check_in_wl_tree(ubi, e, &ubi->erroneous);
1163 rb_erase(&e->u.rb, &ubi->erroneous);
1164 ubi->erroneous_peb_count -= 1;
1165 ubi_assert(ubi->erroneous_peb_count >= 0);
1166 /* Erroneous PEBs should be tortured */
1169 err = prot_queue_del(ubi, e->pnum);
1171 ubi_err("PEB %d not found", pnum);
1173 spin_unlock(&ubi->wl_lock);
1178 spin_unlock(&ubi->wl_lock);
1180 err = schedule_erase(ubi, e, vol_id, lnum, torture);
1182 spin_lock(&ubi->wl_lock);
1183 wl_tree_add(e, &ubi->used);
1184 spin_unlock(&ubi->wl_lock);
1191 * ubi_wl_scrub_peb - schedule a physical eraseblock for scrubbing.
1192 * @ubi: UBI device description object
1193 * @pnum: the physical eraseblock to schedule
1195 * If a bit-flip in a physical eraseblock is detected, this physical eraseblock
1196 * needs scrubbing. This function schedules a physical eraseblock for
1197 * scrubbing which is done in background. This function returns zero in case of
1198 * success and a negative error code in case of failure.
1200 int ubi_wl_scrub_peb(struct ubi_device *ubi, int pnum)
1202 struct ubi_wl_entry *e;
1204 ubi_msg("schedule PEB %d for scrubbing", pnum);
1207 spin_lock(&ubi->wl_lock);
1208 e = ubi->lookuptbl[pnum];
1209 if (e == ubi->move_from || in_wl_tree(e, &ubi->scrub) ||
1210 in_wl_tree(e, &ubi->erroneous)) {
1211 spin_unlock(&ubi->wl_lock);
1215 if (e == ubi->move_to) {
1217 * This physical eraseblock was used to move data to. The data
1218 * was moved but the PEB was not yet inserted to the proper
1219 * tree. We should just wait a little and let the WL worker
1222 spin_unlock(&ubi->wl_lock);
1223 dbg_wl("the PEB %d is not in proper tree, retry", pnum);
1228 if (in_wl_tree(e, &ubi->used)) {
1229 self_check_in_wl_tree(ubi, e, &ubi->used);
1230 rb_erase(&e->u.rb, &ubi->used);
1234 err = prot_queue_del(ubi, e->pnum);
1236 ubi_err("PEB %d not found", pnum);
1238 spin_unlock(&ubi->wl_lock);
1243 wl_tree_add(e, &ubi->scrub);
1244 spin_unlock(&ubi->wl_lock);
1247 * Technically scrubbing is the same as wear-leveling, so it is done
1250 return ensure_wear_leveling(ubi);
1254 * ubi_wl_flush - flush all pending works.
1255 * @ubi: UBI device description object
1256 * @vol_id: the volume id to flush for
1257 * @lnum: the logical eraseblock number to flush for
1259 * This function executes all pending works for a particular volume id /
1260 * logical eraseblock number pair. If either value is set to %UBI_ALL, then it
1261 * acts as a wildcard for all of the corresponding volume numbers or logical
1262 * eraseblock numbers. It returns zero in case of success and a negative error
1263 * code in case of failure.
1265 int ubi_wl_flush(struct ubi_device *ubi, int vol_id, int lnum)
1271 * Erase while the pending works queue is not empty, but not more than
1272 * the number of currently pending works.
1274 dbg_wl("flush pending work for LEB %d:%d (%d pending works)",
1275 vol_id, lnum, ubi->works_count);
1278 struct ubi_work *wrk;
1281 down_read(&ubi->work_sem);
1282 spin_lock(&ubi->wl_lock);
1283 list_for_each_entry(wrk, &ubi->works, list) {
1284 if ((vol_id == UBI_ALL || wrk->vol_id == vol_id) &&
1285 (lnum == UBI_ALL || wrk->lnum == lnum)) {
1286 list_del(&wrk->list);
1287 ubi->works_count -= 1;
1288 ubi_assert(ubi->works_count >= 0);
1289 spin_unlock(&ubi->wl_lock);
1291 err = wrk->func(ubi, wrk, 0);
1293 up_read(&ubi->work_sem);
1297 spin_lock(&ubi->wl_lock);
1302 spin_unlock(&ubi->wl_lock);
1303 up_read(&ubi->work_sem);
1307 * Make sure all the works which have been done in parallel are
1310 down_write(&ubi->work_sem);
1311 up_write(&ubi->work_sem);
1317 * tree_destroy - destroy an RB-tree.
1318 * @root: the root of the tree to destroy
1320 static void tree_destroy(struct rb_root *root)
1323 struct ubi_wl_entry *e;
1329 else if (rb->rb_right)
1332 e = rb_entry(rb, struct ubi_wl_entry, u.rb);
1336 if (rb->rb_left == &e->u.rb)
1339 rb->rb_right = NULL;
1342 kmem_cache_free(ubi_wl_entry_slab, e);
1348 * ubi_thread - UBI background thread.
1349 * @u: the UBI device description object pointer
1351 int ubi_thread(void *u)
1354 struct ubi_device *ubi = u;
1356 ubi_msg("background thread \"%s\" started, PID %d",
1357 ubi->bgt_name, task_pid_nr(current));
1363 if (kthread_should_stop())
1366 if (try_to_freeze())
1369 spin_lock(&ubi->wl_lock);
1370 if (list_empty(&ubi->works) || ubi->ro_mode ||
1371 !ubi->thread_enabled || ubi_dbg_is_bgt_disabled(ubi)) {
1372 set_current_state(TASK_INTERRUPTIBLE);
1373 spin_unlock(&ubi->wl_lock);
1377 spin_unlock(&ubi->wl_lock);
1381 ubi_err("%s: work failed with error code %d",
1382 ubi->bgt_name, err);
1383 if (failures++ > WL_MAX_FAILURES) {
1385 * Too many failures, disable the thread and
1386 * switch to read-only mode.
1388 ubi_msg("%s: %d consecutive failures",
1389 ubi->bgt_name, WL_MAX_FAILURES);
1391 ubi->thread_enabled = 0;
1400 dbg_wl("background thread \"%s\" is killed", ubi->bgt_name);
1405 * cancel_pending - cancel all pending works.
1406 * @ubi: UBI device description object
1408 static void cancel_pending(struct ubi_device *ubi)
1410 while (!list_empty(&ubi->works)) {
1411 struct ubi_work *wrk;
1413 wrk = list_entry(ubi->works.next, struct ubi_work, list);
1414 list_del(&wrk->list);
1415 wrk->func(ubi, wrk, 1);
1416 ubi->works_count -= 1;
1417 ubi_assert(ubi->works_count >= 0);
1422 * ubi_wl_init - initialize the WL sub-system using attaching information.
1423 * @ubi: UBI device description object
1424 * @ai: attaching information
1426 * This function returns zero in case of success, and a negative error code in
1429 int ubi_wl_init(struct ubi_device *ubi, struct ubi_attach_info *ai)
1432 struct rb_node *rb1, *rb2;
1433 struct ubi_ainf_volume *av;
1434 struct ubi_ainf_peb *aeb, *tmp;
1435 struct ubi_wl_entry *e;
1437 ubi->used = ubi->erroneous = ubi->free = ubi->scrub = RB_ROOT;
1438 spin_lock_init(&ubi->wl_lock);
1439 mutex_init(&ubi->move_mutex);
1440 init_rwsem(&ubi->work_sem);
1441 ubi->max_ec = ai->max_ec;
1442 INIT_LIST_HEAD(&ubi->works);
1444 sprintf(ubi->bgt_name, UBI_BGT_NAME_PATTERN, ubi->ubi_num);
1447 ubi->lookuptbl = kzalloc(ubi->peb_count * sizeof(void *), GFP_KERNEL);
1448 if (!ubi->lookuptbl)
1451 for (i = 0; i < UBI_PROT_QUEUE_LEN; i++)
1452 INIT_LIST_HEAD(&ubi->pq[i]);
1455 list_for_each_entry_safe(aeb, tmp, &ai->erase, u.list) {
1458 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1462 e->pnum = aeb->pnum;
1464 ubi->lookuptbl[e->pnum] = e;
1465 if (schedule_erase(ubi, e, aeb->vol_id, aeb->lnum, 0)) {
1466 kmem_cache_free(ubi_wl_entry_slab, e);
1471 list_for_each_entry(aeb, &ai->free, u.list) {
1474 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1478 e->pnum = aeb->pnum;
1480 ubi_assert(e->ec >= 0);
1481 wl_tree_add(e, &ubi->free);
1482 ubi->lookuptbl[e->pnum] = e;
1485 ubi_rb_for_each_entry(rb1, av, &ai->volumes, rb) {
1486 ubi_rb_for_each_entry(rb2, aeb, &av->root, u.rb) {
1489 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1493 e->pnum = aeb->pnum;
1495 ubi->lookuptbl[e->pnum] = e;
1497 dbg_wl("add PEB %d EC %d to the used tree",
1499 wl_tree_add(e, &ubi->used);
1501 dbg_wl("add PEB %d EC %d to the scrub tree",
1503 wl_tree_add(e, &ubi->scrub);
1508 if (ubi->avail_pebs < WL_RESERVED_PEBS) {
1509 ubi_err("no enough physical eraseblocks (%d, need %d)",
1510 ubi->avail_pebs, WL_RESERVED_PEBS);
1511 if (ubi->corr_peb_count)
1512 ubi_err("%d PEBs are corrupted and not used",
1513 ubi->corr_peb_count);
1516 ubi->avail_pebs -= WL_RESERVED_PEBS;
1517 ubi->rsvd_pebs += WL_RESERVED_PEBS;
1519 /* Schedule wear-leveling if needed */
1520 err = ensure_wear_leveling(ubi);
1527 cancel_pending(ubi);
1528 tree_destroy(&ubi->used);
1529 tree_destroy(&ubi->free);
1530 tree_destroy(&ubi->scrub);
1531 kfree(ubi->lookuptbl);
1536 * protection_queue_destroy - destroy the protection queue.
1537 * @ubi: UBI device description object
1539 static void protection_queue_destroy(struct ubi_device *ubi)
1542 struct ubi_wl_entry *e, *tmp;
1544 for (i = 0; i < UBI_PROT_QUEUE_LEN; ++i) {
1545 list_for_each_entry_safe(e, tmp, &ubi->pq[i], u.list) {
1546 list_del(&e->u.list);
1547 kmem_cache_free(ubi_wl_entry_slab, e);
1553 * ubi_wl_close - close the wear-leveling sub-system.
1554 * @ubi: UBI device description object
1556 void ubi_wl_close(struct ubi_device *ubi)
1558 dbg_wl("close the WL sub-system");
1559 cancel_pending(ubi);
1560 protection_queue_destroy(ubi);
1561 tree_destroy(&ubi->used);
1562 tree_destroy(&ubi->erroneous);
1563 tree_destroy(&ubi->free);
1564 tree_destroy(&ubi->scrub);
1565 kfree(ubi->lookuptbl);
1569 * self_check_ec - make sure that the erase counter of a PEB is correct.
1570 * @ubi: UBI device description object
1571 * @pnum: the physical eraseblock number to check
1572 * @ec: the erase counter to check
1574 * This function returns zero if the erase counter of physical eraseblock @pnum
1575 * is equivalent to @ec, and a negative error code if not or if an error
1578 static int self_check_ec(struct ubi_device *ubi, int pnum, int ec)
1582 struct ubi_ec_hdr *ec_hdr;
1584 if (!ubi->dbg->chk_gen)
1587 ec_hdr = kzalloc(ubi->ec_hdr_alsize, GFP_NOFS);
1591 err = ubi_io_read_ec_hdr(ubi, pnum, ec_hdr, 0);
1592 if (err && err != UBI_IO_BITFLIPS) {
1593 /* The header does not have to exist */
1598 read_ec = be64_to_cpu(ec_hdr->ec);
1599 if (ec != read_ec) {
1600 ubi_err("self-check failed for PEB %d", pnum);
1601 ubi_err("read EC is %lld, should be %d", read_ec, ec);
1613 * self_check_in_wl_tree - check that wear-leveling entry is in WL RB-tree.
1614 * @ubi: UBI device description object
1615 * @e: the wear-leveling entry to check
1616 * @root: the root of the tree
1618 * This function returns zero if @e is in the @root RB-tree and %-EINVAL if it
1621 static int self_check_in_wl_tree(const struct ubi_device *ubi,
1622 struct ubi_wl_entry *e, struct rb_root *root)
1624 if (!ubi->dbg->chk_gen)
1627 if (in_wl_tree(e, root))
1630 ubi_err("self-check failed for PEB %d, EC %d, RB-tree %p ",
1631 e->pnum, e->ec, root);
1637 * self_check_in_pq - check if wear-leveling entry is in the protection
1639 * @ubi: UBI device description object
1640 * @e: the wear-leveling entry to check
1642 * This function returns zero if @e is in @ubi->pq and %-EINVAL if it is not.
1644 static int self_check_in_pq(const struct ubi_device *ubi,
1645 struct ubi_wl_entry *e)
1647 struct ubi_wl_entry *p;
1650 if (!ubi->dbg->chk_gen)
1653 for (i = 0; i < UBI_PROT_QUEUE_LEN; ++i)
1654 list_for_each_entry(p, &ubi->pq[i], u.list)
1658 ubi_err("self-check failed for PEB %d, EC %d, Protect queue",